1 //===--- llvm/ADT/SparseMultiSet.h - Sparse multiset ------------*- 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 SparseMultiSet class, which adds multiset behavior to
13 // A sparse multiset holds a small number of objects identified by integer keys
14 // from a moderately sized universe. The sparse multiset uses more memory than
15 // other containers in order to provide faster operations. Any key can map to
16 // multiple values. A SparseMultiSetNode class is provided, which serves as a
17 // convenient base class for the contents of a SparseMultiSet.
19 //===----------------------------------------------------------------------===//
21 #ifndef LLVM_ADT_SPARSEMULTISET_H
22 #define LLVM_ADT_SPARSEMULTISET_H
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/SparseSet.h"
26 #include "llvm/ADT/STLExtras.h"
36 /// Fast multiset implementation for objects that can be identified by small
39 /// SparseMultiSet allocates memory proportional to the size of the key
40 /// universe, so it is not recommended for building composite data structures.
41 /// It is useful for algorithms that require a single set with fast operations.
43 /// Compared to DenseSet and DenseMap, SparseMultiSet provides constant-time
44 /// fast clear() as fast as a vector. The find(), insert(), and erase()
45 /// operations are all constant time, and typically faster than a hash table.
46 /// The iteration order doesn't depend on numerical key values, it only depends
47 /// on the order of insert() and erase() operations. Iteration order is the
48 /// insertion order. Iteration is only provided over elements of equivalent
49 /// keys, but iterators are bidirectional.
51 /// Compared to BitVector, SparseMultiSet<unsigned> uses 8x-40x more memory, but
52 /// offers constant-time clear() and size() operations as well as fast iteration
53 /// independent on the size of the universe.
55 /// SparseMultiSet contains a dense vector holding all the objects and a sparse
56 /// array holding indexes into the dense vector. Most of the memory is used by
57 /// the sparse array which is the size of the key universe. The SparseT template
58 /// parameter provides a space/speed tradeoff for sets holding many elements.
60 /// When SparseT is uint32_t, find() only touches up to 3 cache lines, but the
61 /// sparse array uses 4 x Universe bytes.
63 /// When SparseT is uint8_t (the default), find() touches up to 3+[N/256] cache
64 /// lines, but the sparse array is 4x smaller. N is the number of elements in
67 /// For sets that may grow to thousands of elements, SparseT should be set to
68 /// uint16_t or uint32_t.
70 /// Multiset behavior is provided by providing doubly linked lists for values
71 /// that are inlined in the dense vector. SparseMultiSet is a good choice when
72 /// one desires a growable number of entries per key, as it will retain the
73 /// SparseSet algorithmic properties despite being growable. Thus, it is often a
74 /// better choice than a SparseSet of growable containers or a vector of
75 /// vectors. SparseMultiSet also keeps iterators valid after erasure (provided
76 /// the iterators don't point to the element erased), allowing for more
77 /// intuitive and fast removal.
79 /// @tparam ValueT The type of objects in the set.
80 /// @tparam KeyFunctorT A functor that computes an unsigned index from KeyT.
81 /// @tparam SparseT An unsigned integer type. See above.
83 template<typename ValueT,
84 typename KeyFunctorT = identity<unsigned>,
85 typename SparseT = uint8_t>
86 class SparseMultiSet {
87 static_assert(std::numeric_limits<SparseT>::is_integer &&
88 !std::numeric_limits<SparseT>::is_signed,
89 "SparseT must be an unsigned integer type");
91 /// The actual data that's stored, as a doubly-linked list implemented via
92 /// indices into the DenseVector. The doubly linked list is implemented
93 /// circular in Prev indices, and INVALID-terminated in Next indices. This
94 /// provides efficient access to list tails. These nodes can also be
95 /// tombstones, in which case they are actually nodes in a single-linked
96 /// freelist of recyclable slots.
98 static const unsigned INVALID = ~0U;
104 SMSNode(ValueT D, unsigned P, unsigned N) : Data(D), Prev(P), Next(N) { }
106 /// List tails have invalid Nexts.
107 bool isTail() const {
108 return Next == INVALID;
111 /// Whether this node is a tombstone node, and thus is in our freelist.
112 bool isTombstone() const {
113 return Prev == INVALID;
116 /// Since the list is circular in Prev, all non-tombstone nodes have a valid
118 bool isValid() const { return Prev != INVALID; }
121 typedef typename KeyFunctorT::argument_type KeyT;
122 typedef SmallVector<SMSNode, 8> DenseT;
124 SparseT *Sparse = nullptr;
125 unsigned Universe = 0;
126 KeyFunctorT KeyIndexOf;
127 SparseSetValFunctor<KeyT, ValueT, KeyFunctorT> ValIndexOf;
129 /// We have a built-in recycler for reusing tombstone slots. This recycler
130 /// puts a singly-linked free list into tombstone slots, allowing us quick
131 /// erasure, iterator preservation, and dense size.
132 unsigned FreelistIdx = SMSNode::INVALID;
133 unsigned NumFree = 0;
135 unsigned sparseIndex(const ValueT &Val) const {
136 assert(ValIndexOf(Val) < Universe &&
137 "Invalid key in set. Did object mutate?");
138 return ValIndexOf(Val);
140 unsigned sparseIndex(const SMSNode &N) const { return sparseIndex(N.Data); }
142 /// Whether the given entry is the head of the list. List heads's previous
143 /// pointers are to the tail of the list, allowing for efficient access to the
144 /// list tail. D must be a valid entry node.
145 bool isHead(const SMSNode &D) const {
146 assert(D.isValid() && "Invalid node for head");
147 return Dense[D.Prev].isTail();
150 /// Whether the given entry is a singleton entry, i.e. the only entry with
152 bool isSingleton(const SMSNode &N) const {
153 assert(N.isValid() && "Invalid node for singleton");
154 // Is N its own predecessor?
155 return &Dense[N.Prev] == &N;
158 /// Add in the given SMSNode. Uses a free entry in our freelist if
159 /// available. Returns the index of the added node.
160 unsigned addValue(const ValueT& V, unsigned Prev, unsigned Next) {
162 Dense.push_back(SMSNode(V, Prev, Next));
163 return Dense.size() - 1;
166 // Peel off a free slot
167 unsigned Idx = FreelistIdx;
168 unsigned NextFree = Dense[Idx].Next;
169 assert(Dense[Idx].isTombstone() && "Non-tombstone free?");
171 Dense[Idx] = SMSNode(V, Prev, Next);
172 FreelistIdx = NextFree;
177 /// Make the current index a new tombstone. Pushes it onto the freelist.
178 void makeTombstone(unsigned Idx) {
179 Dense[Idx].Prev = SMSNode::INVALID;
180 Dense[Idx].Next = FreelistIdx;
186 typedef ValueT value_type;
187 typedef ValueT &reference;
188 typedef const ValueT &const_reference;
189 typedef ValueT *pointer;
190 typedef const ValueT *const_pointer;
191 typedef unsigned size_type;
193 SparseMultiSet() = default;
194 SparseMultiSet(const SparseMultiSet &) = delete;
195 SparseMultiSet &operator=(const SparseMultiSet &) = delete;
196 ~SparseMultiSet() { free(Sparse); }
198 /// Set the universe size which determines the largest key the set can hold.
199 /// The universe must be sized before any elements can be added.
201 /// @param U Universe size. All object keys must be less than U.
203 void setUniverse(unsigned U) {
204 // It's not hard to resize the universe on a non-empty set, but it doesn't
205 // seem like a likely use case, so we can add that code when we need it.
206 assert(empty() && "Can only resize universe on an empty map");
207 // Hysteresis prevents needless reallocations.
208 if (U >= Universe/4 && U <= Universe)
211 // The Sparse array doesn't actually need to be initialized, so malloc
212 // would be enough here, but that will cause tools like valgrind to
213 // complain about branching on uninitialized data.
214 Sparse = reinterpret_cast<SparseT*>(calloc(U, sizeof(SparseT)));
218 /// Our iterators are iterators over the collection of objects that share a
220 template<typename SMSPtrTy>
221 class iterator_base : public std::iterator<std::bidirectional_iterator_tag,
223 friend class SparseMultiSet;
229 iterator_base(SMSPtrTy P, unsigned I, unsigned SI)
230 : SMS(P), Idx(I), SparseIdx(SI) { }
232 /// Whether our iterator has fallen outside our dense vector.
234 if (Idx == SMSNode::INVALID)
237 assert(Idx < SMS->Dense.size() && "Out of range, non-INVALID Idx?");
241 /// Whether our iterator is properly keyed, i.e. the SparseIdx is valid
242 bool isKeyed() const { return SparseIdx < SMS->Universe; }
244 unsigned Prev() const { return SMS->Dense[Idx].Prev; }
245 unsigned Next() const { return SMS->Dense[Idx].Next; }
247 void setPrev(unsigned P) { SMS->Dense[Idx].Prev = P; }
248 void setNext(unsigned N) { SMS->Dense[Idx].Next = N; }
251 typedef std::iterator<std::bidirectional_iterator_tag, ValueT> super;
252 typedef typename super::value_type value_type;
253 typedef typename super::difference_type difference_type;
254 typedef typename super::pointer pointer;
255 typedef typename super::reference reference;
257 reference operator*() const {
258 assert(isKeyed() && SMS->sparseIndex(SMS->Dense[Idx].Data) == SparseIdx &&
259 "Dereferencing iterator of invalid key or index");
261 return SMS->Dense[Idx].Data;
263 pointer operator->() const { return &operator*(); }
265 /// Comparison operators
266 bool operator==(const iterator_base &RHS) const {
267 // end compares equal
268 if (SMS == RHS.SMS && Idx == RHS.Idx) {
269 assert((isEnd() || SparseIdx == RHS.SparseIdx) &&
270 "Same dense entry, but different keys?");
277 bool operator!=(const iterator_base &RHS) const {
278 return !operator==(RHS);
281 /// Increment and decrement operators
282 iterator_base &operator--() { // predecrement - Back up
283 assert(isKeyed() && "Decrementing an invalid iterator");
284 assert((isEnd() || !SMS->isHead(SMS->Dense[Idx])) &&
285 "Decrementing head of list");
287 // If we're at the end, then issue a new find()
289 Idx = SMS->findIndex(SparseIdx).Prev();
295 iterator_base &operator++() { // preincrement - Advance
296 assert(!isEnd() && isKeyed() && "Incrementing an invalid/end iterator");
300 iterator_base operator--(int) { // postdecrement
301 iterator_base I(*this);
305 iterator_base operator++(int) { // postincrement
306 iterator_base I(*this);
311 typedef iterator_base<SparseMultiSet *> iterator;
312 typedef iterator_base<const SparseMultiSet *> const_iterator;
315 typedef std::pair<iterator, iterator> RangePair;
317 /// Returns an iterator past this container. Note that such an iterator cannot
318 /// be decremented, but will compare equal to other end iterators.
319 iterator end() { return iterator(this, SMSNode::INVALID, SMSNode::INVALID); }
320 const_iterator end() const {
321 return const_iterator(this, SMSNode::INVALID, SMSNode::INVALID);
324 /// Returns true if the set is empty.
326 /// This is not the same as BitVector::empty().
328 bool empty() const { return size() == 0; }
330 /// Returns the number of elements in the set.
332 /// This is not the same as BitVector::size() which returns the size of the
335 size_type size() const {
336 assert(NumFree <= Dense.size() && "Out-of-bounds free entries");
337 return Dense.size() - NumFree;
340 /// Clears the set. This is a very fast constant time operation.
343 // Sparse does not need to be cleared, see find().
346 FreelistIdx = SMSNode::INVALID;
349 /// Find an element by its index.
351 /// @param Idx A valid index to find.
352 /// @returns An iterator to the element identified by key, or end().
354 iterator findIndex(unsigned Idx) {
355 assert(Idx < Universe && "Key out of range");
356 const unsigned Stride = std::numeric_limits<SparseT>::max() + 1u;
357 for (unsigned i = Sparse[Idx], e = Dense.size(); i < e; i += Stride) {
358 const unsigned FoundIdx = sparseIndex(Dense[i]);
359 // Check that we're pointing at the correct entry and that it is the head
361 if (Idx == FoundIdx && Dense[i].isValid() && isHead(Dense[i]))
362 return iterator(this, i, Idx);
363 // Stride is 0 when SparseT >= unsigned. We don't need to loop.
370 /// Find an element by its key.
372 /// @param Key A valid key to find.
373 /// @returns An iterator to the element identified by key, or end().
375 iterator find(const KeyT &Key) {
376 return findIndex(KeyIndexOf(Key));
379 const_iterator find(const KeyT &Key) const {
380 iterator I = const_cast<SparseMultiSet*>(this)->findIndex(KeyIndexOf(Key));
381 return const_iterator(I.SMS, I.Idx, KeyIndexOf(Key));
384 /// Returns the number of elements identified by Key. This will be linear in
385 /// the number of elements of that key.
386 size_type count(const KeyT &Key) const {
388 for (const_iterator It = find(Key); It != end(); ++It)
394 /// Returns true if this set contains an element identified by Key.
395 bool contains(const KeyT &Key) const {
396 return find(Key) != end();
399 /// Return the head and tail of the subset's list, otherwise returns end().
400 iterator getHead(const KeyT &Key) { return find(Key); }
401 iterator getTail(const KeyT &Key) {
402 iterator I = find(Key);
404 I = iterator(this, I.Prev(), KeyIndexOf(Key));
408 /// The bounds of the range of items sharing Key K. First member is the head
409 /// of the list, and the second member is a decrementable end iterator for
411 RangePair equal_range(const KeyT &K) {
412 iterator B = find(K);
413 iterator E = iterator(this, SMSNode::INVALID, B.SparseIdx);
414 return make_pair(B, E);
417 /// Insert a new element at the tail of the subset list. Returns an iterator
418 /// to the newly added entry.
419 iterator insert(const ValueT &Val) {
420 unsigned Idx = sparseIndex(Val);
421 iterator I = findIndex(Idx);
423 unsigned NodeIdx = addValue(Val, SMSNode::INVALID, SMSNode::INVALID);
426 // Make a singleton list
427 Sparse[Idx] = NodeIdx;
428 Dense[NodeIdx].Prev = NodeIdx;
429 return iterator(this, NodeIdx, Idx);
432 // Stick it at the end.
433 unsigned HeadIdx = I.Idx;
434 unsigned TailIdx = I.Prev();
435 Dense[TailIdx].Next = NodeIdx;
436 Dense[HeadIdx].Prev = NodeIdx;
437 Dense[NodeIdx].Prev = TailIdx;
439 return iterator(this, NodeIdx, Idx);
442 /// Erases an existing element identified by a valid iterator.
444 /// This invalidates iterators pointing at the same entry, but erase() returns
445 /// an iterator pointing to the next element in the subset's list. This makes
446 /// it possible to erase selected elements while iterating over the subset:
448 /// tie(I, E) = Set.equal_range(Key);
451 /// I = Set.erase(I);
455 /// Note that if the last element in the subset list is erased, this will
456 /// return an end iterator which can be decremented to get the new tail (if it
459 /// tie(B, I) = Set.equal_range(Key);
460 /// for (bool isBegin = B == I; !isBegin; /* empty */) {
461 /// isBegin = (--I) == B;
466 iterator erase(iterator I) {
467 assert(I.isKeyed() && !I.isEnd() && !Dense[I.Idx].isTombstone() &&
468 "erasing invalid/end/tombstone iterator");
470 // First, unlink the node from its list. Then swap the node out with the
471 // dense vector's last entry
472 iterator NextI = unlink(Dense[I.Idx]);
474 // Put in a tombstone.
475 makeTombstone(I.Idx);
480 /// Erase all elements with the given key. This invalidates all
481 /// iterators of that key.
482 void eraseAll(const KeyT &K) {
483 for (iterator I = find(K); I != end(); /* empty */)
488 /// Unlink the node from its list. Returns the next node in the list.
489 iterator unlink(const SMSNode &N) {
490 if (isSingleton(N)) {
491 // Singleton is already unlinked
492 assert(N.Next == SMSNode::INVALID && "Singleton has next?");
493 return iterator(this, SMSNode::INVALID, ValIndexOf(N.Data));
497 // If we're the head, then update the sparse array and our next.
498 Sparse[sparseIndex(N)] = N.Next;
499 Dense[N.Next].Prev = N.Prev;
500 return iterator(this, N.Next, ValIndexOf(N.Data));
504 // If we're the tail, then update our head and our previous.
505 findIndex(sparseIndex(N)).setPrev(N.Prev);
506 Dense[N.Prev].Next = N.Next;
508 // Give back an end iterator that can be decremented
509 iterator I(this, N.Prev, ValIndexOf(N.Data));
513 // Otherwise, just drop us
514 Dense[N.Next].Prev = N.Prev;
515 Dense[N.Prev].Next = N.Next;
516 return iterator(this, N.Next, ValIndexOf(N.Data));
520 } // end namespace llvm
522 #endif // LLVM_ADT_SPARSEMULTISET_H