1 //===--- CaptureTracking.cpp - Determine whether a pointer is captured ----===//
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
7 //===----------------------------------------------------------------------===//
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.
16 //===----------------------------------------------------------------------===//
18 #include "llvm/Analysis/CaptureTracking.h"
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/Analysis/AliasAnalysis.h"
22 #include "llvm/Analysis/CFG.h"
23 #include "llvm/Analysis/OrderedBasicBlock.h"
24 #include "llvm/Analysis/ValueTracking.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"
32 CaptureTracker::~CaptureTracker() {}
34 bool CaptureTracker::shouldExplore(const Use *U) { return true; }
37 struct SimpleCaptureTracker : public CaptureTracker {
38 explicit SimpleCaptureTracker(bool ReturnCaptures)
39 : ReturnCaptures(ReturnCaptures), Captured(false) {}
41 void tooManyUses() override { Captured = true; }
43 bool captured(const Use *U) override {
44 if (isa<ReturnInst>(U->getUser()) && !ReturnCaptures)
56 /// Only find pointer captures which happen before the given instruction. Uses
57 /// the dominator tree to determine whether one instruction is before another.
58 /// Only support the case where the Value is defined in the same basic block
59 /// as the given instruction and the use.
60 struct CapturesBefore : public CaptureTracker {
62 CapturesBefore(bool ReturnCaptures, const Instruction *I, const DominatorTree *DT,
63 bool IncludeI, OrderedBasicBlock *IC)
64 : OrderedBB(IC), BeforeHere(I), DT(DT),
65 ReturnCaptures(ReturnCaptures), IncludeI(IncludeI), Captured(false) {}
67 void tooManyUses() override { Captured = true; }
69 bool isSafeToPrune(Instruction *I) {
70 BasicBlock *BB = I->getParent();
71 // We explore this usage only if the usage can reach "BeforeHere".
72 // If use is not reachable from entry, there is no need to explore.
73 if (BeforeHere != I && !DT->isReachableFromEntry(BB))
76 // Compute the case where both instructions are inside the same basic
77 // block. Since instructions in the same BB as BeforeHere are numbered in
78 // 'OrderedBB', avoid using 'dominates' and 'isPotentiallyReachable'
79 // which are very expensive for large basic blocks.
80 if (BB == BeforeHere->getParent()) {
81 // 'I' dominates 'BeforeHere' => not safe to prune.
83 // The value defined by an invoke dominates an instruction only
84 // if it dominates every instruction in UseBB. A PHI is dominated only
85 // if the instruction dominates every possible use in the UseBB. Since
86 // UseBB == BB, avoid pruning.
87 if (isa<InvokeInst>(BeforeHere) || isa<PHINode>(I) || I == BeforeHere)
89 if (!OrderedBB->dominates(BeforeHere, I))
92 // 'BeforeHere' comes before 'I', it's safe to prune if we also
93 // guarantee that 'I' never reaches 'BeforeHere' through a back-edge or
94 // by its successors, i.e, prune if:
96 // (1) BB is an entry block or have no successors.
97 // (2) There's no path coming back through BB successors.
98 if (BB == &BB->getParent()->getEntryBlock() ||
99 !BB->getTerminator()->getNumSuccessors())
102 SmallVector<BasicBlock*, 32> Worklist;
103 Worklist.append(succ_begin(BB), succ_end(BB));
104 return !isPotentiallyReachableFromMany(Worklist, BB, nullptr, DT);
107 // If the value is defined in the same basic block as use and BeforeHere,
108 // there is no need to explore the use if BeforeHere dominates use.
109 // Check whether there is a path from I to BeforeHere.
110 if (BeforeHere != I && DT->dominates(BeforeHere, I) &&
111 !isPotentiallyReachable(I, BeforeHere, nullptr, DT))
117 bool shouldExplore(const Use *U) override {
118 Instruction *I = cast<Instruction>(U->getUser());
120 if (BeforeHere == I && !IncludeI)
123 if (isSafeToPrune(I))
129 bool captured(const Use *U) override {
130 if (isa<ReturnInst>(U->getUser()) && !ReturnCaptures)
133 if (!shouldExplore(U))
140 OrderedBasicBlock *OrderedBB;
141 const Instruction *BeforeHere;
142 const DominatorTree *DT;
151 /// PointerMayBeCaptured - Return true if this pointer value may be captured
152 /// by the enclosing function (which is required to exist). This routine can
153 /// be expensive, so consider caching the results. The boolean ReturnCaptures
154 /// specifies whether returning the value (or part of it) from the function
155 /// counts as capturing it or not. The boolean StoreCaptures specified whether
156 /// storing the value (or part of it) into memory anywhere automatically
157 /// counts as capturing it or not.
158 bool llvm::PointerMayBeCaptured(const Value *V,
159 bool ReturnCaptures, bool StoreCaptures,
160 unsigned MaxUsesToExplore) {
161 assert(!isa<GlobalValue>(V) &&
162 "It doesn't make sense to ask whether a global is captured.");
164 // TODO: If StoreCaptures is not true, we could do Fancy analysis
165 // to determine whether this store is not actually an escape point.
166 // In that case, BasicAliasAnalysis should be updated as well to
167 // take advantage of this.
170 SimpleCaptureTracker SCT(ReturnCaptures);
171 PointerMayBeCaptured(V, &SCT, MaxUsesToExplore);
175 /// PointerMayBeCapturedBefore - Return true if this pointer value may be
176 /// captured by the enclosing function (which is required to exist). If a
177 /// DominatorTree is provided, only captures which happen before the given
178 /// instruction are considered. This routine can be expensive, so consider
179 /// caching the results. The boolean ReturnCaptures specifies whether
180 /// returning the value (or part of it) from the function counts as capturing
181 /// it or not. The boolean StoreCaptures specified whether storing the value
182 /// (or part of it) into memory anywhere automatically counts as capturing it
183 /// or not. A ordered basic block \p OBB can be used in order to speed up
184 /// queries about relative order among instructions in the same basic block.
185 bool llvm::PointerMayBeCapturedBefore(const Value *V, bool ReturnCaptures,
186 bool StoreCaptures, const Instruction *I,
187 const DominatorTree *DT, bool IncludeI,
188 OrderedBasicBlock *OBB,
189 unsigned MaxUsesToExplore) {
190 assert(!isa<GlobalValue>(V) &&
191 "It doesn't make sense to ask whether a global is captured.");
192 bool UseNewOBB = OBB == nullptr;
195 return PointerMayBeCaptured(V, ReturnCaptures, StoreCaptures,
198 OBB = new OrderedBasicBlock(I->getParent());
200 // TODO: See comment in PointerMayBeCaptured regarding what could be done
201 // with StoreCaptures.
203 CapturesBefore CB(ReturnCaptures, I, DT, IncludeI, OBB);
204 PointerMayBeCaptured(V, &CB, MaxUsesToExplore);
211 void llvm::PointerMayBeCaptured(const Value *V, CaptureTracker *Tracker,
212 unsigned MaxUsesToExplore) {
213 assert(V->getType()->isPointerTy() && "Capture is for pointers only!");
214 SmallVector<const Use *, DefaultMaxUsesToExplore> Worklist;
215 SmallSet<const Use *, DefaultMaxUsesToExplore> Visited;
217 auto AddUses = [&](const Value *V) {
219 for (const Use &U : V->uses()) {
220 // If there are lots of uses, conservatively say that the value
221 // is captured to avoid taking too much compile time.
222 if (Count++ >= MaxUsesToExplore)
223 return Tracker->tooManyUses();
224 if (!Visited.insert(&U).second)
226 if (!Tracker->shouldExplore(&U))
228 Worklist.push_back(&U);
233 while (!Worklist.empty()) {
234 const Use *U = Worklist.pop_back_val();
235 Instruction *I = cast<Instruction>(U->getUser());
238 switch (I->getOpcode()) {
239 case Instruction::Call:
240 case Instruction::Invoke: {
241 auto *Call = cast<CallBase>(I);
242 // Not captured if the callee is readonly, doesn't return a copy through
243 // its return value and doesn't unwind (a readonly function can leak bits
244 // by throwing an exception or not depending on the input value).
245 if (Call->onlyReadsMemory() && Call->doesNotThrow() &&
246 Call->getType()->isVoidTy())
249 // The pointer is not captured if returned pointer is not captured.
250 // NOTE: CaptureTracking users should not assume that only functions
251 // marked with nocapture do not capture. This means that places like
252 // GetUnderlyingObject in ValueTracking or DecomposeGEPExpression
253 // in BasicAA also need to know about this property.
254 if (isIntrinsicReturningPointerAliasingArgumentWithoutCapturing(Call)) {
259 // Volatile operations effectively capture the memory location that they
260 // load and store to.
261 if (auto *MI = dyn_cast<MemIntrinsic>(Call))
262 if (MI->isVolatile())
263 if (Tracker->captured(U))
266 // Not captured if only passed via 'nocapture' arguments. Note that
267 // calling a function pointer does not in itself cause the pointer to
268 // be captured. This is a subtle point considering that (for example)
269 // the callee might return its own address. It is analogous to saying
270 // that loading a value from a pointer does not cause the pointer to be
271 // captured, even though the loaded value might be the pointer itself
272 // (think of self-referential objects).
273 for (auto IdxOpPair : enumerate(Call->data_ops())) {
274 int Idx = IdxOpPair.index();
275 Value *A = IdxOpPair.value();
276 if (A == V && !Call->doesNotCapture(Idx))
277 // The parameter is not marked 'nocapture' - captured.
278 if (Tracker->captured(U))
283 case Instruction::Load:
284 // Volatile loads make the address observable.
285 if (cast<LoadInst>(I)->isVolatile())
286 if (Tracker->captured(U))
289 case Instruction::VAArg:
290 // "va-arg" from a pointer does not cause it to be captured.
292 case Instruction::Store:
293 // Stored the pointer - conservatively assume it may be captured.
294 // Volatile stores make the address observable.
295 if (V == I->getOperand(0) || cast<StoreInst>(I)->isVolatile())
296 if (Tracker->captured(U))
299 case Instruction::AtomicRMW: {
300 // atomicrmw conceptually includes both a load and store from
301 // the same location.
302 // As with a store, the location being accessed is not captured,
303 // but the value being stored is.
304 // Volatile stores make the address observable.
305 auto *ARMWI = cast<AtomicRMWInst>(I);
306 if (ARMWI->getValOperand() == V || ARMWI->isVolatile())
307 if (Tracker->captured(U))
311 case Instruction::AtomicCmpXchg: {
312 // cmpxchg conceptually includes both a load and store from
313 // the same location.
314 // As with a store, the location being accessed is not captured,
315 // but the value being stored is.
316 // Volatile stores make the address observable.
317 auto *ACXI = cast<AtomicCmpXchgInst>(I);
318 if (ACXI->getCompareOperand() == V || ACXI->getNewValOperand() == V ||
320 if (Tracker->captured(U))
324 case Instruction::BitCast:
325 case Instruction::GetElementPtr:
326 case Instruction::PHI:
327 case Instruction::Select:
328 case Instruction::AddrSpaceCast:
329 // The original value is not captured via this if the new value isn't.
332 case Instruction::ICmp: {
333 if (auto *CPN = dyn_cast<ConstantPointerNull>(I->getOperand(1))) {
334 // Don't count comparisons of a no-alias return value against null as
335 // captures. This allows us to ignore comparisons of malloc results
336 // with null, for example.
337 if (CPN->getType()->getAddressSpace() == 0)
338 if (isNoAliasCall(V->stripPointerCasts()))
340 if (!I->getFunction()->nullPointerIsDefined()) {
341 auto *O = I->getOperand(0)->stripPointerCastsSameRepresentation();
342 // An inbounds GEP can either be a valid pointer (pointing into
343 // or to the end of an allocation), or be null in the default
344 // address space. So for an inbounds GEPs there is no way to let
345 // the pointer escape using clever GEP hacking because doing so
346 // would make the pointer point outside of the allocated object
347 // and thus make the GEP result a poison value.
348 if (auto *GEP = dyn_cast<GetElementPtrInst>(O))
349 if (GEP->isInBounds())
351 // Comparing a dereferenceable_or_null argument against null
352 // cannot lead to pointer escapes, because if it is not null it
353 // must be a valid (in-bounds) pointer.
355 if (O->getPointerDereferenceableBytes(I->getModule()->getDataLayout(), CanBeNull))
359 // Comparison against value stored in global variable. Given the pointer
360 // does not escape, its value cannot be guessed and stored separately in a
362 unsigned OtherIndex = (I->getOperand(0) == V) ? 1 : 0;
363 auto *LI = dyn_cast<LoadInst>(I->getOperand(OtherIndex));
364 if (LI && isa<GlobalVariable>(LI->getPointerOperand()))
366 // Otherwise, be conservative. There are crazy ways to capture pointers
367 // using comparisons.
368 if (Tracker->captured(U))
373 // Something else - be conservative and say it is captured.
374 if (Tracker->captured(U))
380 // All uses examined.