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_LAZYCALLGRAPH_H 36 #define LLVM_ANALYSIS_LAZYCALLGRAPH_H 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/IR/PassManager.h" 50 #include "llvm/Support/Allocator.h" 51 #include <iterator> 52 53 namespace llvm { 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 make_range(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 /// \brief Short name useful for debugging or logging. 256 /// 257 /// We use the name of the first function in the SCC to name the SCC for 258 /// the purposes of debugging and logging. getName()259 StringRef getName() const { return (*begin())->getFunction().getName(); } 260 261 ///@{ 262 /// \name Mutation API 263 /// 264 /// These methods provide the core API for updating the call graph in the 265 /// presence of a (potentially still in-flight) DFS-found SCCs. 266 /// 267 /// Note that these methods sometimes have complex runtimes, so be careful 268 /// how you call them. 269 270 /// \brief Insert an edge from one node in this SCC to another in this SCC. 271 /// 272 /// By the definition of an SCC, this does not change the nature or make-up 273 /// of any SCCs. 274 void insertIntraSCCEdge(Node &CallerN, Node &CalleeN); 275 276 /// \brief Insert an edge whose tail is in this SCC and head is in some 277 /// child SCC. 278 /// 279 /// There must be an existing path from the caller to the callee. This 280 /// operation is inexpensive and does not change the set of SCCs in the 281 /// graph. 282 void insertOutgoingEdge(Node &CallerN, Node &CalleeN); 283 284 /// \brief Insert an edge whose tail is in a descendant SCC and head is in 285 /// this SCC. 286 /// 287 /// There must be an existing path from the callee to the caller in this 288 /// case. NB! This is has the potential to be a very expensive function. It 289 /// inherently forms a cycle in the prior SCC DAG and we have to merge SCCs 290 /// to resolve that cycle. But finding all of the SCCs which participate in 291 /// the cycle can in the worst case require traversing every SCC in the 292 /// graph. Every attempt is made to avoid that, but passes must still 293 /// exercise caution calling this routine repeatedly. 294 /// 295 /// FIXME: We could possibly optimize this quite a bit for cases where the 296 /// caller and callee are very nearby in the graph. See comments in the 297 /// implementation for details, but that use case might impact users. 298 SmallVector<SCC *, 1> insertIncomingEdge(Node &CallerN, Node &CalleeN); 299 300 /// \brief Remove an edge whose source is in this SCC and target is *not*. 301 /// 302 /// This removes an inter-SCC edge. All inter-SCC edges originating from 303 /// this SCC have been fully explored by any in-flight DFS SCC formation, 304 /// so this is always safe to call once you have the source SCC. 305 /// 306 /// This operation does not change the set of SCCs or the members of the 307 /// SCCs and so is very inexpensive. It may change the connectivity graph 308 /// of the SCCs though, so be careful calling this while iterating over 309 /// them. 310 void removeInterSCCEdge(Node &CallerN, Node &CalleeN); 311 312 /// \brief Remove an edge which is entirely within this SCC. 313 /// 314 /// Both the \a Caller and the \a Callee must be within this SCC. Removing 315 /// such an edge make break cycles that form this SCC and thus this 316 /// operation may change the SCC graph significantly. In particular, this 317 /// operation will re-form new SCCs based on the remaining connectivity of 318 /// the graph. The following invariants are guaranteed to hold after 319 /// calling this method: 320 /// 321 /// 1) This SCC is still an SCC in the graph. 322 /// 2) This SCC will be the parent of any new SCCs. Thus, this SCC is 323 /// preserved as the root of any new SCC directed graph formed. 324 /// 3) No SCC other than this SCC has its member set changed (this is 325 /// inherent in the definition of removing such an edge). 326 /// 4) All of the parent links of the SCC graph will be updated to reflect 327 /// the new SCC structure. 328 /// 5) All SCCs formed out of this SCC, excluding this SCC, will be 329 /// returned in a vector. 330 /// 6) The order of the SCCs in the vector will be a valid postorder 331 /// traversal of the new SCCs. 332 /// 333 /// These invariants are very important to ensure that we can build 334 /// optimization pipeliens on top of the CGSCC pass manager which 335 /// intelligently update the SCC graph without invalidating other parts of 336 /// the SCC graph. 337 /// 338 /// The runtime complexity of this method is, in the worst case, O(V+E) 339 /// where V is the number of nodes in this SCC and E is the number of edges 340 /// leaving the nodes in this SCC. Note that E includes both edges within 341 /// this SCC and edges from this SCC to child SCCs. Some effort has been 342 /// made to minimize the overhead of common cases such as self-edges and 343 /// edge removals which result in a spanning tree with no more cycles. 344 SmallVector<SCC *, 1> removeIntraSCCEdge(Node &CallerN, Node &CalleeN); 345 346 ///@} 347 }; 348 349 /// \brief A post-order depth-first SCC iterator over the call graph. 350 /// 351 /// This iterator triggers the Tarjan DFS-based formation of the SCC DAG for 352 /// the call graph, walking it lazily in depth-first post-order. That is, it 353 /// always visits SCCs for a callee prior to visiting the SCC for a caller 354 /// (when they are in different SCCs). 355 class postorder_scc_iterator 356 : public iterator_facade_base<postorder_scc_iterator, 357 std::forward_iterator_tag, SCC> { 358 friend class LazyCallGraph; 359 friend class LazyCallGraph::Node; 360 361 /// \brief Nonce type to select the constructor for the end iterator. 362 struct IsAtEndT {}; 363 364 LazyCallGraph *G; 365 SCC *C; 366 367 // Build the begin iterator for a node. postorder_scc_iterator(LazyCallGraph & G)368 postorder_scc_iterator(LazyCallGraph &G) : G(&G) { 369 C = G.getNextSCCInPostOrder(); 370 } 371 372 // Build the end iterator for a node. This is selected purely by overload. postorder_scc_iterator(LazyCallGraph & G,IsAtEndT)373 postorder_scc_iterator(LazyCallGraph &G, IsAtEndT /*Nonce*/) 374 : G(&G), C(nullptr) {} 375 376 public: 377 bool operator==(const postorder_scc_iterator &Arg) const { 378 return G == Arg.G && C == Arg.C; 379 } 380 381 reference operator*() const { return *C; } 382 383 using iterator_facade_base::operator++; 384 postorder_scc_iterator &operator++() { 385 C = G->getNextSCCInPostOrder(); 386 return *this; 387 } 388 }; 389 390 /// \brief Construct a graph for the given module. 391 /// 392 /// This sets up the graph and computes all of the entry points of the graph. 393 /// No function definitions are scanned until their nodes in the graph are 394 /// requested during traversal. 395 LazyCallGraph(Module &M); 396 397 LazyCallGraph(LazyCallGraph &&G); 398 LazyCallGraph &operator=(LazyCallGraph &&RHS); 399 begin()400 iterator begin() { 401 return iterator(*this, EntryNodes.begin(), EntryNodes.end()); 402 } end()403 iterator end() { return iterator(*this, EntryNodes.end(), EntryNodes.end()); } 404 postorder_scc_begin()405 postorder_scc_iterator postorder_scc_begin() { 406 return postorder_scc_iterator(*this); 407 } postorder_scc_end()408 postorder_scc_iterator postorder_scc_end() { 409 return postorder_scc_iterator(*this, postorder_scc_iterator::IsAtEndT()); 410 } 411 postorder_sccs()412 iterator_range<postorder_scc_iterator> postorder_sccs() { 413 return make_range(postorder_scc_begin(), postorder_scc_end()); 414 } 415 416 /// \brief Lookup a function in the graph which has already been scanned and 417 /// added. lookup(const Function & F)418 Node *lookup(const Function &F) const { return NodeMap.lookup(&F); } 419 420 /// \brief Lookup a function's SCC in the graph. 421 /// 422 /// \returns null if the function hasn't been assigned an SCC via the SCC 423 /// iterator walk. lookupSCC(Node & N)424 SCC *lookupSCC(Node &N) const { return SCCMap.lookup(&N); } 425 426 /// \brief Get a graph node for a given function, scanning it to populate the 427 /// graph data as necessary. get(Function & F)428 Node &get(Function &F) { 429 Node *&N = NodeMap[&F]; 430 if (N) 431 return *N; 432 433 return insertInto(F, N); 434 } 435 436 ///@{ 437 /// \name Pre-SCC Mutation API 438 /// 439 /// These methods are only valid to call prior to forming any SCCs for this 440 /// call graph. They can be used to update the core node-graph during 441 /// a node-based inorder traversal that precedes any SCC-based traversal. 442 /// 443 /// Once you begin manipulating a call graph's SCCs, you must perform all 444 /// mutation of the graph via the SCC methods. 445 446 /// \brief Update the call graph after inserting a new edge. 447 void insertEdge(Node &Caller, Function &Callee); 448 449 /// \brief Update the call graph after inserting a new edge. insertEdge(Function & Caller,Function & Callee)450 void insertEdge(Function &Caller, Function &Callee) { 451 return insertEdge(get(Caller), Callee); 452 } 453 454 /// \brief Update the call graph after deleting an edge. 455 void removeEdge(Node &Caller, Function &Callee); 456 457 /// \brief Update the call graph after deleting an edge. removeEdge(Function & Caller,Function & Callee)458 void removeEdge(Function &Caller, Function &Callee) { 459 return removeEdge(get(Caller), Callee); 460 } 461 462 ///@} 463 464 private: 465 /// \brief Allocator that holds all the call graph nodes. 466 SpecificBumpPtrAllocator<Node> BPA; 467 468 /// \brief Maps function->node for fast lookup. 469 DenseMap<const Function *, Node *> NodeMap; 470 471 /// \brief The entry nodes to the graph. 472 /// 473 /// These nodes are reachable through "external" means. Put another way, they 474 /// escape at the module scope. 475 NodeVectorT EntryNodes; 476 477 /// \brief Map of the entry nodes in the graph to their indices in 478 /// \c EntryNodes. 479 DenseMap<Function *, size_t> EntryIndexMap; 480 481 /// \brief Allocator that holds all the call graph SCCs. 482 SpecificBumpPtrAllocator<SCC> SCCBPA; 483 484 /// \brief Maps Function -> SCC for fast lookup. 485 DenseMap<Node *, SCC *> SCCMap; 486 487 /// \brief The leaf SCCs of the graph. 488 /// 489 /// These are all of the SCCs which have no children. 490 SmallVector<SCC *, 4> LeafSCCs; 491 492 /// \brief Stack of nodes in the DFS walk. 493 SmallVector<std::pair<Node *, iterator>, 4> DFSStack; 494 495 /// \brief Set of entry nodes not-yet-processed into SCCs. 496 SmallVector<Function *, 4> SCCEntryNodes; 497 498 /// \brief Stack of nodes the DFS has walked but not yet put into a SCC. 499 SmallVector<Node *, 4> PendingSCCStack; 500 501 /// \brief Counter for the next DFS number to assign. 502 int NextDFSNumber; 503 504 /// \brief Helper to insert a new function, with an already looked-up entry in 505 /// the NodeMap. 506 Node &insertInto(Function &F, Node *&MappedN); 507 508 /// \brief Helper to update pointers back to the graph object during moves. 509 void updateGraphPtrs(); 510 511 /// \brief Helper to form a new SCC out of the top of a DFSStack-like 512 /// structure. 513 SCC *formSCC(Node *RootN, SmallVectorImpl<Node *> &NodeStack); 514 515 /// \brief Retrieve the next node in the post-order SCC walk of the call graph. 516 SCC *getNextSCCInPostOrder(); 517 }; 518 519 // Provide GraphTraits specializations for call graphs. 520 template <> struct GraphTraits<LazyCallGraph::Node *> { 521 typedef LazyCallGraph::Node NodeType; 522 typedef LazyCallGraph::iterator ChildIteratorType; 523 524 static NodeType *getEntryNode(NodeType *N) { return N; } 525 static ChildIteratorType child_begin(NodeType *N) { return N->begin(); } 526 static ChildIteratorType child_end(NodeType *N) { return N->end(); } 527 }; 528 template <> struct GraphTraits<LazyCallGraph *> { 529 typedef LazyCallGraph::Node NodeType; 530 typedef LazyCallGraph::iterator ChildIteratorType; 531 532 static NodeType *getEntryNode(NodeType *N) { return N; } 533 static ChildIteratorType child_begin(NodeType *N) { return N->begin(); } 534 static ChildIteratorType child_end(NodeType *N) { return N->end(); } 535 }; 536 537 /// \brief An analysis pass which computes the call graph for a module. 538 class LazyCallGraphAnalysis { 539 public: 540 /// \brief Inform generic clients of the result type. 541 typedef LazyCallGraph Result; 542 543 static void *ID() { return (void *)&PassID; } 544 545 static StringRef name() { return "Lazy CallGraph Analysis"; } 546 547 /// \brief Compute the \c LazyCallGraph for the module \c M. 548 /// 549 /// This just builds the set of entry points to the call graph. The rest is 550 /// built lazily as it is walked. 551 LazyCallGraph run(Module &M) { return LazyCallGraph(M); } 552 553 private: 554 static char PassID; 555 }; 556 557 /// \brief A pass which prints the call graph to a \c raw_ostream. 558 /// 559 /// This is primarily useful for testing the analysis. 560 class LazyCallGraphPrinterPass { 561 raw_ostream &OS; 562 563 public: 564 explicit LazyCallGraphPrinterPass(raw_ostream &OS); 565 566 PreservedAnalyses run(Module &M, ModuleAnalysisManager *AM); 567 568 static StringRef name() { return "LazyCallGraphPrinterPass"; } 569 }; 570 571 } 572 573 #endif 574