1 //=-- ExplodedGraph.cpp - Local, Path-Sens. "Exploded Graph" -*- C++ -*------=//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines the template classes ExplodedNode and ExplodedGraph,
11 // which represent a path-sensitive, intra-procedural "exploded graph."
12 //
13 //===----------------------------------------------------------------------===//
14
15 #include "clang/StaticAnalyzer/Core/PathSensitive/ExplodedGraph.h"
16 #include "clang/AST/ParentMap.h"
17 #include "clang/AST/Stmt.h"
18 #include "clang/StaticAnalyzer/Core/PathSensitive/CallEvent.h"
19 #include "clang/StaticAnalyzer/Core/PathSensitive/ProgramState.h"
20 #include "llvm/ADT/DenseMap.h"
21 #include "llvm/ADT/DenseSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/Statistic.h"
24 #include <vector>
25
26 using namespace clang;
27 using namespace ento;
28
29 //===----------------------------------------------------------------------===//
30 // Node auditing.
31 //===----------------------------------------------------------------------===//
32
33 // An out of line virtual method to provide a home for the class vtable.
~Auditor()34 ExplodedNode::Auditor::~Auditor() {}
35
36 #ifndef NDEBUG
37 static ExplodedNode::Auditor* NodeAuditor = nullptr;
38 #endif
39
SetAuditor(ExplodedNode::Auditor * A)40 void ExplodedNode::SetAuditor(ExplodedNode::Auditor* A) {
41 #ifndef NDEBUG
42 NodeAuditor = A;
43 #endif
44 }
45
46 //===----------------------------------------------------------------------===//
47 // Cleanup.
48 //===----------------------------------------------------------------------===//
49
ExplodedGraph()50 ExplodedGraph::ExplodedGraph()
51 : NumNodes(0), ReclaimNodeInterval(0) {}
52
~ExplodedGraph()53 ExplodedGraph::~ExplodedGraph() {}
54
55 //===----------------------------------------------------------------------===//
56 // Node reclamation.
57 //===----------------------------------------------------------------------===//
58
isInterestingLValueExpr(const Expr * Ex)59 bool ExplodedGraph::isInterestingLValueExpr(const Expr *Ex) {
60 if (!Ex->isLValue())
61 return false;
62 return isa<DeclRefExpr>(Ex) ||
63 isa<MemberExpr>(Ex) ||
64 isa<ObjCIvarRefExpr>(Ex);
65 }
66
shouldCollect(const ExplodedNode * node)67 bool ExplodedGraph::shouldCollect(const ExplodedNode *node) {
68 // First, we only consider nodes for reclamation of the following
69 // conditions apply:
70 //
71 // (1) 1 predecessor (that has one successor)
72 // (2) 1 successor (that has one predecessor)
73 //
74 // If a node has no successor it is on the "frontier", while a node
75 // with no predecessor is a root.
76 //
77 // After these prerequisites, we discard all "filler" nodes that
78 // are used only for intermediate processing, and are not essential
79 // for analyzer history:
80 //
81 // (a) PreStmtPurgeDeadSymbols
82 //
83 // We then discard all other nodes where *all* of the following conditions
84 // apply:
85 //
86 // (3) The ProgramPoint is for a PostStmt, but not a PostStore.
87 // (4) There is no 'tag' for the ProgramPoint.
88 // (5) The 'store' is the same as the predecessor.
89 // (6) The 'GDM' is the same as the predecessor.
90 // (7) The LocationContext is the same as the predecessor.
91 // (8) Expressions that are *not* lvalue expressions.
92 // (9) The PostStmt isn't for a non-consumed Stmt or Expr.
93 // (10) The successor is neither a CallExpr StmtPoint nor a CallEnter or
94 // PreImplicitCall (so that we would be able to find it when retrying a
95 // call with no inlining).
96 // FIXME: It may be safe to reclaim PreCall and PostCall nodes as well.
97
98 // Conditions 1 and 2.
99 if (node->pred_size() != 1 || node->succ_size() != 1)
100 return false;
101
102 const ExplodedNode *pred = *(node->pred_begin());
103 if (pred->succ_size() != 1)
104 return false;
105
106 const ExplodedNode *succ = *(node->succ_begin());
107 if (succ->pred_size() != 1)
108 return false;
109
110 // Now reclaim any nodes that are (by definition) not essential to
111 // analysis history and are not consulted by any client code.
112 ProgramPoint progPoint = node->getLocation();
113 if (progPoint.getAs<PreStmtPurgeDeadSymbols>())
114 return !progPoint.getTag();
115
116 // Condition 3.
117 if (!progPoint.getAs<PostStmt>() || progPoint.getAs<PostStore>())
118 return false;
119
120 // Condition 4.
121 if (progPoint.getTag())
122 return false;
123
124 // Conditions 5, 6, and 7.
125 ProgramStateRef state = node->getState();
126 ProgramStateRef pred_state = pred->getState();
127 if (state->store != pred_state->store || state->GDM != pred_state->GDM ||
128 progPoint.getLocationContext() != pred->getLocationContext())
129 return false;
130
131 // All further checks require expressions. As per #3, we know that we have
132 // a PostStmt.
133 const Expr *Ex = dyn_cast<Expr>(progPoint.castAs<PostStmt>().getStmt());
134 if (!Ex)
135 return false;
136
137 // Condition 8.
138 // Do not collect nodes for "interesting" lvalue expressions since they are
139 // used extensively for generating path diagnostics.
140 if (isInterestingLValueExpr(Ex))
141 return false;
142
143 // Condition 9.
144 // Do not collect nodes for non-consumed Stmt or Expr to ensure precise
145 // diagnostic generation; specifically, so that we could anchor arrows
146 // pointing to the beginning of statements (as written in code).
147 ParentMap &PM = progPoint.getLocationContext()->getParentMap();
148 if (!PM.isConsumedExpr(Ex))
149 return false;
150
151 // Condition 10.
152 const ProgramPoint SuccLoc = succ->getLocation();
153 if (Optional<StmtPoint> SP = SuccLoc.getAs<StmtPoint>())
154 if (CallEvent::isCallStmt(SP->getStmt()))
155 return false;
156
157 // Condition 10, continuation.
158 if (SuccLoc.getAs<CallEnter>() || SuccLoc.getAs<PreImplicitCall>())
159 return false;
160
161 return true;
162 }
163
collectNode(ExplodedNode * node)164 void ExplodedGraph::collectNode(ExplodedNode *node) {
165 // Removing a node means:
166 // (a) changing the predecessors successor to the successor of this node
167 // (b) changing the successors predecessor to the predecessor of this node
168 // (c) Putting 'node' onto freeNodes.
169 assert(node->pred_size() == 1 || node->succ_size() == 1);
170 ExplodedNode *pred = *(node->pred_begin());
171 ExplodedNode *succ = *(node->succ_begin());
172 pred->replaceSuccessor(succ);
173 succ->replacePredecessor(pred);
174 FreeNodes.push_back(node);
175 Nodes.RemoveNode(node);
176 --NumNodes;
177 node->~ExplodedNode();
178 }
179
reclaimRecentlyAllocatedNodes()180 void ExplodedGraph::reclaimRecentlyAllocatedNodes() {
181 if (ChangedNodes.empty())
182 return;
183
184 // Only periodically reclaim nodes so that we can build up a set of
185 // nodes that meet the reclamation criteria. Freshly created nodes
186 // by definition have no successor, and thus cannot be reclaimed (see below).
187 assert(ReclaimCounter > 0);
188 if (--ReclaimCounter != 0)
189 return;
190 ReclaimCounter = ReclaimNodeInterval;
191
192 for (NodeVector::iterator it = ChangedNodes.begin(), et = ChangedNodes.end();
193 it != et; ++it) {
194 ExplodedNode *node = *it;
195 if (shouldCollect(node))
196 collectNode(node);
197 }
198 ChangedNodes.clear();
199 }
200
201 //===----------------------------------------------------------------------===//
202 // ExplodedNode.
203 //===----------------------------------------------------------------------===//
204
205 // An NodeGroup's storage type is actually very much like a TinyPtrVector:
206 // it can be either a pointer to a single ExplodedNode, or a pointer to a
207 // BumpVector allocated with the ExplodedGraph's allocator. This allows the
208 // common case of single-node NodeGroups to be implemented with no extra memory.
209 //
210 // Consequently, each of the NodeGroup methods have up to four cases to handle:
211 // 1. The flag is set and this group does not actually contain any nodes.
212 // 2. The group is empty, in which case the storage value is null.
213 // 3. The group contains a single node.
214 // 4. The group contains more than one node.
215 typedef BumpVector<ExplodedNode *> ExplodedNodeVector;
216 typedef llvm::PointerUnion<ExplodedNode *, ExplodedNodeVector *> GroupStorage;
217
addPredecessor(ExplodedNode * V,ExplodedGraph & G)218 void ExplodedNode::addPredecessor(ExplodedNode *V, ExplodedGraph &G) {
219 assert (!V->isSink());
220 Preds.addNode(V, G);
221 V->Succs.addNode(this, G);
222 #ifndef NDEBUG
223 if (NodeAuditor) NodeAuditor->AddEdge(V, this);
224 #endif
225 }
226
replaceNode(ExplodedNode * node)227 void ExplodedNode::NodeGroup::replaceNode(ExplodedNode *node) {
228 assert(!getFlag());
229
230 GroupStorage &Storage = reinterpret_cast<GroupStorage&>(P);
231 assert(Storage.is<ExplodedNode *>());
232 Storage = node;
233 assert(Storage.is<ExplodedNode *>());
234 }
235
addNode(ExplodedNode * N,ExplodedGraph & G)236 void ExplodedNode::NodeGroup::addNode(ExplodedNode *N, ExplodedGraph &G) {
237 assert(!getFlag());
238
239 GroupStorage &Storage = reinterpret_cast<GroupStorage&>(P);
240 if (Storage.isNull()) {
241 Storage = N;
242 assert(Storage.is<ExplodedNode *>());
243 return;
244 }
245
246 ExplodedNodeVector *V = Storage.dyn_cast<ExplodedNodeVector *>();
247
248 if (!V) {
249 // Switch from single-node to multi-node representation.
250 ExplodedNode *Old = Storage.get<ExplodedNode *>();
251
252 BumpVectorContext &Ctx = G.getNodeAllocator();
253 V = G.getAllocator().Allocate<ExplodedNodeVector>();
254 new (V) ExplodedNodeVector(Ctx, 4);
255 V->push_back(Old, Ctx);
256
257 Storage = V;
258 assert(!getFlag());
259 assert(Storage.is<ExplodedNodeVector *>());
260 }
261
262 V->push_back(N, G.getNodeAllocator());
263 }
264
size() const265 unsigned ExplodedNode::NodeGroup::size() const {
266 if (getFlag())
267 return 0;
268
269 const GroupStorage &Storage = reinterpret_cast<const GroupStorage &>(P);
270 if (Storage.isNull())
271 return 0;
272 if (ExplodedNodeVector *V = Storage.dyn_cast<ExplodedNodeVector *>())
273 return V->size();
274 return 1;
275 }
276
begin() const277 ExplodedNode * const *ExplodedNode::NodeGroup::begin() const {
278 if (getFlag())
279 return nullptr;
280
281 const GroupStorage &Storage = reinterpret_cast<const GroupStorage &>(P);
282 if (Storage.isNull())
283 return nullptr;
284 if (ExplodedNodeVector *V = Storage.dyn_cast<ExplodedNodeVector *>())
285 return V->begin();
286 return Storage.getAddrOfPtr1();
287 }
288
end() const289 ExplodedNode * const *ExplodedNode::NodeGroup::end() const {
290 if (getFlag())
291 return nullptr;
292
293 const GroupStorage &Storage = reinterpret_cast<const GroupStorage &>(P);
294 if (Storage.isNull())
295 return nullptr;
296 if (ExplodedNodeVector *V = Storage.dyn_cast<ExplodedNodeVector *>())
297 return V->end();
298 return Storage.getAddrOfPtr1() + 1;
299 }
300
getNode(const ProgramPoint & L,ProgramStateRef State,bool IsSink,bool * IsNew)301 ExplodedNode *ExplodedGraph::getNode(const ProgramPoint &L,
302 ProgramStateRef State,
303 bool IsSink,
304 bool* IsNew) {
305 // Profile 'State' to determine if we already have an existing node.
306 llvm::FoldingSetNodeID profile;
307 void *InsertPos = nullptr;
308
309 NodeTy::Profile(profile, L, State, IsSink);
310 NodeTy* V = Nodes.FindNodeOrInsertPos(profile, InsertPos);
311
312 if (!V) {
313 if (!FreeNodes.empty()) {
314 V = FreeNodes.back();
315 FreeNodes.pop_back();
316 }
317 else {
318 // Allocate a new node.
319 V = (NodeTy*) getAllocator().Allocate<NodeTy>();
320 }
321
322 new (V) NodeTy(L, State, IsSink);
323
324 if (ReclaimNodeInterval)
325 ChangedNodes.push_back(V);
326
327 // Insert the node into the node set and return it.
328 Nodes.InsertNode(V, InsertPos);
329 ++NumNodes;
330
331 if (IsNew) *IsNew = true;
332 }
333 else
334 if (IsNew) *IsNew = false;
335
336 return V;
337 }
338
339 ExplodedGraph *
trim(ArrayRef<const NodeTy * > Sinks,InterExplodedGraphMap * ForwardMap,InterExplodedGraphMap * InverseMap) const340 ExplodedGraph::trim(ArrayRef<const NodeTy *> Sinks,
341 InterExplodedGraphMap *ForwardMap,
342 InterExplodedGraphMap *InverseMap) const{
343
344 if (Nodes.empty())
345 return nullptr;
346
347 typedef llvm::DenseSet<const ExplodedNode*> Pass1Ty;
348 Pass1Ty Pass1;
349
350 typedef InterExplodedGraphMap Pass2Ty;
351 InterExplodedGraphMap Pass2Scratch;
352 Pass2Ty &Pass2 = ForwardMap ? *ForwardMap : Pass2Scratch;
353
354 SmallVector<const ExplodedNode*, 10> WL1, WL2;
355
356 // ===- Pass 1 (reverse DFS) -===
357 for (ArrayRef<const NodeTy *>::iterator I = Sinks.begin(), E = Sinks.end();
358 I != E; ++I) {
359 if (*I)
360 WL1.push_back(*I);
361 }
362
363 // Process the first worklist until it is empty.
364 while (!WL1.empty()) {
365 const ExplodedNode *N = WL1.pop_back_val();
366
367 // Have we already visited this node? If so, continue to the next one.
368 if (Pass1.count(N))
369 continue;
370
371 // Otherwise, mark this node as visited.
372 Pass1.insert(N);
373
374 // If this is a root enqueue it to the second worklist.
375 if (N->Preds.empty()) {
376 WL2.push_back(N);
377 continue;
378 }
379
380 // Visit our predecessors and enqueue them.
381 for (ExplodedNode::pred_iterator I = N->Preds.begin(), E = N->Preds.end();
382 I != E; ++I)
383 WL1.push_back(*I);
384 }
385
386 // We didn't hit a root? Return with a null pointer for the new graph.
387 if (WL2.empty())
388 return nullptr;
389
390 // Create an empty graph.
391 ExplodedGraph* G = MakeEmptyGraph();
392
393 // ===- Pass 2 (forward DFS to construct the new graph) -===
394 while (!WL2.empty()) {
395 const ExplodedNode *N = WL2.pop_back_val();
396
397 // Skip this node if we have already processed it.
398 if (Pass2.find(N) != Pass2.end())
399 continue;
400
401 // Create the corresponding node in the new graph and record the mapping
402 // from the old node to the new node.
403 ExplodedNode *NewN = G->getNode(N->getLocation(), N->State, N->isSink(),
404 nullptr);
405 Pass2[N] = NewN;
406
407 // Also record the reverse mapping from the new node to the old node.
408 if (InverseMap) (*InverseMap)[NewN] = N;
409
410 // If this node is a root, designate it as such in the graph.
411 if (N->Preds.empty())
412 G->addRoot(NewN);
413
414 // In the case that some of the intended predecessors of NewN have already
415 // been created, we should hook them up as predecessors.
416
417 // Walk through the predecessors of 'N' and hook up their corresponding
418 // nodes in the new graph (if any) to the freshly created node.
419 for (ExplodedNode::pred_iterator I = N->Preds.begin(), E = N->Preds.end();
420 I != E; ++I) {
421 Pass2Ty::iterator PI = Pass2.find(*I);
422 if (PI == Pass2.end())
423 continue;
424
425 NewN->addPredecessor(const_cast<ExplodedNode *>(PI->second), *G);
426 }
427
428 // In the case that some of the intended successors of NewN have already
429 // been created, we should hook them up as successors. Otherwise, enqueue
430 // the new nodes from the original graph that should have nodes created
431 // in the new graph.
432 for (ExplodedNode::succ_iterator I = N->Succs.begin(), E = N->Succs.end();
433 I != E; ++I) {
434 Pass2Ty::iterator PI = Pass2.find(*I);
435 if (PI != Pass2.end()) {
436 const_cast<ExplodedNode *>(PI->second)->addPredecessor(NewN, *G);
437 continue;
438 }
439
440 // Enqueue nodes to the worklist that were marked during pass 1.
441 if (Pass1.count(*I))
442 WL2.push_back(*I);
443 }
444 }
445
446 return G;
447 }
448
449