1 //===-- Support/TarjanSCCIterator.h - Tarjan SCC iterator -------*- C++ -*-===//
3 // This builds on the Support/GraphTraits.h file to find the strongly
4 // connected components (SCCs) of a graph in O(N+E) time using
5 // Tarjan's DFS algorithm.
7 // The SCC iterator has the important property that if a node in SCC S1
8 // has an edge to a node in SCC S2, then it visits S1 *after* S2.
10 // To visit S1 *before* S2, use the TarjanSCCIterator on the Inverse graph.
11 // (NOTE: This requires some simple wrappers and is not supported yet.)
13 //===----------------------------------------------------------------------===//
15 #ifndef SUPPORT_TARJANSCCITERATOR_H
16 #define SUPPORT_TARJANSCCITERATOR_H
18 #include "Support/GraphTraits.h"
19 #include "Support/Statistic.h"
20 #include "Support/iterator"
26 //--------------------------------------------------------------------------
27 // class SCC : A simple representation of an SCC in a generic Graph.
28 //--------------------------------------------------------------------------
30 template<class GraphT, class GT = GraphTraits<GraphT> >
31 struct SCC: public std::vector<typename GT::NodeType*> {
33 typedef typename GT::NodeType NodeType;
34 typedef typename GT::ChildIteratorType ChildItTy;
36 typedef std::vector<typename GT::NodeType*> super;
37 typedef typename super::iterator iterator;
38 typedef typename super::const_iterator const_iterator;
39 typedef typename super::reverse_iterator reverse_iterator;
40 typedef typename super::const_reverse_iterator const_reverse_iterator;
42 // HasLoop() -- Test if this SCC has a loop. If it has more than one
43 // node, this is trivially true. If not, it may still contain a loop
44 // if the node has an edge back to itself.
45 bool HasLoop() const {
46 if (size() > 1) return true;
47 NodeType* N = front();
48 for (ChildItTy CI=GT::child_begin(N), CE=GT::child_end(N); CI != CE; ++CI)
55 //--------------------------------------------------------------------------
56 // class TarjanSCC_iterator: Enumerate the SCCs of a directed graph, in
57 // reverse topological order of the SCC DAG.
58 //--------------------------------------------------------------------------
61 Statistic<> NumSCCs("NumSCCs", "Number of Strongly Connected Components");
62 Statistic<> MaxSCCSize("MaxSCCSize", "Size of largest Strongly Connected Component");
65 template<class GraphT, class GT = GraphTraits<GraphT> >
66 class TarjanSCC_iterator : public forward_iterator<SCC<GraphT, GT>, ptrdiff_t>
68 typedef SCC<GraphT, GT> SccTy;
69 typedef forward_iterator<SccTy, ptrdiff_t> super;
70 typedef typename super::reference reference;
71 typedef typename super::pointer pointer;
72 typedef typename GT::NodeType NodeType;
73 typedef typename GT::ChildIteratorType ChildItTy;
75 // The visit counters used to detect when a complete SCC is on the stack.
76 // visitNum is the global counter.
77 // nodeVisitNumbers are per-node visit numbers, also used as DFS flags.
78 unsigned long visitNum;
79 std::map<NodeType *, unsigned long> nodeVisitNumbers;
81 // SCCNodeStack - Stack holding nodes of the SCC.
82 std::stack<NodeType *> SCCNodeStack;
84 // CurrentSCC - The current SCC, retrieved using operator*().
87 // VisitStack - Used to maintain the ordering. Top = current block
88 // First element is basic block pointer, second is the 'next child' to visit
89 std::stack<std::pair<NodeType *, ChildItTy> > VisitStack;
91 // MinVistNumStack - Stack holding the "min" values for each node in the DFS.
92 // This is used to track the minimum uplink values for all children of
93 // the corresponding node on the VisitStack.
94 std::stack<unsigned long> MinVisitNumStack;
96 // A single "visit" within the non-recursive DFS traversal.
97 void DFSVisitOne(NodeType* N) {
98 ++visitNum; // Global counter for the visit order
99 nodeVisitNumbers[N] = visitNum;
100 SCCNodeStack.push(N);
101 MinVisitNumStack.push(visitNum);
102 VisitStack.push(make_pair(N, GT::child_begin(N)));
103 DEBUG(std::cerr << "TarjanSCC: Node " << N <<
104 " : visitNum = " << visitNum << "\n");
107 // The stack-based DFS traversal; defined below.
108 void DFSVisitChildren() {
109 assert(!VisitStack.empty());
110 while (VisitStack.top().second != GT::child_end(VisitStack.top().first))
111 { // TOS has at least one more child so continue DFS
112 NodeType *childN = *VisitStack.top().second++;
113 if (nodeVisitNumbers.find(childN) == nodeVisitNumbers.end())
114 { // this node has never been seen
119 unsigned long childNum = nodeVisitNumbers[childN];
120 if (MinVisitNumStack.top() > childNum)
121 MinVisitNumStack.top() = childNum;
126 // Compute the next SCC using the DFS traversal.
128 assert(VisitStack.size() == MinVisitNumStack.size());
129 CurrentSCC.clear(); // Prepare to compute the next SCC
130 while (! VisitStack.empty())
134 assert(VisitStack.top().second==GT::child_end(VisitStack.top().first));
135 NodeType* visitingN = VisitStack.top().first;
136 unsigned long minVisitNum = MinVisitNumStack.top();
138 MinVisitNumStack.pop();
139 if (! MinVisitNumStack.empty() && MinVisitNumStack.top() > minVisitNum)
140 MinVisitNumStack.top() = minVisitNum;
142 DEBUG(std::cerr << "TarjanSCC: Popped node " << visitingN <<
143 " : minVisitNum = " << minVisitNum << "; Node visit num = " <<
144 nodeVisitNumbers[visitingN] << "\n");
146 if (minVisitNum == nodeVisitNumbers[visitingN])
147 { // A full SCC is on the SCCNodeStack! It includes all nodes below
148 // visitingN on the stack. Copy those nodes to CurrentSCC,
149 // reset their minVisit values, and return (this suspends
150 // the DFS traversal till the next ++).
152 CurrentSCC.push_back(SCCNodeStack.top());
154 nodeVisitNumbers[CurrentSCC.back()] = ~0UL;
155 } while (CurrentSCC.back() != visitingN);
158 if (CurrentSCC.size() > MaxSCCSize) MaxSCCSize = CurrentSCC.size();
165 inline TarjanSCC_iterator(NodeType *entryN) : visitNum(0) {
169 inline TarjanSCC_iterator() { /* End is when DFS stack is empty */ }
172 typedef TarjanSCC_iterator<GraphT, GT> _Self;
174 // Provide static "constructors"...
175 static inline _Self begin(GraphT& G) { return _Self(GT::getEntryNode(G)); }
176 static inline _Self end (GraphT& G) { return _Self(); }
178 // Direct loop termination test (I.fini() is more efficient than I == end())
179 inline bool fini() const {
180 assert(!CurrentSCC.empty() || VisitStack.empty());
181 return CurrentSCC.empty();
184 inline bool operator==(const _Self& x) const {
185 return VisitStack == x.VisitStack && CurrentSCC == x.CurrentSCC;
187 inline bool operator!=(const _Self& x) const { return !operator==(x); }
189 // Iterator traversal: forward iteration only
190 inline _Self& operator++() { // Preincrement
194 inline _Self operator++(int) { // Postincrement
195 _Self tmp = *this; ++*this; return tmp;
198 // Retrieve a pointer to the current SCC. Returns NULL when done.
199 inline const SccTy* operator*() const {
200 assert(!CurrentSCC.empty() || VisitStack.empty());
201 return CurrentSCC.empty()? NULL : &CurrentSCC;
203 inline SccTy* operator*() {
204 assert(!CurrentSCC.empty() || VisitStack.empty());
205 return CurrentSCC.empty()? NULL : &CurrentSCC;
210 // Global constructor for the Tarjan SCC iterator. Use *I == NULL or I.fini()
211 // to test termination efficiently, instead of I == the "end" iterator.
213 TarjanSCC_iterator<T> tarj_begin(T G)
215 return TarjanSCC_iterator<T>::begin(G);
219 TarjanSCC_iterator<T> tarj_end(T G)
221 return TarjanSCC_iterator<T>::end(G);
224 //===----------------------------------------------------------------------===//