1 //===- CallGraph.h - Build a Module's call graph ----------------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// This file provides interfaces used to build and manipulate a call graph,
12 /// which is a very useful tool for interprocedural optimization.
14 /// Every function in a module is represented as a node in the call graph. The
15 /// callgraph node keeps track of which functions are called by the function
16 /// corresponding to the node.
18 /// A call graph may contain nodes where the function that they correspond to
19 /// is null. These 'external' nodes are used to represent control flow that is
20 /// not represented (or analyzable) in the module. In particular, this
21 /// analysis builds one external node such that:
22 /// 1. All functions in the module without internal linkage will have edges
23 /// from this external node, indicating that they could be called by
24 /// functions outside of the module.
25 /// 2. All functions whose address is used for something more than a direct
26 /// call, for example being stored into a memory location will also have
27 /// an edge from this external node. Since they may be called by an
28 /// unknown caller later, they must be tracked as such.
30 /// There is a second external node added for calls that leave this module.
31 /// Functions have a call edge to the external node iff:
32 /// 1. The function is external, reflecting the fact that they could call
33 /// anything without internal linkage or that has its address taken.
34 /// 2. The function contains an indirect function call.
36 /// As an extension in the future, there may be multiple nodes with a null
37 /// function. These will be used when we can prove (through pointer analysis)
38 /// that an indirect call site can call only a specific set of functions.
40 /// Because of these properties, the CallGraph captures a conservative superset
41 /// of all of the caller-callee relationships, which is useful for
44 /// The CallGraph class also attempts to figure out what the root of the
45 /// CallGraph is, which it currently does by looking for a function named
46 /// 'main'. If no function named 'main' is found, the external node is used as
47 /// the entry node, reflecting the fact that any function without internal
48 /// linkage could be called into (which is common for libraries).
50 //===----------------------------------------------------------------------===//
52 #ifndef LLVM_ANALYSIS_CALLGRAPH_H
53 #define LLVM_ANALYSIS_CALLGRAPH_H
55 #include "llvm/ADT/GraphTraits.h"
56 #include "llvm/ADT/STLExtras.h"
57 #include "llvm/IR/CallSite.h"
58 #include "llvm/IR/Function.h"
59 #include "llvm/IR/Intrinsics.h"
60 #include "llvm/IR/ValueHandle.h"
61 #include "llvm/Pass.h"
70 /// \brief The basic data container for the call graph of a \c Module of IR.
72 /// This class exposes both the interface to the call graph for a module of IR.
74 /// The core call graph itself can also be updated to reflect changes to the IR.
78 typedef std::map<const Function *, std::unique_ptr<CallGraphNode>>
81 /// \brief A map from \c Function* to \c CallGraphNode*.
82 FunctionMapTy FunctionMap;
84 /// \brief Root is root of the call graph, or the external node if a 'main'
85 /// function couldn't be found.
88 /// \brief This node has edges to all external functions and those internal
89 /// functions that have their address taken.
90 CallGraphNode *ExternalCallingNode;
92 /// \brief This node has edges to it from all functions making indirect calls
93 /// or calling an external function.
94 std::unique_ptr<CallGraphNode> CallsExternalNode;
96 /// \brief Replace the function represented by this node by another.
98 /// This does not rescan the body of the function, so it is suitable when
99 /// splicing the body of one function to another while also updating all
100 /// callers from the old function to the new.
101 void spliceFunction(const Function *From, const Function *To);
103 /// \brief Add a function to the call graph, and link the node to all of the
104 /// functions that it calls.
105 void addToCallGraph(Function *F);
108 CallGraph(Module &M);
109 // Default move ctor is fine, the dtor just does things to CallsExternalNode
110 // (if non-null) and the values in the FunctionMap, all of which should be
111 // null post-move, so no-op the dtor on the moved-from side.
112 CallGraph(CallGraph &&) = default;
115 void print(raw_ostream &OS) const;
118 typedef FunctionMapTy::iterator iterator;
119 typedef FunctionMapTy::const_iterator const_iterator;
121 /// \brief Returns the module the call graph corresponds to.
122 Module &getModule() const { return M; }
124 inline iterator begin() { return FunctionMap.begin(); }
125 inline iterator end() { return FunctionMap.end(); }
126 inline const_iterator begin() const { return FunctionMap.begin(); }
127 inline const_iterator end() const { return FunctionMap.end(); }
129 /// \brief Returns the call graph node for the provided function.
130 inline const CallGraphNode *operator[](const Function *F) const {
131 const_iterator I = FunctionMap.find(F);
132 assert(I != FunctionMap.end() && "Function not in callgraph!");
133 return I->second.get();
136 /// \brief Returns the call graph node for the provided function.
137 inline CallGraphNode *operator[](const Function *F) {
138 const_iterator I = FunctionMap.find(F);
139 assert(I != FunctionMap.end() && "Function not in callgraph!");
140 return I->second.get();
143 /// \brief Returns the \c CallGraphNode which is used to represent
144 /// undetermined calls into the callgraph.
145 CallGraphNode *getExternalCallingNode() const { return ExternalCallingNode; }
147 CallGraphNode *getCallsExternalNode() const {
148 return CallsExternalNode.get();
151 //===---------------------------------------------------------------------
152 // Functions to keep a call graph up to date with a function that has been
156 /// \brief Unlink the function from this module, returning it.
158 /// Because this removes the function from the module, the call graph node is
159 /// destroyed. This is only valid if the function does not call any other
160 /// functions (ie, there are no edges in it's CGN). The easiest way to do
161 /// this is to dropAllReferences before calling this.
162 Function *removeFunctionFromModule(CallGraphNode *CGN);
164 /// \brief Similar to operator[], but this will insert a new CallGraphNode for
165 /// \c F if one does not already exist.
166 CallGraphNode *getOrInsertFunction(const Function *F);
169 /// \brief A node in the call graph for a module.
171 /// Typically represents a function in the call graph. There are also special
172 /// "null" nodes used to represent theoretical entries in the call graph.
173 class CallGraphNode {
175 /// \brief A pair of the calling instruction (a call or invoke)
176 /// and the call graph node being called.
177 typedef std::pair<WeakVH, CallGraphNode *> CallRecord;
180 typedef std::vector<CallRecord> CalledFunctionsVector;
182 /// \brief Creates a node for the specified function.
183 inline CallGraphNode(Function *F) : F(F), NumReferences(0) {}
186 assert(NumReferences == 0 && "Node deleted while references remain");
189 typedef std::vector<CallRecord>::iterator iterator;
190 typedef std::vector<CallRecord>::const_iterator const_iterator;
192 /// \brief Returns the function that this call graph node represents.
193 Function *getFunction() const { return F; }
195 inline iterator begin() { return CalledFunctions.begin(); }
196 inline iterator end() { return CalledFunctions.end(); }
197 inline const_iterator begin() const { return CalledFunctions.begin(); }
198 inline const_iterator end() const { return CalledFunctions.end(); }
199 inline bool empty() const { return CalledFunctions.empty(); }
200 inline unsigned size() const { return (unsigned)CalledFunctions.size(); }
202 /// \brief Returns the number of other CallGraphNodes in this CallGraph that
203 /// reference this node in their callee list.
204 unsigned getNumReferences() const { return NumReferences; }
206 /// \brief Returns the i'th called function.
207 CallGraphNode *operator[](unsigned i) const {
208 assert(i < CalledFunctions.size() && "Invalid index");
209 return CalledFunctions[i].second;
212 /// \brief Print out this call graph node.
214 void print(raw_ostream &OS) const;
216 //===---------------------------------------------------------------------
217 // Methods to keep a call graph up to date with a function that has been
221 /// \brief Removes all edges from this CallGraphNode to any functions it
223 void removeAllCalledFunctions() {
224 while (!CalledFunctions.empty()) {
225 CalledFunctions.back().second->DropRef();
226 CalledFunctions.pop_back();
230 /// \brief Moves all the callee information from N to this node.
231 void stealCalledFunctionsFrom(CallGraphNode *N) {
232 assert(CalledFunctions.empty() &&
233 "Cannot steal callsite information if I already have some");
234 std::swap(CalledFunctions, N->CalledFunctions);
237 /// \brief Adds a function to the list of functions called by this one.
238 void addCalledFunction(CallSite CS, CallGraphNode *M) {
239 assert(!CS.getInstruction() || !CS.getCalledFunction() ||
240 !CS.getCalledFunction()->isIntrinsic() ||
241 !Intrinsic::isLeaf(CS.getCalledFunction()->getIntrinsicID()));
242 CalledFunctions.emplace_back(CS.getInstruction(), M);
246 void removeCallEdge(iterator I) {
247 I->second->DropRef();
248 *I = CalledFunctions.back();
249 CalledFunctions.pop_back();
252 /// \brief Removes the edge in the node for the specified call site.
254 /// Note that this method takes linear time, so it should be used sparingly.
255 void removeCallEdgeFor(CallSite CS);
257 /// \brief Removes all call edges from this node to the specified callee
260 /// This takes more time to execute than removeCallEdgeTo, so it should not
261 /// be used unless necessary.
262 void removeAnyCallEdgeTo(CallGraphNode *Callee);
264 /// \brief Removes one edge associated with a null callsite from this node to
265 /// the specified callee function.
266 void removeOneAbstractEdgeTo(CallGraphNode *Callee);
268 /// \brief Replaces the edge in the node for the specified call site with a
271 /// Note that this method takes linear time, so it should be used sparingly.
272 void replaceCallEdge(CallSite CS, CallSite NewCS, CallGraphNode *NewNode);
275 friend class CallGraph;
277 AssertingVH<Function> F;
279 std::vector<CallRecord> CalledFunctions;
281 /// \brief The number of times that this CallGraphNode occurs in the
282 /// CalledFunctions array of this or other CallGraphNodes.
283 unsigned NumReferences;
285 CallGraphNode(const CallGraphNode &) = delete;
286 void operator=(const CallGraphNode &) = delete;
288 void DropRef() { --NumReferences; }
289 void AddRef() { ++NumReferences; }
291 /// \brief A special function that should only be used by the CallGraph class.
292 void allReferencesDropped() { NumReferences = 0; }
295 /// \brief An analysis pass to compute the \c CallGraph for a \c Module.
297 /// This class implements the concept of an analysis pass used by the \c
298 /// ModuleAnalysisManager to run an analysis over a module and cache the
300 class CallGraphAnalysis {
302 /// \brief A formulaic typedef to inform clients of the result type.
303 typedef CallGraph Result;
305 static void *ID() { return (void *)&PassID; }
307 /// \brief Compute the \c CallGraph for the module \c M.
309 /// The real work here is done in the \c CallGraph constructor.
310 CallGraph run(Module *M) { return CallGraph(*M); }
316 /// \brief The \c ModulePass which wraps up a \c CallGraph and the logic to
319 /// This class exposes both the interface to the call graph container and the
320 /// module pass which runs over a module of IR and produces the call graph. The
321 /// call graph interface is entirelly a wrapper around a \c CallGraph object
322 /// which is stored internally for each module.
323 class CallGraphWrapperPass : public ModulePass {
324 std::unique_ptr<CallGraph> G;
327 static char ID; // Class identification, replacement for typeinfo
329 CallGraphWrapperPass();
330 ~CallGraphWrapperPass() override;
332 /// \brief The internal \c CallGraph around which the rest of this interface
334 const CallGraph &getCallGraph() const { return *G; }
335 CallGraph &getCallGraph() { return *G; }
337 typedef CallGraph::iterator iterator;
338 typedef CallGraph::const_iterator const_iterator;
340 /// \brief Returns the module the call graph corresponds to.
341 Module &getModule() const { return G->getModule(); }
343 inline iterator begin() { return G->begin(); }
344 inline iterator end() { return G->end(); }
345 inline const_iterator begin() const { return G->begin(); }
346 inline const_iterator end() const { return G->end(); }
348 /// \brief Returns the call graph node for the provided function.
349 inline const CallGraphNode *operator[](const Function *F) const {
353 /// \brief Returns the call graph node for the provided function.
354 inline CallGraphNode *operator[](const Function *F) { return (*G)[F]; }
356 /// \brief Returns the \c CallGraphNode which is used to represent
357 /// undetermined calls into the callgraph.
358 CallGraphNode *getExternalCallingNode() const {
359 return G->getExternalCallingNode();
362 CallGraphNode *getCallsExternalNode() const {
363 return G->getCallsExternalNode();
366 //===---------------------------------------------------------------------
367 // Functions to keep a call graph up to date with a function that has been
371 /// \brief Unlink the function from this module, returning it.
373 /// Because this removes the function from the module, the call graph node is
374 /// destroyed. This is only valid if the function does not call any other
375 /// functions (ie, there are no edges in it's CGN). The easiest way to do
376 /// this is to dropAllReferences before calling this.
377 Function *removeFunctionFromModule(CallGraphNode *CGN) {
378 return G->removeFunctionFromModule(CGN);
381 /// \brief Similar to operator[], but this will insert a new CallGraphNode for
382 /// \c F if one does not already exist.
383 CallGraphNode *getOrInsertFunction(const Function *F) {
384 return G->getOrInsertFunction(F);
387 //===---------------------------------------------------------------------
388 // Implementation of the ModulePass interface needed here.
391 void getAnalysisUsage(AnalysisUsage &AU) const override;
392 bool runOnModule(Module &M) override;
393 void releaseMemory() override;
395 void print(raw_ostream &o, const Module *) const override;
399 //===----------------------------------------------------------------------===//
400 // GraphTraits specializations for call graphs so that they can be treated as
401 // graphs by the generic graph algorithms.
404 // Provide graph traits for tranversing call graphs using standard graph
406 template <> struct GraphTraits<CallGraphNode *> {
407 typedef CallGraphNode NodeType;
409 typedef CallGraphNode::CallRecord CGNPairTy;
410 typedef std::pointer_to_unary_function<CGNPairTy, CallGraphNode *>
413 static NodeType *getEntryNode(CallGraphNode *CGN) { return CGN; }
415 typedef mapped_iterator<NodeType::iterator, CGNDerefFun> ChildIteratorType;
417 static inline ChildIteratorType child_begin(NodeType *N) {
418 return map_iterator(N->begin(), CGNDerefFun(CGNDeref));
420 static inline ChildIteratorType child_end(NodeType *N) {
421 return map_iterator(N->end(), CGNDerefFun(CGNDeref));
424 static CallGraphNode *CGNDeref(CGNPairTy P) { return P.second; }
427 template <> struct GraphTraits<const CallGraphNode *> {
428 typedef const CallGraphNode NodeType;
430 typedef CallGraphNode::CallRecord CGNPairTy;
431 typedef std::pointer_to_unary_function<CGNPairTy, const CallGraphNode *>
434 static NodeType *getEntryNode(const CallGraphNode *CGN) { return CGN; }
436 typedef mapped_iterator<NodeType::const_iterator, CGNDerefFun>
439 static inline ChildIteratorType child_begin(NodeType *N) {
440 return map_iterator(N->begin(), CGNDerefFun(CGNDeref));
442 static inline ChildIteratorType child_end(NodeType *N) {
443 return map_iterator(N->end(), CGNDerefFun(CGNDeref));
446 static const CallGraphNode *CGNDeref(CGNPairTy P) { return P.second; }
450 struct GraphTraits<CallGraph *> : public GraphTraits<CallGraphNode *> {
451 static NodeType *getEntryNode(CallGraph *CGN) {
452 return CGN->getExternalCallingNode(); // Start at the external node!
454 typedef std::pair<const Function *const, std::unique_ptr<CallGraphNode>>
456 typedef std::pointer_to_unary_function<const PairTy &, CallGraphNode &>
459 // nodes_iterator/begin/end - Allow iteration over all nodes in the graph
460 typedef mapped_iterator<CallGraph::iterator, DerefFun> nodes_iterator;
461 static nodes_iterator nodes_begin(CallGraph *CG) {
462 return map_iterator(CG->begin(), DerefFun(CGdereference));
464 static nodes_iterator nodes_end(CallGraph *CG) {
465 return map_iterator(CG->end(), DerefFun(CGdereference));
468 static CallGraphNode &CGdereference(const PairTy &P) { return *P.second; }
472 struct GraphTraits<const CallGraph *> : public GraphTraits<
473 const CallGraphNode *> {
474 static NodeType *getEntryNode(const CallGraph *CGN) {
475 return CGN->getExternalCallingNode(); // Start at the external node!
477 typedef std::pair<const Function *const, std::unique_ptr<CallGraphNode>>
479 typedef std::pointer_to_unary_function<const PairTy &, const CallGraphNode &>
482 // nodes_iterator/begin/end - Allow iteration over all nodes in the graph
483 typedef mapped_iterator<CallGraph::const_iterator, DerefFun> nodes_iterator;
484 static nodes_iterator nodes_begin(const CallGraph *CG) {
485 return map_iterator(CG->begin(), DerefFun(CGdereference));
487 static nodes_iterator nodes_end(const CallGraph *CG) {
488 return map_iterator(CG->end(), DerefFun(CGdereference));
491 static const CallGraphNode &CGdereference(const PairTy &P) {
496 } // End llvm namespace