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/ADT/SmallPtrSet.h"
17 #include "llvm/Analysis/AssumptionCache.h"
18 #include "llvm/Analysis/InstructionSimplify.h"
19 #include "llvm/Analysis/MemoryBuiltins.h"
20 #include "llvm/IR/CallSite.h"
21 #include "llvm/IR/ConstantRange.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/Dominators.h"
25 #include "llvm/IR/GetElementPtrTypeIterator.h"
26 #include "llvm/IR/GlobalAlias.h"
27 #include "llvm/IR/GlobalVariable.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/IntrinsicInst.h"
30 #include "llvm/IR/LLVMContext.h"
31 #include "llvm/IR/Metadata.h"
32 #include "llvm/IR/Operator.h"
33 #include "llvm/IR/PatternMatch.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/MathExtras.h"
38 using namespace llvm::PatternMatch;
40 const unsigned MaxDepth = 6;
42 /// Returns the bitwidth of the given scalar or pointer type (if unknown returns
43 /// 0). For vector types, returns the element type's bitwidth.
44 static unsigned getBitWidth(Type *Ty, const DataLayout *TD) {
45 if (unsigned BitWidth = Ty->getScalarSizeInBits())
48 return TD ? TD->getPointerTypeSizeInBits(Ty) : 0;
51 // Many of these functions have internal versions that take an assumption
52 // exclusion set. This is because of the potential for mutual recursion to
53 // cause computeKnownBits to repeatedly visit the same assume intrinsic. The
54 // classic case of this is assume(x = y), which will attempt to determine
55 // bits in x from bits in y, which will attempt to determine bits in y from
56 // bits in x, etc. Regarding the mutual recursion, computeKnownBits can call
57 // isKnownNonZero, which calls computeKnownBits and ComputeSignBit and
58 // isKnownToBeAPowerOfTwo (all of which can call computeKnownBits), and so on.
59 typedef SmallPtrSet<const Value *, 8> ExclInvsSet;
62 // Simplifying using an assume can only be done in a particular control-flow
63 // context (the context instruction provides that context). If an assume and
64 // the context instruction are not in the same block then the DT helps in
65 // figuring out if we can use it.
69 const Instruction *CxtI;
70 const DominatorTree *DT;
72 Query(AssumptionCache *AC = nullptr, const Instruction *CxtI = nullptr,
73 const DominatorTree *DT = nullptr)
74 : AC(AC), CxtI(CxtI), DT(DT) {}
76 Query(const Query &Q, const Value *NewExcl)
77 : ExclInvs(Q.ExclInvs), AC(Q.AC), CxtI(Q.CxtI), DT(Q.DT) {
78 ExclInvs.insert(NewExcl);
81 } // end anonymous namespace
83 // Given the provided Value and, potentially, a context instruction, return
84 // the preferred context instruction (if any).
85 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) {
86 // If we've been provided with a context instruction, then use that (provided
87 // it has been inserted).
88 if (CxtI && CxtI->getParent())
91 // If the value is really an already-inserted instruction, then use that.
92 CxtI = dyn_cast<Instruction>(V);
93 if (CxtI && CxtI->getParent())
99 static void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
100 const DataLayout *TD, unsigned Depth,
103 void llvm::computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
104 const DataLayout *TD, unsigned Depth,
105 AssumptionCache *AC, const Instruction *CxtI,
106 const DominatorTree *DT) {
107 ::computeKnownBits(V, KnownZero, KnownOne, TD, Depth,
108 Query(AC, safeCxtI(V, CxtI), DT));
111 static void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
112 const DataLayout *TD, unsigned Depth,
115 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
116 const DataLayout *TD, unsigned Depth,
117 AssumptionCache *AC, const Instruction *CxtI,
118 const DominatorTree *DT) {
119 ::ComputeSignBit(V, KnownZero, KnownOne, TD, Depth,
120 Query(AC, safeCxtI(V, CxtI), DT));
123 static bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
126 bool llvm::isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
127 AssumptionCache *AC, const Instruction *CxtI,
128 const DominatorTree *DT) {
129 return ::isKnownToBeAPowerOfTwo(V, OrZero, Depth,
130 Query(AC, safeCxtI(V, CxtI), DT));
133 static bool isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth,
136 bool llvm::isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth,
137 AssumptionCache *AC, const Instruction *CxtI,
138 const DominatorTree *DT) {
139 return ::isKnownNonZero(V, TD, Depth, Query(AC, safeCxtI(V, CxtI), DT));
142 static bool MaskedValueIsZero(Value *V, const APInt &Mask,
143 const DataLayout *TD, unsigned Depth,
146 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout *TD,
147 unsigned Depth, AssumptionCache *AC,
148 const Instruction *CxtI, const DominatorTree *DT) {
149 return ::MaskedValueIsZero(V, Mask, TD, Depth,
150 Query(AC, safeCxtI(V, CxtI), DT));
153 static unsigned ComputeNumSignBits(Value *V, const DataLayout *TD,
154 unsigned Depth, const Query &Q);
156 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout *TD,
157 unsigned Depth, AssumptionCache *AC,
158 const Instruction *CxtI,
159 const DominatorTree *DT) {
160 return ::ComputeNumSignBits(V, TD, Depth, Query(AC, safeCxtI(V, CxtI), DT));
163 static void computeKnownBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW,
164 APInt &KnownZero, APInt &KnownOne,
165 APInt &KnownZero2, APInt &KnownOne2,
166 const DataLayout *TD, unsigned Depth,
169 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(Op0)) {
170 // We know that the top bits of C-X are clear if X contains less bits
171 // than C (i.e. no wrap-around can happen). For example, 20-X is
172 // positive if we can prove that X is >= 0 and < 16.
173 if (!CLHS->getValue().isNegative()) {
174 unsigned BitWidth = KnownZero.getBitWidth();
175 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
176 // NLZ can't be BitWidth with no sign bit
177 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
178 computeKnownBits(Op1, KnownZero2, KnownOne2, TD, Depth+1, Q);
180 // If all of the MaskV bits are known to be zero, then we know the
181 // output top bits are zero, because we now know that the output is
183 if ((KnownZero2 & MaskV) == MaskV) {
184 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
185 // Top bits known zero.
186 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2);
192 unsigned BitWidth = KnownZero.getBitWidth();
194 // If an initial sequence of bits in the result is not needed, the
195 // corresponding bits in the operands are not needed.
196 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
197 computeKnownBits(Op0, LHSKnownZero, LHSKnownOne, TD, Depth+1, Q);
198 computeKnownBits(Op1, KnownZero2, KnownOne2, TD, Depth+1, Q);
200 // Carry in a 1 for a subtract, rather than a 0.
201 APInt CarryIn(BitWidth, 0);
203 // Sum = LHS + ~RHS + 1
204 std::swap(KnownZero2, KnownOne2);
208 APInt PossibleSumZero = ~LHSKnownZero + ~KnownZero2 + CarryIn;
209 APInt PossibleSumOne = LHSKnownOne + KnownOne2 + CarryIn;
211 // Compute known bits of the carry.
212 APInt CarryKnownZero = ~(PossibleSumZero ^ LHSKnownZero ^ KnownZero2);
213 APInt CarryKnownOne = PossibleSumOne ^ LHSKnownOne ^ KnownOne2;
215 // Compute set of known bits (where all three relevant bits are known).
216 APInt LHSKnown = LHSKnownZero | LHSKnownOne;
217 APInt RHSKnown = KnownZero2 | KnownOne2;
218 APInt CarryKnown = CarryKnownZero | CarryKnownOne;
219 APInt Known = LHSKnown & RHSKnown & CarryKnown;
221 assert((PossibleSumZero & Known) == (PossibleSumOne & Known) &&
222 "known bits of sum differ");
224 // Compute known bits of the result.
225 KnownZero = ~PossibleSumOne & Known;
226 KnownOne = PossibleSumOne & Known;
228 // Are we still trying to solve for the sign bit?
229 if (!Known.isNegative()) {
231 // Adding two non-negative numbers, or subtracting a negative number from
232 // a non-negative one, can't wrap into negative.
233 if (LHSKnownZero.isNegative() && KnownZero2.isNegative())
234 KnownZero |= APInt::getSignBit(BitWidth);
235 // Adding two negative numbers, or subtracting a non-negative number from
236 // a negative one, can't wrap into non-negative.
237 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative())
238 KnownOne |= APInt::getSignBit(BitWidth);
243 static void computeKnownBitsMul(Value *Op0, Value *Op1, bool NSW,
244 APInt &KnownZero, APInt &KnownOne,
245 APInt &KnownZero2, APInt &KnownOne2,
246 const DataLayout *TD, unsigned Depth,
248 unsigned BitWidth = KnownZero.getBitWidth();
249 computeKnownBits(Op1, KnownZero, KnownOne, TD, Depth+1, Q);
250 computeKnownBits(Op0, KnownZero2, KnownOne2, TD, Depth+1, Q);
252 bool isKnownNegative = false;
253 bool isKnownNonNegative = false;
254 // If the multiplication is known not to overflow, compute the sign bit.
257 // The product of a number with itself is non-negative.
258 isKnownNonNegative = true;
260 bool isKnownNonNegativeOp1 = KnownZero.isNegative();
261 bool isKnownNonNegativeOp0 = KnownZero2.isNegative();
262 bool isKnownNegativeOp1 = KnownOne.isNegative();
263 bool isKnownNegativeOp0 = KnownOne2.isNegative();
264 // The product of two numbers with the same sign is non-negative.
265 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
266 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
267 // The product of a negative number and a non-negative number is either
269 if (!isKnownNonNegative)
270 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
271 isKnownNonZero(Op0, TD, Depth, Q)) ||
272 (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
273 isKnownNonZero(Op1, TD, Depth, Q));
277 // If low bits are zero in either operand, output low known-0 bits.
278 // Also compute a conserative estimate for high known-0 bits.
279 // More trickiness is possible, but this is sufficient for the
280 // interesting case of alignment computation.
281 KnownOne.clearAllBits();
282 unsigned TrailZ = KnownZero.countTrailingOnes() +
283 KnownZero2.countTrailingOnes();
284 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
285 KnownZero2.countLeadingOnes(),
286 BitWidth) - BitWidth;
288 TrailZ = std::min(TrailZ, BitWidth);
289 LeadZ = std::min(LeadZ, BitWidth);
290 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
291 APInt::getHighBitsSet(BitWidth, LeadZ);
293 // Only make use of no-wrap flags if we failed to compute the sign bit
294 // directly. This matters if the multiplication always overflows, in
295 // which case we prefer to follow the result of the direct computation,
296 // though as the program is invoking undefined behaviour we can choose
297 // whatever we like here.
298 if (isKnownNonNegative && !KnownOne.isNegative())
299 KnownZero.setBit(BitWidth - 1);
300 else if (isKnownNegative && !KnownZero.isNegative())
301 KnownOne.setBit(BitWidth - 1);
304 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
306 unsigned BitWidth = KnownZero.getBitWidth();
307 unsigned NumRanges = Ranges.getNumOperands() / 2;
308 assert(NumRanges >= 1);
310 // Use the high end of the ranges to find leading zeros.
311 unsigned MinLeadingZeros = BitWidth;
312 for (unsigned i = 0; i < NumRanges; ++i) {
314 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
316 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
317 ConstantRange Range(Lower->getValue(), Upper->getValue());
318 if (Range.isWrappedSet())
319 MinLeadingZeros = 0; // -1 has no zeros
320 unsigned LeadingZeros = (Upper->getValue() - 1).countLeadingZeros();
321 MinLeadingZeros = std::min(LeadingZeros, MinLeadingZeros);
324 KnownZero = APInt::getHighBitsSet(BitWidth, MinLeadingZeros);
327 static bool isEphemeralValueOf(Instruction *I, const Value *E) {
328 SmallVector<const Value *, 16> WorkSet(1, I);
329 SmallPtrSet<const Value *, 32> Visited;
330 SmallPtrSet<const Value *, 16> EphValues;
332 while (!WorkSet.empty()) {
333 const Value *V = WorkSet.pop_back_val();
334 if (!Visited.insert(V).second)
337 // If all uses of this value are ephemeral, then so is this value.
338 bool FoundNEUse = false;
339 for (const User *I : V->users())
340 if (!EphValues.count(I)) {
350 if (const User *U = dyn_cast<User>(V))
351 for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
353 if (isSafeToSpeculativelyExecute(*J))
354 WorkSet.push_back(*J);
362 // Is this an intrinsic that cannot be speculated but also cannot trap?
363 static bool isAssumeLikeIntrinsic(const Instruction *I) {
364 if (const CallInst *CI = dyn_cast<CallInst>(I))
365 if (Function *F = CI->getCalledFunction())
366 switch (F->getIntrinsicID()) {
368 // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
369 case Intrinsic::assume:
370 case Intrinsic::dbg_declare:
371 case Intrinsic::dbg_value:
372 case Intrinsic::invariant_start:
373 case Intrinsic::invariant_end:
374 case Intrinsic::lifetime_start:
375 case Intrinsic::lifetime_end:
376 case Intrinsic::objectsize:
377 case Intrinsic::ptr_annotation:
378 case Intrinsic::var_annotation:
385 static bool isValidAssumeForContext(Value *V, const Query &Q,
386 const DataLayout *DL) {
387 Instruction *Inv = cast<Instruction>(V);
389 // There are two restrictions on the use of an assume:
390 // 1. The assume must dominate the context (or the control flow must
391 // reach the assume whenever it reaches the context).
392 // 2. The context must not be in the assume's set of ephemeral values
393 // (otherwise we will use the assume to prove that the condition
394 // feeding the assume is trivially true, thus causing the removal of
398 if (Q.DT->dominates(Inv, Q.CxtI)) {
400 } else if (Inv->getParent() == Q.CxtI->getParent()) {
401 // The context comes first, but they're both in the same block. Make sure
402 // there is nothing in between that might interrupt the control flow.
403 for (BasicBlock::const_iterator I =
404 std::next(BasicBlock::const_iterator(Q.CxtI)),
405 IE(Inv); I != IE; ++I)
406 if (!isSafeToSpeculativelyExecute(I, DL) &&
407 !isAssumeLikeIntrinsic(I))
410 return !isEphemeralValueOf(Inv, Q.CxtI);
416 // When we don't have a DT, we do a limited search...
417 if (Inv->getParent() == Q.CxtI->getParent()->getSinglePredecessor()) {
419 } else if (Inv->getParent() == Q.CxtI->getParent()) {
420 // Search forward from the assume until we reach the context (or the end
421 // of the block); the common case is that the assume will come first.
422 for (BasicBlock::iterator I = std::next(BasicBlock::iterator(Inv)),
423 IE = Inv->getParent()->end(); I != IE; ++I)
427 // The context must come first...
428 for (BasicBlock::const_iterator I =
429 std::next(BasicBlock::const_iterator(Q.CxtI)),
430 IE(Inv); I != IE; ++I)
431 if (!isSafeToSpeculativelyExecute(I, DL) &&
432 !isAssumeLikeIntrinsic(I))
435 return !isEphemeralValueOf(Inv, Q.CxtI);
441 bool llvm::isValidAssumeForContext(const Instruction *I,
442 const Instruction *CxtI,
443 const DataLayout *DL,
444 const DominatorTree *DT) {
445 return ::isValidAssumeForContext(const_cast<Instruction*>(I),
446 Query(nullptr, CxtI, DT), DL);
449 template<typename LHS, typename RHS>
450 inline match_combine_or<CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>,
451 CmpClass_match<RHS, LHS, ICmpInst, ICmpInst::Predicate>>
452 m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
453 return m_CombineOr(m_ICmp(Pred, L, R), m_ICmp(Pred, R, L));
456 template<typename LHS, typename RHS>
457 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::And>,
458 BinaryOp_match<RHS, LHS, Instruction::And>>
459 m_c_And(const LHS &L, const RHS &R) {
460 return m_CombineOr(m_And(L, R), m_And(R, L));
463 template<typename LHS, typename RHS>
464 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Or>,
465 BinaryOp_match<RHS, LHS, Instruction::Or>>
466 m_c_Or(const LHS &L, const RHS &R) {
467 return m_CombineOr(m_Or(L, R), m_Or(R, L));
470 template<typename LHS, typename RHS>
471 inline match_combine_or<BinaryOp_match<LHS, RHS, Instruction::Xor>,
472 BinaryOp_match<RHS, LHS, Instruction::Xor>>
473 m_c_Xor(const LHS &L, const RHS &R) {
474 return m_CombineOr(m_Xor(L, R), m_Xor(R, L));
477 static void computeKnownBitsFromAssume(Value *V, APInt &KnownZero,
479 const DataLayout *DL,
480 unsigned Depth, const Query &Q) {
481 // Use of assumptions is context-sensitive. If we don't have a context, we
483 if (!Q.AC || !Q.CxtI)
486 unsigned BitWidth = KnownZero.getBitWidth();
488 for (auto &AssumeVH : Q.AC->assumptions()) {
491 CallInst *I = cast<CallInst>(AssumeVH);
492 assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
493 "Got assumption for the wrong function!");
494 if (Q.ExclInvs.count(I))
497 // Warning: This loop can end up being somewhat performance sensetive.
498 // We're running this loop for once for each value queried resulting in a
499 // runtime of ~O(#assumes * #values).
501 assert(isa<IntrinsicInst>(I) &&
502 dyn_cast<IntrinsicInst>(I)->getIntrinsicID() == Intrinsic::assume &&
503 "must be an assume intrinsic");
505 Value *Arg = I->getArgOperand(0);
508 isValidAssumeForContext(I, Q, DL)) {
509 assert(BitWidth == 1 && "assume operand is not i1?");
510 KnownZero.clearAllBits();
511 KnownOne.setAllBits();
515 // The remaining tests are all recursive, so bail out if we hit the limit.
516 if (Depth == MaxDepth)
520 auto m_V = m_CombineOr(m_Specific(V),
521 m_CombineOr(m_PtrToInt(m_Specific(V)),
522 m_BitCast(m_Specific(V))));
524 CmpInst::Predicate Pred;
527 if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
528 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
529 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
530 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
531 KnownZero |= RHSKnownZero;
532 KnownOne |= RHSKnownOne;
534 } else if (match(Arg, m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)),
536 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
537 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
538 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
539 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
540 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
542 // For those bits in the mask that are known to be one, we can propagate
543 // known bits from the RHS to V.
544 KnownZero |= RHSKnownZero & MaskKnownOne;
545 KnownOne |= RHSKnownOne & MaskKnownOne;
546 // assume(~(v & b) = a)
547 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
549 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
550 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
551 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
552 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0);
553 computeKnownBits(B, MaskKnownZero, MaskKnownOne, DL, Depth+1, Query(Q, I));
555 // For those bits in the mask that are known to be one, we can propagate
556 // inverted known bits from the RHS to V.
557 KnownZero |= RHSKnownOne & MaskKnownOne;
558 KnownOne |= RHSKnownZero & MaskKnownOne;
560 } else if (match(Arg, m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)),
562 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
563 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
564 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
565 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
566 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
568 // For those bits in B that are known to be zero, we can propagate known
569 // bits from the RHS to V.
570 KnownZero |= RHSKnownZero & BKnownZero;
571 KnownOne |= RHSKnownOne & BKnownZero;
572 // assume(~(v | b) = a)
573 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
575 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
576 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
577 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
578 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
579 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
581 // For those bits in B that are known to be zero, we can propagate
582 // inverted known bits from the RHS to V.
583 KnownZero |= RHSKnownOne & BKnownZero;
584 KnownOne |= RHSKnownZero & BKnownZero;
586 } else if (match(Arg, m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)),
588 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
589 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
590 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
591 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
592 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
594 // For those bits in B that are known to be zero, we can propagate known
595 // bits from the RHS to V. For those bits in B that are known to be one,
596 // we can propagate inverted known bits from the RHS to V.
597 KnownZero |= RHSKnownZero & BKnownZero;
598 KnownOne |= RHSKnownOne & BKnownZero;
599 KnownZero |= RHSKnownOne & BKnownOne;
600 KnownOne |= RHSKnownZero & BKnownOne;
601 // assume(~(v ^ b) = a)
602 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
604 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
605 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
606 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
607 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0);
608 computeKnownBits(B, BKnownZero, BKnownOne, DL, Depth+1, Query(Q, I));
610 // For those bits in B that are known to be zero, we can propagate
611 // inverted known bits from the RHS to V. For those bits in B that are
612 // known to be one, we can propagate known bits from the RHS to V.
613 KnownZero |= RHSKnownOne & BKnownZero;
614 KnownOne |= RHSKnownZero & BKnownZero;
615 KnownZero |= RHSKnownZero & BKnownOne;
616 KnownOne |= RHSKnownOne & BKnownOne;
617 // assume(v << c = a)
618 } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
620 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
621 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
622 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
623 // For those bits in RHS that are known, we can propagate them to known
624 // bits in V shifted to the right by C.
625 KnownZero |= RHSKnownZero.lshr(C->getZExtValue());
626 KnownOne |= RHSKnownOne.lshr(C->getZExtValue());
627 // assume(~(v << c) = a)
628 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
630 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
631 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
632 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
633 // For those bits in RHS that are known, we can propagate them inverted
634 // to known bits in V shifted to the right by C.
635 KnownZero |= RHSKnownOne.lshr(C->getZExtValue());
636 KnownOne |= RHSKnownZero.lshr(C->getZExtValue());
637 // assume(v >> c = a)
638 } else if (match(Arg,
639 m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)),
643 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
644 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
645 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
646 // For those bits in RHS that are known, we can propagate them to known
647 // bits in V shifted to the right by C.
648 KnownZero |= RHSKnownZero << C->getZExtValue();
649 KnownOne |= RHSKnownOne << C->getZExtValue();
650 // assume(~(v >> c) = a)
651 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr(
652 m_LShr(m_V, m_ConstantInt(C)),
653 m_AShr(m_V, m_ConstantInt(C)))),
655 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q, DL)) {
656 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
657 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
658 // For those bits in RHS that are known, we can propagate them inverted
659 // to known bits in V shifted to the right by C.
660 KnownZero |= RHSKnownOne << C->getZExtValue();
661 KnownOne |= RHSKnownZero << C->getZExtValue();
662 // assume(v >=_s c) where c is non-negative
663 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
664 Pred == ICmpInst::ICMP_SGE &&
665 isValidAssumeForContext(I, Q, DL)) {
666 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
667 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
669 if (RHSKnownZero.isNegative()) {
670 // We know that the sign bit is zero.
671 KnownZero |= APInt::getSignBit(BitWidth);
673 // assume(v >_s c) where c is at least -1.
674 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
675 Pred == ICmpInst::ICMP_SGT &&
676 isValidAssumeForContext(I, Q, DL)) {
677 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
678 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
680 if (RHSKnownOne.isAllOnesValue() || RHSKnownZero.isNegative()) {
681 // We know that the sign bit is zero.
682 KnownZero |= APInt::getSignBit(BitWidth);
684 // assume(v <=_s c) where c is negative
685 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
686 Pred == ICmpInst::ICMP_SLE &&
687 isValidAssumeForContext(I, Q, DL)) {
688 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
689 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
691 if (RHSKnownOne.isNegative()) {
692 // We know that the sign bit is one.
693 KnownOne |= APInt::getSignBit(BitWidth);
695 // assume(v <_s c) where c is non-positive
696 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
697 Pred == ICmpInst::ICMP_SLT &&
698 isValidAssumeForContext(I, Q, DL)) {
699 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
700 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
702 if (RHSKnownZero.isAllOnesValue() || RHSKnownOne.isNegative()) {
703 // We know that the sign bit is one.
704 KnownOne |= APInt::getSignBit(BitWidth);
707 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
708 Pred == ICmpInst::ICMP_ULE &&
709 isValidAssumeForContext(I, Q, DL)) {
710 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
711 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
713 // Whatever high bits in c are zero are known to be zero.
715 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
717 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
718 Pred == ICmpInst::ICMP_ULT &&
719 isValidAssumeForContext(I, Q, DL)) {
720 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
721 computeKnownBits(A, RHSKnownZero, RHSKnownOne, DL, Depth+1, Query(Q, I));
723 // Whatever high bits in c are zero are known to be zero (if c is a power
724 // of 2, then one more).
725 if (isKnownToBeAPowerOfTwo(A, false, Depth+1, Query(Q, I)))
727 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()+1);
730 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes());
735 /// Determine which bits of V are known to be either zero or one and return
736 /// them in the KnownZero/KnownOne bit sets.
738 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
739 /// we cannot optimize based on the assumption that it is zero without changing
740 /// it to be an explicit zero. If we don't change it to zero, other code could
741 /// optimized based on the contradictory assumption that it is non-zero.
742 /// Because instcombine aggressively folds operations with undef args anyway,
743 /// this won't lose us code quality.
745 /// This function is defined on values with integer type, values with pointer
746 /// type (but only if TD is non-null), and vectors of integers. In the case
747 /// where V is a vector, known zero, and known one values are the
748 /// same width as the vector element, and the bit is set only if it is true
749 /// for all of the elements in the vector.
750 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
751 const DataLayout *TD, unsigned Depth,
753 assert(V && "No Value?");
754 assert(Depth <= MaxDepth && "Limit Search Depth");
755 unsigned BitWidth = KnownZero.getBitWidth();
757 assert((V->getType()->isIntOrIntVectorTy() ||
758 V->getType()->getScalarType()->isPointerTy()) &&
759 "Not integer or pointer type!");
761 TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
762 (!V->getType()->isIntOrIntVectorTy() ||
763 V->getType()->getScalarSizeInBits() == BitWidth) &&
764 KnownZero.getBitWidth() == BitWidth &&
765 KnownOne.getBitWidth() == BitWidth &&
766 "V, KnownOne and KnownZero should have same BitWidth");
768 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
769 // We know all of the bits for a constant!
770 KnownOne = CI->getValue();
771 KnownZero = ~KnownOne;
774 // Null and aggregate-zero are all-zeros.
775 if (isa<ConstantPointerNull>(V) ||
776 isa<ConstantAggregateZero>(V)) {
777 KnownOne.clearAllBits();
778 KnownZero = APInt::getAllOnesValue(BitWidth);
781 // Handle a constant vector by taking the intersection of the known bits of
782 // each element. There is no real need to handle ConstantVector here, because
783 // we don't handle undef in any particularly useful way.
784 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
785 // We know that CDS must be a vector of integers. Take the intersection of
787 KnownZero.setAllBits(); KnownOne.setAllBits();
788 APInt Elt(KnownZero.getBitWidth(), 0);
789 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
790 Elt = CDS->getElementAsInteger(i);
797 // The address of an aligned GlobalValue has trailing zeros.
798 if (auto *GO = dyn_cast<GlobalObject>(V)) {
799 unsigned Align = GO->getAlignment();
800 if (Align == 0 && TD) {
801 if (auto *GVar = dyn_cast<GlobalVariable>(GO)) {
802 Type *ObjectType = GVar->getType()->getElementType();
803 if (ObjectType->isSized()) {
804 // If the object is defined in the current Module, we'll be giving
805 // it the preferred alignment. Otherwise, we have to assume that it
806 // may only have the minimum ABI alignment.
807 if (!GVar->isDeclaration() && !GVar->isWeakForLinker())
808 Align = TD->getPreferredAlignment(GVar);
810 Align = TD->getABITypeAlignment(ObjectType);
815 KnownZero = APInt::getLowBitsSet(BitWidth,
816 countTrailingZeros(Align));
818 KnownZero.clearAllBits();
819 KnownOne.clearAllBits();
823 if (Argument *A = dyn_cast<Argument>(V)) {
824 unsigned Align = A->getType()->isPointerTy() ? A->getParamAlignment() : 0;
826 if (!Align && TD && A->hasStructRetAttr()) {
827 // An sret parameter has at least the ABI alignment of the return type.
828 Type *EltTy = cast<PointerType>(A->getType())->getElementType();
829 if (EltTy->isSized())
830 Align = TD->getABITypeAlignment(EltTy);
834 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
836 KnownZero.clearAllBits();
837 KnownOne.clearAllBits();
839 // Don't give up yet... there might be an assumption that provides more
841 computeKnownBitsFromAssume(V, KnownZero, KnownOne, TD, Depth, Q);
845 // Start out not knowing anything.
846 KnownZero.clearAllBits(); KnownOne.clearAllBits();
848 // Limit search depth.
849 // All recursive calls that increase depth must come after this.
850 if (Depth == MaxDepth)
853 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
854 // the bits of its aliasee.
855 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
856 if (!GA->mayBeOverridden())
857 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, TD, Depth + 1, Q);
861 // Check whether a nearby assume intrinsic can determine some known bits.
862 computeKnownBitsFromAssume(V, KnownZero, KnownOne, TD, Depth, Q);
864 Operator *I = dyn_cast<Operator>(V);
867 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
868 switch (I->getOpcode()) {
870 case Instruction::Load:
871 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
872 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
874 case Instruction::And: {
875 // If either the LHS or the RHS are Zero, the result is zero.
876 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
877 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
879 // Output known-1 bits are only known if set in both the LHS & RHS.
880 KnownOne &= KnownOne2;
881 // Output known-0 are known to be clear if zero in either the LHS | RHS.
882 KnownZero |= KnownZero2;
885 case Instruction::Or: {
886 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
887 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
889 // Output known-0 bits are only known if clear in both the LHS & RHS.
890 KnownZero &= KnownZero2;
891 // Output known-1 are known to be set if set in either the LHS | RHS.
892 KnownOne |= KnownOne2;
895 case Instruction::Xor: {
896 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
897 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
899 // Output known-0 bits are known if clear or set in both the LHS & RHS.
900 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
901 // Output known-1 are known to be set if set in only one of the LHS, RHS.
902 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
903 KnownZero = KnownZeroOut;
906 case Instruction::Mul: {
907 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
908 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW,
909 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
913 case Instruction::UDiv: {
914 // For the purposes of computing leading zeros we can conservatively
915 // treat a udiv as a logical right shift by the power of 2 known to
916 // be less than the denominator.
917 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD, Depth+1, Q);
918 unsigned LeadZ = KnownZero2.countLeadingOnes();
920 KnownOne2.clearAllBits();
921 KnownZero2.clearAllBits();
922 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
923 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
924 if (RHSUnknownLeadingOnes != BitWidth)
925 LeadZ = std::min(BitWidth,
926 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
928 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ);
931 case Instruction::Select:
932 computeKnownBits(I->getOperand(2), KnownZero, KnownOne, TD, Depth+1, Q);
933 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
935 // Only known if known in both the LHS and RHS.
936 KnownOne &= KnownOne2;
937 KnownZero &= KnownZero2;
939 case Instruction::FPTrunc:
940 case Instruction::FPExt:
941 case Instruction::FPToUI:
942 case Instruction::FPToSI:
943 case Instruction::SIToFP:
944 case Instruction::UIToFP:
945 break; // Can't work with floating point.
946 case Instruction::PtrToInt:
947 case Instruction::IntToPtr:
948 case Instruction::AddrSpaceCast: // Pointers could be different sizes.
949 // We can't handle these if we don't know the pointer size.
951 // FALL THROUGH and handle them the same as zext/trunc.
952 case Instruction::ZExt:
953 case Instruction::Trunc: {
954 Type *SrcTy = I->getOperand(0)->getType();
956 unsigned SrcBitWidth;
957 // Note that we handle pointer operands here because of inttoptr/ptrtoint
958 // which fall through here.
960 SrcBitWidth = TD->getTypeSizeInBits(SrcTy->getScalarType());
962 SrcBitWidth = SrcTy->getScalarSizeInBits();
963 if (!SrcBitWidth) break;
966 assert(SrcBitWidth && "SrcBitWidth can't be zero");
967 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth);
968 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth);
969 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
970 KnownZero = KnownZero.zextOrTrunc(BitWidth);
971 KnownOne = KnownOne.zextOrTrunc(BitWidth);
972 // Any top bits are known to be zero.
973 if (BitWidth > SrcBitWidth)
974 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
977 case Instruction::BitCast: {
978 Type *SrcTy = I->getOperand(0)->getType();
979 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
980 // TODO: For now, not handling conversions like:
981 // (bitcast i64 %x to <2 x i32>)
982 !I->getType()->isVectorTy()) {
983 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
988 case Instruction::SExt: {
989 // Compute the bits in the result that are not present in the input.
990 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
992 KnownZero = KnownZero.trunc(SrcBitWidth);
993 KnownOne = KnownOne.trunc(SrcBitWidth);
994 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
995 KnownZero = KnownZero.zext(BitWidth);
996 KnownOne = KnownOne.zext(BitWidth);
998 // If the sign bit of the input is known set or clear, then we know the
999 // top bits of the result.
1000 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
1001 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1002 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
1003 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1006 case Instruction::Shl:
1007 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1008 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1009 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1010 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1011 KnownZero <<= ShiftAmt;
1012 KnownOne <<= ShiftAmt;
1013 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
1016 case Instruction::LShr:
1017 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1018 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1019 // Compute the new bits that are at the top now.
1020 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1022 // Unsigned shift right.
1023 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1024 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1025 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1026 // high bits known zero.
1027 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
1030 case Instruction::AShr:
1031 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1032 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1033 // Compute the new bits that are at the top now.
1034 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
1036 // Signed shift right.
1037 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1038 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
1039 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
1041 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1042 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
1043 KnownZero |= HighBits;
1044 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
1045 KnownOne |= HighBits;
1048 case Instruction::Sub: {
1049 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1050 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1051 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
1055 case Instruction::Add: {
1056 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1057 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1058 KnownZero, KnownOne, KnownZero2, KnownOne2, TD,
1062 case Instruction::SRem:
1063 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1064 APInt RA = Rem->getValue().abs();
1065 if (RA.isPowerOf2()) {
1066 APInt LowBits = RA - 1;
1067 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, TD,
1070 // The low bits of the first operand are unchanged by the srem.
1071 KnownZero = KnownZero2 & LowBits;
1072 KnownOne = KnownOne2 & LowBits;
1074 // If the first operand is non-negative or has all low bits zero, then
1075 // the upper bits are all zero.
1076 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
1077 KnownZero |= ~LowBits;
1079 // If the first operand is negative and not all low bits are zero, then
1080 // the upper bits are all one.
1081 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
1082 KnownOne |= ~LowBits;
1084 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1088 // The sign bit is the LHS's sign bit, except when the result of the
1089 // remainder is zero.
1090 if (KnownZero.isNonNegative()) {
1091 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1092 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, TD,
1094 // If it's known zero, our sign bit is also zero.
1095 if (LHSKnownZero.isNegative())
1096 KnownZero.setBit(BitWidth - 1);
1100 case Instruction::URem: {
1101 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1102 APInt RA = Rem->getValue();
1103 if (RA.isPowerOf2()) {
1104 APInt LowBits = (RA - 1);
1105 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD,
1107 KnownZero |= ~LowBits;
1108 KnownOne &= LowBits;
1113 // Since the result is less than or equal to either operand, any leading
1114 // zero bits in either operand must also exist in the result.
1115 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, TD, Depth+1, Q);
1116 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, TD, Depth+1, Q);
1118 unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
1119 KnownZero2.countLeadingOnes());
1120 KnownOne.clearAllBits();
1121 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders);
1125 case Instruction::Alloca: {
1126 AllocaInst *AI = cast<AllocaInst>(V);
1127 unsigned Align = AI->getAlignment();
1128 if (Align == 0 && TD)
1129 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
1132 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
1135 case Instruction::GetElementPtr: {
1136 // Analyze all of the subscripts of this getelementptr instruction
1137 // to determine if we can prove known low zero bits.
1138 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1139 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, TD,
1141 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1143 gep_type_iterator GTI = gep_type_begin(I);
1144 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1145 Value *Index = I->getOperand(i);
1146 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1147 // Handle struct member offset arithmetic.
1153 // Handle case when index is vector zeroinitializer
1154 Constant *CIndex = cast<Constant>(Index);
1155 if (CIndex->isZeroValue())
1158 if (CIndex->getType()->isVectorTy())
1159 Index = CIndex->getSplatValue();
1161 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1162 const StructLayout *SL = TD->getStructLayout(STy);
1163 uint64_t Offset = SL->getElementOffset(Idx);
1164 TrailZ = std::min<unsigned>(TrailZ,
1165 countTrailingZeros(Offset));
1167 // Handle array index arithmetic.
1168 Type *IndexedTy = GTI.getIndexedType();
1169 if (!IndexedTy->isSized()) {
1173 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1174 uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
1175 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1176 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, TD, Depth+1, Q);
1177 TrailZ = std::min(TrailZ,
1178 unsigned(countTrailingZeros(TypeSize) +
1179 LocalKnownZero.countTrailingOnes()));
1183 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
1186 case Instruction::PHI: {
1187 PHINode *P = cast<PHINode>(I);
1188 // Handle the case of a simple two-predecessor recurrence PHI.
1189 // There's a lot more that could theoretically be done here, but
1190 // this is sufficient to catch some interesting cases.
1191 if (P->getNumIncomingValues() == 2) {
1192 for (unsigned i = 0; i != 2; ++i) {
1193 Value *L = P->getIncomingValue(i);
1194 Value *R = P->getIncomingValue(!i);
1195 Operator *LU = dyn_cast<Operator>(L);
1198 unsigned Opcode = LU->getOpcode();
1199 // Check for operations that have the property that if
1200 // both their operands have low zero bits, the result
1201 // will have low zero bits.
1202 if (Opcode == Instruction::Add ||
1203 Opcode == Instruction::Sub ||
1204 Opcode == Instruction::And ||
1205 Opcode == Instruction::Or ||
1206 Opcode == Instruction::Mul) {
1207 Value *LL = LU->getOperand(0);
1208 Value *LR = LU->getOperand(1);
1209 // Find a recurrence.
1216 // Ok, we have a PHI of the form L op= R. Check for low
1218 computeKnownBits(R, KnownZero2, KnownOne2, TD, Depth+1, Q);
1220 // We need to take the minimum number of known bits
1221 APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
1222 computeKnownBits(L, KnownZero3, KnownOne3, TD, Depth+1, Q);
1224 KnownZero = APInt::getLowBitsSet(BitWidth,
1225 std::min(KnownZero2.countTrailingOnes(),
1226 KnownZero3.countTrailingOnes()));
1232 // Unreachable blocks may have zero-operand PHI nodes.
1233 if (P->getNumIncomingValues() == 0)
1236 // Otherwise take the unions of the known bit sets of the operands,
1237 // taking conservative care to avoid excessive recursion.
1238 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
1239 // Skip if every incoming value references to ourself.
1240 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1243 KnownZero = APInt::getAllOnesValue(BitWidth);
1244 KnownOne = APInt::getAllOnesValue(BitWidth);
1245 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
1246 // Skip direct self references.
1247 if (P->getIncomingValue(i) == P) continue;
1249 KnownZero2 = APInt(BitWidth, 0);
1250 KnownOne2 = APInt(BitWidth, 0);
1251 // Recurse, but cap the recursion to one level, because we don't
1252 // want to waste time spinning around in loops.
1253 computeKnownBits(P->getIncomingValue(i), KnownZero2, KnownOne2, TD,
1255 KnownZero &= KnownZero2;
1256 KnownOne &= KnownOne2;
1257 // If all bits have been ruled out, there's no need to check
1259 if (!KnownZero && !KnownOne)
1265 case Instruction::Call:
1266 case Instruction::Invoke:
1267 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1268 computeKnownBitsFromRangeMetadata(*MD, KnownZero);
1269 // If a range metadata is attached to this IntrinsicInst, intersect the
1270 // explicit range specified by the metadata and the implicit range of
1272 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1273 switch (II->getIntrinsicID()) {
1275 case Intrinsic::ctlz:
1276 case Intrinsic::cttz: {
1277 unsigned LowBits = Log2_32(BitWidth)+1;
1278 // If this call is undefined for 0, the result will be less than 2^n.
1279 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1281 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1284 case Intrinsic::ctpop: {
1285 unsigned LowBits = Log2_32(BitWidth)+1;
1286 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1289 case Intrinsic::x86_sse42_crc32_64_64:
1290 KnownZero |= APInt::getHighBitsSet(64, 32);
1295 case Instruction::ExtractValue:
1296 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1297 ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1298 if (EVI->getNumIndices() != 1) break;
1299 if (EVI->getIndices()[0] == 0) {
1300 switch (II->getIntrinsicID()) {
1302 case Intrinsic::uadd_with_overflow:
1303 case Intrinsic::sadd_with_overflow:
1304 computeKnownBitsAddSub(true, II->getArgOperand(0),
1305 II->getArgOperand(1), false, KnownZero,
1306 KnownOne, KnownZero2, KnownOne2, TD, Depth, Q);
1308 case Intrinsic::usub_with_overflow:
1309 case Intrinsic::ssub_with_overflow:
1310 computeKnownBitsAddSub(false, II->getArgOperand(0),
1311 II->getArgOperand(1), false, KnownZero,
1312 KnownOne, KnownZero2, KnownOne2, TD, Depth, Q);
1314 case Intrinsic::umul_with_overflow:
1315 case Intrinsic::smul_with_overflow:
1316 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1),
1317 false, KnownZero, KnownOne,
1318 KnownZero2, KnownOne2, TD, Depth, Q);
1325 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1328 /// Determine whether the sign bit is known to be zero or one.
1329 /// Convenience wrapper around computeKnownBits.
1330 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne,
1331 const DataLayout *TD, unsigned Depth,
1333 unsigned BitWidth = getBitWidth(V->getType(), TD);
1339 APInt ZeroBits(BitWidth, 0);
1340 APInt OneBits(BitWidth, 0);
1341 computeKnownBits(V, ZeroBits, OneBits, TD, Depth, Q);
1342 KnownOne = OneBits[BitWidth - 1];
1343 KnownZero = ZeroBits[BitWidth - 1];
1346 /// Return true if the given value is known to have exactly one
1347 /// bit set when defined. For vectors return true if every element is known to
1348 /// be a power of two when defined. Supports values with integer or pointer
1349 /// types and vectors of integers.
1350 bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth,
1352 if (Constant *C = dyn_cast<Constant>(V)) {
1353 if (C->isNullValue())
1355 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1356 return CI->getValue().isPowerOf2();
1357 // TODO: Handle vector constants.
1360 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1361 // it is shifted off the end then the result is undefined.
1362 if (match(V, m_Shl(m_One(), m_Value())))
1365 // (signbit) >>l X is clearly a power of two if the one is not shifted off the
1366 // bottom. If it is shifted off the bottom then the result is undefined.
1367 if (match(V, m_LShr(m_SignBit(), m_Value())))
1370 // The remaining tests are all recursive, so bail out if we hit the limit.
1371 if (Depth++ == MaxDepth)
1374 Value *X = nullptr, *Y = nullptr;
1375 // A shift of a power of two is a power of two or zero.
1376 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1377 match(V, m_Shr(m_Value(X), m_Value()))))
1378 return isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth, Q);
1380 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1381 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
1383 if (SelectInst *SI = dyn_cast<SelectInst>(V))
1385 isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
1386 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
1388 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1389 // A power of two and'd with anything is a power of two or zero.
1390 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/true, Depth, Q) ||
1391 isKnownToBeAPowerOfTwo(Y, /*OrZero*/true, Depth, Q))
1393 // X & (-X) is always a power of two or zero.
1394 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1399 // Adding a power-of-two or zero to the same power-of-two or zero yields
1400 // either the original power-of-two, a larger power-of-two or zero.
1401 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1402 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1403 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1404 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1405 match(X, m_And(m_Value(), m_Specific(Y))))
1406 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
1408 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1409 match(Y, m_And(m_Value(), m_Specific(X))))
1410 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
1413 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1414 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0);
1415 computeKnownBits(X, LHSZeroBits, LHSOneBits, nullptr, Depth, Q);
1417 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0);
1418 computeKnownBits(Y, RHSZeroBits, RHSOneBits, nullptr, Depth, Q);
1419 // If i8 V is a power of two or zero:
1420 // ZeroBits: 1 1 1 0 1 1 1 1
1421 // ~ZeroBits: 0 0 0 1 0 0 0 0
1422 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2())
1423 // If OrZero isn't set, we cannot give back a zero result.
1424 // Make sure either the LHS or RHS has a bit set.
1425 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue())
1430 // An exact divide or right shift can only shift off zero bits, so the result
1431 // is a power of two only if the first operand is a power of two and not
1432 // copying a sign bit (sdiv int_min, 2).
1433 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1434 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1435 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1442 /// \brief Test whether a GEP's result is known to be non-null.
1444 /// Uses properties inherent in a GEP to try to determine whether it is known
1447 /// Currently this routine does not support vector GEPs.
1448 static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout *DL,
1449 unsigned Depth, const Query &Q) {
1450 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
1453 // FIXME: Support vector-GEPs.
1454 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1456 // If the base pointer is non-null, we cannot walk to a null address with an
1457 // inbounds GEP in address space zero.
1458 if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth, Q))
1461 // Past this, if we don't have DataLayout, we can't do much.
1465 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1466 // If so, then the GEP cannot produce a null pointer, as doing so would
1467 // inherently violate the inbounds contract within address space zero.
1468 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1469 GTI != GTE; ++GTI) {
1470 // Struct types are easy -- they must always be indexed by a constant.
1471 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1472 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1473 unsigned ElementIdx = OpC->getZExtValue();
1474 const StructLayout *SL = DL->getStructLayout(STy);
1475 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1476 if (ElementOffset > 0)
1481 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1482 if (DL->getTypeAllocSize(GTI.getIndexedType()) == 0)
1485 // Fast path the constant operand case both for efficiency and so we don't
1486 // increment Depth when just zipping down an all-constant GEP.
1487 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1493 // We post-increment Depth here because while isKnownNonZero increments it
1494 // as well, when we pop back up that increment won't persist. We don't want
1495 // to recurse 10k times just because we have 10k GEP operands. We don't
1496 // bail completely out because we want to handle constant GEPs regardless
1498 if (Depth++ >= MaxDepth)
1501 if (isKnownNonZero(GTI.getOperand(), DL, Depth, Q))
1508 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1509 /// ensure that the value it's attached to is never Value? 'RangeType' is
1510 /// is the type of the value described by the range.
1511 static bool rangeMetadataExcludesValue(MDNode* Ranges,
1512 const APInt& Value) {
1513 const unsigned NumRanges = Ranges->getNumOperands() / 2;
1514 assert(NumRanges >= 1);
1515 for (unsigned i = 0; i < NumRanges; ++i) {
1516 ConstantInt *Lower =
1517 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1518 ConstantInt *Upper =
1519 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1520 ConstantRange Range(Lower->getValue(), Upper->getValue());
1521 if (Range.contains(Value))
1527 /// Return true if the given value is known to be non-zero when defined.
1528 /// For vectors return true if every element is known to be non-zero when
1529 /// defined. Supports values with integer or pointer type and vectors of
1531 bool isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth,
1533 if (Constant *C = dyn_cast<Constant>(V)) {
1534 if (C->isNullValue())
1536 if (isa<ConstantInt>(C))
1537 // Must be non-zero due to null test above.
1539 // TODO: Handle vectors
1543 if (Instruction* I = dyn_cast<Instruction>(V)) {
1544 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1545 // If the possible ranges don't contain zero, then the value is
1546 // definitely non-zero.
1547 if (IntegerType* Ty = dyn_cast<IntegerType>(V->getType())) {
1548 const APInt ZeroValue(Ty->getBitWidth(), 0);
1549 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1555 // The remaining tests are all recursive, so bail out if we hit the limit.
1556 if (Depth++ >= MaxDepth)
1559 // Check for pointer simplifications.
1560 if (V->getType()->isPointerTy()) {
1561 if (isKnownNonNull(V))
1563 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1564 if (isGEPKnownNonNull(GEP, TD, Depth, Q))
1568 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), TD);
1570 // X | Y != 0 if X != 0 or Y != 0.
1571 Value *X = nullptr, *Y = nullptr;
1572 if (match(V, m_Or(m_Value(X), m_Value(Y))))
1573 return isKnownNonZero(X, TD, Depth, Q) ||
1574 isKnownNonZero(Y, TD, Depth, Q);
1576 // ext X != 0 if X != 0.
1577 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1578 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), TD, Depth, Q);
1580 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
1581 // if the lowest bit is shifted off the end.
1582 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) {
1583 // shl nuw can't remove any non-zero bits.
1584 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1585 if (BO->hasNoUnsignedWrap())
1586 return isKnownNonZero(X, TD, Depth, Q);
1588 APInt KnownZero(BitWidth, 0);
1589 APInt KnownOne(BitWidth, 0);
1590 computeKnownBits(X, KnownZero, KnownOne, TD, Depth, Q);
1594 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
1595 // defined if the sign bit is shifted off the end.
1596 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
1597 // shr exact can only shift out zero bits.
1598 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
1600 return isKnownNonZero(X, TD, Depth, Q);
1602 bool XKnownNonNegative, XKnownNegative;
1603 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth, Q);
1607 // div exact can only produce a zero if the dividend is zero.
1608 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
1609 return isKnownNonZero(X, TD, Depth, Q);
1612 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1613 bool XKnownNonNegative, XKnownNegative;
1614 bool YKnownNonNegative, YKnownNegative;
1615 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, TD, Depth, Q);
1616 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, TD, Depth, Q);
1618 // If X and Y are both non-negative (as signed values) then their sum is not
1619 // zero unless both X and Y are zero.
1620 if (XKnownNonNegative && YKnownNonNegative)
1621 if (isKnownNonZero(X, TD, Depth, Q) ||
1622 isKnownNonZero(Y, TD, Depth, Q))
1625 // If X and Y are both negative (as signed values) then their sum is not
1626 // zero unless both X and Y equal INT_MIN.
1627 if (BitWidth && XKnownNegative && YKnownNegative) {
1628 APInt KnownZero(BitWidth, 0);
1629 APInt KnownOne(BitWidth, 0);
1630 APInt Mask = APInt::getSignedMaxValue(BitWidth);
1631 // The sign bit of X is set. If some other bit is set then X is not equal
1633 computeKnownBits(X, KnownZero, KnownOne, TD, Depth, Q);
1634 if ((KnownOne & Mask) != 0)
1636 // The sign bit of Y is set. If some other bit is set then Y is not equal
1638 computeKnownBits(Y, KnownZero, KnownOne, TD, Depth, Q);
1639 if ((KnownOne & Mask) != 0)
1643 // The sum of a non-negative number and a power of two is not zero.
1644 if (XKnownNonNegative &&
1645 isKnownToBeAPowerOfTwo(Y, /*OrZero*/false, Depth, Q))
1647 if (YKnownNonNegative &&
1648 isKnownToBeAPowerOfTwo(X, /*OrZero*/false, Depth, Q))
1652 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
1653 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
1654 // If X and Y are non-zero then so is X * Y as long as the multiplication
1655 // does not overflow.
1656 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
1657 isKnownNonZero(X, TD, Depth, Q) &&
1658 isKnownNonZero(Y, TD, Depth, Q))
1661 // (C ? X : Y) != 0 if X != 0 and Y != 0.
1662 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
1663 if (isKnownNonZero(SI->getTrueValue(), TD, Depth, Q) &&
1664 isKnownNonZero(SI->getFalseValue(), TD, Depth, Q))
1668 if (!BitWidth) return false;
1669 APInt KnownZero(BitWidth, 0);
1670 APInt KnownOne(BitWidth, 0);
1671 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1672 return KnownOne != 0;
1675 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
1676 /// simplify operations downstream. Mask is known to be zero for bits that V
1679 /// This function is defined on values with integer type, values with pointer
1680 /// type (but only if TD is non-null), and vectors of integers. In the case
1681 /// where V is a vector, the mask, known zero, and known one values are the
1682 /// same width as the vector element, and the bit is set only if it is true
1683 /// for all of the elements in the vector.
1684 bool MaskedValueIsZero(Value *V, const APInt &Mask,
1685 const DataLayout *TD, unsigned Depth,
1687 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1688 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1689 return (KnownZero & Mask) == Mask;
1694 /// Return the number of times the sign bit of the register is replicated into
1695 /// the other bits. We know that at least 1 bit is always equal to the sign bit
1696 /// (itself), but other cases can give us information. For example, immediately
1697 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
1698 /// other, so we return 3.
1700 /// 'Op' must have a scalar integer type.
1702 unsigned ComputeNumSignBits(Value *V, const DataLayout *TD,
1703 unsigned Depth, const Query &Q) {
1704 assert((TD || V->getType()->isIntOrIntVectorTy()) &&
1705 "ComputeNumSignBits requires a DataLayout object to operate "
1706 "on non-integer values!");
1707 Type *Ty = V->getType();
1708 unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
1709 Ty->getScalarSizeInBits();
1711 unsigned FirstAnswer = 1;
1713 // Note that ConstantInt is handled by the general computeKnownBits case
1717 return 1; // Limit search depth.
1719 Operator *U = dyn_cast<Operator>(V);
1720 switch (Operator::getOpcode(V)) {
1722 case Instruction::SExt:
1723 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
1724 return ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q) + Tmp;
1726 case Instruction::SDiv:
1727 const APInt *Denominator;
1728 // sdiv X, C -> adds log(C) sign bits.
1729 if (match(U->getOperand(1), m_APInt(Denominator))) {
1731 // Ignore non-positive denominator.
1732 if (!Denominator->isStrictlyPositive())
1735 // Calculate the incoming numerator bits.
1736 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1738 // Add floor(log(C)) bits to the numerator bits.
1739 return std::min(TyBits, NumBits + Denominator->logBase2());
1743 case Instruction::AShr: {
1744 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1745 // ashr X, C -> adds C sign bits. Vectors too.
1747 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1748 Tmp += ShAmt->getZExtValue();
1749 if (Tmp > TyBits) Tmp = TyBits;
1753 case Instruction::Shl: {
1755 if (match(U->getOperand(1), m_APInt(ShAmt))) {
1756 // shl destroys sign bits.
1757 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1758 Tmp2 = ShAmt->getZExtValue();
1759 if (Tmp2 >= TyBits || // Bad shift.
1760 Tmp2 >= Tmp) break; // Shifted all sign bits out.
1765 case Instruction::And:
1766 case Instruction::Or:
1767 case Instruction::Xor: // NOT is handled here.
1768 // Logical binary ops preserve the number of sign bits at the worst.
1769 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1771 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1772 FirstAnswer = std::min(Tmp, Tmp2);
1773 // We computed what we know about the sign bits as our first
1774 // answer. Now proceed to the generic code that uses
1775 // computeKnownBits, and pick whichever answer is better.
1779 case Instruction::Select:
1780 Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1781 if (Tmp == 1) return 1; // Early out.
1782 Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1, Q);
1783 return std::min(Tmp, Tmp2);
1785 case Instruction::Add:
1786 // Add can have at most one carry bit. Thus we know that the output
1787 // is, at worst, one more bit than the inputs.
1788 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1789 if (Tmp == 1) return 1; // Early out.
1791 // Special case decrementing a value (ADD X, -1):
1792 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
1793 if (CRHS->isAllOnesValue()) {
1794 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1795 computeKnownBits(U->getOperand(0), 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 we are subtracting one from a positive number, there is no carry
1803 // out of the result.
1804 if (KnownZero.isNegative())
1808 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1809 if (Tmp2 == 1) return 1;
1810 return std::min(Tmp, Tmp2)-1;
1812 case Instruction::Sub:
1813 Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1, Q);
1814 if (Tmp2 == 1) return 1;
1817 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
1818 if (CLHS->isNullValue()) {
1819 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1820 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, TD, Depth+1, Q);
1821 // If the input is known to be 0 or 1, the output is 0/-1, which is all
1823 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue())
1826 // If the input is known to be positive (the sign bit is known clear),
1827 // the output of the NEG has the same number of sign bits as the input.
1828 if (KnownZero.isNegative())
1831 // Otherwise, we treat this like a SUB.
1834 // Sub can have at most one carry bit. Thus we know that the output
1835 // is, at worst, one more bit than the inputs.
1836 Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1, Q);
1837 if (Tmp == 1) return 1; // Early out.
1838 return std::min(Tmp, Tmp2)-1;
1840 case Instruction::PHI: {
1841 PHINode *PN = cast<PHINode>(U);
1842 unsigned NumIncomingValues = PN->getNumIncomingValues();
1843 // Don't analyze large in-degree PHIs.
1844 if (NumIncomingValues > 4) break;
1845 // Unreachable blocks may have zero-operand PHI nodes.
1846 if (NumIncomingValues == 0) break;
1848 // Take the minimum of all incoming values. This can't infinitely loop
1849 // because of our depth threshold.
1850 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1, Q);
1851 for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
1852 if (Tmp == 1) return Tmp;
1854 ComputeNumSignBits(PN->getIncomingValue(i), TD,
1860 case Instruction::Trunc:
1861 // FIXME: it's tricky to do anything useful for this, but it is an important
1862 // case for targets like X86.
1866 // Finally, if we can prove that the top bits of the result are 0's or 1's,
1867 // use this information.
1868 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
1870 computeKnownBits(V, KnownZero, KnownOne, TD, Depth, Q);
1872 if (KnownZero.isNegative()) { // sign bit is 0
1874 } else if (KnownOne.isNegative()) { // sign bit is 1;
1881 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
1882 // the number of identical bits in the top of the input value.
1884 Mask <<= Mask.getBitWidth()-TyBits;
1885 // Return # leading zeros. We use 'min' here in case Val was zero before
1886 // shifting. We don't want to return '64' as for an i32 "0".
1887 return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
1890 /// This function computes the integer multiple of Base that equals V.
1891 /// If successful, it returns true and returns the multiple in
1892 /// Multiple. If unsuccessful, it returns false. It looks
1893 /// through SExt instructions only if LookThroughSExt is true.
1894 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
1895 bool LookThroughSExt, unsigned Depth) {
1896 const unsigned MaxDepth = 6;
1898 assert(V && "No Value?");
1899 assert(Depth <= MaxDepth && "Limit Search Depth");
1900 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
1902 Type *T = V->getType();
1904 ConstantInt *CI = dyn_cast<ConstantInt>(V);
1914 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
1915 Constant *BaseVal = ConstantInt::get(T, Base);
1916 if (CO && CO == BaseVal) {
1918 Multiple = ConstantInt::get(T, 1);
1922 if (CI && CI->getZExtValue() % Base == 0) {
1923 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
1927 if (Depth == MaxDepth) return false; // Limit search depth.
1929 Operator *I = dyn_cast<Operator>(V);
1930 if (!I) return false;
1932 switch (I->getOpcode()) {
1934 case Instruction::SExt:
1935 if (!LookThroughSExt) return false;
1936 // otherwise fall through to ZExt
1937 case Instruction::ZExt:
1938 return ComputeMultiple(I->getOperand(0), Base, Multiple,
1939 LookThroughSExt, Depth+1);
1940 case Instruction::Shl:
1941 case Instruction::Mul: {
1942 Value *Op0 = I->getOperand(0);
1943 Value *Op1 = I->getOperand(1);
1945 if (I->getOpcode() == Instruction::Shl) {
1946 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
1947 if (!Op1CI) return false;
1948 // Turn Op0 << Op1 into Op0 * 2^Op1
1949 APInt Op1Int = Op1CI->getValue();
1950 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
1951 APInt API(Op1Int.getBitWidth(), 0);
1952 API.setBit(BitToSet);
1953 Op1 = ConstantInt::get(V->getContext(), API);
1956 Value *Mul0 = nullptr;
1957 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
1958 if (Constant *Op1C = dyn_cast<Constant>(Op1))
1959 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
1960 if (Op1C->getType()->getPrimitiveSizeInBits() <
1961 MulC->getType()->getPrimitiveSizeInBits())
1962 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
1963 if (Op1C->getType()->getPrimitiveSizeInBits() >
1964 MulC->getType()->getPrimitiveSizeInBits())
1965 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
1967 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
1968 Multiple = ConstantExpr::getMul(MulC, Op1C);
1972 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
1973 if (Mul0CI->getValue() == 1) {
1974 // V == Base * Op1, so return Op1
1980 Value *Mul1 = nullptr;
1981 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
1982 if (Constant *Op0C = dyn_cast<Constant>(Op0))
1983 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
1984 if (Op0C->getType()->getPrimitiveSizeInBits() <
1985 MulC->getType()->getPrimitiveSizeInBits())
1986 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
1987 if (Op0C->getType()->getPrimitiveSizeInBits() >
1988 MulC->getType()->getPrimitiveSizeInBits())
1989 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
1991 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
1992 Multiple = ConstantExpr::getMul(MulC, Op0C);
1996 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
1997 if (Mul1CI->getValue() == 1) {
1998 // V == Base * Op0, so return Op0
2006 // We could not determine if V is a multiple of Base.
2010 /// Return true if we can prove that the specified FP value is never equal to
2013 /// NOTE: this function will need to be revisited when we support non-default
2016 bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
2017 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2018 return !CFP->getValueAPF().isNegZero();
2020 // FIXME: Magic number! At the least, this should be given a name because it's
2021 // used similarly in CannotBeOrderedLessThanZero(). A better fix may be to
2022 // expose it as a parameter, so it can be used for testing / experimenting.
2024 return false; // Limit search depth.
2026 const Operator *I = dyn_cast<Operator>(V);
2027 if (!I) return false;
2029 // Check if the nsz fast-math flag is set
2030 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I))
2031 if (FPO->hasNoSignedZeros())
2034 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2035 if (I->getOpcode() == Instruction::FAdd)
2036 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1)))
2037 if (CFP->isNullValue())
2040 // sitofp and uitofp turn into +0.0 for zero.
2041 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
2044 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2045 // sqrt(-0.0) = -0.0, no other negative results are possible.
2046 if (II->getIntrinsicID() == Intrinsic::sqrt)
2047 return CannotBeNegativeZero(II->getArgOperand(0), Depth+1);
2049 if (const CallInst *CI = dyn_cast<CallInst>(I))
2050 if (const Function *F = CI->getCalledFunction()) {
2051 if (F->isDeclaration()) {
2053 if (F->getName() == "abs") return true;
2054 // fabs[lf](x) != -0.0
2055 if (F->getName() == "fabs") return true;
2056 if (F->getName() == "fabsf") return true;
2057 if (F->getName() == "fabsl") return true;
2058 if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
2059 F->getName() == "sqrtl")
2060 return CannotBeNegativeZero(CI->getArgOperand(0), Depth+1);
2067 bool llvm::CannotBeOrderedLessThanZero(const Value *V, unsigned Depth) {
2068 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2069 return !CFP->getValueAPF().isNegative() || CFP->getValueAPF().isZero();
2071 // FIXME: Magic number! At the least, this should be given a name because it's
2072 // used similarly in CannotBeNegativeZero(). A better fix may be to
2073 // expose it as a parameter, so it can be used for testing / experimenting.
2075 return false; // Limit search depth.
2077 const Operator *I = dyn_cast<Operator>(V);
2078 if (!I) return false;
2080 switch (I->getOpcode()) {
2082 case Instruction::FMul:
2083 // x*x is always non-negative or a NaN.
2084 if (I->getOperand(0) == I->getOperand(1))
2087 case Instruction::FAdd:
2088 case Instruction::FDiv:
2089 case Instruction::FRem:
2090 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1) &&
2091 CannotBeOrderedLessThanZero(I->getOperand(1), Depth+1);
2092 case Instruction::FPExt:
2093 case Instruction::FPTrunc:
2094 // Widening/narrowing never change sign.
2095 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2096 case Instruction::Call:
2097 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2098 switch (II->getIntrinsicID()) {
2100 case Intrinsic::exp:
2101 case Intrinsic::exp2:
2102 case Intrinsic::fabs:
2103 case Intrinsic::sqrt:
2105 case Intrinsic::powi:
2106 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
2107 // powi(x,n) is non-negative if n is even.
2108 if (CI->getBitWidth() <= 64 && CI->getSExtValue() % 2u == 0)
2111 return CannotBeOrderedLessThanZero(I->getOperand(0), Depth+1);
2112 case Intrinsic::fma:
2113 case Intrinsic::fmuladd:
2114 // x*x+y is non-negative if y is non-negative.
2115 return I->getOperand(0) == I->getOperand(1) &&
2116 CannotBeOrderedLessThanZero(I->getOperand(2), Depth+1);
2123 /// If the specified value can be set by repeating the same byte in memory,
2124 /// return the i8 value that it is represented with. This is
2125 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2126 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
2127 /// byte store (e.g. i16 0x1234), return null.
2128 Value *llvm::isBytewiseValue(Value *V) {
2129 // All byte-wide stores are splatable, even of arbitrary variables.
2130 if (V->getType()->isIntegerTy(8)) return V;
2132 // Handle 'null' ConstantArrayZero etc.
2133 if (Constant *C = dyn_cast<Constant>(V))
2134 if (C->isNullValue())
2135 return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2137 // Constant float and double values can be handled as integer values if the
2138 // corresponding integer value is "byteable". An important case is 0.0.
2139 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2140 if (CFP->getType()->isFloatTy())
2141 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2142 if (CFP->getType()->isDoubleTy())
2143 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2144 // Don't handle long double formats, which have strange constraints.
2147 // We can handle constant integers that are multiple of 8 bits.
2148 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2149 if (CI->getBitWidth() % 8 == 0) {
2150 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
2152 // We can check that all bytes of an integer are equal by making use of a
2153 // little trick: rotate by 8 and check if it's still the same value.
2154 if (CI->getValue() != CI->getValue().rotl(8))
2156 return ConstantInt::get(V->getContext(), CI->getValue().trunc(8));
2160 // A ConstantDataArray/Vector is splatable if all its members are equal and
2162 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2163 Value *Elt = CA->getElementAsConstant(0);
2164 Value *Val = isBytewiseValue(Elt);
2168 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2169 if (CA->getElementAsConstant(I) != Elt)
2175 // Conceptually, we could handle things like:
2176 // %a = zext i8 %X to i16
2177 // %b = shl i16 %a, 8
2178 // %c = or i16 %a, %b
2179 // but until there is an example that actually needs this, it doesn't seem
2180 // worth worrying about.
2185 // This is the recursive version of BuildSubAggregate. It takes a few different
2186 // arguments. Idxs is the index within the nested struct From that we are
2187 // looking at now (which is of type IndexedType). IdxSkip is the number of
2188 // indices from Idxs that should be left out when inserting into the resulting
2189 // struct. To is the result struct built so far, new insertvalue instructions
2191 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2192 SmallVectorImpl<unsigned> &Idxs,
2194 Instruction *InsertBefore) {
2195 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType);
2197 // Save the original To argument so we can modify it
2199 // General case, the type indexed by Idxs is a struct
2200 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2201 // Process each struct element recursively
2204 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
2208 // Couldn't find any inserted value for this index? Cleanup
2209 while (PrevTo != OrigTo) {
2210 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
2211 PrevTo = Del->getAggregateOperand();
2212 Del->eraseFromParent();
2214 // Stop processing elements
2218 // If we successfully found a value for each of our subaggregates
2222 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
2223 // the struct's elements had a value that was inserted directly. In the latter
2224 // case, perhaps we can't determine each of the subelements individually, but
2225 // we might be able to find the complete struct somewhere.
2227 // Find the value that is at that particular spot
2228 Value *V = FindInsertedValue(From, Idxs);
2233 // Insert the value in the new (sub) aggregrate
2234 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
2235 "tmp", InsertBefore);
2238 // This helper takes a nested struct and extracts a part of it (which is again a
2239 // struct) into a new value. For example, given the struct:
2240 // { a, { b, { c, d }, e } }
2241 // and the indices "1, 1" this returns
2244 // It does this by inserting an insertvalue for each element in the resulting
2245 // struct, as opposed to just inserting a single struct. This will only work if
2246 // each of the elements of the substruct are known (ie, inserted into From by an
2247 // insertvalue instruction somewhere).
2249 // All inserted insertvalue instructions are inserted before InsertBefore
2250 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
2251 Instruction *InsertBefore) {
2252 assert(InsertBefore && "Must have someplace to insert!");
2253 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
2255 Value *To = UndefValue::get(IndexedType);
2256 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
2257 unsigned IdxSkip = Idxs.size();
2259 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
2262 /// Given an aggregrate and an sequence of indices, see if
2263 /// the scalar value indexed is already around as a register, for example if it
2264 /// were inserted directly into the aggregrate.
2266 /// If InsertBefore is not null, this function will duplicate (modified)
2267 /// insertvalues when a part of a nested struct is extracted.
2268 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
2269 Instruction *InsertBefore) {
2270 // Nothing to index? Just return V then (this is useful at the end of our
2272 if (idx_range.empty())
2274 // We have indices, so V should have an indexable type.
2275 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
2276 "Not looking at a struct or array?");
2277 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
2278 "Invalid indices for type?");
2280 if (Constant *C = dyn_cast<Constant>(V)) {
2281 C = C->getAggregateElement(idx_range[0]);
2282 if (!C) return nullptr;
2283 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
2286 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
2287 // Loop the indices for the insertvalue instruction in parallel with the
2288 // requested indices
2289 const unsigned *req_idx = idx_range.begin();
2290 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
2291 i != e; ++i, ++req_idx) {
2292 if (req_idx == idx_range.end()) {
2293 // We can't handle this without inserting insertvalues
2297 // The requested index identifies a part of a nested aggregate. Handle
2298 // this specially. For example,
2299 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
2300 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
2301 // %C = extractvalue {i32, { i32, i32 } } %B, 1
2302 // This can be changed into
2303 // %A = insertvalue {i32, i32 } undef, i32 10, 0
2304 // %C = insertvalue {i32, i32 } %A, i32 11, 1
2305 // which allows the unused 0,0 element from the nested struct to be
2307 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
2311 // This insert value inserts something else than what we are looking for.
2312 // See if the (aggregrate) value inserted into has the value we are
2313 // looking for, then.
2315 return FindInsertedValue(I->getAggregateOperand(), idx_range,
2318 // If we end up here, the indices of the insertvalue match with those
2319 // requested (though possibly only partially). Now we recursively look at
2320 // the inserted value, passing any remaining indices.
2321 return FindInsertedValue(I->getInsertedValueOperand(),
2322 makeArrayRef(req_idx, idx_range.end()),
2326 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
2327 // If we're extracting a value from an aggregrate that was extracted from
2328 // something else, we can extract from that something else directly instead.
2329 // However, we will need to chain I's indices with the requested indices.
2331 // Calculate the number of indices required
2332 unsigned size = I->getNumIndices() + idx_range.size();
2333 // Allocate some space to put the new indices in
2334 SmallVector<unsigned, 5> Idxs;
2336 // Add indices from the extract value instruction
2337 Idxs.append(I->idx_begin(), I->idx_end());
2339 // Add requested indices
2340 Idxs.append(idx_range.begin(), idx_range.end());
2342 assert(Idxs.size() == size
2343 && "Number of indices added not correct?");
2345 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
2347 // Otherwise, we don't know (such as, extracting from a function return value
2348 // or load instruction)
2352 /// Analyze the specified pointer to see if it can be expressed as a base
2353 /// pointer plus a constant offset. Return the base and offset to the caller.
2354 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
2355 const DataLayout *DL) {
2356 // Without DataLayout, conservatively assume 64-bit offsets, which is
2357 // the widest we support.
2358 unsigned BitWidth = DL ? DL->getPointerTypeSizeInBits(Ptr->getType()) : 64;
2359 APInt ByteOffset(BitWidth, 0);
2361 if (Ptr->getType()->isVectorTy())
2364 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
2366 APInt GEPOffset(BitWidth, 0);
2367 if (!GEP->accumulateConstantOffset(*DL, GEPOffset))
2370 ByteOffset += GEPOffset;
2373 Ptr = GEP->getPointerOperand();
2374 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
2375 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
2376 Ptr = cast<Operator>(Ptr)->getOperand(0);
2377 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
2378 if (GA->mayBeOverridden())
2380 Ptr = GA->getAliasee();
2385 Offset = ByteOffset.getSExtValue();
2390 /// This function computes the length of a null-terminated C string pointed to
2391 /// by V. If successful, it returns true and returns the string in Str.
2392 /// If unsuccessful, it returns false.
2393 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
2394 uint64_t Offset, bool TrimAtNul) {
2397 // Look through bitcast instructions and geps.
2398 V = V->stripPointerCasts();
2400 // If the value is a GEP instructionor constant expression, treat it as an
2402 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2403 // Make sure the GEP has exactly three arguments.
2404 if (GEP->getNumOperands() != 3)
2407 // Make sure the index-ee is a pointer to array of i8.
2408 PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
2409 ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
2410 if (!AT || !AT->getElementType()->isIntegerTy(8))
2413 // Check to make sure that the first operand of the GEP is an integer and
2414 // has value 0 so that we are sure we're indexing into the initializer.
2415 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
2416 if (!FirstIdx || !FirstIdx->isZero())
2419 // If the second index isn't a ConstantInt, then this is a variable index
2420 // into the array. If this occurs, we can't say anything meaningful about
2422 uint64_t StartIdx = 0;
2423 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
2424 StartIdx = CI->getZExtValue();
2427 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset);
2430 // The GEP instruction, constant or instruction, must reference a global
2431 // variable that is a constant and is initialized. The referenced constant
2432 // initializer is the array that we'll use for optimization.
2433 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2434 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
2437 // Handle the all-zeros case
2438 if (GV->getInitializer()->isNullValue()) {
2439 // This is a degenerate case. The initializer is constant zero so the
2440 // length of the string must be zero.
2445 // Must be a Constant Array
2446 const ConstantDataArray *Array =
2447 dyn_cast<ConstantDataArray>(GV->getInitializer());
2448 if (!Array || !Array->isString())
2451 // Get the number of elements in the array
2452 uint64_t NumElts = Array->getType()->getArrayNumElements();
2454 // Start out with the entire array in the StringRef.
2455 Str = Array->getAsString();
2457 if (Offset > NumElts)
2460 // Skip over 'offset' bytes.
2461 Str = Str.substr(Offset);
2464 // Trim off the \0 and anything after it. If the array is not nul
2465 // terminated, we just return the whole end of string. The client may know
2466 // some other way that the string is length-bound.
2467 Str = Str.substr(0, Str.find('\0'));
2472 // These next two are very similar to the above, but also look through PHI
2474 // TODO: See if we can integrate these two together.
2476 /// If we can compute the length of the string pointed to by
2477 /// the specified pointer, return 'len+1'. If we can't, return 0.
2478 static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) {
2479 // Look through noop bitcast instructions.
2480 V = V->stripPointerCasts();
2482 // If this is a PHI node, there are two cases: either we have already seen it
2484 if (PHINode *PN = dyn_cast<PHINode>(V)) {
2485 if (!PHIs.insert(PN).second)
2486 return ~0ULL; // already in the set.
2488 // If it was new, see if all the input strings are the same length.
2489 uint64_t LenSoFar = ~0ULL;
2490 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2491 uint64_t Len = GetStringLengthH(PN->getIncomingValue(i), PHIs);
2492 if (Len == 0) return 0; // Unknown length -> unknown.
2494 if (Len == ~0ULL) continue;
2496 if (Len != LenSoFar && LenSoFar != ~0ULL)
2497 return 0; // Disagree -> unknown.
2501 // Success, all agree.
2505 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
2506 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
2507 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs);
2508 if (Len1 == 0) return 0;
2509 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs);
2510 if (Len2 == 0) return 0;
2511 if (Len1 == ~0ULL) return Len2;
2512 if (Len2 == ~0ULL) return Len1;
2513 if (Len1 != Len2) return 0;
2517 // Otherwise, see if we can read the string.
2519 if (!getConstantStringInfo(V, StrData))
2522 return StrData.size()+1;
2525 /// If we can compute the length of the string pointed to by
2526 /// the specified pointer, return 'len+1'. If we can't, return 0.
2527 uint64_t llvm::GetStringLength(Value *V) {
2528 if (!V->getType()->isPointerTy()) return 0;
2530 SmallPtrSet<PHINode*, 32> PHIs;
2531 uint64_t Len = GetStringLengthH(V, PHIs);
2532 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
2533 // an empty string as a length.
2534 return Len == ~0ULL ? 1 : Len;
2538 llvm::GetUnderlyingObject(Value *V, const DataLayout *TD, unsigned MaxLookup) {
2539 if (!V->getType()->isPointerTy())
2541 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
2542 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
2543 V = GEP->getPointerOperand();
2544 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
2545 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
2546 V = cast<Operator>(V)->getOperand(0);
2547 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
2548 if (GA->mayBeOverridden())
2550 V = GA->getAliasee();
2552 // See if InstructionSimplify knows any relevant tricks.
2553 if (Instruction *I = dyn_cast<Instruction>(V))
2554 // TODO: Acquire a DominatorTree and AssumptionCache and use them.
2555 if (Value *Simplified = SimplifyInstruction(I, TD, nullptr)) {
2562 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
2568 llvm::GetUnderlyingObjects(Value *V,
2569 SmallVectorImpl<Value *> &Objects,
2570 const DataLayout *TD,
2571 unsigned MaxLookup) {
2572 SmallPtrSet<Value *, 4> Visited;
2573 SmallVector<Value *, 4> Worklist;
2574 Worklist.push_back(V);
2576 Value *P = Worklist.pop_back_val();
2577 P = GetUnderlyingObject(P, TD, MaxLookup);
2579 if (!Visited.insert(P).second)
2582 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
2583 Worklist.push_back(SI->getTrueValue());
2584 Worklist.push_back(SI->getFalseValue());
2588 if (PHINode *PN = dyn_cast<PHINode>(P)) {
2589 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
2590 Worklist.push_back(PN->getIncomingValue(i));
2594 Objects.push_back(P);
2595 } while (!Worklist.empty());
2598 /// Return true if the only users of this pointer are lifetime markers.
2599 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
2600 for (const User *U : V->users()) {
2601 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
2602 if (!II) return false;
2604 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2605 II->getIntrinsicID() != Intrinsic::lifetime_end)
2611 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
2612 const DataLayout *TD) {
2613 const Operator *Inst = dyn_cast<Operator>(V);
2617 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
2618 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
2622 switch (Inst->getOpcode()) {
2625 case Instruction::UDiv:
2626 case Instruction::URem: {
2627 // x / y is undefined if y == 0.
2629 if (match(Inst->getOperand(1), m_APInt(V)))
2633 case Instruction::SDiv:
2634 case Instruction::SRem: {
2635 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
2636 const APInt *Numerator, *Denominator;
2637 if (!match(Inst->getOperand(1), m_APInt(Denominator)))
2639 // We cannot hoist this division if the denominator is 0.
2640 if (*Denominator == 0)
2642 // It's safe to hoist if the denominator is not 0 or -1.
2643 if (*Denominator != -1)
2645 // At this point we know that the denominator is -1. It is safe to hoist as
2646 // long we know that the numerator is not INT_MIN.
2647 if (match(Inst->getOperand(0), m_APInt(Numerator)))
2648 return !Numerator->isMinSignedValue();
2649 // The numerator *might* be MinSignedValue.
2652 case Instruction::Load: {
2653 const LoadInst *LI = cast<LoadInst>(Inst);
2654 if (!LI->isUnordered() ||
2655 // Speculative load may create a race that did not exist in the source.
2656 LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
2658 return LI->getPointerOperand()->isDereferenceablePointer(TD);
2660 case Instruction::Call: {
2661 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
2662 switch (II->getIntrinsicID()) {
2663 // These synthetic intrinsics have no side-effects and just mark
2664 // information about their operands.
2665 // FIXME: There are other no-op synthetic instructions that potentially
2666 // should be considered at least *safe* to speculate...
2667 case Intrinsic::dbg_declare:
2668 case Intrinsic::dbg_value:
2671 case Intrinsic::bswap:
2672 case Intrinsic::ctlz:
2673 case Intrinsic::ctpop:
2674 case Intrinsic::cttz:
2675 case Intrinsic::objectsize:
2676 case Intrinsic::sadd_with_overflow:
2677 case Intrinsic::smul_with_overflow:
2678 case Intrinsic::ssub_with_overflow:
2679 case Intrinsic::uadd_with_overflow:
2680 case Intrinsic::umul_with_overflow:
2681 case Intrinsic::usub_with_overflow:
2683 // Sqrt should be OK, since the llvm sqrt intrinsic isn't defined to set
2684 // errno like libm sqrt would.
2685 case Intrinsic::sqrt:
2686 case Intrinsic::fma:
2687 case Intrinsic::fmuladd:
2688 case Intrinsic::fabs:
2689 case Intrinsic::minnum:
2690 case Intrinsic::maxnum:
2692 // TODO: some fp intrinsics are marked as having the same error handling
2693 // as libm. They're safe to speculate when they won't error.
2694 // TODO: are convert_{from,to}_fp16 safe?
2695 // TODO: can we list target-specific intrinsics here?
2699 return false; // The called function could have undefined behavior or
2700 // side-effects, even if marked readnone nounwind.
2702 case Instruction::VAArg:
2703 case Instruction::Alloca:
2704 case Instruction::Invoke:
2705 case Instruction::PHI:
2706 case Instruction::Store:
2707 case Instruction::Ret:
2708 case Instruction::Br:
2709 case Instruction::IndirectBr:
2710 case Instruction::Switch:
2711 case Instruction::Unreachable:
2712 case Instruction::Fence:
2713 case Instruction::LandingPad:
2714 case Instruction::AtomicRMW:
2715 case Instruction::AtomicCmpXchg:
2716 case Instruction::Resume:
2717 return false; // Misc instructions which have effects
2721 /// Return true if we know that the specified value is never null.
2722 bool llvm::isKnownNonNull(const Value *V, const TargetLibraryInfo *TLI) {
2723 // Alloca never returns null, malloc might.
2724 if (isa<AllocaInst>(V)) return true;
2726 // A byval, inalloca, or nonnull argument is never null.
2727 if (const Argument *A = dyn_cast<Argument>(V))
2728 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
2730 // Global values are not null unless extern weak.
2731 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2732 return !GV->hasExternalWeakLinkage();
2734 // A Load tagged w/nonnull metadata is never null.
2735 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
2736 return LI->getMetadata(LLVMContext::MD_nonnull);
2738 if (ImmutableCallSite CS = V)
2739 if (CS.isReturnNonNull())
2742 // operator new never returns null.
2743 if (isOperatorNewLikeFn(V, TLI, /*LookThroughBitCast=*/true))
2749 OverflowResult llvm::computeOverflowForUnsignedMul(Value *LHS, Value *RHS,
2750 const DataLayout *DL,
2751 AssumptionCache *AC,
2752 const Instruction *CxtI,
2753 const DominatorTree *DT) {
2754 // Multiplying n * m significant bits yields a result of n + m significant
2755 // bits. If the total number of significant bits does not exceed the
2756 // result bit width (minus 1), there is no overflow.
2757 // This means if we have enough leading zero bits in the operands
2758 // we can guarantee that the result does not overflow.
2759 // Ref: "Hacker's Delight" by Henry Warren
2760 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
2761 APInt LHSKnownZero(BitWidth, 0);
2762 APInt LHSKnownOne(BitWidth, 0);
2763 APInt RHSKnownZero(BitWidth, 0);
2764 APInt RHSKnownOne(BitWidth, 0);
2765 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
2767 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, /*Depth=*/0, AC, CxtI,
2769 // Note that underestimating the number of zero bits gives a more
2770 // conservative answer.
2771 unsigned ZeroBits = LHSKnownZero.countLeadingOnes() +
2772 RHSKnownZero.countLeadingOnes();
2773 // First handle the easy case: if we have enough zero bits there's
2774 // definitely no overflow.
2775 if (ZeroBits >= BitWidth)
2776 return OverflowResult::NeverOverflows;
2778 // Get the largest possible values for each operand.
2779 APInt LHSMax = ~LHSKnownZero;
2780 APInt RHSMax = ~RHSKnownZero;
2782 // We know the multiply operation doesn't overflow if the maximum values for
2783 // each operand will not overflow after we multiply them together.
2785 LHSMax.umul_ov(RHSMax, MaxOverflow);
2787 return OverflowResult::NeverOverflows;
2789 // We know it always overflows if multiplying the smallest possible values for
2790 // the operands also results in overflow.
2792 LHSKnownOne.umul_ov(RHSKnownOne, MinOverflow);
2794 return OverflowResult::AlwaysOverflows;
2796 return OverflowResult::MayOverflow;
2799 OverflowResult llvm::computeOverflowForUnsignedAdd(Value *LHS, Value *RHS,
2800 const DataLayout *DL,
2801 AssumptionCache *AC,
2802 const Instruction *CxtI,
2803 const DominatorTree *DT) {
2804 bool LHSKnownNonNegative, LHSKnownNegative;
2805 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0,
2807 if (LHSKnownNonNegative || LHSKnownNegative) {
2808 bool RHSKnownNonNegative, RHSKnownNegative;
2809 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0,
2812 if (LHSKnownNegative && RHSKnownNegative) {
2813 // The sign bit is set in both cases: this MUST overflow.
2814 // Create a simple add instruction, and insert it into the struct.
2815 return OverflowResult::AlwaysOverflows;
2818 if (LHSKnownNonNegative && RHSKnownNonNegative) {
2819 // The sign bit is clear in both cases: this CANNOT overflow.
2820 // Create a simple add instruction, and insert it into the struct.
2821 return OverflowResult::NeverOverflows;
2825 return OverflowResult::MayOverflow;