1//===- CallGraph.h - Build a Module's call 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/// \file 10/// 11/// This file provides interfaces used to build and manipulate a call graph, 12/// which is a very useful tool for interprocedural optimization. 13/// 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. 17/// 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. 29/// 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. 35/// 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. 39/// 40/// Because of these properties, the CallGraph captures a conservative superset 41/// of all of the caller-callee relationships, which is useful for 42/// transformations. 43/// 44//===----------------------------------------------------------------------===// 45 46#ifndef LLVM_ANALYSIS_CALLGRAPH_H 47#define LLVM_ANALYSIS_CALLGRAPH_H 48 49#include "llvm/ADT/GraphTraits.h" 50#include "llvm/ADT/STLExtras.h" 51#include "llvm/IR/CallSite.h" 52#include "llvm/IR/Function.h" 53#include "llvm/IR/Intrinsics.h" 54#include "llvm/IR/PassManager.h" 55#include "llvm/IR/ValueHandle.h" 56#include "llvm/Pass.h" 57#include <map> 58 59namespace llvm { 60 61class Function; 62class Module; 63class CallGraphNode; 64 65/// \brief The basic data container for the call graph of a \c Module of IR. 66/// 67/// This class exposes both the interface to the call graph for a module of IR. 68/// 69/// The core call graph itself can also be updated to reflect changes to the IR. 70class CallGraph { 71 Module &M; 72 73 typedef std::map<const Function *, std::unique_ptr<CallGraphNode>> 74 FunctionMapTy; 75 76 /// \brief A map from \c Function* to \c CallGraphNode*. 77 FunctionMapTy FunctionMap; 78 79 /// \brief This node has edges to all external functions and those internal 80 /// functions that have their address taken. 81 CallGraphNode *ExternalCallingNode; 82 83 /// \brief This node has edges to it from all functions making indirect calls 84 /// or calling an external function. 85 std::unique_ptr<CallGraphNode> CallsExternalNode; 86 87 /// \brief Replace the function represented by this node by another. 88 /// 89 /// This does not rescan the body of the function, so it is suitable when 90 /// splicing the body of one function to another while also updating all 91 /// callers from the old function to the new. 92 void spliceFunction(const Function *From, const Function *To); 93 94 /// \brief Add a function to the call graph, and link the node to all of the 95 /// functions that it calls. 96 void addToCallGraph(Function *F); 97 98public: 99 explicit CallGraph(Module &M); 100 CallGraph(CallGraph &&Arg); 101 ~CallGraph(); 102 103 void print(raw_ostream &OS) const; 104 void dump() const; 105 106 typedef FunctionMapTy::iterator iterator; 107 typedef FunctionMapTy::const_iterator const_iterator; 108 109 /// \brief Returns the module the call graph corresponds to. 110 Module &getModule() const { return M; } 111 112 inline iterator begin() { return FunctionMap.begin(); } 113 inline iterator end() { return FunctionMap.end(); } 114 inline const_iterator begin() const { return FunctionMap.begin(); } 115 inline const_iterator end() const { return FunctionMap.end(); } 116 117 /// \brief Returns the call graph node for the provided function. 118 inline const CallGraphNode *operator[](const Function *F) const { 119 const_iterator I = FunctionMap.find(F); 120 assert(I != FunctionMap.end() && "Function not in callgraph!"); 121 return I->second.get(); 122 } 123 124 /// \brief Returns the call graph node for the provided function. 125 inline CallGraphNode *operator[](const Function *F) { 126 const_iterator I = FunctionMap.find(F); 127 assert(I != FunctionMap.end() && "Function not in callgraph!"); 128 return I->second.get(); 129 } 130 131 /// \brief Returns the \c CallGraphNode which is used to represent 132 /// undetermined calls into the callgraph. 133 CallGraphNode *getExternalCallingNode() const { return ExternalCallingNode; } 134 135 CallGraphNode *getCallsExternalNode() const { 136 return CallsExternalNode.get(); 137 } 138 139 //===--------------------------------------------------------------------- 140 // Functions to keep a call graph up to date with a function that has been 141 // modified. 142 // 143 144 /// \brief Unlink the function from this module, returning it. 145 /// 146 /// Because this removes the function from the module, the call graph node is 147 /// destroyed. This is only valid if the function does not call any other 148 /// functions (ie, there are no edges in it's CGN). The easiest way to do 149 /// this is to dropAllReferences before calling this. 150 Function *removeFunctionFromModule(CallGraphNode *CGN); 151 152 /// \brief Similar to operator[], but this will insert a new CallGraphNode for 153 /// \c F if one does not already exist. 154 CallGraphNode *getOrInsertFunction(const Function *F); 155}; 156 157/// \brief A node in the call graph for a module. 158/// 159/// Typically represents a function in the call graph. There are also special 160/// "null" nodes used to represent theoretical entries in the call graph. 161class CallGraphNode { 162public: 163 /// \brief A pair of the calling instruction (a call or invoke) 164 /// and the call graph node being called. 165 typedef std::pair<WeakTrackingVH, CallGraphNode *> CallRecord; 166 167public: 168 typedef std::vector<CallRecord> CalledFunctionsVector; 169 170 /// \brief Creates a node for the specified function. 171 inline CallGraphNode(Function *F) : F(F), NumReferences(0) {} 172 173 ~CallGraphNode() { 174 assert(NumReferences == 0 && "Node deleted while references remain"); 175 } 176 177 typedef std::vector<CallRecord>::iterator iterator; 178 typedef std::vector<CallRecord>::const_iterator const_iterator; 179 180 /// \brief Returns the function that this call graph node represents. 181 Function *getFunction() const { return F; } 182 183 inline iterator begin() { return CalledFunctions.begin(); } 184 inline iterator end() { return CalledFunctions.end(); } 185 inline const_iterator begin() const { return CalledFunctions.begin(); } 186 inline const_iterator end() const { return CalledFunctions.end(); } 187 inline bool empty() const { return CalledFunctions.empty(); } 188 inline unsigned size() const { return (unsigned)CalledFunctions.size(); } 189 190 /// \brief Returns the number of other CallGraphNodes in this CallGraph that 191 /// reference this node in their callee list. 192 unsigned getNumReferences() const { return NumReferences; } 193 194 /// \brief Returns the i'th called function. 195 CallGraphNode *operator[](unsigned i) const { 196 assert(i < CalledFunctions.size() && "Invalid index"); 197 return CalledFunctions[i].second; 198 } 199 200 /// \brief Print out this call graph node. 201 void dump() const; 202 void print(raw_ostream &OS) const; 203 204 //===--------------------------------------------------------------------- 205 // Methods to keep a call graph up to date with a function that has been 206 // modified 207 // 208 209 /// \brief Removes all edges from this CallGraphNode to any functions it 210 /// calls. 211 void removeAllCalledFunctions() { 212 while (!CalledFunctions.empty()) { 213 CalledFunctions.back().second->DropRef(); 214 CalledFunctions.pop_back(); 215 } 216 } 217 218 /// \brief Moves all the callee information from N to this node. 219 void stealCalledFunctionsFrom(CallGraphNode *N) { 220 assert(CalledFunctions.empty() && 221 "Cannot steal callsite information if I already have some"); 222 std::swap(CalledFunctions, N->CalledFunctions); 223 } 224 225 /// \brief Adds a function to the list of functions called by this one. 226 void addCalledFunction(CallSite CS, CallGraphNode *M) { 227 assert(!CS.getInstruction() || !CS.getCalledFunction() || 228 !CS.getCalledFunction()->isIntrinsic() || 229 !Intrinsic::isLeaf(CS.getCalledFunction()->getIntrinsicID())); 230 CalledFunctions.emplace_back(CS.getInstruction(), M); 231 M->AddRef(); 232 } 233 234 void removeCallEdge(iterator I) { 235 I->second->DropRef(); 236 *I = CalledFunctions.back(); 237 CalledFunctions.pop_back(); 238 } 239 240 /// \brief Removes the edge in the node for the specified call site. 241 /// 242 /// Note that this method takes linear time, so it should be used sparingly. 243 void removeCallEdgeFor(CallSite CS); 244 245 /// \brief Removes all call edges from this node to the specified callee 246 /// function. 247 /// 248 /// This takes more time to execute than removeCallEdgeTo, so it should not 249 /// be used unless necessary. 250 void removeAnyCallEdgeTo(CallGraphNode *Callee); 251 252 /// \brief Removes one edge associated with a null callsite from this node to 253 /// the specified callee function. 254 void removeOneAbstractEdgeTo(CallGraphNode *Callee); 255 256 /// \brief Replaces the edge in the node for the specified call site with a 257 /// new one. 258 /// 259 /// Note that this method takes linear time, so it should be used sparingly. 260 void replaceCallEdge(CallSite CS, CallSite NewCS, CallGraphNode *NewNode); 261 262private: 263 friend class CallGraph; 264 265 Function *F; 266 267 std::vector<CallRecord> CalledFunctions; 268 269 /// \brief The number of times that this CallGraphNode occurs in the 270 /// CalledFunctions array of this or other CallGraphNodes. 271 unsigned NumReferences; 272 273 CallGraphNode(const CallGraphNode &) = delete; 274 void operator=(const CallGraphNode &) = delete; 275 276 void DropRef() { --NumReferences; } 277 void AddRef() { ++NumReferences; } 278 279 /// \brief A special function that should only be used by the CallGraph class. 280 void allReferencesDropped() { NumReferences = 0; } 281}; 282 283/// \brief An analysis pass to compute the \c CallGraph for a \c Module. 284/// 285/// This class implements the concept of an analysis pass used by the \c 286/// ModuleAnalysisManager to run an analysis over a module and cache the 287/// resulting data. 288class CallGraphAnalysis : public AnalysisInfoMixin<CallGraphAnalysis> { 289 friend AnalysisInfoMixin<CallGraphAnalysis>; 290 static AnalysisKey Key; 291 292public: 293 /// \brief A formulaic typedef to inform clients of the result type. 294 typedef CallGraph Result; 295 296 /// \brief Compute the \c CallGraph for the module \c M. 297 /// 298 /// The real work here is done in the \c CallGraph constructor. 299 CallGraph run(Module &M, ModuleAnalysisManager &) { return CallGraph(M); } 300}; 301 302/// \brief Printer pass for the \c CallGraphAnalysis results. 303class CallGraphPrinterPass : public PassInfoMixin<CallGraphPrinterPass> { 304 raw_ostream &OS; 305 306public: 307 explicit CallGraphPrinterPass(raw_ostream &OS) : OS(OS) {} 308 PreservedAnalyses run(Module &M, ModuleAnalysisManager &AM); 309}; 310 311/// \brief The \c ModulePass which wraps up a \c CallGraph and the logic to 312/// build it. 313/// 314/// This class exposes both the interface to the call graph container and the 315/// module pass which runs over a module of IR and produces the call graph. The 316/// call graph interface is entirelly a wrapper around a \c CallGraph object 317/// which is stored internally for each module. 318class CallGraphWrapperPass : public ModulePass { 319 std::unique_ptr<CallGraph> G; 320 321public: 322 static char ID; // Class identification, replacement for typeinfo 323 324 CallGraphWrapperPass(); 325 ~CallGraphWrapperPass() override; 326 327 /// \brief The internal \c CallGraph around which the rest of this interface 328 /// is wrapped. 329 const CallGraph &getCallGraph() const { return *G; } 330 CallGraph &getCallGraph() { return *G; } 331 332 typedef CallGraph::iterator iterator; 333 typedef CallGraph::const_iterator const_iterator; 334 335 /// \brief Returns the module the call graph corresponds to. 336 Module &getModule() const { return G->getModule(); } 337 338 inline iterator begin() { return G->begin(); } 339 inline iterator end() { return G->end(); } 340 inline const_iterator begin() const { return G->begin(); } 341 inline const_iterator end() const { return G->end(); } 342 343 /// \brief Returns the call graph node for the provided function. 344 inline const CallGraphNode *operator[](const Function *F) const { 345 return (*G)[F]; 346 } 347 348 /// \brief Returns the call graph node for the provided function. 349 inline CallGraphNode *operator[](const Function *F) { return (*G)[F]; } 350 351 /// \brief Returns the \c CallGraphNode which is used to represent 352 /// undetermined calls into the callgraph. 353 CallGraphNode *getExternalCallingNode() const { 354 return G->getExternalCallingNode(); 355 } 356 357 CallGraphNode *getCallsExternalNode() const { 358 return G->getCallsExternalNode(); 359 } 360 361 //===--------------------------------------------------------------------- 362 // Functions to keep a call graph up to date with a function that has been 363 // modified. 364 // 365 366 /// \brief Unlink the function from this module, returning it. 367 /// 368 /// Because this removes the function from the module, the call graph node is 369 /// destroyed. This is only valid if the function does not call any other 370 /// functions (ie, there are no edges in it's CGN). The easiest way to do 371 /// this is to dropAllReferences before calling this. 372 Function *removeFunctionFromModule(CallGraphNode *CGN) { 373 return G->removeFunctionFromModule(CGN); 374 } 375 376 /// \brief Similar to operator[], but this will insert a new CallGraphNode for 377 /// \c F if one does not already exist. 378 CallGraphNode *getOrInsertFunction(const Function *F) { 379 return G->getOrInsertFunction(F); 380 } 381 382 //===--------------------------------------------------------------------- 383 // Implementation of the ModulePass interface needed here. 384 // 385 386 void getAnalysisUsage(AnalysisUsage &AU) const override; 387 bool runOnModule(Module &M) override; 388 void releaseMemory() override; 389 390 void print(raw_ostream &o, const Module *) const override; 391 void dump() const; 392}; 393 394//===----------------------------------------------------------------------===// 395// GraphTraits specializations for call graphs so that they can be treated as 396// graphs by the generic graph algorithms. 397// 398 399// Provide graph traits for tranversing call graphs using standard graph 400// traversals. 401template <> struct GraphTraits<CallGraphNode *> { 402 typedef CallGraphNode *NodeRef; 403 404 typedef CallGraphNode::CallRecord CGNPairTy; 405 406 static NodeRef getEntryNode(CallGraphNode *CGN) { return CGN; } 407 408 static CallGraphNode *CGNGetValue(CGNPairTy P) { return P.second; } 409 410 typedef mapped_iterator<CallGraphNode::iterator, decltype(&CGNGetValue)> 411 ChildIteratorType; 412 413 static ChildIteratorType child_begin(NodeRef N) { 414 return ChildIteratorType(N->begin(), &CGNGetValue); 415 } 416 static ChildIteratorType child_end(NodeRef N) { 417 return ChildIteratorType(N->end(), &CGNGetValue); 418 } 419}; 420 421template <> struct GraphTraits<const CallGraphNode *> { 422 typedef const CallGraphNode *NodeRef; 423 424 typedef CallGraphNode::CallRecord CGNPairTy; 425 426 static NodeRef getEntryNode(const CallGraphNode *CGN) { return CGN; } 427 428 static const CallGraphNode *CGNGetValue(CGNPairTy P) { return P.second; } 429 430 typedef mapped_iterator<CallGraphNode::const_iterator, decltype(&CGNGetValue)> 431 ChildIteratorType; 432 433 static ChildIteratorType child_begin(NodeRef N) { 434 return ChildIteratorType(N->begin(), &CGNGetValue); 435 } 436 static ChildIteratorType child_end(NodeRef N) { 437 return ChildIteratorType(N->end(), &CGNGetValue); 438 } 439}; 440 441template <> 442struct GraphTraits<CallGraph *> : public GraphTraits<CallGraphNode *> { 443 static NodeRef getEntryNode(CallGraph *CGN) { 444 return CGN->getExternalCallingNode(); // Start at the external node! 445 } 446 typedef std::pair<const Function *const, std::unique_ptr<CallGraphNode>> 447 PairTy; 448 static CallGraphNode *CGGetValuePtr(const PairTy &P) { 449 return P.second.get(); 450 } 451 452 // nodes_iterator/begin/end - Allow iteration over all nodes in the graph 453 typedef mapped_iterator<CallGraph::iterator, decltype(&CGGetValuePtr)> 454 nodes_iterator; 455 static nodes_iterator nodes_begin(CallGraph *CG) { 456 return nodes_iterator(CG->begin(), &CGGetValuePtr); 457 } 458 static nodes_iterator nodes_end(CallGraph *CG) { 459 return nodes_iterator(CG->end(), &CGGetValuePtr); 460 } 461}; 462 463template <> 464struct GraphTraits<const CallGraph *> : public GraphTraits< 465 const CallGraphNode *> { 466 static NodeRef getEntryNode(const CallGraph *CGN) { 467 return CGN->getExternalCallingNode(); // Start at the external node! 468 } 469 typedef std::pair<const Function *const, std::unique_ptr<CallGraphNode>> 470 PairTy; 471 static const CallGraphNode *CGGetValuePtr(const PairTy &P) { 472 return P.second.get(); 473 } 474 475 // nodes_iterator/begin/end - Allow iteration over all nodes in the graph 476 typedef mapped_iterator<CallGraph::const_iterator, decltype(&CGGetValuePtr)> 477 nodes_iterator; 478 static nodes_iterator nodes_begin(const CallGraph *CG) { 479 return nodes_iterator(CG->begin(), &CGGetValuePtr); 480 } 481 static nodes_iterator nodes_end(const CallGraph *CG) { 482 return nodes_iterator(CG->end(), &CGGetValuePtr); 483 } 484}; 485 486} // End llvm namespace 487 488#endif 489