xref: /freebsd/contrib/llvm-project/llvm/lib/Analysis/LazyCallGraph.cpp (revision 68d75eff68281c1b445e3010bb975eae07aac225)
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/ScopeExit.h"
13 #include "llvm/ADT/Sequence.h"
14 #include "llvm/ADT/SmallPtrSet.h"
15 #include "llvm/ADT/SmallVector.h"
16 #include "llvm/ADT/iterator_range.h"
17 #include "llvm/Analysis/TargetLibraryInfo.h"
18 #include "llvm/Config/llvm-config.h"
19 #include "llvm/IR/CallSite.h"
20 #include "llvm/IR/Function.h"
21 #include "llvm/IR/GlobalVariable.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 <cstddef>
33 #include <iterator>
34 #include <string>
35 #include <tuple>
36 #include <utility>
37 
38 using namespace llvm;
39 
40 #define DEBUG_TYPE "lcg"
41 
42 void LazyCallGraph::EdgeSequence::insertEdgeInternal(Node &TargetN,
43                                                      Edge::Kind EK) {
44   EdgeIndexMap.insert({&TargetN, Edges.size()});
45   Edges.emplace_back(TargetN, EK);
46 }
47 
48 void LazyCallGraph::EdgeSequence::setEdgeKind(Node &TargetN, Edge::Kind EK) {
49   Edges[EdgeIndexMap.find(&TargetN)->second].setKind(EK);
50 }
51 
52 bool LazyCallGraph::EdgeSequence::removeEdgeInternal(Node &TargetN) {
53   auto IndexMapI = EdgeIndexMap.find(&TargetN);
54   if (IndexMapI == EdgeIndexMap.end())
55     return false;
56 
57   Edges[IndexMapI->second] = Edge();
58   EdgeIndexMap.erase(IndexMapI);
59   return true;
60 }
61 
62 static void addEdge(SmallVectorImpl<LazyCallGraph::Edge> &Edges,
63                     DenseMap<LazyCallGraph::Node *, int> &EdgeIndexMap,
64                     LazyCallGraph::Node &N, LazyCallGraph::Edge::Kind EK) {
65   if (!EdgeIndexMap.insert({&N, Edges.size()}).second)
66     return;
67 
68   LLVM_DEBUG(dbgs() << "    Added callable function: " << N.getName() << "\n");
69   Edges.emplace_back(LazyCallGraph::Edge(N, EK));
70 }
71 
72 LazyCallGraph::EdgeSequence &LazyCallGraph::Node::populateSlow() {
73   assert(!Edges && "Must not have already populated the edges for this node!");
74 
75   LLVM_DEBUG(dbgs() << "  Adding functions called by '" << getName()
76                     << "' to the graph.\n");
77 
78   Edges = EdgeSequence();
79 
80   SmallVector<Constant *, 16> Worklist;
81   SmallPtrSet<Function *, 4> Callees;
82   SmallPtrSet<Constant *, 16> Visited;
83 
84   // Find all the potential call graph edges in this function. We track both
85   // actual call edges and indirect references to functions. The direct calls
86   // are trivially added, but to accumulate the latter we walk the instructions
87   // and add every operand which is a constant to the worklist to process
88   // afterward.
89   //
90   // Note that we consider *any* function with a definition to be a viable
91   // edge. Even if the function's definition is subject to replacement by
92   // some other module (say, a weak definition) there may still be
93   // optimizations which essentially speculate based on the definition and
94   // a way to check that the specific definition is in fact the one being
95   // used. For example, this could be done by moving the weak definition to
96   // a strong (internal) definition and making the weak definition be an
97   // alias. Then a test of the address of the weak function against the new
98   // strong definition's address would be an effective way to determine the
99   // safety of optimizing a direct call edge.
100   for (BasicBlock &BB : *F)
101     for (Instruction &I : BB) {
102       if (auto CS = CallSite(&I))
103         if (Function *Callee = CS.getCalledFunction())
104           if (!Callee->isDeclaration())
105             if (Callees.insert(Callee).second) {
106               Visited.insert(Callee);
107               addEdge(Edges->Edges, Edges->EdgeIndexMap, G->get(*Callee),
108                       LazyCallGraph::Edge::Call);
109             }
110 
111       for (Value *Op : I.operand_values())
112         if (Constant *C = dyn_cast<Constant>(Op))
113           if (Visited.insert(C).second)
114             Worklist.push_back(C);
115     }
116 
117   // We've collected all the constant (and thus potentially function or
118   // function containing) operands to all of the instructions in the function.
119   // Process them (recursively) collecting every function found.
120   visitReferences(Worklist, Visited, [&](Function &F) {
121     addEdge(Edges->Edges, Edges->EdgeIndexMap, G->get(F),
122             LazyCallGraph::Edge::Ref);
123   });
124 
125   // Add implicit reference edges to any defined libcall functions (if we
126   // haven't found an explicit edge).
127   for (auto *F : G->LibFunctions)
128     if (!Visited.count(F))
129       addEdge(Edges->Edges, Edges->EdgeIndexMap, G->get(*F),
130               LazyCallGraph::Edge::Ref);
131 
132   return *Edges;
133 }
134 
135 void LazyCallGraph::Node::replaceFunction(Function &NewF) {
136   assert(F != &NewF && "Must not replace a function with itself!");
137   F = &NewF;
138 }
139 
140 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
141 LLVM_DUMP_METHOD void LazyCallGraph::Node::dump() const {
142   dbgs() << *this << '\n';
143 }
144 #endif
145 
146 static bool isKnownLibFunction(Function &F, TargetLibraryInfo &TLI) {
147   LibFunc LF;
148 
149   // Either this is a normal library function or a "vectorizable" function.
150   return TLI.getLibFunc(F, LF) || TLI.isFunctionVectorizable(F.getName());
151 }
152 
153 LazyCallGraph::LazyCallGraph(
154     Module &M, function_ref<TargetLibraryInfo &(Function &)> GetTLI) {
155   LLVM_DEBUG(dbgs() << "Building CG for module: " << M.getModuleIdentifier()
156                     << "\n");
157   for (Function &F : M) {
158     if (F.isDeclaration())
159       continue;
160     // If this function is a known lib function to LLVM then we want to
161     // synthesize reference edges to it to model the fact that LLVM can turn
162     // arbitrary code into a library function call.
163     if (isKnownLibFunction(F, GetTLI(F)))
164       LibFunctions.insert(&F);
165 
166     if (F.hasLocalLinkage())
167       continue;
168 
169     // External linkage defined functions have edges to them from other
170     // modules.
171     LLVM_DEBUG(dbgs() << "  Adding '" << F.getName()
172                       << "' to entry set of the graph.\n");
173     addEdge(EntryEdges.Edges, EntryEdges.EdgeIndexMap, get(F), Edge::Ref);
174   }
175 
176   // Externally visible aliases of internal functions are also viable entry
177   // edges to the module.
178   for (auto &A : M.aliases()) {
179     if (A.hasLocalLinkage())
180       continue;
181     if (Function* F = dyn_cast<Function>(A.getAliasee())) {
182       LLVM_DEBUG(dbgs() << "  Adding '" << F->getName()
183                         << "' with alias '" << A.getName()
184                         << "' to entry set of the graph.\n");
185       addEdge(EntryEdges.Edges, EntryEdges.EdgeIndexMap, get(*F), Edge::Ref);
186     }
187   }
188 
189   // Now add entry nodes for functions reachable via initializers to globals.
190   SmallVector<Constant *, 16> Worklist;
191   SmallPtrSet<Constant *, 16> Visited;
192   for (GlobalVariable &GV : M.globals())
193     if (GV.hasInitializer())
194       if (Visited.insert(GV.getInitializer()).second)
195         Worklist.push_back(GV.getInitializer());
196 
197   LLVM_DEBUG(
198       dbgs() << "  Adding functions referenced by global initializers to the "
199                 "entry set.\n");
200   visitReferences(Worklist, Visited, [&](Function &F) {
201     addEdge(EntryEdges.Edges, EntryEdges.EdgeIndexMap, get(F),
202             LazyCallGraph::Edge::Ref);
203   });
204 }
205 
206 LazyCallGraph::LazyCallGraph(LazyCallGraph &&G)
207     : BPA(std::move(G.BPA)), NodeMap(std::move(G.NodeMap)),
208       EntryEdges(std::move(G.EntryEdges)), SCCBPA(std::move(G.SCCBPA)),
209       SCCMap(std::move(G.SCCMap)),
210       LibFunctions(std::move(G.LibFunctions)) {
211   updateGraphPtrs();
212 }
213 
214 LazyCallGraph &LazyCallGraph::operator=(LazyCallGraph &&G) {
215   BPA = std::move(G.BPA);
216   NodeMap = std::move(G.NodeMap);
217   EntryEdges = std::move(G.EntryEdges);
218   SCCBPA = std::move(G.SCCBPA);
219   SCCMap = std::move(G.SCCMap);
220   LibFunctions = std::move(G.LibFunctions);
221   updateGraphPtrs();
222   return *this;
223 }
224 
225 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
226 LLVM_DUMP_METHOD void LazyCallGraph::SCC::dump() const {
227   dbgs() << *this << '\n';
228 }
229 #endif
230 
231 #ifndef NDEBUG
232 void LazyCallGraph::SCC::verify() {
233   assert(OuterRefSCC && "Can't have a null RefSCC!");
234   assert(!Nodes.empty() && "Can't have an empty SCC!");
235 
236   for (Node *N : Nodes) {
237     assert(N && "Can't have a null node!");
238     assert(OuterRefSCC->G->lookupSCC(*N) == this &&
239            "Node does not map to this SCC!");
240     assert(N->DFSNumber == -1 &&
241            "Must set DFS numbers to -1 when adding a node to an SCC!");
242     assert(N->LowLink == -1 &&
243            "Must set low link to -1 when adding a node to an SCC!");
244     for (Edge &E : **N)
245       assert(E.getNode().isPopulated() && "Can't have an unpopulated node!");
246   }
247 }
248 #endif
249 
250 bool LazyCallGraph::SCC::isParentOf(const SCC &C) const {
251   if (this == &C)
252     return false;
253 
254   for (Node &N : *this)
255     for (Edge &E : N->calls())
256       if (OuterRefSCC->G->lookupSCC(E.getNode()) == &C)
257         return true;
258 
259   // No edges found.
260   return false;
261 }
262 
263 bool LazyCallGraph::SCC::isAncestorOf(const SCC &TargetC) const {
264   if (this == &TargetC)
265     return false;
266 
267   LazyCallGraph &G = *OuterRefSCC->G;
268 
269   // Start with this SCC.
270   SmallPtrSet<const SCC *, 16> Visited = {this};
271   SmallVector<const SCC *, 16> Worklist = {this};
272 
273   // Walk down the graph until we run out of edges or find a path to TargetC.
274   do {
275     const SCC &C = *Worklist.pop_back_val();
276     for (Node &N : C)
277       for (Edge &E : N->calls()) {
278         SCC *CalleeC = G.lookupSCC(E.getNode());
279         if (!CalleeC)
280           continue;
281 
282         // If the callee's SCC is the TargetC, we're done.
283         if (CalleeC == &TargetC)
284           return true;
285 
286         // If this is the first time we've reached this SCC, put it on the
287         // worklist to recurse through.
288         if (Visited.insert(CalleeC).second)
289           Worklist.push_back(CalleeC);
290       }
291   } while (!Worklist.empty());
292 
293   // No paths found.
294   return false;
295 }
296 
297 LazyCallGraph::RefSCC::RefSCC(LazyCallGraph &G) : G(&G) {}
298 
299 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
300 LLVM_DUMP_METHOD void LazyCallGraph::RefSCC::dump() const {
301   dbgs() << *this << '\n';
302 }
303 #endif
304 
305 #ifndef NDEBUG
306 void LazyCallGraph::RefSCC::verify() {
307   assert(G && "Can't have a null graph!");
308   assert(!SCCs.empty() && "Can't have an empty SCC!");
309 
310   // Verify basic properties of the SCCs.
311   SmallPtrSet<SCC *, 4> SCCSet;
312   for (SCC *C : SCCs) {
313     assert(C && "Can't have a null SCC!");
314     C->verify();
315     assert(&C->getOuterRefSCC() == this &&
316            "SCC doesn't think it is inside this RefSCC!");
317     bool Inserted = SCCSet.insert(C).second;
318     assert(Inserted && "Found a duplicate SCC!");
319     auto IndexIt = SCCIndices.find(C);
320     assert(IndexIt != SCCIndices.end() &&
321            "Found an SCC that doesn't have an index!");
322   }
323 
324   // Check that our indices map correctly.
325   for (auto &SCCIndexPair : SCCIndices) {
326     SCC *C = SCCIndexPair.first;
327     int i = SCCIndexPair.second;
328     assert(C && "Can't have a null SCC in the indices!");
329     assert(SCCSet.count(C) && "Found an index for an SCC not in the RefSCC!");
330     assert(SCCs[i] == C && "Index doesn't point to SCC!");
331   }
332 
333   // Check that the SCCs are in fact in post-order.
334   for (int i = 0, Size = SCCs.size(); i < Size; ++i) {
335     SCC &SourceSCC = *SCCs[i];
336     for (Node &N : SourceSCC)
337       for (Edge &E : *N) {
338         if (!E.isCall())
339           continue;
340         SCC &TargetSCC = *G->lookupSCC(E.getNode());
341         if (&TargetSCC.getOuterRefSCC() == this) {
342           assert(SCCIndices.find(&TargetSCC)->second <= i &&
343                  "Edge between SCCs violates post-order relationship.");
344           continue;
345         }
346       }
347   }
348 }
349 #endif
350 
351 bool LazyCallGraph::RefSCC::isParentOf(const RefSCC &RC) const {
352   if (&RC == this)
353     return false;
354 
355   // Search all edges to see if this is a parent.
356   for (SCC &C : *this)
357     for (Node &N : C)
358       for (Edge &E : *N)
359         if (G->lookupRefSCC(E.getNode()) == &RC)
360           return true;
361 
362   return false;
363 }
364 
365 bool LazyCallGraph::RefSCC::isAncestorOf(const RefSCC &RC) const {
366   if (&RC == this)
367     return false;
368 
369   // For each descendant of this RefSCC, see if one of its children is the
370   // argument. If not, add that descendant to the worklist and continue
371   // searching.
372   SmallVector<const RefSCC *, 4> Worklist;
373   SmallPtrSet<const RefSCC *, 4> Visited;
374   Worklist.push_back(this);
375   Visited.insert(this);
376   do {
377     const RefSCC &DescendantRC = *Worklist.pop_back_val();
378     for (SCC &C : DescendantRC)
379       for (Node &N : C)
380         for (Edge &E : *N) {
381           auto *ChildRC = G->lookupRefSCC(E.getNode());
382           if (ChildRC == &RC)
383             return true;
384           if (!ChildRC || !Visited.insert(ChildRC).second)
385             continue;
386           Worklist.push_back(ChildRC);
387         }
388   } while (!Worklist.empty());
389 
390   return false;
391 }
392 
393 /// Generic helper that updates a postorder sequence of SCCs for a potentially
394 /// cycle-introducing edge insertion.
395 ///
396 /// A postorder sequence of SCCs of a directed graph has one fundamental
397 /// property: all deges in the DAG of SCCs point "up" the sequence. That is,
398 /// all edges in the SCC DAG point to prior SCCs in the sequence.
399 ///
400 /// This routine both updates a postorder sequence and uses that sequence to
401 /// compute the set of SCCs connected into a cycle. It should only be called to
402 /// insert a "downward" edge which will require changing the sequence to
403 /// restore it to a postorder.
404 ///
405 /// When inserting an edge from an earlier SCC to a later SCC in some postorder
406 /// sequence, all of the SCCs which may be impacted are in the closed range of
407 /// those two within the postorder sequence. The algorithm used here to restore
408 /// the state is as follows:
409 ///
410 /// 1) Starting from the source SCC, construct a set of SCCs which reach the
411 ///    source SCC consisting of just the source SCC. Then scan toward the
412 ///    target SCC in postorder and for each SCC, if it has an edge to an SCC
413 ///    in the set, add it to the set. Otherwise, the source SCC is not
414 ///    a successor, move it in the postorder sequence to immediately before
415 ///    the source SCC, shifting the source SCC and all SCCs in the set one
416 ///    position toward the target SCC. Stop scanning after processing the
417 ///    target SCC.
418 /// 2) If the source SCC is now past the target SCC in the postorder sequence,
419 ///    and thus the new edge will flow toward the start, we are done.
420 /// 3) Otherwise, starting from the target SCC, walk all edges which reach an
421 ///    SCC between the source and the target, and add them to the set of
422 ///    connected SCCs, then recurse through them. Once a complete set of the
423 ///    SCCs the target connects to is known, hoist the remaining SCCs between
424 ///    the source and the target to be above the target. Note that there is no
425 ///    need to process the source SCC, it is already known to connect.
426 /// 4) At this point, all of the SCCs in the closed range between the source
427 ///    SCC and the target SCC in the postorder sequence are connected,
428 ///    including the target SCC and the source SCC. Inserting the edge from
429 ///    the source SCC to the target SCC will form a cycle out of precisely
430 ///    these SCCs. Thus we can merge all of the SCCs in this closed range into
431 ///    a single SCC.
432 ///
433 /// This process has various important properties:
434 /// - Only mutates the SCCs when adding the edge actually changes the SCC
435 ///   structure.
436 /// - Never mutates SCCs which are unaffected by the change.
437 /// - Updates the postorder sequence to correctly satisfy the postorder
438 ///   constraint after the edge is inserted.
439 /// - Only reorders SCCs in the closed postorder sequence from the source to
440 ///   the target, so easy to bound how much has changed even in the ordering.
441 /// - Big-O is the number of edges in the closed postorder range of SCCs from
442 ///   source to target.
443 ///
444 /// This helper routine, in addition to updating the postorder sequence itself
445 /// will also update a map from SCCs to indices within that sequence.
446 ///
447 /// The sequence and the map must operate on pointers to the SCC type.
448 ///
449 /// Two callbacks must be provided. The first computes the subset of SCCs in
450 /// the postorder closed range from the source to the target which connect to
451 /// the source SCC via some (transitive) set of edges. The second computes the
452 /// subset of the same range which the target SCC connects to via some
453 /// (transitive) set of edges. Both callbacks should populate the set argument
454 /// provided.
455 template <typename SCCT, typename PostorderSequenceT, typename SCCIndexMapT,
456           typename ComputeSourceConnectedSetCallableT,
457           typename ComputeTargetConnectedSetCallableT>
458 static iterator_range<typename PostorderSequenceT::iterator>
459 updatePostorderSequenceForEdgeInsertion(
460     SCCT &SourceSCC, SCCT &TargetSCC, PostorderSequenceT &SCCs,
461     SCCIndexMapT &SCCIndices,
462     ComputeSourceConnectedSetCallableT ComputeSourceConnectedSet,
463     ComputeTargetConnectedSetCallableT ComputeTargetConnectedSet) {
464   int SourceIdx = SCCIndices[&SourceSCC];
465   int TargetIdx = SCCIndices[&TargetSCC];
466   assert(SourceIdx < TargetIdx && "Cannot have equal indices here!");
467 
468   SmallPtrSet<SCCT *, 4> ConnectedSet;
469 
470   // Compute the SCCs which (transitively) reach the source.
471   ComputeSourceConnectedSet(ConnectedSet);
472 
473   // Partition the SCCs in this part of the port-order sequence so only SCCs
474   // connecting to the source remain between it and the target. This is
475   // a benign partition as it preserves postorder.
476   auto SourceI = std::stable_partition(
477       SCCs.begin() + SourceIdx, SCCs.begin() + TargetIdx + 1,
478       [&ConnectedSet](SCCT *C) { return !ConnectedSet.count(C); });
479   for (int i = SourceIdx, e = TargetIdx + 1; i < e; ++i)
480     SCCIndices.find(SCCs[i])->second = i;
481 
482   // If the target doesn't connect to the source, then we've corrected the
483   // post-order and there are no cycles formed.
484   if (!ConnectedSet.count(&TargetSCC)) {
485     assert(SourceI > (SCCs.begin() + SourceIdx) &&
486            "Must have moved the source to fix the post-order.");
487     assert(*std::prev(SourceI) == &TargetSCC &&
488            "Last SCC to move should have bene the target.");
489 
490     // Return an empty range at the target SCC indicating there is nothing to
491     // merge.
492     return make_range(std::prev(SourceI), std::prev(SourceI));
493   }
494 
495   assert(SCCs[TargetIdx] == &TargetSCC &&
496          "Should not have moved target if connected!");
497   SourceIdx = SourceI - SCCs.begin();
498   assert(SCCs[SourceIdx] == &SourceSCC &&
499          "Bad updated index computation for the source SCC!");
500 
501 
502   // See whether there are any remaining intervening SCCs between the source
503   // and target. If so we need to make sure they all are reachable form the
504   // target.
505   if (SourceIdx + 1 < TargetIdx) {
506     ConnectedSet.clear();
507     ComputeTargetConnectedSet(ConnectedSet);
508 
509     // Partition SCCs so that only SCCs reached from the target remain between
510     // the source and the target. This preserves postorder.
511     auto TargetI = std::stable_partition(
512         SCCs.begin() + SourceIdx + 1, SCCs.begin() + TargetIdx + 1,
513         [&ConnectedSet](SCCT *C) { return ConnectedSet.count(C); });
514     for (int i = SourceIdx + 1, e = TargetIdx + 1; i < e; ++i)
515       SCCIndices.find(SCCs[i])->second = i;
516     TargetIdx = std::prev(TargetI) - SCCs.begin();
517     assert(SCCs[TargetIdx] == &TargetSCC &&
518            "Should always end with the target!");
519   }
520 
521   // At this point, we know that connecting source to target forms a cycle
522   // because target connects back to source, and we know that all of the SCCs
523   // between the source and target in the postorder sequence participate in that
524   // cycle.
525   return make_range(SCCs.begin() + SourceIdx, SCCs.begin() + TargetIdx);
526 }
527 
528 bool
529 LazyCallGraph::RefSCC::switchInternalEdgeToCall(
530     Node &SourceN, Node &TargetN,
531     function_ref<void(ArrayRef<SCC *> MergeSCCs)> MergeCB) {
532   assert(!(*SourceN)[TargetN].isCall() && "Must start with a ref edge!");
533   SmallVector<SCC *, 1> DeletedSCCs;
534 
535 #ifndef NDEBUG
536   // In a debug build, verify the RefSCC is valid to start with and when this
537   // routine finishes.
538   verify();
539   auto VerifyOnExit = make_scope_exit([&]() { verify(); });
540 #endif
541 
542   SCC &SourceSCC = *G->lookupSCC(SourceN);
543   SCC &TargetSCC = *G->lookupSCC(TargetN);
544 
545   // If the two nodes are already part of the same SCC, we're also done as
546   // we've just added more connectivity.
547   if (&SourceSCC == &TargetSCC) {
548     SourceN->setEdgeKind(TargetN, Edge::Call);
549     return false; // No new cycle.
550   }
551 
552   // At this point we leverage the postorder list of SCCs to detect when the
553   // insertion of an edge changes the SCC structure in any way.
554   //
555   // First and foremost, we can eliminate the need for any changes when the
556   // edge is toward the beginning of the postorder sequence because all edges
557   // flow in that direction already. Thus adding a new one cannot form a cycle.
558   int SourceIdx = SCCIndices[&SourceSCC];
559   int TargetIdx = SCCIndices[&TargetSCC];
560   if (TargetIdx < SourceIdx) {
561     SourceN->setEdgeKind(TargetN, Edge::Call);
562     return false; // No new cycle.
563   }
564 
565   // Compute the SCCs which (transitively) reach the source.
566   auto ComputeSourceConnectedSet = [&](SmallPtrSetImpl<SCC *> &ConnectedSet) {
567 #ifndef NDEBUG
568     // Check that the RefSCC is still valid before computing this as the
569     // results will be nonsensical of we've broken its invariants.
570     verify();
571 #endif
572     ConnectedSet.insert(&SourceSCC);
573     auto IsConnected = [&](SCC &C) {
574       for (Node &N : C)
575         for (Edge &E : N->calls())
576           if (ConnectedSet.count(G->lookupSCC(E.getNode())))
577             return true;
578 
579       return false;
580     };
581 
582     for (SCC *C :
583          make_range(SCCs.begin() + SourceIdx + 1, SCCs.begin() + TargetIdx + 1))
584       if (IsConnected(*C))
585         ConnectedSet.insert(C);
586   };
587 
588   // Use a normal worklist to find which SCCs the target connects to. We still
589   // bound the search based on the range in the postorder list we care about,
590   // but because this is forward connectivity we just "recurse" through the
591   // edges.
592   auto ComputeTargetConnectedSet = [&](SmallPtrSetImpl<SCC *> &ConnectedSet) {
593 #ifndef NDEBUG
594     // Check that the RefSCC is still valid before computing this as the
595     // results will be nonsensical of we've broken its invariants.
596     verify();
597 #endif
598     ConnectedSet.insert(&TargetSCC);
599     SmallVector<SCC *, 4> Worklist;
600     Worklist.push_back(&TargetSCC);
601     do {
602       SCC &C = *Worklist.pop_back_val();
603       for (Node &N : C)
604         for (Edge &E : *N) {
605           if (!E.isCall())
606             continue;
607           SCC &EdgeC = *G->lookupSCC(E.getNode());
608           if (&EdgeC.getOuterRefSCC() != this)
609             // Not in this RefSCC...
610             continue;
611           if (SCCIndices.find(&EdgeC)->second <= SourceIdx)
612             // Not in the postorder sequence between source and target.
613             continue;
614 
615           if (ConnectedSet.insert(&EdgeC).second)
616             Worklist.push_back(&EdgeC);
617         }
618     } while (!Worklist.empty());
619   };
620 
621   // Use a generic helper to update the postorder sequence of SCCs and return
622   // a range of any SCCs connected into a cycle by inserting this edge. This
623   // routine will also take care of updating the indices into the postorder
624   // sequence.
625   auto MergeRange = updatePostorderSequenceForEdgeInsertion(
626       SourceSCC, TargetSCC, SCCs, SCCIndices, ComputeSourceConnectedSet,
627       ComputeTargetConnectedSet);
628 
629   // Run the user's callback on the merged SCCs before we actually merge them.
630   if (MergeCB)
631     MergeCB(makeArrayRef(MergeRange.begin(), MergeRange.end()));
632 
633   // If the merge range is empty, then adding the edge didn't actually form any
634   // new cycles. We're done.
635   if (MergeRange.empty()) {
636     // Now that the SCC structure is finalized, flip the kind to call.
637     SourceN->setEdgeKind(TargetN, Edge::Call);
638     return false; // No new cycle.
639   }
640 
641 #ifndef NDEBUG
642   // Before merging, check that the RefSCC remains valid after all the
643   // postorder updates.
644   verify();
645 #endif
646 
647   // Otherwise we need to merge all of the SCCs in the cycle into a single
648   // result SCC.
649   //
650   // NB: We merge into the target because all of these functions were already
651   // reachable from the target, meaning any SCC-wide properties deduced about it
652   // other than the set of functions within it will not have changed.
653   for (SCC *C : MergeRange) {
654     assert(C != &TargetSCC &&
655            "We merge *into* the target and shouldn't process it here!");
656     SCCIndices.erase(C);
657     TargetSCC.Nodes.append(C->Nodes.begin(), C->Nodes.end());
658     for (Node *N : C->Nodes)
659       G->SCCMap[N] = &TargetSCC;
660     C->clear();
661     DeletedSCCs.push_back(C);
662   }
663 
664   // Erase the merged SCCs from the list and update the indices of the
665   // remaining SCCs.
666   int IndexOffset = MergeRange.end() - MergeRange.begin();
667   auto EraseEnd = SCCs.erase(MergeRange.begin(), MergeRange.end());
668   for (SCC *C : make_range(EraseEnd, SCCs.end()))
669     SCCIndices[C] -= IndexOffset;
670 
671   // Now that the SCC structure is finalized, flip the kind to call.
672   SourceN->setEdgeKind(TargetN, Edge::Call);
673 
674   // And we're done, but we did form a new cycle.
675   return true;
676 }
677 
678 void LazyCallGraph::RefSCC::switchTrivialInternalEdgeToRef(Node &SourceN,
679                                                            Node &TargetN) {
680   assert((*SourceN)[TargetN].isCall() && "Must start with a call edge!");
681 
682 #ifndef NDEBUG
683   // In a debug build, verify the RefSCC is valid to start with and when this
684   // routine finishes.
685   verify();
686   auto VerifyOnExit = make_scope_exit([&]() { verify(); });
687 #endif
688 
689   assert(G->lookupRefSCC(SourceN) == this &&
690          "Source must be in this RefSCC.");
691   assert(G->lookupRefSCC(TargetN) == this &&
692          "Target must be in this RefSCC.");
693   assert(G->lookupSCC(SourceN) != G->lookupSCC(TargetN) &&
694          "Source and Target must be in separate SCCs for this to be trivial!");
695 
696   // Set the edge kind.
697   SourceN->setEdgeKind(TargetN, Edge::Ref);
698 }
699 
700 iterator_range<LazyCallGraph::RefSCC::iterator>
701 LazyCallGraph::RefSCC::switchInternalEdgeToRef(Node &SourceN, Node &TargetN) {
702   assert((*SourceN)[TargetN].isCall() && "Must start with a call edge!");
703 
704 #ifndef NDEBUG
705   // In a debug build, verify the RefSCC is valid to start with and when this
706   // routine finishes.
707   verify();
708   auto VerifyOnExit = make_scope_exit([&]() { verify(); });
709 #endif
710 
711   assert(G->lookupRefSCC(SourceN) == this &&
712          "Source must be in this RefSCC.");
713   assert(G->lookupRefSCC(TargetN) == this &&
714          "Target must be in this RefSCC.");
715 
716   SCC &TargetSCC = *G->lookupSCC(TargetN);
717   assert(G->lookupSCC(SourceN) == &TargetSCC && "Source and Target must be in "
718                                                 "the same SCC to require the "
719                                                 "full CG update.");
720 
721   // Set the edge kind.
722   SourceN->setEdgeKind(TargetN, Edge::Ref);
723 
724   // Otherwise we are removing a call edge from a single SCC. This may break
725   // the cycle. In order to compute the new set of SCCs, we need to do a small
726   // DFS over the nodes within the SCC to form any sub-cycles that remain as
727   // distinct SCCs and compute a postorder over the resulting SCCs.
728   //
729   // However, we specially handle the target node. The target node is known to
730   // reach all other nodes in the original SCC by definition. This means that
731   // we want the old SCC to be replaced with an SCC containing that node as it
732   // will be the root of whatever SCC DAG results from the DFS. Assumptions
733   // about an SCC such as the set of functions called will continue to hold,
734   // etc.
735 
736   SCC &OldSCC = TargetSCC;
737   SmallVector<std::pair<Node *, EdgeSequence::call_iterator>, 16> DFSStack;
738   SmallVector<Node *, 16> PendingSCCStack;
739   SmallVector<SCC *, 4> NewSCCs;
740 
741   // Prepare the nodes for a fresh DFS.
742   SmallVector<Node *, 16> Worklist;
743   Worklist.swap(OldSCC.Nodes);
744   for (Node *N : Worklist) {
745     N->DFSNumber = N->LowLink = 0;
746     G->SCCMap.erase(N);
747   }
748 
749   // Force the target node to be in the old SCC. This also enables us to take
750   // a very significant short-cut in the standard Tarjan walk to re-form SCCs
751   // below: whenever we build an edge that reaches the target node, we know
752   // that the target node eventually connects back to all other nodes in our
753   // walk. As a consequence, we can detect and handle participants in that
754   // cycle without walking all the edges that form this connection, and instead
755   // by relying on the fundamental guarantee coming into this operation (all
756   // nodes are reachable from the target due to previously forming an SCC).
757   TargetN.DFSNumber = TargetN.LowLink = -1;
758   OldSCC.Nodes.push_back(&TargetN);
759   G->SCCMap[&TargetN] = &OldSCC;
760 
761   // Scan down the stack and DFS across the call edges.
762   for (Node *RootN : Worklist) {
763     assert(DFSStack.empty() &&
764            "Cannot begin a new root with a non-empty DFS stack!");
765     assert(PendingSCCStack.empty() &&
766            "Cannot begin a new root with pending nodes for an SCC!");
767 
768     // Skip any nodes we've already reached in the DFS.
769     if (RootN->DFSNumber != 0) {
770       assert(RootN->DFSNumber == -1 &&
771              "Shouldn't have any mid-DFS root nodes!");
772       continue;
773     }
774 
775     RootN->DFSNumber = RootN->LowLink = 1;
776     int NextDFSNumber = 2;
777 
778     DFSStack.push_back({RootN, (*RootN)->call_begin()});
779     do {
780       Node *N;
781       EdgeSequence::call_iterator I;
782       std::tie(N, I) = DFSStack.pop_back_val();
783       auto E = (*N)->call_end();
784       while (I != E) {
785         Node &ChildN = I->getNode();
786         if (ChildN.DFSNumber == 0) {
787           // We haven't yet visited this child, so descend, pushing the current
788           // node onto the stack.
789           DFSStack.push_back({N, I});
790 
791           assert(!G->SCCMap.count(&ChildN) &&
792                  "Found a node with 0 DFS number but already in an SCC!");
793           ChildN.DFSNumber = ChildN.LowLink = NextDFSNumber++;
794           N = &ChildN;
795           I = (*N)->call_begin();
796           E = (*N)->call_end();
797           continue;
798         }
799 
800         // Check for the child already being part of some component.
801         if (ChildN.DFSNumber == -1) {
802           if (G->lookupSCC(ChildN) == &OldSCC) {
803             // If the child is part of the old SCC, we know that it can reach
804             // every other node, so we have formed a cycle. Pull the entire DFS
805             // and pending stacks into it. See the comment above about setting
806             // up the old SCC for why we do this.
807             int OldSize = OldSCC.size();
808             OldSCC.Nodes.push_back(N);
809             OldSCC.Nodes.append(PendingSCCStack.begin(), PendingSCCStack.end());
810             PendingSCCStack.clear();
811             while (!DFSStack.empty())
812               OldSCC.Nodes.push_back(DFSStack.pop_back_val().first);
813             for (Node &N : make_range(OldSCC.begin() + OldSize, OldSCC.end())) {
814               N.DFSNumber = N.LowLink = -1;
815               G->SCCMap[&N] = &OldSCC;
816             }
817             N = nullptr;
818             break;
819           }
820 
821           // If the child has already been added to some child component, it
822           // couldn't impact the low-link of this parent because it isn't
823           // connected, and thus its low-link isn't relevant so skip it.
824           ++I;
825           continue;
826         }
827 
828         // Track the lowest linked child as the lowest link for this node.
829         assert(ChildN.LowLink > 0 && "Must have a positive low-link number!");
830         if (ChildN.LowLink < N->LowLink)
831           N->LowLink = ChildN.LowLink;
832 
833         // Move to the next edge.
834         ++I;
835       }
836       if (!N)
837         // Cleared the DFS early, start another round.
838         break;
839 
840       // We've finished processing N and its descendants, put it on our pending
841       // SCC stack to eventually get merged into an SCC of nodes.
842       PendingSCCStack.push_back(N);
843 
844       // If this node is linked to some lower entry, continue walking up the
845       // stack.
846       if (N->LowLink != N->DFSNumber)
847         continue;
848 
849       // Otherwise, we've completed an SCC. Append it to our post order list of
850       // SCCs.
851       int RootDFSNumber = N->DFSNumber;
852       // Find the range of the node stack by walking down until we pass the
853       // root DFS number.
854       auto SCCNodes = make_range(
855           PendingSCCStack.rbegin(),
856           find_if(reverse(PendingSCCStack), [RootDFSNumber](const Node *N) {
857             return N->DFSNumber < RootDFSNumber;
858           }));
859 
860       // Form a new SCC out of these nodes and then clear them off our pending
861       // stack.
862       NewSCCs.push_back(G->createSCC(*this, SCCNodes));
863       for (Node &N : *NewSCCs.back()) {
864         N.DFSNumber = N.LowLink = -1;
865         G->SCCMap[&N] = NewSCCs.back();
866       }
867       PendingSCCStack.erase(SCCNodes.end().base(), PendingSCCStack.end());
868     } while (!DFSStack.empty());
869   }
870 
871   // Insert the remaining SCCs before the old one. The old SCC can reach all
872   // other SCCs we form because it contains the target node of the removed edge
873   // of the old SCC. This means that we will have edges into all of the new
874   // SCCs, which means the old one must come last for postorder.
875   int OldIdx = SCCIndices[&OldSCC];
876   SCCs.insert(SCCs.begin() + OldIdx, NewSCCs.begin(), NewSCCs.end());
877 
878   // Update the mapping from SCC* to index to use the new SCC*s, and remove the
879   // old SCC from the mapping.
880   for (int Idx = OldIdx, Size = SCCs.size(); Idx < Size; ++Idx)
881     SCCIndices[SCCs[Idx]] = Idx;
882 
883   return make_range(SCCs.begin() + OldIdx,
884                     SCCs.begin() + OldIdx + NewSCCs.size());
885 }
886 
887 void LazyCallGraph::RefSCC::switchOutgoingEdgeToCall(Node &SourceN,
888                                                      Node &TargetN) {
889   assert(!(*SourceN)[TargetN].isCall() && "Must start with a ref edge!");
890 
891   assert(G->lookupRefSCC(SourceN) == this && "Source must be in this RefSCC.");
892   assert(G->lookupRefSCC(TargetN) != this &&
893          "Target must not be in this RefSCC.");
894 #ifdef EXPENSIVE_CHECKS
895   assert(G->lookupRefSCC(TargetN)->isDescendantOf(*this) &&
896          "Target must be a descendant of the Source.");
897 #endif
898 
899   // Edges between RefSCCs are the same regardless of call or ref, so we can
900   // just flip the edge here.
901   SourceN->setEdgeKind(TargetN, Edge::Call);
902 
903 #ifndef NDEBUG
904   // Check that the RefSCC is still valid.
905   verify();
906 #endif
907 }
908 
909 void LazyCallGraph::RefSCC::switchOutgoingEdgeToRef(Node &SourceN,
910                                                     Node &TargetN) {
911   assert((*SourceN)[TargetN].isCall() && "Must start with a call edge!");
912 
913   assert(G->lookupRefSCC(SourceN) == this && "Source must be in this RefSCC.");
914   assert(G->lookupRefSCC(TargetN) != this &&
915          "Target must not be in this RefSCC.");
916 #ifdef EXPENSIVE_CHECKS
917   assert(G->lookupRefSCC(TargetN)->isDescendantOf(*this) &&
918          "Target must be a descendant of the Source.");
919 #endif
920 
921   // Edges between RefSCCs are the same regardless of call or ref, so we can
922   // just flip the edge here.
923   SourceN->setEdgeKind(TargetN, Edge::Ref);
924 
925 #ifndef NDEBUG
926   // Check that the RefSCC is still valid.
927   verify();
928 #endif
929 }
930 
931 void LazyCallGraph::RefSCC::insertInternalRefEdge(Node &SourceN,
932                                                   Node &TargetN) {
933   assert(G->lookupRefSCC(SourceN) == this && "Source must be in this RefSCC.");
934   assert(G->lookupRefSCC(TargetN) == this && "Target must be in this RefSCC.");
935 
936   SourceN->insertEdgeInternal(TargetN, Edge::Ref);
937 
938 #ifndef NDEBUG
939   // Check that the RefSCC is still valid.
940   verify();
941 #endif
942 }
943 
944 void LazyCallGraph::RefSCC::insertOutgoingEdge(Node &SourceN, Node &TargetN,
945                                                Edge::Kind EK) {
946   // First insert it into the caller.
947   SourceN->insertEdgeInternal(TargetN, EK);
948 
949   assert(G->lookupRefSCC(SourceN) == this && "Source must be in this RefSCC.");
950 
951   assert(G->lookupRefSCC(TargetN) != this &&
952          "Target must not be in this RefSCC.");
953 #ifdef EXPENSIVE_CHECKS
954   assert(G->lookupRefSCC(TargetN)->isDescendantOf(*this) &&
955          "Target must be a descendant of the Source.");
956 #endif
957 
958 #ifndef NDEBUG
959   // Check that the RefSCC is still valid.
960   verify();
961 #endif
962 }
963 
964 SmallVector<LazyCallGraph::RefSCC *, 1>
965 LazyCallGraph::RefSCC::insertIncomingRefEdge(Node &SourceN, Node &TargetN) {
966   assert(G->lookupRefSCC(TargetN) == this && "Target must be in this RefSCC.");
967   RefSCC &SourceC = *G->lookupRefSCC(SourceN);
968   assert(&SourceC != this && "Source must not be in this RefSCC.");
969 #ifdef EXPENSIVE_CHECKS
970   assert(SourceC.isDescendantOf(*this) &&
971          "Source must be a descendant of the Target.");
972 #endif
973 
974   SmallVector<RefSCC *, 1> DeletedRefSCCs;
975 
976 #ifndef NDEBUG
977   // In a debug build, verify the RefSCC is valid to start with and when this
978   // routine finishes.
979   verify();
980   auto VerifyOnExit = make_scope_exit([&]() { verify(); });
981 #endif
982 
983   int SourceIdx = G->RefSCCIndices[&SourceC];
984   int TargetIdx = G->RefSCCIndices[this];
985   assert(SourceIdx < TargetIdx &&
986          "Postorder list doesn't see edge as incoming!");
987 
988   // Compute the RefSCCs which (transitively) reach the source. We do this by
989   // working backwards from the source using the parent set in each RefSCC,
990   // skipping any RefSCCs that don't fall in the postorder range. This has the
991   // advantage of walking the sparser parent edge (in high fan-out graphs) but
992   // more importantly this removes examining all forward edges in all RefSCCs
993   // within the postorder range which aren't in fact connected. Only connected
994   // RefSCCs (and their edges) are visited here.
995   auto ComputeSourceConnectedSet = [&](SmallPtrSetImpl<RefSCC *> &Set) {
996     Set.insert(&SourceC);
997     auto IsConnected = [&](RefSCC &RC) {
998       for (SCC &C : RC)
999         for (Node &N : C)
1000           for (Edge &E : *N)
1001             if (Set.count(G->lookupRefSCC(E.getNode())))
1002               return true;
1003 
1004       return false;
1005     };
1006 
1007     for (RefSCC *C : make_range(G->PostOrderRefSCCs.begin() + SourceIdx + 1,
1008                                 G->PostOrderRefSCCs.begin() + TargetIdx + 1))
1009       if (IsConnected(*C))
1010         Set.insert(C);
1011   };
1012 
1013   // Use a normal worklist to find which SCCs the target connects to. We still
1014   // bound the search based on the range in the postorder list we care about,
1015   // but because this is forward connectivity we just "recurse" through the
1016   // edges.
1017   auto ComputeTargetConnectedSet = [&](SmallPtrSetImpl<RefSCC *> &Set) {
1018     Set.insert(this);
1019     SmallVector<RefSCC *, 4> Worklist;
1020     Worklist.push_back(this);
1021     do {
1022       RefSCC &RC = *Worklist.pop_back_val();
1023       for (SCC &C : RC)
1024         for (Node &N : C)
1025           for (Edge &E : *N) {
1026             RefSCC &EdgeRC = *G->lookupRefSCC(E.getNode());
1027             if (G->getRefSCCIndex(EdgeRC) <= SourceIdx)
1028               // Not in the postorder sequence between source and target.
1029               continue;
1030 
1031             if (Set.insert(&EdgeRC).second)
1032               Worklist.push_back(&EdgeRC);
1033           }
1034     } while (!Worklist.empty());
1035   };
1036 
1037   // Use a generic helper to update the postorder sequence of RefSCCs and return
1038   // a range of any RefSCCs connected into a cycle by inserting this edge. This
1039   // routine will also take care of updating the indices into the postorder
1040   // sequence.
1041   iterator_range<SmallVectorImpl<RefSCC *>::iterator> MergeRange =
1042       updatePostorderSequenceForEdgeInsertion(
1043           SourceC, *this, G->PostOrderRefSCCs, G->RefSCCIndices,
1044           ComputeSourceConnectedSet, ComputeTargetConnectedSet);
1045 
1046   // Build a set so we can do fast tests for whether a RefSCC will end up as
1047   // part of the merged RefSCC.
1048   SmallPtrSet<RefSCC *, 16> MergeSet(MergeRange.begin(), MergeRange.end());
1049 
1050   // This RefSCC will always be part of that set, so just insert it here.
1051   MergeSet.insert(this);
1052 
1053   // Now that we have identified all of the SCCs which need to be merged into
1054   // a connected set with the inserted edge, merge all of them into this SCC.
1055   SmallVector<SCC *, 16> MergedSCCs;
1056   int SCCIndex = 0;
1057   for (RefSCC *RC : MergeRange) {
1058     assert(RC != this && "We're merging into the target RefSCC, so it "
1059                          "shouldn't be in the range.");
1060 
1061     // Walk the inner SCCs to update their up-pointer and walk all the edges to
1062     // update any parent sets.
1063     // FIXME: We should try to find a way to avoid this (rather expensive) edge
1064     // walk by updating the parent sets in some other manner.
1065     for (SCC &InnerC : *RC) {
1066       InnerC.OuterRefSCC = this;
1067       SCCIndices[&InnerC] = SCCIndex++;
1068       for (Node &N : InnerC)
1069         G->SCCMap[&N] = &InnerC;
1070     }
1071 
1072     // Now merge in the SCCs. We can actually move here so try to reuse storage
1073     // the first time through.
1074     if (MergedSCCs.empty())
1075       MergedSCCs = std::move(RC->SCCs);
1076     else
1077       MergedSCCs.append(RC->SCCs.begin(), RC->SCCs.end());
1078     RC->SCCs.clear();
1079     DeletedRefSCCs.push_back(RC);
1080   }
1081 
1082   // Append our original SCCs to the merged list and move it into place.
1083   for (SCC &InnerC : *this)
1084     SCCIndices[&InnerC] = SCCIndex++;
1085   MergedSCCs.append(SCCs.begin(), SCCs.end());
1086   SCCs = std::move(MergedSCCs);
1087 
1088   // Remove the merged away RefSCCs from the post order sequence.
1089   for (RefSCC *RC : MergeRange)
1090     G->RefSCCIndices.erase(RC);
1091   int IndexOffset = MergeRange.end() - MergeRange.begin();
1092   auto EraseEnd =
1093       G->PostOrderRefSCCs.erase(MergeRange.begin(), MergeRange.end());
1094   for (RefSCC *RC : make_range(EraseEnd, G->PostOrderRefSCCs.end()))
1095     G->RefSCCIndices[RC] -= IndexOffset;
1096 
1097   // At this point we have a merged RefSCC with a post-order SCCs list, just
1098   // connect the nodes to form the new edge.
1099   SourceN->insertEdgeInternal(TargetN, Edge::Ref);
1100 
1101   // We return the list of SCCs which were merged so that callers can
1102   // invalidate any data they have associated with those SCCs. Note that these
1103   // SCCs are no longer in an interesting state (they are totally empty) but
1104   // the pointers will remain stable for the life of the graph itself.
1105   return DeletedRefSCCs;
1106 }
1107 
1108 void LazyCallGraph::RefSCC::removeOutgoingEdge(Node &SourceN, Node &TargetN) {
1109   assert(G->lookupRefSCC(SourceN) == this &&
1110          "The source must be a member of this RefSCC.");
1111   assert(G->lookupRefSCC(TargetN) != this &&
1112          "The target must not be a member of this RefSCC");
1113 
1114 #ifndef NDEBUG
1115   // In a debug build, verify the RefSCC is valid to start with and when this
1116   // routine finishes.
1117   verify();
1118   auto VerifyOnExit = make_scope_exit([&]() { verify(); });
1119 #endif
1120 
1121   // First remove it from the node.
1122   bool Removed = SourceN->removeEdgeInternal(TargetN);
1123   (void)Removed;
1124   assert(Removed && "Target not in the edge set for this caller?");
1125 }
1126 
1127 SmallVector<LazyCallGraph::RefSCC *, 1>
1128 LazyCallGraph::RefSCC::removeInternalRefEdge(Node &SourceN,
1129                                              ArrayRef<Node *> TargetNs) {
1130   // We return a list of the resulting *new* RefSCCs in post-order.
1131   SmallVector<RefSCC *, 1> Result;
1132 
1133 #ifndef NDEBUG
1134   // In a debug build, verify the RefSCC is valid to start with and that either
1135   // we return an empty list of result RefSCCs and this RefSCC remains valid,
1136   // or we return new RefSCCs and this RefSCC is dead.
1137   verify();
1138   auto VerifyOnExit = make_scope_exit([&]() {
1139     // If we didn't replace our RefSCC with new ones, check that this one
1140     // remains valid.
1141     if (G)
1142       verify();
1143   });
1144 #endif
1145 
1146   // First remove the actual edges.
1147   for (Node *TargetN : TargetNs) {
1148     assert(!(*SourceN)[*TargetN].isCall() &&
1149            "Cannot remove a call edge, it must first be made a ref edge");
1150 
1151     bool Removed = SourceN->removeEdgeInternal(*TargetN);
1152     (void)Removed;
1153     assert(Removed && "Target not in the edge set for this caller?");
1154   }
1155 
1156   // Direct self references don't impact the ref graph at all.
1157   if (llvm::all_of(TargetNs,
1158                    [&](Node *TargetN) { return &SourceN == TargetN; }))
1159     return Result;
1160 
1161   // If all targets are in the same SCC as the source, because no call edges
1162   // were removed there is no RefSCC structure change.
1163   SCC &SourceC = *G->lookupSCC(SourceN);
1164   if (llvm::all_of(TargetNs, [&](Node *TargetN) {
1165         return G->lookupSCC(*TargetN) == &SourceC;
1166       }))
1167     return Result;
1168 
1169   // We build somewhat synthetic new RefSCCs by providing a postorder mapping
1170   // for each inner SCC. We store these inside the low-link field of the nodes
1171   // rather than associated with SCCs because this saves a round-trip through
1172   // the node->SCC map and in the common case, SCCs are small. We will verify
1173   // that we always give the same number to every node in the SCC such that
1174   // these are equivalent.
1175   int PostOrderNumber = 0;
1176 
1177   // Reset all the other nodes to prepare for a DFS over them, and add them to
1178   // our worklist.
1179   SmallVector<Node *, 8> Worklist;
1180   for (SCC *C : SCCs) {
1181     for (Node &N : *C)
1182       N.DFSNumber = N.LowLink = 0;
1183 
1184     Worklist.append(C->Nodes.begin(), C->Nodes.end());
1185   }
1186 
1187   // Track the number of nodes in this RefSCC so that we can quickly recognize
1188   // an important special case of the edge removal not breaking the cycle of
1189   // this RefSCC.
1190   const int NumRefSCCNodes = Worklist.size();
1191 
1192   SmallVector<std::pair<Node *, EdgeSequence::iterator>, 4> DFSStack;
1193   SmallVector<Node *, 4> PendingRefSCCStack;
1194   do {
1195     assert(DFSStack.empty() &&
1196            "Cannot begin a new root with a non-empty DFS stack!");
1197     assert(PendingRefSCCStack.empty() &&
1198            "Cannot begin a new root with pending nodes for an SCC!");
1199 
1200     Node *RootN = Worklist.pop_back_val();
1201     // Skip any nodes we've already reached in the DFS.
1202     if (RootN->DFSNumber != 0) {
1203       assert(RootN->DFSNumber == -1 &&
1204              "Shouldn't have any mid-DFS root nodes!");
1205       continue;
1206     }
1207 
1208     RootN->DFSNumber = RootN->LowLink = 1;
1209     int NextDFSNumber = 2;
1210 
1211     DFSStack.push_back({RootN, (*RootN)->begin()});
1212     do {
1213       Node *N;
1214       EdgeSequence::iterator I;
1215       std::tie(N, I) = DFSStack.pop_back_val();
1216       auto E = (*N)->end();
1217 
1218       assert(N->DFSNumber != 0 && "We should always assign a DFS number "
1219                                   "before processing a node.");
1220 
1221       while (I != E) {
1222         Node &ChildN = I->getNode();
1223         if (ChildN.DFSNumber == 0) {
1224           // Mark that we should start at this child when next this node is the
1225           // top of the stack. We don't start at the next child to ensure this
1226           // child's lowlink is reflected.
1227           DFSStack.push_back({N, I});
1228 
1229           // Continue, resetting to the child node.
1230           ChildN.LowLink = ChildN.DFSNumber = NextDFSNumber++;
1231           N = &ChildN;
1232           I = ChildN->begin();
1233           E = ChildN->end();
1234           continue;
1235         }
1236         if (ChildN.DFSNumber == -1) {
1237           // If this child isn't currently in this RefSCC, no need to process
1238           // it.
1239           ++I;
1240           continue;
1241         }
1242 
1243         // Track the lowest link of the children, if any are still in the stack.
1244         // Any child not on the stack will have a LowLink of -1.
1245         assert(ChildN.LowLink != 0 &&
1246                "Low-link must not be zero with a non-zero DFS number.");
1247         if (ChildN.LowLink >= 0 && ChildN.LowLink < N->LowLink)
1248           N->LowLink = ChildN.LowLink;
1249         ++I;
1250       }
1251 
1252       // We've finished processing N and its descendants, put it on our pending
1253       // stack to eventually get merged into a RefSCC.
1254       PendingRefSCCStack.push_back(N);
1255 
1256       // If this node is linked to some lower entry, continue walking up the
1257       // stack.
1258       if (N->LowLink != N->DFSNumber) {
1259         assert(!DFSStack.empty() &&
1260                "We never found a viable root for a RefSCC to pop off!");
1261         continue;
1262       }
1263 
1264       // Otherwise, form a new RefSCC from the top of the pending node stack.
1265       int RefSCCNumber = PostOrderNumber++;
1266       int RootDFSNumber = N->DFSNumber;
1267 
1268       // Find the range of the node stack by walking down until we pass the
1269       // root DFS number. Update the DFS numbers and low link numbers in the
1270       // process to avoid re-walking this list where possible.
1271       auto StackRI = find_if(reverse(PendingRefSCCStack), [&](Node *N) {
1272         if (N->DFSNumber < RootDFSNumber)
1273           // We've found the bottom.
1274           return true;
1275 
1276         // Update this node and keep scanning.
1277         N->DFSNumber = -1;
1278         // Save the post-order number in the lowlink field so that we can use
1279         // it to map SCCs into new RefSCCs after we finish the DFS.
1280         N->LowLink = RefSCCNumber;
1281         return false;
1282       });
1283       auto RefSCCNodes = make_range(StackRI.base(), PendingRefSCCStack.end());
1284 
1285       // If we find a cycle containing all nodes originally in this RefSCC then
1286       // the removal hasn't changed the structure at all. This is an important
1287       // special case and we can directly exit the entire routine more
1288       // efficiently as soon as we discover it.
1289       if (llvm::size(RefSCCNodes) == NumRefSCCNodes) {
1290         // Clear out the low link field as we won't need it.
1291         for (Node *N : RefSCCNodes)
1292           N->LowLink = -1;
1293         // Return the empty result immediately.
1294         return Result;
1295       }
1296 
1297       // We've already marked the nodes internally with the RefSCC number so
1298       // just clear them off the stack and continue.
1299       PendingRefSCCStack.erase(RefSCCNodes.begin(), PendingRefSCCStack.end());
1300     } while (!DFSStack.empty());
1301 
1302     assert(DFSStack.empty() && "Didn't flush the entire DFS stack!");
1303     assert(PendingRefSCCStack.empty() && "Didn't flush all pending nodes!");
1304   } while (!Worklist.empty());
1305 
1306   assert(PostOrderNumber > 1 &&
1307          "Should never finish the DFS when the existing RefSCC remains valid!");
1308 
1309   // Otherwise we create a collection of new RefSCC nodes and build
1310   // a radix-sort style map from postorder number to these new RefSCCs. We then
1311   // append SCCs to each of these RefSCCs in the order they occurred in the
1312   // original SCCs container.
1313   for (int i = 0; i < PostOrderNumber; ++i)
1314     Result.push_back(G->createRefSCC(*G));
1315 
1316   // Insert the resulting postorder sequence into the global graph postorder
1317   // sequence before the current RefSCC in that sequence, and then remove the
1318   // current one.
1319   //
1320   // FIXME: It'd be nice to change the APIs so that we returned an iterator
1321   // range over the global postorder sequence and generally use that sequence
1322   // rather than building a separate result vector here.
1323   int Idx = G->getRefSCCIndex(*this);
1324   G->PostOrderRefSCCs.erase(G->PostOrderRefSCCs.begin() + Idx);
1325   G->PostOrderRefSCCs.insert(G->PostOrderRefSCCs.begin() + Idx, Result.begin(),
1326                              Result.end());
1327   for (int i : seq<int>(Idx, G->PostOrderRefSCCs.size()))
1328     G->RefSCCIndices[G->PostOrderRefSCCs[i]] = i;
1329 
1330   for (SCC *C : SCCs) {
1331     // We store the SCC number in the node's low-link field above.
1332     int SCCNumber = C->begin()->LowLink;
1333     // Clear out all of the SCC's node's low-link fields now that we're done
1334     // using them as side-storage.
1335     for (Node &N : *C) {
1336       assert(N.LowLink == SCCNumber &&
1337              "Cannot have different numbers for nodes in the same SCC!");
1338       N.LowLink = -1;
1339     }
1340 
1341     RefSCC &RC = *Result[SCCNumber];
1342     int SCCIndex = RC.SCCs.size();
1343     RC.SCCs.push_back(C);
1344     RC.SCCIndices[C] = SCCIndex;
1345     C->OuterRefSCC = &RC;
1346   }
1347 
1348   // Now that we've moved things into the new RefSCCs, clear out our current
1349   // one.
1350   G = nullptr;
1351   SCCs.clear();
1352   SCCIndices.clear();
1353 
1354 #ifndef NDEBUG
1355   // Verify the new RefSCCs we've built.
1356   for (RefSCC *RC : Result)
1357     RC->verify();
1358 #endif
1359 
1360   // Return the new list of SCCs.
1361   return Result;
1362 }
1363 
1364 void LazyCallGraph::RefSCC::handleTrivialEdgeInsertion(Node &SourceN,
1365                                                        Node &TargetN) {
1366   // The only trivial case that requires any graph updates is when we add new
1367   // ref edge and may connect different RefSCCs along that path. This is only
1368   // because of the parents set. Every other part of the graph remains constant
1369   // after this edge insertion.
1370   assert(G->lookupRefSCC(SourceN) == this && "Source must be in this RefSCC.");
1371   RefSCC &TargetRC = *G->lookupRefSCC(TargetN);
1372   if (&TargetRC == this)
1373     return;
1374 
1375 #ifdef EXPENSIVE_CHECKS
1376   assert(TargetRC.isDescendantOf(*this) &&
1377          "Target must be a descendant of the Source.");
1378 #endif
1379 }
1380 
1381 void LazyCallGraph::RefSCC::insertTrivialCallEdge(Node &SourceN,
1382                                                   Node &TargetN) {
1383 #ifndef NDEBUG
1384   // Check that the RefSCC is still valid when we finish.
1385   auto ExitVerifier = make_scope_exit([this] { verify(); });
1386 
1387 #ifdef EXPENSIVE_CHECKS
1388   // Check that we aren't breaking some invariants of the SCC graph. Note that
1389   // this is quadratic in the number of edges in the call graph!
1390   SCC &SourceC = *G->lookupSCC(SourceN);
1391   SCC &TargetC = *G->lookupSCC(TargetN);
1392   if (&SourceC != &TargetC)
1393     assert(SourceC.isAncestorOf(TargetC) &&
1394            "Call edge is not trivial in the SCC graph!");
1395 #endif // EXPENSIVE_CHECKS
1396 #endif // NDEBUG
1397 
1398   // First insert it into the source or find the existing edge.
1399   auto InsertResult =
1400       SourceN->EdgeIndexMap.insert({&TargetN, SourceN->Edges.size()});
1401   if (!InsertResult.second) {
1402     // Already an edge, just update it.
1403     Edge &E = SourceN->Edges[InsertResult.first->second];
1404     if (E.isCall())
1405       return; // Nothing to do!
1406     E.setKind(Edge::Call);
1407   } else {
1408     // Create the new edge.
1409     SourceN->Edges.emplace_back(TargetN, Edge::Call);
1410   }
1411 
1412   // Now that we have the edge, handle the graph fallout.
1413   handleTrivialEdgeInsertion(SourceN, TargetN);
1414 }
1415 
1416 void LazyCallGraph::RefSCC::insertTrivialRefEdge(Node &SourceN, Node &TargetN) {
1417 #ifndef NDEBUG
1418   // Check that the RefSCC is still valid when we finish.
1419   auto ExitVerifier = make_scope_exit([this] { verify(); });
1420 
1421 #ifdef EXPENSIVE_CHECKS
1422   // Check that we aren't breaking some invariants of the RefSCC graph.
1423   RefSCC &SourceRC = *G->lookupRefSCC(SourceN);
1424   RefSCC &TargetRC = *G->lookupRefSCC(TargetN);
1425   if (&SourceRC != &TargetRC)
1426     assert(SourceRC.isAncestorOf(TargetRC) &&
1427            "Ref edge is not trivial in the RefSCC graph!");
1428 #endif // EXPENSIVE_CHECKS
1429 #endif // NDEBUG
1430 
1431   // First insert it into the source or find the existing edge.
1432   auto InsertResult =
1433       SourceN->EdgeIndexMap.insert({&TargetN, SourceN->Edges.size()});
1434   if (!InsertResult.second)
1435     // Already an edge, we're done.
1436     return;
1437 
1438   // Create the new edge.
1439   SourceN->Edges.emplace_back(TargetN, Edge::Ref);
1440 
1441   // Now that we have the edge, handle the graph fallout.
1442   handleTrivialEdgeInsertion(SourceN, TargetN);
1443 }
1444 
1445 void LazyCallGraph::RefSCC::replaceNodeFunction(Node &N, Function &NewF) {
1446   Function &OldF = N.getFunction();
1447 
1448 #ifndef NDEBUG
1449   // Check that the RefSCC is still valid when we finish.
1450   auto ExitVerifier = make_scope_exit([this] { verify(); });
1451 
1452   assert(G->lookupRefSCC(N) == this &&
1453          "Cannot replace the function of a node outside this RefSCC.");
1454 
1455   assert(G->NodeMap.find(&NewF) == G->NodeMap.end() &&
1456          "Must not have already walked the new function!'");
1457 
1458   // It is important that this replacement not introduce graph changes so we
1459   // insist that the caller has already removed every use of the original
1460   // function and that all uses of the new function correspond to existing
1461   // edges in the graph. The common and expected way to use this is when
1462   // replacing the function itself in the IR without changing the call graph
1463   // shape and just updating the analysis based on that.
1464   assert(&OldF != &NewF && "Cannot replace a function with itself!");
1465   assert(OldF.use_empty() &&
1466          "Must have moved all uses from the old function to the new!");
1467 #endif
1468 
1469   N.replaceFunction(NewF);
1470 
1471   // Update various call graph maps.
1472   G->NodeMap.erase(&OldF);
1473   G->NodeMap[&NewF] = &N;
1474 }
1475 
1476 void LazyCallGraph::insertEdge(Node &SourceN, Node &TargetN, Edge::Kind EK) {
1477   assert(SCCMap.empty() &&
1478          "This method cannot be called after SCCs have been formed!");
1479 
1480   return SourceN->insertEdgeInternal(TargetN, EK);
1481 }
1482 
1483 void LazyCallGraph::removeEdge(Node &SourceN, Node &TargetN) {
1484   assert(SCCMap.empty() &&
1485          "This method cannot be called after SCCs have been formed!");
1486 
1487   bool Removed = SourceN->removeEdgeInternal(TargetN);
1488   (void)Removed;
1489   assert(Removed && "Target not in the edge set for this caller?");
1490 }
1491 
1492 void LazyCallGraph::removeDeadFunction(Function &F) {
1493   // FIXME: This is unnecessarily restrictive. We should be able to remove
1494   // functions which recursively call themselves.
1495   assert(F.use_empty() &&
1496          "This routine should only be called on trivially dead functions!");
1497 
1498   // We shouldn't remove library functions as they are never really dead while
1499   // the call graph is in use -- every function definition refers to them.
1500   assert(!isLibFunction(F) &&
1501          "Must not remove lib functions from the call graph!");
1502 
1503   auto NI = NodeMap.find(&F);
1504   if (NI == NodeMap.end())
1505     // Not in the graph at all!
1506     return;
1507 
1508   Node &N = *NI->second;
1509   NodeMap.erase(NI);
1510 
1511   // Remove this from the entry edges if present.
1512   EntryEdges.removeEdgeInternal(N);
1513 
1514   if (SCCMap.empty()) {
1515     // No SCCs have been formed, so removing this is fine and there is nothing
1516     // else necessary at this point but clearing out the node.
1517     N.clear();
1518     return;
1519   }
1520 
1521   // Cannot remove a function which has yet to be visited in the DFS walk, so
1522   // if we have a node at all then we must have an SCC and RefSCC.
1523   auto CI = SCCMap.find(&N);
1524   assert(CI != SCCMap.end() &&
1525          "Tried to remove a node without an SCC after DFS walk started!");
1526   SCC &C = *CI->second;
1527   SCCMap.erase(CI);
1528   RefSCC &RC = C.getOuterRefSCC();
1529 
1530   // This node must be the only member of its SCC as it has no callers, and
1531   // that SCC must be the only member of a RefSCC as it has no references.
1532   // Validate these properties first.
1533   assert(C.size() == 1 && "Dead functions must be in a singular SCC");
1534   assert(RC.size() == 1 && "Dead functions must be in a singular RefSCC");
1535 
1536   auto RCIndexI = RefSCCIndices.find(&RC);
1537   int RCIndex = RCIndexI->second;
1538   PostOrderRefSCCs.erase(PostOrderRefSCCs.begin() + RCIndex);
1539   RefSCCIndices.erase(RCIndexI);
1540   for (int i = RCIndex, Size = PostOrderRefSCCs.size(); i < Size; ++i)
1541     RefSCCIndices[PostOrderRefSCCs[i]] = i;
1542 
1543   // Finally clear out all the data structures from the node down through the
1544   // components.
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 LazyCallGraph::Node &LazyCallGraph::insertInto(Function &F, Node *&MappedN) {
1557   return *new (MappedN = BPA.Allocate()) Node(*this, F);
1558 }
1559 
1560 void LazyCallGraph::updateGraphPtrs() {
1561   // Walk the node map to update their graph pointers. While this iterates in
1562   // an unstable order, the order has no effect so it remains correct.
1563   for (auto &FunctionNodePair : NodeMap)
1564     FunctionNodePair.second->G = this;
1565 
1566   for (auto *RC : PostOrderRefSCCs)
1567     RC->G = this;
1568 }
1569 
1570 template <typename RootsT, typename GetBeginT, typename GetEndT,
1571           typename GetNodeT, typename FormSCCCallbackT>
1572 void LazyCallGraph::buildGenericSCCs(RootsT &&Roots, GetBeginT &&GetBegin,
1573                                      GetEndT &&GetEnd, GetNodeT &&GetNode,
1574                                      FormSCCCallbackT &&FormSCC) {
1575   using EdgeItT = decltype(GetBegin(std::declval<Node &>()));
1576 
1577   SmallVector<std::pair<Node *, EdgeItT>, 16> DFSStack;
1578   SmallVector<Node *, 16> PendingSCCStack;
1579 
1580   // Scan down the stack and DFS across the call edges.
1581   for (Node *RootN : Roots) {
1582     assert(DFSStack.empty() &&
1583            "Cannot begin a new root with a non-empty DFS stack!");
1584     assert(PendingSCCStack.empty() &&
1585            "Cannot begin a new root with pending nodes for an SCC!");
1586 
1587     // Skip any nodes we've already reached in the DFS.
1588     if (RootN->DFSNumber != 0) {
1589       assert(RootN->DFSNumber == -1 &&
1590              "Shouldn't have any mid-DFS root nodes!");
1591       continue;
1592     }
1593 
1594     RootN->DFSNumber = RootN->LowLink = 1;
1595     int NextDFSNumber = 2;
1596 
1597     DFSStack.push_back({RootN, GetBegin(*RootN)});
1598     do {
1599       Node *N;
1600       EdgeItT I;
1601       std::tie(N, I) = DFSStack.pop_back_val();
1602       auto E = GetEnd(*N);
1603       while (I != E) {
1604         Node &ChildN = GetNode(I);
1605         if (ChildN.DFSNumber == 0) {
1606           // We haven't yet visited this child, so descend, pushing the current
1607           // node onto the stack.
1608           DFSStack.push_back({N, I});
1609 
1610           ChildN.DFSNumber = ChildN.LowLink = NextDFSNumber++;
1611           N = &ChildN;
1612           I = GetBegin(*N);
1613           E = GetEnd(*N);
1614           continue;
1615         }
1616 
1617         // If the child has already been added to some child component, it
1618         // couldn't impact the low-link of this parent because it isn't
1619         // connected, and thus its low-link isn't relevant so skip it.
1620         if (ChildN.DFSNumber == -1) {
1621           ++I;
1622           continue;
1623         }
1624 
1625         // Track the lowest linked child as the lowest link for this node.
1626         assert(ChildN.LowLink > 0 && "Must have a positive low-link number!");
1627         if (ChildN.LowLink < N->LowLink)
1628           N->LowLink = ChildN.LowLink;
1629 
1630         // Move to the next edge.
1631         ++I;
1632       }
1633 
1634       // We've finished processing N and its descendants, put it on our pending
1635       // SCC stack to eventually get merged into an SCC of nodes.
1636       PendingSCCStack.push_back(N);
1637 
1638       // If this node is linked to some lower entry, continue walking up the
1639       // stack.
1640       if (N->LowLink != N->DFSNumber)
1641         continue;
1642 
1643       // Otherwise, we've completed an SCC. Append it to our post order list of
1644       // SCCs.
1645       int RootDFSNumber = N->DFSNumber;
1646       // Find the range of the node stack by walking down until we pass the
1647       // root DFS number.
1648       auto SCCNodes = make_range(
1649           PendingSCCStack.rbegin(),
1650           find_if(reverse(PendingSCCStack), [RootDFSNumber](const Node *N) {
1651             return N->DFSNumber < RootDFSNumber;
1652           }));
1653       // Form a new SCC out of these nodes and then clear them off our pending
1654       // stack.
1655       FormSCC(SCCNodes);
1656       PendingSCCStack.erase(SCCNodes.end().base(), PendingSCCStack.end());
1657     } while (!DFSStack.empty());
1658   }
1659 }
1660 
1661 /// Build the internal SCCs for a RefSCC from a sequence of nodes.
1662 ///
1663 /// Appends the SCCs to the provided vector and updates the map with their
1664 /// indices. Both the vector and map must be empty when passed into this
1665 /// routine.
1666 void LazyCallGraph::buildSCCs(RefSCC &RC, node_stack_range Nodes) {
1667   assert(RC.SCCs.empty() && "Already built SCCs!");
1668   assert(RC.SCCIndices.empty() && "Already mapped SCC indices!");
1669 
1670   for (Node *N : Nodes) {
1671     assert(N->LowLink >= (*Nodes.begin())->LowLink &&
1672            "We cannot have a low link in an SCC lower than its root on the "
1673            "stack!");
1674 
1675     // This node will go into the next RefSCC, clear out its DFS and low link
1676     // as we scan.
1677     N->DFSNumber = N->LowLink = 0;
1678   }
1679 
1680   // Each RefSCC contains a DAG of the call SCCs. To build these, we do
1681   // a direct walk of the call edges using Tarjan's algorithm. We reuse the
1682   // internal storage as we won't need it for the outer graph's DFS any longer.
1683   buildGenericSCCs(
1684       Nodes, [](Node &N) { return N->call_begin(); },
1685       [](Node &N) { return N->call_end(); },
1686       [](EdgeSequence::call_iterator I) -> Node & { return I->getNode(); },
1687       [this, &RC](node_stack_range Nodes) {
1688         RC.SCCs.push_back(createSCC(RC, Nodes));
1689         for (Node &N : *RC.SCCs.back()) {
1690           N.DFSNumber = N.LowLink = -1;
1691           SCCMap[&N] = RC.SCCs.back();
1692         }
1693       });
1694 
1695   // Wire up the SCC indices.
1696   for (int i = 0, Size = RC.SCCs.size(); i < Size; ++i)
1697     RC.SCCIndices[RC.SCCs[i]] = i;
1698 }
1699 
1700 void LazyCallGraph::buildRefSCCs() {
1701   if (EntryEdges.empty() || !PostOrderRefSCCs.empty())
1702     // RefSCCs are either non-existent or already built!
1703     return;
1704 
1705   assert(RefSCCIndices.empty() && "Already mapped RefSCC indices!");
1706 
1707   SmallVector<Node *, 16> Roots;
1708   for (Edge &E : *this)
1709     Roots.push_back(&E.getNode());
1710 
1711   // The roots will be popped of a stack, so use reverse to get a less
1712   // surprising order. This doesn't change any of the semantics anywhere.
1713   std::reverse(Roots.begin(), Roots.end());
1714 
1715   buildGenericSCCs(
1716       Roots,
1717       [](Node &N) {
1718         // We need to populate each node as we begin to walk its edges.
1719         N.populate();
1720         return N->begin();
1721       },
1722       [](Node &N) { return N->end(); },
1723       [](EdgeSequence::iterator I) -> Node & { return I->getNode(); },
1724       [this](node_stack_range Nodes) {
1725         RefSCC *NewRC = createRefSCC(*this);
1726         buildSCCs(*NewRC, Nodes);
1727 
1728         // Push the new node into the postorder list and remember its position
1729         // in the index map.
1730         bool Inserted =
1731             RefSCCIndices.insert({NewRC, PostOrderRefSCCs.size()}).second;
1732         (void)Inserted;
1733         assert(Inserted && "Cannot already have this RefSCC in the index map!");
1734         PostOrderRefSCCs.push_back(NewRC);
1735 #ifndef NDEBUG
1736         NewRC->verify();
1737 #endif
1738       });
1739 }
1740 
1741 AnalysisKey LazyCallGraphAnalysis::Key;
1742 
1743 LazyCallGraphPrinterPass::LazyCallGraphPrinterPass(raw_ostream &OS) : OS(OS) {}
1744 
1745 static void printNode(raw_ostream &OS, LazyCallGraph::Node &N) {
1746   OS << "  Edges in function: " << N.getFunction().getName() << "\n";
1747   for (LazyCallGraph::Edge &E : N.populate())
1748     OS << "    " << (E.isCall() ? "call" : "ref ") << " -> "
1749        << E.getFunction().getName() << "\n";
1750 
1751   OS << "\n";
1752 }
1753 
1754 static void printSCC(raw_ostream &OS, LazyCallGraph::SCC &C) {
1755   OS << "    SCC with " << C.size() << " functions:\n";
1756 
1757   for (LazyCallGraph::Node &N : C)
1758     OS << "      " << N.getFunction().getName() << "\n";
1759 }
1760 
1761 static void printRefSCC(raw_ostream &OS, LazyCallGraph::RefSCC &C) {
1762   OS << "  RefSCC with " << C.size() << " call SCCs:\n";
1763 
1764   for (LazyCallGraph::SCC &InnerC : C)
1765     printSCC(OS, InnerC);
1766 
1767   OS << "\n";
1768 }
1769 
1770 PreservedAnalyses LazyCallGraphPrinterPass::run(Module &M,
1771                                                 ModuleAnalysisManager &AM) {
1772   LazyCallGraph &G = AM.getResult<LazyCallGraphAnalysis>(M);
1773 
1774   OS << "Printing the call graph for module: " << M.getModuleIdentifier()
1775      << "\n\n";
1776 
1777   for (Function &F : M)
1778     printNode(OS, G.get(F));
1779 
1780   G.buildRefSCCs();
1781   for (LazyCallGraph::RefSCC &C : G.postorder_ref_sccs())
1782     printRefSCC(OS, C);
1783 
1784   return PreservedAnalyses::all();
1785 }
1786 
1787 LazyCallGraphDOTPrinterPass::LazyCallGraphDOTPrinterPass(raw_ostream &OS)
1788     : OS(OS) {}
1789 
1790 static void printNodeDOT(raw_ostream &OS, LazyCallGraph::Node &N) {
1791   std::string Name = "\"" + DOT::EscapeString(N.getFunction().getName()) + "\"";
1792 
1793   for (LazyCallGraph::Edge &E : N.populate()) {
1794     OS << "  " << Name << " -> \""
1795        << DOT::EscapeString(E.getFunction().getName()) << "\"";
1796     if (!E.isCall()) // It is a ref edge.
1797       OS << " [style=dashed,label=\"ref\"]";
1798     OS << ";\n";
1799   }
1800 
1801   OS << "\n";
1802 }
1803 
1804 PreservedAnalyses LazyCallGraphDOTPrinterPass::run(Module &M,
1805                                                    ModuleAnalysisManager &AM) {
1806   LazyCallGraph &G = AM.getResult<LazyCallGraphAnalysis>(M);
1807 
1808   OS << "digraph \"" << DOT::EscapeString(M.getModuleIdentifier()) << "\" {\n";
1809 
1810   for (Function &F : M)
1811     printNodeDOT(OS, G.get(F));
1812 
1813   OS << "}\n";
1814 
1815   return PreservedAnalyses::all();
1816 }
1817