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 <cassert> 58#include <map> 59#include <memory> 60#include <utility> 61#include <vector> 62 63namespace llvm { 64 65class CallGraphNode; 66class Module; 67class raw_ostream; 68 69/// \brief The basic data container for the call graph of a \c Module of IR. 70/// 71/// This class exposes both the interface to the call graph for a module of IR. 72/// 73/// The core call graph itself can also be updated to reflect changes to the IR. 74class CallGraph { 75 Module &M; 76 77 using FunctionMapTy = 78 std::map<const Function *, std::unique_ptr<CallGraphNode>>; 79 80 /// \brief A map from \c Function* to \c CallGraphNode*. 81 FunctionMapTy FunctionMap; 82 83 /// \brief This node has edges to all external functions and those internal 84 /// functions that have their address taken. 85 CallGraphNode *ExternalCallingNode; 86 87 /// \brief This node has edges to it from all functions making indirect calls 88 /// or calling an external function. 89 std::unique_ptr<CallGraphNode> CallsExternalNode; 90 91 /// \brief Replace the function represented by this node by another. 92 /// 93 /// This does not rescan the body of the function, so it is suitable when 94 /// splicing the body of one function to another while also updating all 95 /// callers from the old function to the new. 96 void spliceFunction(const Function *From, const Function *To); 97 98 /// \brief Add a function to the call graph, and link the node to all of the 99 /// functions that it calls. 100 void addToCallGraph(Function *F); 101 102public: 103 explicit CallGraph(Module &M); 104 CallGraph(CallGraph &&Arg); 105 ~CallGraph(); 106 107 void print(raw_ostream &OS) const; 108 void dump() const; 109 110 using iterator = FunctionMapTy::iterator; 111 using const_iterator = FunctionMapTy::const_iterator; 112 113 /// \brief Returns the module the call graph corresponds to. 114 Module &getModule() const { return M; } 115 116 inline iterator begin() { return FunctionMap.begin(); } 117 inline iterator end() { return FunctionMap.end(); } 118 inline const_iterator begin() const { return FunctionMap.begin(); } 119 inline const_iterator end() const { return FunctionMap.end(); } 120 121 /// \brief Returns the call graph node for the provided function. 122 inline const CallGraphNode *operator[](const Function *F) const { 123 const_iterator I = FunctionMap.find(F); 124 assert(I != FunctionMap.end() && "Function not in callgraph!"); 125 return I->second.get(); 126 } 127 128 /// \brief Returns the call graph node for the provided function. 129 inline CallGraphNode *operator[](const Function *F) { 130 const_iterator I = FunctionMap.find(F); 131 assert(I != FunctionMap.end() && "Function not in callgraph!"); 132 return I->second.get(); 133 } 134 135 /// \brief Returns the \c CallGraphNode which is used to represent 136 /// undetermined calls into the callgraph. 137 CallGraphNode *getExternalCallingNode() const { return ExternalCallingNode; } 138 139 CallGraphNode *getCallsExternalNode() const { 140 return CallsExternalNode.get(); 141 } 142 143 //===--------------------------------------------------------------------- 144 // Functions to keep a call graph up to date with a function that has been 145 // modified. 146 // 147 148 /// \brief Unlink the function from this module, returning it. 149 /// 150 /// Because this removes the function from the module, the call graph node is 151 /// destroyed. This is only valid if the function does not call any other 152 /// functions (ie, there are no edges in it's CGN). The easiest way to do 153 /// this is to dropAllReferences before calling this. 154 Function *removeFunctionFromModule(CallGraphNode *CGN); 155 156 /// \brief Similar to operator[], but this will insert a new CallGraphNode for 157 /// \c F if one does not already exist. 158 CallGraphNode *getOrInsertFunction(const Function *F); 159}; 160 161/// \brief A node in the call graph for a module. 162/// 163/// Typically represents a function in the call graph. There are also special 164/// "null" nodes used to represent theoretical entries in the call graph. 165class CallGraphNode { 166public: 167 /// \brief A pair of the calling instruction (a call or invoke) 168 /// and the call graph node being called. 169 using CallRecord = std::pair<WeakTrackingVH, CallGraphNode *>; 170 171public: 172 using CalledFunctionsVector = std::vector<CallRecord>; 173 174 /// \brief Creates a node for the specified function. 175 inline CallGraphNode(Function *F) : F(F) {} 176 177 CallGraphNode(const CallGraphNode &) = delete; 178 CallGraphNode &operator=(const CallGraphNode &) = delete; 179 180 ~CallGraphNode() { 181 assert(NumReferences == 0 && "Node deleted while references remain"); 182 } 183 184 using iterator = std::vector<CallRecord>::iterator; 185 using const_iterator = std::vector<CallRecord>::const_iterator; 186 187 /// \brief Returns the function that this call graph node represents. 188 Function *getFunction() const { return F; } 189 190 inline iterator begin() { return CalledFunctions.begin(); } 191 inline iterator end() { return CalledFunctions.end(); } 192 inline const_iterator begin() const { return CalledFunctions.begin(); } 193 inline const_iterator end() const { return CalledFunctions.end(); } 194 inline bool empty() const { return CalledFunctions.empty(); } 195 inline unsigned size() const { return (unsigned)CalledFunctions.size(); } 196 197 /// \brief Returns the number of other CallGraphNodes in this CallGraph that 198 /// reference this node in their callee list. 199 unsigned getNumReferences() const { return NumReferences; } 200 201 /// \brief Returns the i'th called function. 202 CallGraphNode *operator[](unsigned i) const { 203 assert(i < CalledFunctions.size() && "Invalid index"); 204 return CalledFunctions[i].second; 205 } 206 207 /// \brief Print out this call graph node. 208 void dump() const; 209 void print(raw_ostream &OS) const; 210 211 //===--------------------------------------------------------------------- 212 // Methods to keep a call graph up to date with a function that has been 213 // modified 214 // 215 216 /// \brief Removes all edges from this CallGraphNode to any functions it 217 /// calls. 218 void removeAllCalledFunctions() { 219 while (!CalledFunctions.empty()) { 220 CalledFunctions.back().second->DropRef(); 221 CalledFunctions.pop_back(); 222 } 223 } 224 225 /// \brief Moves all the callee information from N to this node. 226 void stealCalledFunctionsFrom(CallGraphNode *N) { 227 assert(CalledFunctions.empty() && 228 "Cannot steal callsite information if I already have some"); 229 std::swap(CalledFunctions, N->CalledFunctions); 230 } 231 232 /// \brief Adds a function to the list of functions called by this one. 233 void addCalledFunction(CallSite CS, CallGraphNode *M) { 234 assert(!CS.getInstruction() || !CS.getCalledFunction() || 235 !CS.getCalledFunction()->isIntrinsic() || 236 !Intrinsic::isLeaf(CS.getCalledFunction()->getIntrinsicID())); 237 CalledFunctions.emplace_back(CS.getInstruction(), M); 238 M->AddRef(); 239 } 240 241 void removeCallEdge(iterator I) { 242 I->second->DropRef(); 243 *I = CalledFunctions.back(); 244 CalledFunctions.pop_back(); 245 } 246 247 /// \brief Removes the edge in the node for the specified call site. 248 /// 249 /// Note that this method takes linear time, so it should be used sparingly. 250 void removeCallEdgeFor(CallSite CS); 251 252 /// \brief Removes all call edges from this node to the specified callee 253 /// function. 254 /// 255 /// This takes more time to execute than removeCallEdgeTo, so it should not 256 /// be used unless necessary. 257 void removeAnyCallEdgeTo(CallGraphNode *Callee); 258 259 /// \brief Removes one edge associated with a null callsite from this node to 260 /// the specified callee function. 261 void removeOneAbstractEdgeTo(CallGraphNode *Callee); 262 263 /// \brief Replaces the edge in the node for the specified call site with a 264 /// new one. 265 /// 266 /// Note that this method takes linear time, so it should be used sparingly. 267 void replaceCallEdge(CallSite CS, CallSite NewCS, CallGraphNode *NewNode); 268 269private: 270 friend class CallGraph; 271 272 Function *F; 273 274 std::vector<CallRecord> CalledFunctions; 275 276 /// \brief The number of times that this CallGraphNode occurs in the 277 /// CalledFunctions array of this or other CallGraphNodes. 278 unsigned NumReferences = 0; 279 280 void DropRef() { --NumReferences; } 281 void AddRef() { ++NumReferences; } 282 283 /// \brief A special function that should only be used by the CallGraph class. 284 void allReferencesDropped() { NumReferences = 0; } 285}; 286 287/// \brief An analysis pass to compute the \c CallGraph for a \c Module. 288/// 289/// This class implements the concept of an analysis pass used by the \c 290/// ModuleAnalysisManager to run an analysis over a module and cache the 291/// resulting data. 292class CallGraphAnalysis : public AnalysisInfoMixin<CallGraphAnalysis> { 293 friend AnalysisInfoMixin<CallGraphAnalysis>; 294 295 static AnalysisKey Key; 296 297public: 298 /// \brief A formulaic type to inform clients of the result type. 299 using Result = CallGraph; 300 301 /// \brief Compute the \c CallGraph for the module \c M. 302 /// 303 /// The real work here is done in the \c CallGraph constructor. 304 CallGraph run(Module &M, ModuleAnalysisManager &) { return CallGraph(M); } 305}; 306 307/// \brief Printer pass for the \c CallGraphAnalysis results. 308class CallGraphPrinterPass : public PassInfoMixin<CallGraphPrinterPass> { 309 raw_ostream &OS; 310 311public: 312 explicit CallGraphPrinterPass(raw_ostream &OS) : OS(OS) {} 313 314 PreservedAnalyses run(Module &M, ModuleAnalysisManager &AM); 315}; 316 317/// \brief The \c ModulePass which wraps up a \c CallGraph and the logic to 318/// build it. 319/// 320/// This class exposes both the interface to the call graph container and the 321/// module pass which runs over a module of IR and produces the call graph. The 322/// call graph interface is entirelly a wrapper around a \c CallGraph object 323/// which is stored internally for each module. 324class CallGraphWrapperPass : public ModulePass { 325 std::unique_ptr<CallGraph> G; 326 327public: 328 static char ID; // Class identification, replacement for typeinfo 329 330 CallGraphWrapperPass(); 331 ~CallGraphWrapperPass() override; 332 333 /// \brief The internal \c CallGraph around which the rest of this interface 334 /// is wrapped. 335 const CallGraph &getCallGraph() const { return *G; } 336 CallGraph &getCallGraph() { return *G; } 337 338 using iterator = CallGraph::iterator; 339 using const_iterator = CallGraph::const_iterator; 340 341 /// \brief Returns the module the call graph corresponds to. 342 Module &getModule() const { return G->getModule(); } 343 344 inline iterator begin() { return G->begin(); } 345 inline iterator end() { return G->end(); } 346 inline const_iterator begin() const { return G->begin(); } 347 inline const_iterator end() const { return G->end(); } 348 349 /// \brief Returns the call graph node for the provided function. 350 inline const CallGraphNode *operator[](const Function *F) const { 351 return (*G)[F]; 352 } 353 354 /// \brief Returns the call graph node for the provided function. 355 inline CallGraphNode *operator[](const Function *F) { return (*G)[F]; } 356 357 /// \brief Returns the \c CallGraphNode which is used to represent 358 /// undetermined calls into the callgraph. 359 CallGraphNode *getExternalCallingNode() const { 360 return G->getExternalCallingNode(); 361 } 362 363 CallGraphNode *getCallsExternalNode() const { 364 return G->getCallsExternalNode(); 365 } 366 367 //===--------------------------------------------------------------------- 368 // Functions to keep a call graph up to date with a function that has been 369 // modified. 370 // 371 372 /// \brief Unlink the function from this module, returning it. 373 /// 374 /// Because this removes the function from the module, the call graph node is 375 /// destroyed. This is only valid if the function does not call any other 376 /// functions (ie, there are no edges in it's CGN). The easiest way to do 377 /// this is to dropAllReferences before calling this. 378 Function *removeFunctionFromModule(CallGraphNode *CGN) { 379 return G->removeFunctionFromModule(CGN); 380 } 381 382 /// \brief Similar to operator[], but this will insert a new CallGraphNode for 383 /// \c F if one does not already exist. 384 CallGraphNode *getOrInsertFunction(const Function *F) { 385 return G->getOrInsertFunction(F); 386 } 387 388 //===--------------------------------------------------------------------- 389 // Implementation of the ModulePass interface needed here. 390 // 391 392 void getAnalysisUsage(AnalysisUsage &AU) const override; 393 bool runOnModule(Module &M) override; 394 void releaseMemory() override; 395 396 void print(raw_ostream &o, const Module *) const override; 397 void dump() const; 398}; 399 400//===----------------------------------------------------------------------===// 401// GraphTraits specializations for call graphs so that they can be treated as 402// graphs by the generic graph algorithms. 403// 404 405// Provide graph traits for tranversing call graphs using standard graph 406// traversals. 407template <> struct GraphTraits<CallGraphNode *> { 408 using NodeRef = CallGraphNode *; 409 using CGNPairTy = CallGraphNode::CallRecord; 410 411 static NodeRef getEntryNode(CallGraphNode *CGN) { return CGN; } 412 static CallGraphNode *CGNGetValue(CGNPairTy P) { return P.second; } 413 414 using ChildIteratorType = 415 mapped_iterator<CallGraphNode::iterator, decltype(&CGNGetValue)>; 416 417 static ChildIteratorType child_begin(NodeRef N) { 418 return ChildIteratorType(N->begin(), &CGNGetValue); 419 } 420 421 static ChildIteratorType child_end(NodeRef N) { 422 return ChildIteratorType(N->end(), &CGNGetValue); 423 } 424}; 425 426template <> struct GraphTraits<const CallGraphNode *> { 427 using NodeRef = const CallGraphNode *; 428 using CGNPairTy = CallGraphNode::CallRecord; 429 430 static NodeRef getEntryNode(const CallGraphNode *CGN) { return CGN; } 431 static const CallGraphNode *CGNGetValue(CGNPairTy P) { return P.second; } 432 433 using ChildIteratorType = 434 mapped_iterator<CallGraphNode::const_iterator, decltype(&CGNGetValue)>; 435 436 static ChildIteratorType child_begin(NodeRef N) { 437 return ChildIteratorType(N->begin(), &CGNGetValue); 438 } 439 440 static ChildIteratorType child_end(NodeRef N) { 441 return ChildIteratorType(N->end(), &CGNGetValue); 442 } 443}; 444 445template <> 446struct GraphTraits<CallGraph *> : public GraphTraits<CallGraphNode *> { 447 using PairTy = 448 std::pair<const Function *const, std::unique_ptr<CallGraphNode>>; 449 450 static NodeRef getEntryNode(CallGraph *CGN) { 451 return CGN->getExternalCallingNode(); // Start at the external node! 452 } 453 454 static CallGraphNode *CGGetValuePtr(const PairTy &P) { 455 return P.second.get(); 456 } 457 458 // nodes_iterator/begin/end - Allow iteration over all nodes in the graph 459 using nodes_iterator = 460 mapped_iterator<CallGraph::iterator, decltype(&CGGetValuePtr)>; 461 462 static nodes_iterator nodes_begin(CallGraph *CG) { 463 return nodes_iterator(CG->begin(), &CGGetValuePtr); 464 } 465 466 static nodes_iterator nodes_end(CallGraph *CG) { 467 return nodes_iterator(CG->end(), &CGGetValuePtr); 468 } 469}; 470 471template <> 472struct GraphTraits<const CallGraph *> : public GraphTraits< 473 const CallGraphNode *> { 474 using PairTy = 475 std::pair<const Function *const, std::unique_ptr<CallGraphNode>>; 476 477 static NodeRef getEntryNode(const CallGraph *CGN) { 478 return CGN->getExternalCallingNode(); // Start at the external node! 479 } 480 481 static const CallGraphNode *CGGetValuePtr(const PairTy &P) { 482 return P.second.get(); 483 } 484 485 // nodes_iterator/begin/end - Allow iteration over all nodes in the graph 486 using nodes_iterator = 487 mapped_iterator<CallGraph::const_iterator, decltype(&CGGetValuePtr)>; 488 489 static nodes_iterator nodes_begin(const CallGraph *CG) { 490 return nodes_iterator(CG->begin(), &CGGetValuePtr); 491 } 492 493 static nodes_iterator nodes_end(const CallGraph *CG) { 494 return nodes_iterator(CG->end(), &CGGetValuePtr); 495 } 496}; 497 498} // end namespace llvm 499 500#endif // LLVM_ANALYSIS_CALLGRAPH_H 501