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 SmallPtrSetImpl<Function *> &CalleeSet) {
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() && CalleeSet.insert(F)) {
42 DEBUG(dbgs() << " Added callable function: " << F->getName()
49 for (Value *Op : C->operand_values())
50 if (Visited.insert(cast<Constant>(Op)))
51 Worklist.push_back(cast<Constant>(Op));
55 LazyCallGraph::Node::Node(LazyCallGraph &G, Function &F)
56 : G(&G), F(F), DFSNumber(0), LowLink(0) {
57 DEBUG(dbgs() << " Adding functions called by '" << F.getName()
58 << "' to the graph.\n");
60 SmallVector<Constant *, 16> Worklist;
61 SmallPtrSet<Constant *, 16> Visited;
62 // Find all the potential callees in this function. First walk the
63 // instructions and add every operand which is a constant to the worklist.
64 for (BasicBlock &BB : F)
65 for (Instruction &I : BB)
66 for (Value *Op : I.operand_values())
67 if (Constant *C = dyn_cast<Constant>(Op))
68 if (Visited.insert(C))
69 Worklist.push_back(C);
71 // We've collected all the constant (and thus potentially function or
72 // function containing) operands to all of the instructions in the function.
73 // Process them (recursively) collecting every function found.
74 findCallees(Worklist, Visited, Callees, CalleeSet);
77 LazyCallGraph::LazyCallGraph(Module &M) : NextDFSNumber(0) {
78 DEBUG(dbgs() << "Building CG for module: " << M.getModuleIdentifier()
81 if (!F.isDeclaration() && !F.hasLocalLinkage())
82 if (EntryNodeSet.insert(&F)) {
83 DEBUG(dbgs() << " Adding '" << F.getName()
84 << "' to entry set of the graph.\n");
85 EntryNodes.push_back(&F);
88 // Now add entry nodes for functions reachable via initializers to globals.
89 SmallVector<Constant *, 16> Worklist;
90 SmallPtrSet<Constant *, 16> Visited;
91 for (GlobalVariable &GV : M.globals())
92 if (GV.hasInitializer())
93 if (Visited.insert(GV.getInitializer()))
94 Worklist.push_back(GV.getInitializer());
96 DEBUG(dbgs() << " Adding functions referenced by global initializers to the "
98 findCallees(Worklist, Visited, EntryNodes, EntryNodeSet);
100 for (auto &Entry : EntryNodes)
101 if (Function *F = Entry.dyn_cast<Function *>())
102 SCCEntryNodes.insert(F);
104 SCCEntryNodes.insert(&Entry.get<Node *>()->getFunction());
107 LazyCallGraph::LazyCallGraph(LazyCallGraph &&G)
108 : BPA(std::move(G.BPA)), NodeMap(std::move(G.NodeMap)),
109 EntryNodes(std::move(G.EntryNodes)),
110 EntryNodeSet(std::move(G.EntryNodeSet)), SCCBPA(std::move(G.SCCBPA)),
111 SCCMap(std::move(G.SCCMap)), LeafSCCs(std::move(G.LeafSCCs)),
112 DFSStack(std::move(G.DFSStack)),
113 SCCEntryNodes(std::move(G.SCCEntryNodes)),
114 NextDFSNumber(G.NextDFSNumber) {
118 LazyCallGraph &LazyCallGraph::operator=(LazyCallGraph &&G) {
119 BPA = std::move(G.BPA);
120 NodeMap = std::move(G.NodeMap);
121 EntryNodes = std::move(G.EntryNodes);
122 EntryNodeSet = std::move(G.EntryNodeSet);
123 SCCBPA = std::move(G.SCCBPA);
124 SCCMap = std::move(G.SCCMap);
125 LeafSCCs = std::move(G.LeafSCCs);
126 DFSStack = std::move(G.DFSStack);
127 SCCEntryNodes = std::move(G.SCCEntryNodes);
128 NextDFSNumber = G.NextDFSNumber;
133 LazyCallGraph::Node *LazyCallGraph::insertInto(Function &F, Node *&MappedN) {
134 return new (MappedN = BPA.Allocate()) Node(*this, F);
137 void LazyCallGraph::updateGraphPtrs() {
138 // Process all nodes updating the graph pointers.
139 SmallVector<Node *, 16> Worklist;
140 for (auto &Entry : EntryNodes)
141 if (Node *EntryN = Entry.dyn_cast<Node *>())
142 Worklist.push_back(EntryN);
144 while (!Worklist.empty()) {
145 Node *N = Worklist.pop_back_val();
147 for (auto &Callee : N->Callees)
148 if (Node *CalleeN = Callee.dyn_cast<Node *>())
149 Worklist.push_back(CalleeN);
153 LazyCallGraph::SCC *LazyCallGraph::getNextSCCInPostOrder() {
154 // When the stack is empty, there are no more SCCs to walk in this graph.
155 if (DFSStack.empty()) {
156 // If we've handled all candidate entry nodes to the SCC forest, we're done.
157 if (SCCEntryNodes.empty())
160 Node *N = get(*SCCEntryNodes.pop_back_val());
161 DFSStack.push_back(std::make_pair(N, N->begin()));
164 Node *N = DFSStack.back().first;
165 if (N->DFSNumber == 0) {
166 // This node hasn't been visited before, assign it a DFS number and remove
167 // it from the entry set.
168 N->LowLink = N->DFSNumber = NextDFSNumber++;
169 SCCEntryNodes.remove(&N->getFunction());
172 for (auto I = DFSStack.back().second, E = N->end(); I != E; ++I) {
174 if (ChildN->DFSNumber == 0) {
175 // Mark that we should start at this child when next this node is the
176 // top of the stack. We don't start at the next child to ensure this
177 // child's lowlink is reflected.
178 // FIXME: I don't actually think this is required, and we could start
179 // at the next child.
180 DFSStack.back().second = I;
182 // Recurse onto this node via a tail call.
183 DFSStack.push_back(std::make_pair(ChildN, ChildN->begin()));
184 return LazyCallGraph::getNextSCCInPostOrder();
187 // Track the lowest link of the childen, if any are still in the stack.
188 if (ChildN->LowLink < N->LowLink && !SCCMap.count(&ChildN->getFunction()))
189 N->LowLink = ChildN->LowLink;
192 // The tail of the stack is the new SCC. Allocate the SCC and pop the stack
194 SCC *NewSCC = new (SCCBPA.Allocate()) SCC();
196 // Because we don't follow the strict Tarjan recursive formulation, walk
197 // from the top of the stack down, propagating the lowest link and stopping
198 // when the DFS number is the lowest link.
199 int LowestLink = N->LowLink;
201 Node *SCCN = DFSStack.pop_back_val().first;
202 SCCMap.insert(std::make_pair(&SCCN->getFunction(), NewSCC));
203 NewSCC->Nodes.push_back(SCCN);
204 LowestLink = std::min(LowestLink, SCCN->LowLink);
206 NewSCC->NodeSet.insert(&SCCN->getFunction());
208 assert(Inserted && "Cannot have duplicates in the DFSStack!");
209 } while (!DFSStack.empty() && LowestLink <= DFSStack.back().first->DFSNumber);
210 assert(LowestLink == NewSCC->Nodes.back()->DFSNumber &&
211 "Cannot stop with a DFS number greater than the lowest link!");
213 // A final pass over all edges in the SCC (this remains linear as we only
214 // do this once when we build the SCC) to connect it to the parent sets of
216 bool IsLeafSCC = true;
217 for (Node *SCCN : NewSCC->Nodes)
218 for (Node *SCCChildN : *SCCN) {
219 if (NewSCC->NodeSet.count(&SCCChildN->getFunction()))
221 SCC *ChildSCC = SCCMap.lookup(&SCCChildN->getFunction());
223 "Must have all child SCCs processed when building a new SCC!");
224 ChildSCC->ParentSCCs.insert(NewSCC);
228 // For the SCCs where we fine no child SCCs, add them to the leaf list.
230 LeafSCCs.push_back(NewSCC);
235 char LazyCallGraphAnalysis::PassID;
237 LazyCallGraphPrinterPass::LazyCallGraphPrinterPass(raw_ostream &OS) : OS(OS) {}
239 static void printNodes(raw_ostream &OS, LazyCallGraph::Node &N,
240 SmallPtrSetImpl<LazyCallGraph::Node *> &Printed) {
241 // Recurse depth first through the nodes.
242 for (LazyCallGraph::Node *ChildN : N)
243 if (Printed.insert(ChildN))
244 printNodes(OS, *ChildN, Printed);
246 OS << " Call edges in function: " << N.getFunction().getName() << "\n";
247 for (LazyCallGraph::iterator I = N.begin(), E = N.end(); I != E; ++I)
248 OS << " -> " << I->getFunction().getName() << "\n";
253 static void printSCC(raw_ostream &OS, LazyCallGraph::SCC &SCC) {
254 ptrdiff_t SCCSize = std::distance(SCC.begin(), SCC.end());
255 OS << " SCC with " << SCCSize << " functions:\n";
257 for (LazyCallGraph::Node *N : SCC)
258 OS << " " << N->getFunction().getName() << "\n";
263 PreservedAnalyses LazyCallGraphPrinterPass::run(Module *M,
264 ModuleAnalysisManager *AM) {
265 LazyCallGraph &G = AM->getResult<LazyCallGraphAnalysis>(M);
267 OS << "Printing the call graph for module: " << M->getModuleIdentifier()
270 SmallPtrSet<LazyCallGraph::Node *, 16> Printed;
271 for (LazyCallGraph::Node *N : G)
272 if (Printed.insert(N))
273 printNodes(OS, *N, Printed);
275 for (LazyCallGraph::SCC *SCC : G.postorder_sccs())
278 return PreservedAnalyses::all();