1 //===--- llvm/ADT/SparseSet.h - Sparse set ----------------------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the SparseSet class derived from the version described in
11 // Briggs, Torczon, "An efficient representation for sparse sets", ACM Letters
12 // on Programming Languages and Systems, Volume 2 Issue 1-4, March-Dec. 1993.
14 // A sparse set holds a small number of objects identified by integer keys from
15 // a moderately sized universe. The sparse set uses more memory than other
16 // containers in order to provide faster operations.
18 //===----------------------------------------------------------------------===//
20 #ifndef LLVM_ADT_SPARSESET_H
21 #define LLVM_ADT_SPARSESET_H
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/Support/DataTypes.h"
30 /// SparseSetValTraits - Objects in a SparseSet are identified by keys that can
31 /// be uniquely converted to a small integer less than the set's universe. This
32 /// class allows the set to hold values that differ from the set's key type as
33 /// long as an index can still be derived from the value. SparseSet never
34 /// directly compares ValueT, only their indices, so it can map keys to
35 /// arbitrary values. SparseSetValTraits computes the index from the value
36 /// object. To compute the index from a key, SparseSet uses a separate
37 /// KeyFunctorT template argument.
39 /// A simple type declaration, SparseSet<Type>, handles these cases:
40 /// - unsigned key, identity index, identity value
41 /// - unsigned key, identity index, fat value providing getSparseSetIndex()
43 /// The type declaration SparseSet<Type, UnaryFunction> handles:
44 /// - unsigned key, remapped index, identity value (virtual registers)
45 /// - pointer key, pointer-derived index, identity value (node+ID)
46 /// - pointer key, pointer-derived index, fat value with getSparseSetIndex()
48 /// Only other, unexpected cases require specializing SparseSetValTraits.
50 /// For best results, ValueT should not require a destructor.
52 template<typename ValueT>
53 struct SparseSetValTraits {
54 static unsigned getValIndex(const ValueT &Val) {
55 return Val.getSparseSetIndex();
59 /// SparseSetValFunctor - Helper class for selecting SparseSetValTraits. The
60 /// generic implementation handles ValueT classes which either provide
61 /// getSparseSetIndex() or specialize SparseSetValTraits<>.
63 template<typename KeyT, typename ValueT, typename KeyFunctorT>
64 struct SparseSetValFunctor {
65 unsigned operator()(const ValueT &Val) const {
66 return SparseSetValTraits<ValueT>::getValIndex(Val);
70 /// SparseSetValFunctor<KeyT, KeyT> - Helper class for the common case of
71 /// identity key/value sets.
72 template<typename KeyT, typename KeyFunctorT>
73 struct SparseSetValFunctor<KeyT, KeyT, KeyFunctorT> {
74 unsigned operator()(const KeyT &Key) const {
75 return KeyFunctorT()(Key);
79 /// SparseSet - Fast set implmentation for objects that can be identified by
80 /// small unsigned keys.
82 /// SparseSet allocates memory proportional to the size of the key universe, so
83 /// it is not recommended for building composite data structures. It is useful
84 /// for algorithms that require a single set with fast operations.
86 /// Compared to DenseSet and DenseMap, SparseSet provides constant-time fast
87 /// clear() and iteration as fast as a vector. The find(), insert(), and
88 /// erase() operations are all constant time, and typically faster than a hash
89 /// table. The iteration order doesn't depend on numerical key values, it only
90 /// depends on the order of insert() and erase() operations. When no elements
91 /// have been erased, the iteration order is the insertion order.
93 /// Compared to BitVector, SparseSet<unsigned> uses 8x-40x more memory, but
94 /// offers constant-time clear() and size() operations as well as fast
95 /// iteration independent on the size of the universe.
97 /// SparseSet contains a dense vector holding all the objects and a sparse
98 /// array holding indexes into the dense vector. Most of the memory is used by
99 /// the sparse array which is the size of the key universe. The SparseT
100 /// template parameter provides a space/speed tradeoff for sets holding many
103 /// When SparseT is uint32_t, find() only touches 2 cache lines, but the sparse
104 /// array uses 4 x Universe bytes.
106 /// When SparseT is uint8_t (the default), find() touches up to 2+[N/256] cache
107 /// lines, but the sparse array is 4x smaller. N is the number of elements in
110 /// For sets that may grow to thousands of elements, SparseT should be set to
111 /// uint16_t or uint32_t.
113 /// @param ValueT The type of objects in the set.
114 /// @param KeyFunctorT A functor that computes an unsigned index from KeyT.
115 /// @param SparseT An unsigned integer type. See above.
117 template<typename ValueT,
118 typename KeyFunctorT = llvm::identity<unsigned>,
119 typename SparseT = uint8_t>
121 typedef typename KeyFunctorT::argument_type KeyT;
122 typedef SmallVector<ValueT, 8> DenseT;
126 KeyFunctorT KeyIndexOf;
127 SparseSetValFunctor<KeyT, ValueT, KeyFunctorT> ValIndexOf;
129 // Disable copy construction and assignment.
130 // This data structure is not meant to be used that way.
131 SparseSet(const SparseSet&); // DO NOT IMPLEMENT.
132 SparseSet &operator=(const SparseSet&); // DO NOT IMPLEMENT.
135 typedef ValueT value_type;
136 typedef ValueT &reference;
137 typedef const ValueT &const_reference;
138 typedef ValueT *pointer;
139 typedef const ValueT *const_pointer;
141 SparseSet() : Sparse(0), Universe(0) {}
142 ~SparseSet() { free(Sparse); }
144 /// setUniverse - Set the universe size which determines the largest key the
145 /// set can hold. The universe must be sized before any elements can be
148 /// @param U Universe size. All object keys must be less than U.
150 void setUniverse(unsigned U) {
151 // It's not hard to resize the universe on a non-empty set, but it doesn't
152 // seem like a likely use case, so we can add that code when we need it.
153 assert(empty() && "Can only resize universe on an empty map");
154 // Hysteresis prevents needless reallocations.
155 if (U >= Universe/4 && U <= Universe)
158 // The Sparse array doesn't actually need to be initialized, so malloc
159 // would be enough here, but that will cause tools like valgrind to
160 // complain about branching on uninitialized data.
161 Sparse = reinterpret_cast<SparseT*>(calloc(U, sizeof(SparseT)));
165 // Import trivial vector stuff from DenseT.
166 typedef typename DenseT::iterator iterator;
167 typedef typename DenseT::const_iterator const_iterator;
169 const_iterator begin() const { return Dense.begin(); }
170 const_iterator end() const { return Dense.end(); }
171 iterator begin() { return Dense.begin(); }
172 iterator end() { return Dense.end(); }
174 /// empty - Returns true if the set is empty.
176 /// This is not the same as BitVector::empty().
178 bool empty() const { return Dense.empty(); }
180 /// size - Returns the number of elements in the set.
182 /// This is not the same as BitVector::size() which returns the size of the
185 unsigned size() const { return Dense.size(); }
187 /// clear - Clears the set. This is a very fast constant time operation.
190 // Sparse does not need to be cleared, see find().
194 /// findIndex - Find an element by its index.
196 /// @param Idx A valid index to find.
197 /// @returns An iterator to the element identified by key, or end().
199 iterator findIndex(unsigned Idx) {
200 assert(Idx < Universe && "Key out of range");
201 assert(std::numeric_limits<SparseT>::is_integer &&
202 !std::numeric_limits<SparseT>::is_signed &&
203 "SparseT must be an unsigned integer type");
204 const unsigned Stride = std::numeric_limits<SparseT>::max() + 1u;
205 for (unsigned i = Sparse[Idx], e = size(); i < e; i += Stride) {
206 const unsigned FoundIdx = ValIndexOf(Dense[i]);
207 assert(FoundIdx < Universe && "Invalid key in set. Did object mutate?");
210 // Stride is 0 when SparseT >= unsigned. We don't need to loop.
217 /// find - Find an element by its key.
219 /// @param Key A valid key to find.
220 /// @returns An iterator to the element identified by key, or end().
222 iterator find(const KeyT &Key) {
223 return findIndex(KeyIndexOf(Key));
226 const_iterator find(const KeyT &Key) const {
227 return const_cast<SparseSet*>(this)->findIndex(KeyIndexOf(Key));
230 /// count - Returns true if this set contains an element identified by Key.
232 bool count(const KeyT &Key) const {
233 return find(Key) != end();
236 /// insert - Attempts to insert a new element.
238 /// If Val is successfully inserted, return (I, true), where I is an iterator
239 /// pointing to the newly inserted element.
241 /// If the set already contains an element with the same key as Val, return
242 /// (I, false), where I is an iterator pointing to the existing element.
244 /// Insertion invalidates all iterators.
246 std::pair<iterator, bool> insert(const ValueT &Val) {
247 unsigned Idx = ValIndexOf(Val);
248 iterator I = findIndex(Idx);
250 return std::make_pair(I, false);
251 Sparse[Idx] = size();
252 Dense.push_back(Val);
253 return std::make_pair(end() - 1, true);
256 /// array subscript - If an element already exists with this key, return it.
257 /// Otherwise, automatically construct a new value from Key, insert it,
258 /// and return the newly inserted element.
259 ValueT &operator[](const KeyT &Key) {
260 return *insert(ValueT(Key)).first;
263 /// erase - Erases an existing element identified by a valid iterator.
265 /// This invalidates all iterators, but erase() returns an iterator pointing
266 /// to the next element. This makes it possible to erase selected elements
267 /// while iterating over the set:
269 /// for (SparseSet::iterator I = Set.begin(); I != Set.end();)
271 /// I = Set.erase(I);
275 /// Note that end() changes when elements are erased, unlike std::list.
277 iterator erase(iterator I) {
278 assert(unsigned(I - begin()) < size() && "Invalid iterator");
279 if (I != end() - 1) {
281 unsigned BackIdx = ValIndexOf(Dense.back());
282 assert(BackIdx < Universe && "Invalid key in set. Did object mutate?");
283 Sparse[BackIdx] = I - begin();
285 // This depends on SmallVector::pop_back() not invalidating iterators.
286 // std::vector::pop_back() doesn't give that guarantee.
291 /// erase - Erases an element identified by Key, if it exists.
293 /// @param Key The key identifying the element to erase.
294 /// @returns True when an element was erased, false if no element was found.
296 bool erase(const KeyT &Key) {
297 iterator I = find(Key);
306 } // end namespace llvm