1 //===- LazyCallGraph.cpp - Analysis of a Module's call graph --------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
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"
19 using namespace llvm;
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()
44 << "\n");
45 Callees.push_back(F);
46 }
47 continue;
48 }
50 for (Value *Op : C->operand_values())
51 if (Visited.insert(cast<Constant>(Op)))
52 Worklist.push_back(cast<Constant>(Op));
53 }
54 }
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);
76 }
78 LazyCallGraph::LazyCallGraph(Module &M) : NextDFSNumber(0) {
79 DEBUG(dbgs() << "Building CG for module: " << M.getModuleIdentifier()
80 << "\n");
81 for (Function &F : M)
82 if (!F.isDeclaration() && !F.hasLocalLinkage())
83 if (EntryIndexMap.insert(std::make_pair(&F, EntryNodes.size())).second) {
84 DEBUG(dbgs() << " Adding '" << F.getName()
85 << "' to entry set of the graph.\n");
86 EntryNodes.push_back(&F);
87 }
89 // Now add entry nodes for functions reachable via initializers to globals.
90 SmallVector<Constant *, 16> Worklist;
91 SmallPtrSet<Constant *, 16> Visited;
92 for (GlobalVariable &GV : M.globals())
93 if (GV.hasInitializer())
94 if (Visited.insert(GV.getInitializer()))
95 Worklist.push_back(GV.getInitializer());
97 DEBUG(dbgs() << " Adding functions referenced by global initializers to the "
98 "entry set.\n");
99 findCallees(Worklist, Visited, EntryNodes, EntryIndexMap);
101 for (auto &Entry : EntryNodes)
102 if (Function *F = Entry.dyn_cast<Function *>())
103 SCCEntryNodes.push_back(F);
104 else
105 SCCEntryNodes.push_back(&Entry.get<Node *>()->getFunction());
106 }
108 LazyCallGraph::LazyCallGraph(LazyCallGraph &&G)
109 : BPA(std::move(G.BPA)), NodeMap(std::move(G.NodeMap)),
110 EntryNodes(std::move(G.EntryNodes)),
111 EntryIndexMap(std::move(G.EntryIndexMap)), SCCBPA(std::move(G.SCCBPA)),
112 SCCMap(std::move(G.SCCMap)), LeafSCCs(std::move(G.LeafSCCs)),
113 DFSStack(std::move(G.DFSStack)),
114 SCCEntryNodes(std::move(G.SCCEntryNodes)),
115 NextDFSNumber(G.NextDFSNumber) {
116 updateGraphPtrs();
117 }
119 LazyCallGraph &LazyCallGraph::operator=(LazyCallGraph &&G) {
120 BPA = std::move(G.BPA);
121 NodeMap = std::move(G.NodeMap);
122 EntryNodes = std::move(G.EntryNodes);
123 EntryIndexMap = std::move(G.EntryIndexMap);
124 SCCBPA = std::move(G.SCCBPA);
125 SCCMap = std::move(G.SCCMap);
126 LeafSCCs = std::move(G.LeafSCCs);
127 DFSStack = std::move(G.DFSStack);
128 SCCEntryNodes = std::move(G.SCCEntryNodes);
129 NextDFSNumber = G.NextDFSNumber;
130 updateGraphPtrs();
131 return *this;
132 }
134 void LazyCallGraph::SCC::insert(LazyCallGraph &G, Node &N) {
135 N.DFSNumber = N.LowLink = -1;
136 Nodes.push_back(&N);
137 G.SCCMap[&N] = this;
138 }
140 void LazyCallGraph::SCC::removeEdge(LazyCallGraph &G, Function &Caller,
141 Function &Callee, SCC &CalleeC) {
142 assert(std::find(G.LeafSCCs.begin(), G.LeafSCCs.end(), this) ==
143 G.LeafSCCs.end() &&
144 "Cannot have a leaf SCC caller with a different SCC callee.");
146 bool HasOtherCallToCalleeC = false;
147 bool HasOtherCallOutsideSCC = false;
148 for (Node *N : *this) {
149 for (Node &Callee : *N) {
150 SCC &OtherCalleeC = *G.SCCMap.lookup(&Callee);
151 if (&OtherCalleeC == &CalleeC) {
152 HasOtherCallToCalleeC = true;
153 break;
154 }
155 if (&OtherCalleeC != this)
156 HasOtherCallOutsideSCC = true;
157 }
158 if (HasOtherCallToCalleeC)
159 break;
160 }
161 // Because the SCCs form a DAG, deleting such an edge cannot change the set
162 // of SCCs in the graph. However, it may cut an edge of the SCC DAG, making
163 // the caller no longer a parent of the callee. Walk the other call edges
164 // in the caller to tell.
165 if (!HasOtherCallToCalleeC) {
166 bool Removed = CalleeC.ParentSCCs.erase(this);
167 (void)Removed;
168 assert(Removed &&
169 "Did not find the caller SCC in the callee SCC's parent list!");
171 // It may orphan an SCC if it is the last edge reaching it, but that does
172 // not violate any invariants of the graph.
173 if (CalleeC.ParentSCCs.empty())
174 DEBUG(dbgs() << "LCG: Update removing " << Caller.getName() << " -> "
175 << Callee.getName() << " edge orphaned the callee's SCC!\n");
176 }
178 // It may make the Caller SCC a leaf SCC.
179 if (!HasOtherCallOutsideSCC)
180 G.LeafSCCs.push_back(this);
181 }
183 void LazyCallGraph::SCC::internalDFS(
184 LazyCallGraph &G,
185 SmallVectorImpl<std::pair<Node *, Node::iterator>> &DFSStack,
186 SmallVectorImpl<Node *> &PendingSCCStack, Node *N,
187 SmallVectorImpl<SCC *> &ResultSCCs) {
188 Node::iterator I = N->begin();
189 N->LowLink = N->DFSNumber = 1;
190 int NextDFSNumber = 2;
191 for (;;) {
192 assert(N->DFSNumber != 0 && "We should always assign a DFS number "
193 "before processing a node.");
195 // We simulate recursion by popping out of the nested loop and continuing.
196 Node::iterator E = N->end();
197 while (I != E) {
198 Node &ChildN = *I;
199 if (SCC *ChildSCC = G.SCCMap.lookup(&ChildN)) {
200 // Check if we have reached a node in the new (known connected) set of
201 // this SCC. If so, the entire stack is necessarily in that set and we
202 // can re-start.
203 if (ChildSCC == this) {
204 insert(G, *N);
205 while (!PendingSCCStack.empty())
206 insert(G, *PendingSCCStack.pop_back_val());
207 while (!DFSStack.empty())
208 insert(G, *DFSStack.pop_back_val().first);
209 return;
210 }
212 // If this child isn't currently in this SCC, no need to process it.
213 // However, we do need to remove this SCC from its SCC's parent set.
214 ChildSCC->ParentSCCs.erase(this);
215 ++I;
216 continue;
217 }
219 if (ChildN.DFSNumber == 0) {
220 // Mark that we should start at this child when next this node is the
221 // top of the stack. We don't start at the next child to ensure this
222 // child's lowlink is reflected.
223 DFSStack.push_back(std::make_pair(N, I));
225 // Continue, resetting to the child node.
226 ChildN.LowLink = ChildN.DFSNumber = NextDFSNumber++;
227 N = &ChildN;
228 I = ChildN.begin();
229 E = ChildN.end();
230 continue;
231 }
233 // Track the lowest link of the childen, if any are still in the stack.
234 // Any child not on the stack will have a LowLink of -1.
235 assert(ChildN.LowLink != 0 &&
236 "Low-link must not be zero with a non-zero DFS number.");
237 if (ChildN.LowLink >= 0 && ChildN.LowLink < N->LowLink)
238 N->LowLink = ChildN.LowLink;
239 ++I;
240 }
242 if (N->LowLink == N->DFSNumber) {
243 ResultSCCs.push_back(G.formSCC(N, PendingSCCStack));
244 if (DFSStack.empty())
245 return;
246 } else {
247 // At this point we know that N cannot ever be an SCC root. Its low-link
248 // is not its dfs-number, and we've processed all of its children. It is
249 // just sitting here waiting until some node further down the stack gets
250 // low-link == dfs-number and pops it off as well. Move it to the pending
251 // stack which is pulled into the next SCC to be formed.
252 PendingSCCStack.push_back(N);
254 assert(!DFSStack.empty() && "We shouldn't have an empty stack!");
255 }
257 N = DFSStack.back().first;
258 I = DFSStack.back().second;
259 DFSStack.pop_back();
260 }
261 }
263 SmallVector<LazyCallGraph::SCC *, 1>
264 LazyCallGraph::SCC::removeInternalEdge(LazyCallGraph &G, Node &Caller,
265 Node &Callee) {
266 // We return a list of the resulting SCCs, where 'this' is always the first
267 // element.
268 SmallVector<SCC *, 1> ResultSCCs;
269 ResultSCCs.push_back(this);
271 // Direct recursion doesn't impact the SCC graph at all.
272 if (&Caller == &Callee)
273 return ResultSCCs;
275 // The worklist is every node in the original SCC.
276 SmallVector<Node *, 1> Worklist;
277 Worklist.swap(Nodes);
278 for (Node *N : Worklist) {
279 // The nodes formerly in this SCC are no longer in any SCC.
280 N->DFSNumber = 0;
281 N->LowLink = 0;
282 G.SCCMap.erase(N);
283 }
284 assert(Worklist.size() > 1 && "We have to have at least two nodes to have an "
285 "edge between them that is within the SCC.");
287 // The callee can already reach every node in this SCC (by definition). It is
288 // the only node we know will stay inside this SCC. Everything which
289 // transitively reaches Callee will also remain in the SCC. To model this we
290 // incrementally add any chain of nodes which reaches something in the new
291 // node set to the new node set. This short circuits one side of the Tarjan's
292 // walk.
293 insert(G, Callee);
295 // We're going to do a full mini-Tarjan's walk using a local stack here.
296 SmallVector<std::pair<Node *, Node::iterator>, 4> DFSStack;
297 SmallVector<Node *, 4> PendingSCCStack;
298 do {
299 Node *N = Worklist.pop_back_val();
300 if (N->DFSNumber == 0)
301 internalDFS(G, DFSStack, PendingSCCStack, N, ResultSCCs);
303 assert(DFSStack.empty() && "Didn't flush the entire DFS stack!");
304 assert(PendingSCCStack.empty() && "Didn't flush all pending SCC nodes!");
305 } while (!Worklist.empty());
307 // Now we need to reconnect the current SCC to the graph.
308 bool IsLeafSCC = true;
309 for (Node *N : Nodes) {
310 for (Node &ChildN : *N) {
311 SCC &ChildSCC = *G.SCCMap.lookup(&ChildN);
312 if (&ChildSCC == this)
313 continue;
314 ChildSCC.ParentSCCs.insert(this);
315 IsLeafSCC = false;
316 }
317 }
318 #ifndef NDEBUG
319 if (ResultSCCs.size() > 1)
320 assert(!IsLeafSCC && "This SCC cannot be a leaf as we have split out new "
321 "SCCs by removing this edge.");
322 if (!std::any_of(G.LeafSCCs.begin(), G.LeafSCCs.end(),
323 [&](SCC *C) { return C == this; }))
324 assert(!IsLeafSCC && "This SCC cannot be a leaf as it already had child "
325 "SCCs before we removed this edge.");
326 #endif
327 // If this SCC stopped being a leaf through this edge removal, remove it from
328 // the leaf SCC list.
329 if (!IsLeafSCC && ResultSCCs.size() > 1)
330 G.LeafSCCs.erase(std::remove(G.LeafSCCs.begin(), G.LeafSCCs.end(), this),
331 G.LeafSCCs.end());
333 // Return the new list of SCCs.
334 return ResultSCCs;
335 }
337 void LazyCallGraph::removeEdge(Node &CallerN, Function &Callee) {
338 auto IndexMapI = CallerN.CalleeIndexMap.find(&Callee);
339 assert(IndexMapI != CallerN.CalleeIndexMap.end() &&
340 "Callee not in the callee set for the caller?");
342 Node *CalleeN = CallerN.Callees[IndexMapI->second].dyn_cast<Node *>();
343 CallerN.Callees.erase(CallerN.Callees.begin() + IndexMapI->second);
344 CallerN.CalleeIndexMap.erase(IndexMapI);
346 SCC *CallerC = SCCMap.lookup(&CallerN);
347 if (!CallerC) {
348 // We can only remove edges when the edge isn't actively participating in
349 // a DFS walk. Either it must have been popped into an SCC, or it must not
350 // yet have been reached by the DFS walk. Assert the latter here.
351 assert(std::all_of(DFSStack.begin(), DFSStack.end(),
352 [&](const std::pair<Node *, iterator> &StackEntry) {
353 return StackEntry.first != &CallerN;
354 }) &&
355 "Found the caller on the DFSStack!");
356 return;
357 }
359 assert(CalleeN && "If the caller is in an SCC, we have to have explored all "
360 "its transitively called functions.");
362 SCC *CalleeC = SCCMap.lookup(CalleeN);
363 assert(CalleeC &&
364 "The caller has an SCC, and thus by necessity so does the callee.");
366 // The easy case is when they are different SCCs.
367 if (CallerC != CalleeC) {
368 CallerC->removeEdge(*this, CallerN.getFunction(), Callee, *CalleeC);
369 return;
370 }
372 // The hard case is when we remove an edge within a SCC. This may cause new
373 // SCCs to need to be added to the graph.
374 CallerC->removeInternalEdge(*this, CallerN, *CalleeN);
375 }
377 LazyCallGraph::Node &LazyCallGraph::insertInto(Function &F, Node *&MappedN) {
378 return *new (MappedN = BPA.Allocate()) Node(*this, F);
379 }
381 void LazyCallGraph::updateGraphPtrs() {
382 // Process all nodes updating the graph pointers.
383 SmallVector<Node *, 16> Worklist;
384 for (auto &Entry : EntryNodes)
385 if (Node *EntryN = Entry.dyn_cast<Node *>())
386 Worklist.push_back(EntryN);
388 while (!Worklist.empty()) {
389 Node *N = Worklist.pop_back_val();
390 N->G = this;
391 for (auto &Callee : N->Callees)
392 if (Node *CalleeN = Callee.dyn_cast<Node *>())
393 Worklist.push_back(CalleeN);
394 }
395 }
397 LazyCallGraph::SCC *LazyCallGraph::formSCC(Node *RootN,
398 SmallVectorImpl<Node *> &NodeStack) {
399 // The tail of the stack is the new SCC. Allocate the SCC and pop the stack
400 // into it.
401 SCC *NewSCC = new (SCCBPA.Allocate()) SCC();
403 while (!NodeStack.empty() && NodeStack.back()->DFSNumber > RootN->DFSNumber) {
404 assert(NodeStack.back()->LowLink >= RootN->LowLink &&
405 "We cannot have a low link in an SCC lower than its root on the "
406 "stack!");
407 NewSCC->insert(*this, *NodeStack.pop_back_val());
408 }
409 NewSCC->insert(*this, *RootN);
411 // A final pass over all edges in the SCC (this remains linear as we only
412 // do this once when we build the SCC) to connect it to the parent sets of
413 // its children.
414 bool IsLeafSCC = true;
415 for (Node *SCCN : NewSCC->Nodes)
416 for (Node &SCCChildN : *SCCN) {
417 if (SCCMap.lookup(&SCCChildN) == NewSCC)
418 continue;
419 SCC &ChildSCC = *SCCMap.lookup(&SCCChildN);
420 ChildSCC.ParentSCCs.insert(NewSCC);
421 IsLeafSCC = false;
422 }
424 // For the SCCs where we fine no child SCCs, add them to the leaf list.
425 if (IsLeafSCC)
426 LeafSCCs.push_back(NewSCC);
428 return NewSCC;
429 }
431 LazyCallGraph::SCC *LazyCallGraph::getNextSCCInPostOrder() {
432 Node *N;
433 Node::iterator I;
434 if (!DFSStack.empty()) {
435 N = DFSStack.back().first;
436 I = DFSStack.back().second;
437 DFSStack.pop_back();
438 } else {
439 // If we've handled all candidate entry nodes to the SCC forest, we're done.
440 do {
441 if (SCCEntryNodes.empty())
442 return nullptr;
444 N = &get(*SCCEntryNodes.pop_back_val());
445 } while (N->DFSNumber != 0);
446 I = N->begin();
447 N->LowLink = N->DFSNumber = 1;
448 NextDFSNumber = 2;
449 }
451 for (;;) {
452 assert(N->DFSNumber != 0 && "We should always assign a DFS number "
453 "before placing a node onto the stack.");
455 Node::iterator E = N->end();
456 while (I != E) {
457 Node &ChildN = *I;
458 if (ChildN.DFSNumber == 0) {
459 // Mark that we should start at this child when next this node is the
460 // top of the stack. We don't start at the next child to ensure this
461 // child's lowlink is reflected.
462 DFSStack.push_back(std::make_pair(N, N->begin()));
464 // Recurse onto this node via a tail call.
465 assert(!SCCMap.count(&ChildN) &&
466 "Found a node with 0 DFS number but already in an SCC!");
467 ChildN.LowLink = ChildN.DFSNumber = NextDFSNumber++;
468 N = &ChildN;
469 I = ChildN.begin();
470 E = ChildN.end();
471 continue;
472 }
474 // Track the lowest link of the childen, if any are still in the stack.
475 assert(ChildN.LowLink != 0 &&
476 "Low-link must not be zero with a non-zero DFS number.");
477 if (ChildN.LowLink >= 0 && ChildN.LowLink < N->LowLink)
478 N->LowLink = ChildN.LowLink;
479 ++I;
480 }
482 if (N->LowLink == N->DFSNumber)
483 // Form the new SCC out of the top of the DFS stack.
484 return formSCC(N, PendingSCCStack);
486 // At this point we know that N cannot ever be an SCC root. Its low-link
487 // is not its dfs-number, and we've processed all of its children. It is
488 // just sitting here waiting until some node further down the stack gets
489 // low-link == dfs-number and pops it off as well. Move it to the pending
490 // stack which is pulled into the next SCC to be formed.
491 PendingSCCStack.push_back(N);
493 assert(!DFSStack.empty() && "We never found a viable root!");
494 N = DFSStack.back().first;
495 I = DFSStack.back().second;
496 DFSStack.pop_back();
497 }
498 }
500 char LazyCallGraphAnalysis::PassID;
502 LazyCallGraphPrinterPass::LazyCallGraphPrinterPass(raw_ostream &OS) : OS(OS) {}
504 static void printNodes(raw_ostream &OS, LazyCallGraph::Node &N,
505 SmallPtrSetImpl<LazyCallGraph::Node *> &Printed) {
506 // Recurse depth first through the nodes.
507 for (LazyCallGraph::Node &ChildN : N)
508 if (Printed.insert(&ChildN))
509 printNodes(OS, ChildN, Printed);
511 OS << " Call edges in function: " << N.getFunction().getName() << "\n";
512 for (LazyCallGraph::iterator I = N.begin(), E = N.end(); I != E; ++I)
513 OS << " -> " << I->getFunction().getName() << "\n";
515 OS << "\n";
516 }
518 static void printSCC(raw_ostream &OS, LazyCallGraph::SCC &SCC) {
519 ptrdiff_t SCCSize = std::distance(SCC.begin(), SCC.end());
520 OS << " SCC with " << SCCSize << " functions:\n";
522 for (LazyCallGraph::Node *N : SCC)
523 OS << " " << N->getFunction().getName() << "\n";
525 OS << "\n";
526 }
528 PreservedAnalyses LazyCallGraphPrinterPass::run(Module *M,
529 ModuleAnalysisManager *AM) {
530 LazyCallGraph &G = AM->getResult<LazyCallGraphAnalysis>(M);
532 OS << "Printing the call graph for module: " << M->getModuleIdentifier()
533 << "\n\n";
535 SmallPtrSet<LazyCallGraph::Node *, 16> Printed;
536 for (LazyCallGraph::Node &N : G)
537 if (Printed.insert(&N))
538 printNodes(OS, N, Printed);
540 for (LazyCallGraph::SCC &SCC : G.postorder_sccs())
541 printSCC(OS, SCC);
543 return PreservedAnalyses::all();
545 }