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 
3252  const TargetLibraryInfo *TLI) {
3253  const Function *F = CB.getCalledFunction();
3254  if (!F)
3255  return Intrinsic::not_intrinsic;
3256 
3257  if (F->isIntrinsic())
3258  return F->getIntrinsicID();
3259 
3260  // We are going to infer semantics of a library function based on mapping it
3261  // to an LLVM intrinsic. Check that the library function is available from
3262  // this callbase and in this environment.
3263  LibFunc Func;
3264  if (F->hasLocalLinkage() || !TLI || !TLI->getLibFunc(CB, Func) ||
3265  !CB.onlyReadsMemory())
3266  return Intrinsic::not_intrinsic;
3267 
3268  switch (Func) {
3269  default:
3270  break;
3271  case LibFunc_sin:
3272  case LibFunc_sinf:
3273  case LibFunc_sinl:
3274  return Intrinsic::sin;
3275  case LibFunc_cos:
3276  case LibFunc_cosf:
3277  case LibFunc_cosl:
3278  return Intrinsic::cos;
3279  case LibFunc_exp:
3280  case LibFunc_expf:
3281  case LibFunc_expl:
3282  return Intrinsic::exp;
3283  case LibFunc_exp2:
3284  case LibFunc_exp2f:
3285  case LibFunc_exp2l:
3286  return Intrinsic::exp2;
3287  case LibFunc_log:
3288  case LibFunc_logf:
3289  case LibFunc_logl:
3290  return Intrinsic::log;
3291  case LibFunc_log10:
3292  case LibFunc_log10f:
3293  case LibFunc_log10l:
3294  return Intrinsic::log10;
3295  case LibFunc_log2:
3296  case LibFunc_log2f:
3297  case LibFunc_log2l:
3298  return Intrinsic::log2;
3299  case LibFunc_fabs:
3300  case LibFunc_fabsf:
3301  case LibFunc_fabsl:
3302  return Intrinsic::fabs;
3303  case LibFunc_fmin:
3304  case LibFunc_fminf:
3305  case LibFunc_fminl:
3306  return Intrinsic::minnum;
3307  case LibFunc_fmax:
3308  case LibFunc_fmaxf:
3309  case LibFunc_fmaxl:
3310  return Intrinsic::maxnum;
3311  case LibFunc_copysign:
3312  case LibFunc_copysignf:
3313  case LibFunc_copysignl:
3314  return Intrinsic::copysign;
3315  case LibFunc_floor:
3316  case LibFunc_floorf:
3317  case LibFunc_floorl:
3318  return Intrinsic::floor;
3319  case LibFunc_ceil:
3320  case LibFunc_ceilf:
3321  case LibFunc_ceill:
3322  return Intrinsic::ceil;
3323  case LibFunc_trunc:
3324  case LibFunc_truncf:
3325  case LibFunc_truncl:
3326  return Intrinsic::trunc;
3327  case LibFunc_rint:
3328  case LibFunc_rintf:
3329  case LibFunc_rintl:
3330  return Intrinsic::rint;
3331  case LibFunc_nearbyint:
3332  case LibFunc_nearbyintf:
3333  case LibFunc_nearbyintl:
3334  return Intrinsic::nearbyint;
3335  case LibFunc_round:
3336  case LibFunc_roundf:
3337  case LibFunc_roundl:
3338  return Intrinsic::round;
3339  case LibFunc_roundeven:
3340  case LibFunc_roundevenf:
3341  case LibFunc_roundevenl:
3342  return Intrinsic::roundeven;
3343  case LibFunc_pow:
3344  case LibFunc_powf:
3345  case LibFunc_powl:
3346  return Intrinsic::pow;
3347  case LibFunc_sqrt:
3348  case LibFunc_sqrtf:
3349  case LibFunc_sqrtl:
3350  return Intrinsic::sqrt;
3351  }
3352 
3353  return Intrinsic::not_intrinsic;
3354 }
3355 
3356 /// Return true if we can prove that the specified FP value is never equal to
3357 /// -0.0.
3358 /// NOTE: Do not check 'nsz' here because that fast-math-flag does not guarantee
3359 /// that a value is not -0.0. It only guarantees that -0.0 may be treated
3360 /// the same as +0.0 in floating-point ops.
3361 ///
3362 /// NOTE: this function will need to be revisited when we support non-default
3363 /// rounding modes!
3365  unsigned Depth) {
3366  if (auto *CFP = dyn_cast<ConstantFP>(V))
3367  return !CFP->getValueAPF().isNegZero();
3368 
3370  return false;
3371 
3372  auto *Op = dyn_cast<Operator>(V);
3373  if (!Op)
3374  return false;
3375 
3376  // (fadd x, 0.0) is guaranteed to return +0.0, not -0.0.
3377  if (match(Op, m_FAdd(m_Value(), m_PosZeroFP())))
3378  return true;
3379 
3380  // sitofp and uitofp turn into +0.0 for zero.
3381  if (isa<SIToFPInst>(Op) || isa<UIToFPInst>(Op))
3382  return true;
3383 
3384  if (auto *Call = dyn_cast<CallInst>(Op)) {
3385  Intrinsic::ID IID = getIntrinsicForCallSite(*Call, TLI);
3386  switch (IID) {
3387  default:
3388  break;
3389  // sqrt(-0.0) = -0.0, no other negative results are possible.
3390  case Intrinsic::sqrt:
3391  case Intrinsic::canonicalize:
3392  return CannotBeNegativeZero(Call->getArgOperand(0), TLI, Depth + 1);
3393  // fabs(x) != -0.0
3394  case Intrinsic::fabs:
3395  return true;
3396  }
3397  }
3398 
3399  return false;
3400 }
3401 
3402 /// If \p SignBitOnly is true, test for a known 0 sign bit rather than a
3403 /// standard ordered compare. e.g. make -0.0 olt 0.0 be true because of the sign
3404 /// bit despite comparing equal.
3406  const TargetLibraryInfo *TLI,
3407  bool SignBitOnly,
3408  unsigned Depth) {
3409  // TODO: This function does not do the right thing when SignBitOnly is true
3410  // and we're lowering to a hypothetical IEEE 754-compliant-but-evil platform
3411  // which flips the sign bits of NaNs. See
3412  // https://llvm.org/bugs/show_bug.cgi?id=31702.
3413 
3414  if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
3415  return !CFP->getValueAPF().isNegative() ||
3416  (!SignBitOnly && CFP->getValueAPF().isZero());
3417  }
3418 
3419  // Handle vector of constants.
3420  if (auto *CV = dyn_cast<Constant>(V)) {
3421  if (auto *CVFVTy = dyn_cast<FixedVectorType>(CV->getType())) {
3422  unsigned NumElts = CVFVTy->getNumElements();
3423  for (unsigned i = 0; i != NumElts; ++i) {
3424  auto *CFP = dyn_cast_or_null<ConstantFP>(CV->getAggregateElement(i));
3425  if (!CFP)
3426  return false;
3427  if (CFP->getValueAPF().isNegative() &&
3428  (SignBitOnly || !CFP->getValueAPF().isZero()))
3429  return false;
3430  }
3431 
3432  // All non-negative ConstantFPs.
3433  return true;
3434  }
3435  }
3436 
3438  return false;
3439 
3440  const Operator *I = dyn_cast<Operator>(V);
3441  if (!I)
3442  return false;
3443 
3444  switch (I->getOpcode()) {
3445  default:
3446  break;
3447  // Unsigned integers are always nonnegative.
3448  case Instruction::UIToFP:
3449  return true;
3450  case Instruction::FMul:
3451  case Instruction::FDiv:
3452  // X * X is always non-negative or a NaN.
3453  // X / X is always exactly 1.0 or a NaN.
3454  if (I->getOperand(0) == I->getOperand(1) &&
3455  (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()))
3456  return true;
3457 
3459  case Instruction::FAdd:
3460  case Instruction::FRem:
3461  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3462  Depth + 1) &&
3463  cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
3464  Depth + 1);
3465  case Instruction::Select:
3466  return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
3467  Depth + 1) &&
3468  cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
3469  Depth + 1);
3470  case Instruction::FPExt:
3471  case Instruction::FPTrunc:
3472  // Widening/narrowing never change sign.
3473  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3474  Depth + 1);
3475  case Instruction::ExtractElement:
3476  // Look through extract element. At the moment we keep this simple and skip
3477  // tracking the specific element. But at least we might find information
3478  // valid for all elements of the vector.
3479  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3480  Depth + 1);
3481  case Instruction::Call:
3482  const auto *CI = cast<CallInst>(I);
3483  Intrinsic::ID IID = getIntrinsicForCallSite(*CI, TLI);
3484  switch (IID) {
3485  default:
3486  break;
3487  case Intrinsic::maxnum: {
3488  Value *V0 = I->getOperand(0), *V1 = I->getOperand(1);
3489  auto isPositiveNum = [&](Value *V) {
3490  if (SignBitOnly) {
3491  // With SignBitOnly, this is tricky because the result of
3492  // maxnum(+0.0, -0.0) is unspecified. Just check if the operand is
3493  // a constant strictly greater than 0.0.
3494  const APFloat *C;
3495  return match(V, m_APFloat(C)) &&
3496  *C > APFloat::getZero(C->getSemantics());
3497  }
3498 
3499  // -0.0 compares equal to 0.0, so if this operand is at least -0.0,
3500  // maxnum can't be ordered-less-than-zero.
3501  return isKnownNeverNaN(V, TLI) &&
3502  cannotBeOrderedLessThanZeroImpl(V, TLI, false, Depth + 1);
3503  };
3504 
3505  // TODO: This could be improved. We could also check that neither operand
3506  // has its sign bit set (and at least 1 is not-NAN?).
3507  return isPositiveNum(V0) || isPositiveNum(V1);
3508  }
3509 
3510  case Intrinsic::maximum:
3511  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3512  Depth + 1) ||
3513  cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
3514  Depth + 1);
3515  case Intrinsic::minnum:
3516  case Intrinsic::minimum:
3517  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3518  Depth + 1) &&
3519  cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
3520  Depth + 1);
3521  case Intrinsic::exp:
3522  case Intrinsic::exp2:
3523  case Intrinsic::fabs:
3524  return true;
3525 
3526  case Intrinsic::sqrt:
3527  // sqrt(x) is always >= -0 or NaN. Moreover, sqrt(x) == -0 iff x == -0.
3528  if (!SignBitOnly)
3529  return true;
3530  return CI->hasNoNaNs() && (CI->hasNoSignedZeros() ||
3531  CannotBeNegativeZero(CI->getOperand(0), TLI));
3532 
3533  case Intrinsic::powi:
3534  if (ConstantInt *Exponent = dyn_cast<ConstantInt>(I->getOperand(1))) {
3535  // powi(x,n) is non-negative if n is even.
3536  if (Exponent->getBitWidth() <= 64 && Exponent->getSExtValue() % 2u == 0)
3537  return true;
3538  }
3539  // TODO: This is not correct. Given that exp is an integer, here are the
3540  // ways that pow can return a negative value:
3541  //
3542  // pow(x, exp) --> negative if exp is odd and x is negative.
3543  // pow(-0, exp) --> -inf if exp is negative odd.
3544  // pow(-0, exp) --> -0 if exp is positive odd.
3545  // pow(-inf, exp) --> -0 if exp is negative odd.
3546  // pow(-inf, exp) --> -inf if exp is positive odd.
3547  //
3548  // Therefore, if !SignBitOnly, we can return true if x >= +0 or x is NaN,
3549  // but we must return false if x == -0. Unfortunately we do not currently
3550  // have a way of expressing this constraint. See details in
3551  // https://llvm.org/bugs/show_bug.cgi?id=31702.
3552  return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3553  Depth + 1);
3554 
3555  case Intrinsic::fma:
3556  case Intrinsic::fmuladd:
3557  // x*x+y is non-negative if y is non-negative.
3558  return I->getOperand(0) == I->getOperand(1) &&
3559  (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()) &&
3560  cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
3561  Depth + 1);
3562  }
3563  break;
3564  }
3565  return false;
3566 }
3567 
3569  const TargetLibraryInfo *TLI) {
3570  return cannotBeOrderedLessThanZeroImpl(V, TLI, false, 0);
3571 }
3572 
3574  return cannotBeOrderedLessThanZeroImpl(V, TLI, true, 0);
3575 }
3576 
3578  unsigned Depth) {
3579  assert(V->getType()->isFPOrFPVectorTy() && "Querying for Inf on non-FP type");
3580 
3581  // If we're told that infinities won't happen, assume they won't.
3582  if (auto *FPMathOp = dyn_cast<FPMathOperator>(V))
3583  if (FPMathOp->hasNoInfs())
3584  return true;
3585 
3586  // Handle scalar constants.
3587  if (auto *CFP = dyn_cast<ConstantFP>(V))
3588  return !CFP->isInfinity();
3589 
3591  return false;
3592 
3593  if (auto *Inst = dyn_cast<Instruction>(V)) {
3594  switch (Inst->getOpcode()) {
3595  case Instruction::Select: {
3596  return isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1) &&
3597  isKnownNeverInfinity(Inst->getOperand(2), TLI, Depth + 1);
3598  }
3599  case Instruction::SIToFP:
3600  case Instruction::UIToFP: {
3601  // Get width of largest magnitude integer (remove a bit if signed).
3602  // This still works for a signed minimum value because the largest FP
3603  // value is scaled by some fraction close to 2.0 (1.0 + 0.xxxx).
3604  int IntSize = Inst->getOperand(0)->getType()->getScalarSizeInBits();
3605  if (Inst->getOpcode() == Instruction::SIToFP)
3606  --IntSize;
3607 
3608  // If the exponent of the largest finite FP value can hold the largest
3609  // integer, the result of the cast must be finite.
3610  Type *FPTy = Inst->getType()->getScalarType();
3611  return ilogb(APFloat::getLargest(FPTy->getFltSemantics())) >= IntSize;
3612  }
3613  default:
3614  break;
3615  }
3616  }
3617 
3618  // try to handle fixed width vector constants
3619  auto *VFVTy = dyn_cast<FixedVectorType>(V->getType());
3620  if (VFVTy && isa<Constant>(V)) {
3621  // For vectors, verify that each element is not infinity.
3622  unsigned NumElts = VFVTy->getNumElements();
3623  for (unsigned i = 0; i != NumElts; ++i) {
3624  Constant *Elt = cast<Constant>(V)->getAggregateElement(i);
3625  if (!Elt)
3626  return false;
3627  if (isa<UndefValue>(Elt))
3628  continue;
3629  auto *CElt = dyn_cast<ConstantFP>(Elt);
3630  if (!CElt || CElt->isInfinity())
3631  return false;
3632  }
3633  // All elements were confirmed non-infinity or undefined.
3634  return true;
3635  }
3636 
3637  // was not able to prove that V never contains infinity
3638  return false;
3639 }
3640 
3642  unsigned Depth) {
3643  assert(V->getType()->isFPOrFPVectorTy() && "Querying for NaN on non-FP type");
3644 
3645  // If we're told that NaNs won't happen, assume they won't.
3646  if (auto *FPMathOp = dyn_cast<FPMathOperator>(V))
3647  if (FPMathOp->hasNoNaNs())
3648  return true;
3649 
3650  // Handle scalar constants.
3651  if (auto *CFP = dyn_cast<ConstantFP>(V))
3652  return !CFP->isNaN();
3653 
3655  return false;
3656 
3657  if (auto *Inst = dyn_cast<Instruction>(V)) {
3658  switch (Inst->getOpcode()) {
3659  case Instruction::FAdd:
3660  case Instruction::FSub:
3661  // Adding positive and negative infinity produces NaN.
3662  return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1) &&
3663  isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) &&
3664  (isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) ||
3665  isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1));
3666 
3667  case Instruction::FMul:
3668  // Zero multiplied with infinity produces NaN.
3669  // FIXME: If neither side can be zero fmul never produces NaN.
3670  return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1) &&
3671  isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) &&
3672  isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) &&
3673  isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1);
3674 
3675  case Instruction::FDiv:
3676  case Instruction::FRem:
3677  // FIXME: Only 0/0, Inf/Inf, Inf REM x and x REM 0 produce NaN.
3678  return false;
3679 
3680  case Instruction::Select: {
3681  return isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) &&
3682  isKnownNeverNaN(Inst->getOperand(2), TLI, Depth + 1);
3683  }
3684  case Instruction::SIToFP:
3685  case Instruction::UIToFP:
3686  return true;
3687  case Instruction::FPTrunc:
3688  case Instruction::FPExt:
3689  return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1);
3690  default:
3691  break;
3692  }
3693  }
3694 
3695  if (const auto *II = dyn_cast<IntrinsicInst>(V)) {
3696  switch (II->getIntrinsicID()) {
3697  case Intrinsic::canonicalize:
3698  case Intrinsic::fabs:
3699  case Intrinsic::copysign:
3700  case Intrinsic::exp:
3701  case Intrinsic::exp2:
3702  case Intrinsic::floor:
3703  case Intrinsic::ceil:
3704  case Intrinsic::trunc:
3705  case Intrinsic::rint:
3706  case Intrinsic::nearbyint:
3707  case Intrinsic::round:
3708  case Intrinsic::roundeven:
3709  return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1);
3710  case Intrinsic::sqrt:
3711  return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) &&
3712  CannotBeOrderedLessThanZero(II->getArgOperand(0), TLI);
3713  case Intrinsic::minnum:
3714  case Intrinsic::maxnum:
3715  // If either operand is not NaN, the result is not NaN.
3716  return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) ||
3717  isKnownNeverNaN(II->getArgOperand(1), TLI, Depth + 1);
3718  default:
3719  return false;
3720  }
3721  }
3722 
3723  // Try to handle fixed width vector constants
3724  auto *VFVTy = dyn_cast<FixedVectorType>(V->getType());
3725  if (VFVTy && isa<Constant>(V)) {
3726  // For vectors, verify that each element is not NaN.
3727  unsigned NumElts = VFVTy->getNumElements();
3728  for (unsigned i = 0; i != NumElts; ++i) {
3729  Constant *Elt = cast<Constant>(V)->getAggregateElement(i);
3730  if (!Elt)
3731  return false;
3732  if (isa<UndefValue>(Elt))
3733  continue;
3734  auto *CElt = dyn_cast<ConstantFP>(Elt);
3735  if (!CElt || CElt->isNaN())
3736  return false;
3737  }
3738  // All elements were confirmed not-NaN or undefined.
3739  return true;
3740  }
3741 
3742  // Was not able to prove that V never contains NaN
3743  return false;
3744 }
3745 
3747 
3748  // All byte-wide stores are splatable, even of arbitrary variables.
3749  if (V->getType()->isIntegerTy(8))
3750  return V;
3751 
3752  LLVMContext &Ctx = V->getContext();
3753 
3754  // Undef don't care.
3755  auto *UndefInt8 = UndefValue::get(Type::getInt8Ty(Ctx));
3756  if (isa<UndefValue>(V))
3757  return UndefInt8;
3758 
3759  // Return Undef for zero-sized type.
3760  if (!DL.getTypeStoreSize(V->getType()).isNonZero())
3761  return UndefInt8;
3762 
3763  Constant *C = dyn_cast<Constant>(V);
3764  if (!C) {
3765  // Conceptually, we could handle things like:
3766  // %a = zext i8 %X to i16
3767  // %b = shl i16 %a, 8
3768  // %c = or i16 %a, %b
3769  // but until there is an example that actually needs this, it doesn't seem
3770  // worth worrying about.
3771  return nullptr;
3772  }
3773 
3774  // Handle 'null' ConstantArrayZero etc.
3775  if (C->isNullValue())
3777 
3778  // Constant floating-point values can be handled as integer values if the
3779  // corresponding integer value is "byteable". An important case is 0.0.
3780  if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
3781  Type *Ty = nullptr;
3782  if (CFP->getType()->isHalfTy())
3783  Ty = Type::getInt16Ty(Ctx);
3784  else if (CFP->getType()->isFloatTy())
3785  Ty = Type::getInt32Ty(Ctx);
3786  else if (CFP->getType()->isDoubleTy())
3787  Ty = Type::getInt64Ty(Ctx);
3788  // Don't handle long double formats, which have strange constraints.
3789  return Ty ? isBytewiseValue(ConstantExpr::getBitCast(CFP, Ty), DL)
3790  : nullptr;
3791  }
3792 
3793  // We can handle constant integers that are multiple of 8 bits.
3794  if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
3795  if (CI->getBitWidth() % 8 == 0) {
3796  assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
3797  if (!CI->getValue().isSplat(8))
3798  return nullptr;
3799  return ConstantInt::get(Ctx, CI->getValue().trunc(8));
3800  }
3801  }
3802 
3803  if (auto *CE = dyn_cast<ConstantExpr>(C)) {
3804  if (CE->getOpcode() == Instruction::IntToPtr) {
3805  if (auto *PtrTy = dyn_cast<PointerType>(CE->getType())) {
3806  unsigned BitWidth = DL.getPointerSizeInBits(PtrTy->getAddressSpace());
3807  return isBytewiseValue(
3808  ConstantExpr::getIntegerCast(CE->getOperand(0),
3809  Type::getIntNTy(Ctx, BitWidth), false),
3810  DL);
3811  }
3812  }
3813  }
3814 
3815  auto Merge = [&](Value *LHS, Value *RHS) -> Value * {
3816  if (LHS == RHS)
3817  return LHS;
3818  if (!LHS || !RHS)
3819  return nullptr;
3820  if (LHS == UndefInt8)
3821  return RHS;
3822  if (RHS == UndefInt8)
3823  return LHS;
3824  return nullptr;
3825  };
3826 
3827  if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(C)) {
3828  Value *Val = UndefInt8;
3829  for (unsigned I = 0, E = CA->getNumElements(); I != E; ++I)
3830  if (!(Val = Merge(Val, isBytewiseValue(CA->getElementAsConstant(I), DL))))
3831  return nullptr;
3832  return Val;
3833  }
3834 
3835  if (isa<ConstantAggregate>(C)) {
3836  Value *Val = UndefInt8;
3837  for (unsigned I = 0, E = C->getNumOperands(); I != E; ++I)
3838  if (!(Val = Merge(Val, isBytewiseValue(C->getOperand(I), DL))))
3839  return nullptr;
3840  return Val;
3841  }
3842 
3843  // Don't try to handle the handful of other constants.
3844  return nullptr;
3845 }
3846 
3847 // This is the recursive version of BuildSubAggregate. It takes a few different
3848 // arguments. Idxs is the index within the nested struct From that we are
3849 // looking at now (which is of type IndexedType). IdxSkip is the number of
3850 // indices from Idxs that should be left out when inserting into the resulting
3851 // struct. To is the result struct built so far, new insertvalue instructions
3852 // build on that.
3853 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
3855  unsigned IdxSkip,
3856  Instruction *InsertBefore) {
3857  StructType *STy = dyn_cast<StructType>(IndexedType);
3858  if (STy) {
3859  // Save the original To argument so we can modify it
3860  Value *OrigTo = To;
3861  // General case, the type indexed by Idxs is a struct
3862  for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
3863  // Process each struct element recursively
3864  Idxs.push_back(i);
3865  Value *PrevTo = To;
3866  To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
3867  InsertBefore);
3868  Idxs.pop_back();
3869  if (!To) {
3870  // Couldn't find any inserted value for this index? Cleanup
3871  while (PrevTo != OrigTo) {
3872  InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
3873  PrevTo = Del->getAggregateOperand();
3874  Del->eraseFromParent();
3875  }
3876  // Stop processing elements
3877  break;
3878  }
3879  }
3880  // If we successfully found a value for each of our subaggregates
3881  if (To)
3882  return To;
3883  }
3884  // Base case, the type indexed by SourceIdxs is not a struct, or not all of
3885  // the struct's elements had a value that was inserted directly. In the latter
3886  // case, perhaps we can't determine each of the subelements individually, but
3887  // we might be able to find the complete struct somewhere.
3888 
3889  // Find the value that is at that particular spot
3890  Value *V = FindInsertedValue(From, Idxs);
3891 
3892  if (!V)
3893  return nullptr;
3894 
3895  // Insert the value in the new (sub) aggregate
3896  return InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
3897  "tmp", InsertBefore);
3898 }
3899 
3900 // This helper takes a nested struct and extracts a part of it (which is again a
3901 // struct) into a new value. For example, given the struct:
3902 // { a, { b, { c, d }, e } }
3903 // and the indices "1, 1" this returns
3904 // { c, d }.
3905 //
3906 // It does this by inserting an insertvalue for each element in the resulting
3907 // struct, as opposed to just inserting a single struct. This will only work if
3908 // each of the elements of the substruct are known (ie, inserted into From by an
3909 // insertvalue instruction somewhere).
3910 //
3911 // All inserted insertvalue instructions are inserted before InsertBefore
3913  Instruction *InsertBefore) {
3914  assert(InsertBefore && "Must have someplace to insert!");
3915  Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
3916  idx_range);
3917  Value *To = UndefValue::get(IndexedType);
3918  SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
3919  unsigned IdxSkip = Idxs.size();
3920 
3921  return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
3922 }
3923 
3924 /// Given an aggregate and a sequence of indices, see if the scalar value
3925 /// indexed is already around as a register, for example if it was inserted
3926 /// directly into the aggregate.
3927 ///
3928 /// If InsertBefore is not null, this function will duplicate (modified)
3929 /// insertvalues when a part of a nested struct is extracted.
3931  Instruction *InsertBefore) {
3932  // Nothing to index? Just return V then (this is useful at the end of our
3933  // recursion).
3934  if (idx_range.empty())
3935  return V;
3936  // We have indices, so V should have an indexable type.
3937  assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
3938  "Not looking at a struct or array?");
3939  assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
3940  "Invalid indices for type?");
3941 
3942  if (Constant *C = dyn_cast<Constant>(V)) {
3943  C = C->getAggregateElement(idx_range[0]);
3944  if (!C) return nullptr;
3945  return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
3946  }
3947 
3948  if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
3949  // Loop the indices for the insertvalue instruction in parallel with the
3950  // requested indices
3951  const unsigned *req_idx = idx_range.begin();
3952  for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
3953  i != e; ++i, ++req_idx) {
3954  if (req_idx == idx_range.end()) {
3955  // We can't handle this without inserting insertvalues
3956  if (!InsertBefore)
3957  return nullptr;
3958 
3959  // The requested index identifies a part of a nested aggregate. Handle
3960  // this specially. For example,
3961  // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
3962  // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
3963  // %C = extractvalue {i32, { i32, i32 } } %B, 1
3964  // This can be changed into
3965  // %A = insertvalue {i32, i32 } undef, i32 10, 0
3966  // %C = insertvalue {i32, i32 } %A, i32 11, 1
3967  // which allows the unused 0,0 element from the nested struct to be
3968  // removed.
3969  return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
3970  InsertBefore);
3971  }
3972 
3973  // This insert value inserts something else than what we are looking for.
3974  // See if the (aggregate) value inserted into has the value we are
3975  // looking for, then.
3976  if (*req_idx != *i)
3977  return FindInsertedValue(I->getAggregateOperand(), idx_range,
3978  InsertBefore);
3979  }
3980  // If we end up here, the indices of the insertvalue match with those
3981  // requested (though possibly only partially). Now we recursively look at
3982  // the inserted value, passing any remaining indices.
3983  return FindInsertedValue(I->getInsertedValueOperand(),
3984  makeArrayRef(req_idx, idx_range.end()),
3985  InsertBefore);
3986  }
3987 
3988  if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
3989  // If we're extracting a value from an aggregate that was extracted from
3990  // something else, we can extract from that something else directly instead.
3991  // However, we will need to chain I's indices with the requested indices.
3992 
3993  // Calculate the number of indices required
3994  unsigned size = I->getNumIndices() + idx_range.size();
3995  // Allocate some space to put the new indices in
3997  Idxs.reserve(size);
3998  // Add indices from the extract value instruction
3999  Idxs.append(I->idx_begin(), I->idx_end());
4000 
4001  // Add requested indices
4002  Idxs.append(idx_range.begin(), idx_range.end());
4003 
4004  assert(Idxs.size() == size
4005  && "Number of indices added not correct?");
4006 
4007  return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
4008  }
4009  // Otherwise, we don't know (such as, extracting from a function return value
4010  // or load instruction)
4011  return nullptr;
4012 }
4013 
4015  unsigned CharSize) {
4016  // Make sure the GEP has exactly three arguments.
4017  if (GEP->getNumOperands() != 3)
4018  return false;
4019 
4020  // Make sure the index-ee is a pointer to array of \p CharSize integers.
4021  // CharSize.
4022  ArrayType *AT = dyn_cast<ArrayType>(GEP->getSourceElementType());
4023  if (!AT || !AT->getElementType()->isIntegerTy(CharSize))
4024  return false;
4025 
4026  // Check to make sure that the first operand of the GEP is an integer and
4027  // has value 0 so that we are sure we're indexing into the initializer.
4028  const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
4029  if (!FirstIdx || !FirstIdx->isZero())
4030  return false;
4031 
4032  return true;
4033 }
4034 
4036  ConstantDataArraySlice &Slice,
4037  unsigned ElementSize, uint64_t Offset) {
4038  assert(V);
4039 
4040  // Look through bitcast instructions and geps.
4041  V = V->stripPointerCasts();
4042 
4043  // If the value is a GEP instruction or constant expression, treat it as an
4044  // offset.
4045  if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
4046  // The GEP operator should be based on a pointer to string constant, and is
4047  // indexing into the string constant.
4048  if (!isGEPBasedOnPointerToString(GEP, ElementSize))
4049  return false;
4050 
4051  // If the second index isn't a ConstantInt, then this is a variable index
4052  // into the array. If this occurs, we can't say anything meaningful about
4053  // the string.
4054  uint64_t StartIdx = 0;
4055  if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
4056  StartIdx = CI->getZExtValue();
4057  else
4058  return false;
4059  return getConstantDataArrayInfo(GEP->getOperand(0), Slice, ElementSize,
4060  StartIdx + Offset);
4061  }
4062 
4063  // The GEP instruction, constant or instruction, must reference a global
4064  // variable that is a constant and is initialized. The referenced constant
4065  // initializer is the array that we'll use for optimization.
4066  const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
4067  if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
4068  return false;
4069 
4070  const ConstantDataArray *Array;
4071  ArrayType *ArrayTy;
4072  if (GV->getInitializer()->isNullValue()) {
4073  Type *GVTy = GV->getValueType();
4074  if ( (ArrayTy = dyn_cast<ArrayType>(GVTy)) ) {
4075  // A zeroinitializer for the array; there is no ConstantDataArray.
4076  Array = nullptr;
4077  } else {
4078  const DataLayout &DL = GV->getParent()->getDataLayout();
4079  uint64_t SizeInBytes = DL.getTypeStoreSize(GVTy).getFixedSize();
4080  uint64_t Length = SizeInBytes / (ElementSize / 8);
4081  if (Length <= Offset)
4082  return false;
4083 
4084  Slice.Array = nullptr;
4085  Slice.Offset = 0;
4086  Slice.Length = Length - Offset;
4087  return true;
4088  }
4089  } else {
4090  // This must be a ConstantDataArray.
4091  Array = dyn_cast<ConstantDataArray>(GV->getInitializer());
4092  if (!Array)
4093  return false;
4094  ArrayTy = Array->getType();
4095  }
4096  if (!ArrayTy->getElementType()->isIntegerTy(ElementSize))
4097  return false;
4098 
4099  uint64_t NumElts = ArrayTy->getArrayNumElements();
4100  if (Offset > NumElts)
4101  return false;
4102 
4103  Slice.Array = Array;
4104  Slice.Offset = Offset;
4105  Slice.Length = NumElts - Offset;
4106  return true;
4107 }
4108 
4109 /// This function computes the length of a null-terminated C string pointed to
4110 /// by V. If successful, it returns true and returns the string in Str.
4111 /// If unsuccessful, it returns false.
4113  uint64_t Offset, bool TrimAtNul) {
4114  ConstantDataArraySlice Slice;
4115  if (!getConstantDataArrayInfo(V, Slice, 8, Offset))
4116  return false;
4117 
4118  if (Slice.Array == nullptr) {
4119  if (TrimAtNul) {
4120  Str = StringRef();
4121  return true;
4122  }
4123  if (Slice.Length == 1) {
4124  Str = StringRef("", 1);
4125  return true;
4126  }
4127  // We cannot instantiate a StringRef as we do not have an appropriate string
4128  // of 0s at hand.
4129  return false;
4130  }
4131 
4132  // Start out with the entire array in the StringRef.
4133  Str = Slice.Array->getAsString();
4134  // Skip over 'offset' bytes.
4135  Str = Str.substr(Slice.Offset);
4136 
4137  if (TrimAtNul) {
4138  // Trim off the \0 and anything after it. If the array is not nul
4139  // terminated, we just return the whole end of string. The client may know
4140  // some other way that the string is length-bound.
4141  Str = Str.substr(0, Str.find('\0'));
4142  }
4143  return true;
4144 }
4145 
4146 // These next two are very similar to the above, but also look through PHI
4147 // nodes.
4148 // TODO: See if we can integrate these two together.
4149 
4150 /// If we can compute the length of the string pointed to by
4151 /// the specified pointer, return 'len+1'. If we can't, return 0.
4154  unsigned CharSize) {
4155  // Look through noop bitcast instructions.
4156  V = V->stripPointerCasts();
4157 
4158  // If this is a PHI node, there are two cases: either we have already seen it
4159  // or we haven't.
4160  if (const PHINode *PN = dyn_cast<PHINode>(V)) {
4161  if (!PHIs.insert(PN).second)
4162  return ~0ULL; // already in the set.
4163 
4164  // If it was new, see if all the input strings are the same length.
4165  uint64_t LenSoFar = ~0ULL;
4166  for (Value *IncValue : PN->incoming_values()) {
4167  uint64_t Len = GetStringLengthH(IncValue, PHIs, CharSize);
4168  if (Len == 0) return 0; // Unknown length -> unknown.
4169 
4170  if (Len == ~0ULL) continue;
4171 
4172  if (Len != LenSoFar && LenSoFar != ~0ULL)
4173  return 0; // Disagree -> unknown.
4174  LenSoFar = Len;
4175  }
4176 
4177  // Success, all agree.
4178  return LenSoFar;
4179  }
4180 
4181  // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
4182  if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
4183  uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs, CharSize);
4184  if (Len1 == 0) return 0;
4185  uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs, CharSize);
4186  if (Len2 == 0) return 0;
4187  if (Len1 == ~0ULL) return Len2;
4188  if (Len2 == ~0ULL) return Len1;
4189  if (Len1 != Len2) return 0;
4190  return Len1;
4191  }
4192 
4193  // Otherwise, see if we can read the string.
4194  ConstantDataArraySlice Slice;
4195  if (!getConstantDataArrayInfo(V, Slice, CharSize))
4196  return 0;
4197 
4198  if (Slice.Array == nullptr)
4199  return 1;
4200 
4201  // Search for nul characters
4202  unsigned NullIndex = 0;
4203  for (unsigned E = Slice.Length; NullIndex < E; ++NullIndex) {
4204  if (Slice.Array->getElementAsInteger(Slice.Offset + NullIndex) == 0)
4205  break;
4206  }
4207 
4208  return NullIndex + 1;
4209 }
4210 
4211 /// If we can compute the length of the string pointed to by
4212 /// the specified pointer, return 'len+1'. If we can't, return 0.
4213 uint64_t llvm::GetStringLength(const Value *V, unsigned CharSize) {
4214  if (!V->getType()->isPointerTy())
4215  return 0;
4216 
4218  uint64_t Len = GetStringLengthH(V, PHIs, CharSize);
4219  // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
4220  // an empty string as a length.
4221  return Len == ~0ULL ? 1 : Len;
4222 }
4223 
4224 const Value *
4226  bool MustPreserveNullness) {
4227  assert(Call &&
4228  "getArgumentAliasingToReturnedPointer only works on nonnull calls");
4229  if (const Value *RV = Call->getReturnedArgOperand())
4230  return RV;
4231  // This can be used only as a aliasing property.
4233  Call, MustPreserveNullness))
4234  return Call->getArgOperand(0);
4235  return nullptr;
4236 }
4237 
4239  const CallBase *Call, bool MustPreserveNullness) {
4240  switch (Call->getIntrinsicID()) {
4241  case Intrinsic::launder_invariant_group:
4242  case Intrinsic::strip_invariant_group:
4243  case Intrinsic::aarch64_irg:
4244  case Intrinsic::aarch64_tagp:
4245  return true;
4246  case Intrinsic::ptrmask:
4247  return !MustPreserveNullness;
4248  default:
4249  return false;
4250  }
4251 }
4252 
4253 /// \p PN defines a loop-variant pointer to an object. Check if the
4254 /// previous iteration of the loop was referring to the same object as \p PN.
4256  const LoopInfo *LI) {
4257  // Find the loop-defined value.
4258  Loop *L = LI->getLoopFor(PN->getParent());
4259  if (PN->getNumIncomingValues() != 2)
4260  return true;
4261 
4262  // Find the value from previous iteration.
4263  auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
4264  if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
4265  PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
4266  if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
4267  return true;
4268 
4269  // If a new pointer is loaded in the loop, the pointer references a different
4270  // object in every iteration. E.g.:
4271  // for (i)
4272  // int *p = a[i];
4273  // ...
4274  if (auto *Load = dyn_cast<LoadInst>(PrevValue))
4275  if (!L->isLoopInvariant(Load->getPointerOperand()))
4276  return false;
4277  return true;
4278 }
4279 
4280 const Value *llvm::getUnderlyingObject(const Value *V, unsigned MaxLookup) {
4281  if (!V->getType()->isPointerTy())
4282  return V;
4283  for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
4284  if (auto *GEP = dyn_cast<GEPOperator>(V)) {
4285  V = GEP->getPointerOperand();
4286  } else if (Operator::getOpcode(V) == Instruction::BitCast ||
4287  Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
4288  V = cast<Operator>(V)->getOperand(0);
4289  if (!V->getType()->isPointerTy())
4290  return V;
4291  } else if (auto *GA = dyn_cast<GlobalAlias>(V)) {
4292  if (GA->isInterposable())
4293  return V;
4294  V = GA->getAliasee();
4295  } else {
4296  if (auto *PHI = dyn_cast<PHINode>(V)) {
4297  // Look through single-arg phi nodes created by LCSSA.
4298  if (PHI->getNumIncomingValues() == 1) {
4299  V = PHI->getIncomingValue(0);
4300  continue;
4301  }
4302  } else if (auto *Call = dyn_cast<CallBase>(V)) {
4303  // CaptureTracking can know about special capturing properties of some
4304  // intrinsics like launder.invariant.group, that can't be expressed with
4305  // the attributes, but have properties like returning aliasing pointer.
4306  // Because some analysis may assume that nocaptured pointer is not
4307  // returned from some special intrinsic (because function would have to
4308  // be marked with returns attribute), it is crucial to use this function
4309  // because it should be in sync with CaptureTracking. Not using it may
4310  // cause weird miscompilations where 2 aliasing pointers are assumed to
4311  // noalias.
4312  if (auto *RP = getArgumentAliasingToReturnedPointer(Call, false)) {
4313  V = RP;
4314  continue;
4315  }
4316  }
4317 
4318  return V;
4319  }
4320  assert(V->getType()->isPointerTy() && "Unexpected operand type!");
4321  }
4322  return V;
4323 }
4324 
4327  LoopInfo *LI, unsigned MaxLookup) {
4330  Worklist.push_back(V);
4331  do {
4332  const Value *P = Worklist.pop_back_val();
4333  P = getUnderlyingObject(P, MaxLookup);
4334 
4335  if (!Visited.insert(P).second)
4336  continue;
4337 
4338  if (auto *SI = dyn_cast<SelectInst>(P)) {
4339  Worklist.push_back(SI->getTrueValue());
4340  Worklist.push_back(SI->getFalseValue());
4341  continue;
4342  }
4343 
4344  if (auto *PN = dyn_cast<PHINode>(P)) {
4345  // If this PHI changes the underlying object in every iteration of the
4346  // loop, don't look through it. Consider:
4347  // int **A;
4348  // for (i) {
4349  // Prev = Curr; // Prev = PHI (Prev_0, Curr)
4350  // Curr = A[i];
4351  // *Prev, *Curr;
4352  //
4353  // Prev is tracking Curr one iteration behind so they refer to different
4354  // underlying objects.
4355  if (!LI || !LI->isLoopHeader(PN->getParent()) ||
4357  append_range(Worklist, PN->incoming_values());
4358  continue;
4359  }
4360 
4361  Objects.push_back(P);
4362  } while (!Worklist.empty());
4363 }
4364 
4365 /// This is the function that does the work of looking through basic
4366 /// ptrtoint+arithmetic+inttoptr sequences.
4367 static const Value *getUnderlyingObjectFromInt(const Value *V) {
4368  do {
4369  if (const Operator *U = dyn_cast<Operator>(V)) {
4370  // If we find a ptrtoint, we can transfer control back to the
4371  // regular getUnderlyingObjectFromInt.
4372  if (U->getOpcode() == Instruction::PtrToInt)
4373  return U->getOperand(0);
4374  // If we find an add of a constant, a multiplied value, or a phi, it's
4375  // likely that the other operand will lead us to the base
4376  // object. We don't have to worry about the case where the
4377  // object address is somehow being computed by the multiply,
4378  // because our callers only care when the result is an
4379  // identifiable object.
4380  if (U->getOpcode() != Instruction::Add ||
4381  (!isa<ConstantInt>(U->getOperand(1)) &&
4382  Operator::getOpcode(U->getOperand(1)) != Instruction::Mul &&
4383  !isa<PHINode>(U->getOperand(1))))
4384  return V;
4385  V = U->getOperand(0);
4386  } else {
4387  return V;
4388  }
4389  assert(V->getType()->isIntegerTy() && "Unexpected operand type!");
4390  } while (true);
4391 }
4392 
4393 /// This is a wrapper around getUnderlyingObjects and adds support for basic
4394 /// ptrtoint+arithmetic+inttoptr sequences.
4395 /// It returns false if unidentified object is found in getUnderlyingObjects.
4397  SmallVectorImpl<Value *> &Objects) {
4399  SmallVector<const Value *, 4> Working(1, V);
4400  do {
4401  V = Working.pop_back_val();
4402 
4404  getUnderlyingObjects(V, Objs);
4405 
4406  for (const Value *V : Objs) {
4407  if (!Visited.insert(V).second)
4408  continue;
4409  if (Operator::getOpcode(V) == Instruction::IntToPtr) {
4410  const Value *O =
4411  getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0));
4412  if (O->getType()->isPointerTy()) {
4413  Working.push_back(O);
4414  continue;
4415  }
4416  }
4417  // If getUnderlyingObjects fails to find an identifiable object,
4418  // getUnderlyingObjectsForCodeGen also fails for safety.
4419  if (!isIdentifiedObject(V)) {
4420  Objects.clear();
4421  return false;
4422  }
4423  Objects.push_back(const_cast<Value *>(V));
4424  }
4425  } while (!Working.empty());
4426  return true;
4427 }
4428 
4430  AllocaInst *Result = nullptr;
4431  SmallPtrSet<Value *, 4> Visited;
4432  SmallVector<Value *, 4> Worklist;
4433 
4434  auto AddWork = [&](Value *V) {
4435  if (Visited.insert(V).second)
4436  Worklist.push_back(V);
4437  };
4438 
4439  AddWork(V);
4440  do {
4441  V = Worklist.pop_back_val();
4442  assert(Visited.count(V));
4443 
4444  if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
4445  if (Result && Result != AI)
4446  return nullptr;
4447  Result = AI;
4448  } else if (CastInst *CI = dyn_cast<CastInst>(V)) {
4449  AddWork(CI->getOperand(0));
4450  } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
4451  for (Value *IncValue : PN->incoming_values())
4452  AddWork(IncValue);
4453  } else if (auto *SI = dyn_cast<SelectInst>(V)) {
4454  AddWork(SI->getTrueValue());
4455  AddWork(SI->getFalseValue());
4456  } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
4457  if (OffsetZero && !GEP->hasAllZeroIndices())
4458  return nullptr;
4459  AddWork(GEP->getPointerOperand());
4460  } else if (CallBase *CB = dyn_cast<CallBase>(V)) {
4461  Value *Returned = CB->getReturnedArgOperand();
4462  if (Returned)
4463  AddWork(Returned);
4464  else
4465  return nullptr;
4466  } else {
4467  return nullptr;
4468  }
4469  } while (!Worklist.empty());
4470 
4471  return Result;
4472 }
4473 
4475  const Value *V, bool AllowLifetime, bool AllowDroppable) {
4476  for (const User *U : V->users()) {
4477  const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
4478  if (!II)
4479  return false;
4480 
4481  if (AllowLifetime && II->isLifetimeStartOrEnd())
4482  continue;
4483 
4484  if (AllowDroppable && II->isDroppable())
4485  continue;
4486 
4487  return false;
4488  }
4489  return true;
4490 }
4491 
4494  V, /* AllowLifetime */ true, /* AllowDroppable */ false);
4495 }
4498  V, /* AllowLifetime */ true, /* AllowDroppable */ true);
4499 }
4500 
4502  if (!LI.isUnordered())
4503  return true;
4504  const Function &F = *LI.getFunction();
4505  // Speculative load may create a race that did not exist in the source.
4506  return F.hasFnAttribute(Attribute::SanitizeThread) ||
4507  // Speculative load may load data from dirty regions.
4508  F.hasFnAttribute(Attribute::SanitizeAddress) ||
4509  F.hasFnAttribute(Attribute::SanitizeHWAddress);
4510 }
4511 
4512 
4514  const Instruction *CtxI,
4515  const DominatorTree *DT,
4516  const TargetLibraryInfo *TLI) {
4517  const Operator *Inst = dyn_cast<Operator>(V);
4518  if (!Inst)
4519  return false;
4520 
4521  for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
4522  if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
4523  if (C->canTrap())
4524  return false;
4525 
4526  switch (Inst->getOpcode()) {
4527  default:
4528  return true;
4529  case Instruction::UDiv:
4530  case Instruction::URem: {
4531  // x / y is undefined if y == 0.
4532  const APInt *V;
4533  if (match(Inst->getOperand(1), m_APInt(V)))
4534  return *V != 0;
4535  return false;
4536  }
4537  case Instruction::SDiv:
4538  case Instruction::SRem: {
4539  // x / y is undefined if y == 0 or x == INT_MIN and y == -1
4540  const APInt *Numerator, *Denominator;
4541  if (!match(Inst->getOperand(1), m_APInt(Denominator)))
4542  return false;
4543  // We cannot hoist this division if the denominator is 0.
4544  if (*Denominator == 0)
4545  return false;
4546  // It's safe to hoist if the denominator is not 0 or -1.
4547  if (!Denominator->isAllOnes())
4548  return true;
4549  // At this point we know that the denominator is -1. It is safe to hoist as
4550  // long we know that the numerator is not INT_MIN.
4551  if (match(Inst->getOperand(0), m_APInt(Numerator)))
4552  return !Numerator->isMinSignedValue();
4553  // The numerator *might* be MinSignedValue.
4554  return false;
4555  }
4556  case Instruction::Load: {
4557  const LoadInst *LI = cast<LoadInst>(Inst);
4558  if (mustSuppressSpeculation(*LI))
4559  return false;
4560  const DataLayout &DL = LI->getModule()->getDataLayout();
4562  LI->getPointerOperand(), LI->getType(), MaybeAlign(LI->getAlign()), DL,
4563  CtxI, DT, TLI);
4564  }
4565  case Instruction::Call: {
4566  auto *CI = cast<const CallInst>(Inst);
4567  const Function *Callee = CI->getCalledFunction();
4568 
4569  // The called function could have undefined behavior or side-effects, even
4570  // if marked readnone nounwind.
4571  return Callee && Callee->isSpeculatable();
4572  }
4573  case Instruction::VAArg:
4574  case Instruction::Alloca:
4575  case Instruction::Invoke:
4576  case Instruction::CallBr:
4577  case Instruction::PHI:
4578  case Instruction::Store:
4579  case Instruction::Ret:
4580  case Instruction::Br:
4581  case Instruction::IndirectBr:
4582  case Instruction::Switch:
4583  case Instruction::Unreachable:
4584  case Instruction::Fence:
4585  case Instruction::AtomicRMW:
4586  case Instruction::AtomicCmpXchg:
4587  case Instruction::LandingPad:
4588  case Instruction::Resume:
4589  case Instruction::CatchSwitch:
4590  case Instruction::CatchPad:
4591  case Instruction::CatchRet:
4592  case Instruction::CleanupPad:
4593  case Instruction::CleanupRet:
4594  return false; // Misc instructions which have effects
4595  }
4596 }
4597 
4599  return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I);
4600 }
4601 
4602 /// Convert ConstantRange OverflowResult into ValueTracking OverflowResult.
4604  switch (OR) {
4613  }
4614  llvm_unreachable("Unknown OverflowResult");
4615 }
4616 
4617 /// Combine constant ranges from computeConstantRange() and computeKnownBits().
4619  const Value *V, bool ForSigned, const DataLayout &DL, unsigned Depth,
4620  AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT,
4621  OptimizationRemarkEmitter *ORE = nullptr, bool UseInstrInfo = true) {
4622  KnownBits Known = computeKnownBits(
4623  V, DL, Depth, AC, CxtI, DT, ORE, UseInstrInfo);
4624  ConstantRange CR1 = ConstantRange::fromKnownBits(Known, ForSigned);
4625  ConstantRange CR2 = computeConstantRange(V, UseInstrInfo);
4628  return CR1.intersectWith(CR2, RangeType);
4629 }
4630 
4632  const Value *LHS, const Value *RHS, const DataLayout &DL,
4633  AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT,
4634  bool UseInstrInfo) {
4635  KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT,
4636  nullptr, UseInstrInfo);
4637  KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT,
4638  nullptr, UseInstrInfo);
4639  ConstantRange LHSRange = ConstantRange::fromKnownBits(LHSKnown, false);
4640  ConstantRange RHSRange = ConstantRange::fromKnownBits(RHSKnown, false);
4641  return mapOverflowResult(LHSRange.unsignedMulMayOverflow(RHSRange));
4642 }
4643 
4646  const DataLayout &DL, AssumptionCache *AC,
4647  const Instruction *CxtI,
4648  const DominatorTree *DT, bool UseInstrInfo) {
4649  // Multiplying n * m significant bits yields a result of n + m significant
4650  // bits. If the total number of significant bits does not exceed the
4651  // result bit width (minus 1), there is no overflow.
4652  // This means if we have enough leading sign bits in the operands
4653  // we can guarantee that the result does not overflow.
4654  // Ref: "Hacker's Delight" by Henry Warren
4655  unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
4656 
4657  // Note that underestimating the number of sign bits gives a more
4658  // conservative answer.
4659  unsigned SignBits = ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) +
4660  ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT);
4661 
4662  // First handle the easy case: if we have enough sign bits there's
4663  // definitely no overflow.
4664  if (SignBits > BitWidth + 1)
4666 
4667  // There are two ambiguous cases where there can be no overflow:
4668  // SignBits == BitWidth + 1 and
4669  // SignBits == BitWidth
4670  // The second case is difficult to check, therefore we only handle the
4671  // first case.
4672  if (SignBits == BitWidth + 1) {
4673  // It overflows only when both arguments are negative and the true
4674  // product is exactly the minimum negative number.
4675  // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000
4676  // For simplicity we just check if at least one side is not negative.
4677  KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT,
4678  nullptr, UseInstrInfo);
4679  KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT,
4680  nullptr, UseInstrInfo);
4681  if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative())
4683  }
4685 }
4686 
4688  const Value *LHS, const Value *RHS, const DataLayout &DL,
4689  AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT,
4690  bool UseInstrInfo) {
4692  LHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT,
4693  nullptr, UseInstrInfo);
4695  RHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT,
4696  nullptr, UseInstrInfo);
4697  return mapOverflowResult(LHSRange.unsignedAddMayOverflow(RHSRange));
4698 }
4699 
4701  const Value *RHS,
4702  const AddOperator *Add,
4703  const DataLayout &DL,
4704  AssumptionCache *AC,
4705  const Instruction *CxtI,
4706  const DominatorTree *DT) {
4707  if (Add && Add->hasNoSignedWrap()) {
4709  }
4710 
4711  // If LHS and RHS each have at least two sign bits, the addition will look
4712  // like
4713  //
4714  // XX..... +
4715  // YY.....
4716  //
4717  // If the carry into the most significant position is 0, X and Y can't both
4718  // be 1 and therefore the carry out of the addition is also 0.
4719  //
4720  // If the carry into the most significant position is 1, X and Y can't both
4721  // be 0 and therefore the carry out of the addition is also 1.
4722  //
4723  // Since the carry into the most significant position is always equal to
4724  // the carry out of the addition, there is no signed overflow.
4725  if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 &&
4726  ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1)
4728 
4730  LHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT);
4732  RHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT);
4733  OverflowResult OR =
4734  mapOverflowResult(LHSRange.signedAddMayOverflow(RHSRange));
4736  return OR;
4737 
4738  // The remaining code needs Add to be available. Early returns if not so.
4739  if (!Add)
4741 
4742  // If the sign of Add is the same as at least one of the operands, this add
4743  // CANNOT overflow. If this can be determined from the known bits of the
4744  // operands the above signedAddMayOverflow() check will have already done so.
4745  // The only other way to improve on the known bits is from an assumption, so
4746  // call computeKnownBitsFromAssume() directly.
4747  bool LHSOrRHSKnownNonNegative =
4748  (LHSRange.isAllNonNegative() || RHSRange.isAllNonNegative());
4749  bool LHSOrRHSKnownNegative =
4750  (LHSRange.isAllNegative() || RHSRange.isAllNegative());
4751  if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
4752  KnownBits AddKnown(LHSRange.getBitWidth());
4754  Add, AddKnown, /*Depth=*/0, Query(DL, AC, CxtI, DT, true));
4755  if ((AddKnown.isNonNegative() && LHSOrRHSKnownNonNegative) ||
4756  (AddKnown.isNegative() && LHSOrRHSKnownNegative))
4758  }
4759 
4761 }
4762 
4764  const Value *RHS,
4765  const DataLayout &DL,
4766  AssumptionCache *AC,
4767  const Instruction *CxtI,
4768  const DominatorTree *DT) {
4769  // Checking for conditions implied by dominating conditions may be expensive.
4770  // Limit it to usub_with_overflow calls for now.
4771  if (match(CxtI,
4772  m_Intrinsic<Intrinsic::usub_with_overflow>(m_Value(), m_Value())))
4773  if (auto C =
4775  if (*C)
4778  }
4780  LHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT);
4782  RHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT);
4783  return mapOverflowResult(LHSRange.unsignedSubMayOverflow(RHSRange));
4784 }
4785 
4787  const Value *RHS,
4788  const DataLayout &DL,
4789  AssumptionCache *AC,
4790  const Instruction *CxtI,
4791  const DominatorTree *DT) {
4792  // If LHS and RHS each have at least two sign bits, the subtraction
4793  // cannot overflow.
4794  if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 &&
4795  ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1)
4797 
4799  LHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT);
4801  RHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT);
4802  return mapOverflowResult(LHSRange.signedSubMayOverflow(RHSRange));
4803 }
4804 
4806  const DominatorTree &DT) {
4807  SmallVector<const BranchInst *, 2> GuardingBranches;
4809 
4810  for (const User *U : WO->users()) {
4811  if (const auto *EVI = dyn_cast<ExtractValueInst>(U)) {
4812  assert(EVI->getNumIndices() == 1 && "Obvious from CI's type");
4813 
4814  if (EVI->getIndices()[0] == 0)
4815  Results.push_back(EVI);
4816  else {
4817  assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type");
4818 
4819  for (const auto *U : EVI->users())
4820  if (const auto *B = dyn_cast<BranchInst>(U)) {
4821  assert(B->isConditional() && "How else is it using an i1?");
4822  GuardingBranches.push_back(B);
4823  }
4824  }
4825  } else {
4826  // We are using the aggregate directly in a way we don't want to analyze
4827  // here (storing it to a global, say).
4828  return false;
4829  }
4830  }
4831 
4832  auto AllUsesGuardedByBranch = [&](const BranchInst *BI) {
4833  BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(1));
4834  if (!NoWrapEdge.isSingleEdge())
4835  return false;
4836 
4837  // Check if all users of the add are provably no-wrap.
4838  for (const auto *Result : Results) {
4839  // If the extractvalue itself is not executed on overflow, the we don't
4840  // need to check each use separately, since domination is transitive.
4841  if (DT.dominates(NoWrapEdge, Result->getParent()))
4842  continue;
4843 
4844  for (auto &RU : Result->uses())
4845  if (!DT.dominates(NoWrapEdge, RU))
4846  return false;
4847  }
4848 
4849  return true;
4850  };
4851 
4852  return llvm::any_of(GuardingBranches, AllUsesGuardedByBranch);
4853 }
4854 
4855 static bool canCreateUndefOrPoison(const Operator *Op, bool PoisonOnly,
4856  bool ConsiderFlags) {
4857 
4858  if (ConsiderFlags && Op->hasPoisonGeneratingFlags())
4859  return true;
4860 
4861  unsigned Opcode = Op->getOpcode();
4862 
4863  // Check whether opcode is a poison/undef-generating operation
4864  switch (Opcode) {
4865  case Instruction::Shl:
4866  case Instruction::AShr:
4867  case Instruction::LShr: {
4868  // Shifts return poison if shiftwidth is larger than the bitwidth.
4869  if (auto *C = dyn_cast<Constant>(Op->getOperand(1))) {
4870  SmallVector<Constant *, 4> ShiftAmounts;
4871  if (auto *FVTy = dyn_cast<FixedVectorType>(C->getType())) {
4872  unsigned NumElts = FVTy->getNumElements();
4873  for (unsigned i = 0; i < NumElts; ++i)
4874  ShiftAmounts.push_back(C->getAggregateElement(i));
4875  } else if (isa<ScalableVectorType>(C->getType()))
4876  return true; // Can't tell, just return true to be safe
4877  else
4878  ShiftAmounts.push_back(C);
4879 
4880  bool Safe = llvm::all_of(ShiftAmounts, [](Constant *C) {
4881  auto *CI = dyn_cast_or_null<ConstantInt>(C);
4882  return CI && CI->getValue().ult(C->getType()->getIntegerBitWidth());
4883  });
4884  return !Safe;
4885  }
4886  return true;
4887  }
4888  case Instruction::FPToSI:
4889  case Instruction::FPToUI:
4890  // fptosi/ui yields poison if the resulting value does not fit in the
4891  // destination type.
4892  return true;