1 //===- LazyCallGraph.cpp - Analysis of a Module's call graph --------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 9 #include "llvm/Analysis/LazyCallGraph.h" 10 #include "llvm/ADT/ArrayRef.h" 11 #include "llvm/ADT/STLExtras.h" 12 #include "llvm/ADT/Sequence.h" 13 #include "llvm/ADT/SmallPtrSet.h" 14 #include "llvm/ADT/SmallVector.h" 15 #include "llvm/ADT/iterator_range.h" 16 #include "llvm/Analysis/TargetLibraryInfo.h" 17 #include "llvm/Config/llvm-config.h" 18 #include "llvm/IR/Constants.h" 19 #include "llvm/IR/Function.h" 20 #include "llvm/IR/GlobalVariable.h" 21 #include "llvm/IR/InstIterator.h" 22 #include "llvm/IR/Instruction.h" 23 #include "llvm/IR/Module.h" 24 #include "llvm/IR/PassManager.h" 25 #include "llvm/Support/Casting.h" 26 #include "llvm/Support/Compiler.h" 27 #include "llvm/Support/Debug.h" 28 #include "llvm/Support/GraphWriter.h" 29 #include "llvm/Support/raw_ostream.h" 30 #include <algorithm> 31 #include <cassert> 32 #include <iterator> 33 #include <string> 34 #include <tuple> 35 #include <utility> 36 37 #ifdef EXPENSIVE_CHECKS 38 #include "llvm/ADT/ScopeExit.h" 39 #endif 40 41 using namespace llvm; 42 43 #define DEBUG_TYPE "lcg" 44 45 void LazyCallGraph::EdgeSequence::insertEdgeInternal(Node &TargetN, 46 Edge::Kind EK) { 47 EdgeIndexMap.insert({&TargetN, Edges.size()}); 48 Edges.emplace_back(TargetN, EK); 49 } 50 51 void LazyCallGraph::EdgeSequence::setEdgeKind(Node &TargetN, Edge::Kind EK) { 52 Edges[EdgeIndexMap.find(&TargetN)->second].setKind(EK); 53 } 54 55 bool LazyCallGraph::EdgeSequence::removeEdgeInternal(Node &TargetN) { 56 auto IndexMapI = EdgeIndexMap.find(&TargetN); 57 if (IndexMapI == EdgeIndexMap.end()) 58 return false; 59 60 Edges[IndexMapI->second] = Edge(); 61 EdgeIndexMap.erase(IndexMapI); 62 return true; 63 } 64 65 static void addEdge(SmallVectorImpl<LazyCallGraph::Edge> &Edges, 66 DenseMap<LazyCallGraph::Node *, int> &EdgeIndexMap, 67 LazyCallGraph::Node &N, LazyCallGraph::Edge::Kind EK) { 68 if (!EdgeIndexMap.insert({&N, Edges.size()}).second) 69 return; 70 71 LLVM_DEBUG(dbgs() << " Added callable function: " << N.getName() << "\n"); 72 Edges.emplace_back(LazyCallGraph::Edge(N, EK)); 73 } 74 75 LazyCallGraph::EdgeSequence &LazyCallGraph::Node::populateSlow() { 76 assert(!Edges && "Must not have already populated the edges for this node!"); 77 78 LLVM_DEBUG(dbgs() << " Adding functions called by '" << getName() 79 << "' to the graph.\n"); 80 81 Edges = EdgeSequence(); 82 83 SmallVector<Constant *, 16> Worklist; 84 SmallPtrSet<Function *, 4> Callees; 85 SmallPtrSet<Constant *, 16> Visited; 86 87 // Find all the potential call graph edges in this function. We track both 88 // actual call edges and indirect references to functions. The direct calls 89 // are trivially added, but to accumulate the latter we walk the instructions 90 // and add every operand which is a constant to the worklist to process 91 // afterward. 92 // 93 // Note that we consider *any* function with a definition to be a viable 94 // edge. Even if the function's definition is subject to replacement by 95 // some other module (say, a weak definition) there may still be 96 // optimizations which essentially speculate based on the definition and 97 // a way to check that the specific definition is in fact the one being 98 // used. For example, this could be done by moving the weak definition to 99 // a strong (internal) definition and making the weak definition be an 100 // alias. Then a test of the address of the weak function against the new 101 // strong definition's address would be an effective way to determine the 102 // safety of optimizing a direct call edge. 103 for (BasicBlock &BB : *F) 104 for (Instruction &I : BB) { 105 if (auto *CB = dyn_cast<CallBase>(&I)) 106 if (Function *Callee = CB->getCalledFunction()) 107 if (!Callee->isDeclaration()) 108 if (Callees.insert(Callee).second) { 109 Visited.insert(Callee); 110 addEdge(Edges->Edges, Edges->EdgeIndexMap, G->get(*Callee), 111 LazyCallGraph::Edge::Call); 112 } 113 114 for (Value *Op : I.operand_values()) 115 if (Constant *C = dyn_cast<Constant>(Op)) 116 if (Visited.insert(C).second) 117 Worklist.push_back(C); 118 } 119 120 // We've collected all the constant (and thus potentially function or 121 // function containing) operands to all of the instructions in the function. 122 // Process them (recursively) collecting every function found. 123 visitReferences(Worklist, Visited, [&](Function &F) { 124 addEdge(Edges->Edges, Edges->EdgeIndexMap, G->get(F), 125 LazyCallGraph::Edge::Ref); 126 }); 127 128 // Add implicit reference edges to any defined libcall functions (if we 129 // haven't found an explicit edge). 130 for (auto *F : G->LibFunctions) 131 if (!Visited.count(F)) 132 addEdge(Edges->Edges, Edges->EdgeIndexMap, G->get(*F), 133 LazyCallGraph::Edge::Ref); 134 135 return *Edges; 136 } 137 138 void LazyCallGraph::Node::replaceFunction(Function &NewF) { 139 assert(F != &NewF && "Must not replace a function with itself!"); 140 F = &NewF; 141 } 142 143 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 144 LLVM_DUMP_METHOD void LazyCallGraph::Node::dump() const { 145 dbgs() << *this << '\n'; 146 } 147 #endif 148 149 static bool isKnownLibFunction(Function &F, TargetLibraryInfo &TLI) { 150 LibFunc LF; 151 152 // Either this is a normal library function or a "vectorizable" 153 // function. Not using the VFDatabase here because this query 154 // is related only to libraries handled via the TLI. 155 return TLI.getLibFunc(F, LF) || 156 TLI.isKnownVectorFunctionInLibrary(F.getName()); 157 } 158 159 LazyCallGraph::LazyCallGraph( 160 Module &M, function_ref<TargetLibraryInfo &(Function &)> GetTLI) { 161 LLVM_DEBUG(dbgs() << "Building CG for module: " << M.getModuleIdentifier() 162 << "\n"); 163 for (Function &F : M) { 164 if (F.isDeclaration()) 165 continue; 166 // If this function is a known lib function to LLVM then we want to 167 // synthesize reference edges to it to model the fact that LLVM can turn 168 // arbitrary code into a library function call. 169 if (isKnownLibFunction(F, GetTLI(F))) 170 LibFunctions.insert(&F); 171 172 if (F.hasLocalLinkage()) 173 continue; 174 175 // External linkage defined functions have edges to them from other 176 // modules. 177 LLVM_DEBUG(dbgs() << " Adding '" << F.getName() 178 << "' to entry set of the graph.\n"); 179 addEdge(EntryEdges.Edges, EntryEdges.EdgeIndexMap, get(F), Edge::Ref); 180 } 181 182 // Externally visible aliases of internal functions are also viable entry 183 // edges to the module. 184 for (auto &A : M.aliases()) { 185 if (A.hasLocalLinkage()) 186 continue; 187 if (Function* F = dyn_cast<Function>(A.getAliasee())) { 188 LLVM_DEBUG(dbgs() << " Adding '" << F->getName() 189 << "' with alias '" << A.getName() 190 << "' to entry set of the graph.\n"); 191 addEdge(EntryEdges.Edges, EntryEdges.EdgeIndexMap, get(*F), Edge::Ref); 192 } 193 } 194 195 // Now add entry nodes for functions reachable via initializers to globals. 196 SmallVector<Constant *, 16> Worklist; 197 SmallPtrSet<Constant *, 16> Visited; 198 for (GlobalVariable &GV : M.globals()) 199 if (GV.hasInitializer()) 200 if (Visited.insert(GV.getInitializer()).second) 201 Worklist.push_back(GV.getInitializer()); 202 203 LLVM_DEBUG( 204 dbgs() << " Adding functions referenced by global initializers to the " 205 "entry set.\n"); 206 visitReferences(Worklist, Visited, [&](Function &F) { 207 addEdge(EntryEdges.Edges, EntryEdges.EdgeIndexMap, get(F), 208 LazyCallGraph::Edge::Ref); 209 }); 210 } 211 212 LazyCallGraph::LazyCallGraph(LazyCallGraph &&G) 213 : BPA(std::move(G.BPA)), NodeMap(std::move(G.NodeMap)), 214 EntryEdges(std::move(G.EntryEdges)), SCCBPA(std::move(G.SCCBPA)), 215 SCCMap(std::move(G.SCCMap)), 216 LibFunctions(std::move(G.LibFunctions)) { 217 updateGraphPtrs(); 218 } 219 220 bool LazyCallGraph::invalidate(Module &, const PreservedAnalyses &PA, 221 ModuleAnalysisManager::Invalidator &) { 222 // Check whether the analysis, all analyses on functions, or the function's 223 // CFG have been preserved. 224 auto PAC = PA.getChecker<llvm::LazyCallGraphAnalysis>(); 225 return !(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Module>>()); 226 } 227 228 LazyCallGraph &LazyCallGraph::operator=(LazyCallGraph &&G) { 229 BPA = std::move(G.BPA); 230 NodeMap = std::move(G.NodeMap); 231 EntryEdges = std::move(G.EntryEdges); 232 SCCBPA = std::move(G.SCCBPA); 233 SCCMap = std::move(G.SCCMap); 234 LibFunctions = std::move(G.LibFunctions); 235 updateGraphPtrs(); 236 return *this; 237 } 238 239 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 240 LLVM_DUMP_METHOD void LazyCallGraph::SCC::dump() const { 241 dbgs() << *this << '\n'; 242 } 243 #endif 244 245 #if !defined(NDEBUG) || defined(EXPENSIVE_CHECKS) 246 void LazyCallGraph::SCC::verify() { 247 assert(OuterRefSCC && "Can't have a null RefSCC!"); 248 assert(!Nodes.empty() && "Can't have an empty SCC!"); 249 250 for (Node *N : Nodes) { 251 assert(N && "Can't have a null node!"); 252 assert(OuterRefSCC->G->lookupSCC(*N) == this && 253 "Node does not map to this SCC!"); 254 assert(N->DFSNumber == -1 && 255 "Must set DFS numbers to -1 when adding a node to an SCC!"); 256 assert(N->LowLink == -1 && 257 "Must set low link to -1 when adding a node to an SCC!"); 258 for (Edge &E : **N) 259 assert(E.getNode().isPopulated() && "Can't have an unpopulated node!"); 260 261 #ifdef EXPENSIVE_CHECKS 262 // Verify that all nodes in this SCC can reach all other nodes. 263 SmallVector<Node *, 4> Worklist; 264 SmallPtrSet<Node *, 4> Visited; 265 Worklist.push_back(N); 266 while (!Worklist.empty()) { 267 Node *VisitingNode = Worklist.pop_back_val(); 268 if (!Visited.insert(VisitingNode).second) 269 continue; 270 for (Edge &E : (*VisitingNode)->calls()) 271 Worklist.push_back(&E.getNode()); 272 } 273 for (Node *NodeToVisit : Nodes) { 274 assert(Visited.contains(NodeToVisit) && 275 "Cannot reach all nodes within SCC"); 276 } 277 #endif 278 } 279 } 280 #endif 281 282 bool LazyCallGraph::SCC::isParentOf(const SCC &C) const { 283 if (this == &C) 284 return false; 285 286 for (Node &N : *this) 287 for (Edge &E : N->calls()) 288 if (OuterRefSCC->G->lookupSCC(E.getNode()) == &C) 289 return true; 290 291 // No edges found. 292 return false; 293 } 294 295 bool LazyCallGraph::SCC::isAncestorOf(const SCC &TargetC) const { 296 if (this == &TargetC) 297 return false; 298 299 LazyCallGraph &G = *OuterRefSCC->G; 300 301 // Start with this SCC. 302 SmallPtrSet<const SCC *, 16> Visited = {this}; 303 SmallVector<const SCC *, 16> Worklist = {this}; 304 305 // Walk down the graph until we run out of edges or find a path to TargetC. 306 do { 307 const SCC &C = *Worklist.pop_back_val(); 308 for (Node &N : C) 309 for (Edge &E : N->calls()) { 310 SCC *CalleeC = G.lookupSCC(E.getNode()); 311 if (!CalleeC) 312 continue; 313 314 // If the callee's SCC is the TargetC, we're done. 315 if (CalleeC == &TargetC) 316 return true; 317 318 // If this is the first time we've reached this SCC, put it on the 319 // worklist to recurse through. 320 if (Visited.insert(CalleeC).second) 321 Worklist.push_back(CalleeC); 322 } 323 } while (!Worklist.empty()); 324 325 // No paths found. 326 return false; 327 } 328 329 LazyCallGraph::RefSCC::RefSCC(LazyCallGraph &G) : G(&G) {} 330 331 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 332 LLVM_DUMP_METHOD void LazyCallGraph::RefSCC::dump() const { 333 dbgs() << *this << '\n'; 334 } 335 #endif 336 337 #if !defined(NDEBUG) || defined(EXPENSIVE_CHECKS) 338 void LazyCallGraph::RefSCC::verify() { 339 assert(G && "Can't have a null graph!"); 340 assert(!SCCs.empty() && "Can't have an empty SCC!"); 341 342 // Verify basic properties of the SCCs. 343 SmallPtrSet<SCC *, 4> SCCSet; 344 for (SCC *C : SCCs) { 345 assert(C && "Can't have a null SCC!"); 346 C->verify(); 347 assert(&C->getOuterRefSCC() == this && 348 "SCC doesn't think it is inside this RefSCC!"); 349 bool Inserted = SCCSet.insert(C).second; 350 assert(Inserted && "Found a duplicate SCC!"); 351 auto IndexIt = SCCIndices.find(C); 352 assert(IndexIt != SCCIndices.end() && 353 "Found an SCC that doesn't have an index!"); 354 } 355 356 // Check that our indices map correctly. 357 for (auto &SCCIndexPair : SCCIndices) { 358 SCC *C = SCCIndexPair.first; 359 int i = SCCIndexPair.second; 360 assert(C && "Can't have a null SCC in the indices!"); 361 assert(SCCSet.count(C) && "Found an index for an SCC not in the RefSCC!"); 362 assert(SCCs[i] == C && "Index doesn't point to SCC!"); 363 } 364 365 // Check that the SCCs are in fact in post-order. 366 for (int i = 0, Size = SCCs.size(); i < Size; ++i) { 367 SCC &SourceSCC = *SCCs[i]; 368 for (Node &N : SourceSCC) 369 for (Edge &E : *N) { 370 if (!E.isCall()) 371 continue; 372 SCC &TargetSCC = *G->lookupSCC(E.getNode()); 373 if (&TargetSCC.getOuterRefSCC() == this) { 374 assert(SCCIndices.find(&TargetSCC)->second <= i && 375 "Edge between SCCs violates post-order relationship."); 376 continue; 377 } 378 } 379 } 380 381 #ifdef EXPENSIVE_CHECKS 382 // Verify that all nodes in this RefSCC can reach all other nodes. 383 SmallVector<Node *> Nodes; 384 for (SCC *C : SCCs) { 385 for (Node &N : *C) 386 Nodes.push_back(&N); 387 } 388 for (Node *N : Nodes) { 389 SmallVector<Node *, 4> Worklist; 390 SmallPtrSet<Node *, 4> Visited; 391 Worklist.push_back(N); 392 while (!Worklist.empty()) { 393 Node *VisitingNode = Worklist.pop_back_val(); 394 if (!Visited.insert(VisitingNode).second) 395 continue; 396 for (Edge &E : **VisitingNode) 397 Worklist.push_back(&E.getNode()); 398 } 399 for (Node *NodeToVisit : Nodes) { 400 assert(Visited.contains(NodeToVisit) && 401 "Cannot reach all nodes within RefSCC"); 402 } 403 } 404 #endif 405 } 406 #endif 407 408 bool LazyCallGraph::RefSCC::isParentOf(const RefSCC &RC) const { 409 if (&RC == this) 410 return false; 411 412 // Search all edges to see if this is a parent. 413 for (SCC &C : *this) 414 for (Node &N : C) 415 for (Edge &E : *N) 416 if (G->lookupRefSCC(E.getNode()) == &RC) 417 return true; 418 419 return false; 420 } 421 422 bool LazyCallGraph::RefSCC::isAncestorOf(const RefSCC &RC) const { 423 if (&RC == this) 424 return false; 425 426 // For each descendant of this RefSCC, see if one of its children is the 427 // argument. If not, add that descendant to the worklist and continue 428 // searching. 429 SmallVector<const RefSCC *, 4> Worklist; 430 SmallPtrSet<const RefSCC *, 4> Visited; 431 Worklist.push_back(this); 432 Visited.insert(this); 433 do { 434 const RefSCC &DescendantRC = *Worklist.pop_back_val(); 435 for (SCC &C : DescendantRC) 436 for (Node &N : C) 437 for (Edge &E : *N) { 438 auto *ChildRC = G->lookupRefSCC(E.getNode()); 439 if (ChildRC == &RC) 440 return true; 441 if (!ChildRC || !Visited.insert(ChildRC).second) 442 continue; 443 Worklist.push_back(ChildRC); 444 } 445 } while (!Worklist.empty()); 446 447 return false; 448 } 449 450 /// Generic helper that updates a postorder sequence of SCCs for a potentially 451 /// cycle-introducing edge insertion. 452 /// 453 /// A postorder sequence of SCCs of a directed graph has one fundamental 454 /// property: all deges in the DAG of SCCs point "up" the sequence. That is, 455 /// all edges in the SCC DAG point to prior SCCs in the sequence. 456 /// 457 /// This routine both updates a postorder sequence and uses that sequence to 458 /// compute the set of SCCs connected into a cycle. It should only be called to 459 /// insert a "downward" edge which will require changing the sequence to 460 /// restore it to a postorder. 461 /// 462 /// When inserting an edge from an earlier SCC to a later SCC in some postorder 463 /// sequence, all of the SCCs which may be impacted are in the closed range of 464 /// those two within the postorder sequence. The algorithm used here to restore 465 /// the state is as follows: 466 /// 467 /// 1) Starting from the source SCC, construct a set of SCCs which reach the 468 /// source SCC consisting of just the source SCC. Then scan toward the 469 /// target SCC in postorder and for each SCC, if it has an edge to an SCC 470 /// in the set, add it to the set. Otherwise, the source SCC is not 471 /// a successor, move it in the postorder sequence to immediately before 472 /// the source SCC, shifting the source SCC and all SCCs in the set one 473 /// position toward the target SCC. Stop scanning after processing the 474 /// target SCC. 475 /// 2) If the source SCC is now past the target SCC in the postorder sequence, 476 /// and thus the new edge will flow toward the start, we are done. 477 /// 3) Otherwise, starting from the target SCC, walk all edges which reach an 478 /// SCC between the source and the target, and add them to the set of 479 /// connected SCCs, then recurse through them. Once a complete set of the 480 /// SCCs the target connects to is known, hoist the remaining SCCs between 481 /// the source and the target to be above the target. Note that there is no 482 /// need to process the source SCC, it is already known to connect. 483 /// 4) At this point, all of the SCCs in the closed range between the source 484 /// SCC and the target SCC in the postorder sequence are connected, 485 /// including the target SCC and the source SCC. Inserting the edge from 486 /// the source SCC to the target SCC will form a cycle out of precisely 487 /// these SCCs. Thus we can merge all of the SCCs in this closed range into 488 /// a single SCC. 489 /// 490 /// This process has various important properties: 491 /// - Only mutates the SCCs when adding the edge actually changes the SCC 492 /// structure. 493 /// - Never mutates SCCs which are unaffected by the change. 494 /// - Updates the postorder sequence to correctly satisfy the postorder 495 /// constraint after the edge is inserted. 496 /// - Only reorders SCCs in the closed postorder sequence from the source to 497 /// the target, so easy to bound how much has changed even in the ordering. 498 /// - Big-O is the number of edges in the closed postorder range of SCCs from 499 /// source to target. 500 /// 501 /// This helper routine, in addition to updating the postorder sequence itself 502 /// will also update a map from SCCs to indices within that sequence. 503 /// 504 /// The sequence and the map must operate on pointers to the SCC type. 505 /// 506 /// Two callbacks must be provided. The first computes the subset of SCCs in 507 /// the postorder closed range from the source to the target which connect to 508 /// the source SCC via some (transitive) set of edges. The second computes the 509 /// subset of the same range which the target SCC connects to via some 510 /// (transitive) set of edges. Both callbacks should populate the set argument 511 /// provided. 512 template <typename SCCT, typename PostorderSequenceT, typename SCCIndexMapT, 513 typename ComputeSourceConnectedSetCallableT, 514 typename ComputeTargetConnectedSetCallableT> 515 static iterator_range<typename PostorderSequenceT::iterator> 516 updatePostorderSequenceForEdgeInsertion( 517 SCCT &SourceSCC, SCCT &TargetSCC, PostorderSequenceT &SCCs, 518 SCCIndexMapT &SCCIndices, 519 ComputeSourceConnectedSetCallableT ComputeSourceConnectedSet, 520 ComputeTargetConnectedSetCallableT ComputeTargetConnectedSet) { 521 int SourceIdx = SCCIndices[&SourceSCC]; 522 int TargetIdx = SCCIndices[&TargetSCC]; 523 assert(SourceIdx < TargetIdx && "Cannot have equal indices here!"); 524 525 SmallPtrSet<SCCT *, 4> ConnectedSet; 526 527 // Compute the SCCs which (transitively) reach the source. 528 ComputeSourceConnectedSet(ConnectedSet); 529 530 // Partition the SCCs in this part of the port-order sequence so only SCCs 531 // connecting to the source remain between it and the target. This is 532 // a benign partition as it preserves postorder. 533 auto SourceI = std::stable_partition( 534 SCCs.begin() + SourceIdx, SCCs.begin() + TargetIdx + 1, 535 [&ConnectedSet](SCCT *C) { return !ConnectedSet.count(C); }); 536 for (int i = SourceIdx, e = TargetIdx + 1; i < e; ++i) 537 SCCIndices.find(SCCs[i])->second = i; 538 539 // If the target doesn't connect to the source, then we've corrected the 540 // post-order and there are no cycles formed. 541 if (!ConnectedSet.count(&TargetSCC)) { 542 assert(SourceI > (SCCs.begin() + SourceIdx) && 543 "Must have moved the source to fix the post-order."); 544 assert(*std::prev(SourceI) == &TargetSCC && 545 "Last SCC to move should have bene the target."); 546 547 // Return an empty range at the target SCC indicating there is nothing to 548 // merge. 549 return make_range(std::prev(SourceI), std::prev(SourceI)); 550 } 551 552 assert(SCCs[TargetIdx] == &TargetSCC && 553 "Should not have moved target if connected!"); 554 SourceIdx = SourceI - SCCs.begin(); 555 assert(SCCs[SourceIdx] == &SourceSCC && 556 "Bad updated index computation for the source SCC!"); 557 558 559 // See whether there are any remaining intervening SCCs between the source 560 // and target. If so we need to make sure they all are reachable form the 561 // target. 562 if (SourceIdx + 1 < TargetIdx) { 563 ConnectedSet.clear(); 564 ComputeTargetConnectedSet(ConnectedSet); 565 566 // Partition SCCs so that only SCCs reached from the target remain between 567 // the source and the target. This preserves postorder. 568 auto TargetI = std::stable_partition( 569 SCCs.begin() + SourceIdx + 1, SCCs.begin() + TargetIdx + 1, 570 [&ConnectedSet](SCCT *C) { return ConnectedSet.count(C); }); 571 for (int i = SourceIdx + 1, e = TargetIdx + 1; i < e; ++i) 572 SCCIndices.find(SCCs[i])->second = i; 573 TargetIdx = std::prev(TargetI) - SCCs.begin(); 574 assert(SCCs[TargetIdx] == &TargetSCC && 575 "Should always end with the target!"); 576 } 577 578 // At this point, we know that connecting source to target forms a cycle 579 // because target connects back to source, and we know that all of the SCCs 580 // between the source and target in the postorder sequence participate in that 581 // cycle. 582 return make_range(SCCs.begin() + SourceIdx, SCCs.begin() + TargetIdx); 583 } 584 585 bool 586 LazyCallGraph::RefSCC::switchInternalEdgeToCall( 587 Node &SourceN, Node &TargetN, 588 function_ref<void(ArrayRef<SCC *> MergeSCCs)> MergeCB) { 589 assert(!(*SourceN)[TargetN].isCall() && "Must start with a ref edge!"); 590 SmallVector<SCC *, 1> DeletedSCCs; 591 592 #ifdef EXPENSIVE_CHECKS 593 verify(); 594 auto VerifyOnExit = make_scope_exit([&]() { verify(); }); 595 #endif 596 597 SCC &SourceSCC = *G->lookupSCC(SourceN); 598 SCC &TargetSCC = *G->lookupSCC(TargetN); 599 600 // If the two nodes are already part of the same SCC, we're also done as 601 // we've just added more connectivity. 602 if (&SourceSCC == &TargetSCC) { 603 SourceN->setEdgeKind(TargetN, Edge::Call); 604 return false; // No new cycle. 605 } 606 607 // At this point we leverage the postorder list of SCCs to detect when the 608 // insertion of an edge changes the SCC structure in any way. 609 // 610 // First and foremost, we can eliminate the need for any changes when the 611 // edge is toward the beginning of the postorder sequence because all edges 612 // flow in that direction already. Thus adding a new one cannot form a cycle. 613 int SourceIdx = SCCIndices[&SourceSCC]; 614 int TargetIdx = SCCIndices[&TargetSCC]; 615 if (TargetIdx < SourceIdx) { 616 SourceN->setEdgeKind(TargetN, Edge::Call); 617 return false; // No new cycle. 618 } 619 620 // Compute the SCCs which (transitively) reach the source. 621 auto ComputeSourceConnectedSet = [&](SmallPtrSetImpl<SCC *> &ConnectedSet) { 622 #ifdef EXPENSIVE_CHECKS 623 // Check that the RefSCC is still valid before computing this as the 624 // results will be nonsensical of we've broken its invariants. 625 verify(); 626 #endif 627 ConnectedSet.insert(&SourceSCC); 628 auto IsConnected = [&](SCC &C) { 629 for (Node &N : C) 630 for (Edge &E : N->calls()) 631 if (ConnectedSet.count(G->lookupSCC(E.getNode()))) 632 return true; 633 634 return false; 635 }; 636 637 for (SCC *C : 638 make_range(SCCs.begin() + SourceIdx + 1, SCCs.begin() + TargetIdx + 1)) 639 if (IsConnected(*C)) 640 ConnectedSet.insert(C); 641 }; 642 643 // Use a normal worklist to find which SCCs the target connects to. We still 644 // bound the search based on the range in the postorder list we care about, 645 // but because this is forward connectivity we just "recurse" through the 646 // edges. 647 auto ComputeTargetConnectedSet = [&](SmallPtrSetImpl<SCC *> &ConnectedSet) { 648 #ifdef EXPENSIVE_CHECKS 649 // Check that the RefSCC is still valid before computing this as the 650 // results will be nonsensical of we've broken its invariants. 651 verify(); 652 #endif 653 ConnectedSet.insert(&TargetSCC); 654 SmallVector<SCC *, 4> Worklist; 655 Worklist.push_back(&TargetSCC); 656 do { 657 SCC &C = *Worklist.pop_back_val(); 658 for (Node &N : C) 659 for (Edge &E : *N) { 660 if (!E.isCall()) 661 continue; 662 SCC &EdgeC = *G->lookupSCC(E.getNode()); 663 if (&EdgeC.getOuterRefSCC() != this) 664 // Not in this RefSCC... 665 continue; 666 if (SCCIndices.find(&EdgeC)->second <= SourceIdx) 667 // Not in the postorder sequence between source and target. 668 continue; 669 670 if (ConnectedSet.insert(&EdgeC).second) 671 Worklist.push_back(&EdgeC); 672 } 673 } while (!Worklist.empty()); 674 }; 675 676 // Use a generic helper to update the postorder sequence of SCCs and return 677 // a range of any SCCs connected into a cycle by inserting this edge. This 678 // routine will also take care of updating the indices into the postorder 679 // sequence. 680 auto MergeRange = updatePostorderSequenceForEdgeInsertion( 681 SourceSCC, TargetSCC, SCCs, SCCIndices, ComputeSourceConnectedSet, 682 ComputeTargetConnectedSet); 683 684 // Run the user's callback on the merged SCCs before we actually merge them. 685 if (MergeCB) 686 MergeCB(makeArrayRef(MergeRange.begin(), MergeRange.end())); 687 688 // If the merge range is empty, then adding the edge didn't actually form any 689 // new cycles. We're done. 690 if (MergeRange.empty()) { 691 // Now that the SCC structure is finalized, flip the kind to call. 692 SourceN->setEdgeKind(TargetN, Edge::Call); 693 return false; // No new cycle. 694 } 695 696 #ifdef EXPENSIVE_CHECKS 697 // Before merging, check that the RefSCC remains valid after all the 698 // postorder updates. 699 verify(); 700 #endif 701 702 // Otherwise we need to merge all of the SCCs in the cycle into a single 703 // result SCC. 704 // 705 // NB: We merge into the target because all of these functions were already 706 // reachable from the target, meaning any SCC-wide properties deduced about it 707 // other than the set of functions within it will not have changed. 708 for (SCC *C : MergeRange) { 709 assert(C != &TargetSCC && 710 "We merge *into* the target and shouldn't process it here!"); 711 SCCIndices.erase(C); 712 TargetSCC.Nodes.append(C->Nodes.begin(), C->Nodes.end()); 713 for (Node *N : C->Nodes) 714 G->SCCMap[N] = &TargetSCC; 715 C->clear(); 716 DeletedSCCs.push_back(C); 717 } 718 719 // Erase the merged SCCs from the list and update the indices of the 720 // remaining SCCs. 721 int IndexOffset = MergeRange.end() - MergeRange.begin(); 722 auto EraseEnd = SCCs.erase(MergeRange.begin(), MergeRange.end()); 723 for (SCC *C : make_range(EraseEnd, SCCs.end())) 724 SCCIndices[C] -= IndexOffset; 725 726 // Now that the SCC structure is finalized, flip the kind to call. 727 SourceN->setEdgeKind(TargetN, Edge::Call); 728 729 // And we're done, but we did form a new cycle. 730 return true; 731 } 732 733 void LazyCallGraph::RefSCC::switchTrivialInternalEdgeToRef(Node &SourceN, 734 Node &TargetN) { 735 assert((*SourceN)[TargetN].isCall() && "Must start with a call edge!"); 736 737 #ifdef EXPENSIVE_CHECKS 738 verify(); 739 auto VerifyOnExit = make_scope_exit([&]() { verify(); }); 740 #endif 741 742 assert(G->lookupRefSCC(SourceN) == this && 743 "Source must be in this RefSCC."); 744 assert(G->lookupRefSCC(TargetN) == this && 745 "Target must be in this RefSCC."); 746 assert(G->lookupSCC(SourceN) != G->lookupSCC(TargetN) && 747 "Source and Target must be in separate SCCs for this to be trivial!"); 748 749 // Set the edge kind. 750 SourceN->setEdgeKind(TargetN, Edge::Ref); 751 } 752 753 iterator_range<LazyCallGraph::RefSCC::iterator> 754 LazyCallGraph::RefSCC::switchInternalEdgeToRef(Node &SourceN, Node &TargetN) { 755 assert((*SourceN)[TargetN].isCall() && "Must start with a call edge!"); 756 757 #ifdef EXPENSIVE_CHECKS 758 verify(); 759 auto VerifyOnExit = make_scope_exit([&]() { verify(); }); 760 #endif 761 762 assert(G->lookupRefSCC(SourceN) == this && 763 "Source must be in this RefSCC."); 764 assert(G->lookupRefSCC(TargetN) == this && 765 "Target must be in this RefSCC."); 766 767 SCC &TargetSCC = *G->lookupSCC(TargetN); 768 assert(G->lookupSCC(SourceN) == &TargetSCC && "Source and Target must be in " 769 "the same SCC to require the " 770 "full CG update."); 771 772 // Set the edge kind. 773 SourceN->setEdgeKind(TargetN, Edge::Ref); 774 775 // Otherwise we are removing a call edge from a single SCC. This may break 776 // the cycle. In order to compute the new set of SCCs, we need to do a small 777 // DFS over the nodes within the SCC to form any sub-cycles that remain as 778 // distinct SCCs and compute a postorder over the resulting SCCs. 779 // 780 // However, we specially handle the target node. The target node is known to 781 // reach all other nodes in the original SCC by definition. This means that 782 // we want the old SCC to be replaced with an SCC containing that node as it 783 // will be the root of whatever SCC DAG results from the DFS. Assumptions 784 // about an SCC such as the set of functions called will continue to hold, 785 // etc. 786 787 SCC &OldSCC = TargetSCC; 788 SmallVector<std::pair<Node *, EdgeSequence::call_iterator>, 16> DFSStack; 789 SmallVector<Node *, 16> PendingSCCStack; 790 SmallVector<SCC *, 4> NewSCCs; 791 792 // Prepare the nodes for a fresh DFS. 793 SmallVector<Node *, 16> Worklist; 794 Worklist.swap(OldSCC.Nodes); 795 for (Node *N : Worklist) { 796 N->DFSNumber = N->LowLink = 0; 797 G->SCCMap.erase(N); 798 } 799 800 // Force the target node to be in the old SCC. This also enables us to take 801 // a very significant short-cut in the standard Tarjan walk to re-form SCCs 802 // below: whenever we build an edge that reaches the target node, we know 803 // that the target node eventually connects back to all other nodes in our 804 // walk. As a consequence, we can detect and handle participants in that 805 // cycle without walking all the edges that form this connection, and instead 806 // by relying on the fundamental guarantee coming into this operation (all 807 // nodes are reachable from the target due to previously forming an SCC). 808 TargetN.DFSNumber = TargetN.LowLink = -1; 809 OldSCC.Nodes.push_back(&TargetN); 810 G->SCCMap[&TargetN] = &OldSCC; 811 812 // Scan down the stack and DFS across the call edges. 813 for (Node *RootN : Worklist) { 814 assert(DFSStack.empty() && 815 "Cannot begin a new root with a non-empty DFS stack!"); 816 assert(PendingSCCStack.empty() && 817 "Cannot begin a new root with pending nodes for an SCC!"); 818 819 // Skip any nodes we've already reached in the DFS. 820 if (RootN->DFSNumber != 0) { 821 assert(RootN->DFSNumber == -1 && 822 "Shouldn't have any mid-DFS root nodes!"); 823 continue; 824 } 825 826 RootN->DFSNumber = RootN->LowLink = 1; 827 int NextDFSNumber = 2; 828 829 DFSStack.push_back({RootN, (*RootN)->call_begin()}); 830 do { 831 Node *N; 832 EdgeSequence::call_iterator I; 833 std::tie(N, I) = DFSStack.pop_back_val(); 834 auto E = (*N)->call_end(); 835 while (I != E) { 836 Node &ChildN = I->getNode(); 837 if (ChildN.DFSNumber == 0) { 838 // We haven't yet visited this child, so descend, pushing the current 839 // node onto the stack. 840 DFSStack.push_back({N, I}); 841 842 assert(!G->SCCMap.count(&ChildN) && 843 "Found a node with 0 DFS number but already in an SCC!"); 844 ChildN.DFSNumber = ChildN.LowLink = NextDFSNumber++; 845 N = &ChildN; 846 I = (*N)->call_begin(); 847 E = (*N)->call_end(); 848 continue; 849 } 850 851 // Check for the child already being part of some component. 852 if (ChildN.DFSNumber == -1) { 853 if (G->lookupSCC(ChildN) == &OldSCC) { 854 // If the child is part of the old SCC, we know that it can reach 855 // every other node, so we have formed a cycle. Pull the entire DFS 856 // and pending stacks into it. See the comment above about setting 857 // up the old SCC for why we do this. 858 int OldSize = OldSCC.size(); 859 OldSCC.Nodes.push_back(N); 860 OldSCC.Nodes.append(PendingSCCStack.begin(), PendingSCCStack.end()); 861 PendingSCCStack.clear(); 862 while (!DFSStack.empty()) 863 OldSCC.Nodes.push_back(DFSStack.pop_back_val().first); 864 for (Node &N : drop_begin(OldSCC, OldSize)) { 865 N.DFSNumber = N.LowLink = -1; 866 G->SCCMap[&N] = &OldSCC; 867 } 868 N = nullptr; 869 break; 870 } 871 872 // If the child has already been added to some child component, it 873 // couldn't impact the low-link of this parent because it isn't 874 // connected, and thus its low-link isn't relevant so skip it. 875 ++I; 876 continue; 877 } 878 879 // Track the lowest linked child as the lowest link for this node. 880 assert(ChildN.LowLink > 0 && "Must have a positive low-link number!"); 881 if (ChildN.LowLink < N->LowLink) 882 N->LowLink = ChildN.LowLink; 883 884 // Move to the next edge. 885 ++I; 886 } 887 if (!N) 888 // Cleared the DFS early, start another round. 889 break; 890 891 // We've finished processing N and its descendants, put it on our pending 892 // SCC stack to eventually get merged into an SCC of nodes. 893 PendingSCCStack.push_back(N); 894 895 // If this node is linked to some lower entry, continue walking up the 896 // stack. 897 if (N->LowLink != N->DFSNumber) 898 continue; 899 900 // Otherwise, we've completed an SCC. Append it to our post order list of 901 // SCCs. 902 int RootDFSNumber = N->DFSNumber; 903 // Find the range of the node stack by walking down until we pass the 904 // root DFS number. 905 auto SCCNodes = make_range( 906 PendingSCCStack.rbegin(), 907 find_if(reverse(PendingSCCStack), [RootDFSNumber](const Node *N) { 908 return N->DFSNumber < RootDFSNumber; 909 })); 910 911 // Form a new SCC out of these nodes and then clear them off our pending 912 // stack. 913 NewSCCs.push_back(G->createSCC(*this, SCCNodes)); 914 for (Node &N : *NewSCCs.back()) { 915 N.DFSNumber = N.LowLink = -1; 916 G->SCCMap[&N] = NewSCCs.back(); 917 } 918 PendingSCCStack.erase(SCCNodes.end().base(), PendingSCCStack.end()); 919 } while (!DFSStack.empty()); 920 } 921 922 // Insert the remaining SCCs before the old one. The old SCC can reach all 923 // other SCCs we form because it contains the target node of the removed edge 924 // of the old SCC. This means that we will have edges into all of the new 925 // SCCs, which means the old one must come last for postorder. 926 int OldIdx = SCCIndices[&OldSCC]; 927 SCCs.insert(SCCs.begin() + OldIdx, NewSCCs.begin(), NewSCCs.end()); 928 929 // Update the mapping from SCC* to index to use the new SCC*s, and remove the 930 // old SCC from the mapping. 931 for (int Idx = OldIdx, Size = SCCs.size(); Idx < Size; ++Idx) 932 SCCIndices[SCCs[Idx]] = Idx; 933 934 return make_range(SCCs.begin() + OldIdx, 935 SCCs.begin() + OldIdx + NewSCCs.size()); 936 } 937 938 void LazyCallGraph::RefSCC::switchOutgoingEdgeToCall(Node &SourceN, 939 Node &TargetN) { 940 assert(!(*SourceN)[TargetN].isCall() && "Must start with a ref edge!"); 941 942 assert(G->lookupRefSCC(SourceN) == this && "Source must be in this RefSCC."); 943 assert(G->lookupRefSCC(TargetN) != this && 944 "Target must not be in this RefSCC."); 945 #ifdef EXPENSIVE_CHECKS 946 assert(G->lookupRefSCC(TargetN)->isDescendantOf(*this) && 947 "Target must be a descendant of the Source."); 948 #endif 949 950 // Edges between RefSCCs are the same regardless of call or ref, so we can 951 // just flip the edge here. 952 SourceN->setEdgeKind(TargetN, Edge::Call); 953 954 #ifdef EXPENSIVE_CHECKS 955 verify(); 956 #endif 957 } 958 959 void LazyCallGraph::RefSCC::switchOutgoingEdgeToRef(Node &SourceN, 960 Node &TargetN) { 961 assert((*SourceN)[TargetN].isCall() && "Must start with a call edge!"); 962 963 assert(G->lookupRefSCC(SourceN) == this && "Source must be in this RefSCC."); 964 assert(G->lookupRefSCC(TargetN) != this && 965 "Target must not be in this RefSCC."); 966 #ifdef EXPENSIVE_CHECKS 967 assert(G->lookupRefSCC(TargetN)->isDescendantOf(*this) && 968 "Target must be a descendant of the Source."); 969 #endif 970 971 // Edges between RefSCCs are the same regardless of call or ref, so we can 972 // just flip the edge here. 973 SourceN->setEdgeKind(TargetN, Edge::Ref); 974 975 #ifdef EXPENSIVE_CHECKS 976 verify(); 977 #endif 978 } 979 980 void LazyCallGraph::RefSCC::insertInternalRefEdge(Node &SourceN, 981 Node &TargetN) { 982 assert(G->lookupRefSCC(SourceN) == this && "Source must be in this RefSCC."); 983 assert(G->lookupRefSCC(TargetN) == this && "Target must be in this RefSCC."); 984 985 SourceN->insertEdgeInternal(TargetN, Edge::Ref); 986 987 #ifdef EXPENSIVE_CHECKS 988 verify(); 989 #endif 990 } 991 992 void LazyCallGraph::RefSCC::insertOutgoingEdge(Node &SourceN, Node &TargetN, 993 Edge::Kind EK) { 994 // First insert it into the caller. 995 SourceN->insertEdgeInternal(TargetN, EK); 996 997 assert(G->lookupRefSCC(SourceN) == this && "Source must be in this RefSCC."); 998 999 assert(G->lookupRefSCC(TargetN) != this && 1000 "Target must not be in this RefSCC."); 1001 #ifdef EXPENSIVE_CHECKS 1002 assert(G->lookupRefSCC(TargetN)->isDescendantOf(*this) && 1003 "Target must be a descendant of the Source."); 1004 #endif 1005 1006 #ifdef EXPENSIVE_CHECKS 1007 verify(); 1008 #endif 1009 } 1010 1011 SmallVector<LazyCallGraph::RefSCC *, 1> 1012 LazyCallGraph::RefSCC::insertIncomingRefEdge(Node &SourceN, Node &TargetN) { 1013 assert(G->lookupRefSCC(TargetN) == this && "Target must be in this RefSCC."); 1014 RefSCC &SourceC = *G->lookupRefSCC(SourceN); 1015 assert(&SourceC != this && "Source must not be in this RefSCC."); 1016 #ifdef EXPENSIVE_CHECKS 1017 assert(SourceC.isDescendantOf(*this) && 1018 "Source must be a descendant of the Target."); 1019 #endif 1020 1021 SmallVector<RefSCC *, 1> DeletedRefSCCs; 1022 1023 #ifdef EXPENSIVE_CHECKS 1024 verify(); 1025 auto VerifyOnExit = make_scope_exit([&]() { verify(); }); 1026 #endif 1027 1028 int SourceIdx = G->RefSCCIndices[&SourceC]; 1029 int TargetIdx = G->RefSCCIndices[this]; 1030 assert(SourceIdx < TargetIdx && 1031 "Postorder list doesn't see edge as incoming!"); 1032 1033 // Compute the RefSCCs which (transitively) reach the source. We do this by 1034 // working backwards from the source using the parent set in each RefSCC, 1035 // skipping any RefSCCs that don't fall in the postorder range. This has the 1036 // advantage of walking the sparser parent edge (in high fan-out graphs) but 1037 // more importantly this removes examining all forward edges in all RefSCCs 1038 // within the postorder range which aren't in fact connected. Only connected 1039 // RefSCCs (and their edges) are visited here. 1040 auto ComputeSourceConnectedSet = [&](SmallPtrSetImpl<RefSCC *> &Set) { 1041 Set.insert(&SourceC); 1042 auto IsConnected = [&](RefSCC &RC) { 1043 for (SCC &C : RC) 1044 for (Node &N : C) 1045 for (Edge &E : *N) 1046 if (Set.count(G->lookupRefSCC(E.getNode()))) 1047 return true; 1048 1049 return false; 1050 }; 1051 1052 for (RefSCC *C : make_range(G->PostOrderRefSCCs.begin() + SourceIdx + 1, 1053 G->PostOrderRefSCCs.begin() + TargetIdx + 1)) 1054 if (IsConnected(*C)) 1055 Set.insert(C); 1056 }; 1057 1058 // Use a normal worklist to find which SCCs the target connects to. We still 1059 // bound the search based on the range in the postorder list we care about, 1060 // but because this is forward connectivity we just "recurse" through the 1061 // edges. 1062 auto ComputeTargetConnectedSet = [&](SmallPtrSetImpl<RefSCC *> &Set) { 1063 Set.insert(this); 1064 SmallVector<RefSCC *, 4> Worklist; 1065 Worklist.push_back(this); 1066 do { 1067 RefSCC &RC = *Worklist.pop_back_val(); 1068 for (SCC &C : RC) 1069 for (Node &N : C) 1070 for (Edge &E : *N) { 1071 RefSCC &EdgeRC = *G->lookupRefSCC(E.getNode()); 1072 if (G->getRefSCCIndex(EdgeRC) <= SourceIdx) 1073 // Not in the postorder sequence between source and target. 1074 continue; 1075 1076 if (Set.insert(&EdgeRC).second) 1077 Worklist.push_back(&EdgeRC); 1078 } 1079 } while (!Worklist.empty()); 1080 }; 1081 1082 // Use a generic helper to update the postorder sequence of RefSCCs and return 1083 // a range of any RefSCCs connected into a cycle by inserting this edge. This 1084 // routine will also take care of updating the indices into the postorder 1085 // sequence. 1086 iterator_range<SmallVectorImpl<RefSCC *>::iterator> MergeRange = 1087 updatePostorderSequenceForEdgeInsertion( 1088 SourceC, *this, G->PostOrderRefSCCs, G->RefSCCIndices, 1089 ComputeSourceConnectedSet, ComputeTargetConnectedSet); 1090 1091 // Build a set so we can do fast tests for whether a RefSCC will end up as 1092 // part of the merged RefSCC. 1093 SmallPtrSet<RefSCC *, 16> MergeSet(MergeRange.begin(), MergeRange.end()); 1094 1095 // This RefSCC will always be part of that set, so just insert it here. 1096 MergeSet.insert(this); 1097 1098 // Now that we have identified all of the SCCs which need to be merged into 1099 // a connected set with the inserted edge, merge all of them into this SCC. 1100 SmallVector<SCC *, 16> MergedSCCs; 1101 int SCCIndex = 0; 1102 for (RefSCC *RC : MergeRange) { 1103 assert(RC != this && "We're merging into the target RefSCC, so it " 1104 "shouldn't be in the range."); 1105 1106 // Walk the inner SCCs to update their up-pointer and walk all the edges to 1107 // update any parent sets. 1108 // FIXME: We should try to find a way to avoid this (rather expensive) edge 1109 // walk by updating the parent sets in some other manner. 1110 for (SCC &InnerC : *RC) { 1111 InnerC.OuterRefSCC = this; 1112 SCCIndices[&InnerC] = SCCIndex++; 1113 for (Node &N : InnerC) 1114 G->SCCMap[&N] = &InnerC; 1115 } 1116 1117 // Now merge in the SCCs. We can actually move here so try to reuse storage 1118 // the first time through. 1119 if (MergedSCCs.empty()) 1120 MergedSCCs = std::move(RC->SCCs); 1121 else 1122 MergedSCCs.append(RC->SCCs.begin(), RC->SCCs.end()); 1123 RC->SCCs.clear(); 1124 DeletedRefSCCs.push_back(RC); 1125 } 1126 1127 // Append our original SCCs to the merged list and move it into place. 1128 for (SCC &InnerC : *this) 1129 SCCIndices[&InnerC] = SCCIndex++; 1130 MergedSCCs.append(SCCs.begin(), SCCs.end()); 1131 SCCs = std::move(MergedSCCs); 1132 1133 // Remove the merged away RefSCCs from the post order sequence. 1134 for (RefSCC *RC : MergeRange) 1135 G->RefSCCIndices.erase(RC); 1136 int IndexOffset = MergeRange.end() - MergeRange.begin(); 1137 auto EraseEnd = 1138 G->PostOrderRefSCCs.erase(MergeRange.begin(), MergeRange.end()); 1139 for (RefSCC *RC : make_range(EraseEnd, G->PostOrderRefSCCs.end())) 1140 G->RefSCCIndices[RC] -= IndexOffset; 1141 1142 // At this point we have a merged RefSCC with a post-order SCCs list, just 1143 // connect the nodes to form the new edge. 1144 SourceN->insertEdgeInternal(TargetN, Edge::Ref); 1145 1146 // We return the list of SCCs which were merged so that callers can 1147 // invalidate any data they have associated with those SCCs. Note that these 1148 // SCCs are no longer in an interesting state (they are totally empty) but 1149 // the pointers will remain stable for the life of the graph itself. 1150 return DeletedRefSCCs; 1151 } 1152 1153 void LazyCallGraph::RefSCC::removeOutgoingEdge(Node &SourceN, Node &TargetN) { 1154 assert(G->lookupRefSCC(SourceN) == this && 1155 "The source must be a member of this RefSCC."); 1156 assert(G->lookupRefSCC(TargetN) != this && 1157 "The target must not be a member of this RefSCC"); 1158 1159 #ifdef EXPENSIVE_CHECKS 1160 verify(); 1161 auto VerifyOnExit = make_scope_exit([&]() { verify(); }); 1162 #endif 1163 1164 // First remove it from the node. 1165 bool Removed = SourceN->removeEdgeInternal(TargetN); 1166 (void)Removed; 1167 assert(Removed && "Target not in the edge set for this caller?"); 1168 } 1169 1170 SmallVector<LazyCallGraph::RefSCC *, 1> 1171 LazyCallGraph::RefSCC::removeInternalRefEdge(Node &SourceN, 1172 ArrayRef<Node *> TargetNs) { 1173 // We return a list of the resulting *new* RefSCCs in post-order. 1174 SmallVector<RefSCC *, 1> Result; 1175 1176 #ifdef EXPENSIVE_CHECKS 1177 // Verify the RefSCC is valid to start with and that either we return an empty 1178 // list of result RefSCCs and this RefSCC remains valid, or we return new 1179 // RefSCCs and this RefSCC is dead. 1180 verify(); 1181 auto VerifyOnExit = make_scope_exit([&]() { 1182 // If we didn't replace our RefSCC with new ones, check that this one 1183 // remains valid. 1184 if (G) 1185 verify(); 1186 }); 1187 #endif 1188 1189 // First remove the actual edges. 1190 for (Node *TargetN : TargetNs) { 1191 assert(!(*SourceN)[*TargetN].isCall() && 1192 "Cannot remove a call edge, it must first be made a ref edge"); 1193 1194 bool Removed = SourceN->removeEdgeInternal(*TargetN); 1195 (void)Removed; 1196 assert(Removed && "Target not in the edge set for this caller?"); 1197 } 1198 1199 // Direct self references don't impact the ref graph at all. 1200 if (llvm::all_of(TargetNs, 1201 [&](Node *TargetN) { return &SourceN == TargetN; })) 1202 return Result; 1203 1204 // If all targets are in the same SCC as the source, because no call edges 1205 // were removed there is no RefSCC structure change. 1206 SCC &SourceC = *G->lookupSCC(SourceN); 1207 if (llvm::all_of(TargetNs, [&](Node *TargetN) { 1208 return G->lookupSCC(*TargetN) == &SourceC; 1209 })) 1210 return Result; 1211 1212 // We build somewhat synthetic new RefSCCs by providing a postorder mapping 1213 // for each inner SCC. We store these inside the low-link field of the nodes 1214 // rather than associated with SCCs because this saves a round-trip through 1215 // the node->SCC map and in the common case, SCCs are small. We will verify 1216 // that we always give the same number to every node in the SCC such that 1217 // these are equivalent. 1218 int PostOrderNumber = 0; 1219 1220 // Reset all the other nodes to prepare for a DFS over them, and add them to 1221 // our worklist. 1222 SmallVector<Node *, 8> Worklist; 1223 for (SCC *C : SCCs) { 1224 for (Node &N : *C) 1225 N.DFSNumber = N.LowLink = 0; 1226 1227 Worklist.append(C->Nodes.begin(), C->Nodes.end()); 1228 } 1229 1230 // Track the number of nodes in this RefSCC so that we can quickly recognize 1231 // an important special case of the edge removal not breaking the cycle of 1232 // this RefSCC. 1233 const int NumRefSCCNodes = Worklist.size(); 1234 1235 SmallVector<std::pair<Node *, EdgeSequence::iterator>, 4> DFSStack; 1236 SmallVector<Node *, 4> PendingRefSCCStack; 1237 do { 1238 assert(DFSStack.empty() && 1239 "Cannot begin a new root with a non-empty DFS stack!"); 1240 assert(PendingRefSCCStack.empty() && 1241 "Cannot begin a new root with pending nodes for an SCC!"); 1242 1243 Node *RootN = Worklist.pop_back_val(); 1244 // Skip any nodes we've already reached in the DFS. 1245 if (RootN->DFSNumber != 0) { 1246 assert(RootN->DFSNumber == -1 && 1247 "Shouldn't have any mid-DFS root nodes!"); 1248 continue; 1249 } 1250 1251 RootN->DFSNumber = RootN->LowLink = 1; 1252 int NextDFSNumber = 2; 1253 1254 DFSStack.push_back({RootN, (*RootN)->begin()}); 1255 do { 1256 Node *N; 1257 EdgeSequence::iterator I; 1258 std::tie(N, I) = DFSStack.pop_back_val(); 1259 auto E = (*N)->end(); 1260 1261 assert(N->DFSNumber != 0 && "We should always assign a DFS number " 1262 "before processing a node."); 1263 1264 while (I != E) { 1265 Node &ChildN = I->getNode(); 1266 if (ChildN.DFSNumber == 0) { 1267 // Mark that we should start at this child when next this node is the 1268 // top of the stack. We don't start at the next child to ensure this 1269 // child's lowlink is reflected. 1270 DFSStack.push_back({N, I}); 1271 1272 // Continue, resetting to the child node. 1273 ChildN.LowLink = ChildN.DFSNumber = NextDFSNumber++; 1274 N = &ChildN; 1275 I = ChildN->begin(); 1276 E = ChildN->end(); 1277 continue; 1278 } 1279 if (ChildN.DFSNumber == -1) { 1280 // If this child isn't currently in this RefSCC, no need to process 1281 // it. 1282 ++I; 1283 continue; 1284 } 1285 1286 // Track the lowest link of the children, if any are still in the stack. 1287 // Any child not on the stack will have a LowLink of -1. 1288 assert(ChildN.LowLink != 0 && 1289 "Low-link must not be zero with a non-zero DFS number."); 1290 if (ChildN.LowLink >= 0 && ChildN.LowLink < N->LowLink) 1291 N->LowLink = ChildN.LowLink; 1292 ++I; 1293 } 1294 1295 // We've finished processing N and its descendants, put it on our pending 1296 // stack to eventually get merged into a RefSCC. 1297 PendingRefSCCStack.push_back(N); 1298 1299 // If this node is linked to some lower entry, continue walking up the 1300 // stack. 1301 if (N->LowLink != N->DFSNumber) { 1302 assert(!DFSStack.empty() && 1303 "We never found a viable root for a RefSCC to pop off!"); 1304 continue; 1305 } 1306 1307 // Otherwise, form a new RefSCC from the top of the pending node stack. 1308 int RefSCCNumber = PostOrderNumber++; 1309 int RootDFSNumber = N->DFSNumber; 1310 1311 // Find the range of the node stack by walking down until we pass the 1312 // root DFS number. Update the DFS numbers and low link numbers in the 1313 // process to avoid re-walking this list where possible. 1314 auto StackRI = find_if(reverse(PendingRefSCCStack), [&](Node *N) { 1315 if (N->DFSNumber < RootDFSNumber) 1316 // We've found the bottom. 1317 return true; 1318 1319 // Update this node and keep scanning. 1320 N->DFSNumber = -1; 1321 // Save the post-order number in the lowlink field so that we can use 1322 // it to map SCCs into new RefSCCs after we finish the DFS. 1323 N->LowLink = RefSCCNumber; 1324 return false; 1325 }); 1326 auto RefSCCNodes = make_range(StackRI.base(), PendingRefSCCStack.end()); 1327 1328 // If we find a cycle containing all nodes originally in this RefSCC then 1329 // the removal hasn't changed the structure at all. This is an important 1330 // special case and we can directly exit the entire routine more 1331 // efficiently as soon as we discover it. 1332 if (llvm::size(RefSCCNodes) == NumRefSCCNodes) { 1333 // Clear out the low link field as we won't need it. 1334 for (Node *N : RefSCCNodes) 1335 N->LowLink = -1; 1336 // Return the empty result immediately. 1337 return Result; 1338 } 1339 1340 // We've already marked the nodes internally with the RefSCC number so 1341 // just clear them off the stack and continue. 1342 PendingRefSCCStack.erase(RefSCCNodes.begin(), PendingRefSCCStack.end()); 1343 } while (!DFSStack.empty()); 1344 1345 assert(DFSStack.empty() && "Didn't flush the entire DFS stack!"); 1346 assert(PendingRefSCCStack.empty() && "Didn't flush all pending nodes!"); 1347 } while (!Worklist.empty()); 1348 1349 assert(PostOrderNumber > 1 && 1350 "Should never finish the DFS when the existing RefSCC remains valid!"); 1351 1352 // Otherwise we create a collection of new RefSCC nodes and build 1353 // a radix-sort style map from postorder number to these new RefSCCs. We then 1354 // append SCCs to each of these RefSCCs in the order they occurred in the 1355 // original SCCs container. 1356 for (int i = 0; i < PostOrderNumber; ++i) 1357 Result.push_back(G->createRefSCC(*G)); 1358 1359 // Insert the resulting postorder sequence into the global graph postorder 1360 // sequence before the current RefSCC in that sequence, and then remove the 1361 // current one. 1362 // 1363 // FIXME: It'd be nice to change the APIs so that we returned an iterator 1364 // range over the global postorder sequence and generally use that sequence 1365 // rather than building a separate result vector here. 1366 int Idx = G->getRefSCCIndex(*this); 1367 G->PostOrderRefSCCs.erase(G->PostOrderRefSCCs.begin() + Idx); 1368 G->PostOrderRefSCCs.insert(G->PostOrderRefSCCs.begin() + Idx, Result.begin(), 1369 Result.end()); 1370 for (int i : seq<int>(Idx, G->PostOrderRefSCCs.size())) 1371 G->RefSCCIndices[G->PostOrderRefSCCs[i]] = i; 1372 1373 for (SCC *C : SCCs) { 1374 // We store the SCC number in the node's low-link field above. 1375 int SCCNumber = C->begin()->LowLink; 1376 // Clear out all of the SCC's node's low-link fields now that we're done 1377 // using them as side-storage. 1378 for (Node &N : *C) { 1379 assert(N.LowLink == SCCNumber && 1380 "Cannot have different numbers for nodes in the same SCC!"); 1381 N.LowLink = -1; 1382 } 1383 1384 RefSCC &RC = *Result[SCCNumber]; 1385 int SCCIndex = RC.SCCs.size(); 1386 RC.SCCs.push_back(C); 1387 RC.SCCIndices[C] = SCCIndex; 1388 C->OuterRefSCC = &RC; 1389 } 1390 1391 // Now that we've moved things into the new RefSCCs, clear out our current 1392 // one. 1393 G = nullptr; 1394 SCCs.clear(); 1395 SCCIndices.clear(); 1396 1397 #ifdef EXPENSIVE_CHECKS 1398 // Verify the new RefSCCs we've built. 1399 for (RefSCC *RC : Result) 1400 RC->verify(); 1401 #endif 1402 1403 // Return the new list of SCCs. 1404 return Result; 1405 } 1406 1407 void LazyCallGraph::RefSCC::insertTrivialCallEdge(Node &SourceN, 1408 Node &TargetN) { 1409 #ifdef EXPENSIVE_CHECKS 1410 auto ExitVerifier = make_scope_exit([this] { verify(); }); 1411 1412 // Check that we aren't breaking some invariants of the SCC graph. Note that 1413 // this is quadratic in the number of edges in the call graph! 1414 SCC &SourceC = *G->lookupSCC(SourceN); 1415 SCC &TargetC = *G->lookupSCC(TargetN); 1416 if (&SourceC != &TargetC) 1417 assert(SourceC.isAncestorOf(TargetC) && 1418 "Call edge is not trivial in the SCC graph!"); 1419 #endif 1420 1421 // First insert it into the source or find the existing edge. 1422 auto InsertResult = 1423 SourceN->EdgeIndexMap.insert({&TargetN, SourceN->Edges.size()}); 1424 if (!InsertResult.second) { 1425 // Already an edge, just update it. 1426 Edge &E = SourceN->Edges[InsertResult.first->second]; 1427 if (E.isCall()) 1428 return; // Nothing to do! 1429 E.setKind(Edge::Call); 1430 } else { 1431 // Create the new edge. 1432 SourceN->Edges.emplace_back(TargetN, Edge::Call); 1433 } 1434 } 1435 1436 void LazyCallGraph::RefSCC::insertTrivialRefEdge(Node &SourceN, Node &TargetN) { 1437 #ifdef EXPENSIVE_CHECKS 1438 auto ExitVerifier = make_scope_exit([this] { verify(); }); 1439 1440 // Check that we aren't breaking some invariants of the RefSCC graph. 1441 RefSCC &SourceRC = *G->lookupRefSCC(SourceN); 1442 RefSCC &TargetRC = *G->lookupRefSCC(TargetN); 1443 if (&SourceRC != &TargetRC) 1444 assert(SourceRC.isAncestorOf(TargetRC) && 1445 "Ref edge is not trivial in the RefSCC graph!"); 1446 #endif 1447 1448 // First insert it into the source or find the existing edge. 1449 auto InsertResult = 1450 SourceN->EdgeIndexMap.insert({&TargetN, SourceN->Edges.size()}); 1451 if (!InsertResult.second) 1452 // Already an edge, we're done. 1453 return; 1454 1455 // Create the new edge. 1456 SourceN->Edges.emplace_back(TargetN, Edge::Ref); 1457 } 1458 1459 void LazyCallGraph::RefSCC::replaceNodeFunction(Node &N, Function &NewF) { 1460 Function &OldF = N.getFunction(); 1461 1462 #ifdef EXPENSIVE_CHECKS 1463 auto ExitVerifier = make_scope_exit([this] { verify(); }); 1464 1465 assert(G->lookupRefSCC(N) == this && 1466 "Cannot replace the function of a node outside this RefSCC."); 1467 1468 assert(G->NodeMap.find(&NewF) == G->NodeMap.end() && 1469 "Must not have already walked the new function!'"); 1470 1471 // It is important that this replacement not introduce graph changes so we 1472 // insist that the caller has already removed every use of the original 1473 // function and that all uses of the new function correspond to existing 1474 // edges in the graph. The common and expected way to use this is when 1475 // replacing the function itself in the IR without changing the call graph 1476 // shape and just updating the analysis based on that. 1477 assert(&OldF != &NewF && "Cannot replace a function with itself!"); 1478 assert(OldF.use_empty() && 1479 "Must have moved all uses from the old function to the new!"); 1480 #endif 1481 1482 N.replaceFunction(NewF); 1483 1484 // Update various call graph maps. 1485 G->NodeMap.erase(&OldF); 1486 G->NodeMap[&NewF] = &N; 1487 } 1488 1489 void LazyCallGraph::insertEdge(Node &SourceN, Node &TargetN, Edge::Kind EK) { 1490 assert(SCCMap.empty() && 1491 "This method cannot be called after SCCs have been formed!"); 1492 1493 return SourceN->insertEdgeInternal(TargetN, EK); 1494 } 1495 1496 void LazyCallGraph::removeEdge(Node &SourceN, Node &TargetN) { 1497 assert(SCCMap.empty() && 1498 "This method cannot be called after SCCs have been formed!"); 1499 1500 bool Removed = SourceN->removeEdgeInternal(TargetN); 1501 (void)Removed; 1502 assert(Removed && "Target not in the edge set for this caller?"); 1503 } 1504 1505 void LazyCallGraph::removeDeadFunction(Function &F) { 1506 // FIXME: This is unnecessarily restrictive. We should be able to remove 1507 // functions which recursively call themselves. 1508 assert(F.hasZeroLiveUses() && 1509 "This routine should only be called on trivially dead functions!"); 1510 1511 // We shouldn't remove library functions as they are never really dead while 1512 // the call graph is in use -- every function definition refers to them. 1513 assert(!isLibFunction(F) && 1514 "Must not remove lib functions from the call graph!"); 1515 1516 auto NI = NodeMap.find(&F); 1517 if (NI == NodeMap.end()) 1518 // Not in the graph at all! 1519 return; 1520 1521 Node &N = *NI->second; 1522 NodeMap.erase(NI); 1523 1524 // Remove this from the entry edges if present. 1525 EntryEdges.removeEdgeInternal(N); 1526 1527 // Cannot remove a function which has yet to be visited in the DFS walk, so 1528 // if we have a node at all then we must have an SCC and RefSCC. 1529 auto CI = SCCMap.find(&N); 1530 assert(CI != SCCMap.end() && 1531 "Tried to remove a node without an SCC after DFS walk started!"); 1532 SCC &C = *CI->second; 1533 SCCMap.erase(CI); 1534 RefSCC &RC = C.getOuterRefSCC(); 1535 1536 // This node must be the only member of its SCC as it has no callers, and 1537 // that SCC must be the only member of a RefSCC as it has no references. 1538 // Validate these properties first. 1539 assert(C.size() == 1 && "Dead functions must be in a singular SCC"); 1540 assert(RC.size() == 1 && "Dead functions must be in a singular RefSCC"); 1541 1542 // Finally clear out all the data structures from the node down through the 1543 // components. postorder_ref_scc_iterator will skip empty RefSCCs, so no need 1544 // to adjust LazyCallGraph data structures. 1545 N.clear(); 1546 N.G = nullptr; 1547 N.F = nullptr; 1548 C.clear(); 1549 RC.clear(); 1550 RC.G = nullptr; 1551 1552 // Nothing to delete as all the objects are allocated in stable bump pointer 1553 // allocators. 1554 } 1555 1556 // Gets the Edge::Kind from one function to another by looking at the function's 1557 // instructions. Asserts if there is no edge. 1558 // Useful for determining what type of edge should exist between functions when 1559 // the edge hasn't been created yet. 1560 static LazyCallGraph::Edge::Kind getEdgeKind(Function &OriginalFunction, 1561 Function &NewFunction) { 1562 // In release builds, assume that if there are no direct calls to the new 1563 // function, then there is a ref edge. In debug builds, keep track of 1564 // references to assert that there is actually a ref edge if there is no call 1565 // edge. 1566 #ifndef NDEBUG 1567 SmallVector<Constant *, 16> Worklist; 1568 SmallPtrSet<Constant *, 16> Visited; 1569 #endif 1570 1571 for (Instruction &I : instructions(OriginalFunction)) { 1572 if (auto *CB = dyn_cast<CallBase>(&I)) { 1573 if (Function *Callee = CB->getCalledFunction()) { 1574 if (Callee == &NewFunction) 1575 return LazyCallGraph::Edge::Kind::Call; 1576 } 1577 } 1578 #ifndef NDEBUG 1579 for (Value *Op : I.operand_values()) { 1580 if (Constant *C = dyn_cast<Constant>(Op)) { 1581 if (Visited.insert(C).second) 1582 Worklist.push_back(C); 1583 } 1584 } 1585 #endif 1586 } 1587 1588 #ifndef NDEBUG 1589 bool FoundNewFunction = false; 1590 LazyCallGraph::visitReferences(Worklist, Visited, [&](Function &F) { 1591 if (&F == &NewFunction) 1592 FoundNewFunction = true; 1593 }); 1594 assert(FoundNewFunction && "No edge from original function to new function"); 1595 #endif 1596 1597 return LazyCallGraph::Edge::Kind::Ref; 1598 } 1599 1600 void LazyCallGraph::addSplitFunction(Function &OriginalFunction, 1601 Function &NewFunction) { 1602 assert(lookup(OriginalFunction) && 1603 "Original function's node should already exist"); 1604 Node &OriginalN = get(OriginalFunction); 1605 SCC *OriginalC = lookupSCC(OriginalN); 1606 RefSCC *OriginalRC = lookupRefSCC(OriginalN); 1607 1608 #ifdef EXPENSIVE_CHECKS 1609 OriginalRC->verify(); 1610 auto VerifyOnExit = make_scope_exit([&]() { OriginalRC->verify(); }); 1611 #endif 1612 1613 assert(!lookup(NewFunction) && 1614 "New function's node should not already exist"); 1615 Node &NewN = initNode(NewFunction); 1616 1617 Edge::Kind EK = getEdgeKind(OriginalFunction, NewFunction); 1618 1619 SCC *NewC = nullptr; 1620 for (Edge &E : *NewN) { 1621 Node &EN = E.getNode(); 1622 if (EK == Edge::Kind::Call && E.isCall() && lookupSCC(EN) == OriginalC) { 1623 // If the edge to the new function is a call edge and there is a call edge 1624 // from the new function to any function in the original function's SCC, 1625 // it is in the same SCC (and RefSCC) as the original function. 1626 NewC = OriginalC; 1627 NewC->Nodes.push_back(&NewN); 1628 break; 1629 } 1630 } 1631 1632 if (!NewC) { 1633 for (Edge &E : *NewN) { 1634 Node &EN = E.getNode(); 1635 if (lookupRefSCC(EN) == OriginalRC) { 1636 // If there is any edge from the new function to any function in the 1637 // original function's RefSCC, it is in the same RefSCC as the original 1638 // function but a new SCC. 1639 RefSCC *NewRC = OriginalRC; 1640 NewC = createSCC(*NewRC, SmallVector<Node *, 1>({&NewN})); 1641 1642 // The new function's SCC is not the same as the original function's 1643 // SCC, since that case was handled earlier. If the edge from the 1644 // original function to the new function was a call edge, then we need 1645 // to insert the newly created function's SCC before the original 1646 // function's SCC. Otherwise either the new SCC comes after the original 1647 // function's SCC, or it doesn't matter, and in both cases we can add it 1648 // to the very end. 1649 int InsertIndex = EK == Edge::Kind::Call ? NewRC->SCCIndices[OriginalC] 1650 : NewRC->SCCIndices.size(); 1651 NewRC->SCCs.insert(NewRC->SCCs.begin() + InsertIndex, NewC); 1652 for (int I = InsertIndex, Size = NewRC->SCCs.size(); I < Size; ++I) 1653 NewRC->SCCIndices[NewRC->SCCs[I]] = I; 1654 1655 break; 1656 } 1657 } 1658 } 1659 1660 if (!NewC) { 1661 // We didn't find any edges back to the original function's RefSCC, so the 1662 // new function belongs in a new RefSCC. The new RefSCC goes before the 1663 // original function's RefSCC. 1664 RefSCC *NewRC = createRefSCC(*this); 1665 NewC = createSCC(*NewRC, SmallVector<Node *, 1>({&NewN})); 1666 NewRC->SCCIndices[NewC] = 0; 1667 NewRC->SCCs.push_back(NewC); 1668 auto OriginalRCIndex = RefSCCIndices.find(OriginalRC)->second; 1669 PostOrderRefSCCs.insert(PostOrderRefSCCs.begin() + OriginalRCIndex, NewRC); 1670 for (int I = OriginalRCIndex, Size = PostOrderRefSCCs.size(); I < Size; ++I) 1671 RefSCCIndices[PostOrderRefSCCs[I]] = I; 1672 } 1673 1674 SCCMap[&NewN] = NewC; 1675 1676 OriginalN->insertEdgeInternal(NewN, EK); 1677 } 1678 1679 void LazyCallGraph::addSplitRefRecursiveFunctions( 1680 Function &OriginalFunction, ArrayRef<Function *> NewFunctions) { 1681 assert(!NewFunctions.empty() && "Can't add zero functions"); 1682 assert(lookup(OriginalFunction) && 1683 "Original function's node should already exist"); 1684 Node &OriginalN = get(OriginalFunction); 1685 RefSCC *OriginalRC = lookupRefSCC(OriginalN); 1686 1687 #ifdef EXPENSIVE_CHECKS 1688 OriginalRC->verify(); 1689 auto VerifyOnExit = make_scope_exit([&]() { 1690 OriginalRC->verify(); 1691 for (Function *NewFunction : NewFunctions) 1692 lookupRefSCC(get(*NewFunction))->verify(); 1693 }); 1694 #endif 1695 1696 bool ExistsRefToOriginalRefSCC = false; 1697 1698 for (Function *NewFunction : NewFunctions) { 1699 Node &NewN = initNode(*NewFunction); 1700 1701 OriginalN->insertEdgeInternal(NewN, Edge::Kind::Ref); 1702 1703 // Check if there is any edge from any new function back to any function in 1704 // the original function's RefSCC. 1705 for (Edge &E : *NewN) { 1706 if (lookupRefSCC(E.getNode()) == OriginalRC) { 1707 ExistsRefToOriginalRefSCC = true; 1708 break; 1709 } 1710 } 1711 } 1712 1713 RefSCC *NewRC; 1714 if (ExistsRefToOriginalRefSCC) { 1715 // If there is any edge from any new function to any function in the 1716 // original function's RefSCC, all new functions will be in the same RefSCC 1717 // as the original function. 1718 NewRC = OriginalRC; 1719 } else { 1720 // Otherwise the new functions are in their own RefSCC. 1721 NewRC = createRefSCC(*this); 1722 // The new RefSCC goes before the original function's RefSCC in postorder 1723 // since there are only edges from the original function's RefSCC to the new 1724 // RefSCC. 1725 auto OriginalRCIndex = RefSCCIndices.find(OriginalRC)->second; 1726 PostOrderRefSCCs.insert(PostOrderRefSCCs.begin() + OriginalRCIndex, NewRC); 1727 for (int I = OriginalRCIndex, Size = PostOrderRefSCCs.size(); I < Size; ++I) 1728 RefSCCIndices[PostOrderRefSCCs[I]] = I; 1729 } 1730 1731 for (Function *NewFunction : NewFunctions) { 1732 Node &NewN = get(*NewFunction); 1733 // Each new function is in its own new SCC. The original function can only 1734 // have a ref edge to new functions, and no other existing functions can 1735 // have references to new functions. Each new function only has a ref edge 1736 // to the other new functions. 1737 SCC *NewC = createSCC(*NewRC, SmallVector<Node *, 1>({&NewN})); 1738 // The new SCCs are either sibling SCCs or parent SCCs to all other existing 1739 // SCCs in the RefSCC. Either way, they can go at the back of the postorder 1740 // SCC list. 1741 auto Index = NewRC->SCCIndices.size(); 1742 NewRC->SCCIndices[NewC] = Index; 1743 NewRC->SCCs.push_back(NewC); 1744 SCCMap[&NewN] = NewC; 1745 } 1746 1747 #ifndef NDEBUG 1748 for (Function *F1 : NewFunctions) { 1749 assert(getEdgeKind(OriginalFunction, *F1) == Edge::Kind::Ref && 1750 "Expected ref edges from original function to every new function"); 1751 Node &N1 = get(*F1); 1752 for (Function *F2 : NewFunctions) { 1753 if (F1 == F2) 1754 continue; 1755 Node &N2 = get(*F2); 1756 assert(!N1->lookup(N2)->isCall() && 1757 "Edges between new functions must be ref edges"); 1758 } 1759 } 1760 #endif 1761 } 1762 1763 LazyCallGraph::Node &LazyCallGraph::insertInto(Function &F, Node *&MappedN) { 1764 return *new (MappedN = BPA.Allocate()) Node(*this, F); 1765 } 1766 1767 void LazyCallGraph::updateGraphPtrs() { 1768 // Walk the node map to update their graph pointers. While this iterates in 1769 // an unstable order, the order has no effect so it remains correct. 1770 for (auto &FunctionNodePair : NodeMap) 1771 FunctionNodePair.second->G = this; 1772 1773 for (auto *RC : PostOrderRefSCCs) 1774 RC->G = this; 1775 } 1776 1777 LazyCallGraph::Node &LazyCallGraph::initNode(Function &F) { 1778 Node &N = get(F); 1779 N.DFSNumber = N.LowLink = -1; 1780 N.populate(); 1781 NodeMap[&F] = &N; 1782 return N; 1783 } 1784 1785 template <typename RootsT, typename GetBeginT, typename GetEndT, 1786 typename GetNodeT, typename FormSCCCallbackT> 1787 void LazyCallGraph::buildGenericSCCs(RootsT &&Roots, GetBeginT &&GetBegin, 1788 GetEndT &&GetEnd, GetNodeT &&GetNode, 1789 FormSCCCallbackT &&FormSCC) { 1790 using EdgeItT = decltype(GetBegin(std::declval<Node &>())); 1791 1792 SmallVector<std::pair<Node *, EdgeItT>, 16> DFSStack; 1793 SmallVector<Node *, 16> PendingSCCStack; 1794 1795 // Scan down the stack and DFS across the call edges. 1796 for (Node *RootN : Roots) { 1797 assert(DFSStack.empty() && 1798 "Cannot begin a new root with a non-empty DFS stack!"); 1799 assert(PendingSCCStack.empty() && 1800 "Cannot begin a new root with pending nodes for an SCC!"); 1801 1802 // Skip any nodes we've already reached in the DFS. 1803 if (RootN->DFSNumber != 0) { 1804 assert(RootN->DFSNumber == -1 && 1805 "Shouldn't have any mid-DFS root nodes!"); 1806 continue; 1807 } 1808 1809 RootN->DFSNumber = RootN->LowLink = 1; 1810 int NextDFSNumber = 2; 1811 1812 DFSStack.push_back({RootN, GetBegin(*RootN)}); 1813 do { 1814 Node *N; 1815 EdgeItT I; 1816 std::tie(N, I) = DFSStack.pop_back_val(); 1817 auto E = GetEnd(*N); 1818 while (I != E) { 1819 Node &ChildN = GetNode(I); 1820 if (ChildN.DFSNumber == 0) { 1821 // We haven't yet visited this child, so descend, pushing the current 1822 // node onto the stack. 1823 DFSStack.push_back({N, I}); 1824 1825 ChildN.DFSNumber = ChildN.LowLink = NextDFSNumber++; 1826 N = &ChildN; 1827 I = GetBegin(*N); 1828 E = GetEnd(*N); 1829 continue; 1830 } 1831 1832 // If the child has already been added to some child component, it 1833 // couldn't impact the low-link of this parent because it isn't 1834 // connected, and thus its low-link isn't relevant so skip it. 1835 if (ChildN.DFSNumber == -1) { 1836 ++I; 1837 continue; 1838 } 1839 1840 // Track the lowest linked child as the lowest link for this node. 1841 assert(ChildN.LowLink > 0 && "Must have a positive low-link number!"); 1842 if (ChildN.LowLink < N->LowLink) 1843 N->LowLink = ChildN.LowLink; 1844 1845 // Move to the next edge. 1846 ++I; 1847 } 1848 1849 // We've finished processing N and its descendants, put it on our pending 1850 // SCC stack to eventually get merged into an SCC of nodes. 1851 PendingSCCStack.push_back(N); 1852 1853 // If this node is linked to some lower entry, continue walking up the 1854 // stack. 1855 if (N->LowLink != N->DFSNumber) 1856 continue; 1857 1858 // Otherwise, we've completed an SCC. Append it to our post order list of 1859 // SCCs. 1860 int RootDFSNumber = N->DFSNumber; 1861 // Find the range of the node stack by walking down until we pass the 1862 // root DFS number. 1863 auto SCCNodes = make_range( 1864 PendingSCCStack.rbegin(), 1865 find_if(reverse(PendingSCCStack), [RootDFSNumber](const Node *N) { 1866 return N->DFSNumber < RootDFSNumber; 1867 })); 1868 // Form a new SCC out of these nodes and then clear them off our pending 1869 // stack. 1870 FormSCC(SCCNodes); 1871 PendingSCCStack.erase(SCCNodes.end().base(), PendingSCCStack.end()); 1872 } while (!DFSStack.empty()); 1873 } 1874 } 1875 1876 /// Build the internal SCCs for a RefSCC from a sequence of nodes. 1877 /// 1878 /// Appends the SCCs to the provided vector and updates the map with their 1879 /// indices. Both the vector and map must be empty when passed into this 1880 /// routine. 1881 void LazyCallGraph::buildSCCs(RefSCC &RC, node_stack_range Nodes) { 1882 assert(RC.SCCs.empty() && "Already built SCCs!"); 1883 assert(RC.SCCIndices.empty() && "Already mapped SCC indices!"); 1884 1885 for (Node *N : Nodes) { 1886 assert(N->LowLink >= (*Nodes.begin())->LowLink && 1887 "We cannot have a low link in an SCC lower than its root on the " 1888 "stack!"); 1889 1890 // This node will go into the next RefSCC, clear out its DFS and low link 1891 // as we scan. 1892 N->DFSNumber = N->LowLink = 0; 1893 } 1894 1895 // Each RefSCC contains a DAG of the call SCCs. To build these, we do 1896 // a direct walk of the call edges using Tarjan's algorithm. We reuse the 1897 // internal storage as we won't need it for the outer graph's DFS any longer. 1898 buildGenericSCCs( 1899 Nodes, [](Node &N) { return N->call_begin(); }, 1900 [](Node &N) { return N->call_end(); }, 1901 [](EdgeSequence::call_iterator I) -> Node & { return I->getNode(); }, 1902 [this, &RC](node_stack_range Nodes) { 1903 RC.SCCs.push_back(createSCC(RC, Nodes)); 1904 for (Node &N : *RC.SCCs.back()) { 1905 N.DFSNumber = N.LowLink = -1; 1906 SCCMap[&N] = RC.SCCs.back(); 1907 } 1908 }); 1909 1910 // Wire up the SCC indices. 1911 for (int i = 0, Size = RC.SCCs.size(); i < Size; ++i) 1912 RC.SCCIndices[RC.SCCs[i]] = i; 1913 } 1914 1915 void LazyCallGraph::buildRefSCCs() { 1916 if (EntryEdges.empty() || !PostOrderRefSCCs.empty()) 1917 // RefSCCs are either non-existent or already built! 1918 return; 1919 1920 assert(RefSCCIndices.empty() && "Already mapped RefSCC indices!"); 1921 1922 SmallVector<Node *, 16> Roots; 1923 for (Edge &E : *this) 1924 Roots.push_back(&E.getNode()); 1925 1926 // The roots will be iterated in order. 1927 buildGenericSCCs( 1928 Roots, 1929 [](Node &N) { 1930 // We need to populate each node as we begin to walk its edges. 1931 N.populate(); 1932 return N->begin(); 1933 }, 1934 [](Node &N) { return N->end(); }, 1935 [](EdgeSequence::iterator I) -> Node & { return I->getNode(); }, 1936 [this](node_stack_range Nodes) { 1937 RefSCC *NewRC = createRefSCC(*this); 1938 buildSCCs(*NewRC, Nodes); 1939 1940 // Push the new node into the postorder list and remember its position 1941 // in the index map. 1942 bool Inserted = 1943 RefSCCIndices.insert({NewRC, PostOrderRefSCCs.size()}).second; 1944 (void)Inserted; 1945 assert(Inserted && "Cannot already have this RefSCC in the index map!"); 1946 PostOrderRefSCCs.push_back(NewRC); 1947 #ifdef EXPENSIVE_CHECKS 1948 NewRC->verify(); 1949 #endif 1950 }); 1951 } 1952 1953 void LazyCallGraph::visitReferences(SmallVectorImpl<Constant *> &Worklist, 1954 SmallPtrSetImpl<Constant *> &Visited, 1955 function_ref<void(Function &)> Callback) { 1956 while (!Worklist.empty()) { 1957 Constant *C = Worklist.pop_back_val(); 1958 1959 if (Function *F = dyn_cast<Function>(C)) { 1960 if (!F->isDeclaration()) 1961 Callback(*F); 1962 continue; 1963 } 1964 1965 // blockaddresses are weird and don't participate in the call graph anyway, 1966 // skip them. 1967 if (isa<BlockAddress>(C)) 1968 continue; 1969 1970 for (Value *Op : C->operand_values()) 1971 if (Visited.insert(cast<Constant>(Op)).second) 1972 Worklist.push_back(cast<Constant>(Op)); 1973 } 1974 } 1975 1976 AnalysisKey LazyCallGraphAnalysis::Key; 1977 1978 LazyCallGraphPrinterPass::LazyCallGraphPrinterPass(raw_ostream &OS) : OS(OS) {} 1979 1980 static void printNode(raw_ostream &OS, LazyCallGraph::Node &N) { 1981 OS << " Edges in function: " << N.getFunction().getName() << "\n"; 1982 for (LazyCallGraph::Edge &E : N.populate()) 1983 OS << " " << (E.isCall() ? "call" : "ref ") << " -> " 1984 << E.getFunction().getName() << "\n"; 1985 1986 OS << "\n"; 1987 } 1988 1989 static void printSCC(raw_ostream &OS, LazyCallGraph::SCC &C) { 1990 OS << " SCC with " << C.size() << " functions:\n"; 1991 1992 for (LazyCallGraph::Node &N : C) 1993 OS << " " << N.getFunction().getName() << "\n"; 1994 } 1995 1996 static void printRefSCC(raw_ostream &OS, LazyCallGraph::RefSCC &C) { 1997 OS << " RefSCC with " << C.size() << " call SCCs:\n"; 1998 1999 for (LazyCallGraph::SCC &InnerC : C) 2000 printSCC(OS, InnerC); 2001 2002 OS << "\n"; 2003 } 2004 2005 PreservedAnalyses LazyCallGraphPrinterPass::run(Module &M, 2006 ModuleAnalysisManager &AM) { 2007 LazyCallGraph &G = AM.getResult<LazyCallGraphAnalysis>(M); 2008 2009 OS << "Printing the call graph for module: " << M.getModuleIdentifier() 2010 << "\n\n"; 2011 2012 for (Function &F : M) 2013 printNode(OS, G.get(F)); 2014 2015 G.buildRefSCCs(); 2016 for (LazyCallGraph::RefSCC &C : G.postorder_ref_sccs()) 2017 printRefSCC(OS, C); 2018 2019 return PreservedAnalyses::all(); 2020 } 2021 2022 LazyCallGraphDOTPrinterPass::LazyCallGraphDOTPrinterPass(raw_ostream &OS) 2023 : OS(OS) {} 2024 2025 static void printNodeDOT(raw_ostream &OS, LazyCallGraph::Node &N) { 2026 std::string Name = 2027 "\"" + DOT::EscapeString(std::string(N.getFunction().getName())) + "\""; 2028 2029 for (LazyCallGraph::Edge &E : N.populate()) { 2030 OS << " " << Name << " -> \"" 2031 << DOT::EscapeString(std::string(E.getFunction().getName())) << "\""; 2032 if (!E.isCall()) // It is a ref edge. 2033 OS << " [style=dashed,label=\"ref\"]"; 2034 OS << ";\n"; 2035 } 2036 2037 OS << "\n"; 2038 } 2039 2040 PreservedAnalyses LazyCallGraphDOTPrinterPass::run(Module &M, 2041 ModuleAnalysisManager &AM) { 2042 LazyCallGraph &G = AM.getResult<LazyCallGraphAnalysis>(M); 2043 2044 OS << "digraph \"" << DOT::EscapeString(M.getModuleIdentifier()) << "\" {\n"; 2045 2046 for (Function &F : M) 2047 printNodeDOT(OS, G.get(F)); 2048 2049 OS << "}\n"; 2050 2051 return PreservedAnalyses::all(); 2052 } 2053