LLVM  8.0.0svn
SparseSet.h
Go to the documentation of this file.
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/Allocator.h"
26 #include <cassert>
27 #include <cstdint>
28 #include <cstdlib>
29 #include <limits>
30 #include <utility>
31 
32 namespace llvm {
33 
34 /// SparseSetValTraits - Objects in a SparseSet are identified by keys that can
35 /// be uniquely converted to a small integer less than the set's universe. This
36 /// class allows the set to hold values that differ from the set's key type as
37 /// long as an index can still be derived from the value. SparseSet never
38 /// directly compares ValueT, only their indices, so it can map keys to
39 /// arbitrary values. SparseSetValTraits computes the index from the value
40 /// object. To compute the index from a key, SparseSet uses a separate
41 /// KeyFunctorT template argument.
42 ///
43 /// A simple type declaration, SparseSet<Type>, handles these cases:
44 /// - unsigned key, identity index, identity value
45 /// - unsigned key, identity index, fat value providing getSparseSetIndex()
46 ///
47 /// The type declaration SparseSet<Type, UnaryFunction> handles:
48 /// - unsigned key, remapped index, identity value (virtual registers)
49 /// - pointer key, pointer-derived index, identity value (node+ID)
50 /// - pointer key, pointer-derived index, fat value with getSparseSetIndex()
51 ///
52 /// Only other, unexpected cases require specializing SparseSetValTraits.
53 ///
54 /// For best results, ValueT should not require a destructor.
55 ///
56 template<typename ValueT>
58  static unsigned getValIndex(const ValueT &Val) {
59  return Val.getSparseSetIndex();
60  }
61 };
62 
63 /// SparseSetValFunctor - Helper class for selecting SparseSetValTraits. The
64 /// generic implementation handles ValueT classes which either provide
65 /// getSparseSetIndex() or specialize SparseSetValTraits<>.
66 ///
67 template<typename KeyT, typename ValueT, typename KeyFunctorT>
69  unsigned operator()(const ValueT &Val) const {
71  }
72 };
73 
74 /// SparseSetValFunctor<KeyT, KeyT> - Helper class for the common case of
75 /// identity key/value sets.
76 template<typename KeyT, typename KeyFunctorT>
77 struct SparseSetValFunctor<KeyT, KeyT, KeyFunctorT> {
78  unsigned operator()(const KeyT &Key) const {
79  return KeyFunctorT()(Key);
80  }
81 };
82 
83 /// SparseSet - Fast set implmentation for objects that can be identified by
84 /// small unsigned keys.
85 ///
86 /// SparseSet allocates memory proportional to the size of the key universe, so
87 /// it is not recommended for building composite data structures. It is useful
88 /// for algorithms that require a single set with fast operations.
89 ///
90 /// Compared to DenseSet and DenseMap, SparseSet provides constant-time fast
91 /// clear() and iteration as fast as a vector. The find(), insert(), and
92 /// erase() operations are all constant time, and typically faster than a hash
93 /// table. The iteration order doesn't depend on numerical key values, it only
94 /// depends on the order of insert() and erase() operations. When no elements
95 /// have been erased, the iteration order is the insertion order.
96 ///
97 /// Compared to BitVector, SparseSet<unsigned> uses 8x-40x more memory, but
98 /// offers constant-time clear() and size() operations as well as fast
99 /// iteration independent on the size of the universe.
100 ///
101 /// SparseSet contains a dense vector holding all the objects and a sparse
102 /// array holding indexes into the dense vector. Most of the memory is used by
103 /// the sparse array which is the size of the key universe. The SparseT
104 /// template parameter provides a space/speed tradeoff for sets holding many
105 /// elements.
106 ///
107 /// When SparseT is uint32_t, find() only touches 2 cache lines, but the sparse
108 /// array uses 4 x Universe bytes.
109 ///
110 /// When SparseT is uint8_t (the default), find() touches up to 2+[N/256] cache
111 /// lines, but the sparse array is 4x smaller. N is the number of elements in
112 /// the set.
113 ///
114 /// For sets that may grow to thousands of elements, SparseT should be set to
115 /// uint16_t or uint32_t.
116 ///
117 /// @tparam ValueT The type of objects in the set.
118 /// @tparam KeyFunctorT A functor that computes an unsigned index from KeyT.
119 /// @tparam SparseT An unsigned integer type. See above.
120 ///
121 template<typename ValueT,
122  typename KeyFunctorT = identity<unsigned>,
123  typename SparseT = uint8_t>
124 class SparseSet {
125  static_assert(std::numeric_limits<SparseT>::is_integer &&
126  !std::numeric_limits<SparseT>::is_signed,
127  "SparseT must be an unsigned integer type");
128 
129  using KeyT = typename KeyFunctorT::argument_type;
131  using size_type = unsigned;
132  DenseT Dense;
133  SparseT *Sparse = nullptr;
134  unsigned Universe = 0;
135  KeyFunctorT KeyIndexOf;
137 
138 public:
140  using reference = ValueT &;
141  using const_reference = const ValueT &;
142  using pointer = ValueT *;
143  using const_pointer = const ValueT *;
144 
145  SparseSet() = default;
146  SparseSet(const SparseSet &) = delete;
147  SparseSet &operator=(const SparseSet &) = delete;
148  ~SparseSet() { free(Sparse); }
149 
150  /// setUniverse - Set the universe size which determines the largest key the
151  /// set can hold. The universe must be sized before any elements can be
152  /// added.
153  ///
154  /// @param U Universe size. All object keys must be less than U.
155  ///
156  void setUniverse(unsigned U) {
157  // It's not hard to resize the universe on a non-empty set, but it doesn't
158  // seem like a likely use case, so we can add that code when we need it.
159  assert(empty() && "Can only resize universe on an empty map");
160  // Hysteresis prevents needless reallocations.
161  if (U >= Universe/4 && U <= Universe)
162  return;
163  free(Sparse);
164  // The Sparse array doesn't actually need to be initialized, so malloc
165  // would be enough here, but that will cause tools like valgrind to
166  // complain about branching on uninitialized data.
167  Sparse = static_cast<SparseT*>(safe_calloc(U, sizeof(SparseT)));
168  Universe = U;
169  }
170 
171  // Import trivial vector stuff from DenseT.
172  using iterator = typename DenseT::iterator;
174 
175  const_iterator begin() const { return Dense.begin(); }
176  const_iterator end() const { return Dense.end(); }
177  iterator begin() { return Dense.begin(); }
178  iterator end() { return Dense.end(); }
179 
180  /// empty - Returns true if the set is empty.
181  ///
182  /// This is not the same as BitVector::empty().
183  ///
184  bool empty() const { return Dense.empty(); }
185 
186  /// size - Returns the number of elements in the set.
187  ///
188  /// This is not the same as BitVector::size() which returns the size of the
189  /// universe.
190  ///
191  size_type size() const { return Dense.size(); }
192 
193  /// clear - Clears the set. This is a very fast constant time operation.
194  ///
195  void clear() {
196  // Sparse does not need to be cleared, see find().
197  Dense.clear();
198  }
199 
200  /// findIndex - Find an element by its index.
201  ///
202  /// @param Idx A valid index to find.
203  /// @returns An iterator to the element identified by key, or end().
204  ///
205  iterator findIndex(unsigned Idx) {
206  assert(Idx < Universe && "Key out of range");
207  const unsigned Stride = std::numeric_limits<SparseT>::max() + 1u;
208  for (unsigned i = Sparse[Idx], e = size(); i < e; i += Stride) {
209  const unsigned FoundIdx = ValIndexOf(Dense[i]);
210  assert(FoundIdx < Universe && "Invalid key in set. Did object mutate?");
211  if (Idx == FoundIdx)
212  return begin() + i;
213  // Stride is 0 when SparseT >= unsigned. We don't need to loop.
214  if (!Stride)
215  break;
216  }
217  return end();
218  }
219 
220  /// find - Find an element by its key.
221  ///
222  /// @param Key A valid key to find.
223  /// @returns An iterator to the element identified by key, or end().
224  ///
225  iterator find(const KeyT &Key) {
226  return findIndex(KeyIndexOf(Key));
227  }
228 
229  const_iterator find(const KeyT &Key) const {
230  return const_cast<SparseSet*>(this)->findIndex(KeyIndexOf(Key));
231  }
232 
233  /// count - Returns 1 if this set contains an element identified by Key,
234  /// 0 otherwise.
235  ///
236  size_type count(const KeyT &Key) const {
237  return find(Key) == end() ? 0 : 1;
238  }
239 
240  /// insert - Attempts to insert a new element.
241  ///
242  /// If Val is successfully inserted, return (I, true), where I is an iterator
243  /// pointing to the newly inserted element.
244  ///
245  /// If the set already contains an element with the same key as Val, return
246  /// (I, false), where I is an iterator pointing to the existing element.
247  ///
248  /// Insertion invalidates all iterators.
249  ///
250  std::pair<iterator, bool> insert(const ValueT &Val) {
251  unsigned Idx = ValIndexOf(Val);
252  iterator I = findIndex(Idx);
253  if (I != end())
254  return std::make_pair(I, false);
255  Sparse[Idx] = size();
256  Dense.push_back(Val);
257  return std::make_pair(end() - 1, true);
258  }
259 
260  /// array subscript - If an element already exists with this key, return it.
261  /// Otherwise, automatically construct a new value from Key, insert it,
262  /// and return the newly inserted element.
263  ValueT &operator[](const KeyT &Key) {
264  return *insert(ValueT(Key)).first;
265  }
266 
268  // Sparse does not need to be cleared, see find().
269  return Dense.pop_back_val();
270  }
271 
272  /// erase - Erases an existing element identified by a valid iterator.
273  ///
274  /// This invalidates all iterators, but erase() returns an iterator pointing
275  /// to the next element. This makes it possible to erase selected elements
276  /// while iterating over the set:
277  ///
278  /// for (SparseSet::iterator I = Set.begin(); I != Set.end();)
279  /// if (test(*I))
280  /// I = Set.erase(I);
281  /// else
282  /// ++I;
283  ///
284  /// Note that end() changes when elements are erased, unlike std::list.
285  ///
287  assert(unsigned(I - begin()) < size() && "Invalid iterator");
288  if (I != end() - 1) {
289  *I = Dense.back();
290  unsigned BackIdx = ValIndexOf(Dense.back());
291  assert(BackIdx < Universe && "Invalid key in set. Did object mutate?");
292  Sparse[BackIdx] = I - begin();
293  }
294  // This depends on SmallVector::pop_back() not invalidating iterators.
295  // std::vector::pop_back() doesn't give that guarantee.
296  Dense.pop_back();
297  return I;
298  }
299 
300  /// erase - Erases an element identified by Key, if it exists.
301  ///
302  /// @param Key The key identifying the element to erase.
303  /// @returns True when an element was erased, false if no element was found.
304  ///
305  bool erase(const KeyT &Key) {
306  iterator I = find(Key);
307  if (I == end())
308  return false;
309  erase(I);
310  return true;
311  }
312 };
313 
314 } // end namespace llvm
315 
316 #endif // LLVM_ADT_SPARSESET_H
ValueT & operator[](const KeyT &Key)
array subscript - If an element already exists with this key, return it.
Definition: SparseSet.h:263
const_iterator end(StringRef path)
Get end iterator over path.
Definition: Path.cpp:259
GCNRegPressure max(const GCNRegPressure &P1, const GCNRegPressure &P2)
const_iterator begin(StringRef path, Style style=Style::native)
Get begin iterator over path.
Definition: Path.cpp:250
typename SuperClass::const_iterator const_iterator
Definition: SmallVector.h:328
iterator end()
Definition: SparseSet.h:178
This class represents lattice values for constants.
Definition: AllocatorList.h:24
std::pair< iterator, bool > insert(const ValueT &Val)
insert - Attempts to insert a new element.
Definition: SparseSet.h:250
bool erase(const KeyT &Key)
erase - Erases an element identified by Key, if it exists.
Definition: SparseSet.h:305
void push_back(const T &Elt)
Definition: SmallVector.h:218
SparseSetValTraits - Objects in a SparseSet are identified by keys that can be uniquely converted to ...
Definition: SparseSet.h:57
This file defines the MallocAllocator and BumpPtrAllocator interfaces.
size_type count(const KeyT &Key) const
count - Returns 1 if this set contains an element identified by Key, 0 otherwise. ...
Definition: SparseSet.h:236
Key
PAL metadata keys.
size_type size() const
size - Returns the number of elements in the set.
Definition: SparseSet.h:191
iterator findIndex(unsigned Idx)
findIndex - Find an element by its index.
Definition: SparseSet.h:205
iterator erase(iterator I)
erase - Erases an existing element identified by a valid iterator.
Definition: SparseSet.h:286
bool empty() const
empty - Returns true if the set is empty.
Definition: SparseSet.h:184
typename DenseT::const_iterator const_iterator
Definition: SparseSet.h:173
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator begin()
Definition: SmallVector.h:129
unsigned operator()(const ValueT &Val) const
Definition: SparseSet.h:69
void setUniverse(unsigned U)
setUniverse - Set the universe size which determines the largest key the set can hold.
Definition: SparseSet.h:156
size_t size() const
Definition: SmallVector.h:53
auto find(R &&Range, const T &Val) -> decltype(adl_begin(Range))
Provide wrappers to std::find which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:1201
const_iterator end() const
Definition: SparseSet.h:176
constexpr bool empty(const T &RangeOrContainer)
Test whether RangeOrContainer is empty. Similar to C++17 std::empty.
Definition: STLExtras.h:204
auto size(R &&Range, typename std::enable_if< std::is_same< typename std::iterator_traits< decltype(Range.begin())>::iterator_category, std::random_access_iterator_tag >::value, void >::type *=nullptr) -> decltype(std::distance(Range.begin(), Range.end()))
Get the size of a range.
Definition: STLExtras.h:1161
iterator begin()
Definition: SparseSet.h:177
const_iterator begin() const
Definition: SparseSet.h:175
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:381
typename SuperClass::iterator iterator
Definition: SmallVector.h:327
SparseSetValFunctor - Helper class for selecting SparseSetValTraits.
Definition: SparseSet.h:68
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator end()
Definition: SmallVector.h:133
SparseSet - Fast set implmentation for objects that can be identified by small unsigned keys...
Definition: SparseSet.h:124
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:56
iterator find(const KeyT &Key)
find - Find an element by its key.
Definition: SparseSet.h:225
#define I(x, y, z)
Definition: MD5.cpp:58
const_iterator find(const KeyT &Key) const
Definition: SparseSet.h:229
void clear()
clear - Clears the set.
Definition: SparseSet.h:195
unsigned operator()(const KeyT &Key) const
Definition: SparseSet.h:78
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
LLVM_ATTRIBUTE_RETURNS_NONNULL void * safe_calloc(size_t Count, size_t Sz)
Definition: MemAlloc.h:33
ValueT pop_back_val()
Definition: SparseSet.h:267
static unsigned getValIndex(const ValueT &Val)
Definition: SparseSet.h:58