xref: /freebsd/contrib/llvm-project/llvm/lib/Analysis/LazyCallGraph.cpp (revision ebacd8013fe5f7fdf9f6a5b286f6680dd2891036)
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/Sequence.h"
13 #include "llvm/ADT/SmallPtrSet.h"
14 #include "llvm/ADT/SmallVector.h"
15 #include "llvm/ADT/iterator_range.h"
16 #include "llvm/Analysis/TargetLibraryInfo.h"
17 #include "llvm/Config/llvm-config.h"
18 #include "llvm/IR/Constants.h"
19 #include "llvm/IR/Function.h"
20 #include "llvm/IR/GlobalVariable.h"
21 #include "llvm/IR/InstIterator.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 <iterator>
33 #include <string>
34 #include <tuple>
35 #include <utility>
36 
37 #ifdef EXPENSIVE_CHECKS
38 #include "llvm/ADT/ScopeExit.h"
39 #endif
40 
41 using namespace llvm;
42 
43 #define DEBUG_TYPE "lcg"
44 
45 void LazyCallGraph::EdgeSequence::insertEdgeInternal(Node &TargetN,
46                                                      Edge::Kind EK) {
47   EdgeIndexMap.insert({&TargetN, Edges.size()});
48   Edges.emplace_back(TargetN, EK);
49 }
50 
51 void LazyCallGraph::EdgeSequence::setEdgeKind(Node &TargetN, Edge::Kind EK) {
52   Edges[EdgeIndexMap.find(&TargetN)->second].setKind(EK);
53 }
54 
55 bool LazyCallGraph::EdgeSequence::removeEdgeInternal(Node &TargetN) {
56   auto IndexMapI = EdgeIndexMap.find(&TargetN);
57   if (IndexMapI == EdgeIndexMap.end())
58     return false;
59 
60   Edges[IndexMapI->second] = Edge();
61   EdgeIndexMap.erase(IndexMapI);
62   return true;
63 }
64 
65 static void addEdge(SmallVectorImpl<LazyCallGraph::Edge> &Edges,
66                     DenseMap<LazyCallGraph::Node *, int> &EdgeIndexMap,
67                     LazyCallGraph::Node &N, LazyCallGraph::Edge::Kind EK) {
68   if (!EdgeIndexMap.insert({&N, Edges.size()}).second)
69     return;
70 
71   LLVM_DEBUG(dbgs() << "    Added callable function: " << N.getName() << "\n");
72   Edges.emplace_back(LazyCallGraph::Edge(N, EK));
73 }
74 
75 LazyCallGraph::EdgeSequence &LazyCallGraph::Node::populateSlow() {
76   assert(!Edges && "Must not have already populated the edges for this node!");
77 
78   LLVM_DEBUG(dbgs() << "  Adding functions called by '" << getName()
79                     << "' to the graph.\n");
80 
81   Edges = EdgeSequence();
82 
83   SmallVector<Constant *, 16> Worklist;
84   SmallPtrSet<Function *, 4> Callees;
85   SmallPtrSet<Constant *, 16> Visited;
86 
87   // Find all the potential call graph edges in this function. We track both
88   // actual call edges and indirect references to functions. The direct calls
89   // are trivially added, but to accumulate the latter we walk the instructions
90   // and add every operand which is a constant to the worklist to process
91   // afterward.
92   //
93   // Note that we consider *any* function with a definition to be a viable
94   // edge. Even if the function's definition is subject to replacement by
95   // some other module (say, a weak definition) there may still be
96   // optimizations which essentially speculate based on the definition and
97   // a way to check that the specific definition is in fact the one being
98   // used. For example, this could be done by moving the weak definition to
99   // a strong (internal) definition and making the weak definition be an
100   // alias. Then a test of the address of the weak function against the new
101   // strong definition's address would be an effective way to determine the
102   // safety of optimizing a direct call edge.
103   for (BasicBlock &BB : *F)
104     for (Instruction &I : BB) {
105       if (auto *CB = dyn_cast<CallBase>(&I))
106         if (Function *Callee = CB->getCalledFunction())
107           if (!Callee->isDeclaration())
108             if (Callees.insert(Callee).second) {
109               Visited.insert(Callee);
110               addEdge(Edges->Edges, Edges->EdgeIndexMap, G->get(*Callee),
111                       LazyCallGraph::Edge::Call);
112             }
113 
114       for (Value *Op : I.operand_values())
115         if (Constant *C = dyn_cast<Constant>(Op))
116           if (Visited.insert(C).second)
117             Worklist.push_back(C);
118     }
119 
120   // We've collected all the constant (and thus potentially function or
121   // function containing) operands to all of the instructions in the function.
122   // Process them (recursively) collecting every function found.
123   visitReferences(Worklist, Visited, [&](Function &F) {
124     addEdge(Edges->Edges, Edges->EdgeIndexMap, G->get(F),
125             LazyCallGraph::Edge::Ref);
126   });
127 
128   // Add implicit reference edges to any defined libcall functions (if we
129   // haven't found an explicit edge).
130   for (auto *F : G->LibFunctions)
131     if (!Visited.count(F))
132       addEdge(Edges->Edges, Edges->EdgeIndexMap, G->get(*F),
133               LazyCallGraph::Edge::Ref);
134 
135   return *Edges;
136 }
137 
138 void LazyCallGraph::Node::replaceFunction(Function &NewF) {
139   assert(F != &NewF && "Must not replace a function with itself!");
140   F = &NewF;
141 }
142 
143 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
144 LLVM_DUMP_METHOD void LazyCallGraph::Node::dump() const {
145   dbgs() << *this << '\n';
146 }
147 #endif
148 
149 static bool isKnownLibFunction(Function &F, TargetLibraryInfo &TLI) {
150   LibFunc LF;
151 
152   // Either this is a normal library function or a "vectorizable"
153   // function.  Not using the VFDatabase here because this query
154   // is related only to libraries handled via the TLI.
155   return TLI.getLibFunc(F, LF) ||
156          TLI.isKnownVectorFunctionInLibrary(F.getName());
157 }
158 
159 LazyCallGraph::LazyCallGraph(
160     Module &M, function_ref<TargetLibraryInfo &(Function &)> GetTLI) {
161   LLVM_DEBUG(dbgs() << "Building CG for module: " << M.getModuleIdentifier()
162                     << "\n");
163   for (Function &F : M) {
164     if (F.isDeclaration())
165       continue;
166     // If this function is a known lib function to LLVM then we want to
167     // synthesize reference edges to it to model the fact that LLVM can turn
168     // arbitrary code into a library function call.
169     if (isKnownLibFunction(F, GetTLI(F)))
170       LibFunctions.insert(&F);
171 
172     if (F.hasLocalLinkage())
173       continue;
174 
175     // External linkage defined functions have edges to them from other
176     // modules.
177     LLVM_DEBUG(dbgs() << "  Adding '" << F.getName()
178                       << "' to entry set of the graph.\n");
179     addEdge(EntryEdges.Edges, EntryEdges.EdgeIndexMap, get(F), Edge::Ref);
180   }
181 
182   // Externally visible aliases of internal functions are also viable entry
183   // edges to the module.
184   for (auto &A : M.aliases()) {
185     if (A.hasLocalLinkage())
186       continue;
187     if (Function* F = dyn_cast<Function>(A.getAliasee())) {
188       LLVM_DEBUG(dbgs() << "  Adding '" << F->getName()
189                         << "' with alias '" << A.getName()
190                         << "' to entry set of the graph.\n");
191       addEdge(EntryEdges.Edges, EntryEdges.EdgeIndexMap, get(*F), Edge::Ref);
192     }
193   }
194 
195   // Now add entry nodes for functions reachable via initializers to globals.
196   SmallVector<Constant *, 16> Worklist;
197   SmallPtrSet<Constant *, 16> Visited;
198   for (GlobalVariable &GV : M.globals())
199     if (GV.hasInitializer())
200       if (Visited.insert(GV.getInitializer()).second)
201         Worklist.push_back(GV.getInitializer());
202 
203   LLVM_DEBUG(
204       dbgs() << "  Adding functions referenced by global initializers to the "
205                 "entry set.\n");
206   visitReferences(Worklist, Visited, [&](Function &F) {
207     addEdge(EntryEdges.Edges, EntryEdges.EdgeIndexMap, get(F),
208             LazyCallGraph::Edge::Ref);
209   });
210 }
211 
212 LazyCallGraph::LazyCallGraph(LazyCallGraph &&G)
213     : BPA(std::move(G.BPA)), NodeMap(std::move(G.NodeMap)),
214       EntryEdges(std::move(G.EntryEdges)), SCCBPA(std::move(G.SCCBPA)),
215       SCCMap(std::move(G.SCCMap)),
216       LibFunctions(std::move(G.LibFunctions)) {
217   updateGraphPtrs();
218 }
219 
220 bool LazyCallGraph::invalidate(Module &, const PreservedAnalyses &PA,
221                                ModuleAnalysisManager::Invalidator &) {
222   // Check whether the analysis, all analyses on functions, or the function's
223   // CFG have been preserved.
224   auto PAC = PA.getChecker<llvm::LazyCallGraphAnalysis>();
225   return !(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Module>>());
226 }
227 
228 LazyCallGraph &LazyCallGraph::operator=(LazyCallGraph &&G) {
229   BPA = std::move(G.BPA);
230   NodeMap = std::move(G.NodeMap);
231   EntryEdges = std::move(G.EntryEdges);
232   SCCBPA = std::move(G.SCCBPA);
233   SCCMap = std::move(G.SCCMap);
234   LibFunctions = std::move(G.LibFunctions);
235   updateGraphPtrs();
236   return *this;
237 }
238 
239 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
240 LLVM_DUMP_METHOD void LazyCallGraph::SCC::dump() const {
241   dbgs() << *this << '\n';
242 }
243 #endif
244 
245 #if !defined(NDEBUG) || defined(EXPENSIVE_CHECKS)
246 void LazyCallGraph::SCC::verify() {
247   assert(OuterRefSCC && "Can't have a null RefSCC!");
248   assert(!Nodes.empty() && "Can't have an empty SCC!");
249 
250   for (Node *N : Nodes) {
251     assert(N && "Can't have a null node!");
252     assert(OuterRefSCC->G->lookupSCC(*N) == this &&
253            "Node does not map to this SCC!");
254     assert(N->DFSNumber == -1 &&
255            "Must set DFS numbers to -1 when adding a node to an SCC!");
256     assert(N->LowLink == -1 &&
257            "Must set low link to -1 when adding a node to an SCC!");
258     for (Edge &E : **N)
259       assert(E.getNode().isPopulated() && "Can't have an unpopulated node!");
260 
261 #ifdef EXPENSIVE_CHECKS
262     // Verify that all nodes in this SCC can reach all other nodes.
263     SmallVector<Node *, 4> Worklist;
264     SmallPtrSet<Node *, 4> Visited;
265     Worklist.push_back(N);
266     while (!Worklist.empty()) {
267       Node *VisitingNode = Worklist.pop_back_val();
268       if (!Visited.insert(VisitingNode).second)
269         continue;
270       for (Edge &E : (*VisitingNode)->calls())
271         Worklist.push_back(&E.getNode());
272     }
273     for (Node *NodeToVisit : Nodes) {
274       assert(Visited.contains(NodeToVisit) &&
275              "Cannot reach all nodes within SCC");
276     }
277 #endif
278   }
279 }
280 #endif
281 
282 bool LazyCallGraph::SCC::isParentOf(const SCC &C) const {
283   if (this == &C)
284     return false;
285 
286   for (Node &N : *this)
287     for (Edge &E : N->calls())
288       if (OuterRefSCC->G->lookupSCC(E.getNode()) == &C)
289         return true;
290 
291   // No edges found.
292   return false;
293 }
294 
295 bool LazyCallGraph::SCC::isAncestorOf(const SCC &TargetC) const {
296   if (this == &TargetC)
297     return false;
298 
299   LazyCallGraph &G = *OuterRefSCC->G;
300 
301   // Start with this SCC.
302   SmallPtrSet<const SCC *, 16> Visited = {this};
303   SmallVector<const SCC *, 16> Worklist = {this};
304 
305   // Walk down the graph until we run out of edges or find a path to TargetC.
306   do {
307     const SCC &C = *Worklist.pop_back_val();
308     for (Node &N : C)
309       for (Edge &E : N->calls()) {
310         SCC *CalleeC = G.lookupSCC(E.getNode());
311         if (!CalleeC)
312           continue;
313 
314         // If the callee's SCC is the TargetC, we're done.
315         if (CalleeC == &TargetC)
316           return true;
317 
318         // If this is the first time we've reached this SCC, put it on the
319         // worklist to recurse through.
320         if (Visited.insert(CalleeC).second)
321           Worklist.push_back(CalleeC);
322       }
323   } while (!Worklist.empty());
324 
325   // No paths found.
326   return false;
327 }
328 
329 LazyCallGraph::RefSCC::RefSCC(LazyCallGraph &G) : G(&G) {}
330 
331 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
332 LLVM_DUMP_METHOD void LazyCallGraph::RefSCC::dump() const {
333   dbgs() << *this << '\n';
334 }
335 #endif
336 
337 #if !defined(NDEBUG) || defined(EXPENSIVE_CHECKS)
338 void LazyCallGraph::RefSCC::verify() {
339   assert(G && "Can't have a null graph!");
340   assert(!SCCs.empty() && "Can't have an empty SCC!");
341 
342   // Verify basic properties of the SCCs.
343   SmallPtrSet<SCC *, 4> SCCSet;
344   for (SCC *C : SCCs) {
345     assert(C && "Can't have a null SCC!");
346     C->verify();
347     assert(&C->getOuterRefSCC() == this &&
348            "SCC doesn't think it is inside this RefSCC!");
349     bool Inserted = SCCSet.insert(C).second;
350     assert(Inserted && "Found a duplicate SCC!");
351     auto IndexIt = SCCIndices.find(C);
352     assert(IndexIt != SCCIndices.end() &&
353            "Found an SCC that doesn't have an index!");
354   }
355 
356   // Check that our indices map correctly.
357   for (auto &SCCIndexPair : SCCIndices) {
358     SCC *C = SCCIndexPair.first;
359     int i = SCCIndexPair.second;
360     assert(C && "Can't have a null SCC in the indices!");
361     assert(SCCSet.count(C) && "Found an index for an SCC not in the RefSCC!");
362     assert(SCCs[i] == C && "Index doesn't point to SCC!");
363   }
364 
365   // Check that the SCCs are in fact in post-order.
366   for (int i = 0, Size = SCCs.size(); i < Size; ++i) {
367     SCC &SourceSCC = *SCCs[i];
368     for (Node &N : SourceSCC)
369       for (Edge &E : *N) {
370         if (!E.isCall())
371           continue;
372         SCC &TargetSCC = *G->lookupSCC(E.getNode());
373         if (&TargetSCC.getOuterRefSCC() == this) {
374           assert(SCCIndices.find(&TargetSCC)->second <= i &&
375                  "Edge between SCCs violates post-order relationship.");
376           continue;
377         }
378       }
379   }
380 
381 #ifdef EXPENSIVE_CHECKS
382   // Verify that all nodes in this RefSCC can reach all other nodes.
383   SmallVector<Node *> Nodes;
384   for (SCC *C : SCCs) {
385     for (Node &N : *C)
386       Nodes.push_back(&N);
387   }
388   for (Node *N : Nodes) {
389     SmallVector<Node *, 4> Worklist;
390     SmallPtrSet<Node *, 4> Visited;
391     Worklist.push_back(N);
392     while (!Worklist.empty()) {
393       Node *VisitingNode = Worklist.pop_back_val();
394       if (!Visited.insert(VisitingNode).second)
395         continue;
396       for (Edge &E : **VisitingNode)
397         Worklist.push_back(&E.getNode());
398     }
399     for (Node *NodeToVisit : Nodes) {
400       assert(Visited.contains(NodeToVisit) &&
401              "Cannot reach all nodes within RefSCC");
402     }
403   }
404 #endif
405 }
406 #endif
407 
408 bool LazyCallGraph::RefSCC::isParentOf(const RefSCC &RC) const {
409   if (&RC == this)
410     return false;
411 
412   // Search all edges to see if this is a parent.
413   for (SCC &C : *this)
414     for (Node &N : C)
415       for (Edge &E : *N)
416         if (G->lookupRefSCC(E.getNode()) == &RC)
417           return true;
418 
419   return false;
420 }
421 
422 bool LazyCallGraph::RefSCC::isAncestorOf(const RefSCC &RC) const {
423   if (&RC == this)
424     return false;
425 
426   // For each descendant of this RefSCC, see if one of its children is the
427   // argument. If not, add that descendant to the worklist and continue
428   // searching.
429   SmallVector<const RefSCC *, 4> Worklist;
430   SmallPtrSet<const RefSCC *, 4> Visited;
431   Worklist.push_back(this);
432   Visited.insert(this);
433   do {
434     const RefSCC &DescendantRC = *Worklist.pop_back_val();
435     for (SCC &C : DescendantRC)
436       for (Node &N : C)
437         for (Edge &E : *N) {
438           auto *ChildRC = G->lookupRefSCC(E.getNode());
439           if (ChildRC == &RC)
440             return true;
441           if (!ChildRC || !Visited.insert(ChildRC).second)
442             continue;
443           Worklist.push_back(ChildRC);
444         }
445   } while (!Worklist.empty());
446 
447   return false;
448 }
449 
450 /// Generic helper that updates a postorder sequence of SCCs for a potentially
451 /// cycle-introducing edge insertion.
452 ///
453 /// A postorder sequence of SCCs of a directed graph has one fundamental
454 /// property: all deges in the DAG of SCCs point "up" the sequence. That is,
455 /// all edges in the SCC DAG point to prior SCCs in the sequence.
456 ///
457 /// This routine both updates a postorder sequence and uses that sequence to
458 /// compute the set of SCCs connected into a cycle. It should only be called to
459 /// insert a "downward" edge which will require changing the sequence to
460 /// restore it to a postorder.
461 ///
462 /// When inserting an edge from an earlier SCC to a later SCC in some postorder
463 /// sequence, all of the SCCs which may be impacted are in the closed range of
464 /// those two within the postorder sequence. The algorithm used here to restore
465 /// the state is as follows:
466 ///
467 /// 1) Starting from the source SCC, construct a set of SCCs which reach the
468 ///    source SCC consisting of just the source SCC. Then scan toward the
469 ///    target SCC in postorder and for each SCC, if it has an edge to an SCC
470 ///    in the set, add it to the set. Otherwise, the source SCC is not
471 ///    a successor, move it in the postorder sequence to immediately before
472 ///    the source SCC, shifting the source SCC and all SCCs in the set one
473 ///    position toward the target SCC. Stop scanning after processing the
474 ///    target SCC.
475 /// 2) If the source SCC is now past the target SCC in the postorder sequence,
476 ///    and thus the new edge will flow toward the start, we are done.
477 /// 3) Otherwise, starting from the target SCC, walk all edges which reach an
478 ///    SCC between the source and the target, and add them to the set of
479 ///    connected SCCs, then recurse through them. Once a complete set of the
480 ///    SCCs the target connects to is known, hoist the remaining SCCs between
481 ///    the source and the target to be above the target. Note that there is no
482 ///    need to process the source SCC, it is already known to connect.
483 /// 4) At this point, all of the SCCs in the closed range between the source
484 ///    SCC and the target SCC in the postorder sequence are connected,
485 ///    including the target SCC and the source SCC. Inserting the edge from
486 ///    the source SCC to the target SCC will form a cycle out of precisely
487 ///    these SCCs. Thus we can merge all of the SCCs in this closed range into
488 ///    a single SCC.
489 ///
490 /// This process has various important properties:
491 /// - Only mutates the SCCs when adding the edge actually changes the SCC
492 ///   structure.
493 /// - Never mutates SCCs which are unaffected by the change.
494 /// - Updates the postorder sequence to correctly satisfy the postorder
495 ///   constraint after the edge is inserted.
496 /// - Only reorders SCCs in the closed postorder sequence from the source to
497 ///   the target, so easy to bound how much has changed even in the ordering.
498 /// - Big-O is the number of edges in the closed postorder range of SCCs from
499 ///   source to target.
500 ///
501 /// This helper routine, in addition to updating the postorder sequence itself
502 /// will also update a map from SCCs to indices within that sequence.
503 ///
504 /// The sequence and the map must operate on pointers to the SCC type.
505 ///
506 /// Two callbacks must be provided. The first computes the subset of SCCs in
507 /// the postorder closed range from the source to the target which connect to
508 /// the source SCC via some (transitive) set of edges. The second computes the
509 /// subset of the same range which the target SCC connects to via some
510 /// (transitive) set of edges. Both callbacks should populate the set argument
511 /// provided.
512 template <typename SCCT, typename PostorderSequenceT, typename SCCIndexMapT,
513           typename ComputeSourceConnectedSetCallableT,
514           typename ComputeTargetConnectedSetCallableT>
515 static iterator_range<typename PostorderSequenceT::iterator>
516 updatePostorderSequenceForEdgeInsertion(
517     SCCT &SourceSCC, SCCT &TargetSCC, PostorderSequenceT &SCCs,
518     SCCIndexMapT &SCCIndices,
519     ComputeSourceConnectedSetCallableT ComputeSourceConnectedSet,
520     ComputeTargetConnectedSetCallableT ComputeTargetConnectedSet) {
521   int SourceIdx = SCCIndices[&SourceSCC];
522   int TargetIdx = SCCIndices[&TargetSCC];
523   assert(SourceIdx < TargetIdx && "Cannot have equal indices here!");
524 
525   SmallPtrSet<SCCT *, 4> ConnectedSet;
526 
527   // Compute the SCCs which (transitively) reach the source.
528   ComputeSourceConnectedSet(ConnectedSet);
529 
530   // Partition the SCCs in this part of the port-order sequence so only SCCs
531   // connecting to the source remain between it and the target. This is
532   // a benign partition as it preserves postorder.
533   auto SourceI = std::stable_partition(
534       SCCs.begin() + SourceIdx, SCCs.begin() + TargetIdx + 1,
535       [&ConnectedSet](SCCT *C) { return !ConnectedSet.count(C); });
536   for (int i = SourceIdx, e = TargetIdx + 1; i < e; ++i)
537     SCCIndices.find(SCCs[i])->second = i;
538 
539   // If the target doesn't connect to the source, then we've corrected the
540   // post-order and there are no cycles formed.
541   if (!ConnectedSet.count(&TargetSCC)) {
542     assert(SourceI > (SCCs.begin() + SourceIdx) &&
543            "Must have moved the source to fix the post-order.");
544     assert(*std::prev(SourceI) == &TargetSCC &&
545            "Last SCC to move should have bene the target.");
546 
547     // Return an empty range at the target SCC indicating there is nothing to
548     // merge.
549     return make_range(std::prev(SourceI), std::prev(SourceI));
550   }
551 
552   assert(SCCs[TargetIdx] == &TargetSCC &&
553          "Should not have moved target if connected!");
554   SourceIdx = SourceI - SCCs.begin();
555   assert(SCCs[SourceIdx] == &SourceSCC &&
556          "Bad updated index computation for the source SCC!");
557 
558 
559   // See whether there are any remaining intervening SCCs between the source
560   // and target. If so we need to make sure they all are reachable form the
561   // target.
562   if (SourceIdx + 1 < TargetIdx) {
563     ConnectedSet.clear();
564     ComputeTargetConnectedSet(ConnectedSet);
565 
566     // Partition SCCs so that only SCCs reached from the target remain between
567     // the source and the target. This preserves postorder.
568     auto TargetI = std::stable_partition(
569         SCCs.begin() + SourceIdx + 1, SCCs.begin() + TargetIdx + 1,
570         [&ConnectedSet](SCCT *C) { return ConnectedSet.count(C); });
571     for (int i = SourceIdx + 1, e = TargetIdx + 1; i < e; ++i)
572       SCCIndices.find(SCCs[i])->second = i;
573     TargetIdx = std::prev(TargetI) - SCCs.begin();
574     assert(SCCs[TargetIdx] == &TargetSCC &&
575            "Should always end with the target!");
576   }
577 
578   // At this point, we know that connecting source to target forms a cycle
579   // because target connects back to source, and we know that all of the SCCs
580   // between the source and target in the postorder sequence participate in that
581   // cycle.
582   return make_range(SCCs.begin() + SourceIdx, SCCs.begin() + TargetIdx);
583 }
584 
585 bool
586 LazyCallGraph::RefSCC::switchInternalEdgeToCall(
587     Node &SourceN, Node &TargetN,
588     function_ref<void(ArrayRef<SCC *> MergeSCCs)> MergeCB) {
589   assert(!(*SourceN)[TargetN].isCall() && "Must start with a ref edge!");
590   SmallVector<SCC *, 1> DeletedSCCs;
591 
592 #ifdef EXPENSIVE_CHECKS
593   verify();
594   auto VerifyOnExit = make_scope_exit([&]() { verify(); });
595 #endif
596 
597   SCC &SourceSCC = *G->lookupSCC(SourceN);
598   SCC &TargetSCC = *G->lookupSCC(TargetN);
599 
600   // If the two nodes are already part of the same SCC, we're also done as
601   // we've just added more connectivity.
602   if (&SourceSCC == &TargetSCC) {
603     SourceN->setEdgeKind(TargetN, Edge::Call);
604     return false; // No new cycle.
605   }
606 
607   // At this point we leverage the postorder list of SCCs to detect when the
608   // insertion of an edge changes the SCC structure in any way.
609   //
610   // First and foremost, we can eliminate the need for any changes when the
611   // edge is toward the beginning of the postorder sequence because all edges
612   // flow in that direction already. Thus adding a new one cannot form a cycle.
613   int SourceIdx = SCCIndices[&SourceSCC];
614   int TargetIdx = SCCIndices[&TargetSCC];
615   if (TargetIdx < SourceIdx) {
616     SourceN->setEdgeKind(TargetN, Edge::Call);
617     return false; // No new cycle.
618   }
619 
620   // Compute the SCCs which (transitively) reach the source.
621   auto ComputeSourceConnectedSet = [&](SmallPtrSetImpl<SCC *> &ConnectedSet) {
622 #ifdef EXPENSIVE_CHECKS
623     // Check that the RefSCC is still valid before computing this as the
624     // results will be nonsensical of we've broken its invariants.
625     verify();
626 #endif
627     ConnectedSet.insert(&SourceSCC);
628     auto IsConnected = [&](SCC &C) {
629       for (Node &N : C)
630         for (Edge &E : N->calls())
631           if (ConnectedSet.count(G->lookupSCC(E.getNode())))
632             return true;
633 
634       return false;
635     };
636 
637     for (SCC *C :
638          make_range(SCCs.begin() + SourceIdx + 1, SCCs.begin() + TargetIdx + 1))
639       if (IsConnected(*C))
640         ConnectedSet.insert(C);
641   };
642 
643   // Use a normal worklist to find which SCCs the target connects to. We still
644   // bound the search based on the range in the postorder list we care about,
645   // but because this is forward connectivity we just "recurse" through the
646   // edges.
647   auto ComputeTargetConnectedSet = [&](SmallPtrSetImpl<SCC *> &ConnectedSet) {
648 #ifdef EXPENSIVE_CHECKS
649     // Check that the RefSCC is still valid before computing this as the
650     // results will be nonsensical of we've broken its invariants.
651     verify();
652 #endif
653     ConnectedSet.insert(&TargetSCC);
654     SmallVector<SCC *, 4> Worklist;
655     Worklist.push_back(&TargetSCC);
656     do {
657       SCC &C = *Worklist.pop_back_val();
658       for (Node &N : C)
659         for (Edge &E : *N) {
660           if (!E.isCall())
661             continue;
662           SCC &EdgeC = *G->lookupSCC(E.getNode());
663           if (&EdgeC.getOuterRefSCC() != this)
664             // Not in this RefSCC...
665             continue;
666           if (SCCIndices.find(&EdgeC)->second <= SourceIdx)
667             // Not in the postorder sequence between source and target.
668             continue;
669 
670           if (ConnectedSet.insert(&EdgeC).second)
671             Worklist.push_back(&EdgeC);
672         }
673     } while (!Worklist.empty());
674   };
675 
676   // Use a generic helper to update the postorder sequence of SCCs and return
677   // a range of any SCCs connected into a cycle by inserting this edge. This
678   // routine will also take care of updating the indices into the postorder
679   // sequence.
680   auto MergeRange = updatePostorderSequenceForEdgeInsertion(
681       SourceSCC, TargetSCC, SCCs, SCCIndices, ComputeSourceConnectedSet,
682       ComputeTargetConnectedSet);
683 
684   // Run the user's callback on the merged SCCs before we actually merge them.
685   if (MergeCB)
686     MergeCB(makeArrayRef(MergeRange.begin(), MergeRange.end()));
687 
688   // If the merge range is empty, then adding the edge didn't actually form any
689   // new cycles. We're done.
690   if (MergeRange.empty()) {
691     // Now that the SCC structure is finalized, flip the kind to call.
692     SourceN->setEdgeKind(TargetN, Edge::Call);
693     return false; // No new cycle.
694   }
695 
696 #ifdef EXPENSIVE_CHECKS
697   // Before merging, check that the RefSCC remains valid after all the
698   // postorder updates.
699   verify();
700 #endif
701 
702   // Otherwise we need to merge all of the SCCs in the cycle into a single
703   // result SCC.
704   //
705   // NB: We merge into the target because all of these functions were already
706   // reachable from the target, meaning any SCC-wide properties deduced about it
707   // other than the set of functions within it will not have changed.
708   for (SCC *C : MergeRange) {
709     assert(C != &TargetSCC &&
710            "We merge *into* the target and shouldn't process it here!");
711     SCCIndices.erase(C);
712     TargetSCC.Nodes.append(C->Nodes.begin(), C->Nodes.end());
713     for (Node *N : C->Nodes)
714       G->SCCMap[N] = &TargetSCC;
715     C->clear();
716     DeletedSCCs.push_back(C);
717   }
718 
719   // Erase the merged SCCs from the list and update the indices of the
720   // remaining SCCs.
721   int IndexOffset = MergeRange.end() - MergeRange.begin();
722   auto EraseEnd = SCCs.erase(MergeRange.begin(), MergeRange.end());
723   for (SCC *C : make_range(EraseEnd, SCCs.end()))
724     SCCIndices[C] -= IndexOffset;
725 
726   // Now that the SCC structure is finalized, flip the kind to call.
727   SourceN->setEdgeKind(TargetN, Edge::Call);
728 
729   // And we're done, but we did form a new cycle.
730   return true;
731 }
732 
733 void LazyCallGraph::RefSCC::switchTrivialInternalEdgeToRef(Node &SourceN,
734                                                            Node &TargetN) {
735   assert((*SourceN)[TargetN].isCall() && "Must start with a call edge!");
736 
737 #ifdef EXPENSIVE_CHECKS
738   verify();
739   auto VerifyOnExit = make_scope_exit([&]() { verify(); });
740 #endif
741 
742   assert(G->lookupRefSCC(SourceN) == this &&
743          "Source must be in this RefSCC.");
744   assert(G->lookupRefSCC(TargetN) == this &&
745          "Target must be in this RefSCC.");
746   assert(G->lookupSCC(SourceN) != G->lookupSCC(TargetN) &&
747          "Source and Target must be in separate SCCs for this to be trivial!");
748 
749   // Set the edge kind.
750   SourceN->setEdgeKind(TargetN, Edge::Ref);
751 }
752 
753 iterator_range<LazyCallGraph::RefSCC::iterator>
754 LazyCallGraph::RefSCC::switchInternalEdgeToRef(Node &SourceN, Node &TargetN) {
755   assert((*SourceN)[TargetN].isCall() && "Must start with a call edge!");
756 
757 #ifdef EXPENSIVE_CHECKS
758   verify();
759   auto VerifyOnExit = make_scope_exit([&]() { verify(); });
760 #endif
761 
762   assert(G->lookupRefSCC(SourceN) == this &&
763          "Source must be in this RefSCC.");
764   assert(G->lookupRefSCC(TargetN) == this &&
765          "Target must be in this RefSCC.");
766 
767   SCC &TargetSCC = *G->lookupSCC(TargetN);
768   assert(G->lookupSCC(SourceN) == &TargetSCC && "Source and Target must be in "
769                                                 "the same SCC to require the "
770                                                 "full CG update.");
771 
772   // Set the edge kind.
773   SourceN->setEdgeKind(TargetN, Edge::Ref);
774 
775   // Otherwise we are removing a call edge from a single SCC. This may break
776   // the cycle. In order to compute the new set of SCCs, we need to do a small
777   // DFS over the nodes within the SCC to form any sub-cycles that remain as
778   // distinct SCCs and compute a postorder over the resulting SCCs.
779   //
780   // However, we specially handle the target node. The target node is known to
781   // reach all other nodes in the original SCC by definition. This means that
782   // we want the old SCC to be replaced with an SCC containing that node as it
783   // will be the root of whatever SCC DAG results from the DFS. Assumptions
784   // about an SCC such as the set of functions called will continue to hold,
785   // etc.
786 
787   SCC &OldSCC = TargetSCC;
788   SmallVector<std::pair<Node *, EdgeSequence::call_iterator>, 16> DFSStack;
789   SmallVector<Node *, 16> PendingSCCStack;
790   SmallVector<SCC *, 4> NewSCCs;
791 
792   // Prepare the nodes for a fresh DFS.
793   SmallVector<Node *, 16> Worklist;
794   Worklist.swap(OldSCC.Nodes);
795   for (Node *N : Worklist) {
796     N->DFSNumber = N->LowLink = 0;
797     G->SCCMap.erase(N);
798   }
799 
800   // Force the target node to be in the old SCC. This also enables us to take
801   // a very significant short-cut in the standard Tarjan walk to re-form SCCs
802   // below: whenever we build an edge that reaches the target node, we know
803   // that the target node eventually connects back to all other nodes in our
804   // walk. As a consequence, we can detect and handle participants in that
805   // cycle without walking all the edges that form this connection, and instead
806   // by relying on the fundamental guarantee coming into this operation (all
807   // nodes are reachable from the target due to previously forming an SCC).
808   TargetN.DFSNumber = TargetN.LowLink = -1;
809   OldSCC.Nodes.push_back(&TargetN);
810   G->SCCMap[&TargetN] = &OldSCC;
811 
812   // Scan down the stack and DFS across the call edges.
813   for (Node *RootN : Worklist) {
814     assert(DFSStack.empty() &&
815            "Cannot begin a new root with a non-empty DFS stack!");
816     assert(PendingSCCStack.empty() &&
817            "Cannot begin a new root with pending nodes for an SCC!");
818 
819     // Skip any nodes we've already reached in the DFS.
820     if (RootN->DFSNumber != 0) {
821       assert(RootN->DFSNumber == -1 &&
822              "Shouldn't have any mid-DFS root nodes!");
823       continue;
824     }
825 
826     RootN->DFSNumber = RootN->LowLink = 1;
827     int NextDFSNumber = 2;
828 
829     DFSStack.push_back({RootN, (*RootN)->call_begin()});
830     do {
831       Node *N;
832       EdgeSequence::call_iterator I;
833       std::tie(N, I) = DFSStack.pop_back_val();
834       auto E = (*N)->call_end();
835       while (I != E) {
836         Node &ChildN = I->getNode();
837         if (ChildN.DFSNumber == 0) {
838           // We haven't yet visited this child, so descend, pushing the current
839           // node onto the stack.
840           DFSStack.push_back({N, I});
841 
842           assert(!G->SCCMap.count(&ChildN) &&
843                  "Found a node with 0 DFS number but already in an SCC!");
844           ChildN.DFSNumber = ChildN.LowLink = NextDFSNumber++;
845           N = &ChildN;
846           I = (*N)->call_begin();
847           E = (*N)->call_end();
848           continue;
849         }
850 
851         // Check for the child already being part of some component.
852         if (ChildN.DFSNumber == -1) {
853           if (G->lookupSCC(ChildN) == &OldSCC) {
854             // If the child is part of the old SCC, we know that it can reach
855             // every other node, so we have formed a cycle. Pull the entire DFS
856             // and pending stacks into it. See the comment above about setting
857             // up the old SCC for why we do this.
858             int OldSize = OldSCC.size();
859             OldSCC.Nodes.push_back(N);
860             OldSCC.Nodes.append(PendingSCCStack.begin(), PendingSCCStack.end());
861             PendingSCCStack.clear();
862             while (!DFSStack.empty())
863               OldSCC.Nodes.push_back(DFSStack.pop_back_val().first);
864             for (Node &N : drop_begin(OldSCC, OldSize)) {
865               N.DFSNumber = N.LowLink = -1;
866               G->SCCMap[&N] = &OldSCC;
867             }
868             N = nullptr;
869             break;
870           }
871 
872           // If the child has already been added to some child component, it
873           // couldn't impact the low-link of this parent because it isn't
874           // connected, and thus its low-link isn't relevant so skip it.
875           ++I;
876           continue;
877         }
878 
879         // Track the lowest linked child as the lowest link for this node.
880         assert(ChildN.LowLink > 0 && "Must have a positive low-link number!");
881         if (ChildN.LowLink < N->LowLink)
882           N->LowLink = ChildN.LowLink;
883 
884         // Move to the next edge.
885         ++I;
886       }
887       if (!N)
888         // Cleared the DFS early, start another round.
889         break;
890 
891       // We've finished processing N and its descendants, put it on our pending
892       // SCC stack to eventually get merged into an SCC of nodes.
893       PendingSCCStack.push_back(N);
894 
895       // If this node is linked to some lower entry, continue walking up the
896       // stack.
897       if (N->LowLink != N->DFSNumber)
898         continue;
899 
900       // Otherwise, we've completed an SCC. Append it to our post order list of
901       // SCCs.
902       int RootDFSNumber = N->DFSNumber;
903       // Find the range of the node stack by walking down until we pass the
904       // root DFS number.
905       auto SCCNodes = make_range(
906           PendingSCCStack.rbegin(),
907           find_if(reverse(PendingSCCStack), [RootDFSNumber](const Node *N) {
908             return N->DFSNumber < RootDFSNumber;
909           }));
910 
911       // Form a new SCC out of these nodes and then clear them off our pending
912       // stack.
913       NewSCCs.push_back(G->createSCC(*this, SCCNodes));
914       for (Node &N : *NewSCCs.back()) {
915         N.DFSNumber = N.LowLink = -1;
916         G->SCCMap[&N] = NewSCCs.back();
917       }
918       PendingSCCStack.erase(SCCNodes.end().base(), PendingSCCStack.end());
919     } while (!DFSStack.empty());
920   }
921 
922   // Insert the remaining SCCs before the old one. The old SCC can reach all
923   // other SCCs we form because it contains the target node of the removed edge
924   // of the old SCC. This means that we will have edges into all of the new
925   // SCCs, which means the old one must come last for postorder.
926   int OldIdx = SCCIndices[&OldSCC];
927   SCCs.insert(SCCs.begin() + OldIdx, NewSCCs.begin(), NewSCCs.end());
928 
929   // Update the mapping from SCC* to index to use the new SCC*s, and remove the
930   // old SCC from the mapping.
931   for (int Idx = OldIdx, Size = SCCs.size(); Idx < Size; ++Idx)
932     SCCIndices[SCCs[Idx]] = Idx;
933 
934   return make_range(SCCs.begin() + OldIdx,
935                     SCCs.begin() + OldIdx + NewSCCs.size());
936 }
937 
938 void LazyCallGraph::RefSCC::switchOutgoingEdgeToCall(Node &SourceN,
939                                                      Node &TargetN) {
940   assert(!(*SourceN)[TargetN].isCall() && "Must start with a ref edge!");
941 
942   assert(G->lookupRefSCC(SourceN) == this && "Source must be in this RefSCC.");
943   assert(G->lookupRefSCC(TargetN) != this &&
944          "Target must not be in this RefSCC.");
945 #ifdef EXPENSIVE_CHECKS
946   assert(G->lookupRefSCC(TargetN)->isDescendantOf(*this) &&
947          "Target must be a descendant of the Source.");
948 #endif
949 
950   // Edges between RefSCCs are the same regardless of call or ref, so we can
951   // just flip the edge here.
952   SourceN->setEdgeKind(TargetN, Edge::Call);
953 
954 #ifdef EXPENSIVE_CHECKS
955   verify();
956 #endif
957 }
958 
959 void LazyCallGraph::RefSCC::switchOutgoingEdgeToRef(Node &SourceN,
960                                                     Node &TargetN) {
961   assert((*SourceN)[TargetN].isCall() && "Must start with a call edge!");
962 
963   assert(G->lookupRefSCC(SourceN) == this && "Source must be in this RefSCC.");
964   assert(G->lookupRefSCC(TargetN) != this &&
965          "Target must not be in this RefSCC.");
966 #ifdef EXPENSIVE_CHECKS
967   assert(G->lookupRefSCC(TargetN)->isDescendantOf(*this) &&
968          "Target must be a descendant of the Source.");
969 #endif
970 
971   // Edges between RefSCCs are the same regardless of call or ref, so we can
972   // just flip the edge here.
973   SourceN->setEdgeKind(TargetN, Edge::Ref);
974 
975 #ifdef EXPENSIVE_CHECKS
976   verify();
977 #endif
978 }
979 
980 void LazyCallGraph::RefSCC::insertInternalRefEdge(Node &SourceN,
981                                                   Node &TargetN) {
982   assert(G->lookupRefSCC(SourceN) == this && "Source must be in this RefSCC.");
983   assert(G->lookupRefSCC(TargetN) == this && "Target must be in this RefSCC.");
984 
985   SourceN->insertEdgeInternal(TargetN, Edge::Ref);
986 
987 #ifdef EXPENSIVE_CHECKS
988   verify();
989 #endif
990 }
991 
992 void LazyCallGraph::RefSCC::insertOutgoingEdge(Node &SourceN, Node &TargetN,
993                                                Edge::Kind EK) {
994   // First insert it into the caller.
995   SourceN->insertEdgeInternal(TargetN, EK);
996 
997   assert(G->lookupRefSCC(SourceN) == this && "Source must be in this RefSCC.");
998 
999   assert(G->lookupRefSCC(TargetN) != this &&
1000          "Target must not be in this RefSCC.");
1001 #ifdef EXPENSIVE_CHECKS
1002   assert(G->lookupRefSCC(TargetN)->isDescendantOf(*this) &&
1003          "Target must be a descendant of the Source.");
1004 #endif
1005 
1006 #ifdef EXPENSIVE_CHECKS
1007   verify();
1008 #endif
1009 }
1010 
1011 SmallVector<LazyCallGraph::RefSCC *, 1>
1012 LazyCallGraph::RefSCC::insertIncomingRefEdge(Node &SourceN, Node &TargetN) {
1013   assert(G->lookupRefSCC(TargetN) == this && "Target must be in this RefSCC.");
1014   RefSCC &SourceC = *G->lookupRefSCC(SourceN);
1015   assert(&SourceC != this && "Source must not be in this RefSCC.");
1016 #ifdef EXPENSIVE_CHECKS
1017   assert(SourceC.isDescendantOf(*this) &&
1018          "Source must be a descendant of the Target.");
1019 #endif
1020 
1021   SmallVector<RefSCC *, 1> DeletedRefSCCs;
1022 
1023 #ifdef EXPENSIVE_CHECKS
1024   verify();
1025   auto VerifyOnExit = make_scope_exit([&]() { verify(); });
1026 #endif
1027 
1028   int SourceIdx = G->RefSCCIndices[&SourceC];
1029   int TargetIdx = G->RefSCCIndices[this];
1030   assert(SourceIdx < TargetIdx &&
1031          "Postorder list doesn't see edge as incoming!");
1032 
1033   // Compute the RefSCCs which (transitively) reach the source. We do this by
1034   // working backwards from the source using the parent set in each RefSCC,
1035   // skipping any RefSCCs that don't fall in the postorder range. This has the
1036   // advantage of walking the sparser parent edge (in high fan-out graphs) but
1037   // more importantly this removes examining all forward edges in all RefSCCs
1038   // within the postorder range which aren't in fact connected. Only connected
1039   // RefSCCs (and their edges) are visited here.
1040   auto ComputeSourceConnectedSet = [&](SmallPtrSetImpl<RefSCC *> &Set) {
1041     Set.insert(&SourceC);
1042     auto IsConnected = [&](RefSCC &RC) {
1043       for (SCC &C : RC)
1044         for (Node &N : C)
1045           for (Edge &E : *N)
1046             if (Set.count(G->lookupRefSCC(E.getNode())))
1047               return true;
1048 
1049       return false;
1050     };
1051 
1052     for (RefSCC *C : make_range(G->PostOrderRefSCCs.begin() + SourceIdx + 1,
1053                                 G->PostOrderRefSCCs.begin() + TargetIdx + 1))
1054       if (IsConnected(*C))
1055         Set.insert(C);
1056   };
1057 
1058   // Use a normal worklist to find which SCCs the target connects to. We still
1059   // bound the search based on the range in the postorder list we care about,
1060   // but because this is forward connectivity we just "recurse" through the
1061   // edges.
1062   auto ComputeTargetConnectedSet = [&](SmallPtrSetImpl<RefSCC *> &Set) {
1063     Set.insert(this);
1064     SmallVector<RefSCC *, 4> Worklist;
1065     Worklist.push_back(this);
1066     do {
1067       RefSCC &RC = *Worklist.pop_back_val();
1068       for (SCC &C : RC)
1069         for (Node &N : C)
1070           for (Edge &E : *N) {
1071             RefSCC &EdgeRC = *G->lookupRefSCC(E.getNode());
1072             if (G->getRefSCCIndex(EdgeRC) <= SourceIdx)
1073               // Not in the postorder sequence between source and target.
1074               continue;
1075 
1076             if (Set.insert(&EdgeRC).second)
1077               Worklist.push_back(&EdgeRC);
1078           }
1079     } while (!Worklist.empty());
1080   };
1081 
1082   // Use a generic helper to update the postorder sequence of RefSCCs and return
1083   // a range of any RefSCCs connected into a cycle by inserting this edge. This
1084   // routine will also take care of updating the indices into the postorder
1085   // sequence.
1086   iterator_range<SmallVectorImpl<RefSCC *>::iterator> MergeRange =
1087       updatePostorderSequenceForEdgeInsertion(
1088           SourceC, *this, G->PostOrderRefSCCs, G->RefSCCIndices,
1089           ComputeSourceConnectedSet, ComputeTargetConnectedSet);
1090 
1091   // Build a set so we can do fast tests for whether a RefSCC will end up as
1092   // part of the merged RefSCC.
1093   SmallPtrSet<RefSCC *, 16> MergeSet(MergeRange.begin(), MergeRange.end());
1094 
1095   // This RefSCC will always be part of that set, so just insert it here.
1096   MergeSet.insert(this);
1097 
1098   // Now that we have identified all of the SCCs which need to be merged into
1099   // a connected set with the inserted edge, merge all of them into this SCC.
1100   SmallVector<SCC *, 16> MergedSCCs;
1101   int SCCIndex = 0;
1102   for (RefSCC *RC : MergeRange) {
1103     assert(RC != this && "We're merging into the target RefSCC, so it "
1104                          "shouldn't be in the range.");
1105 
1106     // Walk the inner SCCs to update their up-pointer and walk all the edges to
1107     // update any parent sets.
1108     // FIXME: We should try to find a way to avoid this (rather expensive) edge
1109     // walk by updating the parent sets in some other manner.
1110     for (SCC &InnerC : *RC) {
1111       InnerC.OuterRefSCC = this;
1112       SCCIndices[&InnerC] = SCCIndex++;
1113       for (Node &N : InnerC)
1114         G->SCCMap[&N] = &InnerC;
1115     }
1116 
1117     // Now merge in the SCCs. We can actually move here so try to reuse storage
1118     // the first time through.
1119     if (MergedSCCs.empty())
1120       MergedSCCs = std::move(RC->SCCs);
1121     else
1122       MergedSCCs.append(RC->SCCs.begin(), RC->SCCs.end());
1123     RC->SCCs.clear();
1124     DeletedRefSCCs.push_back(RC);
1125   }
1126 
1127   // Append our original SCCs to the merged list and move it into place.
1128   for (SCC &InnerC : *this)
1129     SCCIndices[&InnerC] = SCCIndex++;
1130   MergedSCCs.append(SCCs.begin(), SCCs.end());
1131   SCCs = std::move(MergedSCCs);
1132 
1133   // Remove the merged away RefSCCs from the post order sequence.
1134   for (RefSCC *RC : MergeRange)
1135     G->RefSCCIndices.erase(RC);
1136   int IndexOffset = MergeRange.end() - MergeRange.begin();
1137   auto EraseEnd =
1138       G->PostOrderRefSCCs.erase(MergeRange.begin(), MergeRange.end());
1139   for (RefSCC *RC : make_range(EraseEnd, G->PostOrderRefSCCs.end()))
1140     G->RefSCCIndices[RC] -= IndexOffset;
1141 
1142   // At this point we have a merged RefSCC with a post-order SCCs list, just
1143   // connect the nodes to form the new edge.
1144   SourceN->insertEdgeInternal(TargetN, Edge::Ref);
1145 
1146   // We return the list of SCCs which were merged so that callers can
1147   // invalidate any data they have associated with those SCCs. Note that these
1148   // SCCs are no longer in an interesting state (they are totally empty) but
1149   // the pointers will remain stable for the life of the graph itself.
1150   return DeletedRefSCCs;
1151 }
1152 
1153 void LazyCallGraph::RefSCC::removeOutgoingEdge(Node &SourceN, Node &TargetN) {
1154   assert(G->lookupRefSCC(SourceN) == this &&
1155          "The source must be a member of this RefSCC.");
1156   assert(G->lookupRefSCC(TargetN) != this &&
1157          "The target must not be a member of this RefSCC");
1158 
1159 #ifdef EXPENSIVE_CHECKS
1160   verify();
1161   auto VerifyOnExit = make_scope_exit([&]() { verify(); });
1162 #endif
1163 
1164   // First remove it from the node.
1165   bool Removed = SourceN->removeEdgeInternal(TargetN);
1166   (void)Removed;
1167   assert(Removed && "Target not in the edge set for this caller?");
1168 }
1169 
1170 SmallVector<LazyCallGraph::RefSCC *, 1>
1171 LazyCallGraph::RefSCC::removeInternalRefEdge(Node &SourceN,
1172                                              ArrayRef<Node *> TargetNs) {
1173   // We return a list of the resulting *new* RefSCCs in post-order.
1174   SmallVector<RefSCC *, 1> Result;
1175 
1176 #ifdef EXPENSIVE_CHECKS
1177   // Verify the RefSCC is valid to start with and that either we return an empty
1178   // list of result RefSCCs and this RefSCC remains valid, or we return new
1179   // RefSCCs and this RefSCC is dead.
1180   verify();
1181   auto VerifyOnExit = make_scope_exit([&]() {
1182     // If we didn't replace our RefSCC with new ones, check that this one
1183     // remains valid.
1184     if (G)
1185       verify();
1186   });
1187 #endif
1188 
1189   // First remove the actual edges.
1190   for (Node *TargetN : TargetNs) {
1191     assert(!(*SourceN)[*TargetN].isCall() &&
1192            "Cannot remove a call edge, it must first be made a ref edge");
1193 
1194     bool Removed = SourceN->removeEdgeInternal(*TargetN);
1195     (void)Removed;
1196     assert(Removed && "Target not in the edge set for this caller?");
1197   }
1198 
1199   // Direct self references don't impact the ref graph at all.
1200   if (llvm::all_of(TargetNs,
1201                    [&](Node *TargetN) { return &SourceN == TargetN; }))
1202     return Result;
1203 
1204   // If all targets are in the same SCC as the source, because no call edges
1205   // were removed there is no RefSCC structure change.
1206   SCC &SourceC = *G->lookupSCC(SourceN);
1207   if (llvm::all_of(TargetNs, [&](Node *TargetN) {
1208         return G->lookupSCC(*TargetN) == &SourceC;
1209       }))
1210     return Result;
1211 
1212   // We build somewhat synthetic new RefSCCs by providing a postorder mapping
1213   // for each inner SCC. We store these inside the low-link field of the nodes
1214   // rather than associated with SCCs because this saves a round-trip through
1215   // the node->SCC map and in the common case, SCCs are small. We will verify
1216   // that we always give the same number to every node in the SCC such that
1217   // these are equivalent.
1218   int PostOrderNumber = 0;
1219 
1220   // Reset all the other nodes to prepare for a DFS over them, and add them to
1221   // our worklist.
1222   SmallVector<Node *, 8> Worklist;
1223   for (SCC *C : SCCs) {
1224     for (Node &N : *C)
1225       N.DFSNumber = N.LowLink = 0;
1226 
1227     Worklist.append(C->Nodes.begin(), C->Nodes.end());
1228   }
1229 
1230   // Track the number of nodes in this RefSCC so that we can quickly recognize
1231   // an important special case of the edge removal not breaking the cycle of
1232   // this RefSCC.
1233   const int NumRefSCCNodes = Worklist.size();
1234 
1235   SmallVector<std::pair<Node *, EdgeSequence::iterator>, 4> DFSStack;
1236   SmallVector<Node *, 4> PendingRefSCCStack;
1237   do {
1238     assert(DFSStack.empty() &&
1239            "Cannot begin a new root with a non-empty DFS stack!");
1240     assert(PendingRefSCCStack.empty() &&
1241            "Cannot begin a new root with pending nodes for an SCC!");
1242 
1243     Node *RootN = Worklist.pop_back_val();
1244     // Skip any nodes we've already reached in the DFS.
1245     if (RootN->DFSNumber != 0) {
1246       assert(RootN->DFSNumber == -1 &&
1247              "Shouldn't have any mid-DFS root nodes!");
1248       continue;
1249     }
1250 
1251     RootN->DFSNumber = RootN->LowLink = 1;
1252     int NextDFSNumber = 2;
1253 
1254     DFSStack.push_back({RootN, (*RootN)->begin()});
1255     do {
1256       Node *N;
1257       EdgeSequence::iterator I;
1258       std::tie(N, I) = DFSStack.pop_back_val();
1259       auto E = (*N)->end();
1260 
1261       assert(N->DFSNumber != 0 && "We should always assign a DFS number "
1262                                   "before processing a node.");
1263 
1264       while (I != E) {
1265         Node &ChildN = I->getNode();
1266         if (ChildN.DFSNumber == 0) {
1267           // Mark that we should start at this child when next this node is the
1268           // top of the stack. We don't start at the next child to ensure this
1269           // child's lowlink is reflected.
1270           DFSStack.push_back({N, I});
1271 
1272           // Continue, resetting to the child node.
1273           ChildN.LowLink = ChildN.DFSNumber = NextDFSNumber++;
1274           N = &ChildN;
1275           I = ChildN->begin();
1276           E = ChildN->end();
1277           continue;
1278         }
1279         if (ChildN.DFSNumber == -1) {
1280           // If this child isn't currently in this RefSCC, no need to process
1281           // it.
1282           ++I;
1283           continue;
1284         }
1285 
1286         // Track the lowest link of the children, if any are still in the stack.
1287         // Any child not on the stack will have a LowLink of -1.
1288         assert(ChildN.LowLink != 0 &&
1289                "Low-link must not be zero with a non-zero DFS number.");
1290         if (ChildN.LowLink >= 0 && ChildN.LowLink < N->LowLink)
1291           N->LowLink = ChildN.LowLink;
1292         ++I;
1293       }
1294 
1295       // We've finished processing N and its descendants, put it on our pending
1296       // stack to eventually get merged into a RefSCC.
1297       PendingRefSCCStack.push_back(N);
1298 
1299       // If this node is linked to some lower entry, continue walking up the
1300       // stack.
1301       if (N->LowLink != N->DFSNumber) {
1302         assert(!DFSStack.empty() &&
1303                "We never found a viable root for a RefSCC to pop off!");
1304         continue;
1305       }
1306 
1307       // Otherwise, form a new RefSCC from the top of the pending node stack.
1308       int RefSCCNumber = PostOrderNumber++;
1309       int RootDFSNumber = N->DFSNumber;
1310 
1311       // Find the range of the node stack by walking down until we pass the
1312       // root DFS number. Update the DFS numbers and low link numbers in the
1313       // process to avoid re-walking this list where possible.
1314       auto StackRI = find_if(reverse(PendingRefSCCStack), [&](Node *N) {
1315         if (N->DFSNumber < RootDFSNumber)
1316           // We've found the bottom.
1317           return true;
1318 
1319         // Update this node and keep scanning.
1320         N->DFSNumber = -1;
1321         // Save the post-order number in the lowlink field so that we can use
1322         // it to map SCCs into new RefSCCs after we finish the DFS.
1323         N->LowLink = RefSCCNumber;
1324         return false;
1325       });
1326       auto RefSCCNodes = make_range(StackRI.base(), PendingRefSCCStack.end());
1327 
1328       // If we find a cycle containing all nodes originally in this RefSCC then
1329       // the removal hasn't changed the structure at all. This is an important
1330       // special case and we can directly exit the entire routine more
1331       // efficiently as soon as we discover it.
1332       if (llvm::size(RefSCCNodes) == NumRefSCCNodes) {
1333         // Clear out the low link field as we won't need it.
1334         for (Node *N : RefSCCNodes)
1335           N->LowLink = -1;
1336         // Return the empty result immediately.
1337         return Result;
1338       }
1339 
1340       // We've already marked the nodes internally with the RefSCC number so
1341       // just clear them off the stack and continue.
1342       PendingRefSCCStack.erase(RefSCCNodes.begin(), PendingRefSCCStack.end());
1343     } while (!DFSStack.empty());
1344 
1345     assert(DFSStack.empty() && "Didn't flush the entire DFS stack!");
1346     assert(PendingRefSCCStack.empty() && "Didn't flush all pending nodes!");
1347   } while (!Worklist.empty());
1348 
1349   assert(PostOrderNumber > 1 &&
1350          "Should never finish the DFS when the existing RefSCC remains valid!");
1351 
1352   // Otherwise we create a collection of new RefSCC nodes and build
1353   // a radix-sort style map from postorder number to these new RefSCCs. We then
1354   // append SCCs to each of these RefSCCs in the order they occurred in the
1355   // original SCCs container.
1356   for (int i = 0; i < PostOrderNumber; ++i)
1357     Result.push_back(G->createRefSCC(*G));
1358 
1359   // Insert the resulting postorder sequence into the global graph postorder
1360   // sequence before the current RefSCC in that sequence, and then remove the
1361   // current one.
1362   //
1363   // FIXME: It'd be nice to change the APIs so that we returned an iterator
1364   // range over the global postorder sequence and generally use that sequence
1365   // rather than building a separate result vector here.
1366   int Idx = G->getRefSCCIndex(*this);
1367   G->PostOrderRefSCCs.erase(G->PostOrderRefSCCs.begin() + Idx);
1368   G->PostOrderRefSCCs.insert(G->PostOrderRefSCCs.begin() + Idx, Result.begin(),
1369                              Result.end());
1370   for (int i : seq<int>(Idx, G->PostOrderRefSCCs.size()))
1371     G->RefSCCIndices[G->PostOrderRefSCCs[i]] = i;
1372 
1373   for (SCC *C : SCCs) {
1374     // We store the SCC number in the node's low-link field above.
1375     int SCCNumber = C->begin()->LowLink;
1376     // Clear out all of the SCC's node's low-link fields now that we're done
1377     // using them as side-storage.
1378     for (Node &N : *C) {
1379       assert(N.LowLink == SCCNumber &&
1380              "Cannot have different numbers for nodes in the same SCC!");
1381       N.LowLink = -1;
1382     }
1383 
1384     RefSCC &RC = *Result[SCCNumber];
1385     int SCCIndex = RC.SCCs.size();
1386     RC.SCCs.push_back(C);
1387     RC.SCCIndices[C] = SCCIndex;
1388     C->OuterRefSCC = &RC;
1389   }
1390 
1391   // Now that we've moved things into the new RefSCCs, clear out our current
1392   // one.
1393   G = nullptr;
1394   SCCs.clear();
1395   SCCIndices.clear();
1396 
1397 #ifdef EXPENSIVE_CHECKS
1398   // Verify the new RefSCCs we've built.
1399   for (RefSCC *RC : Result)
1400     RC->verify();
1401 #endif
1402 
1403   // Return the new list of SCCs.
1404   return Result;
1405 }
1406 
1407 void LazyCallGraph::RefSCC::insertTrivialCallEdge(Node &SourceN,
1408                                                   Node &TargetN) {
1409 #ifdef EXPENSIVE_CHECKS
1410   auto ExitVerifier = make_scope_exit([this] { verify(); });
1411 
1412   // Check that we aren't breaking some invariants of the SCC graph. Note that
1413   // this is quadratic in the number of edges in the call graph!
1414   SCC &SourceC = *G->lookupSCC(SourceN);
1415   SCC &TargetC = *G->lookupSCC(TargetN);
1416   if (&SourceC != &TargetC)
1417     assert(SourceC.isAncestorOf(TargetC) &&
1418            "Call edge is not trivial in the SCC graph!");
1419 #endif
1420 
1421   // First insert it into the source or find the existing edge.
1422   auto InsertResult =
1423       SourceN->EdgeIndexMap.insert({&TargetN, SourceN->Edges.size()});
1424   if (!InsertResult.second) {
1425     // Already an edge, just update it.
1426     Edge &E = SourceN->Edges[InsertResult.first->second];
1427     if (E.isCall())
1428       return; // Nothing to do!
1429     E.setKind(Edge::Call);
1430   } else {
1431     // Create the new edge.
1432     SourceN->Edges.emplace_back(TargetN, Edge::Call);
1433   }
1434 }
1435 
1436 void LazyCallGraph::RefSCC::insertTrivialRefEdge(Node &SourceN, Node &TargetN) {
1437 #ifdef EXPENSIVE_CHECKS
1438   auto ExitVerifier = make_scope_exit([this] { verify(); });
1439 
1440   // Check that we aren't breaking some invariants of the RefSCC graph.
1441   RefSCC &SourceRC = *G->lookupRefSCC(SourceN);
1442   RefSCC &TargetRC = *G->lookupRefSCC(TargetN);
1443   if (&SourceRC != &TargetRC)
1444     assert(SourceRC.isAncestorOf(TargetRC) &&
1445            "Ref edge is not trivial in the RefSCC graph!");
1446 #endif
1447 
1448   // First insert it into the source or find the existing edge.
1449   auto InsertResult =
1450       SourceN->EdgeIndexMap.insert({&TargetN, SourceN->Edges.size()});
1451   if (!InsertResult.second)
1452     // Already an edge, we're done.
1453     return;
1454 
1455   // Create the new edge.
1456   SourceN->Edges.emplace_back(TargetN, Edge::Ref);
1457 }
1458 
1459 void LazyCallGraph::RefSCC::replaceNodeFunction(Node &N, Function &NewF) {
1460   Function &OldF = N.getFunction();
1461 
1462 #ifdef EXPENSIVE_CHECKS
1463   auto ExitVerifier = make_scope_exit([this] { verify(); });
1464 
1465   assert(G->lookupRefSCC(N) == this &&
1466          "Cannot replace the function of a node outside this RefSCC.");
1467 
1468   assert(G->NodeMap.find(&NewF) == G->NodeMap.end() &&
1469          "Must not have already walked the new function!'");
1470 
1471   // It is important that this replacement not introduce graph changes so we
1472   // insist that the caller has already removed every use of the original
1473   // function and that all uses of the new function correspond to existing
1474   // edges in the graph. The common and expected way to use this is when
1475   // replacing the function itself in the IR without changing the call graph
1476   // shape and just updating the analysis based on that.
1477   assert(&OldF != &NewF && "Cannot replace a function with itself!");
1478   assert(OldF.use_empty() &&
1479          "Must have moved all uses from the old function to the new!");
1480 #endif
1481 
1482   N.replaceFunction(NewF);
1483 
1484   // Update various call graph maps.
1485   G->NodeMap.erase(&OldF);
1486   G->NodeMap[&NewF] = &N;
1487 }
1488 
1489 void LazyCallGraph::insertEdge(Node &SourceN, Node &TargetN, Edge::Kind EK) {
1490   assert(SCCMap.empty() &&
1491          "This method cannot be called after SCCs have been formed!");
1492 
1493   return SourceN->insertEdgeInternal(TargetN, EK);
1494 }
1495 
1496 void LazyCallGraph::removeEdge(Node &SourceN, Node &TargetN) {
1497   assert(SCCMap.empty() &&
1498          "This method cannot be called after SCCs have been formed!");
1499 
1500   bool Removed = SourceN->removeEdgeInternal(TargetN);
1501   (void)Removed;
1502   assert(Removed && "Target not in the edge set for this caller?");
1503 }
1504 
1505 void LazyCallGraph::removeDeadFunction(Function &F) {
1506   // FIXME: This is unnecessarily restrictive. We should be able to remove
1507   // functions which recursively call themselves.
1508   assert(F.hasZeroLiveUses() &&
1509          "This routine should only be called on trivially dead functions!");
1510 
1511   // We shouldn't remove library functions as they are never really dead while
1512   // the call graph is in use -- every function definition refers to them.
1513   assert(!isLibFunction(F) &&
1514          "Must not remove lib functions from the call graph!");
1515 
1516   auto NI = NodeMap.find(&F);
1517   if (NI == NodeMap.end())
1518     // Not in the graph at all!
1519     return;
1520 
1521   Node &N = *NI->second;
1522   NodeMap.erase(NI);
1523 
1524   // Remove this from the entry edges if present.
1525   EntryEdges.removeEdgeInternal(N);
1526 
1527   // Cannot remove a function which has yet to be visited in the DFS walk, so
1528   // if we have a node at all then we must have an SCC and RefSCC.
1529   auto CI = SCCMap.find(&N);
1530   assert(CI != SCCMap.end() &&
1531          "Tried to remove a node without an SCC after DFS walk started!");
1532   SCC &C = *CI->second;
1533   SCCMap.erase(CI);
1534   RefSCC &RC = C.getOuterRefSCC();
1535 
1536   // This node must be the only member of its SCC as it has no callers, and
1537   // that SCC must be the only member of a RefSCC as it has no references.
1538   // Validate these properties first.
1539   assert(C.size() == 1 && "Dead functions must be in a singular SCC");
1540   assert(RC.size() == 1 && "Dead functions must be in a singular RefSCC");
1541 
1542   // Finally clear out all the data structures from the node down through the
1543   // components. postorder_ref_scc_iterator will skip empty RefSCCs, so no need
1544   // to adjust LazyCallGraph data structures.
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 // Gets the Edge::Kind from one function to another by looking at the function's
1557 // instructions. Asserts if there is no edge.
1558 // Useful for determining what type of edge should exist between functions when
1559 // the edge hasn't been created yet.
1560 static LazyCallGraph::Edge::Kind getEdgeKind(Function &OriginalFunction,
1561                                              Function &NewFunction) {
1562   // In release builds, assume that if there are no direct calls to the new
1563   // function, then there is a ref edge. In debug builds, keep track of
1564   // references to assert that there is actually a ref edge if there is no call
1565   // edge.
1566 #ifndef NDEBUG
1567   SmallVector<Constant *, 16> Worklist;
1568   SmallPtrSet<Constant *, 16> Visited;
1569 #endif
1570 
1571   for (Instruction &I : instructions(OriginalFunction)) {
1572     if (auto *CB = dyn_cast<CallBase>(&I)) {
1573       if (Function *Callee = CB->getCalledFunction()) {
1574         if (Callee == &NewFunction)
1575           return LazyCallGraph::Edge::Kind::Call;
1576       }
1577     }
1578 #ifndef NDEBUG
1579     for (Value *Op : I.operand_values()) {
1580       if (Constant *C = dyn_cast<Constant>(Op)) {
1581         if (Visited.insert(C).second)
1582           Worklist.push_back(C);
1583       }
1584     }
1585 #endif
1586   }
1587 
1588 #ifndef NDEBUG
1589   bool FoundNewFunction = false;
1590   LazyCallGraph::visitReferences(Worklist, Visited, [&](Function &F) {
1591     if (&F == &NewFunction)
1592       FoundNewFunction = true;
1593   });
1594   assert(FoundNewFunction && "No edge from original function to new function");
1595 #endif
1596 
1597   return LazyCallGraph::Edge::Kind::Ref;
1598 }
1599 
1600 void LazyCallGraph::addSplitFunction(Function &OriginalFunction,
1601                                      Function &NewFunction) {
1602   assert(lookup(OriginalFunction) &&
1603          "Original function's node should already exist");
1604   Node &OriginalN = get(OriginalFunction);
1605   SCC *OriginalC = lookupSCC(OriginalN);
1606   RefSCC *OriginalRC = lookupRefSCC(OriginalN);
1607 
1608 #ifdef EXPENSIVE_CHECKS
1609   OriginalRC->verify();
1610   auto VerifyOnExit = make_scope_exit([&]() { OriginalRC->verify(); });
1611 #endif
1612 
1613   assert(!lookup(NewFunction) &&
1614          "New function's node should not already exist");
1615   Node &NewN = initNode(NewFunction);
1616 
1617   Edge::Kind EK = getEdgeKind(OriginalFunction, NewFunction);
1618 
1619   SCC *NewC = nullptr;
1620   for (Edge &E : *NewN) {
1621     Node &EN = E.getNode();
1622     if (EK == Edge::Kind::Call && E.isCall() && lookupSCC(EN) == OriginalC) {
1623       // If the edge to the new function is a call edge and there is a call edge
1624       // from the new function to any function in the original function's SCC,
1625       // it is in the same SCC (and RefSCC) as the original function.
1626       NewC = OriginalC;
1627       NewC->Nodes.push_back(&NewN);
1628       break;
1629     }
1630   }
1631 
1632   if (!NewC) {
1633     for (Edge &E : *NewN) {
1634       Node &EN = E.getNode();
1635       if (lookupRefSCC(EN) == OriginalRC) {
1636         // If there is any edge from the new function to any function in the
1637         // original function's RefSCC, it is in the same RefSCC as the original
1638         // function but a new SCC.
1639         RefSCC *NewRC = OriginalRC;
1640         NewC = createSCC(*NewRC, SmallVector<Node *, 1>({&NewN}));
1641 
1642         // The new function's SCC is not the same as the original function's
1643         // SCC, since that case was handled earlier. If the edge from the
1644         // original function to the new function was a call edge, then we need
1645         // to insert the newly created function's SCC before the original
1646         // function's SCC. Otherwise either the new SCC comes after the original
1647         // function's SCC, or it doesn't matter, and in both cases we can add it
1648         // to the very end.
1649         int InsertIndex = EK == Edge::Kind::Call ? NewRC->SCCIndices[OriginalC]
1650                                                  : NewRC->SCCIndices.size();
1651         NewRC->SCCs.insert(NewRC->SCCs.begin() + InsertIndex, NewC);
1652         for (int I = InsertIndex, Size = NewRC->SCCs.size(); I < Size; ++I)
1653           NewRC->SCCIndices[NewRC->SCCs[I]] = I;
1654 
1655         break;
1656       }
1657     }
1658   }
1659 
1660   if (!NewC) {
1661     // We didn't find any edges back to the original function's RefSCC, so the
1662     // new function belongs in a new RefSCC. The new RefSCC goes before the
1663     // original function's RefSCC.
1664     RefSCC *NewRC = createRefSCC(*this);
1665     NewC = createSCC(*NewRC, SmallVector<Node *, 1>({&NewN}));
1666     NewRC->SCCIndices[NewC] = 0;
1667     NewRC->SCCs.push_back(NewC);
1668     auto OriginalRCIndex = RefSCCIndices.find(OriginalRC)->second;
1669     PostOrderRefSCCs.insert(PostOrderRefSCCs.begin() + OriginalRCIndex, NewRC);
1670     for (int I = OriginalRCIndex, Size = PostOrderRefSCCs.size(); I < Size; ++I)
1671       RefSCCIndices[PostOrderRefSCCs[I]] = I;
1672   }
1673 
1674   SCCMap[&NewN] = NewC;
1675 
1676   OriginalN->insertEdgeInternal(NewN, EK);
1677 }
1678 
1679 void LazyCallGraph::addSplitRefRecursiveFunctions(
1680     Function &OriginalFunction, ArrayRef<Function *> NewFunctions) {
1681   assert(!NewFunctions.empty() && "Can't add zero functions");
1682   assert(lookup(OriginalFunction) &&
1683          "Original function's node should already exist");
1684   Node &OriginalN = get(OriginalFunction);
1685   RefSCC *OriginalRC = lookupRefSCC(OriginalN);
1686 
1687 #ifdef EXPENSIVE_CHECKS
1688   OriginalRC->verify();
1689   auto VerifyOnExit = make_scope_exit([&]() {
1690     OriginalRC->verify();
1691     for (Function *NewFunction : NewFunctions)
1692       lookupRefSCC(get(*NewFunction))->verify();
1693   });
1694 #endif
1695 
1696   bool ExistsRefToOriginalRefSCC = false;
1697 
1698   for (Function *NewFunction : NewFunctions) {
1699     Node &NewN = initNode(*NewFunction);
1700 
1701     OriginalN->insertEdgeInternal(NewN, Edge::Kind::Ref);
1702 
1703     // Check if there is any edge from any new function back to any function in
1704     // the original function's RefSCC.
1705     for (Edge &E : *NewN) {
1706       if (lookupRefSCC(E.getNode()) == OriginalRC) {
1707         ExistsRefToOriginalRefSCC = true;
1708         break;
1709       }
1710     }
1711   }
1712 
1713   RefSCC *NewRC;
1714   if (ExistsRefToOriginalRefSCC) {
1715     // If there is any edge from any new function to any function in the
1716     // original function's RefSCC, all new functions will be in the same RefSCC
1717     // as the original function.
1718     NewRC = OriginalRC;
1719   } else {
1720     // Otherwise the new functions are in their own RefSCC.
1721     NewRC = createRefSCC(*this);
1722     // The new RefSCC goes before the original function's RefSCC in postorder
1723     // since there are only edges from the original function's RefSCC to the new
1724     // RefSCC.
1725     auto OriginalRCIndex = RefSCCIndices.find(OriginalRC)->second;
1726     PostOrderRefSCCs.insert(PostOrderRefSCCs.begin() + OriginalRCIndex, NewRC);
1727     for (int I = OriginalRCIndex, Size = PostOrderRefSCCs.size(); I < Size; ++I)
1728       RefSCCIndices[PostOrderRefSCCs[I]] = I;
1729   }
1730 
1731   for (Function *NewFunction : NewFunctions) {
1732     Node &NewN = get(*NewFunction);
1733     // Each new function is in its own new SCC. The original function can only
1734     // have a ref edge to new functions, and no other existing functions can
1735     // have references to new functions. Each new function only has a ref edge
1736     // to the other new functions.
1737     SCC *NewC = createSCC(*NewRC, SmallVector<Node *, 1>({&NewN}));
1738     // The new SCCs are either sibling SCCs or parent SCCs to all other existing
1739     // SCCs in the RefSCC. Either way, they can go at the back of the postorder
1740     // SCC list.
1741     auto Index = NewRC->SCCIndices.size();
1742     NewRC->SCCIndices[NewC] = Index;
1743     NewRC->SCCs.push_back(NewC);
1744     SCCMap[&NewN] = NewC;
1745   }
1746 
1747 #ifndef NDEBUG
1748   for (Function *F1 : NewFunctions) {
1749     assert(getEdgeKind(OriginalFunction, *F1) == Edge::Kind::Ref &&
1750            "Expected ref edges from original function to every new function");
1751     Node &N1 = get(*F1);
1752     for (Function *F2 : NewFunctions) {
1753       if (F1 == F2)
1754         continue;
1755       Node &N2 = get(*F2);
1756       assert(!N1->lookup(N2)->isCall() &&
1757              "Edges between new functions must be ref edges");
1758     }
1759   }
1760 #endif
1761 }
1762 
1763 LazyCallGraph::Node &LazyCallGraph::insertInto(Function &F, Node *&MappedN) {
1764   return *new (MappedN = BPA.Allocate()) Node(*this, F);
1765 }
1766 
1767 void LazyCallGraph::updateGraphPtrs() {
1768   // Walk the node map to update their graph pointers. While this iterates in
1769   // an unstable order, the order has no effect so it remains correct.
1770   for (auto &FunctionNodePair : NodeMap)
1771     FunctionNodePair.second->G = this;
1772 
1773   for (auto *RC : PostOrderRefSCCs)
1774     RC->G = this;
1775 }
1776 
1777 LazyCallGraph::Node &LazyCallGraph::initNode(Function &F) {
1778   Node &N = get(F);
1779   N.DFSNumber = N.LowLink = -1;
1780   N.populate();
1781   NodeMap[&F] = &N;
1782   return N;
1783 }
1784 
1785 template <typename RootsT, typename GetBeginT, typename GetEndT,
1786           typename GetNodeT, typename FormSCCCallbackT>
1787 void LazyCallGraph::buildGenericSCCs(RootsT &&Roots, GetBeginT &&GetBegin,
1788                                      GetEndT &&GetEnd, GetNodeT &&GetNode,
1789                                      FormSCCCallbackT &&FormSCC) {
1790   using EdgeItT = decltype(GetBegin(std::declval<Node &>()));
1791 
1792   SmallVector<std::pair<Node *, EdgeItT>, 16> DFSStack;
1793   SmallVector<Node *, 16> PendingSCCStack;
1794 
1795   // Scan down the stack and DFS across the call edges.
1796   for (Node *RootN : Roots) {
1797     assert(DFSStack.empty() &&
1798            "Cannot begin a new root with a non-empty DFS stack!");
1799     assert(PendingSCCStack.empty() &&
1800            "Cannot begin a new root with pending nodes for an SCC!");
1801 
1802     // Skip any nodes we've already reached in the DFS.
1803     if (RootN->DFSNumber != 0) {
1804       assert(RootN->DFSNumber == -1 &&
1805              "Shouldn't have any mid-DFS root nodes!");
1806       continue;
1807     }
1808 
1809     RootN->DFSNumber = RootN->LowLink = 1;
1810     int NextDFSNumber = 2;
1811 
1812     DFSStack.push_back({RootN, GetBegin(*RootN)});
1813     do {
1814       Node *N;
1815       EdgeItT I;
1816       std::tie(N, I) = DFSStack.pop_back_val();
1817       auto E = GetEnd(*N);
1818       while (I != E) {
1819         Node &ChildN = GetNode(I);
1820         if (ChildN.DFSNumber == 0) {
1821           // We haven't yet visited this child, so descend, pushing the current
1822           // node onto the stack.
1823           DFSStack.push_back({N, I});
1824 
1825           ChildN.DFSNumber = ChildN.LowLink = NextDFSNumber++;
1826           N = &ChildN;
1827           I = GetBegin(*N);
1828           E = GetEnd(*N);
1829           continue;
1830         }
1831 
1832         // If the child has already been added to some child component, it
1833         // couldn't impact the low-link of this parent because it isn't
1834         // connected, and thus its low-link isn't relevant so skip it.
1835         if (ChildN.DFSNumber == -1) {
1836           ++I;
1837           continue;
1838         }
1839 
1840         // Track the lowest linked child as the lowest link for this node.
1841         assert(ChildN.LowLink > 0 && "Must have a positive low-link number!");
1842         if (ChildN.LowLink < N->LowLink)
1843           N->LowLink = ChildN.LowLink;
1844 
1845         // Move to the next edge.
1846         ++I;
1847       }
1848 
1849       // We've finished processing N and its descendants, put it on our pending
1850       // SCC stack to eventually get merged into an SCC of nodes.
1851       PendingSCCStack.push_back(N);
1852 
1853       // If this node is linked to some lower entry, continue walking up the
1854       // stack.
1855       if (N->LowLink != N->DFSNumber)
1856         continue;
1857 
1858       // Otherwise, we've completed an SCC. Append it to our post order list of
1859       // SCCs.
1860       int RootDFSNumber = N->DFSNumber;
1861       // Find the range of the node stack by walking down until we pass the
1862       // root DFS number.
1863       auto SCCNodes = make_range(
1864           PendingSCCStack.rbegin(),
1865           find_if(reverse(PendingSCCStack), [RootDFSNumber](const Node *N) {
1866             return N->DFSNumber < RootDFSNumber;
1867           }));
1868       // Form a new SCC out of these nodes and then clear them off our pending
1869       // stack.
1870       FormSCC(SCCNodes);
1871       PendingSCCStack.erase(SCCNodes.end().base(), PendingSCCStack.end());
1872     } while (!DFSStack.empty());
1873   }
1874 }
1875 
1876 /// Build the internal SCCs for a RefSCC from a sequence of nodes.
1877 ///
1878 /// Appends the SCCs to the provided vector and updates the map with their
1879 /// indices. Both the vector and map must be empty when passed into this
1880 /// routine.
1881 void LazyCallGraph::buildSCCs(RefSCC &RC, node_stack_range Nodes) {
1882   assert(RC.SCCs.empty() && "Already built SCCs!");
1883   assert(RC.SCCIndices.empty() && "Already mapped SCC indices!");
1884 
1885   for (Node *N : Nodes) {
1886     assert(N->LowLink >= (*Nodes.begin())->LowLink &&
1887            "We cannot have a low link in an SCC lower than its root on the "
1888            "stack!");
1889 
1890     // This node will go into the next RefSCC, clear out its DFS and low link
1891     // as we scan.
1892     N->DFSNumber = N->LowLink = 0;
1893   }
1894 
1895   // Each RefSCC contains a DAG of the call SCCs. To build these, we do
1896   // a direct walk of the call edges using Tarjan's algorithm. We reuse the
1897   // internal storage as we won't need it for the outer graph's DFS any longer.
1898   buildGenericSCCs(
1899       Nodes, [](Node &N) { return N->call_begin(); },
1900       [](Node &N) { return N->call_end(); },
1901       [](EdgeSequence::call_iterator I) -> Node & { return I->getNode(); },
1902       [this, &RC](node_stack_range Nodes) {
1903         RC.SCCs.push_back(createSCC(RC, Nodes));
1904         for (Node &N : *RC.SCCs.back()) {
1905           N.DFSNumber = N.LowLink = -1;
1906           SCCMap[&N] = RC.SCCs.back();
1907         }
1908       });
1909 
1910   // Wire up the SCC indices.
1911   for (int i = 0, Size = RC.SCCs.size(); i < Size; ++i)
1912     RC.SCCIndices[RC.SCCs[i]] = i;
1913 }
1914 
1915 void LazyCallGraph::buildRefSCCs() {
1916   if (EntryEdges.empty() || !PostOrderRefSCCs.empty())
1917     // RefSCCs are either non-existent or already built!
1918     return;
1919 
1920   assert(RefSCCIndices.empty() && "Already mapped RefSCC indices!");
1921 
1922   SmallVector<Node *, 16> Roots;
1923   for (Edge &E : *this)
1924     Roots.push_back(&E.getNode());
1925 
1926   // The roots will be iterated in order.
1927   buildGenericSCCs(
1928       Roots,
1929       [](Node &N) {
1930         // We need to populate each node as we begin to walk its edges.
1931         N.populate();
1932         return N->begin();
1933       },
1934       [](Node &N) { return N->end(); },
1935       [](EdgeSequence::iterator I) -> Node & { return I->getNode(); },
1936       [this](node_stack_range Nodes) {
1937         RefSCC *NewRC = createRefSCC(*this);
1938         buildSCCs(*NewRC, Nodes);
1939 
1940         // Push the new node into the postorder list and remember its position
1941         // in the index map.
1942         bool Inserted =
1943             RefSCCIndices.insert({NewRC, PostOrderRefSCCs.size()}).second;
1944         (void)Inserted;
1945         assert(Inserted && "Cannot already have this RefSCC in the index map!");
1946         PostOrderRefSCCs.push_back(NewRC);
1947 #ifdef EXPENSIVE_CHECKS
1948         NewRC->verify();
1949 #endif
1950       });
1951 }
1952 
1953 void LazyCallGraph::visitReferences(SmallVectorImpl<Constant *> &Worklist,
1954                                     SmallPtrSetImpl<Constant *> &Visited,
1955                                     function_ref<void(Function &)> Callback) {
1956   while (!Worklist.empty()) {
1957     Constant *C = Worklist.pop_back_val();
1958 
1959     if (Function *F = dyn_cast<Function>(C)) {
1960       if (!F->isDeclaration())
1961         Callback(*F);
1962       continue;
1963     }
1964 
1965     // blockaddresses are weird and don't participate in the call graph anyway,
1966     // skip them.
1967     if (isa<BlockAddress>(C))
1968       continue;
1969 
1970     for (Value *Op : C->operand_values())
1971       if (Visited.insert(cast<Constant>(Op)).second)
1972         Worklist.push_back(cast<Constant>(Op));
1973   }
1974 }
1975 
1976 AnalysisKey LazyCallGraphAnalysis::Key;
1977 
1978 LazyCallGraphPrinterPass::LazyCallGraphPrinterPass(raw_ostream &OS) : OS(OS) {}
1979 
1980 static void printNode(raw_ostream &OS, LazyCallGraph::Node &N) {
1981   OS << "  Edges in function: " << N.getFunction().getName() << "\n";
1982   for (LazyCallGraph::Edge &E : N.populate())
1983     OS << "    " << (E.isCall() ? "call" : "ref ") << " -> "
1984        << E.getFunction().getName() << "\n";
1985 
1986   OS << "\n";
1987 }
1988 
1989 static void printSCC(raw_ostream &OS, LazyCallGraph::SCC &C) {
1990   OS << "    SCC with " << C.size() << " functions:\n";
1991 
1992   for (LazyCallGraph::Node &N : C)
1993     OS << "      " << N.getFunction().getName() << "\n";
1994 }
1995 
1996 static void printRefSCC(raw_ostream &OS, LazyCallGraph::RefSCC &C) {
1997   OS << "  RefSCC with " << C.size() << " call SCCs:\n";
1998 
1999   for (LazyCallGraph::SCC &InnerC : C)
2000     printSCC(OS, InnerC);
2001 
2002   OS << "\n";
2003 }
2004 
2005 PreservedAnalyses LazyCallGraphPrinterPass::run(Module &M,
2006                                                 ModuleAnalysisManager &AM) {
2007   LazyCallGraph &G = AM.getResult<LazyCallGraphAnalysis>(M);
2008 
2009   OS << "Printing the call graph for module: " << M.getModuleIdentifier()
2010      << "\n\n";
2011 
2012   for (Function &F : M)
2013     printNode(OS, G.get(F));
2014 
2015   G.buildRefSCCs();
2016   for (LazyCallGraph::RefSCC &C : G.postorder_ref_sccs())
2017     printRefSCC(OS, C);
2018 
2019   return PreservedAnalyses::all();
2020 }
2021 
2022 LazyCallGraphDOTPrinterPass::LazyCallGraphDOTPrinterPass(raw_ostream &OS)
2023     : OS(OS) {}
2024 
2025 static void printNodeDOT(raw_ostream &OS, LazyCallGraph::Node &N) {
2026   std::string Name =
2027       "\"" + DOT::EscapeString(std::string(N.getFunction().getName())) + "\"";
2028 
2029   for (LazyCallGraph::Edge &E : N.populate()) {
2030     OS << "  " << Name << " -> \""
2031        << DOT::EscapeString(std::string(E.getFunction().getName())) << "\"";
2032     if (!E.isCall()) // It is a ref edge.
2033       OS << " [style=dashed,label=\"ref\"]";
2034     OS << ";\n";
2035   }
2036 
2037   OS << "\n";
2038 }
2039 
2040 PreservedAnalyses LazyCallGraphDOTPrinterPass::run(Module &M,
2041                                                    ModuleAnalysisManager &AM) {
2042   LazyCallGraph &G = AM.getResult<LazyCallGraphAnalysis>(M);
2043 
2044   OS << "digraph \"" << DOT::EscapeString(M.getModuleIdentifier()) << "\" {\n";
2045 
2046   for (Function &F : M)
2047     printNodeDOT(OS, G.get(F));
2048 
2049   OS << "}\n";
2050 
2051   return PreservedAnalyses::all();
2052 }
2053