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/SparseSet.h"
28 /// Fast multiset implementation for objects that can be identified by small
31 /// SparseMultiSet allocates memory proportional to the size of the key
32 /// universe, so it is not recommended for building composite data structures.
33 /// It is useful for algorithms that require a single set with fast operations.
35 /// Compared to DenseSet and DenseMap, SparseMultiSet provides constant-time
36 /// fast clear() as fast as a vector. The find(), insert(), and erase()
37 /// operations are all constant time, and typically faster than a hash table.
38 /// The iteration order doesn't depend on numerical key values, it only depends
39 /// on the order of insert() and erase() operations. Iteration order is the
40 /// insertion order. Iteration is only provided over elements of equivalent
41 /// keys, but iterators are bidirectional.
43 /// Compared to BitVector, SparseMultiSet<unsigned> uses 8x-40x more memory, but
44 /// offers constant-time clear() and size() operations as well as fast iteration
45 /// independent on the size of the universe.
47 /// SparseMultiSet contains a dense vector holding all the objects and a sparse
48 /// array holding indexes into the dense vector. Most of the memory is used by
49 /// the sparse array which is the size of the key universe. The SparseT template
50 /// parameter provides a space/speed tradeoff for sets holding many elements.
52 /// When SparseT is uint32_t, find() only touches up to 3 cache lines, but the
53 /// sparse array uses 4 x Universe bytes.
55 /// When SparseT is uint8_t (the default), find() touches up to 3+[N/256] cache
56 /// lines, but the sparse array is 4x smaller. N is the number of elements in
59 /// For sets that may grow to thousands of elements, SparseT should be set to
60 /// uint16_t or uint32_t.
62 /// Multiset behavior is provided by providing doubly linked lists for values
63 /// that are inlined in the dense vector. SparseMultiSet is a good choice when
64 /// one desires a growable number of entries per key, as it will retain the
65 /// SparseSet algorithmic properties despite being growable. Thus, it is often a
66 /// better choice than a SparseSet of growable containers or a vector of
67 /// vectors. SparseMultiSet also keeps iterators valid after erasure (provided
68 /// the iterators don't point to the element erased), allowing for more
69 /// intuitive and fast removal.
71 /// @tparam ValueT The type of objects in the set.
72 /// @tparam KeyFunctorT A functor that computes an unsigned index from KeyT.
73 /// @tparam SparseT An unsigned integer type. See above.
75 template<typename ValueT,
76 typename KeyFunctorT = llvm::identity<unsigned>,
77 typename SparseT = uint8_t>
78 class SparseMultiSet {
79 static_assert(std::numeric_limits<SparseT>::is_integer &&
80 !std::numeric_limits<SparseT>::is_signed,
81 "SparseT must be an unsigned integer type");
83 /// The actual data that's stored, as a doubly-linked list implemented via
84 /// indices into the DenseVector. The doubly linked list is implemented
85 /// circular in Prev indices, and INVALID-terminated in Next indices. This
86 /// provides efficient access to list tails. These nodes can also be
87 /// tombstones, in which case they are actually nodes in a single-linked
88 /// freelist of recyclable slots.
90 static const unsigned INVALID = ~0U;
96 SMSNode(ValueT D, unsigned P, unsigned N) : Data(D), Prev(P), Next(N) { }
98 /// List tails have invalid Nexts.
100 return Next == INVALID;
103 /// Whether this node is a tombstone node, and thus is in our freelist.
104 bool isTombstone() const {
105 return Prev == INVALID;
108 /// Since the list is circular in Prev, all non-tombstone nodes have a valid
110 bool isValid() const { return Prev != INVALID; }
113 typedef typename KeyFunctorT::argument_type KeyT;
114 typedef SmallVector<SMSNode, 8> DenseT;
118 KeyFunctorT KeyIndexOf;
119 SparseSetValFunctor<KeyT, ValueT, KeyFunctorT> ValIndexOf;
121 /// We have a built-in recycler for reusing tombstone slots. This recycler
122 /// puts a singly-linked free list into tombstone slots, allowing us quick
123 /// erasure, iterator preservation, and dense size.
124 unsigned FreelistIdx;
127 unsigned sparseIndex(const ValueT &Val) const {
128 assert(ValIndexOf(Val) < Universe &&
129 "Invalid key in set. Did object mutate?");
130 return ValIndexOf(Val);
132 unsigned sparseIndex(const SMSNode &N) const { return sparseIndex(N.Data); }
134 // Disable copy construction and assignment.
135 // This data structure is not meant to be used that way.
136 SparseMultiSet(const SparseMultiSet&) LLVM_DELETED_FUNCTION;
137 SparseMultiSet &operator=(const SparseMultiSet&) LLVM_DELETED_FUNCTION;
139 /// Whether the given entry is the head of the list. List heads's previous
140 /// pointers are to the tail of the list, allowing for efficient access to the
141 /// list tail. D must be a valid entry node.
142 bool isHead(const SMSNode &D) const {
143 assert(D.isValid() && "Invalid node for head");
144 return Dense[D.Prev].isTail();
147 /// Whether the given entry is a singleton entry, i.e. the only entry with
149 bool isSingleton(const SMSNode &N) const {
150 assert(N.isValid() && "Invalid node for singleton");
151 // Is N its own predecessor?
152 return &Dense[N.Prev] == &N;
155 /// Add in the given SMSNode. Uses a free entry in our freelist if
156 /// available. Returns the index of the added node.
157 unsigned addValue(const ValueT& V, unsigned Prev, unsigned Next) {
159 Dense.push_back(SMSNode(V, Prev, Next));
160 return Dense.size() - 1;
163 // Peel off a free slot
164 unsigned Idx = FreelistIdx;
165 unsigned NextFree = Dense[Idx].Next;
166 assert(Dense[Idx].isTombstone() && "Non-tombstone free?");
168 Dense[Idx] = SMSNode(V, Prev, Next);
169 FreelistIdx = NextFree;
174 /// Make the current index a new tombstone. Pushes it onto the freelist.
175 void makeTombstone(unsigned Idx) {
176 Dense[Idx].Prev = SMSNode::INVALID;
177 Dense[Idx].Next = FreelistIdx;
183 typedef ValueT value_type;
184 typedef ValueT &reference;
185 typedef const ValueT &const_reference;
186 typedef ValueT *pointer;
187 typedef const ValueT *const_pointer;
190 : Sparse(nullptr), Universe(0), FreelistIdx(SMSNode::INVALID), NumFree(0) {}
192 ~SparseMultiSet() { free(Sparse); }
194 /// Set the universe size which determines the largest key the set can hold.
195 /// The universe must be sized before any elements can be added.
197 /// @param U Universe size. All object keys must be less than U.
199 void setUniverse(unsigned U) {
200 // It's not hard to resize the universe on a non-empty set, but it doesn't
201 // seem like a likely use case, so we can add that code when we need it.
202 assert(empty() && "Can only resize universe on an empty map");
203 // Hysteresis prevents needless reallocations.
204 if (U >= Universe/4 && U <= Universe)
207 // The Sparse array doesn't actually need to be initialized, so malloc
208 // would be enough here, but that will cause tools like valgrind to
209 // complain about branching on uninitialized data.
210 Sparse = reinterpret_cast<SparseT*>(calloc(U, sizeof(SparseT)));
214 /// Our iterators are iterators over the collection of objects that share a
216 template<typename SMSPtrTy>
217 class iterator_base : public std::iterator<std::bidirectional_iterator_tag,
219 friend class SparseMultiSet;
224 iterator_base(SMSPtrTy P, unsigned I, unsigned SI)
225 : SMS(P), Idx(I), SparseIdx(SI) { }
227 /// Whether our iterator has fallen outside our dense vector.
229 if (Idx == SMSNode::INVALID)
232 assert(Idx < SMS->Dense.size() && "Out of range, non-INVALID Idx?");
236 /// Whether our iterator is properly keyed, i.e. the SparseIdx is valid
237 bool isKeyed() const { return SparseIdx < SMS->Universe; }
239 unsigned Prev() const { return SMS->Dense[Idx].Prev; }
240 unsigned Next() const { return SMS->Dense[Idx].Next; }
242 void setPrev(unsigned P) { SMS->Dense[Idx].Prev = P; }
243 void setNext(unsigned N) { SMS->Dense[Idx].Next = N; }
246 typedef std::iterator<std::bidirectional_iterator_tag, ValueT> super;
247 typedef typename super::value_type value_type;
248 typedef typename super::difference_type difference_type;
249 typedef typename super::pointer pointer;
250 typedef typename super::reference reference;
252 reference operator*() const {
253 assert(isKeyed() && SMS->sparseIndex(SMS->Dense[Idx].Data) == SparseIdx &&
254 "Dereferencing iterator of invalid key or index");
256 return SMS->Dense[Idx].Data;
258 pointer operator->() const { return &operator*(); }
260 /// Comparison operators
261 bool operator==(const iterator_base &RHS) const {
262 // end compares equal
263 if (SMS == RHS.SMS && Idx == RHS.Idx) {
264 assert((isEnd() || SparseIdx == RHS.SparseIdx) &&
265 "Same dense entry, but different keys?");
272 bool operator!=(const iterator_base &RHS) const {
273 return !operator==(RHS);
276 /// Increment and decrement operators
277 iterator_base &operator--() { // predecrement - Back up
278 assert(isKeyed() && "Decrementing an invalid iterator");
279 assert((isEnd() || !SMS->isHead(SMS->Dense[Idx])) &&
280 "Decrementing head of list");
282 // If we're at the end, then issue a new find()
284 Idx = SMS->findIndex(SparseIdx).Prev();
290 iterator_base &operator++() { // preincrement - Advance
291 assert(!isEnd() && isKeyed() && "Incrementing an invalid/end iterator");
295 iterator_base operator--(int) { // postdecrement
296 iterator_base I(*this);
300 iterator_base operator++(int) { // postincrement
301 iterator_base I(*this);
306 typedef iterator_base<SparseMultiSet *> iterator;
307 typedef iterator_base<const SparseMultiSet *> const_iterator;
310 typedef std::pair<iterator, iterator> RangePair;
312 /// Returns an iterator past this container. Note that such an iterator cannot
313 /// be decremented, but will compare equal to other end iterators.
314 iterator end() { return iterator(this, SMSNode::INVALID, SMSNode::INVALID); }
315 const_iterator end() const {
316 return const_iterator(this, SMSNode::INVALID, SMSNode::INVALID);
319 /// Returns true if the set is empty.
321 /// This is not the same as BitVector::empty().
323 bool empty() const { return size() == 0; }
325 /// Returns the number of elements in the set.
327 /// This is not the same as BitVector::size() which returns the size of the
330 unsigned size() const {
331 assert(NumFree <= Dense.size() && "Out-of-bounds free entries");
332 return Dense.size() - NumFree;
335 /// Clears the set. This is a very fast constant time operation.
338 // Sparse does not need to be cleared, see find().
341 FreelistIdx = SMSNode::INVALID;
344 /// Find an element by its index.
346 /// @param Idx A valid index to find.
347 /// @returns An iterator to the element identified by key, or end().
349 iterator findIndex(unsigned Idx) {
350 assert(Idx < Universe && "Key out of range");
351 const unsigned Stride = std::numeric_limits<SparseT>::max() + 1u;
352 for (unsigned i = Sparse[Idx], e = Dense.size(); i < e; i += Stride) {
353 const unsigned FoundIdx = sparseIndex(Dense[i]);
354 // Check that we're pointing at the correct entry and that it is the head
356 if (Idx == FoundIdx && Dense[i].isValid() && isHead(Dense[i]))
357 return iterator(this, i, Idx);
358 // Stride is 0 when SparseT >= unsigned. We don't need to loop.
365 /// Find an element by its key.
367 /// @param Key A valid key to find.
368 /// @returns An iterator to the element identified by key, or end().
370 iterator find(const KeyT &Key) {
371 return findIndex(KeyIndexOf(Key));
374 const_iterator find(const KeyT &Key) const {
375 iterator I = const_cast<SparseMultiSet*>(this)->findIndex(KeyIndexOf(Key));
376 return const_iterator(I.SMS, I.Idx, KeyIndexOf(Key));
379 /// Returns the number of elements identified by Key. This will be linear in
380 /// the number of elements of that key.
381 unsigned count(const KeyT &Key) const {
383 for (const_iterator It = find(Key); It != end(); ++It)
389 /// Returns true if this set contains an element identified by Key.
390 bool contains(const KeyT &Key) const {
391 return find(Key) != end();
394 /// Return the head and tail of the subset's list, otherwise returns end().
395 iterator getHead(const KeyT &Key) { return find(Key); }
396 iterator getTail(const KeyT &Key) {
397 iterator I = find(Key);
399 I = iterator(this, I.Prev(), KeyIndexOf(Key));
403 /// The bounds of the range of items sharing Key K. First member is the head
404 /// of the list, and the second member is a decrementable end iterator for
406 RangePair equal_range(const KeyT &K) {
407 iterator B = find(K);
408 iterator E = iterator(this, SMSNode::INVALID, B.SparseIdx);
409 return make_pair(B, E);
412 /// Insert a new element at the tail of the subset list. Returns an iterator
413 /// to the newly added entry.
414 iterator insert(const ValueT &Val) {
415 unsigned Idx = sparseIndex(Val);
416 iterator I = findIndex(Idx);
418 unsigned NodeIdx = addValue(Val, SMSNode::INVALID, SMSNode::INVALID);
421 // Make a singleton list
422 Sparse[Idx] = NodeIdx;
423 Dense[NodeIdx].Prev = NodeIdx;
424 return iterator(this, NodeIdx, Idx);
427 // Stick it at the end.
428 unsigned HeadIdx = I.Idx;
429 unsigned TailIdx = I.Prev();
430 Dense[TailIdx].Next = NodeIdx;
431 Dense[HeadIdx].Prev = NodeIdx;
432 Dense[NodeIdx].Prev = TailIdx;
434 return iterator(this, NodeIdx, Idx);
437 /// Erases an existing element identified by a valid iterator.
439 /// This invalidates iterators pointing at the same entry, but erase() returns
440 /// an iterator pointing to the next element in the subset's list. This makes
441 /// it possible to erase selected elements while iterating over the subset:
443 /// tie(I, E) = Set.equal_range(Key);
446 /// I = Set.erase(I);
450 /// Note that if the last element in the subset list is erased, this will
451 /// return an end iterator which can be decremented to get the new tail (if it
454 /// tie(B, I) = Set.equal_range(Key);
455 /// for (bool isBegin = B == I; !isBegin; /* empty */) {
456 /// isBegin = (--I) == B;
461 iterator erase(iterator I) {
462 assert(I.isKeyed() && !I.isEnd() && !Dense[I.Idx].isTombstone() &&
463 "erasing invalid/end/tombstone iterator");
465 // First, unlink the node from its list. Then swap the node out with the
466 // dense vector's last entry
467 iterator NextI = unlink(Dense[I.Idx]);
469 // Put in a tombstone.
470 makeTombstone(I.Idx);
475 /// Erase all elements with the given key. This invalidates all
476 /// iterators of that key.
477 void eraseAll(const KeyT &K) {
478 for (iterator I = find(K); I != end(); /* empty */)
483 /// Unlink the node from its list. Returns the next node in the list.
484 iterator unlink(const SMSNode &N) {
485 if (isSingleton(N)) {
486 // Singleton is already unlinked
487 assert(N.Next == SMSNode::INVALID && "Singleton has next?");
488 return iterator(this, SMSNode::INVALID, ValIndexOf(N.Data));
492 // If we're the head, then update the sparse array and our next.
493 Sparse[sparseIndex(N)] = N.Next;
494 Dense[N.Next].Prev = N.Prev;
495 return iterator(this, N.Next, ValIndexOf(N.Data));
499 // If we're the tail, then update our head and our previous.
500 findIndex(sparseIndex(N)).setPrev(N.Prev);
501 Dense[N.Prev].Next = N.Next;
503 // Give back an end iterator that can be decremented
504 iterator I(this, N.Prev, ValIndexOf(N.Data));
508 // Otherwise, just drop us
509 Dense[N.Next].Prev = N.Prev;
510 Dense[N.Prev].Next = N.Next;
511 return iterator(this, N.Next, ValIndexOf(N.Data));
515 } // end namespace llvm