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.erase(Callees.begin() + IndexMapI->second);
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 if (Function *F = Entry.dyn_cast<Function *>())
120 SCCEntryNodes.push_back(F);
122 SCCEntryNodes.push_back(&Entry.get<Node *>()->getFunction());
125 LazyCallGraph::LazyCallGraph(LazyCallGraph &&G)
126 : BPA(std::move(G.BPA)), NodeMap(std::move(G.NodeMap)),
127 EntryNodes(std::move(G.EntryNodes)),
128 EntryIndexMap(std::move(G.EntryIndexMap)), SCCBPA(std::move(G.SCCBPA)),
129 SCCMap(std::move(G.SCCMap)), LeafSCCs(std::move(G.LeafSCCs)),
130 DFSStack(std::move(G.DFSStack)),
131 SCCEntryNodes(std::move(G.SCCEntryNodes)),
132 NextDFSNumber(G.NextDFSNumber) {
136 LazyCallGraph &LazyCallGraph::operator=(LazyCallGraph &&G) {
137 BPA = std::move(G.BPA);
138 NodeMap = std::move(G.NodeMap);
139 EntryNodes = std::move(G.EntryNodes);
140 EntryIndexMap = std::move(G.EntryIndexMap);
141 SCCBPA = std::move(G.SCCBPA);
142 SCCMap = std::move(G.SCCMap);
143 LeafSCCs = std::move(G.LeafSCCs);
144 DFSStack = std::move(G.DFSStack);
145 SCCEntryNodes = std::move(G.SCCEntryNodes);
146 NextDFSNumber = G.NextDFSNumber;
151 void LazyCallGraph::SCC::insert(Node &N) {
152 N.DFSNumber = N.LowLink = -1;
154 G->SCCMap[&N] = this;
157 void LazyCallGraph::SCC::removeInterSCCEdge(Node &CallerN, Node &CalleeN) {
158 // First remove it from the node.
159 CallerN.removeEdgeInternal(CalleeN.getFunction());
161 assert(G->SCCMap.lookup(&CallerN) == this &&
162 "The caller must be a member of this SCC.");
164 SCC &CalleeC = *G->SCCMap.lookup(&CalleeN);
165 assert(&CalleeC != this &&
166 "This API only supports the rmoval of inter-SCC edges.");
168 assert(std::find(G->LeafSCCs.begin(), G->LeafSCCs.end(), this) ==
170 "Cannot have a leaf SCC caller with a different SCC callee.");
172 bool HasOtherCallToCalleeC = false;
173 bool HasOtherCallOutsideSCC = false;
174 for (Node *N : *this) {
175 for (Node &OtherCalleeN : *N) {
176 SCC &OtherCalleeC = *G->SCCMap.lookup(&OtherCalleeN);
177 if (&OtherCalleeC == &CalleeC) {
178 HasOtherCallToCalleeC = true;
181 if (&OtherCalleeC != this)
182 HasOtherCallOutsideSCC = true;
184 if (HasOtherCallToCalleeC)
187 // Because the SCCs form a DAG, deleting such an edge cannot change the set
188 // of SCCs in the graph. However, it may cut an edge of the SCC DAG, making
189 // the caller no longer a parent of the callee. Walk the other call edges
190 // in the caller to tell.
191 if (!HasOtherCallToCalleeC) {
192 bool Removed = CalleeC.ParentSCCs.erase(this);
195 "Did not find the caller SCC in the callee SCC's parent list!");
197 // It may orphan an SCC if it is the last edge reaching it, but that does
198 // not violate any invariants of the graph.
199 if (CalleeC.ParentSCCs.empty())
200 DEBUG(dbgs() << "LCG: Update removing " << CallerN.getFunction().getName()
201 << " -> " << CalleeN.getFunction().getName()
202 << " edge orphaned the callee's SCC!\n");
205 // It may make the Caller SCC a leaf SCC.
206 if (!HasOtherCallOutsideSCC)
207 G->LeafSCCs.push_back(this);
210 void LazyCallGraph::SCC::internalDFS(
211 SmallVectorImpl<std::pair<Node *, Node::iterator>> &DFSStack,
212 SmallVectorImpl<Node *> &PendingSCCStack, Node *N,
213 SmallVectorImpl<SCC *> &ResultSCCs) {
214 Node::iterator I = N->begin();
215 N->LowLink = N->DFSNumber = 1;
216 int NextDFSNumber = 2;
218 assert(N->DFSNumber != 0 && "We should always assign a DFS number "
219 "before processing a node.");
221 // We simulate recursion by popping out of the nested loop and continuing.
222 Node::iterator E = N->end();
225 if (SCC *ChildSCC = G->SCCMap.lookup(&ChildN)) {
226 // Check if we have reached a node in the new (known connected) set of
227 // this SCC. If so, the entire stack is necessarily in that set and we
229 if (ChildSCC == this) {
231 while (!PendingSCCStack.empty())
232 insert(*PendingSCCStack.pop_back_val());
233 while (!DFSStack.empty())
234 insert(*DFSStack.pop_back_val().first);
238 // If this child isn't currently in this SCC, no need to process it.
239 // However, we do need to remove this SCC from its SCC's parent set.
240 ChildSCC->ParentSCCs.erase(this);
245 if (ChildN.DFSNumber == 0) {
246 // Mark that we should start at this child when next this node is the
247 // top of the stack. We don't start at the next child to ensure this
248 // child's lowlink is reflected.
249 DFSStack.push_back(std::make_pair(N, I));
251 // Continue, resetting to the child node.
252 ChildN.LowLink = ChildN.DFSNumber = NextDFSNumber++;
259 // Track the lowest link of the childen, if any are still in the stack.
260 // Any child not on the stack will have a LowLink of -1.
261 assert(ChildN.LowLink != 0 &&
262 "Low-link must not be zero with a non-zero DFS number.");
263 if (ChildN.LowLink >= 0 && ChildN.LowLink < N->LowLink)
264 N->LowLink = ChildN.LowLink;
268 if (N->LowLink == N->DFSNumber) {
269 ResultSCCs.push_back(G->formSCC(N, PendingSCCStack));
270 if (DFSStack.empty())
273 // At this point we know that N cannot ever be an SCC root. Its low-link
274 // is not its dfs-number, and we've processed all of its children. It is
275 // just sitting here waiting until some node further down the stack gets
276 // low-link == dfs-number and pops it off as well. Move it to the pending
277 // stack which is pulled into the next SCC to be formed.
278 PendingSCCStack.push_back(N);
280 assert(!DFSStack.empty() && "We shouldn't have an empty stack!");
283 N = DFSStack.back().first;
284 I = DFSStack.back().second;
289 SmallVector<LazyCallGraph::SCC *, 1>
290 LazyCallGraph::SCC::removeIntraSCCEdge(Node &CallerN,
292 // First remove it from the node.
293 CallerN.removeEdgeInternal(CalleeN.getFunction());
295 // We return a list of the resulting *new* SCCs in postorder.
296 SmallVector<SCC *, 1> ResultSCCs;
298 // Direct recursion doesn't impact the SCC graph at all.
299 if (&CallerN == &CalleeN)
302 // The worklist is every node in the original SCC.
303 SmallVector<Node *, 1> Worklist;
304 Worklist.swap(Nodes);
305 for (Node *N : Worklist) {
306 // The nodes formerly in this SCC are no longer in any SCC.
311 assert(Worklist.size() > 1 && "We have to have at least two nodes to have an "
312 "edge between them that is within the SCC.");
314 // The callee can already reach every node in this SCC (by definition). It is
315 // the only node we know will stay inside this SCC. Everything which
316 // transitively reaches Callee will also remain in the SCC. To model this we
317 // incrementally add any chain of nodes which reaches something in the new
318 // node set to the new node set. This short circuits one side of the Tarjan's
322 // We're going to do a full mini-Tarjan's walk using a local stack here.
323 SmallVector<std::pair<Node *, Node::iterator>, 4> DFSStack;
324 SmallVector<Node *, 4> PendingSCCStack;
326 Node *N = Worklist.pop_back_val();
327 if (N->DFSNumber == 0)
328 internalDFS(DFSStack, PendingSCCStack, N, ResultSCCs);
330 assert(DFSStack.empty() && "Didn't flush the entire DFS stack!");
331 assert(PendingSCCStack.empty() && "Didn't flush all pending SCC nodes!");
332 } while (!Worklist.empty());
334 // Now we need to reconnect the current SCC to the graph.
335 bool IsLeafSCC = true;
336 for (Node *N : Nodes) {
337 for (Node &ChildN : *N) {
338 SCC &ChildSCC = *G->SCCMap.lookup(&ChildN);
339 if (&ChildSCC == this)
341 ChildSCC.ParentSCCs.insert(this);
346 if (!ResultSCCs.empty())
347 assert(!IsLeafSCC && "This SCC cannot be a leaf as we have split out new "
348 "SCCs by removing this edge.");
349 if (!std::any_of(G->LeafSCCs.begin(), G->LeafSCCs.end(),
350 [&](SCC *C) { return C == this; }))
351 assert(!IsLeafSCC && "This SCC cannot be a leaf as it already had child "
352 "SCCs before we removed this edge.");
354 // If this SCC stopped being a leaf through this edge removal, remove it from
355 // the leaf SCC list.
356 if (!IsLeafSCC && !ResultSCCs.empty())
357 G->LeafSCCs.erase(std::remove(G->LeafSCCs.begin(), G->LeafSCCs.end(), this),
360 // Return the new list of SCCs.
364 void LazyCallGraph::insertEdge(Node &CallerN, Function &Callee) {
365 assert(SCCMap.empty() && DFSStack.empty() &&
366 "This method cannot be called after SCCs have been formed!");
368 return CallerN.insertEdgeInternal(Callee);
371 void LazyCallGraph::removeEdge(Node &CallerN, Function &Callee) {
372 assert(SCCMap.empty() && DFSStack.empty() &&
373 "This method cannot be called after SCCs have been formed!");
375 return CallerN.removeEdgeInternal(Callee);
378 LazyCallGraph::Node &LazyCallGraph::insertInto(Function &F, Node *&MappedN) {
379 return *new (MappedN = BPA.Allocate()) Node(*this, F);
382 void LazyCallGraph::updateGraphPtrs() {
383 // Process all nodes updating the graph pointers.
385 SmallVector<Node *, 16> Worklist;
386 for (auto &Entry : EntryNodes)
387 if (Node *EntryN = Entry.dyn_cast<Node *>())
388 Worklist.push_back(EntryN);
390 while (!Worklist.empty()) {
391 Node *N = Worklist.pop_back_val();
393 for (auto &Callee : N->Callees)
394 if (Node *CalleeN = Callee.dyn_cast<Node *>())
395 Worklist.push_back(CalleeN);
399 // Process all SCCs updating the graph pointers.
401 SmallVector<SCC *, 16> Worklist(LeafSCCs.begin(), LeafSCCs.end());
403 while (!Worklist.empty()) {
404 SCC *C = Worklist.pop_back_val();
406 Worklist.insert(Worklist.end(), C->ParentSCCs.begin(),
407 C->ParentSCCs.end());
412 LazyCallGraph::SCC *LazyCallGraph::formSCC(Node *RootN,
413 SmallVectorImpl<Node *> &NodeStack) {
414 // The tail of the stack is the new SCC. Allocate the SCC and pop the stack
416 SCC *NewSCC = new (SCCBPA.Allocate()) SCC(*this);
418 while (!NodeStack.empty() && NodeStack.back()->DFSNumber > RootN->DFSNumber) {
419 assert(NodeStack.back()->LowLink >= RootN->LowLink &&
420 "We cannot have a low link in an SCC lower than its root on the "
422 NewSCC->insert(*NodeStack.pop_back_val());
424 NewSCC->insert(*RootN);
426 // A final pass over all edges in the SCC (this remains linear as we only
427 // do this once when we build the SCC) to connect it to the parent sets of
429 bool IsLeafSCC = true;
430 for (Node *SCCN : NewSCC->Nodes)
431 for (Node &SCCChildN : *SCCN) {
432 if (SCCMap.lookup(&SCCChildN) == NewSCC)
434 SCC &ChildSCC = *SCCMap.lookup(&SCCChildN);
435 ChildSCC.ParentSCCs.insert(NewSCC);
439 // For the SCCs where we fine no child SCCs, add them to the leaf list.
441 LeafSCCs.push_back(NewSCC);
446 LazyCallGraph::SCC *LazyCallGraph::getNextSCCInPostOrder() {
449 if (!DFSStack.empty()) {
450 N = DFSStack.back().first;
451 I = DFSStack.back().second;
454 // If we've handled all candidate entry nodes to the SCC forest, we're done.
456 if (SCCEntryNodes.empty())
459 N = &get(*SCCEntryNodes.pop_back_val());
460 } while (N->DFSNumber != 0);
462 N->LowLink = N->DFSNumber = 1;
467 assert(N->DFSNumber != 0 && "We should always assign a DFS number "
468 "before placing a node onto the stack.");
470 Node::iterator E = N->end();
473 if (ChildN.DFSNumber == 0) {
474 // Mark that we should start at this child when next this node is the
475 // top of the stack. We don't start at the next child to ensure this
476 // child's lowlink is reflected.
477 DFSStack.push_back(std::make_pair(N, N->begin()));
479 // Recurse onto this node via a tail call.
480 assert(!SCCMap.count(&ChildN) &&
481 "Found a node with 0 DFS number but already in an SCC!");
482 ChildN.LowLink = ChildN.DFSNumber = NextDFSNumber++;
489 // Track the lowest link of the childen, if any are still in the stack.
490 assert(ChildN.LowLink != 0 &&
491 "Low-link must not be zero with a non-zero DFS number.");
492 if (ChildN.LowLink >= 0 && ChildN.LowLink < N->LowLink)
493 N->LowLink = ChildN.LowLink;
497 if (N->LowLink == N->DFSNumber)
498 // Form the new SCC out of the top of the DFS stack.
499 return formSCC(N, PendingSCCStack);
501 // At this point we know that N cannot ever be an SCC root. Its low-link
502 // is not its dfs-number, and we've processed all of its children. It is
503 // just sitting here waiting until some node further down the stack gets
504 // low-link == dfs-number and pops it off as well. Move it to the pending
505 // stack which is pulled into the next SCC to be formed.
506 PendingSCCStack.push_back(N);
508 assert(!DFSStack.empty() && "We never found a viable root!");
509 N = DFSStack.back().first;
510 I = DFSStack.back().second;
515 char LazyCallGraphAnalysis::PassID;
517 LazyCallGraphPrinterPass::LazyCallGraphPrinterPass(raw_ostream &OS) : OS(OS) {}
519 static void printNodes(raw_ostream &OS, LazyCallGraph::Node &N,
520 SmallPtrSetImpl<LazyCallGraph::Node *> &Printed) {
521 // Recurse depth first through the nodes.
522 for (LazyCallGraph::Node &ChildN : N)
523 if (Printed.insert(&ChildN))
524 printNodes(OS, ChildN, Printed);
526 OS << " Call edges in function: " << N.getFunction().getName() << "\n";
527 for (LazyCallGraph::iterator I = N.begin(), E = N.end(); I != E; ++I)
528 OS << " -> " << I->getFunction().getName() << "\n";
533 static void printSCC(raw_ostream &OS, LazyCallGraph::SCC &SCC) {
534 ptrdiff_t SCCSize = std::distance(SCC.begin(), SCC.end());
535 OS << " SCC with " << SCCSize << " functions:\n";
537 for (LazyCallGraph::Node *N : SCC)
538 OS << " " << N->getFunction().getName() << "\n";
543 PreservedAnalyses LazyCallGraphPrinterPass::run(Module *M,
544 ModuleAnalysisManager *AM) {
545 LazyCallGraph &G = AM->getResult<LazyCallGraphAnalysis>(M);
547 OS << "Printing the call graph for module: " << M->getModuleIdentifier()
550 SmallPtrSet<LazyCallGraph::Node *, 16> Printed;
551 for (LazyCallGraph::Node &N : G)
552 if (Printed.insert(&N))
553 printNodes(OS, N, Printed);
555 for (LazyCallGraph::SCC &SCC : G.postorder_sccs())
558 return PreservedAnalyses::all();