1 //===- LazyCallGraph.h - Analysis of 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 /// Implements a lazy call graph analysis and related passes for the new pass 12 /// manager. 13 /// 14 /// NB: This is *not* a traditional call graph! It is a graph which models both 15 /// the current calls and potential calls. As a consequence there are many 16 /// edges in this call graph that do not correspond to a 'call' or 'invoke' 17 /// instruction. 18 /// 19 /// The primary use cases of this graph analysis is to facilitate iterating 20 /// across the functions of a module in ways that ensure all callees are 21 /// visited prior to a caller (given any SCC constraints), or vice versa. As 22 /// such is it particularly well suited to organizing CGSCC optimizations such 23 /// as inlining, outlining, argument promotion, etc. That is its primary use 24 /// case and motivates the design. It may not be appropriate for other 25 /// purposes. The use graph of functions or some other conservative analysis of 26 /// call instructions may be interesting for optimizations and subsequent 27 /// analyses which don't work in the context of an overly specified 28 /// potential-call-edge graph. 29 /// 30 /// To understand the specific rules and nature of this call graph analysis, 31 /// see the documentation of the \c LazyCallGraph below. 32 /// 33 //===----------------------------------------------------------------------===// 34 35 #ifndef LLVM_ANALYSIS_LAZY_CALL_GRAPH 36 #define LLVM_ANALYSIS_LAZY_CALL_GRAPH 37 38 #include "llvm/ADT/DenseMap.h" 39 #include "llvm/ADT/PointerUnion.h" 40 #include "llvm/ADT/STLExtras.h" 41 #include "llvm/ADT/SetVector.h" 42 #include "llvm/ADT/SmallPtrSet.h" 43 #include "llvm/ADT/SmallVector.h" 44 #include "llvm/ADT/iterator.h" 45 #include "llvm/ADT/iterator_range.h" 46 #include "llvm/IR/BasicBlock.h" 47 #include "llvm/IR/Function.h" 48 #include "llvm/IR/Module.h" 49 #include "llvm/Support/Allocator.h" 50 #include <iterator> 51 52 namespace llvm { 53 class ModuleAnalysisManager; 54 class PreservedAnalyses; 55 class raw_ostream; 56 57 /// \brief A lazily constructed view of the call graph of a module. 58 /// 59 /// With the edges of this graph, the motivating constraint that we are 60 /// attempting to maintain is that function-local optimization, CGSCC-local 61 /// optimizations, and optimizations transforming a pair of functions connected 62 /// by an edge in the graph, do not invalidate a bottom-up traversal of the SCC 63 /// DAG. That is, no optimizations will delete, remove, or add an edge such 64 /// that functions already visited in a bottom-up order of the SCC DAG are no 65 /// longer valid to have visited, or such that functions not yet visited in 66 /// a bottom-up order of the SCC DAG are not required to have already been 67 /// visited. 68 /// 69 /// Within this constraint, the desire is to minimize the merge points of the 70 /// SCC DAG. The greater the fanout of the SCC DAG and the fewer merge points 71 /// in the SCC DAG, the more independence there is in optimizing within it. 72 /// There is a strong desire to enable parallelization of optimizations over 73 /// the call graph, and both limited fanout and merge points will (artificially 74 /// in some cases) limit the scaling of such an effort. 75 /// 76 /// To this end, graph represents both direct and any potential resolution to 77 /// an indirect call edge. Another way to think about it is that it represents 78 /// both the direct call edges and any direct call edges that might be formed 79 /// through static optimizations. Specifically, it considers taking the address 80 /// of a function to be an edge in the call graph because this might be 81 /// forwarded to become a direct call by some subsequent function-local 82 /// optimization. The result is that the graph closely follows the use-def 83 /// edges for functions. Walking "up" the graph can be done by looking at all 84 /// of the uses of a function. 85 /// 86 /// The roots of the call graph are the external functions and functions 87 /// escaped into global variables. Those functions can be called from outside 88 /// of the module or via unknowable means in the IR -- we may not be able to 89 /// form even a potential call edge from a function body which may dynamically 90 /// load the function and call it. 91 /// 92 /// This analysis still requires updates to remain valid after optimizations 93 /// which could potentially change the set of potential callees. The 94 /// constraints it operates under only make the traversal order remain valid. 95 /// 96 /// The entire analysis must be re-computed if full interprocedural 97 /// optimizations run at any point. For example, globalopt completely 98 /// invalidates the information in this analysis. 99 /// 100 /// FIXME: This class is named LazyCallGraph in a lame attempt to distinguish 101 /// it from the existing CallGraph. At some point, it is expected that this 102 /// will be the only call graph and it will be renamed accordingly. 103 class LazyCallGraph { 104 public: 105 class Node; 106 class SCC; 107 typedef SmallVector<PointerUnion<Function *, Node *>, 4> NodeVectorT; 108 typedef SmallVectorImpl<PointerUnion<Function *, Node *>> NodeVectorImplT; 109 110 /// \brief A lazy iterator used for both the entry nodes and child nodes. 111 /// 112 /// When this iterator is dereferenced, if not yet available, a function will 113 /// be scanned for "calls" or uses of functions and its child information 114 /// will be constructed. All of these results are accumulated and cached in 115 /// the graph. 116 class iterator 117 : public iterator_adaptor_base<iterator, NodeVectorImplT::iterator, 118 std::forward_iterator_tag, Node> { 119 friend class LazyCallGraph; 120 friend class LazyCallGraph::Node; 121 122 LazyCallGraph *G; 123 NodeVectorImplT::iterator E; 124 125 // Build the iterator for a specific position in a node list. iterator(LazyCallGraph & G,NodeVectorImplT::iterator NI,NodeVectorImplT::iterator E)126 iterator(LazyCallGraph &G, NodeVectorImplT::iterator NI, 127 NodeVectorImplT::iterator E) 128 : iterator_adaptor_base(NI), G(&G), E(E) { 129 while (I != E && I->isNull()) 130 ++I; 131 } 132 133 public: iterator()134 iterator() {} 135 136 using iterator_adaptor_base::operator++; 137 iterator &operator++() { 138 do { 139 ++I; 140 } while (I != E && I->isNull()); 141 return *this; 142 } 143 144 reference operator*() const { 145 if (I->is<Node *>()) 146 return *I->get<Node *>(); 147 148 Function *F = I->get<Function *>(); 149 Node &ChildN = G->get(*F); 150 *I = &ChildN; 151 return ChildN; 152 } 153 }; 154 155 /// \brief A node in the call graph. 156 /// 157 /// This represents a single node. It's primary roles are to cache the list of 158 /// callees, de-duplicate and provide fast testing of whether a function is 159 /// a callee, and facilitate iteration of child nodes in the graph. 160 class Node { 161 friend class LazyCallGraph; 162 friend class LazyCallGraph::SCC; 163 164 LazyCallGraph *G; 165 Function &F; 166 167 // We provide for the DFS numbering and Tarjan walk lowlink numbers to be 168 // stored directly within the node. 169 int DFSNumber; 170 int LowLink; 171 172 mutable NodeVectorT Callees; 173 DenseMap<Function *, size_t> CalleeIndexMap; 174 175 /// \brief Basic constructor implements the scanning of F into Callees and 176 /// CalleeIndexMap. 177 Node(LazyCallGraph &G, Function &F); 178 179 /// \brief Internal helper to insert a callee. 180 void insertEdgeInternal(Function &Callee); 181 182 /// \brief Internal helper to insert a callee. 183 void insertEdgeInternal(Node &CalleeN); 184 185 /// \brief Internal helper to remove a callee from this node. 186 void removeEdgeInternal(Function &Callee); 187 188 public: 189 typedef LazyCallGraph::iterator iterator; 190 getFunction()191 Function &getFunction() const { 192 return F; 193 }; 194 begin()195 iterator begin() const { 196 return iterator(*G, Callees.begin(), Callees.end()); 197 } end()198 iterator end() const { return iterator(*G, Callees.end(), Callees.end()); } 199 200 /// Equality is defined as address equality. 201 bool operator==(const Node &N) const { return this == &N; } 202 bool operator!=(const Node &N) const { return !operator==(N); } 203 }; 204 205 /// \brief An SCC of the call graph. 206 /// 207 /// This represents a Strongly Connected Component of the call graph as 208 /// a collection of call graph nodes. While the order of nodes in the SCC is 209 /// stable, it is not any particular order. 210 class SCC { 211 friend class LazyCallGraph; 212 friend class LazyCallGraph::Node; 213 214 LazyCallGraph *G; 215 SmallPtrSet<SCC *, 1> ParentSCCs; 216 SmallVector<Node *, 1> Nodes; 217 SCC(LazyCallGraph & G)218 SCC(LazyCallGraph &G) : G(&G) {} 219 220 void insert(Node &N); 221 222 void 223 internalDFS(SmallVectorImpl<std::pair<Node *, Node::iterator>> &DFSStack, 224 SmallVectorImpl<Node *> &PendingSCCStack, Node *N, 225 SmallVectorImpl<SCC *> &ResultSCCs); 226 227 public: 228 typedef SmallVectorImpl<Node *>::const_iterator iterator; 229 typedef pointee_iterator<SmallPtrSet<SCC *, 1>::const_iterator> parent_iterator; 230 begin()231 iterator begin() const { return Nodes.begin(); } end()232 iterator end() const { return Nodes.end(); } 233 parent_begin()234 parent_iterator parent_begin() const { return ParentSCCs.begin(); } parent_end()235 parent_iterator parent_end() const { return ParentSCCs.end(); } 236 parents()237 iterator_range<parent_iterator> parents() const { 238 return iterator_range<parent_iterator>(parent_begin(), parent_end()); 239 } 240 241 /// \brief Test if this SCC is a parent of \a C. isParentOf(const SCC & C)242 bool isParentOf(const SCC &C) const { return C.isChildOf(*this); } 243 244 /// \brief Test if this SCC is an ancestor of \a C. isAncestorOf(const SCC & C)245 bool isAncestorOf(const SCC &C) const { return C.isDescendantOf(*this); } 246 247 /// \brief Test if this SCC is a child of \a C. isChildOf(const SCC & C)248 bool isChildOf(const SCC &C) const { 249 return ParentSCCs.count(const_cast<SCC *>(&C)); 250 } 251 252 /// \brief Test if this SCC is a descendant of \a C. 253 bool isDescendantOf(const SCC &C) const; 254 255 ///@{ 256 /// \name Mutation API 257 /// 258 /// These methods provide the core API for updating the call graph in the 259 /// presence of a (potentially still in-flight) DFS-found SCCs. 260 /// 261 /// Note that these methods sometimes have complex runtimes, so be careful 262 /// how you call them. 263 264 /// \brief Insert an edge from one node in this SCC to another in this SCC. 265 /// 266 /// By the definition of an SCC, this does not change the nature or make-up 267 /// of any SCCs. 268 void insertIntraSCCEdge(Node &CallerN, Node &CalleeN); 269 270 /// \brief Insert an edge whose tail is in this SCC and head is in some 271 /// child SCC. 272 /// 273 /// There must be an existing path from the caller to the callee. This 274 /// operation is inexpensive and does not change the set of SCCs in the 275 /// graph. 276 void insertOutgoingEdge(Node &CallerN, Node &CalleeN); 277 278 /// \brief Insert an edge whose tail is in a descendant SCC and head is in 279 /// this SCC. 280 /// 281 /// There must be an existing path from the callee to the caller in this 282 /// case. NB! This is has the potential to be a very expensive function. It 283 /// inherently forms a cycle in the prior SCC DAG and we have to merge SCCs 284 /// to resolve that cycle. But finding all of the SCCs which participate in 285 /// the cycle can in the worst case require traversing every SCC in the 286 /// graph. Every attempt is made to avoid that, but passes must still 287 /// exercise caution calling this routine repeatedly. 288 /// 289 /// FIXME: We could possibly optimize this quite a bit for cases where the 290 /// caller and callee are very nearby in the graph. See comments in the 291 /// implementation for details, but that use case might impact users. 292 SmallVector<SCC *, 1> insertIncomingEdge(Node &CallerN, Node &CalleeN); 293 294 /// \brief Remove an edge whose source is in this SCC and target is *not*. 295 /// 296 /// This removes an inter-SCC edge. All inter-SCC edges originating from 297 /// this SCC have been fully explored by any in-flight DFS SCC formation, 298 /// so this is always safe to call once you have the source SCC. 299 /// 300 /// This operation does not change the set of SCCs or the members of the 301 /// SCCs and so is very inexpensive. It may change the connectivity graph 302 /// of the SCCs though, so be careful calling this while iterating over 303 /// them. 304 void removeInterSCCEdge(Node &CallerN, Node &CalleeN); 305 306 /// \brief Remove an edge which is entirely within this SCC. 307 /// 308 /// Both the \a Caller and the \a Callee must be within this SCC. Removing 309 /// such an edge make break cycles that form this SCC and thus this 310 /// operation may change the SCC graph significantly. In particular, this 311 /// operation will re-form new SCCs based on the remaining connectivity of 312 /// the graph. The following invariants are guaranteed to hold after 313 /// calling this method: 314 /// 315 /// 1) This SCC is still an SCC in the graph. 316 /// 2) This SCC will be the parent of any new SCCs. Thus, this SCC is 317 /// preserved as the root of any new SCC directed graph formed. 318 /// 3) No SCC other than this SCC has its member set changed (this is 319 /// inherent in the definition of removing such an edge). 320 /// 4) All of the parent links of the SCC graph will be updated to reflect 321 /// the new SCC structure. 322 /// 5) All SCCs formed out of this SCC, excluding this SCC, will be 323 /// returned in a vector. 324 /// 6) The order of the SCCs in the vector will be a valid postorder 325 /// traversal of the new SCCs. 326 /// 327 /// These invariants are very important to ensure that we can build 328 /// optimization pipeliens on top of the CGSCC pass manager which 329 /// intelligently update the SCC graph without invalidating other parts of 330 /// the SCC graph. 331 /// 332 /// The runtime complexity of this method is, in the worst case, O(V+E) 333 /// where V is the number of nodes in this SCC and E is the number of edges 334 /// leaving the nodes in this SCC. Note that E includes both edges within 335 /// this SCC and edges from this SCC to child SCCs. Some effort has been 336 /// made to minimize the overhead of common cases such as self-edges and 337 /// edge removals which result in a spanning tree with no more cycles. 338 SmallVector<SCC *, 1> removeIntraSCCEdge(Node &CallerN, Node &CalleeN); 339 340 ///@} 341 }; 342 343 /// \brief A post-order depth-first SCC iterator over the call graph. 344 /// 345 /// This iterator triggers the Tarjan DFS-based formation of the SCC DAG for 346 /// the call graph, walking it lazily in depth-first post-order. That is, it 347 /// always visits SCCs for a callee prior to visiting the SCC for a caller 348 /// (when they are in different SCCs). 349 class postorder_scc_iterator 350 : public iterator_facade_base<postorder_scc_iterator, 351 std::forward_iterator_tag, SCC> { 352 friend class LazyCallGraph; 353 friend class LazyCallGraph::Node; 354 355 /// \brief Nonce type to select the constructor for the end iterator. 356 struct IsAtEndT {}; 357 358 LazyCallGraph *G; 359 SCC *C; 360 361 // Build the begin iterator for a node. postorder_scc_iterator(LazyCallGraph & G)362 postorder_scc_iterator(LazyCallGraph &G) : G(&G) { 363 C = G.getNextSCCInPostOrder(); 364 } 365 366 // Build the end iterator for a node. This is selected purely by overload. postorder_scc_iterator(LazyCallGraph & G,IsAtEndT)367 postorder_scc_iterator(LazyCallGraph &G, IsAtEndT /*Nonce*/) 368 : G(&G), C(nullptr) {} 369 370 public: 371 bool operator==(const postorder_scc_iterator &Arg) const { 372 return G == Arg.G && C == Arg.C; 373 } 374 375 reference operator*() const { return *C; } 376 377 using iterator_facade_base::operator++; 378 postorder_scc_iterator &operator++() { 379 C = G->getNextSCCInPostOrder(); 380 return *this; 381 } 382 }; 383 384 /// \brief Construct a graph for the given module. 385 /// 386 /// This sets up the graph and computes all of the entry points of the graph. 387 /// No function definitions are scanned until their nodes in the graph are 388 /// requested during traversal. 389 LazyCallGraph(Module &M); 390 391 LazyCallGraph(LazyCallGraph &&G); 392 LazyCallGraph &operator=(LazyCallGraph &&RHS); 393 begin()394 iterator begin() { 395 return iterator(*this, EntryNodes.begin(), EntryNodes.end()); 396 } end()397 iterator end() { return iterator(*this, EntryNodes.end(), EntryNodes.end()); } 398 postorder_scc_begin()399 postorder_scc_iterator postorder_scc_begin() { 400 return postorder_scc_iterator(*this); 401 } postorder_scc_end()402 postorder_scc_iterator postorder_scc_end() { 403 return postorder_scc_iterator(*this, postorder_scc_iterator::IsAtEndT()); 404 } 405 postorder_sccs()406 iterator_range<postorder_scc_iterator> postorder_sccs() { 407 return iterator_range<postorder_scc_iterator>(postorder_scc_begin(), 408 postorder_scc_end()); 409 } 410 411 /// \brief Lookup a function in the graph which has already been scanned and 412 /// added. lookup(const Function & F)413 Node *lookup(const Function &F) const { return NodeMap.lookup(&F); } 414 415 /// \brief Lookup a function's SCC in the graph. 416 /// 417 /// \returns null if the function hasn't been assigned an SCC via the SCC 418 /// iterator walk. lookupSCC(Node & N)419 SCC *lookupSCC(Node &N) const { return SCCMap.lookup(&N); } 420 421 /// \brief Get a graph node for a given function, scanning it to populate the 422 /// graph data as necessary. get(Function & F)423 Node &get(Function &F) { 424 Node *&N = NodeMap[&F]; 425 if (N) 426 return *N; 427 428 return insertInto(F, N); 429 } 430 431 ///@{ 432 /// \name Pre-SCC Mutation API 433 /// 434 /// These methods are only valid to call prior to forming any SCCs for this 435 /// call graph. They can be used to update the core node-graph during 436 /// a node-based inorder traversal that precedes any SCC-based traversal. 437 /// 438 /// Once you begin manipulating a call graph's SCCs, you must perform all 439 /// mutation of the graph via the SCC methods. 440 441 /// \brief Update the call graph after inserting a new edge. 442 void insertEdge(Node &Caller, Function &Callee); 443 444 /// \brief Update the call graph after inserting a new edge. insertEdge(Function & Caller,Function & Callee)445 void insertEdge(Function &Caller, Function &Callee) { 446 return insertEdge(get(Caller), Callee); 447 } 448 449 /// \brief Update the call graph after deleting an edge. 450 void removeEdge(Node &Caller, Function &Callee); 451 452 /// \brief Update the call graph after deleting an edge. removeEdge(Function & Caller,Function & Callee)453 void removeEdge(Function &Caller, Function &Callee) { 454 return removeEdge(get(Caller), Callee); 455 } 456 457 ///@} 458 459 private: 460 /// \brief Allocator that holds all the call graph nodes. 461 SpecificBumpPtrAllocator<Node> BPA; 462 463 /// \brief Maps function->node for fast lookup. 464 DenseMap<const Function *, Node *> NodeMap; 465 466 /// \brief The entry nodes to the graph. 467 /// 468 /// These nodes are reachable through "external" means. Put another way, they 469 /// escape at the module scope. 470 NodeVectorT EntryNodes; 471 472 /// \brief Map of the entry nodes in the graph to their indices in 473 /// \c EntryNodes. 474 DenseMap<Function *, size_t> EntryIndexMap; 475 476 /// \brief Allocator that holds all the call graph SCCs. 477 SpecificBumpPtrAllocator<SCC> SCCBPA; 478 479 /// \brief Maps Function -> SCC for fast lookup. 480 DenseMap<Node *, SCC *> SCCMap; 481 482 /// \brief The leaf SCCs of the graph. 483 /// 484 /// These are all of the SCCs which have no children. 485 SmallVector<SCC *, 4> LeafSCCs; 486 487 /// \brief Stack of nodes in the DFS walk. 488 SmallVector<std::pair<Node *, iterator>, 4> DFSStack; 489 490 /// \brief Set of entry nodes not-yet-processed into SCCs. 491 SmallVector<Function *, 4> SCCEntryNodes; 492 493 /// \brief Stack of nodes the DFS has walked but not yet put into a SCC. 494 SmallVector<Node *, 4> PendingSCCStack; 495 496 /// \brief Counter for the next DFS number to assign. 497 int NextDFSNumber; 498 499 /// \brief Helper to insert a new function, with an already looked-up entry in 500 /// the NodeMap. 501 Node &insertInto(Function &F, Node *&MappedN); 502 503 /// \brief Helper to update pointers back to the graph object during moves. 504 void updateGraphPtrs(); 505 506 /// \brief Helper to form a new SCC out of the top of a DFSStack-like 507 /// structure. 508 SCC *formSCC(Node *RootN, SmallVectorImpl<Node *> &NodeStack); 509 510 /// \brief Retrieve the next node in the post-order SCC walk of the call graph. 511 SCC *getNextSCCInPostOrder(); 512 }; 513 514 // Provide GraphTraits specializations for call graphs. 515 template <> struct GraphTraits<LazyCallGraph::Node *> { 516 typedef LazyCallGraph::Node NodeType; 517 typedef LazyCallGraph::iterator ChildIteratorType; 518 519 static NodeType *getEntryNode(NodeType *N) { return N; } 520 static ChildIteratorType child_begin(NodeType *N) { return N->begin(); } 521 static ChildIteratorType child_end(NodeType *N) { return N->end(); } 522 }; 523 template <> struct GraphTraits<LazyCallGraph *> { 524 typedef LazyCallGraph::Node NodeType; 525 typedef LazyCallGraph::iterator ChildIteratorType; 526 527 static NodeType *getEntryNode(NodeType *N) { return N; } 528 static ChildIteratorType child_begin(NodeType *N) { return N->begin(); } 529 static ChildIteratorType child_end(NodeType *N) { return N->end(); } 530 }; 531 532 /// \brief An analysis pass which computes the call graph for a module. 533 class LazyCallGraphAnalysis { 534 public: 535 /// \brief Inform generic clients of the result type. 536 typedef LazyCallGraph Result; 537 538 static void *ID() { return (void *)&PassID; } 539 540 /// \brief Compute the \c LazyCallGraph for a the module \c M. 541 /// 542 /// This just builds the set of entry points to the call graph. The rest is 543 /// built lazily as it is walked. 544 LazyCallGraph run(Module *M) { return LazyCallGraph(*M); } 545 546 private: 547 static char PassID; 548 }; 549 550 /// \brief A pass which prints the call graph to a \c raw_ostream. 551 /// 552 /// This is primarily useful for testing the analysis. 553 class LazyCallGraphPrinterPass { 554 raw_ostream &OS; 555 556 public: 557 explicit LazyCallGraphPrinterPass(raw_ostream &OS); 558 559 PreservedAnalyses run(Module *M, ModuleAnalysisManager *AM); 560 561 static StringRef name() { return "LazyCallGraphPrinterPass"; } 562 }; 563 564 } 565 566 #endif 567