LLVM  14.0.0git
ValueTracking.cpp
Go to the documentation of this file.
1 //===- ValueTracking.cpp - Walk computations to compute properties --------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file contains routines that help analyze properties that chains of
10 // computations have.
11 //
12 //===----------------------------------------------------------------------===//
13 
15 #include "llvm/ADT/APFloat.h"
16 #include "llvm/ADT/APInt.h"
17 #include "llvm/ADT/ArrayRef.h"
18 #include "llvm/ADT/None.h"
19 #include "llvm/ADT/Optional.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/StringRef.h"
32 #include "llvm/Analysis/Loads.h"
33 #include "llvm/Analysis/LoopInfo.h"
36 #include "llvm/IR/Argument.h"
37 #include "llvm/IR/Attributes.h"
38 #include "llvm/IR/BasicBlock.h"
39 #include "llvm/IR/Constant.h"
40 #include "llvm/IR/ConstantRange.h"
41 #include "llvm/IR/Constants.h"
42 #include "llvm/IR/DerivedTypes.h"
43 #include "llvm/IR/DiagnosticInfo.h"
44 #include "llvm/IR/Dominators.h"
45 #include "llvm/IR/Function.h"
47 #include "llvm/IR/GlobalAlias.h"
48 #include "llvm/IR/GlobalValue.h"
49 #include "llvm/IR/GlobalVariable.h"
50 #include "llvm/IR/InstrTypes.h"
51 #include "llvm/IR/Instruction.h"
52 #include "llvm/IR/Instructions.h"
53 #include "llvm/IR/IntrinsicInst.h"
54 #include "llvm/IR/Intrinsics.h"
55 #include "llvm/IR/IntrinsicsAArch64.h"
56 #include "llvm/IR/IntrinsicsRISCV.h"
57 #include "llvm/IR/IntrinsicsX86.h"
58 #include "llvm/IR/LLVMContext.h"
59 #include "llvm/IR/Metadata.h"
60 #include "llvm/IR/Module.h"
61 #include "llvm/IR/Operator.h"
62 #include "llvm/IR/PatternMatch.h"
63 #include "llvm/IR/Type.h"
64 #include "llvm/IR/User.h"
65 #include "llvm/IR/Value.h"
66 #include "llvm/Support/Casting.h"
68 #include "llvm/Support/Compiler.h"
70 #include "llvm/Support/KnownBits.h"
72 #include <algorithm>
73 #include <array>
74 #include <cassert>
75 #include <cstdint>
76 #include <iterator>
77 #include <utility>
78 
79 using namespace llvm;
80 using namespace llvm::PatternMatch;
81 
82 // Controls the number of uses of the value searched for possible
83 // dominating comparisons.
84 static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses",
85  cl::Hidden, cl::init(20));
86 
87 // According to the LangRef, branching on a poison condition is absolutely
88 // immediate full UB. However, historically we haven't implemented that
89 // consistently as we have an important transformation (non-trivial unswitch)
90 // which introduces instances of branch on poison/undef to otherwise well
91 // defined programs. This flag exists to let us test optimization benefit
92 // of exploiting the specified behavior (in combination with enabling the
93 // unswitch fix.)
94 static cl::opt<bool> BranchOnPoisonAsUB("branch-on-poison-as-ub",
95  cl::Hidden, cl::init(false));
96 
97 
98 /// Returns the bitwidth of the given scalar or pointer type. For vector types,
99 /// returns the element type's bitwidth.
100 static unsigned getBitWidth(Type *Ty, const DataLayout &DL) {
101  if (unsigned BitWidth = Ty->getScalarSizeInBits())
102  return BitWidth;
103 
104  return DL.getPointerTypeSizeInBits(Ty);
105 }
106 
107 namespace {
108 
109 // Simplifying using an assume can only be done in a particular control-flow
110 // context (the context instruction provides that context). If an assume and
111 // the context instruction are not in the same block then the DT helps in
112 // figuring out if we can use it.
113 struct Query {
114  const DataLayout &DL;
115  AssumptionCache *AC;
116  const Instruction *CxtI;
117  const DominatorTree *DT;
118 
119  // Unlike the other analyses, this may be a nullptr because not all clients
120  // provide it currently.
122 
123  /// If true, it is safe to use metadata during simplification.
124  InstrInfoQuery IIQ;
125 
126  Query(const DataLayout &DL, AssumptionCache *AC, const Instruction *CxtI,
127  const DominatorTree *DT, bool UseInstrInfo,
128  OptimizationRemarkEmitter *ORE = nullptr)
129  : DL(DL), AC(AC), CxtI(CxtI), DT(DT), ORE(ORE), IIQ(UseInstrInfo) {}
130 };
131 
132 } // end anonymous namespace
133 
134 // Given the provided Value and, potentially, a context instruction, return
135 // the preferred context instruction (if any).
136 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) {
137  // If we've been provided with a context instruction, then use that (provided
138  // it has been inserted).
139  if (CxtI && CxtI->getParent())
140  return CxtI;
141 
142  // If the value is really an already-inserted instruction, then use that.
143  CxtI = dyn_cast<Instruction>(V);
144  if (CxtI && CxtI->getParent())
145  return CxtI;
146 
147  return nullptr;
148 }
149 
150 static const Instruction *safeCxtI(const Value *V1, const Value *V2, const Instruction *CxtI) {
151  // If we've been provided with a context instruction, then use that (provided
152  // it has been inserted).
153  if (CxtI && CxtI->getParent())
154  return CxtI;
155 
156  // If the value is really an already-inserted instruction, then use that.
157  CxtI = dyn_cast<Instruction>(V1);
158  if (CxtI && CxtI->getParent())
159  return CxtI;
160 
161  CxtI = dyn_cast<Instruction>(V2);
162  if (CxtI && CxtI->getParent())
163  return CxtI;
164 
165  return nullptr;
166 }
167 
169  const APInt &DemandedElts,
170  APInt &DemandedLHS, APInt &DemandedRHS) {
171  // The length of scalable vectors is unknown at compile time, thus we
172  // cannot check their values
173  if (isa<ScalableVectorType>(Shuf->getType()))
174  return false;
175 
176  int NumElts =
177  cast<FixedVectorType>(Shuf->getOperand(0)->getType())->getNumElements();
178  int NumMaskElts = cast<FixedVectorType>(Shuf->getType())->getNumElements();
179  DemandedLHS = DemandedRHS = APInt::getZero(NumElts);
180  if (DemandedElts.isZero())
181  return true;
182  // Simple case of a shuffle with zeroinitializer.
183  if (all_of(Shuf->getShuffleMask(), [](int Elt) { return Elt == 0; })) {
184  DemandedLHS.setBit(0);
185  return true;
186  }
187  for (int i = 0; i != NumMaskElts; ++i) {
188  if (!DemandedElts[i])
189  continue;
190  int M = Shuf->getMaskValue(i);
191  assert(M < (NumElts * 2) && "Invalid shuffle mask constant");
192 
193  // For undef elements, we don't know anything about the common state of
194  // the shuffle result.
195  if (M == -1)
196  return false;
197  if (M < NumElts)
198  DemandedLHS.setBit(M % NumElts);
199  else
200  DemandedRHS.setBit(M % NumElts);
201  }
202 
203  return true;
204 }
205 
206 static void computeKnownBits(const Value *V, const APInt &DemandedElts,
207  KnownBits &Known, unsigned Depth, const Query &Q);
208 
209 static void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth,
210  const Query &Q) {
211  // FIXME: We currently have no way to represent the DemandedElts of a scalable
212  // vector
213  if (isa<ScalableVectorType>(V->getType())) {
214  Known.resetAll();
215  return;
216  }
217 
218  auto *FVTy = dyn_cast<FixedVectorType>(V->getType());
219  APInt DemandedElts =
220  FVTy ? APInt::getAllOnes(FVTy->getNumElements()) : APInt(1, 1);
221  computeKnownBits(V, DemandedElts, Known, Depth, Q);
222 }
223 
224 void llvm::computeKnownBits(const Value *V, KnownBits &Known,
225  const DataLayout &DL, unsigned Depth,
226  AssumptionCache *AC, const Instruction *CxtI,
227  const DominatorTree *DT,
228  OptimizationRemarkEmitter *ORE, bool UseInstrInfo) {
229  ::computeKnownBits(V, Known, Depth,
230  Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE));
231 }
232 
233 void llvm::computeKnownBits(const Value *V, const APInt &DemandedElts,
234  KnownBits &Known, const DataLayout &DL,
235  unsigned Depth, AssumptionCache *AC,
236  const Instruction *CxtI, const DominatorTree *DT,
237  OptimizationRemarkEmitter *ORE, bool UseInstrInfo) {
238  ::computeKnownBits(V, DemandedElts, Known, Depth,
239  Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE));
240 }
241 
242 static KnownBits computeKnownBits(const Value *V, const APInt &DemandedElts,
243  unsigned Depth, const Query &Q);
244 
245 static KnownBits computeKnownBits(const Value *V, unsigned Depth,
246  const Query &Q);
247 
249  unsigned Depth, AssumptionCache *AC,
250  const Instruction *CxtI,
251  const DominatorTree *DT,
253  bool UseInstrInfo) {
255  V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE));
256 }
257 
258 KnownBits llvm::computeKnownBits(const Value *V, const APInt &DemandedElts,
259  const DataLayout &DL, unsigned Depth,
260  AssumptionCache *AC, const Instruction *CxtI,
261  const DominatorTree *DT,
263  bool UseInstrInfo) {
265  V, DemandedElts, Depth,
266  Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE));
267 }
268 
270  const DataLayout &DL, AssumptionCache *AC,
271  const Instruction *CxtI, const DominatorTree *DT,
272  bool UseInstrInfo) {
273  assert(LHS->getType() == RHS->getType() &&
274  "LHS and RHS should have the same type");
276  "LHS and RHS should be integers");
277  // Look for an inverted mask: (X & ~M) op (Y & M).
278  Value *M;
279  if (match(LHS, m_c_And(m_Not(m_Value(M)), m_Value())) &&
281  return true;
282  if (match(RHS, m_c_And(m_Not(m_Value(M)), m_Value())) &&
284  return true;
285  IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType());
286  KnownBits LHSKnown(IT->getBitWidth());
287  KnownBits RHSKnown(IT->getBitWidth());
288  computeKnownBits(LHS, LHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo);
289  computeKnownBits(RHS, RHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo);
290  return KnownBits::haveNoCommonBitsSet(LHSKnown, RHSKnown);
291 }
292 
294  return !I->user_empty() && all_of(I->users(), [](const User *U) {
295  ICmpInst::Predicate P;
296  return match(U, m_ICmp(P, m_Value(), m_Zero())) && ICmpInst::isEquality(P);
297  });
298 }
299 
300 static bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
301  const Query &Q);
302 
304  bool OrZero, unsigned Depth,
305  AssumptionCache *AC, const Instruction *CxtI,
306  const DominatorTree *DT, bool UseInstrInfo) {
308  V, OrZero, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
309 }
310 
311 static bool isKnownNonZero(const Value *V, const APInt &DemandedElts,
312  unsigned Depth, const Query &Q);
313 
314 static bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q);
315 
316 bool llvm::isKnownNonZero(const Value *V, const DataLayout &DL, unsigned Depth,
317  AssumptionCache *AC, const Instruction *CxtI,
318  const DominatorTree *DT, bool UseInstrInfo) {
320  Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
321 }
322 
324  unsigned Depth, AssumptionCache *AC,
325  const Instruction *CxtI, const DominatorTree *DT,
326  bool UseInstrInfo) {
327  KnownBits Known =
328  computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo);
329  return Known.isNonNegative();
330 }
331 
332 bool llvm::isKnownPositive(const Value *V, const DataLayout &DL, unsigned Depth,
333  AssumptionCache *AC, const Instruction *CxtI,
334  const DominatorTree *DT, bool UseInstrInfo) {
335  if (auto *CI = dyn_cast<ConstantInt>(V))
336  return CI->getValue().isStrictlyPositive();
337 
338  // TODO: We'd doing two recursive queries here. We should factor this such
339  // that only a single query is needed.
340  return isKnownNonNegative(V, DL, Depth, AC, CxtI, DT, UseInstrInfo) &&
341  isKnownNonZero(V, DL, Depth, AC, CxtI, DT, UseInstrInfo);
342 }
343 
344 bool llvm::isKnownNegative(const Value *V, const DataLayout &DL, unsigned Depth,
345  AssumptionCache *AC, const Instruction *CxtI,
346  const DominatorTree *DT, bool UseInstrInfo) {
347  KnownBits Known =
348  computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo);
349  return Known.isNegative();
350 }
351 
352 static bool isKnownNonEqual(const Value *V1, const Value *V2, unsigned Depth,
353  const Query &Q);
354 
355 bool llvm::isKnownNonEqual(const Value *V1, const Value *V2,
356  const DataLayout &DL, AssumptionCache *AC,
357  const Instruction *CxtI, const DominatorTree *DT,
358  bool UseInstrInfo) {
360  Query(DL, AC, safeCxtI(V2, V1, CxtI), DT,
361  UseInstrInfo, /*ORE=*/nullptr));
362 }
363 
364 static bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
365  const Query &Q);
366 
367 bool llvm::MaskedValueIsZero(const Value *V, const APInt &Mask,
368  const DataLayout &DL, unsigned Depth,
369  AssumptionCache *AC, const Instruction *CxtI,
370  const DominatorTree *DT, bool UseInstrInfo) {
372  V, Mask, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
373 }
374 
375 static unsigned ComputeNumSignBits(const Value *V, const APInt &DemandedElts,
376  unsigned Depth, const Query &Q);
377 
378 static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
379  const Query &Q) {
380  // FIXME: We currently have no way to represent the DemandedElts of a scalable
381  // vector
382  if (isa<ScalableVectorType>(V->getType()))
383  return 1;
384 
385  auto *FVTy = dyn_cast<FixedVectorType>(V->getType());
386  APInt DemandedElts =
387  FVTy ? APInt::getAllOnes(FVTy->getNumElements()) : APInt(1, 1);
388  return ComputeNumSignBits(V, DemandedElts, Depth, Q);
389 }
390 
391 unsigned llvm::ComputeNumSignBits(const Value *V, const DataLayout &DL,
392  unsigned Depth, AssumptionCache *AC,
393  const Instruction *CxtI,
394  const DominatorTree *DT, bool UseInstrInfo) {
396  V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
397 }
398 
400  unsigned Depth, AssumptionCache *AC,
401  const Instruction *CxtI,
402  const DominatorTree *DT) {
403  unsigned SignBits = ComputeNumSignBits(V, DL, Depth, AC, CxtI, DT);
404  return V->getType()->getScalarSizeInBits() - SignBits + 1;
405 }
406 
407 static void computeKnownBitsAddSub(bool Add, const Value *Op0, const Value *Op1,
408  bool NSW, const APInt &DemandedElts,
409  KnownBits &KnownOut, KnownBits &Known2,
410  unsigned Depth, const Query &Q) {
411  computeKnownBits(Op1, DemandedElts, KnownOut, Depth + 1, Q);
412 
413  // If one operand is unknown and we have no nowrap information,
414  // the result will be unknown independently of the second operand.
415  if (KnownOut.isUnknown() && !NSW)
416  return;
417 
418  computeKnownBits(Op0, DemandedElts, Known2, Depth + 1, Q);
419  KnownOut = KnownBits::computeForAddSub(Add, NSW, Known2, KnownOut);
420 }
421 
422 static void computeKnownBitsMul(const Value *Op0, const Value *Op1, bool NSW,
423  const APInt &DemandedElts, KnownBits &Known,
424  KnownBits &Known2, unsigned Depth,
425  const Query &Q) {
426  computeKnownBits(Op1, DemandedElts, Known, Depth + 1, Q);
427  computeKnownBits(Op0, DemandedElts, Known2, Depth + 1, Q);
428 
429  bool isKnownNegative = false;
430  bool isKnownNonNegative = false;
431  // If the multiplication is known not to overflow, compute the sign bit.
432  if (NSW) {
433  if (Op0 == Op1) {
434  // The product of a number with itself is non-negative.
435  isKnownNonNegative = true;
436  } else {
437  bool isKnownNonNegativeOp1 = Known.isNonNegative();
438  bool isKnownNonNegativeOp0 = Known2.isNonNegative();
439  bool isKnownNegativeOp1 = Known.isNegative();
440  bool isKnownNegativeOp0 = Known2.isNegative();
441  // The product of two numbers with the same sign is non-negative.
442  isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
443  (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
444  // The product of a negative number and a non-negative number is either
445  // negative or zero.
446  if (!isKnownNonNegative)
448  (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
449  Known2.isNonZero()) ||
450  (isKnownNegativeOp0 && isKnownNonNegativeOp1 && Known.isNonZero());
451  }
452  }
453 
454  Known = KnownBits::mul(Known, Known2);
455 
456  // Only make use of no-wrap flags if we failed to compute the sign bit
457  // directly. This matters if the multiplication always overflows, in
458  // which case we prefer to follow the result of the direct computation,
459  // though as the program is invoking undefined behaviour we can choose
460  // whatever we like here.
461  if (isKnownNonNegative && !Known.isNegative())
462  Known.makeNonNegative();
463  else if (isKnownNegative && !Known.isNonNegative())
464  Known.makeNegative();
465 }
466 
468  KnownBits &Known) {
469  unsigned BitWidth = Known.getBitWidth();
470  unsigned NumRanges = Ranges.getNumOperands() / 2;
471  assert(NumRanges >= 1);
472 
473  Known.Zero.setAllBits();
474  Known.One.setAllBits();
475 
476  for (unsigned i = 0; i < NumRanges; ++i) {
477  ConstantInt *Lower =
478  mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
479  ConstantInt *Upper =
480  mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
481  ConstantRange Range(Lower->getValue(), Upper->getValue());
482 
483  // The first CommonPrefixBits of all values in Range are equal.
484  unsigned CommonPrefixBits =
485  (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countLeadingZeros();
486  APInt Mask = APInt::getHighBitsSet(BitWidth, CommonPrefixBits);
487  APInt UnsignedMax = Range.getUnsignedMax().zextOrTrunc(BitWidth);
488  Known.One &= UnsignedMax & Mask;
489  Known.Zero &= ~UnsignedMax & Mask;
490  }
491 }
492 
493 static bool isEphemeralValueOf(const Instruction *I, const Value *E) {
494  SmallVector<const Value *, 16> WorkSet(1, I);
497 
498  // The instruction defining an assumption's condition itself is always
499  // considered ephemeral to that assumption (even if it has other
500  // non-ephemeral users). See r246696's test case for an example.
501  if (is_contained(I->operands(), E))
502  return true;
503 
504  while (!WorkSet.empty()) {
505  const Value *V = WorkSet.pop_back_val();
506  if (!Visited.insert(V).second)
507  continue;
508 
509  // If all uses of this value are ephemeral, then so is this value.
510  if (llvm::all_of(V->users(), [&](const User *U) {
511  return EphValues.count(U);
512  })) {
513  if (V == E)
514  return true;
515 
516  if (V == I || (isa<Instruction>(V) &&
517  !cast<Instruction>(V)->mayHaveSideEffects() &&
518  !cast<Instruction>(V)->isTerminator())) {
519  EphValues.insert(V);
520  if (const User *U = dyn_cast<User>(V))
521  append_range(WorkSet, U->operands());
522  }
523  }
524  }
525 
526  return false;
527 }
528 
529 // Is this an intrinsic that cannot be speculated but also cannot trap?
531  if (const IntrinsicInst *CI = dyn_cast<IntrinsicInst>(I))
532  return CI->isAssumeLikeIntrinsic();
533 
534  return false;
535 }
536 
538  const Instruction *CxtI,
539  const DominatorTree *DT) {
540  // There are two restrictions on the use of an assume:
541  // 1. The assume must dominate the context (or the control flow must
542  // reach the assume whenever it reaches the context).
543  // 2. The context must not be in the assume's set of ephemeral values
544  // (otherwise we will use the assume to prove that the condition
545  // feeding the assume is trivially true, thus causing the removal of
546  // the assume).
547 
548  if (Inv->getParent() == CxtI->getParent()) {
549  // If Inv and CtxI are in the same block, check if the assume (Inv) is first
550  // in the BB.
551  if (Inv->comesBefore(CxtI))
552  return true;
553 
554  // Don't let an assume affect itself - this would cause the problems
555  // `isEphemeralValueOf` is trying to prevent, and it would also make
556  // the loop below go out of bounds.
557  if (Inv == CxtI)
558  return false;
559 
560  // The context comes first, but they're both in the same block.
561  // Make sure there is nothing in between that might interrupt
562  // the control flow, not even CxtI itself.
563  // We limit the scan distance between the assume and its context instruction
564  // to avoid a compile-time explosion. This limit is chosen arbitrarily, so
565  // it can be adjusted if needed (could be turned into a cl::opt).
566  auto Range = make_range(CxtI->getIterator(), Inv->getIterator());
568  return false;
569 
570  return !isEphemeralValueOf(Inv, CxtI);
571  }
572 
573  // Inv and CxtI are in different blocks.
574  if (DT) {
575  if (DT->dominates(Inv, CxtI))
576  return true;
577  } else if (Inv->getParent() == CxtI->getParent()->getSinglePredecessor()) {
578  // We don't have a DT, but this trivially dominates.
579  return true;
580  }
581 
582  return false;
583 }
584 
585 static bool cmpExcludesZero(CmpInst::Predicate Pred, const Value *RHS) {
586  // v u> y implies v != 0.
587  if (Pred == ICmpInst::ICMP_UGT)
588  return true;
589 
590  // Special-case v != 0 to also handle v != null.
591  if (Pred == ICmpInst::ICMP_NE)
592  return match(RHS, m_Zero());
593 
594  // All other predicates - rely on generic ConstantRange handling.
595  const APInt *C;
596  if (!match(RHS, m_APInt(C)))
597  return false;
598 
600  return !TrueValues.contains(APInt::getZero(C->getBitWidth()));
601 }
602 
603 static bool isKnownNonZeroFromAssume(const Value *V, const Query &Q) {
604  // Use of assumptions is context-sensitive. If we don't have a context, we
605  // cannot use them!
606  if (!Q.AC || !Q.CxtI)
607  return false;
608 
609  if (Q.CxtI && V->getType()->isPointerTy()) {
610  SmallVector<Attribute::AttrKind, 2> AttrKinds{Attribute::NonNull};
611  if (!NullPointerIsDefined(Q.CxtI->getFunction(),
613  AttrKinds.push_back(Attribute::Dereferenceable);
614 
615  if (getKnowledgeValidInContext(V, AttrKinds, Q.CxtI, Q.DT, Q.AC))
616  return true;
617  }
618 
619  for (auto &AssumeVH : Q.AC->assumptionsFor(V)) {
620  if (!AssumeVH)
621  continue;
622  CallInst *I = cast<CallInst>(AssumeVH);
623  assert(I->getFunction() == Q.CxtI->getFunction() &&
624  "Got assumption for the wrong function!");
625 
626  // Warning: This loop can end up being somewhat performance sensitive.
627  // We're running this loop for once for each value queried resulting in a
628  // runtime of ~O(#assumes * #values).
629 
630  assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
631  "must be an assume intrinsic");
632 
633  Value *RHS;
634  CmpInst::Predicate Pred;
635  auto m_V = m_CombineOr(m_Specific(V), m_PtrToInt(m_Specific(V)));
636  if (!match(I->getArgOperand(0), m_c_ICmp(Pred, m_V, m_Value(RHS))))
637  return false;
638 
639  if (cmpExcludesZero(Pred, RHS) && isValidAssumeForContext(I, Q.CxtI, Q.DT))
640  return true;
641  }
642 
643  return false;
644 }
645 
646 static void computeKnownBitsFromAssume(const Value *V, KnownBits &Known,
647  unsigned Depth, const Query &Q) {
648  // Use of assumptions is context-sensitive. If we don't have a context, we
649  // cannot use them!
650  if (!Q.AC || !Q.CxtI)
651  return;
652 
653  unsigned BitWidth = Known.getBitWidth();
654 
655  // Refine Known set if the pointer alignment is set by assume bundles.
656  if (V->getType()->isPointerTy()) {
658  V, {Attribute::Alignment}, Q.CxtI, Q.DT, Q.AC)) {
659  Known.Zero.setLowBits(Log2_64(RK.ArgValue));
660  }
661  }
662 
663  // Note that the patterns below need to be kept in sync with the code
664  // in AssumptionCache::updateAffectedValues.
665 
666  for (auto &AssumeVH : Q.AC->assumptionsFor(V)) {
667  if (!AssumeVH)
668  continue;
669  CallInst *I = cast<CallInst>(AssumeVH);
670  assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
671  "Got assumption for the wrong function!");
672 
673  // Warning: This loop can end up being somewhat performance sensitive.
674  // We're running this loop for once for each value queried resulting in a
675  // runtime of ~O(#assumes * #values).
676 
677  assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
678  "must be an assume intrinsic");
679 
680  Value *Arg = I->getArgOperand(0);
681 
682  if (Arg == V && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
683  assert(BitWidth == 1 && "assume operand is not i1?");
684  Known.setAllOnes();
685  return;
686  }
687  if (match(Arg, m_Not(m_Specific(V))) &&
688  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
689  assert(BitWidth == 1 && "assume operand is not i1?");
690  Known.setAllZero();
691  return;
692  }
693 
694  // The remaining tests are all recursive, so bail out if we hit the limit.
696  continue;
697 
698  ICmpInst *Cmp = dyn_cast<ICmpInst>(Arg);
699  if (!Cmp)
700  continue;
701 
702  // We are attempting to compute known bits for the operands of an assume.
703  // Do not try to use other assumptions for those recursive calls because
704  // that can lead to mutual recursion and a compile-time explosion.
705  // An example of the mutual recursion: computeKnownBits can call
706  // isKnownNonZero which calls computeKnownBitsFromAssume (this function)
707  // and so on.
708  Query QueryNoAC = Q;
709  QueryNoAC.AC = nullptr;
710 
711  // Note that ptrtoint may change the bitwidth.
712  Value *A, *B;
713  auto m_V = m_CombineOr(m_Specific(V), m_PtrToInt(m_Specific(V)));
714 
715  CmpInst::Predicate Pred;
716  uint64_t C;
717  switch (Cmp->getPredicate()) {
718  default:
719  break;
720  case ICmpInst::ICMP_EQ:
721  // assume(v = a)
722  if (match(Cmp, m_c_ICmp(Pred, m_V, m_Value(A))) &&
723  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
724  KnownBits RHSKnown =
725  computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
726  Known.Zero |= RHSKnown.Zero;
727  Known.One |= RHSKnown.One;
728  // assume(v & b = a)
729  } else if (match(Cmp,
730  m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
731  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
732  KnownBits RHSKnown =
733  computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
734  KnownBits MaskKnown =
735  computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
736 
737  // For those bits in the mask that are known to be one, we can propagate
738  // known bits from the RHS to V.
739  Known.Zero |= RHSKnown.Zero & MaskKnown.One;
740  Known.One |= RHSKnown.One & MaskKnown.One;
741  // assume(~(v & b) = a)
742  } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
743  m_Value(A))) &&
744  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
745  KnownBits RHSKnown =
746  computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
747  KnownBits MaskKnown =
748  computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
749 
750  // For those bits in the mask that are known to be one, we can propagate
751  // inverted known bits from the RHS to V.
752  Known.Zero |= RHSKnown.One & MaskKnown.One;
753  Known.One |= RHSKnown.Zero & MaskKnown.One;
754  // assume(v | b = a)
755  } else if (match(Cmp,
756  m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
757  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
758  KnownBits RHSKnown =
759  computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
760  KnownBits BKnown =
761  computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
762 
763  // For those bits in B that are known to be zero, we can propagate known
764  // bits from the RHS to V.
765  Known.Zero |= RHSKnown.Zero & BKnown.Zero;
766  Known.One |= RHSKnown.One & BKnown.Zero;
767  // assume(~(v | b) = a)
768  } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
769  m_Value(A))) &&
770  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
771  KnownBits RHSKnown =
772  computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
773  KnownBits BKnown =
774  computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
775 
776  // For those bits in B that are known to be zero, we can propagate
777  // inverted known bits from the RHS to V.
778  Known.Zero |= RHSKnown.One & BKnown.Zero;
779  Known.One |= RHSKnown.Zero & BKnown.Zero;
780  // assume(v ^ b = a)
781  } else if (match(Cmp,
782  m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
783  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
784  KnownBits RHSKnown =
785  computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
786  KnownBits BKnown =
787  computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
788 
789  // For those bits in B that are known to be zero, we can propagate known
790  // bits from the RHS to V. For those bits in B that are known to be one,
791  // we can propagate inverted known bits from the RHS to V.
792  Known.Zero |= RHSKnown.Zero & BKnown.Zero;
793  Known.One |= RHSKnown.One & BKnown.Zero;
794  Known.Zero |= RHSKnown.One & BKnown.One;
795  Known.One |= RHSKnown.Zero & BKnown.One;
796  // assume(~(v ^ b) = a)
797  } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
798  m_Value(A))) &&
799  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
800  KnownBits RHSKnown =
801  computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
802  KnownBits BKnown =
803  computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
804 
805  // For those bits in B that are known to be zero, we can propagate
806  // inverted known bits from the RHS to V. For those bits in B that are
807  // known to be one, we can propagate known bits from the RHS to V.
808  Known.Zero |= RHSKnown.One & BKnown.Zero;
809  Known.One |= RHSKnown.Zero & BKnown.Zero;
810  Known.Zero |= RHSKnown.Zero & BKnown.One;
811  Known.One |= RHSKnown.One & BKnown.One;
812  // assume(v << c = a)
813  } else if (match(Cmp, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
814  m_Value(A))) &&
815  isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) {
816  KnownBits RHSKnown =
817  computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
818 
819  // For those bits in RHS that are known, we can propagate them to known
820  // bits in V shifted to the right by C.
821  RHSKnown.Zero.lshrInPlace(C);
822  Known.Zero |= RHSKnown.Zero;
823  RHSKnown.One.lshrInPlace(C);
824  Known.One |= RHSKnown.One;
825  // assume(~(v << c) = a)
826  } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
827  m_Value(A))) &&
828  isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) {
829  KnownBits RHSKnown =
830  computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
831  // For those bits in RHS that are known, we can propagate them inverted
832  // to known bits in V shifted to the right by C.
833  RHSKnown.One.lshrInPlace(C);
834  Known.Zero |= RHSKnown.One;
835  RHSKnown.Zero.lshrInPlace(C);
836  Known.One |= RHSKnown.Zero;
837  // assume(v >> c = a)
838  } else if (match(Cmp, m_c_ICmp(Pred, m_Shr(m_V, m_ConstantInt(C)),
839  m_Value(A))) &&
840  isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) {
841  KnownBits RHSKnown =
842  computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
843  // For those bits in RHS that are known, we can propagate them to known
844  // bits in V shifted to the right by C.
845  Known.Zero |= RHSKnown.Zero << C;
846  Known.One |= RHSKnown.One << C;
847  // assume(~(v >> c) = a)
848  } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_Shr(m_V, m_ConstantInt(C))),
849  m_Value(A))) &&
850  isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) {
851  KnownBits RHSKnown =
852  computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
853  // For those bits in RHS that are known, we can propagate them inverted
854  // to known bits in V shifted to the right by C.
855  Known.Zero |= RHSKnown.One << C;
856  Known.One |= RHSKnown.Zero << C;
857  }
858  break;
859  case ICmpInst::ICMP_SGE:
860  // assume(v >=_s c) where c is non-negative
861  if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) &&
862  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
863  KnownBits RHSKnown =
864  computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth);
865 
866  if (RHSKnown.isNonNegative()) {
867  // We know that the sign bit is zero.
868  Known.makeNonNegative();
869  }
870  }
871  break;
872  case ICmpInst::ICMP_SGT:
873  // assume(v >_s c) where c is at least -1.
874  if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) &&
875  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
876  KnownBits RHSKnown =
877  computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth);
878 
879  if (RHSKnown.isAllOnes() || RHSKnown.isNonNegative()) {
880  // We know that the sign bit is zero.
881  Known.makeNonNegative();
882  }
883  }
884  break;
885  case ICmpInst::ICMP_SLE:
886  // assume(v <=_s c) where c is negative
887  if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) &&
888  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
889  KnownBits RHSKnown =
890  computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth);
891 
892  if (RHSKnown.isNegative()) {
893  // We know that the sign bit is one.
894  Known.makeNegative();
895  }
896  }
897  break;
898  case ICmpInst::ICMP_SLT:
899  // assume(v <_s c) where c is non-positive
900  if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) &&
901  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
902  KnownBits RHSKnown =
903  computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
904 
905  if (RHSKnown.isZero() || RHSKnown.isNegative()) {
906  // We know that the sign bit is one.
907  Known.makeNegative();
908  }
909  }
910  break;
911  case ICmpInst::ICMP_ULE:
912  // assume(v <=_u c)
913  if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) &&
914  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
915  KnownBits RHSKnown =
916  computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
917 
918  // Whatever high bits in c are zero are known to be zero.
919  Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros());
920  }
921  break;
922  case ICmpInst::ICMP_ULT:
923  // assume(v <_u c)
924  if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) &&
925  isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
926  KnownBits RHSKnown =
927  computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
928 
929  // If the RHS is known zero, then this assumption must be wrong (nothing
930  // is unsigned less than zero). Signal a conflict and get out of here.
931  if (RHSKnown.isZero()) {
932  Known.Zero.setAllBits();
933  Known.One.setAllBits();
934  break;
935  }
936 
937  // Whatever high bits in c are zero are known to be zero (if c is a power
938  // of 2, then one more).
939  if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, QueryNoAC))
940  Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros() + 1);
941  else
942  Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros());
943  }
944  break;
945  }
946  }
947 
948  // If assumptions conflict with each other or previous known bits, then we
949  // have a logical fallacy. It's possible that the assumption is not reachable,
950  // so this isn't a real bug. On the other hand, the program may have undefined
951  // behavior, or we might have a bug in the compiler. We can't assert/crash, so
952  // clear out the known bits, try to warn the user, and hope for the best.
953  if (Known.Zero.intersects(Known.One)) {
954  Known.resetAll();
955 
956  if (Q.ORE)
957  Q.ORE->emit([&]() {
958  auto *CxtI = const_cast<Instruction *>(Q.CxtI);
959  return OptimizationRemarkAnalysis("value-tracking", "BadAssumption",
960  CxtI)
961  << "Detected conflicting code assumptions. Program may "
962  "have undefined behavior, or compiler may have "
963  "internal error.";
964  });
965  }
966 }
967 
968 /// Compute known bits from a shift operator, including those with a
969 /// non-constant shift amount. Known is the output of this function. Known2 is a
970 /// pre-allocated temporary with the same bit width as Known and on return
971 /// contains the known bit of the shift value source. KF is an
972 /// operator-specific function that, given the known-bits and a shift amount,
973 /// compute the implied known-bits of the shift operator's result respectively
974 /// for that shift amount. The results from calling KF are conservatively
975 /// combined for all permitted shift amounts.
977  const Operator *I, const APInt &DemandedElts, KnownBits &Known,
978  KnownBits &Known2, unsigned Depth, const Query &Q,
979  function_ref<KnownBits(const KnownBits &, const KnownBits &)> KF) {
980  unsigned BitWidth = Known.getBitWidth();
981  computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q);
982  computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q);
983 
984  // Note: We cannot use Known.Zero.getLimitedValue() here, because if
985  // BitWidth > 64 and any upper bits are known, we'll end up returning the
986  // limit value (which implies all bits are known).
987  uint64_t ShiftAmtKZ = Known.Zero.zextOrTrunc(64).getZExtValue();
988  uint64_t ShiftAmtKO = Known.One.zextOrTrunc(64).getZExtValue();
989  bool ShiftAmtIsConstant = Known.isConstant();
990  bool MaxShiftAmtIsOutOfRange = Known.getMaxValue().uge(BitWidth);
991 
992  if (ShiftAmtIsConstant) {
993  Known = KF(Known2, Known);
994 
995  // If the known bits conflict, this must be an overflowing left shift, so
996  // the shift result is poison. We can return anything we want. Choose 0 for
997  // the best folding opportunity.
998  if (Known.hasConflict())
999  Known.setAllZero();
1000 
1001  return;
1002  }
1003 
1004  // If the shift amount could be greater than or equal to the bit-width of the
1005  // LHS, the value could be poison, but bail out because the check below is
1006  // expensive.
1007  // TODO: Should we just carry on?
1008  if (MaxShiftAmtIsOutOfRange) {
1009  Known.resetAll();
1010  return;
1011  }
1012 
1013  // It would be more-clearly correct to use the two temporaries for this
1014  // calculation. Reusing the APInts here to prevent unnecessary allocations.
1015  Known.resetAll();
1016 
1017  // If we know the shifter operand is nonzero, we can sometimes infer more
1018  // known bits. However this is expensive to compute, so be lazy about it and
1019  // only compute it when absolutely necessary.
1020  Optional<bool> ShifterOperandIsNonZero;
1021 
1022  // Early exit if we can't constrain any well-defined shift amount.
1023  if (!(ShiftAmtKZ & (PowerOf2Ceil(BitWidth) - 1)) &&
1024  !(ShiftAmtKO & (PowerOf2Ceil(BitWidth) - 1))) {
1025  ShifterOperandIsNonZero =
1026  isKnownNonZero(I->getOperand(1), DemandedElts, Depth + 1, Q);
1027  if (!*ShifterOperandIsNonZero)
1028  return;
1029  }
1030 
1031  Known.Zero.setAllBits();
1032  Known.One.setAllBits();
1033  for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) {
1034  // Combine the shifted known input bits only for those shift amounts
1035  // compatible with its known constraints.
1036  if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt)
1037  continue;
1038  if ((ShiftAmt | ShiftAmtKO) != ShiftAmt)
1039  continue;
1040  // If we know the shifter is nonzero, we may be able to infer more known
1041  // bits. This check is sunk down as far as possible to avoid the expensive
1042  // call to isKnownNonZero if the cheaper checks above fail.
1043  if (ShiftAmt == 0) {
1044  if (!ShifterOperandIsNonZero.hasValue())
1045  ShifterOperandIsNonZero =
1046  isKnownNonZero(I->getOperand(1), DemandedElts, Depth + 1, Q);
1047  if (*ShifterOperandIsNonZero)
1048  continue;
1049  }
1050 
1051  Known = KnownBits::commonBits(
1052  Known, KF(Known2, KnownBits::makeConstant(APInt(32, ShiftAmt))));
1053  }
1054 
1055  // If the known bits conflict, the result is poison. Return a 0 and hope the
1056  // caller can further optimize that.
1057  if (Known.hasConflict())
1058  Known.setAllZero();
1059 }
1060 
1062  const APInt &DemandedElts,
1063  KnownBits &Known, unsigned Depth,
1064  const Query &Q) {
1065  unsigned BitWidth = Known.getBitWidth();
1066 
1067  KnownBits Known2(BitWidth);
1068  switch (I->getOpcode()) {
1069  default: break;
1070  case Instruction::Load:
1071  if (MDNode *MD =
1072  Q.IIQ.getMetadata(cast<LoadInst>(I), LLVMContext::MD_range))
1074  break;
1075  case Instruction::And: {
1076  // If either the LHS or the RHS are Zero, the result is zero.
1077  computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q);
1078  computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q);
1079 
1080  Known &= Known2;
1081 
1082  // and(x, add (x, -1)) is a common idiom that always clears the low bit;
1083  // here we handle the more general case of adding any odd number by
1084  // matching the form add(x, add(x, y)) where y is odd.
1085  // TODO: This could be generalized to clearing any bit set in y where the
1086  // following bit is known to be unset in y.
1087  Value *X = nullptr, *Y = nullptr;
1088  if (!Known.Zero[0] && !Known.One[0] &&
1090  Known2.resetAll();
1091  computeKnownBits(Y, DemandedElts, Known2, Depth + 1, Q);
1092  if (Known2.countMinTrailingOnes() > 0)
1093  Known.Zero.setBit(0);
1094  }
1095  break;
1096  }
1097  case Instruction::Or:
1098  computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q);
1099  computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q);
1100 
1101  Known |= Known2;
1102  break;
1103  case Instruction::Xor:
1104  computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q);
1105  computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q);
1106 
1107  Known ^= Known2;
1108  break;
1109  case Instruction::Mul: {
1110  bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1111  computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, DemandedElts,
1112  Known, Known2, Depth, Q);
1113  break;
1114  }
1115  case Instruction::UDiv: {
1116  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1117  computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1118  Known = KnownBits::udiv(Known, Known2);
1119  break;
1120  }
1121  case Instruction::Select: {
1122  const Value *LHS = nullptr, *RHS = nullptr;
1125  computeKnownBits(RHS, Known, Depth + 1, Q);
1126  computeKnownBits(LHS, Known2, Depth + 1, Q);
1127  switch (SPF) {
1128  default:
1129  llvm_unreachable("Unhandled select pattern flavor!");
1130  case SPF_SMAX:
1131  Known = KnownBits::smax(Known, Known2);
1132  break;
1133  case SPF_SMIN:
1134  Known = KnownBits::smin(Known, Known2);
1135  break;
1136  case SPF_UMAX:
1137  Known = KnownBits::umax(Known, Known2);
1138  break;
1139  case SPF_UMIN:
1140  Known = KnownBits::umin(Known, Known2);
1141  break;
1142  }
1143  break;
1144  }
1145 
1146  computeKnownBits(I->getOperand(2), Known, Depth + 1, Q);
1147  computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1148 
1149  // Only known if known in both the LHS and RHS.
1150  Known = KnownBits::commonBits(Known, Known2);
1151 
1152  if (SPF == SPF_ABS) {
1153  // RHS from matchSelectPattern returns the negation part of abs pattern.
1154  // If the negate has an NSW flag we can assume the sign bit of the result
1155  // will be 0 because that makes abs(INT_MIN) undefined.
1156  if (match(RHS, m_Neg(m_Specific(LHS))) &&
1157  Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(RHS)))
1158  Known.Zero.setSignBit();
1159  }
1160 
1161  break;
1162  }
1163  case Instruction::FPTrunc:
1164  case Instruction::FPExt:
1165  case Instruction::FPToUI:
1166  case Instruction::FPToSI:
1167  case Instruction::SIToFP:
1168  case Instruction::UIToFP:
1169  break; // Can't work with floating point.
1170  case Instruction::PtrToInt:
1171  case Instruction::IntToPtr:
1172  // Fall through and handle them the same as zext/trunc.
1174  case Instruction::ZExt:
1175  case Instruction::Trunc: {
1176  Type *SrcTy = I->getOperand(0)->getType();
1177 
1178  unsigned SrcBitWidth;
1179  // Note that we handle pointer operands here because of inttoptr/ptrtoint
1180  // which fall through here.
1181  Type *ScalarTy = SrcTy->getScalarType();
1182  SrcBitWidth = ScalarTy->isPointerTy() ?
1183  Q.DL.getPointerTypeSizeInBits(ScalarTy) :
1184  Q.DL.getTypeSizeInBits(ScalarTy);
1185 
1186  assert(SrcBitWidth && "SrcBitWidth can't be zero");
1187  Known = Known.anyextOrTrunc(SrcBitWidth);
1188  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1189  Known = Known.zextOrTrunc(BitWidth);
1190  break;
1191  }
1192  case Instruction::BitCast: {
1193  Type *SrcTy = I->getOperand(0)->getType();
1194  if (SrcTy->isIntOrPtrTy() &&
1195  // TODO: For now, not handling conversions like:
1196  // (bitcast i64 %x to <2 x i32>)
1197  !I->getType()->isVectorTy()) {
1198  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1199  break;
1200  }
1201 
1202  // Handle cast from vector integer type to scalar or vector integer.
1203  auto *SrcVecTy = dyn_cast<FixedVectorType>(SrcTy);
1204  if (!SrcVecTy || !SrcVecTy->getElementType()->isIntegerTy() ||
1205  !I->getType()->isIntOrIntVectorTy())
1206  break;
1207 
1208  // Look through a cast from narrow vector elements to wider type.
1209  // Examples: v4i32 -> v2i64, v3i8 -> v24
1210  unsigned SubBitWidth = SrcVecTy->getScalarSizeInBits();
1211  if (BitWidth % SubBitWidth == 0) {
1212  // Known bits are automatically intersected across demanded elements of a
1213  // vector. So for example, if a bit is computed as known zero, it must be
1214  // zero across all demanded elements of the vector.
1215  //
1216  // For this bitcast, each demanded element of the output is sub-divided
1217  // across a set of smaller vector elements in the source vector. To get
1218  // the known bits for an entire element of the output, compute the known
1219  // bits for each sub-element sequentially. This is done by shifting the
1220  // one-set-bit demanded elements parameter across the sub-elements for
1221  // consecutive calls to computeKnownBits. We are using the demanded
1222  // elements parameter as a mask operator.
1223  //
1224  // The known bits of each sub-element are then inserted into place
1225  // (dependent on endian) to form the full result of known bits.
1226  unsigned NumElts = DemandedElts.getBitWidth();
1227  unsigned SubScale = BitWidth / SubBitWidth;
1228  APInt SubDemandedElts = APInt::getZero(NumElts * SubScale);
1229  for (unsigned i = 0; i != NumElts; ++i) {
1230  if (DemandedElts[i])
1231  SubDemandedElts.setBit(i * SubScale);
1232  }
1233 
1234  KnownBits KnownSrc(SubBitWidth);
1235  for (unsigned i = 0; i != SubScale; ++i) {
1236  computeKnownBits(I->getOperand(0), SubDemandedElts.shl(i), KnownSrc,
1237  Depth + 1, Q);
1238  unsigned ShiftElt = Q.DL.isLittleEndian() ? i : SubScale - 1 - i;
1239  Known.insertBits(KnownSrc, ShiftElt * SubBitWidth);
1240  }
1241  }
1242  break;
1243  }
1244  case Instruction::SExt: {
1245  // Compute the bits in the result that are not present in the input.
1246  unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1247 
1248  Known = Known.trunc(SrcBitWidth);
1249  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1250  // If the sign bit of the input is known set or clear, then we know the
1251  // top bits of the result.
1252  Known = Known.sext(BitWidth);
1253  break;
1254  }
1255  case Instruction::Shl: {
1256  bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1257  auto KF = [NSW](const KnownBits &KnownVal, const KnownBits &KnownAmt) {
1258  KnownBits Result = KnownBits::shl(KnownVal, KnownAmt);
1259  // If this shift has "nsw" keyword, then the result is either a poison
1260  // value or has the same sign bit as the first operand.
1261  if (NSW) {
1262  if (KnownVal.Zero.isSignBitSet())
1263  Result.Zero.setSignBit();
1264  if (KnownVal.One.isSignBitSet())
1265  Result.One.setSignBit();
1266  }
1267  return Result;
1268  };
1269  computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q,
1270  KF);
1271  // Trailing zeros of a right-shifted constant never decrease.
1272  const APInt *C;
1273  if (match(I->getOperand(0), m_APInt(C)))
1274  Known.Zero.setLowBits(C->countTrailingZeros());
1275  break;
1276  }
1277  case Instruction::LShr: {
1278  auto KF = [](const KnownBits &KnownVal, const KnownBits &KnownAmt) {
1279  return KnownBits::lshr(KnownVal, KnownAmt);
1280  };
1281  computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q,
1282  KF);
1283  // Leading zeros of a left-shifted constant never decrease.
1284  const APInt *C;
1285  if (match(I->getOperand(0), m_APInt(C)))
1286  Known.Zero.setHighBits(C->countLeadingZeros());
1287  break;
1288  }
1289  case Instruction::AShr: {
1290  auto KF = [](const KnownBits &KnownVal, const KnownBits &KnownAmt) {
1291  return KnownBits::ashr(KnownVal, KnownAmt);
1292  };
1293  computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q,
1294  KF);
1295  break;
1296  }
1297  case Instruction::Sub: {
1298  bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1299  computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1300  DemandedElts, Known, Known2, Depth, Q);
1301  break;
1302  }
1303  case Instruction::Add: {
1304  bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1305  computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1306  DemandedElts, Known, Known2, Depth, Q);
1307  break;
1308  }
1309  case Instruction::SRem:
1310  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1311  computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1312  Known = KnownBits::srem(Known, Known2);
1313  break;
1314 
1315  case Instruction::URem:
1316  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1317  computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1318  Known = KnownBits::urem(Known, Known2);
1319  break;
1320  case Instruction::Alloca:
1321  Known.Zero.setLowBits(Log2(cast<AllocaInst>(I)->getAlign()));
1322  break;
1323  case Instruction::GetElementPtr: {
1324  // Analyze all of the subscripts of this getelementptr instruction
1325  // to determine if we can prove known low zero bits.
1326  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1327  // Accumulate the constant indices in a separate variable
1328  // to minimize the number of calls to computeForAddSub.
1329  APInt AccConstIndices(BitWidth, 0, /*IsSigned*/ true);
1330 
1332  for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1333  // TrailZ can only become smaller, short-circuit if we hit zero.
1334  if (Known.isUnknown())
1335  break;
1336 
1337  Value *Index = I->getOperand(i);
1338 
1339  // Handle case when index is zero.
1340  Constant *CIndex = dyn_cast<Constant>(Index);
1341  if (CIndex && CIndex->isZeroValue())
1342  continue;
1343 
1344  if (StructType *STy = GTI.getStructTypeOrNull()) {
1345  // Handle struct member offset arithmetic.
1346 
1347  assert(CIndex &&
1348  "Access to structure field must be known at compile time");
1349 
1350  if (CIndex->getType()->isVectorTy())
1351  Index = CIndex->getSplatValue();
1352 
1353  unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1354  const StructLayout *SL = Q.DL.getStructLayout(STy);
1355  uint64_t Offset = SL->getElementOffset(Idx);
1356  AccConstIndices += Offset;
1357  continue;
1358  }
1359 
1360  // Handle array index arithmetic.
1361  Type *IndexedTy = GTI.getIndexedType();
1362  if (!IndexedTy->isSized()) {
1363  Known.resetAll();
1364  break;
1365  }
1366 
1367  unsigned IndexBitWidth = Index->getType()->getScalarSizeInBits();
1368  KnownBits IndexBits(IndexBitWidth);
1369  computeKnownBits(Index, IndexBits, Depth + 1, Q);
1370  TypeSize IndexTypeSize = Q.DL.getTypeAllocSize(IndexedTy);
1371  uint64_t TypeSizeInBytes = IndexTypeSize.getKnownMinSize();
1372  KnownBits ScalingFactor(IndexBitWidth);
1373  // Multiply by current sizeof type.
1374  // &A[i] == A + i * sizeof(*A[i]).
1375  if (IndexTypeSize.isScalable()) {
1376  // For scalable types the only thing we know about sizeof is
1377  // that this is a multiple of the minimum size.
1378  ScalingFactor.Zero.setLowBits(countTrailingZeros(TypeSizeInBytes));
1379  } else if (IndexBits.isConstant()) {
1380  APInt IndexConst = IndexBits.getConstant();
1381  APInt ScalingFactor(IndexBitWidth, TypeSizeInBytes);
1382  IndexConst *= ScalingFactor;
1383  AccConstIndices += IndexConst.sextOrTrunc(BitWidth);
1384  continue;
1385  } else {
1386  ScalingFactor =
1387  KnownBits::makeConstant(APInt(IndexBitWidth, TypeSizeInBytes));
1388  }
1389  IndexBits = KnownBits::mul(IndexBits, ScalingFactor);
1390 
1391  // If the offsets have a different width from the pointer, according
1392  // to the language reference we need to sign-extend or truncate them
1393  // to the width of the pointer.
1394  IndexBits = IndexBits.sextOrTrunc(BitWidth);
1395 
1396  // Note that inbounds does *not* guarantee nsw for the addition, as only
1397  // the offset is signed, while the base address is unsigned.
1399  /*Add=*/true, /*NSW=*/false, Known, IndexBits);
1400  }
1401  if (!Known.isUnknown() && !AccConstIndices.isZero()) {
1402  KnownBits Index = KnownBits::makeConstant(AccConstIndices);
1404  /*Add=*/true, /*NSW=*/false, Known, Index);
1405  }
1406  break;
1407  }
1408  case Instruction::PHI: {
1409  const PHINode *P = cast<PHINode>(I);
1410  BinaryOperator *BO = nullptr;
1411  Value *R = nullptr, *L = nullptr;
1412  if (matchSimpleRecurrence(P, BO, R, L)) {
1413  // Handle the case of a simple two-predecessor recurrence PHI.
1414  // There's a lot more that could theoretically be done here, but
1415  // this is sufficient to catch some interesting cases.
1416  unsigned Opcode = BO->getOpcode();
1417 
1418  // If this is a shift recurrence, we know the bits being shifted in.
1419  // We can combine that with information about the start value of the
1420  // recurrence to conclude facts about the result.
1421  if ((Opcode == Instruction::LShr || Opcode == Instruction::AShr ||
1422  Opcode == Instruction::Shl) &&
1423  BO->getOperand(0) == I) {
1424 
1425  // We have matched a recurrence of the form:
1426  // %iv = [R, %entry], [%iv.next, %backedge]
1427  // %iv.next = shift_op %iv, L
1428 
1429  // Recurse with the phi context to avoid concern about whether facts
1430  // inferred hold at original context instruction. TODO: It may be
1431  // correct to use the original context. IF warranted, explore and
1432  // add sufficient tests to cover.
1433  Query RecQ = Q;
1434  RecQ.CxtI = P;
1435  computeKnownBits(R, DemandedElts, Known2, Depth + 1, RecQ);
1436  switch (Opcode) {
1437  case Instruction::Shl:
1438  // A shl recurrence will only increase the tailing zeros
1439  Known.Zero.setLowBits(Known2.countMinTrailingZeros());
1440  break;
1441  case Instruction::LShr:
1442  // A lshr recurrence will preserve the leading zeros of the
1443  // start value
1444  Known.Zero.setHighBits(Known2.countMinLeadingZeros());
1445  break;
1446  case Instruction::AShr:
1447  // An ashr recurrence will extend the initial sign bit
1448  Known.Zero.setHighBits(Known2.countMinLeadingZeros());
1449  Known.One.setHighBits(Known2.countMinLeadingOnes());
1450  break;
1451  };
1452  }
1453 
1454  // Check for operations that have the property that if
1455  // both their operands have low zero bits, the result
1456  // will have low zero bits.
1457  if (Opcode == Instruction::Add ||
1458  Opcode == Instruction::Sub ||
1459  Opcode == Instruction::And ||
1460  Opcode == Instruction::Or ||
1461  Opcode == Instruction::Mul) {
1462  // Change the context instruction to the "edge" that flows into the
1463  // phi. This is important because that is where the value is actually
1464  // "evaluated" even though it is used later somewhere else. (see also
1465  // D69571).
1466  Query RecQ = Q;
1467 
1468  unsigned OpNum = P->getOperand(0) == R ? 0 : 1;
1469  Instruction *RInst = P->getIncomingBlock(OpNum)->getTerminator();
1470  Instruction *LInst = P->getIncomingBlock(1-OpNum)->getTerminator();
1471 
1472  // Ok, we have a PHI of the form L op= R. Check for low
1473  // zero bits.
1474  RecQ.CxtI = RInst;
1475  computeKnownBits(R, Known2, Depth + 1, RecQ);
1476 
1477  // We need to take the minimum number of known bits
1478  KnownBits Known3(BitWidth);
1479  RecQ.CxtI = LInst;
1480  computeKnownBits(L, Known3, Depth + 1, RecQ);
1481 
1483  Known3.countMinTrailingZeros()));
1484 
1485  auto *OverflowOp = dyn_cast<OverflowingBinaryOperator>(BO);
1486  if (OverflowOp && Q.IIQ.hasNoSignedWrap(OverflowOp)) {
1487  // If initial value of recurrence is nonnegative, and we are adding
1488  // a nonnegative number with nsw, the result can only be nonnegative
1489  // or poison value regardless of the number of times we execute the
1490  // add in phi recurrence. If initial value is negative and we are
1491  // adding a negative number with nsw, the result can only be
1492  // negative or poison value. Similar arguments apply to sub and mul.
1493  //
1494  // (add non-negative, non-negative) --> non-negative
1495  // (add negative, negative) --> negative
1496  if (Opcode == Instruction::Add) {
1497  if (Known2.isNonNegative() && Known3.isNonNegative())
1498  Known.makeNonNegative();
1499  else if (Known2.isNegative() && Known3.isNegative())
1500  Known.makeNegative();
1501  }
1502 
1503  // (sub nsw non-negative, negative) --> non-negative
1504  // (sub nsw negative, non-negative) --> negative
1505  else if (Opcode == Instruction::Sub && BO->getOperand(0) == I) {
1506  if (Known2.isNonNegative() && Known3.isNegative())
1507  Known.makeNonNegative();
1508  else if (Known2.isNegative() && Known3.isNonNegative())
1509  Known.makeNegative();
1510  }
1511 
1512  // (mul nsw non-negative, non-negative) --> non-negative
1513  else if (Opcode == Instruction::Mul && Known2.isNonNegative() &&
1514  Known3.isNonNegative())
1515  Known.makeNonNegative();
1516  }
1517 
1518  break;
1519  }
1520  }
1521 
1522  // Unreachable blocks may have zero-operand PHI nodes.
1523  if (P->getNumIncomingValues() == 0)
1524  break;
1525 
1526  // Otherwise take the unions of the known bit sets of the operands,
1527  // taking conservative care to avoid excessive recursion.
1528  if (Depth < MaxAnalysisRecursionDepth - 1 && !Known.Zero && !Known.One) {
1529  // Skip if every incoming value references to ourself.
1530  if (isa_and_nonnull<UndefValue>(P->hasConstantValue()))
1531  break;
1532 
1533  Known.Zero.setAllBits();
1534  Known.One.setAllBits();
1535  for (unsigned u = 0, e = P->getNumIncomingValues(); u < e; ++u) {
1536  Value *IncValue = P->getIncomingValue(u);
1537  // Skip direct self references.
1538  if (IncValue == P) continue;
1539 
1540  // Change the context instruction to the "edge" that flows into the
1541  // phi. This is important because that is where the value is actually
1542  // "evaluated" even though it is used later somewhere else. (see also
1543  // D69571).
1544  Query RecQ = Q;
1545  RecQ.CxtI = P->getIncomingBlock(u)->getTerminator();
1546 
1547  Known2 = KnownBits(BitWidth);
1548  // Recurse, but cap the recursion to one level, because we don't
1549  // want to waste time spinning around in loops.
1550  computeKnownBits(IncValue, Known2, MaxAnalysisRecursionDepth - 1, RecQ);
1551  Known = KnownBits::commonBits(Known, Known2);
1552  // If all bits have been ruled out, there's no need to check
1553  // more operands.
1554  if (Known.isUnknown())
1555  break;
1556  }
1557  }
1558  break;
1559  }
1560  case Instruction::Call:
1561  case Instruction::Invoke:
1562  // If range metadata is attached to this call, set known bits from that,
1563  // and then intersect with known bits based on other properties of the
1564  // function.
1565  if (MDNode *MD =
1566  Q.IIQ.getMetadata(cast<Instruction>(I), LLVMContext::MD_range))
1568  if (const Value *RV = cast<CallBase>(I)->getReturnedArgOperand()) {
1569  computeKnownBits(RV, Known2, Depth + 1, Q);
1570  Known.Zero |= Known2.Zero;
1571  Known.One |= Known2.One;
1572  }
1573  if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1574  switch (II->getIntrinsicID()) {
1575  default: break;
1576  case Intrinsic::abs: {
1577  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1578  bool IntMinIsPoison = match(II->getArgOperand(1), m_One());
1579  Known = Known2.abs(IntMinIsPoison);
1580  break;
1581  }
1582  case Intrinsic::bitreverse:
1583  computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q);
1584  Known.Zero |= Known2.Zero.reverseBits();
1585  Known.One |= Known2.One.reverseBits();
1586  break;
1587  case Intrinsic::bswap:
1588  computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q);
1589  Known.Zero |= Known2.Zero.byteSwap();
1590  Known.One |= Known2.One.byteSwap();
1591  break;
1592  case Intrinsic::ctlz: {
1593  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1594  // If we have a known 1, its position is our upper bound.
1595  unsigned PossibleLZ = Known2.countMaxLeadingZeros();
1596  // If this call is undefined for 0, the result will be less than 2^n.
1597  if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1598  PossibleLZ = std::min(PossibleLZ, BitWidth - 1);
1599  unsigned LowBits = Log2_32(PossibleLZ)+1;
1600  Known.Zero.setBitsFrom(LowBits);
1601  break;
1602  }
1603  case Intrinsic::cttz: {
1604  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1605  // If we have a known 1, its position is our upper bound.
1606  unsigned PossibleTZ = Known2.countMaxTrailingZeros();
1607  // If this call is undefined for 0, the result will be less than 2^n.
1608  if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1609  PossibleTZ = std::min(PossibleTZ, BitWidth - 1);
1610  unsigned LowBits = Log2_32(PossibleTZ)+1;
1611  Known.Zero.setBitsFrom(LowBits);
1612  break;
1613  }
1614  case Intrinsic::ctpop: {
1615  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1616  // We can bound the space the count needs. Also, bits known to be zero
1617  // can't contribute to the population.
1618  unsigned BitsPossiblySet = Known2.countMaxPopulation();
1619  unsigned LowBits = Log2_32(BitsPossiblySet)+1;
1620  Known.Zero.setBitsFrom(LowBits);
1621  // TODO: we could bound KnownOne using the lower bound on the number
1622  // of bits which might be set provided by popcnt KnownOne2.
1623  break;
1624  }
1625  case Intrinsic::fshr:
1626  case Intrinsic::fshl: {
1627  const APInt *SA;
1628  if (!match(I->getOperand(2), m_APInt(SA)))
1629  break;
1630 
1631  // Normalize to funnel shift left.
1632  uint64_t ShiftAmt = SA->urem(BitWidth);
1633  if (II->getIntrinsicID() == Intrinsic::fshr)
1634  ShiftAmt = BitWidth - ShiftAmt;
1635 
1636  KnownBits Known3(BitWidth);
1637  computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1638  computeKnownBits(I->getOperand(1), Known3, Depth + 1, Q);
1639 
1640  Known.Zero =
1641  Known2.Zero.shl(ShiftAmt) | Known3.Zero.lshr(BitWidth - ShiftAmt);
1642  Known.One =
1643  Known2.One.shl(ShiftAmt) | Known3.One.lshr(BitWidth - ShiftAmt);
1644  break;
1645  }
1646  case Intrinsic::uadd_sat:
1647  case Intrinsic::usub_sat: {
1648  bool IsAdd = II->getIntrinsicID() == Intrinsic::uadd_sat;
1649  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1650  computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1651 
1652  // Add: Leading ones of either operand are preserved.
1653  // Sub: Leading zeros of LHS and leading ones of RHS are preserved
1654  // as leading zeros in the result.
1655  unsigned LeadingKnown;
1656  if (IsAdd)
1657  LeadingKnown = std::max(Known.countMinLeadingOnes(),
1658  Known2.countMinLeadingOnes());
1659  else
1660  LeadingKnown = std::max(Known.countMinLeadingZeros(),
1661  Known2.countMinLeadingOnes());
1662 
1664  IsAdd, /* NSW */ false, Known, Known2);
1665 
1666  // We select between the operation result and all-ones/zero
1667  // respectively, so we can preserve known ones/zeros.
1668  if (IsAdd) {
1669  Known.One.setHighBits(LeadingKnown);
1670  Known.Zero.clearAllBits();
1671  } else {
1672  Known.Zero.setHighBits(LeadingKnown);
1673  Known.One.clearAllBits();
1674  }
1675  break;
1676  }
1677  case Intrinsic::umin:
1678  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1679  computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1680  Known = KnownBits::umin(Known, Known2);
1681  break;
1682  case Intrinsic::umax:
1683  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1684  computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1685  Known = KnownBits::umax(Known, Known2);
1686  break;
1687  case Intrinsic::smin:
1688  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1689  computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1690  Known = KnownBits::smin(Known, Known2);
1691  break;
1692  case Intrinsic::smax:
1693  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1694  computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1695  Known = KnownBits::smax(Known, Known2);
1696  break;
1697  case Intrinsic::x86_sse42_crc32_64_64:
1698  Known.Zero.setBitsFrom(32);
1699  break;
1700  case Intrinsic::riscv_vsetvli:
1701  case Intrinsic::riscv_vsetvlimax:
1702  // Assume that VL output is positive and would fit in an int32_t.
1703  // TODO: VLEN might be capped at 16 bits in a future V spec update.
1704  if (BitWidth >= 32)
1705  Known.Zero.setBitsFrom(31);
1706  break;
1707  case Intrinsic::vscale: {
1708  if (!II->getParent() || !II->getFunction() ||
1709  !II->getFunction()->hasFnAttribute(Attribute::VScaleRange))
1710  break;
1711 
1712  auto Attr = II->getFunction()->getFnAttribute(Attribute::VScaleRange);
1713  Optional<unsigned> VScaleMax = Attr.getVScaleRangeMax();
1714 
1715  if (!VScaleMax)
1716  break;
1717 
1718  unsigned VScaleMin = Attr.getVScaleRangeMin();
1719 
1720  // If vscale min = max then we know the exact value at compile time
1721  // and hence we know the exact bits.
1722  if (VScaleMin == VScaleMax) {
1723  Known.One = VScaleMin;
1724  Known.Zero = VScaleMin;
1725  Known.Zero.flipAllBits();
1726  break;
1727  }
1728 
1729  unsigned FirstZeroHighBit =
1730  32 - countLeadingZeros(VScaleMax.getValue());
1731  if (FirstZeroHighBit < BitWidth)
1732  Known.Zero.setBitsFrom(FirstZeroHighBit);
1733 
1734  break;
1735  }
1736  }
1737  }
1738  break;
1739  case Instruction::ShuffleVector: {
1740  auto *Shuf = dyn_cast<ShuffleVectorInst>(I);
1741  // FIXME: Do we need to handle ConstantExpr involving shufflevectors?
1742  if (!Shuf) {
1743  Known.resetAll();
1744  return;
1745  }
1746  // For undef elements, we don't know anything about the common state of
1747  // the shuffle result.
1748  APInt DemandedLHS, DemandedRHS;
1749  if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) {
1750  Known.resetAll();
1751  return;
1752  }
1753  Known.One.setAllBits();
1754  Known.Zero.setAllBits();
1755  if (!!DemandedLHS) {
1756  const Value *LHS = Shuf->getOperand(0);
1757  computeKnownBits(LHS, DemandedLHS, Known, Depth + 1, Q);
1758  // If we don't know any bits, early out.
1759  if (Known.isUnknown())
1760  break;
1761  }
1762  if (!!DemandedRHS) {
1763  const Value *RHS = Shuf->getOperand(1);
1764  computeKnownBits(RHS, DemandedRHS, Known2, Depth + 1, Q);
1765  Known = KnownBits::commonBits(Known, Known2);
1766  }
1767  break;
1768  }
1769  case Instruction::InsertElement: {
1770  const Value *Vec = I->getOperand(0);
1771  const Value *Elt = I->getOperand(1);
1772  auto *CIdx = dyn_cast<ConstantInt>(I->getOperand(2));
1773  // Early out if the index is non-constant or out-of-range.
1774  unsigned NumElts = DemandedElts.getBitWidth();
1775  if (!CIdx || CIdx->getValue().uge(NumElts)) {
1776  Known.resetAll();
1777  return;
1778  }
1779  Known.One.setAllBits();
1780  Known.Zero.setAllBits();
1781  unsigned EltIdx = CIdx->getZExtValue();
1782  // Do we demand the inserted element?
1783  if (DemandedElts[EltIdx]) {
1784  computeKnownBits(Elt, Known, Depth + 1, Q);
1785  // If we don't know any bits, early out.
1786  if (Known.isUnknown())
1787  break;
1788  }
1789  // We don't need the base vector element that has been inserted.
1790  APInt DemandedVecElts = DemandedElts;
1791  DemandedVecElts.clearBit(EltIdx);
1792  if (!!DemandedVecElts) {
1793  computeKnownBits(Vec, DemandedVecElts, Known2, Depth + 1, Q);
1794  Known = KnownBits::commonBits(Known, Known2);
1795  }
1796  break;
1797  }
1798  case Instruction::ExtractElement: {
1799  // Look through extract element. If the index is non-constant or
1800  // out-of-range demand all elements, otherwise just the extracted element.
1801  const Value *Vec = I->getOperand(0);
1802  const Value *Idx = I->getOperand(1);
1803  auto *CIdx = dyn_cast<ConstantInt>(Idx);
1804  if (isa<ScalableVectorType>(Vec->getType())) {
1805  // FIXME: there's probably *something* we can do with scalable vectors
1806  Known.resetAll();
1807  break;
1808  }
1809  unsigned NumElts = cast<FixedVectorType>(Vec->getType())->getNumElements();
1810  APInt DemandedVecElts = APInt::getAllOnes(NumElts);
1811  if (CIdx && CIdx->getValue().ult(NumElts))
1812  DemandedVecElts = APInt::getOneBitSet(NumElts, CIdx->getZExtValue());
1813  computeKnownBits(Vec, DemandedVecElts, Known, Depth + 1, Q);
1814  break;
1815  }
1816  case Instruction::ExtractValue:
1817  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1818  const ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1819  if (EVI->getNumIndices() != 1) break;
1820  if (EVI->getIndices()[0] == 0) {
1821  switch (II->getIntrinsicID()) {
1822  default: break;
1823  case Intrinsic::uadd_with_overflow:
1824  case Intrinsic::sadd_with_overflow:
1825  computeKnownBitsAddSub(true, II->getArgOperand(0),
1826  II->getArgOperand(1), false, DemandedElts,
1827  Known, Known2, Depth, Q);
1828  break;
1829  case Intrinsic::usub_with_overflow:
1830  case Intrinsic::ssub_with_overflow:
1831  computeKnownBitsAddSub(false, II->getArgOperand(0),
1832  II->getArgOperand(1), false, DemandedElts,
1833  Known, Known2, Depth, Q);
1834  break;
1835  case Intrinsic::umul_with_overflow:
1836  case Intrinsic::smul_with_overflow:
1837  computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1838  DemandedElts, Known, Known2, Depth, Q);
1839  break;
1840  }
1841  }
1842  }
1843  break;
1844  case Instruction::Freeze:
1845  if (isGuaranteedNotToBePoison(I->getOperand(0), Q.AC, Q.CxtI, Q.DT,
1846  Depth + 1))
1847  computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1848  break;
1849  }
1850 }
1851 
1852 /// Determine which bits of V are known to be either zero or one and return
1853 /// them.
1854 KnownBits computeKnownBits(const Value *V, const APInt &DemandedElts,
1855  unsigned Depth, const Query &Q) {
1856  KnownBits Known(getBitWidth(V->getType(), Q.DL));
1857  computeKnownBits(V, DemandedElts, Known, Depth, Q);
1858  return Known;
1859 }
1860 
1861 /// Determine which bits of V are known to be either zero or one and return
1862 /// them.
1863 KnownBits computeKnownBits(const Value *V, unsigned Depth, const Query &Q) {
1864  KnownBits Known(getBitWidth(V->getType(), Q.DL));
1865  computeKnownBits(V, Known, Depth, Q);
1866  return Known;
1867 }
1868 
1869 /// Determine which bits of V are known to be either zero or one and return
1870 /// them in the Known bit set.
1871 ///
1872 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
1873 /// we cannot optimize based on the assumption that it is zero without changing
1874 /// it to be an explicit zero. If we don't change it to zero, other code could
1875 /// optimized based on the contradictory assumption that it is non-zero.
1876 /// Because instcombine aggressively folds operations with undef args anyway,
1877 /// this won't lose us code quality.
1878 ///
1879 /// This function is defined on values with integer type, values with pointer
1880 /// type, and vectors of integers. In the case
1881 /// where V is a vector, known zero, and known one values are the
1882 /// same width as the vector element, and the bit is set only if it is true
1883 /// for all of the demanded elements in the vector specified by DemandedElts.
1884 void computeKnownBits(const Value *V, const APInt &DemandedElts,
1885  KnownBits &Known, unsigned Depth, const Query &Q) {
1886  if (!DemandedElts || isa<ScalableVectorType>(V->getType())) {
1887  // No demanded elts or V is a scalable vector, better to assume we don't
1888  // know anything.
1889  Known.resetAll();
1890  return;
1891  }
1892 
1893  assert(V && "No Value?");
1894  assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth");
1895 
1896 #ifndef NDEBUG
1897  Type *Ty = V->getType();
1898  unsigned BitWidth = Known.getBitWidth();
1899 
1901  "Not integer or pointer type!");
1902 
1903  if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) {
1904  assert(
1905  FVTy->getNumElements() == DemandedElts.getBitWidth() &&
1906  "DemandedElt width should equal the fixed vector number of elements");
1907  } else {
1908  assert(DemandedElts == APInt(1, 1) &&
1909  "DemandedElt width should be 1 for scalars");
1910  }
1911 
1912  Type *ScalarTy = Ty->getScalarType();
1913  if (ScalarTy->isPointerTy()) {
1914  assert(BitWidth == Q.DL.getPointerTypeSizeInBits(ScalarTy) &&
1915  "V and Known should have same BitWidth");
1916  } else {
1917  assert(BitWidth == Q.DL.getTypeSizeInBits(ScalarTy) &&
1918  "V and Known should have same BitWidth");
1919  }
1920 #endif
1921 
1922  const APInt *C;
1923  if (match(V, m_APInt(C))) {
1924  // We know all of the bits for a scalar constant or a splat vector constant!
1925  Known = KnownBits::makeConstant(*C);
1926  return;
1927  }
1928  // Null and aggregate-zero are all-zeros.
1929  if (isa<ConstantPointerNull>(V) || isa<ConstantAggregateZero>(V)) {
1930  Known.setAllZero();
1931  return;
1932  }
1933  // Handle a constant vector by taking the intersection of the known bits of
1934  // each element.
1935  if (const ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(V)) {
1936  // We know that CDV must be a vector of integers. Take the intersection of
1937  // each element.
1938  Known.Zero.setAllBits(); Known.One.setAllBits();
1939  for (unsigned i = 0, e = CDV->getNumElements(); i != e; ++i) {
1940  if (!DemandedElts[i])
1941  continue;
1942  APInt Elt = CDV->getElementAsAPInt(i);
1943  Known.Zero &= ~Elt;
1944  Known.One &= Elt;
1945  }
1946  return;
1947  }
1948 
1949  if (const auto *CV = dyn_cast<ConstantVector>(V)) {
1950  // We know that CV must be a vector of integers. Take the intersection of
1951  // each element.
1952  Known.Zero.setAllBits(); Known.One.setAllBits();
1953  for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
1954  if (!DemandedElts[i])
1955  continue;
1956  Constant *Element = CV->getAggregateElement(i);
1957  auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);
1958  if (!ElementCI) {
1959  Known.resetAll();
1960  return;
1961  }
1962  const APInt &Elt = ElementCI->getValue();
1963  Known.Zero &= ~Elt;
1964  Known.One &= Elt;
1965  }
1966  return;
1967  }
1968 
1969  // Start out not knowing anything.
1970  Known.resetAll();
1971 
1972  // We can't imply anything about undefs.
1973  if (isa<UndefValue>(V))
1974  return;
1975 
1976  // There's no point in looking through other users of ConstantData for
1977  // assumptions. Confirm that we've handled them all.
1978  assert(!isa<ConstantData>(V) && "Unhandled constant data!");
1979 
1980  // All recursive calls that increase depth must come after this.
1982  return;
1983 
1984  // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1985  // the bits of its aliasee.
1986  if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1987  if (!GA->isInterposable())
1988  computeKnownBits(GA->getAliasee(), Known, Depth + 1, Q);
1989  return;
1990  }
1991 
1992  if (const Operator *I = dyn_cast<Operator>(V))
1993  computeKnownBitsFromOperator(I, DemandedElts, Known, Depth, Q);
1994 
1995  // Aligned pointers have trailing zeros - refine Known.Zero set
1996  if (isa<PointerType>(V->getType())) {
1997  Align Alignment = V->getPointerAlignment(Q.DL);
1998  Known.Zero.setLowBits(Log2(Alignment));
1999  }
2000 
2001  // computeKnownBitsFromAssume strictly refines Known.
2002  // Therefore, we run them after computeKnownBitsFromOperator.
2003 
2004  // Check whether a nearby assume intrinsic can determine some known bits.
2005  computeKnownBitsFromAssume(V, Known, Depth, Q);
2006 
2007  assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?");
2008 }
2009 
2010 /// Return true if the given value is known to have exactly one
2011 /// bit set when defined. For vectors return true if every element is known to
2012 /// be a power of two when defined. Supports values with integer or pointer
2013 /// types and vectors of integers.
2014 bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
2015  const Query &Q) {
2016  assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth");
2017 
2018  // Attempt to match against constants.
2019  if (OrZero && match(V, m_Power2OrZero()))
2020  return true;
2021  if (match(V, m_Power2()))
2022  return true;
2023 
2024  // 1 << X is clearly a power of two if the one is not shifted off the end. If
2025  // it is shifted off the end then the result is undefined.
2026  if (match(V, m_Shl(m_One(), m_Value())))
2027  return true;
2028 
2029  // (signmask) >>l X is clearly a power of two if the one is not shifted off
2030  // the bottom. If it is shifted off the bottom then the result is undefined.
2031  if (match(V, m_LShr(m_SignMask(), m_Value())))
2032  return true;
2033 
2034  // The remaining tests are all recursive, so bail out if we hit the limit.
2036  return false;
2037 
2038  Value *X = nullptr, *Y = nullptr;
2039  // A shift left or a logical shift right of a power of two is a power of two
2040  // or zero.
2041  if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
2042  match(V, m_LShr(m_Value(X), m_Value()))))
2043  return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q);
2044 
2045  if (const ZExtInst *ZI = dyn_cast<ZExtInst>(V))
2046  return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
2047 
2048  if (const SelectInst *SI = dyn_cast<SelectInst>(V))
2049  return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
2050  isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
2051 
2052  // Peek through min/max.
2053  if (match(V, m_MaxOrMin(m_Value(X), m_Value(Y)))) {
2054  return isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q) &&
2055  isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q);
2056  }
2057 
2058  if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
2059  // A power of two and'd with anything is a power of two or zero.
2060  if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q) ||
2061  isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q))
2062  return true;
2063  // X & (-X) is always a power of two or zero.
2064  if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
2065  return true;
2066  return false;
2067  }
2068 
2069  // Adding a power-of-two or zero to the same power-of-two or zero yields
2070  // either the original power-of-two, a larger power-of-two or zero.
2071  if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
2072  const OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
2073  if (OrZero || Q.IIQ.hasNoUnsignedWrap(VOBO) ||
2074  Q.IIQ.hasNoSignedWrap(VOBO)) {
2075  if (match(X, m_And(m_Specific(Y), m_Value())) ||
2076  match(X, m_And(m_Value(), m_Specific(Y))))
2077  if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
2078  return true;
2079  if (match(Y, m_And(m_Specific(X), m_Value())) ||
2080  match(Y, m_And(m_Value(), m_Specific(X))))
2081  if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
2082  return true;
2083 
2084  unsigned BitWidth = V->getType()->getScalarSizeInBits();
2085  KnownBits LHSBits(BitWidth);
2086  computeKnownBits(X, LHSBits, Depth, Q);
2087 
2088  KnownBits RHSBits(BitWidth);
2089  computeKnownBits(Y, RHSBits, Depth, Q);
2090  // If i8 V is a power of two or zero:
2091  // ZeroBits: 1 1 1 0 1 1 1 1
2092  // ~ZeroBits: 0 0 0 1 0 0 0 0
2093  if ((~(LHSBits.Zero & RHSBits.Zero)).isPowerOf2())
2094  // If OrZero isn't set, we cannot give back a zero result.
2095  // Make sure either the LHS or RHS has a bit set.
2096  if (OrZero || RHSBits.One.getBoolValue() || LHSBits.One.getBoolValue())
2097  return true;
2098  }
2099  }
2100 
2101  // An exact divide or right shift can only shift off zero bits, so the result
2102  // is a power of two only if the first operand is a power of two and not
2103  // copying a sign bit (sdiv int_min, 2).
2104  if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
2105  match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
2106  return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
2107  Depth, Q);
2108  }
2109 
2110  return false;
2111 }
2112 
2113 /// Test whether a GEP's result is known to be non-null.
2114 ///
2115 /// Uses properties inherent in a GEP to try to determine whether it is known
2116 /// to be non-null.
2117 ///
2118 /// Currently this routine does not support vector GEPs.
2119 static bool isGEPKnownNonNull(const GEPOperator *GEP, unsigned Depth,
2120  const Query &Q) {
2121  const Function *F = nullptr;
2122  if (const Instruction *I = dyn_cast<Instruction>(GEP))
2123  F = I->getFunction();
2124 
2125  if (!GEP->isInBounds() ||
2126  NullPointerIsDefined(F, GEP->getPointerAddressSpace()))
2127  return false;
2128 
2129  // FIXME: Support vector-GEPs.
2130  assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
2131 
2132  // If the base pointer is non-null, we cannot walk to a null address with an
2133  // inbounds GEP in address space zero.
2134  if (isKnownNonZero(GEP->getPointerOperand(), Depth, Q))
2135  return true;
2136 
2137  // Walk the GEP operands and see if any operand introduces a non-zero offset.
2138  // If so, then the GEP cannot produce a null pointer, as doing so would
2139  // inherently violate the inbounds contract within address space zero.
2140  for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
2141  GTI != GTE; ++GTI) {
2142  // Struct types are easy -- they must always be indexed by a constant.
2143  if (StructType *STy = GTI.getStructTypeOrNull()) {
2144  ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
2145  unsigned ElementIdx = OpC->getZExtValue();
2146  const StructLayout *SL = Q.DL.getStructLayout(STy);
2147  uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
2148  if (ElementOffset > 0)
2149  return true;
2150  continue;
2151  }
2152 
2153  // If we have a zero-sized type, the index doesn't matter. Keep looping.
2154  if (Q.DL.getTypeAllocSize(GTI.getIndexedType()).getKnownMinSize() == 0)
2155  continue;
2156 
2157  // Fast path the constant operand case both for efficiency and so we don't
2158  // increment Depth when just zipping down an all-constant GEP.
2159  if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
2160  if (!OpC->isZero())
2161  return true;
2162  continue;
2163  }
2164 
2165  // We post-increment Depth here because while isKnownNonZero increments it
2166  // as well, when we pop back up that increment won't persist. We don't want
2167  // to recurse 10k times just because we have 10k GEP operands. We don't
2168  // bail completely out because we want to handle constant GEPs regardless
2169  // of depth.
2171  continue;
2172 
2173  if (isKnownNonZero(GTI.getOperand(), Depth, Q))
2174  return true;
2175  }
2176 
2177  return false;
2178 }
2179 
2181  const Instruction *CtxI,
2182  const DominatorTree *DT) {
2183  if (isa<Constant>(V))
2184  return false;
2185 
2186  if (!CtxI || !DT)
2187  return false;
2188 
2189  unsigned NumUsesExplored = 0;
2190  for (auto *U : V->users()) {
2191  // Avoid massive lists
2192  if (NumUsesExplored >= DomConditionsMaxUses)
2193  break;
2194  NumUsesExplored++;
2195 
2196  // If the value is used as an argument to a call or invoke, then argument
2197  // attributes may provide an answer about null-ness.
2198  if (const auto *CB = dyn_cast<CallBase>(U))
2199  if (auto *CalledFunc = CB->getCalledFunction())
2200  for (const Argument &Arg : CalledFunc->args())
2201  if (CB->getArgOperand(Arg.getArgNo()) == V &&
2202  Arg.hasNonNullAttr(/* AllowUndefOrPoison */ false) &&
2203  DT->dominates(CB, CtxI))
2204  return true;
2205 
2206  // If the value is used as a load/store, then the pointer must be non null.
2207  if (V == getLoadStorePointerOperand(U)) {
2208  const Instruction *I = cast<Instruction>(U);
2209  if (!NullPointerIsDefined(I->getFunction(),
2210  V->getType()->getPointerAddressSpace()) &&
2211  DT->dominates(I, CtxI))
2212  return true;
2213  }
2214 
2215  // Consider only compare instructions uniquely controlling a branch
2216  Value *RHS;
2217  CmpInst::Predicate Pred;
2218  if (!match(U, m_c_ICmp(Pred, m_Specific(V), m_Value(RHS))))
2219  continue;
2220 
2221  bool NonNullIfTrue;
2222  if (cmpExcludesZero(Pred, RHS))
2223  NonNullIfTrue = true;
2225  NonNullIfTrue = false;
2226  else
2227  continue;
2228 
2231  for (auto *CmpU : U->users()) {
2232  assert(WorkList.empty() && "Should be!");
2233  if (Visited.insert(CmpU).second)
2234  WorkList.push_back(CmpU);
2235 
2236  while (!WorkList.empty()) {
2237  auto *Curr = WorkList.pop_back_val();
2238 
2239  // If a user is an AND, add all its users to the work list. We only
2240  // propagate "pred != null" condition through AND because it is only
2241  // correct to assume that all conditions of AND are met in true branch.
2242  // TODO: Support similar logic of OR and EQ predicate?
2243  if (NonNullIfTrue)
2244  if (match(Curr, m_LogicalAnd(m_Value(), m_Value()))) {
2245  for (auto *CurrU : Curr->users())
2246  if (Visited.insert(CurrU).second)
2247  WorkList.push_back(CurrU);
2248  continue;
2249  }
2250 
2251  if (const BranchInst *BI = dyn_cast<BranchInst>(Curr)) {
2252  assert(BI->isConditional() && "uses a comparison!");
2253 
2254  BasicBlock *NonNullSuccessor =
2255  BI->getSuccessor(NonNullIfTrue ? 0 : 1);
2256  BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
2257  if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
2258  return true;
2259  } else if (NonNullIfTrue && isGuard(Curr) &&
2260  DT->dominates(cast<Instruction>(Curr), CtxI)) {
2261  return true;
2262  }
2263  }
2264  }
2265  }
2266 
2267  return false;
2268 }
2269 
2270 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
2271 /// ensure that the value it's attached to is never Value? 'RangeType' is
2272 /// is the type of the value described by the range.
2273 static bool rangeMetadataExcludesValue(const MDNode* Ranges, const APInt& Value) {
2274  const unsigned NumRanges = Ranges->getNumOperands() / 2;
2275  assert(NumRanges >= 1);
2276  for (unsigned i = 0; i < NumRanges; ++i) {
2277  ConstantInt *Lower =
2278  mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
2279  ConstantInt *Upper =
2280  mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
2281  ConstantRange Range(Lower->getValue(), Upper->getValue());
2282  if (Range.contains(Value))
2283  return false;
2284  }
2285  return true;
2286 }
2287 
2288 /// Try to detect a recurrence that monotonically increases/decreases from a
2289 /// non-zero starting value. These are common as induction variables.
2290 static bool isNonZeroRecurrence(const PHINode *PN) {
2291  BinaryOperator *BO = nullptr;
2292  Value *Start = nullptr, *Step = nullptr;
2293  const APInt *StartC, *StepC;
2294  if (!matchSimpleRecurrence(PN, BO, Start, Step) ||
2295  !match(Start, m_APInt(StartC)) || StartC->isZero())
2296  return false;
2297 
2298  switch (BO->getOpcode()) {
2299  case Instruction::Add:
2300  // Starting from non-zero and stepping away from zero can never wrap back
2301  // to zero.
2302  return BO->hasNoUnsignedWrap() ||
2303  (BO->hasNoSignedWrap() && match(Step, m_APInt(StepC)) &&
2304  StartC->isNegative() == StepC->isNegative());
2305  case Instruction::Mul:
2306  return (BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap()) &&
2307  match(Step, m_APInt(StepC)) && !StepC->isZero();
2308  case Instruction::Shl:
2309  return BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap();
2310  case Instruction::AShr:
2311  case Instruction::LShr:
2312  return BO->isExact();
2313  default:
2314  return false;
2315  }
2316 }
2317 
2318 /// Return true if the given value is known to be non-zero when defined. For
2319 /// vectors, return true if every demanded element is known to be non-zero when
2320 /// defined. For pointers, if the context instruction and dominator tree are
2321 /// specified, perform context-sensitive analysis and return true if the
2322 /// pointer couldn't possibly be null at the specified instruction.
2323 /// Supports values with integer or pointer type and vectors of integers.
2324 bool isKnownNonZero(const Value *V, const APInt &DemandedElts, unsigned Depth,
2325  const Query &Q) {
2326  // FIXME: We currently have no way to represent the DemandedElts of a scalable
2327  // vector
2328  if (isa<ScalableVectorType>(V->getType()))
2329  return false;
2330 
2331  if (auto *C = dyn_cast<Constant>(V)) {
2332  if (C->isNullValue())
2333  return false;
2334  if (isa<ConstantInt>(C))
2335  // Must be non-zero due to null test above.
2336  return true;
2337 
2338  if (auto *CE = dyn_cast<ConstantExpr>(C)) {
2339  // See the comment for IntToPtr/PtrToInt instructions below.
2340  if (CE->getOpcode() == Instruction::IntToPtr ||
2341  CE->getOpcode() == Instruction::PtrToInt)
2342  if (Q.DL.getTypeSizeInBits(CE->getOperand(0)->getType())
2343  .getFixedSize() <=
2344  Q.DL.getTypeSizeInBits(CE->getType()).getFixedSize())
2345  return isKnownNonZero(CE->getOperand(0), Depth, Q);
2346  }
2347 
2348  // For constant vectors, check that all elements are undefined or known
2349  // non-zero to determine that the whole vector is known non-zero.
2350  if (auto *VecTy = dyn_cast<FixedVectorType>(C->getType())) {
2351  for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) {
2352  if (!DemandedElts[i])
2353  continue;
2354  Constant *Elt = C->getAggregateElement(i);
2355  if (!Elt || Elt->isNullValue())
2356  return false;
2357  if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt))
2358  return false;
2359  }
2360  return true;
2361  }
2362 
2363  // A global variable in address space 0 is non null unless extern weak
2364  // or an absolute symbol reference. Other address spaces may have null as a
2365  // valid address for a global, so we can't assume anything.
2366  if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
2367  if (!GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() &&
2368  GV->getType()->getAddressSpace() == 0)
2369  return true;
2370  } else
2371  return false;
2372  }
2373 
2374  if (auto *I = dyn_cast<Instruction>(V)) {
2375  if (MDNode *Ranges = Q.IIQ.getMetadata(I, LLVMContext::MD_range)) {
2376  // If the possible ranges don't contain zero, then the value is
2377  // definitely non-zero.
2378  if (auto *Ty = dyn_cast<IntegerType>(V->getType())) {
2379  const APInt ZeroValue(Ty->getBitWidth(), 0);
2380  if (rangeMetadataExcludesValue(Ranges, ZeroValue))
2381  return true;
2382  }
2383  }
2384  }
2385 
2386  if (isKnownNonZeroFromAssume(V, Q))
2387  return true;
2388 
2389  // Some of the tests below are recursive, so bail out if we hit the limit.
2391  return false;
2392 
2393  // Check for pointer simplifications.
2394 
2395  if (PointerType *PtrTy = dyn_cast<PointerType>(V->getType())) {
2396  // Alloca never returns null, malloc might.
2397  if (isa<AllocaInst>(V) && Q.DL.getAllocaAddrSpace() == 0)
2398  return true;
2399 
2400  // A byval, inalloca may not be null in a non-default addres space. A
2401  // nonnull argument is assumed never 0.
2402  if (const Argument *A = dyn_cast<Argument>(V)) {
2403  if (((A->hasPassPointeeByValueCopyAttr() &&
2404  !NullPointerIsDefined(A->getParent(), PtrTy->getAddressSpace())) ||
2405  A->hasNonNullAttr()))
2406  return true;
2407  }
2408 
2409  // A Load tagged with nonnull metadata is never null.
2410  if (const LoadInst *LI = dyn_cast<LoadInst>(V))
2411  if (Q.IIQ.getMetadata(LI, LLVMContext::MD_nonnull))
2412  return true;
2413 
2414  if (const auto *Call = dyn_cast<CallBase>(V)) {
2415  if (Call->isReturnNonNull())
2416  return true;
2417  if (const auto *RP = getArgumentAliasingToReturnedPointer(Call, true))
2418  return isKnownNonZero(RP, Depth, Q);
2419  }
2420  }
2421 
2422  if (isKnownNonNullFromDominatingCondition(V, Q.CxtI, Q.DT))
2423  return true;
2424 
2425  // Check for recursive pointer simplifications.
2426  if (V->getType()->isPointerTy()) {
2427  // Look through bitcast operations, GEPs, and int2ptr instructions as they
2428  // do not alter the value, or at least not the nullness property of the
2429  // value, e.g., int2ptr is allowed to zero/sign extend the value.
2430  //
2431  // Note that we have to take special care to avoid looking through
2432  // truncating casts, e.g., int2ptr/ptr2int with appropriate sizes, as well
2433  // as casts that can alter the value, e.g., AddrSpaceCasts.
2434  if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V))
2435  return isGEPKnownNonNull(GEP, Depth, Q);
2436 
2437  if (auto *BCO = dyn_cast<BitCastOperator>(V))
2438  return isKnownNonZero(BCO->getOperand(0), Depth, Q);
2439 
2440  if (auto *I2P = dyn_cast<IntToPtrInst>(V))
2441  if (Q.DL.getTypeSizeInBits(I2P->getSrcTy()).getFixedSize() <=
2442  Q.DL.getTypeSizeInBits(I2P->getDestTy()).getFixedSize())
2443  return isKnownNonZero(I2P->getOperand(0), Depth, Q);
2444  }
2445 
2446  // Similar to int2ptr above, we can look through ptr2int here if the cast
2447  // is a no-op or an extend and not a truncate.
2448  if (auto *P2I = dyn_cast<PtrToIntInst>(V))
2449  if (Q.DL.getTypeSizeInBits(P2I->getSrcTy()).getFixedSize() <=
2450  Q.DL.getTypeSizeInBits(P2I->getDestTy()).getFixedSize())
2451  return isKnownNonZero(P2I->getOperand(0), Depth, Q);
2452 
2453  unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), Q.DL);
2454 
2455  // X | Y != 0 if X != 0 or Y != 0.
2456  Value *X = nullptr, *Y = nullptr;
2457  if (match(V, m_Or(m_Value(X), m_Value(Y))))
2458  return isKnownNonZero(X, DemandedElts, Depth, Q) ||
2459  isKnownNonZero(Y, DemandedElts, Depth, Q);
2460 
2461  // ext X != 0 if X != 0.
2462  if (isa<SExtInst>(V) || isa<ZExtInst>(V))
2463  return isKnownNonZero(cast<Instruction>(V)->getOperand(0), Depth, Q);
2464 
2465  // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
2466  // if the lowest bit is shifted off the end.
2467  if (match(V, m_Shl(m_Value(X), m_Value(Y)))) {
2468  // shl nuw can't remove any non-zero bits.
2469  const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
2470  if (Q.IIQ.hasNoUnsignedWrap(BO))
2471  return isKnownNonZero(X, Depth, Q);
2472 
2473  KnownBits Known(BitWidth);
2474  computeKnownBits(X, DemandedElts, Known, Depth, Q);
2475  if (Known.One[0])
2476  return true;
2477  }
2478  // shr X, Y != 0 if X is negative. Note that the value of the shift is not
2479  // defined if the sign bit is shifted off the end.
2480  else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
2481  // shr exact can only shift out zero bits.
2482  const PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
2483  if (BO->isExact())
2484  return isKnownNonZero(X, Depth, Q);
2485 
2486  KnownBits Known = computeKnownBits(X, DemandedElts, Depth, Q);
2487  if (Known.isNegative())
2488  return true;
2489 
2490  // If the shifter operand is a constant, and all of the bits shifted
2491  // out are known to be zero, and X is known non-zero then at least one
2492  // non-zero bit must remain.
2493  if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) {
2494  auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
2495  // Is there a known one in the portion not shifted out?
2496  if (Known.countMaxLeadingZeros() < BitWidth - ShiftVal)
2497  return true;
2498  // Are all the bits to be shifted out known zero?
2499  if (Known.countMinTrailingZeros() >= ShiftVal)
2500  return isKnownNonZero(X, DemandedElts, Depth, Q);
2501  }
2502  }
2503  // div exact can only produce a zero if the dividend is zero.
2504  else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
2505  return isKnownNonZero(X, DemandedElts, Depth, Q);
2506  }
2507  // X + Y.
2508  else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
2509  KnownBits XKnown = computeKnownBits(X, DemandedElts, Depth, Q);
2510  KnownBits YKnown = computeKnownBits(Y, DemandedElts, Depth, Q);
2511 
2512  // If X and Y are both non-negative (as signed values) then their sum is not
2513  // zero unless both X and Y are zero.
2514  if (XKnown.isNonNegative() && YKnown.isNonNegative())
2515  if (isKnownNonZero(X, DemandedElts, Depth, Q) ||
2516  isKnownNonZero(Y, DemandedElts, Depth, Q))
2517  return true;
2518 
2519  // If X and Y are both negative (as signed values) then their sum is not
2520  // zero unless both X and Y equal INT_MIN.
2521  if (XKnown.isNegative() && YKnown.isNegative()) {
2523  // The sign bit of X is set. If some other bit is set then X is not equal
2524  // to INT_MIN.
2525  if (XKnown.One.intersects(Mask))
2526  return true;
2527  // The sign bit of Y is set. If some other bit is set then Y is not equal
2528  // to INT_MIN.
2529  if (YKnown.One.intersects(Mask))
2530  return true;
2531  }
2532 
2533  // The sum of a non-negative number and a power of two is not zero.
2534  if (XKnown.isNonNegative() &&
2535  isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q))
2536  return true;
2537  if (YKnown.isNonNegative() &&
2538  isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q))
2539  return true;
2540  }
2541  // X * Y.
2542  else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
2543  const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
2544  // If X and Y are non-zero then so is X * Y as long as the multiplication
2545  // does not overflow.
2546  if ((Q.IIQ.hasNoSignedWrap(BO) || Q.IIQ.hasNoUnsignedWrap(BO)) &&
2547  isKnownNonZero(X, DemandedElts, Depth, Q) &&
2548  isKnownNonZero(Y, DemandedElts, Depth, Q))
2549  return true;
2550  }
2551  // (C ? X : Y) != 0 if X != 0 and Y != 0.
2552  else if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
2553  if (isKnownNonZero(SI->getTrueValue(), DemandedElts, Depth, Q) &&
2554  isKnownNonZero(SI->getFalseValue(), DemandedElts, Depth, Q))
2555  return true;
2556  }
2557  // PHI
2558  else if (const PHINode *PN = dyn_cast<PHINode>(V)) {
2559  if (Q.IIQ.UseInstrInfo && isNonZeroRecurrence(PN))
2560  return true;
2561 
2562  // Check if all incoming values are non-zero using recursion.
2563  Query RecQ = Q;
2564  unsigned NewDepth = std::max(Depth, MaxAnalysisRecursionDepth - 1);
2565  return llvm::all_of(PN->operands(), [&](const Use &U) {
2566  if (U.get() == PN)
2567  return true;
2568  RecQ.CxtI = PN->getIncomingBlock(U)->getTerminator();
2569  return isKnownNonZero(U.get(), DemandedElts, NewDepth, RecQ);
2570  });
2571  }
2572  // ExtractElement
2573  else if (const auto *EEI = dyn_cast<ExtractElementInst>(V)) {
2574  const Value *Vec = EEI->getVectorOperand();
2575  const Value *Idx = EEI->getIndexOperand();
2576  auto *CIdx = dyn_cast<ConstantInt>(Idx);
2577  if (auto *VecTy = dyn_cast<FixedVectorType>(Vec->getType())) {
2578  unsigned NumElts = VecTy->getNumElements();
2579  APInt DemandedVecElts = APInt::getAllOnes(NumElts);
2580  if (CIdx && CIdx->getValue().ult(NumElts))
2581  DemandedVecElts = APInt::getOneBitSet(NumElts, CIdx->getZExtValue());
2582  return isKnownNonZero(Vec, DemandedVecElts, Depth, Q);
2583  }
2584  }
2585  // Freeze
2586  else if (const FreezeInst *FI = dyn_cast<FreezeInst>(V)) {
2587  auto *Op = FI->getOperand(0);
2588  if (isKnownNonZero(Op, Depth, Q) &&
2589  isGuaranteedNotToBePoison(Op, Q.AC, Q.CxtI, Q.DT, Depth))
2590  return true;
2591  }
2592 
2593  KnownBits Known(BitWidth);
2594  computeKnownBits(V, DemandedElts, Known, Depth, Q);
2595  return Known.One != 0;
2596 }
2597 
2598 bool isKnownNonZero(const Value* V, unsigned Depth, const Query& Q) {
2599  // FIXME: We currently have no way to represent the DemandedElts of a scalable
2600  // vector
2601  if (isa<ScalableVectorType>(V->getType()))
2602  return false;
2603 
2604  auto *FVTy = dyn_cast<FixedVectorType>(V->getType());
2605  APInt DemandedElts =
2606  FVTy ? APInt::getAllOnes(FVTy->getNumElements()) : APInt(1, 1);
2607  return isKnownNonZero(V, DemandedElts, Depth, Q);
2608 }
2609 
2610 /// If the pair of operators are the same invertible function, return the
2611 /// the operands of the function corresponding to each input. Otherwise,
2612 /// return None. An invertible function is one that is 1-to-1 and maps
2613 /// every input value to exactly one output value. This is equivalent to
2614 /// saying that Op1 and Op2 are equal exactly when the specified pair of
2615 /// operands are equal, (except that Op1 and Op2 may be poison more often.)
2618  const Operator *Op2) {
2619  if (Op1->getOpcode() != Op2->getOpcode())
2620  return None;
2621 
2622  auto getOperands = [&](unsigned OpNum) -> auto {
2623  return std::make_pair(Op1->getOperand(OpNum), Op2->getOperand(OpNum));
2624  };
2625 
2626  switch (Op1->getOpcode()) {
2627  default:
2628  break;
2629  case Instruction::Add:
2630  case Instruction::Sub:
2631  if (Op1->getOperand(0) == Op2->getOperand(0))
2632  return getOperands(1);
2633  if (Op1->getOperand(1) == Op2->getOperand(1))
2634  return getOperands(0);
2635  break;
2636  case Instruction::Mul: {
2637  // invertible if A * B == (A * B) mod 2^N where A, and B are integers
2638  // and N is the bitwdith. The nsw case is non-obvious, but proven by
2639  // alive2: https://alive2.llvm.org/ce/z/Z6D5qK
2640  auto *OBO1 = cast<OverflowingBinaryOperator>(Op1);
2641  auto *OBO2 = cast<OverflowingBinaryOperator>(Op2);
2642  if ((!OBO1->hasNoUnsignedWrap() || !OBO2->hasNoUnsignedWrap()) &&
2643  (!OBO1->hasNoSignedWrap() || !OBO2->hasNoSignedWrap()))
2644  break;
2645 
2646  // Assume operand order has been canonicalized
2647  if (Op1->getOperand(1) == Op2->getOperand(1) &&
2648  isa<ConstantInt>(Op1->getOperand(1)) &&
2649  !cast<ConstantInt>(Op1->getOperand(1))->isZero())
2650  return getOperands(0);
2651  break;
2652  }
2653  case Instruction::Shl: {
2654  // Same as multiplies, with the difference that we don't need to check
2655  // for a non-zero multiply. Shifts always multiply by non-zero.
2656  auto *OBO1 = cast<OverflowingBinaryOperator>(Op1);
2657  auto *OBO2 = cast<OverflowingBinaryOperator>(Op2);
2658  if ((!OBO1->hasNoUnsignedWrap() || !OBO2->hasNoUnsignedWrap()) &&
2659  (!OBO1->hasNoSignedWrap() || !OBO2->hasNoSignedWrap()))
2660  break;
2661 
2662  if (Op1->getOperand(1) == Op2->getOperand(1))
2663  return getOperands(0);
2664  break;
2665  }
2666  case Instruction::AShr:
2667  case Instruction::LShr: {
2668  auto *PEO1 = cast<PossiblyExactOperator>(Op1);
2669  auto *PEO2 = cast<PossiblyExactOperator>(Op2);
2670  if (!PEO1->isExact() || !PEO2->isExact())
2671  break;
2672 
2673  if (Op1->getOperand(1) == Op2->getOperand(1))
2674  return getOperands(0);
2675  break;
2676  }
2677  case Instruction::SExt:
2678  case Instruction::ZExt:
2679  if (Op1->getOperand(0)->getType() == Op2->getOperand(0)->getType())
2680  return getOperands(0);
2681  break;
2682  case Instruction::PHI: {
2683  const PHINode *PN1 = cast<PHINode>(Op1);
2684  const PHINode *PN2 = cast<PHINode>(Op2);
2685 
2686  // If PN1 and PN2 are both recurrences, can we prove the entire recurrences
2687  // are a single invertible function of the start values? Note that repeated
2688  // application of an invertible function is also invertible
2689  BinaryOperator *BO1 = nullptr;
2690  Value *Start1 = nullptr, *Step1 = nullptr;
2691  BinaryOperator *BO2 = nullptr;
2692  Value *Start2 = nullptr, *Step2 = nullptr;
2693  if (PN1->getParent() != PN2->getParent() ||
2694  !matchSimpleRecurrence(PN1, BO1, Start1, Step1) ||
2695  !matchSimpleRecurrence(PN2, BO2, Start2, Step2))
2696  break;
2697 
2698  auto Values = getInvertibleOperands(cast<Operator>(BO1),
2699  cast<Operator>(BO2));
2700  if (!Values)
2701  break;
2702 
2703  // We have to be careful of mutually defined recurrences here. Ex:
2704  // * X_i = X_(i-1) OP Y_(i-1), and Y_i = X_(i-1) OP V
2705  // * X_i = Y_i = X_(i-1) OP Y_(i-1)
2706  // The invertibility of these is complicated, and not worth reasoning
2707  // about (yet?).
2708  if (Values->first != PN1 || Values->second != PN2)
2709  break;
2710 
2711  return std::make_pair(Start1, Start2);
2712  }
2713  }
2714  return None;
2715 }
2716 
2717 /// Return true if V2 == V1 + X, where X is known non-zero.
2718 static bool isAddOfNonZero(const Value *V1, const Value *V2, unsigned Depth,
2719  const Query &Q) {
2720  const BinaryOperator *BO = dyn_cast<BinaryOperator>(V1);
2721  if (!BO || BO->getOpcode() != Instruction::Add)
2722  return false;
2723  Value *Op = nullptr;
2724  if (V2 == BO->getOperand(0))
2725  Op = BO->getOperand(1);
2726  else if (V2 == BO->getOperand(1))
2727  Op = BO->getOperand(0);
2728  else
2729  return false;
2730  return isKnownNonZero(Op, Depth + 1, Q);
2731 }
2732 
2733 /// Return true if V2 == V1 * C, where V1 is known non-zero, C is not 0/1 and
2734 /// the multiplication is nuw or nsw.
2735 static bool isNonEqualMul(const Value *V1, const Value *V2, unsigned Depth,
2736  const Query &Q) {
2737  if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(V2)) {
2738  const APInt *C;
2739  return match(OBO, m_Mul(m_Specific(V1), m_APInt(C))) &&
2740  (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) &&
2741  !C->isZero() && !C->isOne() && isKnownNonZero(V1, Depth + 1, Q);
2742  }
2743  return false;
2744 }
2745 
2746 /// Return true if V2 == V1 << C, where V1 is known non-zero, C is not 0 and
2747 /// the shift is nuw or nsw.
2748 static bool isNonEqualShl(const Value *V1, const Value *V2, unsigned Depth,
2749  const Query &Q) {
2750  if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(V2)) {
2751  const APInt *C;
2752  return match(OBO, m_Shl(m_Specific(V1), m_APInt(C))) &&
2753  (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) &&
2754  !C->isZero() && isKnownNonZero(V1, Depth + 1, Q);
2755  }
2756  return false;
2757 }
2758 
2759 static bool isNonEqualPHIs(const PHINode *PN1, const PHINode *PN2,
2760  unsigned Depth, const Query &Q) {
2761  // Check two PHIs are in same block.
2762  if (PN1->getParent() != PN2->getParent())
2763  return false;
2764 
2766  bool UsedFullRecursion = false;
2767  for (const BasicBlock *IncomBB : PN1->blocks()) {
2768  if (!VisitedBBs.insert(IncomBB).second)
2769  continue; // Don't reprocess blocks that we have dealt with already.
2770  const Value *IV1 = PN1->getIncomingValueForBlock(IncomBB);
2771  const Value *IV2 = PN2->getIncomingValueForBlock(IncomBB);
2772  const APInt *C1, *C2;
2773  if (match(IV1, m_APInt(C1)) && match(IV2, m_APInt(C2)) && *C1 != *C2)
2774  continue;
2775 
2776  // Only one pair of phi operands is allowed for full recursion.
2777  if (UsedFullRecursion)
2778  return false;
2779 
2780  Query RecQ = Q;
2781  RecQ.CxtI = IncomBB->getTerminator();
2782  if (!isKnownNonEqual(IV1, IV2, Depth + 1, RecQ))
2783  return false;
2784  UsedFullRecursion = true;
2785  }
2786  return true;
2787 }
2788 
2789 /// Return true if it is known that V1 != V2.
2790 static bool isKnownNonEqual(const Value *V1, const Value *V2, unsigned Depth,
2791  const Query &Q) {
2792  if (V1 == V2)
2793  return false;
2794  if (V1->getType() != V2->getType())
2795  // We can't look through casts yet.
2796  return false;
2797 
2799  return false;
2800 
2801  // See if we can recurse through (exactly one of) our operands. This
2802  // requires our operation be 1-to-1 and map every input value to exactly
2803  // one output value. Such an operation is invertible.
2804  auto *O1 = dyn_cast<Operator>(V1);
2805  auto *O2 = dyn_cast<Operator>(V2);
2806  if (O1 && O2 && O1->getOpcode() == O2->getOpcode()) {
2807  if (auto Values = getInvertibleOperands(O1, O2))
2808  return isKnownNonEqual(Values->first, Values->second, Depth + 1, Q);
2809 
2810  if (const PHINode *PN1 = dyn_cast<PHINode>(V1)) {
2811  const PHINode *PN2 = cast<PHINode>(V2);
2812  // FIXME: This is missing a generalization to handle the case where one is
2813  // a PHI and another one isn't.
2814  if (isNonEqualPHIs(PN1, PN2, Depth, Q))
2815  return true;
2816  };
2817  }
2818 
2819  if (isAddOfNonZero(V1, V2, Depth, Q) || isAddOfNonZero(V2, V1, Depth, Q))
2820  return true;
2821 
2822  if (isNonEqualMul(V1, V2, Depth, Q) || isNonEqualMul(V2, V1, Depth, Q))
2823  return true;
2824 
2825  if (isNonEqualShl(V1, V2, Depth, Q) || isNonEqualShl(V2, V1, Depth, Q))
2826  return true;
2827 
2828  if (V1->getType()->isIntOrIntVectorTy()) {
2829  // Are any known bits in V1 contradictory to known bits in V2? If V1
2830  // has a known zero where V2 has a known one, they must not be equal.
2831  KnownBits Known1 = computeKnownBits(V1, Depth, Q);
2832  KnownBits Known2 = computeKnownBits(V2, Depth, Q);
2833 
2834  if (Known1.Zero.intersects(Known2.One) ||
2835  Known2.Zero.intersects(Known1.One))
2836  return true;
2837  }
2838  return false;
2839 }
2840 
2841 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
2842 /// simplify operations downstream. Mask is known to be zero for bits that V
2843 /// cannot have.
2844 ///
2845 /// This function is defined on values with integer type, values with pointer
2846 /// type, and vectors of integers. In the case
2847 /// where V is a vector, the mask, known zero, and known one values are the
2848 /// same width as the vector element, and the bit is set only if it is true
2849 /// for all of the elements in the vector.
2850 bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
2851  const Query &Q) {
2852  KnownBits Known(Mask.getBitWidth());
2853  computeKnownBits(V, Known, Depth, Q);
2854  return Mask.isSubsetOf(Known.Zero);
2855 }
2856 
2857 // Match a signed min+max clamp pattern like smax(smin(In, CHigh), CLow).
2858 // Returns the input and lower/upper bounds.
2859 static bool isSignedMinMaxClamp(const Value *Select, const Value *&In,
2860  const APInt *&CLow, const APInt *&CHigh) {
2861  assert(isa<Operator>(Select) &&
2862  cast<Operator>(Select)->getOpcode() == Instruction::Select &&
2863  "Input should be a Select!");
2864 
2865  const Value *LHS = nullptr, *RHS = nullptr;
2867  if (SPF != SPF_SMAX && SPF != SPF_SMIN)
2868  return false;
2869 
2870  if (!match(RHS, m_APInt(CLow)))
2871  return false;
2872 
2873  const Value *LHS2 = nullptr, *RHS2 = nullptr;
2874  SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor;
2875  if (getInverseMinMaxFlavor(SPF) != SPF2)
2876  return false;
2877 
2878  if (!match(RHS2, m_APInt(CHigh)))
2879  return false;
2880 
2881  if (SPF == SPF_SMIN)
2882  std::swap(CLow, CHigh);
2883 
2884  In = LHS2;
2885  return CLow->sle(*CHigh);
2886 }
2887 
2888 /// For vector constants, loop over the elements and find the constant with the
2889 /// minimum number of sign bits. Return 0 if the value is not a vector constant
2890 /// or if any element was not analyzed; otherwise, return the count for the
2891 /// element with the minimum number of sign bits.
2892 static unsigned computeNumSignBitsVectorConstant(const Value *V,
2893  const APInt &DemandedElts,
2894  unsigned TyBits) {
2895  const auto *CV = dyn_cast<Constant>(V);
2896  if (!CV || !isa<FixedVectorType>(CV->getType()))
2897  return 0;
2898 
2899  unsigned MinSignBits = TyBits;
2900  unsigned NumElts = cast<FixedVectorType>(CV->getType())->getNumElements();
2901  for (unsigned i = 0; i != NumElts; ++i) {
2902  if (!DemandedElts[i])
2903  continue;
2904  // If we find a non-ConstantInt, bail out.
2905  auto *Elt = dyn_cast_or_null<ConstantInt>(CV->getAggregateElement(i));
2906  if (!Elt)
2907  return 0;
2908 
2909  MinSignBits = std::min(MinSignBits, Elt->getValue().getNumSignBits());
2910  }
2911 
2912  return MinSignBits;
2913 }
2914 
2915 static unsigned ComputeNumSignBitsImpl(const Value *V,
2916  const APInt &DemandedElts,
2917  unsigned Depth, const Query &Q);
2918 
2919 static unsigned ComputeNumSignBits(const Value *V, const APInt &DemandedElts,
2920  unsigned Depth, const Query &Q) {
2921  unsigned Result = ComputeNumSignBitsImpl(V, DemandedElts, Depth, Q);
2922  assert(Result > 0 && "At least one sign bit needs to be present!");
2923  return Result;
2924 }
2925 
2926 /// Return the number of times the sign bit of the register is replicated into
2927 /// the other bits. We know that at least 1 bit is always equal to the sign bit
2928 /// (itself), but other cases can give us information. For example, immediately
2929 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
2930 /// other, so we return 3. For vectors, return the number of sign bits for the
2931 /// vector element with the minimum number of known sign bits of the demanded
2932 /// elements in the vector specified by DemandedElts.
2933 static unsigned ComputeNumSignBitsImpl(const Value *V,
2934  const APInt &DemandedElts,
2935  unsigned Depth, const Query &Q) {
2936  Type *Ty = V->getType();
2937 
2938  // FIXME: We currently have no way to represent the DemandedElts of a scalable
2939  // vector
2940  if (isa<ScalableVectorType>(Ty))
2941  return 1;
2942 
2943 #ifndef NDEBUG
2944  assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth");
2945 
2946  if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) {
2947  assert(
2948  FVTy->getNumElements() == DemandedElts.getBitWidth() &&
2949  "DemandedElt width should equal the fixed vector number of elements");
2950  } else {
2951  assert(DemandedElts == APInt(1, 1) &&
2952  "DemandedElt width should be 1 for scalars");
2953  }
2954 #endif
2955 
2956  // We return the minimum number of sign bits that are guaranteed to be present
2957  // in V, so for undef we have to conservatively return 1. We don't have the
2958  // same behavior for poison though -- that's a FIXME today.
2959 
2960  Type *ScalarTy = Ty->getScalarType();
2961  unsigned TyBits = ScalarTy->isPointerTy() ?
2962  Q.DL.getPointerTypeSizeInBits(ScalarTy) :
2963  Q.DL.getTypeSizeInBits(ScalarTy);
2964 
2965  unsigned Tmp, Tmp2;
2966  unsigned FirstAnswer = 1;
2967 
2968  // Note that ConstantInt is handled by the general computeKnownBits case
2969  // below.
2970 
2972  return 1;
2973 
2974  if (auto *U = dyn_cast<Operator>(V)) {
2975  switch (Operator::getOpcode(V)) {
2976  default: break;
2977  case Instruction::SExt:
2978  Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
2979  return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q) + Tmp;
2980 
2981  case Instruction::SDiv: {
2982  const APInt *Denominator;
2983  // sdiv X, C -> adds log(C) sign bits.
2984  if (match(U->getOperand(1), m_APInt(Denominator))) {
2985 
2986  // Ignore non-positive denominator.
2987  if (!Denominator->isStrictlyPositive())
2988  break;
2989 
2990  // Calculate the incoming numerator bits.
2991  unsigned NumBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2992 
2993  // Add floor(log(C)) bits to the numerator bits.
2994  return std::min(TyBits, NumBits + Denominator->logBase2());
2995  }
2996  break;
2997  }
2998 
2999  case Instruction::SRem: {
3000  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3001 
3002  const APInt *Denominator;
3003  // srem X, C -> we know that the result is within [-C+1,C) when C is a
3004  // positive constant. This let us put a lower bound on the number of sign
3005  // bits.
3006  if (match(U->getOperand(1), m_APInt(Denominator))) {
3007 
3008  // Ignore non-positive denominator.
3009  if (Denominator->isStrictlyPositive()) {
3010  // Calculate the leading sign bit constraints by examining the
3011  // denominator. Given that the denominator is positive, there are two
3012  // cases:
3013  //
3014  // 1. The numerator is positive. The result range is [0,C) and
3015  // [0,C) u< (1 << ceilLogBase2(C)).
3016  //
3017  // 2. The numerator is negative. Then the result range is (-C,0] and
3018  // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
3019  //
3020  // Thus a lower bound on the number of sign bits is `TyBits -
3021  // ceilLogBase2(C)`.
3022 
3023  unsigned ResBits = TyBits - Denominator->ceilLogBase2();
3024  Tmp = std::max(Tmp, ResBits);
3025  }
3026  }
3027  return Tmp;
3028  }
3029 
3030  case Instruction::AShr: {
3031  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3032  // ashr X, C -> adds C sign bits. Vectors too.
3033  const APInt *ShAmt;
3034  if (match(U->getOperand(1), m_APInt(ShAmt))) {
3035  if (ShAmt->uge(TyBits))
3036  break; // Bad shift.
3037  unsigned ShAmtLimited = ShAmt->getZExtValue();
3038  Tmp += ShAmtLimited;
3039  if (Tmp > TyBits) Tmp = TyBits;
3040  }
3041  return Tmp;
3042  }
3043  case Instruction::Shl: {
3044  const APInt *ShAmt;
3045  if (match(U->getOperand(1), m_APInt(ShAmt))) {
3046  // shl destroys sign bits.
3047  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3048  if (ShAmt->uge(TyBits) || // Bad shift.
3049  ShAmt->uge(Tmp)) break; // Shifted all sign bits out.
3050  Tmp2 = ShAmt->getZExtValue();
3051  return Tmp - Tmp2;
3052  }
3053  break;
3054  }
3055  case Instruction::And:
3056  case Instruction::Or:
3057  case Instruction::Xor: // NOT is handled here.
3058  // Logical binary ops preserve the number of sign bits at the worst.
3059  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3060  if (Tmp != 1) {
3061  Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
3062  FirstAnswer = std::min(Tmp, Tmp2);
3063  // We computed what we know about the sign bits as our first
3064  // answer. Now proceed to the generic code that uses
3065  // computeKnownBits, and pick whichever answer is better.
3066  }
3067  break;
3068 
3069  case Instruction::Select: {
3070  // If we have a clamp pattern, we know that the number of sign bits will
3071  // be the minimum of the clamp min/max range.
3072  const Value *X;
3073  const APInt *CLow, *CHigh;
3074  if (isSignedMinMaxClamp(U, X, CLow, CHigh))
3075  return std::min(CLow->getNumSignBits(), CHigh->getNumSignBits());
3076 
3077  Tmp = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
3078  if (Tmp == 1) break;
3079  Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth + 1, Q);
3080  return std::min(Tmp, Tmp2);
3081  }
3082 
3083  case Instruction::Add:
3084  // Add can have at most one carry bit. Thus we know that the output
3085  // is, at worst, one more bit than the inputs.
3086  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3087  if (Tmp == 1) break;
3088 
3089  // Special case decrementing a value (ADD X, -1):
3090  if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
3091  if (CRHS->isAllOnesValue()) {
3092  KnownBits Known(TyBits);
3093  computeKnownBits(U->getOperand(0), Known, Depth + 1, Q);
3094 
3095  // If the input is known to be 0 or 1, the output is 0/-1, which is
3096  // all sign bits set.
3097  if ((Known.Zero | 1).isAllOnes())
3098  return TyBits;
3099 
3100  // If we are subtracting one from a positive number, there is no carry
3101  // out of the result.
3102  if (Known.isNonNegative())
3103  return Tmp;
3104  }
3105 
3106  Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
3107  if (Tmp2 == 1) break;
3108  return std::min(Tmp, Tmp2) - 1;
3109 
3110  case Instruction::Sub:
3111  Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
3112  if (Tmp2 == 1) break;
3113 
3114  // Handle NEG.
3115  if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
3116  if (CLHS->isNullValue()) {
3117  KnownBits Known(TyBits);
3118  computeKnownBits(U->getOperand(1), Known, Depth + 1, Q);
3119  // If the input is known to be 0 or 1, the output is 0/-1, which is
3120  // all sign bits set.
3121  if ((Known.Zero | 1).isAllOnes())
3122  return TyBits;
3123 
3124  // If the input is known to be positive (the sign bit is known clear),
3125  // the output of the NEG has the same number of sign bits as the
3126  // input.
3127  if (Known.isNonNegative())
3128  return Tmp2;
3129 
3130  // Otherwise, we treat this like a SUB.
3131  }
3132 
3133  // Sub can have at most one carry bit. Thus we know that the output
3134  // is, at worst, one more bit than the inputs.
3135  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3136  if (Tmp == 1) break;
3137  return std::min(Tmp, Tmp2) - 1;
3138 
3139  case Instruction::Mul: {
3140  // The output of the Mul can be at most twice the valid bits in the
3141  // inputs.
3142  unsigned SignBitsOp0 = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3143  if (SignBitsOp0 == 1) break;
3144  unsigned SignBitsOp1 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
3145  if (SignBitsOp1 == 1) break;
3146  unsigned OutValidBits =
3147  (TyBits - SignBitsOp0 + 1) + (TyBits - SignBitsOp1 + 1);
3148  return OutValidBits > TyBits ? 1 : TyBits - OutValidBits + 1;
3149  }
3150 
3151  case Instruction::PHI: {
3152  const PHINode *PN = cast<PHINode>(U);
3153  unsigned NumIncomingValues = PN->getNumIncomingValues();
3154  // Don't analyze large in-degree PHIs.
3155  if (NumIncomingValues > 4) break;
3156  // Unreachable blocks may have zero-operand PHI nodes.
3157  if (NumIncomingValues == 0) break;
3158 
3159  // Take the minimum of all incoming values. This can't infinitely loop
3160  // because of our depth threshold.
3161  Query RecQ = Q;
3162  Tmp = TyBits;
3163  for (unsigned i = 0, e = NumIncomingValues; i != e; ++i) {
3164  if (Tmp == 1) return Tmp;
3165  RecQ.CxtI = PN->getIncomingBlock(i)->getTerminator();
3166  Tmp = std::min(
3167  Tmp, ComputeNumSignBits(PN->getIncomingValue(i), Depth + 1, RecQ));
3168  }
3169  return Tmp;
3170  }
3171 
3172  case Instruction::Trunc:
3173  // FIXME: it's tricky to do anything useful for this, but it is an
3174  // important case for targets like X86.
3175  break;
3176 
3177  case Instruction::ExtractElement:
3178  // Look through extract element. At the moment we keep this simple and
3179  // skip tracking the specific element. But at least we might find
3180  // information valid for all elements of the vector (for example if vector
3181  // is sign extended, shifted, etc).
3182  return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3183 
3184  case Instruction::ShuffleVector: {
3185  // Collect the minimum number of sign bits that are shared by every vector
3186  // element referenced by the shuffle.
3187  auto *Shuf = dyn_cast<ShuffleVectorInst>(U);
3188  if (!Shuf) {
3189  // FIXME: Add support for shufflevector constant expressions.
3190  return 1;
3191  }
3192  APInt DemandedLHS, DemandedRHS;
3193  // For undef elements, we don't know anything about the common state of
3194  // the shuffle result.
3195  if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS))
3196  return 1;
3198  if (!!DemandedLHS) {
3199  const Value *LHS = Shuf->getOperand(0);
3200  Tmp = ComputeNumSignBits(LHS, DemandedLHS, Depth + 1, Q);
3201  }
3202  // If we don't know anything, early out and try computeKnownBits
3203  // fall-back.
3204  if (Tmp == 1)
3205  break;
3206  if (!!DemandedRHS) {
3207  const Value *RHS = Shuf->getOperand(1);
3208  Tmp2 = ComputeNumSignBits(RHS, DemandedRHS, Depth + 1, Q);
3209  Tmp = std::min(Tmp, Tmp2);
3210  }
3211  // If we don't know anything, early out and try computeKnownBits
3212  // fall-back.
3213  if (Tmp == 1)
3214  break;
3215  assert(Tmp <= TyBits && "Failed to determine minimum sign bits");
3216  return Tmp;
3217  }
3218  case Instruction::Call: {
3219  if (const auto *II = dyn_cast<IntrinsicInst>(U)) {
3220  switch (II->getIntrinsicID()) {
3221  default: break;
3222  case Intrinsic::abs:
3223  Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3224  if (Tmp == 1) break;
3225 
3226  // Absolute value reduces number of sign bits by at most 1.
3227  return Tmp - 1;
3228  }
3229  }
3230  }
3231  }
3232  }
3233 
3234  // Finally, if we can prove that the top bits of the result are 0's or 1's,
3235  // use this information.
3236 
3237  // If we can examine all elements of a vector constant successfully, we're
3238  // done (we can't do any better than that). If not, keep trying.
3239  if (unsigned VecSignBits =
3240  computeNumSignBitsVectorConstant(V, DemandedElts, TyBits))
3241  return VecSignBits;
3242 
3243  KnownBits Known(TyBits);
3244  computeKnownBits(V, DemandedElts, Known, Depth, Q);
3245 
3246  // If we know that the sign bit is either zero or one, determine the number of
3247  // identical bits in the top of the input value.
3248  return std::max(FirstAnswer, Known.countMinSignBits());
3249 }
3250 
3251 /// This function computes the integer multiple of Base that equals V.
3252 /// If successful, it returns true and returns the multiple in
3253 /// Multiple. If unsuccessful, it returns false. It looks
3254 /// through SExt instructions only if LookThroughSExt is true.
3255 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
3256  bool LookThroughSExt, unsigned Depth) {
3257  assert(V && "No Value?");
3258  assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth");
3259  assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
3260 
3261  Type *T = V->getType();
3262 
3263  ConstantInt *CI = dyn_cast<ConstantInt>(V);
3264 
3265  if (Base == 0)
3266  return false;
3267 
3268  if (Base == 1) {
3269  Multiple = V;
3270  return true;
3271  }
3272 
3273  ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
3274  Constant *BaseVal = ConstantInt::get(T, Base);
3275  if (CO && CO == BaseVal) {
3276  // Multiple is 1.
3277  Multiple = ConstantInt::get(T, 1);
3278  return true;
3279  }
3280 
3281  if (CI && CI->getZExtValue() % Base == 0) {
3282  Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
3283  return true;
3284  }
3285 
3286  if (Depth == MaxAnalysisRecursionDepth) return false;
3287 
3288  Operator *I = dyn_cast<Operator>(V);
3289  if (!I) return false;
3290 
3291  switch (I->getOpcode()) {
3292  default: break;
3293  case Instruction::SExt:
3294  if (!LookThroughSExt) return false;
3295  // otherwise fall through to ZExt
3297  case Instruction::ZExt:
3298  return ComputeMultiple(I->getOperand(0), Base, Multiple,
3299  LookThroughSExt, Depth+1);
3300  case Instruction::Shl:
3301  case Instruction::Mul: {
3302  Value *Op0 = I->getOperand(0);
3303  Value *Op1 = I->getOperand(1);
3304 
3305  if (I->getOpcode() == Instruction::Shl) {
3306  ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
3307  if (!Op1CI) return false;
3308  // Turn Op0 << Op1 into Op0 * 2^Op1
3309  APInt Op1Int = Op1CI->getValue();
3310  uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
3311  APInt API(Op1Int.getBitWidth(), 0);
3312  API.setBit(BitToSet);
3313  Op1 = ConstantInt::get(V->getContext(), API);
3314  }
3315 
3316  Value *Mul0 = nullptr;
3317  if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
3318  if (Constant *Op1C = dyn_cast<Constant>(Op1))
3319  if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
3320  if (Op1C->getType()->getPrimitiveSizeInBits().getFixedSize() <
3321  MulC->getType()->getPrimitiveSizeInBits().getFixedSize())
3322  Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
3323  if (Op1C->getType()->getPrimitiveSizeInBits().getFixedSize() >
3324  MulC->getType()->getPrimitiveSizeInBits().getFixedSize())
3325  MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
3326 
3327  // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
3328  Multiple = ConstantExpr::getMul(MulC, Op1C);
3329  return true;
3330  }
3331 
3332  if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
3333  if (Mul0CI->getValue() == 1) {
3334  // V == Base * Op1, so return Op1
3335  Multiple = Op1;
3336  return true;
3337  }
3338  }
3339 
3340  Value *Mul1 = nullptr;
3341  if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
3342  if (Constant *Op0C = dyn_cast<Constant>(Op0))
3343  if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
3344  if (Op0C->getType()->getPrimitiveSizeInBits().getFixedSize() <
3345  MulC->getType()->getPrimitiveSizeInBits().getFixedSize())
3346  Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
3347  if (Op0C->getType()->getPrimitiveSizeInBits().getFixedSize() >
3348  MulC->getType()->getPrimitiveSizeInBits().getFixedSize())
3349  MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
3350 
3351  // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
3352  Multiple = ConstantExpr::getMul(MulC, Op0C);
3353  return true;
3354  }
3355 
3356  if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
3357  if (Mul1CI->getValue() == 1) {
3358  // V == Base * Op0, so return Op0
3359  Multiple = Op0;
3360  return true;
3361  }
3362  }
3363  }
3364  }
3365 
3366  // We could not determine if V is a multiple of Base.
3367  return false;
3368 }
3369 
3371  const TargetLibraryInfo *TLI) {
3372  const Function *F = CB.getCalledFunction();
3373  if (!F)
3374  return Intrinsic::not_intrinsic;
3375 
3376  if (F->isIntrinsic())
3377  return F->getIntrinsicID();
3378 
3379  // We are going to infer semantics of a library function based on mapping it
3380  // to an LLVM intrinsic. Check that the library function is available from
3381  // this callbase and in this environment.
3382  LibFunc Func;
3383  if (F->hasLocalLinkage() || !TLI || !TLI->getLibFunc(CB, Func) ||
3384  !CB.onlyReadsMemory())
3385  return Intrinsic::not_intrinsic;
3386 
3387  switch (Func) {
3388  default:
3389  break;
3390  case LibFunc_sin:
3391  case LibFunc_sinf:
3392  case LibFunc_sinl:
3393  return Intrinsic::sin;
3394  case LibFunc_cos:
3395  case LibFunc_cosf:
3396  case LibFunc_cosl:
3397  return Intrinsic::cos;
3398  case LibFunc_exp:
3399  case LibFunc_expf:
3400  case LibFunc_expl:
3401  return Intrinsic::exp;
3402  case LibFunc_exp2:
3403  case LibFunc_exp2f:
3404  case LibFunc_exp2l:
3405  return Intrinsic::exp2;
3406  case LibFunc_log:
3407  case LibFunc_logf:
3408  case LibFunc_logl:
3409  return Intrinsic::log;
3410  case LibFunc_log10:
3411  case LibFunc_log10f:
3412  case LibFunc_log10l:
3413  return Intrinsic::log10;
3414  case LibFunc_log2:
3415  case LibFunc_log2f:
3416  case LibFunc_log2l:
3417  return Intrinsic::log2;
3418  case LibFunc_fabs:
3419  case LibFunc_fabsf:
3420  case LibFunc_fabsl:
3421  return Intrinsic::fabs;
3422  case LibFunc_fmin:
3423  case LibFunc_fminf:
3424  case LibFunc_fminl:
3425  return Intrinsic::minnum;
3426  case LibFunc_fmax:
3427  case LibFunc_fmaxf:
3428  case LibFunc_fmaxl:
3429  return Intrinsic::maxnum;
3430  case LibFunc_copysign:
3431  case LibFunc_copysignf:
3432  case LibFunc_copysignl:
3433  return Intrinsic::copysign;
3434  case LibFunc_floor:
3435  case LibFunc_floorf:
3436  case LibFunc_floorl:
3437  return Intrinsic::floor;
3438  case LibFunc_ceil:
3439  case LibFunc_ceilf:
3440  case LibFunc_ceill:
3441  return Intrinsic::ceil;
3442  case LibFunc_trunc:
3443  case LibFunc_truncf:
3444  case LibFunc_truncl:
3445  return Intrinsic::trunc;
3446  case LibFunc_rint:
3447  case LibFunc_rintf:
3448  case LibFunc_rintl:
3449  return Intrinsic::rint;
3450  case LibFunc_nearbyint:
3451  case LibFunc_nearbyintf:
3452  case LibFunc_nearbyintl:
3453  return Intrinsic::nearbyint;
3454  case LibFunc_round:
3455  case LibFunc_roundf:
3456  case LibFunc_roundl:
3457  return Intrinsic::round;
3458  case LibFunc_roundeven:
3459  case LibFunc_roundevenf:
3460  case LibFunc_roundevenl:
3461  return Intrinsic::roundeven;
3462  case LibFunc_pow:
3463  case LibFunc_powf:
3464  case LibFunc_powl:
3465  return Intrinsic::pow;
3466  case LibFunc_sqrt:
3467  case LibFunc_sqrtf:
3468  case LibFunc_sqrtl:
3469  return Intrinsic::sqrt;
3470  }
3471 
3472  return Intrinsic::not_intrinsic;
3473 }
3474 
3475 /// Return true if we can prove that the specified FP value is never equal to
3476 /// -0.0.
3477 /// NOTE: Do not check 'nsz' here because that fast-math-flag does not guarantee
3478 /// that a value is not -0.0. It only guarantees that -0.0 may be treated
3479 /// the same as +0.0 in floating-point ops.
3480 ///
3481 /// NOTE: this function will need to be revisited when we support non-default
3482 /// rounding modes!
3484  unsigned Depth) {
3485  if (auto *CFP = dyn_cast<ConstantFP>(V))
3486  return !CFP->getValueAPF().isNegZero();
3487 
3489  return false;
3490 
3491  auto *Op = dyn_cast<Operator>(V);
3492  if (!Op)
3493  return false;
3494 
3495  // (fadd x, 0.0) is guaranteed to return +0.0, not -0.0.
3496  if (match(Op, m_FAdd(m_Value(), m_PosZeroFP())))
3497  return true;
3498 
3499  // sitofp and uitofp turn into +0.0 for zero.
3500  if (isa<SIToFPInst>(Op) || isa<UIToFPInst>(Op))
3501  return true;
3502 
3503  if (auto *Call = dyn_cast<CallInst>(Op)) {
3504  Intrinsic::ID IID = getIntrinsicForCallSite(*Call, TLI);
3505  switch (IID) {
3506  default:
3507  break;
3508  // sqrt(-0.0) = -0.0, no other negative results are possible.
3509  case Intrinsic::sqrt:
3510  case Intrinsic::canonicalize:
3511  return CannotBeNegativeZero(Call->getArgOperand(0), TLI, Depth + 1);
3512  // fabs(x) != -0.0
3513  case Intrinsic::fabs:
3514  return true;
3515  }
3516  }
3517 
3518  return false;
3519 }
3520 
3521 /// If \p SignBitOnly is true, test for a known 0 sign bit rather than a
3522 /// standard ordered compare. e.g. make -0.0 olt 0.0 be true because of the sign
3523 /// bit despite comparing equal.
3525  const TargetLibraryInfo *TLI,
3526  bool SignBitOnly,
3527  unsigned Depth) {
3528  // TODO: This function does not do the right thing when SignBitOnly is true
3529  // and we're lowering to a hypothetical IEEE 754-compliant-but-evil platform
3530  // which flips the sign bits of NaNs. See
3531  // https://llvm.org/bugs/show_bug.cgi?id=31702.
3532 
3533  if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
3534  return !CFP->getValueAPF().isNegative() ||
3535  (!SignBitOnly && CFP->getValueAPF().isZero());
3536  }
3537 
3538  // Handle vector of constants.
3539  if (auto *CV = dyn_cast<Constant>(V)) {
3540  if (auto *CVFVTy = dyn_cast<FixedVectorType>(CV->getType())) {
3541  unsigned NumElts = CVFVTy->getNumElements();
3542  for (unsigned i = 0; i != NumElts; ++i) {
3543  auto *CFP = dyn_cast_or_null<ConstantFP>(CV->getAggregateElement(i));
3544  if (!CFP)
3545  return false;
3546  if (CFP->getValueAPF().isNegative() &&
3547  (SignBitOnly || !CFP->getValueAPF().isZero()))
3548  return false;
3549  }
3550 
3551  // All non-negative ConstantFPs.
3552  return true;
3553  }
3554  }
3555 
3557  return false;
3558 
3559  const Operator *I = dyn_cast<Operator>(V);
3560  if (!I)
3561  return false;
3562 
3563  switch (I->getOpcode()) {
3564  default:
3565  break;
3566  // Unsigned integers are always nonnegative.
3567  case Instruction::UIToFP:
3568  return true;
3569  case Instruction::FMul:
3570  case Instruction::FDiv:
3571  // X * X is always non-negative or a NaN.
3572  // X / X is always exactly 1.0 or a NaN.
3573  if (I->getOperand(0) == I->getOperand(1) &&
3574  (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()))
3575  return true;
3576 
3578  case Instruction::FAdd:
3579  case Instruction::FRem:
3580  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3581  Depth + 1) &&
3582  cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
3583  Depth + 1);
3584  case Instruction::Select:
3585  return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
3586  Depth + 1) &&
3587  cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
3588  Depth + 1);
3589  case Instruction::FPExt:
3590  case Instruction::FPTrunc:
3591  // Widening/narrowing never change sign.
3592  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3593  Depth + 1);
3594  case Instruction::ExtractElement:
3595  // Look through extract element. At the moment we keep this simple and skip
3596  // tracking the specific element. But at least we might find information
3597  // valid for all elements of the vector.
3598  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3599  Depth + 1);
3600  case Instruction::Call:
3601  const auto *CI = cast<CallInst>(I);
3602  Intrinsic::ID IID = getIntrinsicForCallSite(*CI, TLI);
3603  switch (IID) {
3604  default:
3605  break;
3606  case Intrinsic::maxnum: {
3607  Value *V0 = I->getOperand(0), *V1 = I->getOperand(1);
3608  auto isPositiveNum = [&](Value *V) {
3609  if (SignBitOnly) {
3610  // With SignBitOnly, this is tricky because the result of
3611  // maxnum(+0.0, -0.0) is unspecified. Just check if the operand is
3612  // a constant strictly greater than 0.0.
3613  const APFloat *C;
3614  return match(V, m_APFloat(C)) &&
3615  *C > APFloat::getZero(C->getSemantics());
3616  }
3617 
3618  // -0.0 compares equal to 0.0, so if this operand is at least -0.0,
3619  // maxnum can't be ordered-less-than-zero.
3620  return isKnownNeverNaN(V, TLI) &&
3621  cannotBeOrderedLessThanZeroImpl(V, TLI, false, Depth + 1);
3622  };
3623 
3624  // TODO: This could be improved. We could also check that neither operand
3625  // has its sign bit set (and at least 1 is not-NAN?).
3626  return isPositiveNum(V0) || isPositiveNum(V1);
3627  }
3628 
3629  case Intrinsic::maximum:
3630  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3631  Depth + 1) ||
3632  cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
3633  Depth + 1);
3634  case Intrinsic::minnum:
3635  case Intrinsic::minimum:
3636  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3637  Depth + 1) &&
3638  cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
3639  Depth + 1);
3640  case Intrinsic::exp:
3641  case Intrinsic::exp2:
3642  case Intrinsic::fabs:
3643  return true;
3644 
3645  case Intrinsic::sqrt:
3646  // sqrt(x) is always >= -0 or NaN. Moreover, sqrt(x) == -0 iff x == -0.
3647  if (!SignBitOnly)
3648  return true;
3649  return CI->hasNoNaNs() && (CI->hasNoSignedZeros() ||
3650  CannotBeNegativeZero(CI->getOperand(0), TLI));
3651 
3652  case Intrinsic::powi:
3653  if (ConstantInt *Exponent = dyn_cast<ConstantInt>(I->getOperand(1))) {
3654  // powi(x,n) is non-negative if n is even.
3655  if (Exponent->getBitWidth() <= 64 && Exponent->getSExtValue() % 2u == 0)
3656  return true;
3657  }
3658  // TODO: This is not correct. Given that exp is an integer, here are the
3659  // ways that pow can return a negative value:
3660  //
3661  // pow(x, exp) --> negative if exp is odd and x is negative.
3662  // pow(-0, exp) --> -inf if exp is negative odd.
3663  // pow(-0, exp) --> -0 if exp is positive odd.
3664  // pow(-inf, exp) --> -0 if exp is negative odd.
3665  // pow(-inf, exp) --> -inf if exp is positive odd.
3666  //
3667  // Therefore, if !SignBitOnly, we can return true if x >= +0 or x is NaN,
3668  // but we must return false if x == -0. Unfortunately we do not currently
3669  // have a way of expressing this constraint. See details in
3670  // https://llvm.org/bugs/show_bug.cgi?id=31702.
3671  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3672  Depth + 1);
3673 
3674  case Intrinsic::fma:
3675  case Intrinsic::fmuladd:
3676  // x*x+y is non-negative if y is non-negative.
3677  return I->getOperand(0) == I->getOperand(1) &&
3678  (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()) &&
3679  cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
3680  Depth + 1);
3681  }
3682  break;
3683  }
3684  return false;
3685 }
3686 
3688  const TargetLibraryInfo *TLI) {
3689  return cannotBeOrderedLessThanZeroImpl(V, TLI, false, 0);
3690 }
3691 
3693  return cannotBeOrderedLessThanZeroImpl(V, TLI, true, 0);
3694 }
3695 
3697  unsigned Depth) {
3698  assert(V->getType()->isFPOrFPVectorTy() && "Querying for Inf on non-FP type");
3699 
3700  // If we're told that infinities won't happen, assume they won't.
3701  if (auto *FPMathOp = dyn_cast<FPMathOperator>(V))
3702  if (FPMathOp->hasNoInfs())
3703  return true;
3704 
3705  // Handle scalar constants.
3706  if (auto *CFP = dyn_cast<ConstantFP>(V))
3707  return !CFP->isInfinity();
3708 
3710  return false;
3711 
3712  if (auto *Inst = dyn_cast<Instruction>(V)) {
3713  switch (Inst->getOpcode()) {
3714  case Instruction::Select: {
3715  return isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1) &&
3716  isKnownNeverInfinity(Inst->getOperand(2), TLI, Depth + 1);
3717  }
3718  case Instruction::SIToFP:
3719  case Instruction::UIToFP: {
3720  // Get width of largest magnitude integer (remove a bit if signed).
3721  // This still works for a signed minimum value because the largest FP
3722  // value is scaled by some fraction close to 2.0 (1.0 + 0.xxxx).
3723  int IntSize = Inst->getOperand(0)->getType()->getScalarSizeInBits();
3724  if (Inst->getOpcode() == Instruction::SIToFP)
3725  --IntSize;
3726 
3727  // If the exponent of the largest finite FP value can hold the largest
3728  // integer, the result of the cast must be finite.
3729  Type *FPTy = Inst->getType()->getScalarType();
3730  return ilogb(APFloat::getLargest(FPTy->getFltSemantics())) >= IntSize;
3731  }
3732  default:
3733  break;
3734  }
3735  }
3736 
3737  // try to handle fixed width vector constants
3738  auto *VFVTy = dyn_cast<FixedVectorType>(V->getType());
3739  if (VFVTy && isa<Constant>(V)) {
3740  // For vectors, verify that each element is not infinity.
3741  unsigned NumElts = VFVTy->getNumElements();
3742  for (unsigned i = 0; i != NumElts; ++i) {
3743  Constant *Elt = cast<Constant>(V)->getAggregateElement(i);
3744  if (!Elt)
3745  return false;
3746  if (isa<UndefValue>(Elt))
3747  continue;
3748  auto *CElt = dyn_cast<ConstantFP>(Elt);
3749  if (!CElt || CElt->isInfinity())
3750  return false;
3751  }
3752  // All elements were confirmed non-infinity or undefined.
3753  return true;
3754  }
3755 
3756  // was not able to prove that V never contains infinity
3757  return false;
3758 }
3759 
3761  unsigned Depth) {
3762  assert(V->getType()->isFPOrFPVectorTy() && "Querying for NaN on non-FP type");
3763 
3764  // If we're told that NaNs won't happen, assume they won't.
3765  if (auto *FPMathOp = dyn_cast<FPMathOperator>(V))
3766  if (FPMathOp->hasNoNaNs())
3767  return true;
3768 
3769  // Handle scalar constants.
3770  if (auto *CFP = dyn_cast<ConstantFP>(V))
3771  return !CFP->isNaN();
3772 
3774  return false;
3775 
3776  if (auto *Inst = dyn_cast<Instruction>(V)) {
3777  switch (Inst->getOpcode()) {
3778  case Instruction::FAdd:
3779  case Instruction::FSub:
3780  // Adding positive and negative infinity produces NaN.
3781  return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1) &&
3782  isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) &&
3783  (isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) ||
3784  isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1));
3785 
3786  case Instruction::FMul:
3787  // Zero multiplied with infinity produces NaN.
3788  // FIXME: If neither side can be zero fmul never produces NaN.
3789  return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1) &&
3790  isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) &&
3791  isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) &&
3792  isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1);
3793 
3794  case Instruction::FDiv:
3795  case Instruction::FRem:
3796  // FIXME: Only 0/0, Inf/Inf, Inf REM x and x REM 0 produce NaN.
3797  return false;
3798 
3799  case Instruction::Select: {
3800  return isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) &&
3801  isKnownNeverNaN(Inst->getOperand(2), TLI, Depth + 1);
3802  }
3803  case Instruction::SIToFP:
3804  case Instruction::UIToFP:
3805  return true;
3806  case Instruction::FPTrunc:
3807  case Instruction::FPExt:
3808  return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1);
3809  default:
3810  break;
3811  }
3812  }
3813 
3814  if (const auto *II = dyn_cast<IntrinsicInst>(V)) {
3815  switch (II->getIntrinsicID()) {
3816  case Intrinsic::canonicalize:
3817  case Intrinsic::fabs:
3818  case Intrinsic::copysign:
3819  case Intrinsic::exp:
3820  case Intrinsic::exp2:
3821  case Intrinsic::floor:
3822  case Intrinsic::ceil:
3823  case Intrinsic::trunc:
3824  case Intrinsic::rint:
3825  case Intrinsic::nearbyint:
3826  case Intrinsic::round:
3827  case Intrinsic::roundeven:
3828  return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1);
3829  case Intrinsic::sqrt:
3830  return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) &&
3831  CannotBeOrderedLessThanZero(II->getArgOperand(0), TLI);
3832  case Intrinsic::minnum:
3833  case Intrinsic::maxnum:
3834  // If either operand is not NaN, the result is not NaN.
3835  return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) ||
3836  isKnownNeverNaN(II->getArgOperand(1), TLI, Depth + 1);
3837  default:
3838  return false;
3839  }
3840  }
3841 
3842  // Try to handle fixed width vector constants
3843  auto *VFVTy = dyn_cast<FixedVectorType>(V->getType());
3844  if (VFVTy && isa<Constant>(V)) {
3845  // For vectors, verify that each element is not NaN.
3846  unsigned NumElts = VFVTy->getNumElements();
3847  for (unsigned i = 0; i != NumElts; ++i) {
3848  Constant *Elt = cast<Constant>(V)->getAggregateElement(i);
3849  if (!Elt)
3850  return false;
3851  if (isa<UndefValue>(Elt))
3852  continue;
3853  auto *CElt = dyn_cast<ConstantFP>(Elt);
3854  if (!CElt || CElt->isNaN())
3855  return false;
3856  }
3857  // All elements were confirmed not-NaN or undefined.
3858  return true;
3859  }
3860 
3861  // Was not able to prove that V never contains NaN
3862  return false;
3863 }
3864 
3866 
3867  // All byte-wide stores are splatable, even of arbitrary variables.
3868  if (V->getType()->isIntegerTy(8))
3869  return V;
3870 
3871  LLVMContext &Ctx = V->getContext();
3872 
3873  // Undef don't care.
3874  auto *UndefInt8 = UndefValue::get(Type::getInt8Ty(Ctx));
3875  if (isa<UndefValue>(V))
3876  return UndefInt8;
3877 
3878  // Return Undef for zero-sized type.
3879  if (!DL.getTypeStoreSize(V->getType()).isNonZero())
3880  return UndefInt8;
3881 
3882  Constant *C = dyn_cast<Constant>(V);
3883  if (!C) {
3884  // Conceptually, we could handle things like:
3885  // %a = zext i8 %X to i16
3886  // %b = shl i16 %a, 8
3887  // %c = or i16 %a, %b
3888  // but until there is an example that actually needs this, it doesn't seem
3889  // worth worrying about.
3890  return nullptr;
3891  }
3892 
3893  // Handle 'null' ConstantArrayZero etc.
3894  if (C->isNullValue())
3896 
3897  // Constant floating-point values can be handled as integer values if the
3898  // corresponding integer value is "byteable". An important case is 0.0.
3899  if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
3900  Type *Ty = nullptr;
3901  if (CFP->getType()->isHalfTy())
3902  Ty = Type::getInt16Ty(Ctx);
3903  else if (CFP->getType()->isFloatTy())
3904  Ty = Type::getInt32Ty(Ctx);
3905  else if (CFP->getType()->isDoubleTy())
3906  Ty = Type::getInt64Ty(Ctx);
3907  // Don't handle long double formats, which have strange constraints.
3908  return Ty ? isBytewiseValue(ConstantExpr::getBitCast(CFP, Ty), DL)
3909  : nullptr;
3910  }
3911 
3912  // We can handle constant integers that are multiple of 8 bits.
3913  if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
3914  if (CI->getBitWidth() % 8 == 0) {
3915  assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
3916  if (!CI->getValue().isSplat(8))
3917  return nullptr;
3918  return ConstantInt::get(Ctx, CI->getValue().trunc(8));
3919  }
3920  }
3921 
3922  if (auto *CE = dyn_cast<ConstantExpr>(C)) {
3923  if (CE->getOpcode() == Instruction::IntToPtr) {
3924  if (auto *PtrTy = dyn_cast<PointerType>(CE->getType())) {
3925  unsigned BitWidth = DL.getPointerSizeInBits(PtrTy->getAddressSpace());
3926  return isBytewiseValue(
3927  ConstantExpr::getIntegerCast(CE->getOperand(0),
3928  Type::getIntNTy(Ctx, BitWidth), false),
3929  DL);
3930  }
3931  }
3932  }
3933 
3934  auto Merge = [&](Value *LHS, Value *RHS) -> Value * {
3935  if (LHS == RHS)
3936  return LHS;
3937  if (!LHS || !RHS)
3938  return nullptr;
3939  if (LHS == UndefInt8)
3940  return RHS;
3941  if (RHS == UndefInt8)
3942  return LHS;
3943  return nullptr;
3944  };
3945 
3946  if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(C)) {
3947  Value *Val = UndefInt8;
3948  for (unsigned I = 0, E = CA->getNumElements(); I != E; ++I)
3949  if (!(Val = Merge(Val, isBytewiseValue(CA->getElementAsConstant(I), DL))))
3950  return nullptr;
3951  return Val;
3952  }
3953 
3954  if (isa<ConstantAggregate>(C)) {
3955  Value *Val = UndefInt8;
3956  for (unsigned I = 0, E = C->getNumOperands(); I != E; ++I)
3957  if (!(Val = Merge(Val, isBytewiseValue(C->getOperand(I), DL))))
3958  return nullptr;
3959  return Val;
3960  }
3961 
3962  // Don't try to handle the handful of other constants.
3963  return nullptr;
3964 }
3965 
3966 // This is the recursive version of BuildSubAggregate. It takes a few different
3967 // arguments. Idxs is the index within the nested struct From that we are
3968 // looking at now (which is of type IndexedType). IdxSkip is the number of
3969 // indices from Idxs that should be left out when inserting into the resulting
3970 // struct. To is the result struct built so far, new insertvalue instructions
3971 // build on that.
3972 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
3974  unsigned IdxSkip,
3975  Instruction *InsertBefore) {
3976  StructType *STy = dyn_cast<StructType>(IndexedType);
3977  if (STy) {
3978  // Save the original To argument so we can modify it
3979  Value *OrigTo = To;
3980  // General case, the type indexed by Idxs is a struct
3981  for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
3982  // Process each struct element recursively
3983  Idxs.push_back(i);
3984  Value *PrevTo = To;
3985  To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
3986  InsertBefore);
3987  Idxs.pop_back();
3988  if (!To) {
3989  // Couldn't find any inserted value for this index? Cleanup
3990  while (PrevTo != OrigTo) {
3991  InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
3992  PrevTo = Del->getAggregateOperand();
3993  Del->eraseFromParent();
3994  }
3995  // Stop processing elements
3996  break;
3997  }
3998  }
3999  // If we successfully found a value for each of our subaggregates
4000  if (To)
4001  return To;
4002  }
4003  // Base case, the type indexed by SourceIdxs is not a struct, or not all of
4004  // the struct's elements had a value that was inserted directly. In the latter
4005  // case, perhaps we can't determine each of the subelements individually, but
4006  // we might be able to find the complete struct somewhere.
4007 
4008  // Find the value that is at that particular spot
4009  Value *V = FindInsertedValue(From, Idxs);
4010 
4011  if (!V)
4012  return nullptr;
4013 
4014  // Insert the value in the new (sub) aggregate
4015  return InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
4016  "tmp", InsertBefore);
4017 }
4018 
4019 // This helper takes a nested struct and extracts a part of it (which is again a
4020 // struct) into a new value. For example, given the struct:
4021 // { a, { b, { c, d }, e } }
4022 // and the indices "1, 1" this returns
4023 // { c, d }.
4024 //
4025 // It does this by inserting an insertvalue for each element in the resulting
4026 // struct, as opposed to just inserting a single struct. This will only work if
4027 // each of the elements of the substruct are known (ie, inserted into From by an
4028 // insertvalue instruction somewhere).
4029 //
4030 // All inserted insertvalue instructions are inserted before InsertBefore
4032  Instruction *InsertBefore) {
4033  assert(InsertBefore && "Must have someplace to insert!");
4034  Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
4035  idx_range);
4036  Value *To = UndefValue::get(IndexedType);
4037  SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
4038  unsigned IdxSkip = Idxs.size();
4039 
4040  return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
4041 }
4042 
4043 /// Given an aggregate and a sequence of indices, see if the scalar value
4044 /// indexed is already around as a register, for example if it was inserted
4045 /// directly into the aggregate.
4046 ///
4047 /// If InsertBefore is not null, this function will duplicate (modified)
4048 /// insertvalues when a part of a nested struct is extracted.
4050  Instruction *InsertBefore) {
4051  // Nothing to index? Just return V then (this is useful at the end of our
4052  // recursion).
4053  if (idx_range.empty())
4054  return V;
4055  // We have indices, so V should have an indexable type.
4056  assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
4057  "Not looking at a struct or array?");
4058  assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
4059  "Invalid indices for type?");
4060 
4061  if (Constant *C = dyn_cast<Constant>(V)) {
4062  C = C->getAggregateElement(idx_range[0]);
4063  if (!C) return nullptr;
4064  return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
4065  }
4066 
4067  if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
4068  // Loop the indices for the insertvalue instruction in parallel with the
4069  // requested indices
4070  const unsigned *req_idx = idx_range.begin();
4071  for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
4072  i != e; ++i, ++req_idx) {
4073  if (req_idx == idx_range.end()) {
4074  // We can't handle this without inserting insertvalues
4075  if (!InsertBefore)
4076  return nullptr;
4077 
4078  // The requested index identifies a part of a nested aggregate. Handle
4079  // this specially. For example,
4080  // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
4081  // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
4082  // %C = extractvalue {i32, { i32, i32 } } %B, 1
4083  // This can be changed into
4084  // %A = insertvalue {i32, i32 } undef, i32 10, 0
4085  // %C = insertvalue {i32, i32 } %A, i32 11, 1
4086  // which allows the unused 0,0 element from the nested struct to be
4087  // removed.
4088  return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
4089  InsertBefore);
4090  }
4091 
4092  // This insert value inserts something else than what we are looking for.
4093  // See if the (aggregate) value inserted into has the value we are
4094  // looking for, then.
4095  if (*req_idx != *i)
4096  return FindInsertedValue(I->getAggregateOperand(), idx_range,
4097  InsertBefore);
4098  }
4099  // If we end up here, the indices of the insertvalue match with those
4100  // requested (though possibly only partially). Now we recursively look at
4101  // the inserted value, passing any remaining indices.
4102  return FindInsertedValue(I->getInsertedValueOperand(),
4103  makeArrayRef(req_idx, idx_range.end()),
4104  InsertBefore);
4105  }
4106 
4107  if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
4108  // If we're extracting a value from an aggregate that was extracted from
4109  // something else, we can extract from that something else directly instead.
4110  // However, we will need to chain I's indices with the requested indices.
4111 
4112  // Calculate the number of indices required
4113  unsigned size = I->getNumIndices() + idx_range.size();
4114  // Allocate some space to put the new indices in
4116  Idxs.reserve(size);
4117  // Add indices from the extract value instruction
4118  Idxs.append(I->idx_begin(), I->idx_end());
4119 
4120  // Add requested indices
4121  Idxs.append(idx_range.begin(), idx_range.end());
4122 
4123  assert(Idxs.size() == size
4124  && "Number of indices added not correct?");
4125 
4126  return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
4127  }
4128  // Otherwise, we don't know (such as, extracting from a function return value
4129  // or load instruction)
4130  return nullptr;
4131 }
4132 
4134  unsigned CharSize) {
4135  // Make sure the GEP has exactly three arguments.
4136  if (GEP->getNumOperands() != 3)
4137  return false;
4138 
4139  // Make sure the index-ee is a pointer to array of \p CharSize integers.
4140  // CharSize.
4141  ArrayType *AT = dyn_cast<ArrayType>(GEP->getSourceElementType());
4142  if (!AT || !AT->getElementType()->isIntegerTy(CharSize))
4143  return false;
4144 
4145  // Check to make sure that the first operand of the GEP is an integer and
4146  // has value 0 so that we are sure we're indexing into the initializer.
4147  const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
4148  if (!FirstIdx || !FirstIdx->isZero())
4149  return false;
4150 
4151  return true;
4152 }
4153 
4155  ConstantDataArraySlice &Slice,
4156  unsigned ElementSize, uint64_t Offset) {
4157  assert(V);
4158 
4159  // Look through bitcast instructions and geps.
4160  V = V->stripPointerCasts();
4161 
4162  // If the value is a GEP instruction or constant expression, treat it as an
4163  // offset.
4164  if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
4165  // The GEP operator should be based on a pointer to string constant, and is
4166  // indexing into the string constant.
4167  if (!isGEPBasedOnPointerToString(GEP, ElementSize))
4168  return false;
4169 
4170  // If the second index isn't a ConstantInt, then this is a variable index
4171  // into the array. If this occurs, we can't say anything meaningful about
4172  // the string.
4173  uint64_t StartIdx = 0;
4174  if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
4175  StartIdx = CI->getZExtValue();
4176  else
4177  return false;
4178  return getConstantDataArrayInfo(GEP->getOperand(0), Slice, ElementSize,
4179  StartIdx + Offset);
4180  }
4181 
4182  // The GEP instruction, constant or instruction, must reference a global
4183  // variable that is a constant and is initialized. The referenced constant
4184  // initializer is the array that we'll use for optimization.
4185  const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
4186  if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
4187  return false;
4188 
4189  const ConstantDataArray *Array;
4190  ArrayType *ArrayTy;
4191  if (GV->getInitializer()->isNullValue()) {
4192  Type *GVTy = GV->getValueType();
4193  if ( (ArrayTy = dyn_cast<ArrayType>(GVTy)) ) {
4194  // A zeroinitializer for the array; there is no ConstantDataArray.
4195  Array = nullptr;
4196  } else {
4197  const DataLayout &DL = GV->getParent()->getDataLayout();
4198  uint64_t SizeInBytes = DL.getTypeStoreSize(GVTy).getFixedSize();
4199  uint64_t Length = SizeInBytes / (ElementSize / 8);
4200  if (Length <= Offset)
4201  return false;
4202 
4203  Slice.Array = nullptr;
4204  Slice.Offset = 0;
4205  Slice.Length = Length - Offset;
4206  return true;
4207  }
4208  } else {
4209  // This must be a ConstantDataArray.
4210  Array = dyn_cast<ConstantDataArray>(GV->getInitializer());
4211  if (!Array)
4212  return false;
4213  ArrayTy = Array->getType();
4214  }
4215  if (!ArrayTy->getElementType()->isIntegerTy(ElementSize))
4216  return false;
4217 
4218  uint64_t NumElts = ArrayTy->getArrayNumElements();
4219  if (Offset > NumElts)
4220  return false;
4221 
4222  Slice.Array = Array;
4223  Slice.Offset = Offset;
4224  Slice.Length = NumElts - Offset;
4225  return true;
4226 }
4227 
4228 /// This function computes the length of a null-terminated C string pointed to
4229 /// by V. If successful, it returns true and returns the string in Str.
4230 /// If unsuccessful, it returns false.
4232  uint64_t Offset, bool TrimAtNul) {
4233  ConstantDataArraySlice Slice;
4234  if (!getConstantDataArrayInfo(V, Slice, 8, Offset))
4235  return false;
4236 
4237  if (Slice.Array == nullptr) {
4238  if (TrimAtNul) {
4239  Str = StringRef();
4240  return true;
4241  }
4242  if (Slice.Length == 1) {
4243  Str = StringRef("", 1);
4244  return true;
4245  }
4246  // We cannot instantiate a StringRef as we do not have an appropriate string
4247  // of 0s at hand.
4248  return false;
4249  }
4250 
4251  // Start out with the entire array in the StringRef.
4252  Str = Slice.Array->getAsString();
4253  // Skip over 'offset' bytes.
4254  Str = Str.substr(Slice.Offset);
4255 
4256  if (TrimAtNul) {
4257  // Trim off the \0 and anything after it. If the array is not nul
4258  // terminated, we just return the whole end of string. The client may know
4259  // some other way that the string is length-bound.
4260  Str = Str.substr(0, Str.find('\0'));
4261  }
4262  return true;
4263 }
4264 
4265 // These next two are very similar to the above, but also look through PHI
4266 // nodes.
4267 // TODO: See if we can integrate these two together.
4268 
4269 /// If we can compute the length of the string pointed to by
4270 /// the specified pointer, return 'len+1'. If we can't, return 0.
4273  unsigned CharSize) {
4274  // Look through noop bitcast instructions.
4275  V = V->stripPointerCasts();
4276 
4277  // If this is a PHI node, there are two cases: either we have already seen it
4278  // or we haven't.
4279  if (const PHINode *PN = dyn_cast<PHINode>(V)) {
4280  if (!PHIs.insert(PN).second)
4281  return ~0ULL; // already in the set.
4282 
4283  // If it was new, see if all the input strings are the same length.
4284  uint64_t LenSoFar = ~0ULL;
4285  for (Value *IncValue : PN->incoming_values()) {
4286  uint64_t Len = GetStringLengthH(IncValue, PHIs, CharSize);
4287  if (Len == 0) return 0; // Unknown length -> unknown.
4288 
4289  if (Len == ~0ULL) continue;
4290 
4291  if (Len != LenSoFar && LenSoFar != ~0ULL)
4292  return 0; // Disagree -> unknown.
4293  LenSoFar = Len;
4294  }
4295 
4296  // Success, all agree.
4297  return LenSoFar;
4298  }
4299 
4300  // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
4301  if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
4302  uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs, CharSize);
4303  if (Len1 == 0) return 0;
4304  uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs, CharSize);
4305  if (Len2 == 0) return 0;
4306  if (Len1 == ~0ULL) return Len2;
4307  if (Len2 == ~0ULL) return Len1;
4308  if (Len1 != Len2) return 0;
4309  return Len1;
4310  }
4311 
4312  // Otherwise, see if we can read the string.
4313  ConstantDataArraySlice Slice;
4314  if (!getConstantDataArrayInfo(V, Slice, CharSize))
4315  return 0;
4316 
4317  if (Slice.Array == nullptr)
4318  return 1;
4319 
4320  // Search for nul characters
4321  unsigned NullIndex = 0;
4322  for (unsigned E = Slice.Length; NullIndex < E; ++NullIndex) {
4323  if (Slice.Array->getElementAsInteger(Slice.Offset + NullIndex) == 0)
4324  break;
4325  }
4326 
4327  return NullIndex + 1;
4328 }
4329 
4330 /// If we can compute the length of the string pointed to by
4331 /// the specified pointer, return 'len+1'. If we can't, return 0.
4332 uint64_t llvm::GetStringLength(const Value *V, unsigned CharSize) {
4333  if (!V->getType()->isPointerTy())
4334  return 0;
4335 
4337  uint64_t Len = GetStringLengthH(V, PHIs, CharSize);
4338  // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
4339  // an empty string as a length.
4340  return Len == ~0ULL ? 1 : Len;
4341 }
4342 
4343 const Value *
4345  bool MustPreserveNullness) {
4346  assert(Call &&
4347  "getArgumentAliasingToReturnedPointer only works on nonnull calls");
4348  if (const Value *RV = Call->getReturnedArgOperand())
4349  return RV;
4350  // This can be used only as a aliasing property.
4352  Call, MustPreserveNullness))
4353  return Call->getArgOperand(0);
4354  return nullptr;
4355 }
4356 
4358  const CallBase *Call, bool MustPreserveNullness) {
4359  switch (Call->getIntrinsicID()) {
4360  case Intrinsic::launder_invariant_group:
4361  case Intrinsic::strip_invariant_group:
4362  case Intrinsic::aarch64_irg:
4363  case Intrinsic::aarch64_tagp:
4364  return true;
4365  case Intrinsic::ptrmask:
4366  return !MustPreserveNullness;
4367  default:
4368  return false;
4369  }
4370 }
4371 
4372 /// \p PN defines a loop-variant pointer to an object. Check if the
4373 /// previous iteration of the loop was referring to the same object as \p PN.
4375  const LoopInfo *LI) {
4376  // Find the loop-defined value.
4377  Loop *L = LI->getLoopFor(PN->getParent());
4378  if (PN->getNumIncomingValues() != 2)
4379  return true;
4380 
4381  // Find the value from previous iteration.
4382  auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
4383  if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
4384  PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
4385  if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
4386  return true;
4387 
4388  // If a new pointer is loaded in the loop, the pointer references a different
4389  // object in every iteration. E.g.:
4390  // for (i)
4391  // int *p = a[i];
4392  // ...
4393  if (auto *Load = dyn_cast<LoadInst>(PrevValue))
4394  if (!L->isLoopInvariant(Load->getPointerOperand()))
4395  return false;
4396  return true;
4397 }
4398 
4399 const Value *llvm::getUnderlyingObject(const Value *V, unsigned MaxLookup) {
4400  if (!V->getType()->isPointerTy())
4401  return V;
4402  for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
4403  if (auto *GEP = dyn_cast<GEPOperator>(V)) {
4404  V = GEP->getPointerOperand();
4405  } else if (Operator::getOpcode(V) == Instruction::BitCast ||
4406  Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
4407  V = cast<Operator>(V)->getOperand(0);
4408  if (!V->getType()->isPointerTy())
4409  return V;
4410  } else if (auto *GA = dyn_cast<GlobalAlias>(V)) {
4411  if (GA->isInterposable())
4412  return V;
4413  V = GA->getAliasee();
4414  } else {
4415  if (auto *PHI = dyn_cast<PHINode>(V)) {
4416  // Look through single-arg phi nodes created by LCSSA.
4417  if (PHI->getNumIncomingValues() == 1) {
4418  V = PHI->getIncomingValue(0);
4419  continue;
4420  }
4421  } else if (auto *Call = dyn_cast<CallBase>(V)) {
4422  // CaptureTracking can know about special capturing properties of some
4423  // intrinsics like launder.invariant.group, that can't be expressed with
4424  // the attributes, but have properties like returning aliasing pointer.
4425  // Because some analysis may assume that nocaptured pointer is not
4426  // returned from some special intrinsic (because function would have to
4427  // be marked with returns attribute), it is crucial to use this function
4428  // because it should be in sync with CaptureTracking. Not using it may
4429  // cause weird miscompilations where 2 aliasing pointers are assumed to
4430  // noalias.
4431  if (auto *RP = getArgumentAliasingToReturnedPointer(Call, false)) {
4432  V = RP;
4433  continue;
4434  }
4435  }
4436 
4437  return V;
4438  }
4439  assert(V->getType()->isPointerTy() && "Unexpected operand type!");
4440  }
4441  return V;
4442 }
4443 
4446  LoopInfo *LI, unsigned MaxLookup) {
4449  Worklist.push_back(V);
4450  do {
4451  const Value *P = Worklist.pop_back_val();
4452  P = getUnderlyingObject(P, MaxLookup);
4453 
4454  if (!Visited.insert(P).second)
4455  continue;
4456 
4457  if (auto *SI = dyn_cast<SelectInst>(P)) {
4458  Worklist.push_back(SI->getTrueValue());
4459  Worklist.push_back(SI->getFalseValue());
4460  continue;
4461  }
4462 
4463  if (auto *PN = dyn_cast<PHINode>(P)) {
4464  // If this PHI changes the underlying object in every iteration of the
4465  // loop, don't look through it. Consider:
4466  // int **A;
4467  // for (i) {
4468  // Prev = Curr; // Prev = PHI (Prev_0, Curr)
4469  // Curr = A[i];
4470  // *Prev, *Curr;
4471  //
4472  // Prev is tracking Curr one iteration behind so they refer to different
4473  // underlying objects.
4474  if (!LI || !LI->isLoopHeader(PN->getParent()) ||
4476  append_range(Worklist, PN->incoming_values());
4477  continue;
4478  }
4479 
4480  Objects.push_back(P);
4481  } while (!Worklist.empty());
4482 }
4483 
4484 /// This is the function that does the work of looking through basic
4485 /// ptrtoint+arithmetic+inttoptr sequences.
4486 static const Value *getUnderlyingObjectFromInt(const Value *V) {
4487  do {
4488  if (const Operator *U = dyn_cast<Operator>(V)) {
4489  // If we find a ptrtoint, we can transfer control back to the
4490  // regular getUnderlyingObjectFromInt.
4491  if (U->getOpcode() == Instruction::PtrToInt)
4492  return U->getOperand(0);
4493  // If we find an add of a constant, a multiplied value, or a phi, it's
4494  // likely that the other operand will lead us to the base
4495  // object. We don't have to worry about the case where the
4496  // object address is somehow being computed by the multiply,
4497  // because our callers only care when the result is an
4498  // identifiable object.
4499  if (U->getOpcode() != Instruction::Add ||
4500  (!isa<ConstantInt>(U->getOperand(1)) &&
4501  Operator::getOpcode(U->getOperand(1)) != Instruction::Mul &&
4502  !isa<PHINode>(U->getOperand(1))))
4503  return V;
4504  V = U->getOperand(0);
4505  } else {
4506  return V;
4507  }
4508  assert(V->getType()->isIntegerTy() && "Unexpected operand type!");
4509  } while (true);
4510 }
4511 
4512 /// This is a wrapper around getUnderlyingObjects and adds support for basic
4513 /// ptrtoint+arithmetic+inttoptr sequences.
4514 /// It returns false if unidentified object is found in getUnderlyingObjects.
4516  SmallVectorImpl<Value *> &Objects) {
4518  SmallVector<const Value *, 4> Working(1, V);
4519  do {
4520  V = Working.pop_back_val();
4521 
4523  getUnderlyingObjects(V, Objs);
4524 
4525  for (const Value *V : Objs) {
4526  if (!Visited.insert(V).second)
4527  continue;
4528  if (Operator::getOpcode(V) == Instruction::IntToPtr) {
4529  const Value *O =
4530  getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0));
4531  if (O->getType()->isPointerTy()) {
4532  Working.push_back(O);
4533  continue;
4534  }
4535  }
4536  // If getUnderlyingObjects fails to find an identifiable object,
4537  // getUnderlyingObjectsForCodeGen also fails for safety.
4538  if (!isIdentifiedObject(V)) {
4539  Objects.clear();
4540  return false;
4541  }
4542  Objects.push_back(const_cast<Value *>(V));
4543  }
4544  } while (!Working.empty());
4545  return true;
4546 }
4547 
4549  AllocaInst *Result = nullptr;
4550  SmallPtrSet<Value *, 4> Visited;
4551  SmallVector<Value *, 4> Worklist;
4552 
4553  auto AddWork = [&](Value *V) {
4554  if (Visited.insert(V).second)
4555  Worklist.push_back(V);
4556  };
4557 
4558  AddWork(V);
4559  do {
4560  V = Worklist.pop_back_val();
4561  assert(Visited.count(V));
4562 
4563  if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
4564  if (Result && Result != AI)
4565  return nullptr;
4566  Result = AI;
4567  } else if (CastInst *CI = dyn_cast<CastInst>(V)) {
4568  AddWork(CI->getOperand(0));
4569  } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
4570  for (Value *IncValue : PN->incoming_values())
4571  AddWork(IncValue);
4572  } else if (auto *SI = dyn_cast<SelectInst>(V)) {
4573  AddWork(SI->getTrueValue());
4574  AddWork(SI->getFalseValue());
4575  } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
4576  if (OffsetZero && !GEP->hasAllZeroIndices())
4577  return nullptr;
4578  AddWork(GEP->getPointerOperand());
4579  } else if (CallBase *CB = dyn_cast<CallBase>(V)) {
4580  Value *Returned = CB->getReturnedArgOperand();
4581  if (Returned)
4582  AddWork(Returned);
4583  else
4584  return nullptr;
4585  } else {
4586  return nullptr;
4587  }
4588  } while (!Worklist.empty());
4589 
4590  return Result;
4591 }
4592 
4594  const Value *V, bool AllowLifetime, bool AllowDroppable) {
4595  for (const User *U : V->users()) {
4596  const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
4597  if (!II)
4598  return false;
4599 
4600  if (AllowLifetime && II->isLifetimeStartOrEnd())
4601  continue;
4602 
4603  if (AllowDroppable && II->isDroppable())
4604  continue;
4605 
4606  return false;
4607  }
4608  return true;
4609 }
4610 
4613  V, /* AllowLifetime */ true, /* AllowDroppable */ false);
4614 }
4617  V, /* AllowLifetime */ true, /* AllowDroppable */ true);
4618 }
4619 
4621  if (!LI.isUnordered())
4622  return true;
4623  const Function &F = *LI.getFunction();
4624  // Speculative load may create a race that did not exist in the source.
4625  return F.hasFnAttribute(Attribute::SanitizeThread) ||
4626  // Speculative load may load data from dirty regions.
4627  F.hasFnAttribute(Attribute::SanitizeAddress) ||
4628  F.hasFnAttribute(Attribute::SanitizeHWAddress);
4629 }
4630 
4631 
4633  const Instruction *CtxI,
4634  const DominatorTree *DT,
4635  const TargetLibraryInfo *TLI) {
4636  const Operator *Inst = dyn_cast<Operator>(V);
4637  if (!Inst)
4638  return false;
4639 
4640  for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
4641  if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
4642  if (C->canTrap())
4643  return false;
4644 
4645  switch (Inst->getOpcode()) {
4646  default:
4647  return true;
4648  case Instruction::UDiv:
4649  case Instruction::URem: {
4650  // x / y is undefined if y == 0.
4651  const APInt *V;
4652  if (match(Inst->getOperand(1), m_APInt(V)))
4653  return *V != 0;
4654  return false;
4655  }
4656  case Instruction::SDiv:
4657  case Instruction::SRem: {
4658  // x / y is undefined if y == 0 or x == INT_MIN and y == -1
4659  const APInt *Numerator, *Denominator;
4660  if (!match(Inst->getOperand(1), m_APInt(Denominator)))
4661  return false;
4662  // We cannot hoist this division if the denominator is 0.
4663  if (*Denominator == 0)
4664  return false;
4665  // It's safe to hoist if the denominator is not 0 or -1.
4666  if (!Denominator->isAllOnes())
4667  return true;
4668  // At this point we know that the denominator is -1. It is safe to hoist as
4669  // long we know that the numerator is not INT_MIN.
4670  if (match(Inst->getOperand(0), m_APInt(Numerator)))
4671  return !Numerator->isMinSignedValue();
4672  // The numerator *might* be MinSignedValue.
4673  return false;
4674  }
4675  case Instruction::Load: {
4676  const LoadInst *LI = cast<LoadInst>(Inst);
4677  if (mustSuppressSpeculation(*LI))
4678  return false;
4679  const DataLayout &DL = LI->getModule()->getDataLayout();
4681  LI->getPointerOperand(), LI->getType(), MaybeAlign(LI->getAlign()), DL,
4682  CtxI, DT, TLI);
4683  }
4684  case Instruction::Call: {
4685  auto *CI = cast<const CallInst>(Inst);
4686  const Function *Callee = CI->getCalledFunction();
4687 
4688  // The called function could have undefined behavior or side-effects, even
4689  // if marked readnone nounwind.
4690  return Callee && Callee->isSpeculatable();
4691  }
4692  case Instruction::VAArg:
4693  case Instruction::Alloca:
4694  case Instruction::Invoke:
4695  case Instruction::CallBr:
4696  case Instruction::PHI:
4697  case Instruction::Store:
4698  case Instruction::Ret:
4699  case Instruction::Br:
4700  case Instruction::IndirectBr:
4701  case Instruction::Switch:
4702  case Instruction::Unreachable:
4703  case Instruction::Fence:
4704  case Instruction::AtomicRMW:
4705  case Instruction::AtomicCmpXchg:
4706  case Instruction::LandingPad:
4707  case Instruction::Resume:
4708  case Instruction::CatchSwitch:
4709  case Instruction::CatchPad:
4710  case Instruction::CatchRet:
4711  case Instruction::CleanupPad:
4712  case Instruction::CleanupRet:
4713  return false; // Misc instructions which have effects
4714  }
4715 }
4716 
4718  return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I);
4719 }
4720 
4721 /// Convert ConstantRange OverflowResult into ValueTracking OverflowResult.
4723  switch (OR) {
4732  }
4733  llvm_unreachable("Unknown OverflowResult");
4734 }
4735 
4736 /// Combine constant ranges from computeConstantRange() and computeKnownBits().
4738  const Value *V, bool ForSigned, const DataLayout &DL, unsigned Depth,
4739  AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT,
4740  OptimizationRemarkEmitter *ORE = nullptr, bool UseInstrInfo = true) {
4741  KnownBits Known = computeKnownBits(
4742  V, DL, Depth, AC, CxtI, DT, ORE, UseInstrInfo);
4743  ConstantRange CR1 = ConstantRange::fromKnownBits(Known, ForSigned);
4744  ConstantRange CR2 = computeConstantRange(V, UseInstrInfo);
4747  return CR1.intersectWith(CR2, RangeType);
4748 }
4749 
4751  const Value *LHS, const Value *RHS, const DataLayout &DL,
4752  AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT,
4753  bool UseInstrInfo) {
4754  KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT,
4755  nullptr, UseInstrInfo);
4756  KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT,
4757  nullptr, UseInstrInfo);
4758  ConstantRange LHSRange = ConstantRange::fromKnownBits(LHSKnown, false);
4759  ConstantRange RHSRange = ConstantRange::fromKnownBits(RHSKnown, false);
4760  return mapOverflowResult(LHSRange.unsignedMulMayOverflow(RHSRange));
4761 }
4762 
4765  const DataLayout &DL, AssumptionCache *AC,
4766  const Instruction *CxtI,
4767  const DominatorTree *DT, bool UseInstrInfo) {
4768  // Multiplying n * m significant bits yields a result of n + m significant
4769  // bits. If the total number of significant bits does not exceed the
4770  // result bit width (minus 1), there is no overflow.
4771  // This means if we have enough leading sign bits in the operands
4772  // we can guarantee that the result does not overflow.
4773  // Ref: "Hacker's Delight" by Henry Warren
4774  unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
4775 
4776  // Note that underestimating the number of sign bits gives a more
4777  // conservative answer.
4778  unsigned SignBits = ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) +
4779  ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT);
4780 
4781  // First handle the easy case: if we have enough sign bits there's
4782  // definitely no overflow.
4783  if (SignBits > BitWidth + 1)
4785 
4786  // There are two ambiguous cases where there can be no overflow:
4787  // SignBits == BitWidth + 1 and
4788  // SignBits == BitWidth
4789  // The second case is difficult to check, therefore we only handle the
4790  // first case.
4791  if (SignBits == BitWidth + 1) {
4792  // It overflows only when both arguments are negative and the true
4793  // product is exactly the minimum negative number.
4794  // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000
4795  // For simplicity we just check if at least one side is not negative.
4796  KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT,
4797  nullptr, UseInstrInfo);
4798  KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT,
4799  nullptr, UseInstrInfo);
4800  if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative())
4802  }
4804 }
4805 
4807  const Value *LHS, const Value *RHS, const DataLayout &DL,
4808  AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT,
4809  bool UseInstrInfo) {
4811  LHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT,
4812  nullptr, UseInstrInfo);
4814  RHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT,
4815  nullptr, UseInstrInfo);
4816  return mapOverflowResult(LHSRange.unsignedAddMayOverflow(RHSRange));
4817 }
4818 
4820  const Value *RHS,
4821  const AddOperator *Add,
4822  const DataLayout &DL,
4823  AssumptionCache *AC,
4824  const Instruction *CxtI,
4825  const DominatorTree *DT) {
4826  if (Add && Add->hasNoSignedWrap()) {
4828  }
4829 
4830  // If LHS and RHS each have at least two sign bits, the addition will look
4831  // like
4832  //
4833  // XX..... +
4834  // YY.....
4835  //
4836  // If the carry into the most significant position is 0, X and Y can't both
4837  // be 1 and therefore the carry out of the addition is also 0.
4838  //
4839  // If the carry into the most significant position is 1, X and Y can't both
4840  // be 0 and therefore the carry out of the addition is also 1.
4841  //
4842  // Since the carry into the most significant position is always equal to
4843  // the carry out of the addition, there is no signed overflow.
4844  if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 &&
4845  ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1)
4847 
4849  LHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT);
4851  RHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT);
4852  OverflowResult OR =
4853  mapOverflowResult(LHSRange.signedAddMayOverflow(RHSRange));
4855  return OR;
4856 
4857  // The remaining code needs Add to be available. Early returns if not so.
4858  if (!Add)
4860 
4861  // If the sign of Add is the same as at least one of the operands, this add
4862  // CANNOT overflow. If this can be determined from the known bits of the
4863  // operands the above signedAddMayOverflow() check will have already done so.
4864  // The only other way to improve on the known bits is from an assumption, so
4865  // call computeKnownBitsFromAssume() directly.
4866  bool LHSOrRHSKnownNonNegative =
4867  (LHSRange.isAllNonNegative() || RHSRange.isAllNonNegative());
4868  bool LHSOrRHSKnownNegative =
4869  (LHSRange.isAllNegative() || RHSRange.isAllNegative());
4870  if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
4871  KnownBits AddKnown(LHSRange.getBitWidth());
4873  Add, AddKnown, /*Depth=*/0, Query(DL, AC, CxtI, DT, true));
4874  if ((AddKnown.isNonNegative() && LHSOrRHSKnownNonNegative) ||
4875  (AddKnown.isNegative() && LHSOrRHSKnownNegative))
4877  }
4878 
4880 }
4881 
4883  const Value *RHS,
4884  const DataLayout &DL,
4885  AssumptionCache *AC,
4886  const Instruction *CxtI,
4887  const DominatorTree *DT) {
4888  // Checking for conditions implied by dominating conditions may be expensive.
4889  // Limit it to usub_with_overflow calls for now.
4890  if (match(CxtI,
4891  m_Intrinsic<Intrinsic::usub_with_overflow>(m_Value(), m_Value())))
4892  if (auto C =
4894  if (*C)
4897  }
4899  LHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT);