]> git.proxmox.com Git - rustc.git/blob - src/llvm/include/llvm/ADT/SparseSet.h
Imported Upstream version 1.0.0+dfsg1
[rustc.git] / src / llvm / include / llvm / ADT / SparseSet.h
1 //===--- llvm/ADT/SparseSet.h - Sparse set ----------------------*- C++ -*-===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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.
13 //
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.
17 //
18 //===----------------------------------------------------------------------===//
19
20 #ifndef LLVM_ADT_SPARSESET_H
21 #define LLVM_ADT_SPARSESET_H
22
23 #include "llvm/ADT/STLExtras.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/Support/DataTypes.h"
26 #include <limits>
27
28 namespace llvm {
29
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.
38 ///
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()
42 ///
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()
47 ///
48 /// Only other, unexpected cases require specializing SparseSetValTraits.
49 ///
50 /// For best results, ValueT should not require a destructor.
51 ///
52 template<typename ValueT>
53 struct SparseSetValTraits {
54 static unsigned getValIndex(const ValueT &Val) {
55 return Val.getSparseSetIndex();
56 }
57 };
58
59 /// SparseSetValFunctor - Helper class for selecting SparseSetValTraits. The
60 /// generic implementation handles ValueT classes which either provide
61 /// getSparseSetIndex() or specialize SparseSetValTraits<>.
62 ///
63 template<typename KeyT, typename ValueT, typename KeyFunctorT>
64 struct SparseSetValFunctor {
65 unsigned operator()(const ValueT &Val) const {
66 return SparseSetValTraits<ValueT>::getValIndex(Val);
67 }
68 };
69
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);
76 }
77 };
78
79 /// SparseSet - Fast set implmentation for objects that can be identified by
80 /// small unsigned keys.
81 ///
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.
85 ///
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.
92 ///
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.
96 ///
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
101 /// elements.
102 ///
103 /// When SparseT is uint32_t, find() only touches 2 cache lines, but the sparse
104 /// array uses 4 x Universe bytes.
105 ///
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
108 /// the set.
109 ///
110 /// For sets that may grow to thousands of elements, SparseT should be set to
111 /// uint16_t or uint32_t.
112 ///
113 /// @tparam ValueT The type of objects in the set.
114 /// @tparam KeyFunctorT A functor that computes an unsigned index from KeyT.
115 /// @tparam SparseT An unsigned integer type. See above.
116 ///
117 template<typename ValueT,
118 typename KeyFunctorT = llvm::identity<unsigned>,
119 typename SparseT = uint8_t>
120 class SparseSet {
121 static_assert(std::numeric_limits<SparseT>::is_integer &&
122 !std::numeric_limits<SparseT>::is_signed,
123 "SparseT must be an unsigned integer type");
124
125 typedef typename KeyFunctorT::argument_type KeyT;
126 typedef SmallVector<ValueT, 8> DenseT;
127 typedef unsigned size_type;
128 DenseT Dense;
129 SparseT *Sparse;
130 unsigned Universe;
131 KeyFunctorT KeyIndexOf;
132 SparseSetValFunctor<KeyT, ValueT, KeyFunctorT> ValIndexOf;
133
134 // Disable copy construction and assignment.
135 // This data structure is not meant to be used that way.
136 SparseSet(const SparseSet&) LLVM_DELETED_FUNCTION;
137 SparseSet &operator=(const SparseSet&) LLVM_DELETED_FUNCTION;
138
139 public:
140 typedef ValueT value_type;
141 typedef ValueT &reference;
142 typedef const ValueT &const_reference;
143 typedef ValueT *pointer;
144 typedef const ValueT *const_pointer;
145
146 SparseSet() : Sparse(nullptr), Universe(0) {}
147 ~SparseSet() { free(Sparse); }
148
149 /// setUniverse - Set the universe size which determines the largest key the
150 /// set can hold. The universe must be sized before any elements can be
151 /// added.
152 ///
153 /// @param U Universe size. All object keys must be less than U.
154 ///
155 void setUniverse(unsigned U) {
156 // It's not hard to resize the universe on a non-empty set, but it doesn't
157 // seem like a likely use case, so we can add that code when we need it.
158 assert(empty() && "Can only resize universe on an empty map");
159 // Hysteresis prevents needless reallocations.
160 if (U >= Universe/4 && U <= Universe)
161 return;
162 free(Sparse);
163 // The Sparse array doesn't actually need to be initialized, so malloc
164 // would be enough here, but that will cause tools like valgrind to
165 // complain about branching on uninitialized data.
166 Sparse = reinterpret_cast<SparseT*>(calloc(U, sizeof(SparseT)));
167 Universe = U;
168 }
169
170 // Import trivial vector stuff from DenseT.
171 typedef typename DenseT::iterator iterator;
172 typedef typename DenseT::const_iterator const_iterator;
173
174 const_iterator begin() const { return Dense.begin(); }
175 const_iterator end() const { return Dense.end(); }
176 iterator begin() { return Dense.begin(); }
177 iterator end() { return Dense.end(); }
178
179 /// empty - Returns true if the set is empty.
180 ///
181 /// This is not the same as BitVector::empty().
182 ///
183 bool empty() const { return Dense.empty(); }
184
185 /// size - Returns the number of elements in the set.
186 ///
187 /// This is not the same as BitVector::size() which returns the size of the
188 /// universe.
189 ///
190 size_type size() const { return Dense.size(); }
191
192 /// clear - Clears the set. This is a very fast constant time operation.
193 ///
194 void clear() {
195 // Sparse does not need to be cleared, see find().
196 Dense.clear();
197 }
198
199 /// findIndex - Find an element by its index.
200 ///
201 /// @param Idx A valid index to find.
202 /// @returns An iterator to the element identified by key, or end().
203 ///
204 iterator findIndex(unsigned Idx) {
205 assert(Idx < Universe && "Key out of range");
206 const unsigned Stride = std::numeric_limits<SparseT>::max() + 1u;
207 for (unsigned i = Sparse[Idx], e = size(); i < e; i += Stride) {
208 const unsigned FoundIdx = ValIndexOf(Dense[i]);
209 assert(FoundIdx < Universe && "Invalid key in set. Did object mutate?");
210 if (Idx == FoundIdx)
211 return begin() + i;
212 // Stride is 0 when SparseT >= unsigned. We don't need to loop.
213 if (!Stride)
214 break;
215 }
216 return end();
217 }
218
219 /// find - Find an element by its key.
220 ///
221 /// @param Key A valid key to find.
222 /// @returns An iterator to the element identified by key, or end().
223 ///
224 iterator find(const KeyT &Key) {
225 return findIndex(KeyIndexOf(Key));
226 }
227
228 const_iterator find(const KeyT &Key) const {
229 return const_cast<SparseSet*>(this)->findIndex(KeyIndexOf(Key));
230 }
231
232 /// count - Returns 1 if this set contains an element identified by Key,
233 /// 0 otherwise.
234 ///
235 size_type count(const KeyT &Key) const {
236 return find(Key) == end() ? 0 : 1;
237 }
238
239 /// insert - Attempts to insert a new element.
240 ///
241 /// If Val is successfully inserted, return (I, true), where I is an iterator
242 /// pointing to the newly inserted element.
243 ///
244 /// If the set already contains an element with the same key as Val, return
245 /// (I, false), where I is an iterator pointing to the existing element.
246 ///
247 /// Insertion invalidates all iterators.
248 ///
249 std::pair<iterator, bool> insert(const ValueT &Val) {
250 unsigned Idx = ValIndexOf(Val);
251 iterator I = findIndex(Idx);
252 if (I != end())
253 return std::make_pair(I, false);
254 Sparse[Idx] = size();
255 Dense.push_back(Val);
256 return std::make_pair(end() - 1, true);
257 }
258
259 /// array subscript - If an element already exists with this key, return it.
260 /// Otherwise, automatically construct a new value from Key, insert it,
261 /// and return the newly inserted element.
262 ValueT &operator[](const KeyT &Key) {
263 return *insert(ValueT(Key)).first;
264 }
265
266 /// erase - Erases an existing element identified by a valid iterator.
267 ///
268 /// This invalidates all iterators, but erase() returns an iterator pointing
269 /// to the next element. This makes it possible to erase selected elements
270 /// while iterating over the set:
271 ///
272 /// for (SparseSet::iterator I = Set.begin(); I != Set.end();)
273 /// if (test(*I))
274 /// I = Set.erase(I);
275 /// else
276 /// ++I;
277 ///
278 /// Note that end() changes when elements are erased, unlike std::list.
279 ///
280 iterator erase(iterator I) {
281 assert(unsigned(I - begin()) < size() && "Invalid iterator");
282 if (I != end() - 1) {
283 *I = Dense.back();
284 unsigned BackIdx = ValIndexOf(Dense.back());
285 assert(BackIdx < Universe && "Invalid key in set. Did object mutate?");
286 Sparse[BackIdx] = I - begin();
287 }
288 // This depends on SmallVector::pop_back() not invalidating iterators.
289 // std::vector::pop_back() doesn't give that guarantee.
290 Dense.pop_back();
291 return I;
292 }
293
294 /// erase - Erases an element identified by Key, if it exists.
295 ///
296 /// @param Key The key identifying the element to erase.
297 /// @returns True when an element was erased, false if no element was found.
298 ///
299 bool erase(const KeyT &Key) {
300 iterator I = find(Key);
301 if (I == end())
302 return false;
303 erase(I);
304 return true;
305 }
306
307 };
308
309 } // end namespace llvm
310
311 #endif