1 //===- LazyCallGraph.cpp - Analysis of a Module's call graph --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 #include "llvm/Analysis/LazyCallGraph.h"
11 #include "llvm/ADT/STLExtras.h"
12 #include "llvm/IR/CallSite.h"
13 #include "llvm/IR/InstVisitor.h"
14 #include "llvm/IR/Instructions.h"
15 #include "llvm/IR/PassManager.h"
16 #include "llvm/Support/Debug.h"
17 #include "llvm/Support/raw_ostream.h"
21 #define DEBUG_TYPE "lcg"
23 static void findCallees(
24 SmallVectorImpl<Constant *> &Worklist, SmallPtrSetImpl<Constant *> &Visited,
25 SmallVectorImpl<PointerUnion<Function *, LazyCallGraph::Node *>> &Callees,
26 DenseMap<Function *, size_t> &CalleeIndexMap) {
27 while (!Worklist.empty()) {
28 Constant *C = Worklist.pop_back_val();
30 if (Function *F = dyn_cast<Function>(C)) {
31 // Note that we consider *any* function with a definition to be a viable
32 // edge. Even if the function's definition is subject to replacement by
33 // some other module (say, a weak definition) there may still be
34 // optimizations which essentially speculate based on the definition and
35 // a way to check that the specific definition is in fact the one being
36 // used. For example, this could be done by moving the weak definition to
37 // a strong (internal) definition and making the weak definition be an
38 // alias. Then a test of the address of the weak function against the new
39 // strong definition's address would be an effective way to determine the
40 // safety of optimizing a direct call edge.
41 if (!F->isDeclaration() &&
42 CalleeIndexMap.insert(std::make_pair(F, Callees.size())).second) {
43 DEBUG(dbgs() << " Added callable function: " << F->getName()
50 for (Value *Op : C->operand_values())
51 if (Visited.insert(cast<Constant>(Op)))
52 Worklist.push_back(cast<Constant>(Op));
56 LazyCallGraph::Node::Node(LazyCallGraph &G, Function &F)
57 : G(&G), F(F), DFSNumber(0), LowLink(0) {
58 DEBUG(dbgs() << " Adding functions called by '" << F.getName()
59 << "' to the graph.\n");
61 SmallVector<Constant *, 16> Worklist;
62 SmallPtrSet<Constant *, 16> Visited;
63 // Find all the potential callees in this function. First walk the
64 // instructions and add every operand which is a constant to the worklist.
65 for (BasicBlock &BB : F)
66 for (Instruction &I : BB)
67 for (Value *Op : I.operand_values())
68 if (Constant *C = dyn_cast<Constant>(Op))
69 if (Visited.insert(C))
70 Worklist.push_back(C);
72 // We've collected all the constant (and thus potentially function or
73 // function containing) operands to all of the instructions in the function.
74 // Process them (recursively) collecting every function found.
75 findCallees(Worklist, Visited, Callees, CalleeIndexMap);
78 void LazyCallGraph::Node::insertEdgeInternal(Function &Callee) {
79 CalleeIndexMap.insert(std::make_pair(&Callee, Callees.size()));
80 if (Node *N = G->lookup(Callee))
83 Callees.push_back(&Callee);
86 void LazyCallGraph::Node::removeEdgeInternal(Function &Callee) {
87 auto IndexMapI = CalleeIndexMap.find(&Callee);
88 assert(IndexMapI != CalleeIndexMap.end() &&
89 "Callee not in the callee set for this caller?");
91 Callees[IndexMapI->second] = nullptr;
92 CalleeIndexMap.erase(IndexMapI);
95 LazyCallGraph::LazyCallGraph(Module &M) : NextDFSNumber(0) {
96 DEBUG(dbgs() << "Building CG for module: " << M.getModuleIdentifier()
99 if (!F.isDeclaration() && !F.hasLocalLinkage())
100 if (EntryIndexMap.insert(std::make_pair(&F, EntryNodes.size())).second) {
101 DEBUG(dbgs() << " Adding '" << F.getName()
102 << "' to entry set of the graph.\n");
103 EntryNodes.push_back(&F);
106 // Now add entry nodes for functions reachable via initializers to globals.
107 SmallVector<Constant *, 16> Worklist;
108 SmallPtrSet<Constant *, 16> Visited;
109 for (GlobalVariable &GV : M.globals())
110 if (GV.hasInitializer())
111 if (Visited.insert(GV.getInitializer()))
112 Worklist.push_back(GV.getInitializer());
114 DEBUG(dbgs() << " Adding functions referenced by global initializers to the "
116 findCallees(Worklist, Visited, EntryNodes, EntryIndexMap);
118 for (auto &Entry : EntryNodes) {
119 assert(!Entry.isNull() &&
120 "We can't have removed edges before we finish the constructor!");
121 if (Function *F = Entry.dyn_cast<Function *>())
122 SCCEntryNodes.push_back(F);
124 SCCEntryNodes.push_back(&Entry.get<Node *>()->getFunction());
128 LazyCallGraph::LazyCallGraph(LazyCallGraph &&G)
129 : BPA(std::move(G.BPA)), NodeMap(std::move(G.NodeMap)),
130 EntryNodes(std::move(G.EntryNodes)),
131 EntryIndexMap(std::move(G.EntryIndexMap)), SCCBPA(std::move(G.SCCBPA)),
132 SCCMap(std::move(G.SCCMap)), LeafSCCs(std::move(G.LeafSCCs)),
133 DFSStack(std::move(G.DFSStack)),
134 SCCEntryNodes(std::move(G.SCCEntryNodes)),
135 NextDFSNumber(G.NextDFSNumber) {
139 LazyCallGraph &LazyCallGraph::operator=(LazyCallGraph &&G) {
140 BPA = std::move(G.BPA);
141 NodeMap = std::move(G.NodeMap);
142 EntryNodes = std::move(G.EntryNodes);
143 EntryIndexMap = std::move(G.EntryIndexMap);
144 SCCBPA = std::move(G.SCCBPA);
145 SCCMap = std::move(G.SCCMap);
146 LeafSCCs = std::move(G.LeafSCCs);
147 DFSStack = std::move(G.DFSStack);
148 SCCEntryNodes = std::move(G.SCCEntryNodes);
149 NextDFSNumber = G.NextDFSNumber;
154 void LazyCallGraph::SCC::insert(Node &N) {
155 N.DFSNumber = N.LowLink = -1;
157 G->SCCMap[&N] = this;
160 void LazyCallGraph::SCC::removeInterSCCEdge(Node &CallerN, Node &CalleeN) {
161 // First remove it from the node.
162 CallerN.removeEdgeInternal(CalleeN.getFunction());
164 assert(G->SCCMap.lookup(&CallerN) == this &&
165 "The caller must be a member of this SCC.");
167 SCC &CalleeC = *G->SCCMap.lookup(&CalleeN);
168 assert(&CalleeC != this &&
169 "This API only supports the rmoval of inter-SCC edges.");
171 assert(std::find(G->LeafSCCs.begin(), G->LeafSCCs.end(), this) ==
173 "Cannot have a leaf SCC caller with a different SCC callee.");
175 bool HasOtherCallToCalleeC = false;
176 bool HasOtherCallOutsideSCC = false;
177 for (Node *N : *this) {
178 for (Node &OtherCalleeN : *N) {
179 SCC &OtherCalleeC = *G->SCCMap.lookup(&OtherCalleeN);
180 if (&OtherCalleeC == &CalleeC) {
181 HasOtherCallToCalleeC = true;
184 if (&OtherCalleeC != this)
185 HasOtherCallOutsideSCC = true;
187 if (HasOtherCallToCalleeC)
190 // Because the SCCs form a DAG, deleting such an edge cannot change the set
191 // of SCCs in the graph. However, it may cut an edge of the SCC DAG, making
192 // the caller no longer a parent of the callee. Walk the other call edges
193 // in the caller to tell.
194 if (!HasOtherCallToCalleeC) {
195 bool Removed = CalleeC.ParentSCCs.erase(this);
198 "Did not find the caller SCC in the callee SCC's parent list!");
200 // It may orphan an SCC if it is the last edge reaching it, but that does
201 // not violate any invariants of the graph.
202 if (CalleeC.ParentSCCs.empty())
203 DEBUG(dbgs() << "LCG: Update removing " << CallerN.getFunction().getName()
204 << " -> " << CalleeN.getFunction().getName()
205 << " edge orphaned the callee's SCC!\n");
208 // It may make the Caller SCC a leaf SCC.
209 if (!HasOtherCallOutsideSCC)
210 G->LeafSCCs.push_back(this);
213 void LazyCallGraph::SCC::internalDFS(
214 SmallVectorImpl<std::pair<Node *, Node::iterator>> &DFSStack,
215 SmallVectorImpl<Node *> &PendingSCCStack, Node *N,
216 SmallVectorImpl<SCC *> &ResultSCCs) {
217 Node::iterator I = N->begin();
218 N->LowLink = N->DFSNumber = 1;
219 int NextDFSNumber = 2;
221 assert(N->DFSNumber != 0 && "We should always assign a DFS number "
222 "before processing a node.");
224 // We simulate recursion by popping out of the nested loop and continuing.
225 Node::iterator E = N->end();
228 if (SCC *ChildSCC = G->SCCMap.lookup(&ChildN)) {
229 // Check if we have reached a node in the new (known connected) set of
230 // this SCC. If so, the entire stack is necessarily in that set and we
232 if (ChildSCC == this) {
234 while (!PendingSCCStack.empty())
235 insert(*PendingSCCStack.pop_back_val());
236 while (!DFSStack.empty())
237 insert(*DFSStack.pop_back_val().first);
241 // If this child isn't currently in this SCC, no need to process it.
242 // However, we do need to remove this SCC from its SCC's parent set.
243 ChildSCC->ParentSCCs.erase(this);
248 if (ChildN.DFSNumber == 0) {
249 // Mark that we should start at this child when next this node is the
250 // top of the stack. We don't start at the next child to ensure this
251 // child's lowlink is reflected.
252 DFSStack.push_back(std::make_pair(N, I));
254 // Continue, resetting to the child node.
255 ChildN.LowLink = ChildN.DFSNumber = NextDFSNumber++;
262 // Track the lowest link of the childen, if any are still in the stack.
263 // Any child not on the stack will have a LowLink of -1.
264 assert(ChildN.LowLink != 0 &&
265 "Low-link must not be zero with a non-zero DFS number.");
266 if (ChildN.LowLink >= 0 && ChildN.LowLink < N->LowLink)
267 N->LowLink = ChildN.LowLink;
271 if (N->LowLink == N->DFSNumber) {
272 ResultSCCs.push_back(G->formSCC(N, PendingSCCStack));
273 if (DFSStack.empty())
276 // At this point we know that N cannot ever be an SCC root. Its low-link
277 // is not its dfs-number, and we've processed all of its children. It is
278 // just sitting here waiting until some node further down the stack gets
279 // low-link == dfs-number and pops it off as well. Move it to the pending
280 // stack which is pulled into the next SCC to be formed.
281 PendingSCCStack.push_back(N);
283 assert(!DFSStack.empty() && "We shouldn't have an empty stack!");
286 N = DFSStack.back().first;
287 I = DFSStack.back().second;
292 SmallVector<LazyCallGraph::SCC *, 1>
293 LazyCallGraph::SCC::removeIntraSCCEdge(Node &CallerN,
295 // First remove it from the node.
296 CallerN.removeEdgeInternal(CalleeN.getFunction());
298 // We return a list of the resulting *new* SCCs in postorder.
299 SmallVector<SCC *, 1> ResultSCCs;
301 // Direct recursion doesn't impact the SCC graph at all.
302 if (&CallerN == &CalleeN)
305 // The worklist is every node in the original SCC.
306 SmallVector<Node *, 1> Worklist;
307 Worklist.swap(Nodes);
308 for (Node *N : Worklist) {
309 // The nodes formerly in this SCC are no longer in any SCC.
314 assert(Worklist.size() > 1 && "We have to have at least two nodes to have an "
315 "edge between them that is within the SCC.");
317 // The callee can already reach every node in this SCC (by definition). It is
318 // the only node we know will stay inside this SCC. Everything which
319 // transitively reaches Callee will also remain in the SCC. To model this we
320 // incrementally add any chain of nodes which reaches something in the new
321 // node set to the new node set. This short circuits one side of the Tarjan's
325 // We're going to do a full mini-Tarjan's walk using a local stack here.
326 SmallVector<std::pair<Node *, Node::iterator>, 4> DFSStack;
327 SmallVector<Node *, 4> PendingSCCStack;
329 Node *N = Worklist.pop_back_val();
330 if (N->DFSNumber == 0)
331 internalDFS(DFSStack, PendingSCCStack, N, ResultSCCs);
333 assert(DFSStack.empty() && "Didn't flush the entire DFS stack!");
334 assert(PendingSCCStack.empty() && "Didn't flush all pending SCC nodes!");
335 } while (!Worklist.empty());
337 // Now we need to reconnect the current SCC to the graph.
338 bool IsLeafSCC = true;
339 for (Node *N : Nodes) {
340 for (Node &ChildN : *N) {
341 SCC &ChildSCC = *G->SCCMap.lookup(&ChildN);
342 if (&ChildSCC == this)
344 ChildSCC.ParentSCCs.insert(this);
349 if (!ResultSCCs.empty())
350 assert(!IsLeafSCC && "This SCC cannot be a leaf as we have split out new "
351 "SCCs by removing this edge.");
352 if (!std::any_of(G->LeafSCCs.begin(), G->LeafSCCs.end(),
353 [&](SCC *C) { return C == this; }))
354 assert(!IsLeafSCC && "This SCC cannot be a leaf as it already had child "
355 "SCCs before we removed this edge.");
357 // If this SCC stopped being a leaf through this edge removal, remove it from
358 // the leaf SCC list.
359 if (!IsLeafSCC && !ResultSCCs.empty())
360 G->LeafSCCs.erase(std::remove(G->LeafSCCs.begin(), G->LeafSCCs.end(), this),
363 // Return the new list of SCCs.
367 void LazyCallGraph::insertEdge(Node &CallerN, Function &Callee) {
368 assert(SCCMap.empty() && DFSStack.empty() &&
369 "This method cannot be called after SCCs have been formed!");
371 return CallerN.insertEdgeInternal(Callee);
374 void LazyCallGraph::removeEdge(Node &CallerN, Function &Callee) {
375 assert(SCCMap.empty() && DFSStack.empty() &&
376 "This method cannot be called after SCCs have been formed!");
378 return CallerN.removeEdgeInternal(Callee);
381 LazyCallGraph::Node &LazyCallGraph::insertInto(Function &F, Node *&MappedN) {
382 return *new (MappedN = BPA.Allocate()) Node(*this, F);
385 void LazyCallGraph::updateGraphPtrs() {
386 // Process all nodes updating the graph pointers.
388 SmallVector<Node *, 16> Worklist;
389 for (auto &Entry : EntryNodes)
390 if (Node *EntryN = Entry.dyn_cast<Node *>())
391 Worklist.push_back(EntryN);
393 while (!Worklist.empty()) {
394 Node *N = Worklist.pop_back_val();
396 for (auto &Callee : N->Callees)
397 if (!Callee.isNull())
398 if (Node *CalleeN = Callee.dyn_cast<Node *>())
399 Worklist.push_back(CalleeN);
403 // Process all SCCs updating the graph pointers.
405 SmallVector<SCC *, 16> Worklist(LeafSCCs.begin(), LeafSCCs.end());
407 while (!Worklist.empty()) {
408 SCC *C = Worklist.pop_back_val();
410 Worklist.insert(Worklist.end(), C->ParentSCCs.begin(),
411 C->ParentSCCs.end());
416 LazyCallGraph::SCC *LazyCallGraph::formSCC(Node *RootN,
417 SmallVectorImpl<Node *> &NodeStack) {
418 // The tail of the stack is the new SCC. Allocate the SCC and pop the stack
420 SCC *NewSCC = new (SCCBPA.Allocate()) SCC(*this);
422 while (!NodeStack.empty() && NodeStack.back()->DFSNumber > RootN->DFSNumber) {
423 assert(NodeStack.back()->LowLink >= RootN->LowLink &&
424 "We cannot have a low link in an SCC lower than its root on the "
426 NewSCC->insert(*NodeStack.pop_back_val());
428 NewSCC->insert(*RootN);
430 // A final pass over all edges in the SCC (this remains linear as we only
431 // do this once when we build the SCC) to connect it to the parent sets of
433 bool IsLeafSCC = true;
434 for (Node *SCCN : NewSCC->Nodes)
435 for (Node &SCCChildN : *SCCN) {
436 if (SCCMap.lookup(&SCCChildN) == NewSCC)
438 SCC &ChildSCC = *SCCMap.lookup(&SCCChildN);
439 ChildSCC.ParentSCCs.insert(NewSCC);
443 // For the SCCs where we fine no child SCCs, add them to the leaf list.
445 LeafSCCs.push_back(NewSCC);
450 LazyCallGraph::SCC *LazyCallGraph::getNextSCCInPostOrder() {
453 if (!DFSStack.empty()) {
454 N = DFSStack.back().first;
455 I = DFSStack.back().second;
458 // If we've handled all candidate entry nodes to the SCC forest, we're done.
460 if (SCCEntryNodes.empty())
463 N = &get(*SCCEntryNodes.pop_back_val());
464 } while (N->DFSNumber != 0);
466 N->LowLink = N->DFSNumber = 1;
471 assert(N->DFSNumber != 0 && "We should always assign a DFS number "
472 "before placing a node onto the stack.");
474 Node::iterator E = N->end();
477 if (ChildN.DFSNumber == 0) {
478 // Mark that we should start at this child when next this node is the
479 // top of the stack. We don't start at the next child to ensure this
480 // child's lowlink is reflected.
481 DFSStack.push_back(std::make_pair(N, N->begin()));
483 // Recurse onto this node via a tail call.
484 assert(!SCCMap.count(&ChildN) &&
485 "Found a node with 0 DFS number but already in an SCC!");
486 ChildN.LowLink = ChildN.DFSNumber = NextDFSNumber++;
493 // Track the lowest link of the childen, if any are still in the stack.
494 assert(ChildN.LowLink != 0 &&
495 "Low-link must not be zero with a non-zero DFS number.");
496 if (ChildN.LowLink >= 0 && ChildN.LowLink < N->LowLink)
497 N->LowLink = ChildN.LowLink;
501 if (N->LowLink == N->DFSNumber)
502 // Form the new SCC out of the top of the DFS stack.
503 return formSCC(N, PendingSCCStack);
505 // At this point we know that N cannot ever be an SCC root. Its low-link
506 // is not its dfs-number, and we've processed all of its children. It is
507 // just sitting here waiting until some node further down the stack gets
508 // low-link == dfs-number and pops it off as well. Move it to the pending
509 // stack which is pulled into the next SCC to be formed.
510 PendingSCCStack.push_back(N);
512 assert(!DFSStack.empty() && "We never found a viable root!");
513 N = DFSStack.back().first;
514 I = DFSStack.back().second;
519 char LazyCallGraphAnalysis::PassID;
521 LazyCallGraphPrinterPass::LazyCallGraphPrinterPass(raw_ostream &OS) : OS(OS) {}
523 static void printNodes(raw_ostream &OS, LazyCallGraph::Node &N,
524 SmallPtrSetImpl<LazyCallGraph::Node *> &Printed) {
525 // Recurse depth first through the nodes.
526 for (LazyCallGraph::Node &ChildN : N)
527 if (Printed.insert(&ChildN))
528 printNodes(OS, ChildN, Printed);
530 OS << " Call edges in function: " << N.getFunction().getName() << "\n";
531 for (LazyCallGraph::iterator I = N.begin(), E = N.end(); I != E; ++I)
532 OS << " -> " << I->getFunction().getName() << "\n";
537 static void printSCC(raw_ostream &OS, LazyCallGraph::SCC &SCC) {
538 ptrdiff_t SCCSize = std::distance(SCC.begin(), SCC.end());
539 OS << " SCC with " << SCCSize << " functions:\n";
541 for (LazyCallGraph::Node *N : SCC)
542 OS << " " << N->getFunction().getName() << "\n";
547 PreservedAnalyses LazyCallGraphPrinterPass::run(Module *M,
548 ModuleAnalysisManager *AM) {
549 LazyCallGraph &G = AM->getResult<LazyCallGraphAnalysis>(M);
551 OS << "Printing the call graph for module: " << M->getModuleIdentifier()
554 SmallPtrSet<LazyCallGraph::Node *, 16> Printed;
555 for (LazyCallGraph::Node &N : G)
556 if (Printed.insert(&N))
557 printNodes(OS, N, Printed);
559 for (LazyCallGraph::SCC &SCC : G.postorder_sccs())
562 return PreservedAnalyses::all();