LLVM  10.0.0svn
CaptureTracking.cpp
Go to the documentation of this file.
1 //===--- CaptureTracking.cpp - Determine whether a pointer is captured ----===//
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 determine which pointers are captured.
10 // A pointer value is captured if the function makes a copy of any part of the
11 // pointer that outlives the call. Not being captured means, more or less, that
12 // the pointer is only dereferenced and not stored in a global. Returning part
13 // of the pointer as the function return value may or may not count as capturing
14 // the pointer, depending on the context.
15 //
16 //===----------------------------------------------------------------------===//
17 
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/Analysis/CFG.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/Dominators.h"
27 #include "llvm/IR/Instructions.h"
28 #include "llvm/IR/IntrinsicInst.h"
29 
30 using namespace llvm;
31 
33 
34 bool CaptureTracker::shouldExplore(const Use *U) { return true; }
35 
37  // An inbounds GEP can either be a valid pointer (pointing into
38  // or to the end of an allocation), or be null in the default
39  // address space. So for an inbounds GEP there is no way to let
40  // the pointer escape using clever GEP hacking because doing so
41  // would make the pointer point outside of the allocated object
42  // and thus make the GEP result a poison value. Similarly, other
43  // dereferenceable pointers cannot be manipulated without producing
44  // poison.
45  if (auto *GEP = dyn_cast<GetElementPtrInst>(O))
46  if (GEP->isInBounds())
47  return true;
48  bool CanBeNull;
49  return O->getPointerDereferenceableBytes(DL, CanBeNull);
50 }
51 
52 namespace {
53  struct SimpleCaptureTracker : public CaptureTracker {
54  explicit SimpleCaptureTracker(bool ReturnCaptures)
55  : ReturnCaptures(ReturnCaptures), Captured(false) {}
56 
57  void tooManyUses() override { Captured = true; }
58 
59  bool captured(const Use *U) override {
60  if (isa<ReturnInst>(U->getUser()) && !ReturnCaptures)
61  return false;
62 
63  Captured = true;
64  return true;
65  }
66 
67  bool ReturnCaptures;
68 
69  bool Captured;
70  };
71 
72  /// Only find pointer captures which happen before the given instruction. Uses
73  /// the dominator tree to determine whether one instruction is before another.
74  /// Only support the case where the Value is defined in the same basic block
75  /// as the given instruction and the use.
76  struct CapturesBefore : public CaptureTracker {
77 
78  CapturesBefore(bool ReturnCaptures, const Instruction *I, const DominatorTree *DT,
79  bool IncludeI, OrderedBasicBlock *IC)
80  : OrderedBB(IC), BeforeHere(I), DT(DT),
81  ReturnCaptures(ReturnCaptures), IncludeI(IncludeI), Captured(false) {}
82 
83  void tooManyUses() override { Captured = true; }
84 
85  bool isSafeToPrune(Instruction *I) {
86  BasicBlock *BB = I->getParent();
87  // We explore this usage only if the usage can reach "BeforeHere".
88  // If use is not reachable from entry, there is no need to explore.
89  if (BeforeHere != I && !DT->isReachableFromEntry(BB))
90  return true;
91 
92  // Compute the case where both instructions are inside the same basic
93  // block. Since instructions in the same BB as BeforeHere are numbered in
94  // 'OrderedBB', avoid using 'dominates' and 'isPotentiallyReachable'
95  // which are very expensive for large basic blocks.
96  if (BB == BeforeHere->getParent()) {
97  // 'I' dominates 'BeforeHere' => not safe to prune.
98  //
99  // The value defined by an invoke dominates an instruction only
100  // if it dominates every instruction in UseBB. A PHI is dominated only
101  // if the instruction dominates every possible use in the UseBB. Since
102  // UseBB == BB, avoid pruning.
103  if (isa<InvokeInst>(BeforeHere) || isa<PHINode>(I) || I == BeforeHere)
104  return false;
105  if (!OrderedBB->dominates(BeforeHere, I))
106  return false;
107 
108  // 'BeforeHere' comes before 'I', it's safe to prune if we also
109  // guarantee that 'I' never reaches 'BeforeHere' through a back-edge or
110  // by its successors, i.e, prune if:
111  //
112  // (1) BB is an entry block or have no successors.
113  // (2) There's no path coming back through BB successors.
114  if (BB == &BB->getParent()->getEntryBlock() ||
116  return true;
117 
119  Worklist.append(succ_begin(BB), succ_end(BB));
120  return !isPotentiallyReachableFromMany(Worklist, BB, nullptr, DT);
121  }
122 
123  // If the value is defined in the same basic block as use and BeforeHere,
124  // there is no need to explore the use if BeforeHere dominates use.
125  // Check whether there is a path from I to BeforeHere.
126  if (BeforeHere != I && DT->dominates(BeforeHere, I) &&
127  !isPotentiallyReachable(I, BeforeHere, nullptr, DT))
128  return true;
129 
130  return false;
131  }
132 
133  bool shouldExplore(const Use *U) override {
134  Instruction *I = cast<Instruction>(U->getUser());
135 
136  if (BeforeHere == I && !IncludeI)
137  return false;
138 
139  if (isSafeToPrune(I))
140  return false;
141 
142  return true;
143  }
144 
145  bool captured(const Use *U) override {
146  if (isa<ReturnInst>(U->getUser()) && !ReturnCaptures)
147  return false;
148 
149  if (!shouldExplore(U))
150  return false;
151 
152  Captured = true;
153  return true;
154  }
155 
156  OrderedBasicBlock *OrderedBB;
157  const Instruction *BeforeHere;
158  const DominatorTree *DT;
159 
160  bool ReturnCaptures;
161  bool IncludeI;
162 
163  bool Captured;
164  };
165 }
166 
167 /// PointerMayBeCaptured - Return true if this pointer value may be captured
168 /// by the enclosing function (which is required to exist). This routine can
169 /// be expensive, so consider caching the results. The boolean ReturnCaptures
170 /// specifies whether returning the value (or part of it) from the function
171 /// counts as capturing it or not. The boolean StoreCaptures specified whether
172 /// storing the value (or part of it) into memory anywhere automatically
173 /// counts as capturing it or not.
175  bool ReturnCaptures, bool StoreCaptures,
176  unsigned MaxUsesToExplore) {
177  assert(!isa<GlobalValue>(V) &&
178  "It doesn't make sense to ask whether a global is captured.");
179 
180  // TODO: If StoreCaptures is not true, we could do Fancy analysis
181  // to determine whether this store is not actually an escape point.
182  // In that case, BasicAliasAnalysis should be updated as well to
183  // take advantage of this.
184  (void)StoreCaptures;
185 
186  SimpleCaptureTracker SCT(ReturnCaptures);
187  PointerMayBeCaptured(V, &SCT, MaxUsesToExplore);
188  return SCT.Captured;
189 }
190 
191 /// PointerMayBeCapturedBefore - Return true if this pointer value may be
192 /// captured by the enclosing function (which is required to exist). If a
193 /// DominatorTree is provided, only captures which happen before the given
194 /// instruction are considered. This routine can be expensive, so consider
195 /// caching the results. The boolean ReturnCaptures specifies whether
196 /// returning the value (or part of it) from the function counts as capturing
197 /// it or not. The boolean StoreCaptures specified whether storing the value
198 /// (or part of it) into memory anywhere automatically counts as capturing it
199 /// or not. A ordered basic block \p OBB can be used in order to speed up
200 /// queries about relative order among instructions in the same basic block.
201 bool llvm::PointerMayBeCapturedBefore(const Value *V, bool ReturnCaptures,
202  bool StoreCaptures, const Instruction *I,
203  const DominatorTree *DT, bool IncludeI,
204  OrderedBasicBlock *OBB,
205  unsigned MaxUsesToExplore) {
206  assert(!isa<GlobalValue>(V) &&
207  "It doesn't make sense to ask whether a global is captured.");
208  bool UseNewOBB = OBB == nullptr;
209 
210  if (!DT)
211  return PointerMayBeCaptured(V, ReturnCaptures, StoreCaptures,
212  MaxUsesToExplore);
213  if (UseNewOBB)
214  OBB = new OrderedBasicBlock(I->getParent());
215 
216  // TODO: See comment in PointerMayBeCaptured regarding what could be done
217  // with StoreCaptures.
218 
219  CapturesBefore CB(ReturnCaptures, I, DT, IncludeI, OBB);
220  PointerMayBeCaptured(V, &CB, MaxUsesToExplore);
221 
222  if (UseNewOBB)
223  delete OBB;
224  return CB.Captured;
225 }
226 
228  unsigned MaxUsesToExplore) {
229  assert(V->getType()->isPointerTy() && "Capture is for pointers only!");
232 
233  auto AddUses = [&](const Value *V) {
234  unsigned Count = 0;
235  for (const Use &U : V->uses()) {
236  // If there are lots of uses, conservatively say that the value
237  // is captured to avoid taking too much compile time.
238  if (Count++ >= MaxUsesToExplore)
239  return Tracker->tooManyUses();
240  if (!Visited.insert(&U).second)
241  continue;
242  if (!Tracker->shouldExplore(&U))
243  continue;
244  Worklist.push_back(&U);
245  }
246  };
247  AddUses(V);
248 
249  while (!Worklist.empty()) {
250  const Use *U = Worklist.pop_back_val();
251  Instruction *I = cast<Instruction>(U->getUser());
252  V = U->get();
253 
254  switch (I->getOpcode()) {
255  case Instruction::Call:
256  case Instruction::Invoke: {
257  auto *Call = cast<CallBase>(I);
258  // Not captured if the callee is readonly, doesn't return a copy through
259  // its return value and doesn't unwind (a readonly function can leak bits
260  // by throwing an exception or not depending on the input value).
261  if (Call->onlyReadsMemory() && Call->doesNotThrow() &&
262  Call->getType()->isVoidTy())
263  break;
264 
265  // The pointer is not captured if returned pointer is not captured.
266  // NOTE: CaptureTracking users should not assume that only functions
267  // marked with nocapture do not capture. This means that places like
268  // GetUnderlyingObject in ValueTracking or DecomposeGEPExpression
269  // in BasicAA also need to know about this property.
271  true)) {
272  AddUses(Call);
273  break;
274  }
275 
276  // Volatile operations effectively capture the memory location that they
277  // load and store to.
278  if (auto *MI = dyn_cast<MemIntrinsic>(Call))
279  if (MI->isVolatile())
280  if (Tracker->captured(U))
281  return;
282 
283  // Not captured if only passed via 'nocapture' arguments. Note that
284  // calling a function pointer does not in itself cause the pointer to
285  // be captured. This is a subtle point considering that (for example)
286  // the callee might return its own address. It is analogous to saying
287  // that loading a value from a pointer does not cause the pointer to be
288  // captured, even though the loaded value might be the pointer itself
289  // (think of self-referential objects).
290  for (auto IdxOpPair : enumerate(Call->data_ops())) {
291  int Idx = IdxOpPair.index();
292  Value *A = IdxOpPair.value();
293  if (A == V && !Call->doesNotCapture(Idx))
294  // The parameter is not marked 'nocapture' - captured.
295  if (Tracker->captured(U))
296  return;
297  }
298  break;
299  }
300  case Instruction::Load:
301  // Volatile loads make the address observable.
302  if (cast<LoadInst>(I)->isVolatile())
303  if (Tracker->captured(U))
304  return;
305  break;
306  case Instruction::VAArg:
307  // "va-arg" from a pointer does not cause it to be captured.
308  break;
309  case Instruction::Store:
310  // Stored the pointer - conservatively assume it may be captured.
311  // Volatile stores make the address observable.
312  if (V == I->getOperand(0) || cast<StoreInst>(I)->isVolatile())
313  if (Tracker->captured(U))
314  return;
315  break;
316  case Instruction::AtomicRMW: {
317  // atomicrmw conceptually includes both a load and store from
318  // the same location.
319  // As with a store, the location being accessed is not captured,
320  // but the value being stored is.
321  // Volatile stores make the address observable.
322  auto *ARMWI = cast<AtomicRMWInst>(I);
323  if (ARMWI->getValOperand() == V || ARMWI->isVolatile())
324  if (Tracker->captured(U))
325  return;
326  break;
327  }
328  case Instruction::AtomicCmpXchg: {
329  // cmpxchg conceptually includes both a load and store from
330  // the same location.
331  // As with a store, the location being accessed is not captured,
332  // but the value being stored is.
333  // Volatile stores make the address observable.
334  auto *ACXI = cast<AtomicCmpXchgInst>(I);
335  if (ACXI->getCompareOperand() == V || ACXI->getNewValOperand() == V ||
336  ACXI->isVolatile())
337  if (Tracker->captured(U))
338  return;
339  break;
340  }
341  case Instruction::BitCast:
342  case Instruction::GetElementPtr:
343  case Instruction::PHI:
344  case Instruction::Select:
345  case Instruction::AddrSpaceCast:
346  // The original value is not captured via this if the new value isn't.
347  AddUses(I);
348  break;
349  case Instruction::ICmp: {
350  unsigned Idx = (I->getOperand(0) == V) ? 0 : 1;
351  unsigned OtherIdx = 1 - Idx;
352  if (auto *CPN = dyn_cast<ConstantPointerNull>(I->getOperand(OtherIdx))) {
353  // Don't count comparisons of a no-alias return value against null as
354  // captures. This allows us to ignore comparisons of malloc results
355  // with null, for example.
356  if (CPN->getType()->getAddressSpace() == 0)
357  if (isNoAliasCall(V->stripPointerCasts()))
358  break;
359  if (!I->getFunction()->nullPointerIsDefined()) {
361  // Comparing a dereferenceable_or_null pointer against null cannot
362  // lead to pointer escapes, because if it is not null it must be a
363  // valid (in-bounds) pointer.
364  if (Tracker->isDereferenceableOrNull(O, I->getModule()->getDataLayout()))
365  break;
366  }
367  }
368  // Comparison against value stored in global variable. Given the pointer
369  // does not escape, its value cannot be guessed and stored separately in a
370  // global variable.
371  auto *LI = dyn_cast<LoadInst>(I->getOperand(OtherIdx));
372  if (LI && isa<GlobalVariable>(LI->getPointerOperand()))
373  break;
374  // Otherwise, be conservative. There are crazy ways to capture pointers
375  // using comparisons.
376  if (Tracker->captured(U))
377  return;
378  break;
379  }
380  default:
381  // Something else - be conservative and say it is captured.
382  if (Tracker->captured(U))
383  return;
384  break;
385  }
386  }
387 
388  // All uses examined.
389 }
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:112
iterator_range< use_iterator > uses()
Definition: Value.h:374
virtual void tooManyUses()=0
tooManyUses - The depth of traversal has breached a limit.
This callback is used in conjunction with PointerMayBeCaptured.
This class represents lattice values for constants.
Definition: AllocatorList.h:23
bool isPotentiallyReachable(const Instruction *From, const Instruction *To, const SmallPtrSetImpl< BasicBlock *> *ExclusionSet=nullptr, const DominatorTree *DT=nullptr, const LoopInfo *LI=nullptr)
Determine whether instruction &#39;To&#39; is reachable from &#39;From&#39;, without passing through any blocks in Ex...
Definition: CFG.cpp:218
virtual bool shouldExplore(const Use *U)
shouldExplore - This is the use of a value derived from the pointer.
bool isNoAliasCall(const Value *V)
Return true if this pointer is returned by a noalias function.
An instruction for reading from memory.
Definition: Instructions.h:169
Hexagon Common GEP
const Instruction * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Definition: BasicBlock.cpp:144
bool isReachableFromEntry(const Use &U) const
Provide an overload for a Use.
Definition: Dominators.cpp:299
bool isIntrinsicReturningPointerAliasingArgumentWithoutCapturing(const CallBase *Call, bool MustPreserveNullness)
{launder,strip}.invariant.group returns pointer that aliases its argument, and it only captures point...
virtual bool isDereferenceableOrNull(Value *O, const DataLayout &DL)
isDereferenceableOrNull - Overload to allow clients with additional knowledge about pointer dereferen...
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:369
A Use represents the edge between a Value definition and its users.
Definition: Use.h:55
virtual bool captured(const Use *U)=0
captured - Information about the pointer was captured by the user of use U.
Interval::succ_iterator succ_begin(Interval *I)
succ_begin/succ_end - define methods so that Intervals may be used just like BasicBlocks can with the...
Definition: Interval.h:102
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:125
bool PointerMayBeCapturedBefore(const Value *V, bool ReturnCaptures, bool StoreCaptures, const Instruction *I, const DominatorTree *DT, bool IncludeI=false, OrderedBasicBlock *OBB=nullptr, unsigned MaxUsesToExplore=DefaultMaxUsesToExplore)
PointerMayBeCapturedBefore - Return true if this pointer value may be captured by the enclosing funct...
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:144
unsigned getNumSuccessors() const
Return the number of successors that this instruction has.
Value * getOperand(unsigned i) const
Definition: User.h:169
Interval::succ_iterator succ_end(Interval *I)
Definition: Interval.h:105
const BasicBlock & getEntryBlock() const
Definition: Function.h:664
LLVM Basic Block Representation.
Definition: BasicBlock.h:57
SmallSet - This maintains a set of unique values, optimizing for the case when the set is small (less...
Definition: SmallSet.h:134
This file contains the declarations for the subclasses of Constant, which represent the different fla...
bool isPointerTy() const
True if this is an instance of PointerType.
Definition: Type.h:224
std::pair< NoneType, bool > insert(const T &V)
insert - Insert an element into the set if it isn&#39;t already there.
Definition: SmallSet.h:180
const Function * getFunction() const
Return the function this instruction belongs to.
Definition: Instruction.cpp:59
const Value * stripPointerCastsSameRepresentation() const
Strip off pointer casts, all-zero GEPs and address space casts but ensures the representation of the ...
Definition: Value.cpp:537
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:837
bool dominates(const Instruction *Def, const Use &U) const
Return true if Def dominates a use in User.
Definition: Dominators.cpp:248
uint64_t getPointerDereferenceableBytes(const DataLayout &DL, bool &CanBeNull) const
Returns the number of bytes known to be dereferenceable for the pointer value.
Definition: Value.cpp:608
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:55
void append(in_iter in_start, in_iter in_end)
Add the specified range to the end of the SmallVector.
Definition: SmallVector.h:387
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:106
#define I(x, y, z)
Definition: MD5.cpp:58
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:332
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
LLVM Value Representation.
Definition: Value.h:73
IRTranslator LLVM IR MI
static bool isVolatile(Instruction *Inst)
bool nullPointerIsDefined() const
Check if null pointer dereferencing is considered undefined behavior for the function.
Definition: Function.cpp:1598
bool isPotentiallyReachableFromMany(SmallVectorImpl< BasicBlock *> &Worklist, BasicBlock *StopBB, const DominatorTree *DT=nullptr, const LoopInfo *LI=nullptr)
Determine whether there is at least one path from a block in &#39;Worklist&#39; to &#39;StopBB&#39;, returning true if uncertain.
detail::enumerator< R > enumerate(R &&TheRange)
Given an input range, returns a new range whose values are are pair (A,B) such that A is the 0-based ...
Definition: STLExtras.h:1500
const BasicBlock * getParent() const
Definition: Instruction.h:66
bool PointerMayBeCaptured(const Value *V, bool ReturnCaptures, bool StoreCaptures, unsigned MaxUsesToExplore=DefaultMaxUsesToExplore)
PointerMayBeCaptured - Return true if this pointer value may be captured by the enclosing function (w...