diff options
Diffstat (limited to 'absl/container/internal/btree.h')
-rw-r--r-- | absl/container/internal/btree.h | 472 |
1 files changed, 234 insertions, 238 deletions
diff --git a/absl/container/internal/btree.h b/absl/container/internal/btree.h index 002ccc1e..0bb38366 100644 --- a/absl/container/internal/btree.h +++ b/absl/container/internal/btree.h @@ -182,15 +182,44 @@ struct key_compare_to_adapter<std::greater<absl::Cord>> { using type = StringBtreeDefaultGreater; }; +// Detects an 'absl_btree_prefer_linear_node_search' member. This is +// a protocol used as an opt-in or opt-out of linear search. +// +// For example, this would be useful for key types that wrap an integer +// and define their own cheap operator<(). For example: +// +// class K { +// public: +// using absl_btree_prefer_linear_node_search = std::true_type; +// ... +// private: +// friend bool operator<(K a, K b) { return a.k_ < b.k_; } +// int k_; +// }; +// +// btree_map<K, V> m; // Uses linear search +// +// If T has the preference tag, then it has a preference. +// Btree will use the tag's truth value. +template <typename T, typename = void> +struct has_linear_node_search_preference : std::false_type {}; +template <typename T, typename = void> +struct prefers_linear_node_search : std::false_type {}; +template <typename T> +struct has_linear_node_search_preference< + T, absl::void_t<typename T::absl_btree_prefer_linear_node_search>> + : std::true_type {}; +template <typename T> +struct prefers_linear_node_search< + T, absl::void_t<typename T::absl_btree_prefer_linear_node_search>> + : T::absl_btree_prefer_linear_node_search {}; + template <typename Key, typename Compare, typename Alloc, int TargetNodeSize, bool Multi, typename SlotPolicy> struct common_params { // If Compare is a common comparator for a string-like type, then we adapt it // to use heterogeneous lookup and to be a key-compare-to comparator. using key_compare = typename key_compare_to_adapter<Compare>::type; - // True when key_compare has been adapted to StringBtreeDefault{Less,Greater}. - using is_key_compare_adapted = - absl::negation<std::is_same<key_compare, Compare>>; // A type which indicates if we have a key-compare-to functor or a plain old // key-compare functor. using is_key_compare_to = btree_is_key_compare_to<key_compare, Key>; @@ -200,9 +229,6 @@ struct common_params { using size_type = std::make_signed<size_t>::type; using difference_type = ptrdiff_t; - // True if this is a multiset or multimap. - using is_multi_container = std::integral_constant<bool, Multi>; - using slot_policy = SlotPolicy; using slot_type = typename slot_policy::slot_type; using value_type = typename slot_policy::value_type; @@ -212,6 +238,23 @@ struct common_params { using reference = value_type &; using const_reference = const value_type &; + // For the given lookup key type, returns whether we can have multiple + // equivalent keys in the btree. If this is a multi-container, then we can. + // Otherwise, we can have multiple equivalent keys only if all of the + // following conditions are met: + // - The comparator is transparent. + // - The lookup key type is not the same as key_type. + // - The comparator is not a StringBtreeDefault{Less,Greater} comparator + // that we know has the same equivalence classes for all lookup types. + template <typename LookupKey> + constexpr static bool can_have_multiple_equivalent_keys() { + return Multi || + (IsTransparent<key_compare>::value && + !std::is_same<LookupKey, Key>::value && + !std::is_same<key_compare, StringBtreeDefaultLess>::value && + !std::is_same<key_compare, StringBtreeDefaultGreater>::value); + } + enum { kTargetNodeSize = TargetNodeSize, @@ -391,6 +434,10 @@ struct SearchResult { // useful information. template <typename V> struct SearchResult<V, false> { + SearchResult() {} + explicit SearchResult(V value) : value(value) {} + SearchResult(V value, MatchKind /*match*/) : value(value) {} + V value; static constexpr bool HasMatch() { return false; } @@ -403,7 +450,6 @@ struct SearchResult<V, false> { template <typename Params> class btree_node { using is_key_compare_to = typename Params::is_key_compare_to; - using is_multi_container = typename Params::is_multi_container; using field_type = typename Params::node_count_type; using allocator_type = typename Params::allocator_type; using slot_type = typename Params::slot_type; @@ -421,15 +467,22 @@ class btree_node { using difference_type = typename Params::difference_type; // Btree decides whether to use linear node search as follows: + // - If the comparator expresses a preference, use that. + // - If the key expresses a preference, use that. // - If the key is arithmetic and the comparator is std::less or // std::greater, choose linear. // - Otherwise, choose binary. // TODO(ezb): Might make sense to add condition(s) based on node-size. using use_linear_search = std::integral_constant< bool, - std::is_arithmetic<key_type>::value && - (std::is_same<std::less<key_type>, key_compare>::value || - std::is_same<std::greater<key_type>, key_compare>::value)>; + has_linear_node_search_preference<key_compare>::value + ? prefers_linear_node_search<key_compare>::value + : has_linear_node_search_preference<key_type>::value + ? prefers_linear_node_search<key_type>::value + : std::is_arithmetic<key_type>::value && + (std::is_same<std::less<key_type>, key_compare>::value || + std::is_same<std::greater<key_type>, + key_compare>::value)>; // This class is organized by gtl::Layout as if it had the following // structure: @@ -446,23 +499,23 @@ class btree_node { // // is the same as the count of values. // field_type finish; // // The maximum number of values the node can hold. This is an integer in - // // [1, kNodeValues] for root leaf nodes, kNodeValues for non-root leaf + // // [1, kNodeSlots] for root leaf nodes, kNodeSlots for non-root leaf // // nodes, and kInternalNodeMaxCount (as a sentinel value) for internal - // // nodes (even though there are still kNodeValues values in the node). + // // nodes (even though there are still kNodeSlots values in the node). // // TODO(ezb): make max_count use only 4 bits and record log2(capacity) // // to free extra bits for is_root, etc. // field_type max_count; // // // The array of values. The capacity is `max_count` for leaf nodes and - // // kNodeValues for internal nodes. Only the values in + // // kNodeSlots for internal nodes. Only the values in // // [start, finish) have been initialized and are valid. // slot_type values[max_count]; // // // The array of child pointers. The keys in children[i] are all less // // than key(i). The keys in children[i + 1] are all greater than key(i). - // // There are 0 children for leaf nodes and kNodeValues + 1 children for + // // There are 0 children for leaf nodes and kNodeSlots + 1 children for // // internal nodes. - // btree_node *children[kNodeValues + 1]; + // btree_node *children[kNodeSlots + 1]; // // This class is only constructed by EmptyNodeType. Normally, pointers to the // layout above are allocated, cast to btree_node*, and de-allocated within @@ -484,57 +537,62 @@ class btree_node { private: using layout_type = absl::container_internal::Layout<btree_node *, field_type, slot_type, btree_node *>; - constexpr static size_type SizeWithNValues(size_type n) { + constexpr static size_type SizeWithNSlots(size_type n) { return layout_type(/*parent*/ 1, /*position, start, finish, max_count*/ 4, - /*values*/ n, + /*slots*/ n, /*children*/ 0) .AllocSize(); } // A lower bound for the overhead of fields other than values in a leaf node. constexpr static size_type MinimumOverhead() { - return SizeWithNValues(1) - sizeof(value_type); + return SizeWithNSlots(1) - sizeof(value_type); } // Compute how many values we can fit onto a leaf node taking into account // padding. - constexpr static size_type NodeTargetValues(const int begin, const int end) { + constexpr static size_type NodeTargetSlots(const int begin, const int end) { return begin == end ? begin - : SizeWithNValues((begin + end) / 2 + 1) > + : SizeWithNSlots((begin + end) / 2 + 1) > params_type::kTargetNodeSize - ? NodeTargetValues(begin, (begin + end) / 2) - : NodeTargetValues((begin + end) / 2 + 1, end); + ? NodeTargetSlots(begin, (begin + end) / 2) + : NodeTargetSlots((begin + end) / 2 + 1, end); } enum { kTargetNodeSize = params_type::kTargetNodeSize, - kNodeTargetValues = NodeTargetValues(0, params_type::kTargetNodeSize), + kNodeTargetSlots = NodeTargetSlots(0, params_type::kTargetNodeSize), - // We need a minimum of 3 values per internal node in order to perform + // We need a minimum of 3 slots per internal node in order to perform // splitting (1 value for the two nodes involved in the split and 1 value - // propagated to the parent as the delimiter for the split). - kNodeValues = kNodeTargetValues >= 3 ? kNodeTargetValues : 3, + // propagated to the parent as the delimiter for the split). For performance + // reasons, we don't allow 3 slots-per-node due to bad worst case occupancy + // of 1/3 (for a node, not a b-tree). + kMinNodeSlots = 4, + + kNodeSlots = + kNodeTargetSlots >= kMinNodeSlots ? kNodeTargetSlots : kMinNodeSlots, // The node is internal (i.e. is not a leaf node) if and only if `max_count` // has this value. kInternalNodeMaxCount = 0, }; - // Leaves can have less than kNodeValues values. - constexpr static layout_type LeafLayout(const int max_values = kNodeValues) { + // Leaves can have less than kNodeSlots values. + constexpr static layout_type LeafLayout(const int slot_count = kNodeSlots) { return layout_type(/*parent*/ 1, /*position, start, finish, max_count*/ 4, - /*values*/ max_values, + /*slots*/ slot_count, /*children*/ 0); } constexpr static layout_type InternalLayout() { return layout_type(/*parent*/ 1, /*position, start, finish, max_count*/ 4, - /*values*/ kNodeValues, - /*children*/ kNodeValues + 1); + /*slots*/ kNodeSlots, + /*children*/ kNodeSlots + 1); } - constexpr static size_type LeafSize(const int max_values = kNodeValues) { - return LeafLayout(max_values).AllocSize(); + constexpr static size_type LeafSize(const int slot_count = kNodeSlots) { + return LeafLayout(slot_count).AllocSize(); } constexpr static size_type InternalSize() { return InternalLayout().AllocSize(); @@ -591,10 +649,10 @@ class btree_node { } field_type max_count() const { // Internal nodes have max_count==kInternalNodeMaxCount. - // Leaf nodes have max_count in [1, kNodeValues]. + // Leaf nodes have max_count in [1, kNodeSlots]. const field_type max_count = GetField<1>()[3]; return max_count == field_type{kInternalNodeMaxCount} - ? field_type{kNodeValues} + ? field_type{kNodeSlots} : max_count; } @@ -672,7 +730,7 @@ class btree_node { } ++s; } - return {s}; + return SearchResult<int, false>{s}; } // Returns the position of the first value whose key is not less than k using @@ -707,7 +765,7 @@ class btree_node { e = mid; } } - return {s}; + return SearchResult<int, false>{s}; } // Returns the position of the first value whose key is not less than k using @@ -716,7 +774,7 @@ class btree_node { SearchResult<int, true> binary_search_impl( const K &k, int s, int e, const CompareTo &comp, std::true_type /* IsCompareTo */) const { - if (is_multi_container::value) { + if (params_type::template can_have_multiple_equivalent_keys<K>()) { MatchKind exact_match = MatchKind::kNe; while (s != e) { const int mid = (s + e) >> 1; @@ -727,14 +785,14 @@ class btree_node { e = mid; if (c == 0) { // Need to return the first value whose key is not less than k, - // which requires continuing the binary search if this is a - // multi-container. + // which requires continuing the binary search if there could be + // multiple equivalent keys. exact_match = MatchKind::kEq; } } } return {s, exact_match}; - } else { // Not a multi-container. + } else { // Can't have multiple equivalent keys. while (s != e) { const int mid = (s + e) >> 1; const absl::weak_ordering c = comp(key(mid), k); @@ -784,12 +842,12 @@ class btree_node { start_slot(), max_count * sizeof(slot_type)); } void init_internal(btree_node *parent) { - init_leaf(parent, kNodeValues); + init_leaf(parent, kNodeSlots); // Set `max_count` to a sentinel value to indicate that this node is // internal. set_max_count(kInternalNodeMaxCount); absl::container_internal::SanitizerPoisonMemoryRegion( - &mutable_child(start()), (kNodeValues + 1) * sizeof(btree_node *)); + &mutable_child(start()), (kNodeSlots + 1) * sizeof(btree_node *)); } static void deallocate(const size_type size, btree_node *node, @@ -800,12 +858,6 @@ class btree_node { // Deletes a node and all of its children. static void clear_and_delete(btree_node *node, allocator_type *alloc); - public: - // Exposed only for tests. - static bool testonly_uses_linear_node_search() { - return use_linear_search::value; - } - private: template <typename... Args> void value_init(const field_type i, allocator_type *alloc, Args &&... args) { @@ -873,6 +925,7 @@ struct btree_iterator { using key_type = typename Node::key_type; using size_type = typename Node::size_type; using params_type = typename Node::params_type; + using is_map_container = typename params_type::is_map_container; using node_type = Node; using normal_node = typename std::remove_const<Node>::type; @@ -884,7 +937,7 @@ struct btree_iterator { using slot_type = typename params_type::slot_type; using iterator = - btree_iterator<normal_node, normal_reference, normal_pointer>; + btree_iterator<normal_node, normal_reference, normal_pointer>; using const_iterator = btree_iterator<const_node, const_reference, const_pointer>; @@ -901,20 +954,19 @@ struct btree_iterator { btree_iterator(Node *n, int p) : node(n), position(p) {} // NOTE: this SFINAE allows for implicit conversions from iterator to - // const_iterator, but it specifically avoids defining copy constructors so - // that btree_iterator can be trivially copyable. This is for performance and - // binary size reasons. + // const_iterator, but it specifically avoids hiding the copy constructor so + // that the trivial one will be used when possible. template <typename N, typename R, typename P, absl::enable_if_t< std::is_same<btree_iterator<N, R, P>, iterator>::value && std::is_same<btree_iterator, const_iterator>::value, int> = 0> - btree_iterator(const btree_iterator<N, R, P> &other) // NOLINT + btree_iterator(const btree_iterator<N, R, P> other) // NOLINT : node(other.node), position(other.position) {} private: // This SFINAE allows explicit conversions from const_iterator to - // iterator, but also avoids defining a copy constructor. + // iterator, but also avoids hiding the copy constructor. // NOTE: the const_cast is safe because this constructor is only called by // non-const methods and the container owns the nodes. template <typename N, typename R, typename P, @@ -922,7 +974,7 @@ struct btree_iterator { std::is_same<btree_iterator<N, R, P>, const_iterator>::value && std::is_same<btree_iterator, iterator>::value, int> = 0> - explicit btree_iterator(const btree_iterator<N, R, P> &other) + explicit btree_iterator(const btree_iterator<N, R, P> other) : node(const_cast<node_type *>(other.node)), position(other.position) {} // Increment/decrement the iterator. @@ -985,6 +1037,8 @@ struct btree_iterator { } private: + friend iterator; + friend const_iterator; template <typename Params> friend class btree; template <typename Tree> @@ -995,8 +1049,6 @@ struct btree_iterator { friend class btree_map_container; template <typename Tree> friend class btree_multiset_container; - template <typename N, typename R, typename P> - friend struct btree_iterator; template <typename TreeType, typename CheckerType> friend class base_checker; @@ -1017,8 +1069,6 @@ class btree { using is_key_compare_to = typename Params::is_key_compare_to; using init_type = typename Params::init_type; using field_type = typename node_type::field_type; - using is_multi_container = typename Params::is_multi_container; - using is_key_compare_adapted = typename Params::is_key_compare_adapted; // We use a static empty node for the root/leftmost/rightmost of empty btrees // in order to avoid branching in begin()/end(). @@ -1054,8 +1104,8 @@ class btree { } enum : uint32_t { - kNodeValues = node_type::kNodeValues, - kMinNodeValues = kNodeValues / 2, + kNodeSlots = node_type::kNodeSlots, + kMinNodeValues = kNodeSlots / 2, }; struct node_stats { @@ -1085,7 +1135,8 @@ class btree { using const_reference = typename Params::const_reference; using pointer = typename Params::pointer; using const_pointer = typename Params::const_pointer; - using iterator = btree_iterator<node_type, reference, pointer>; + using iterator = + typename btree_iterator<node_type, reference, pointer>::iterator; using const_iterator = typename iterator::const_iterator; using reverse_iterator = std::reverse_iterator<iterator>; using const_reverse_iterator = std::reverse_iterator<const_iterator>; @@ -1098,28 +1149,46 @@ class btree { private: // For use in copy_or_move_values_in_order. const value_type &maybe_move_from_iterator(const_iterator it) { return *it; } - value_type &&maybe_move_from_iterator(iterator it) { return std::move(*it); } + value_type &&maybe_move_from_iterator(iterator it) { + // This is a destructive operation on the other container so it's safe for + // us to const_cast and move from the keys here even if it's a set. + return std::move(const_cast<value_type &>(*it)); + } // Copies or moves (depending on the template parameter) the values in // other into this btree in their order in other. This btree must be empty // before this method is called. This method is used in copy construction, // copy assignment, and move assignment. template <typename Btree> - void copy_or_move_values_in_order(Btree *other); + void copy_or_move_values_in_order(Btree &other); // Validates that various assumptions/requirements are true at compile time. constexpr static bool static_assert_validation(); public: - btree(const key_compare &comp, const allocator_type &alloc); + btree(const key_compare &comp, const allocator_type &alloc) + : root_(comp, alloc, EmptyNode()), rightmost_(EmptyNode()), size_(0) {} - btree(const btree &other); + btree(const btree &other) : btree(other, other.allocator()) {} + btree(const btree &other, const allocator_type &alloc) + : btree(other.key_comp(), alloc) { + copy_or_move_values_in_order(other); + } btree(btree &&other) noexcept : root_(std::move(other.root_)), rightmost_(absl::exchange(other.rightmost_, EmptyNode())), size_(absl::exchange(other.size_, 0)) { other.mutable_root() = EmptyNode(); } + btree(btree &&other, const allocator_type &alloc) + : btree(other.key_comp(), alloc) { + if (alloc == other.allocator()) { + swap(other); + } else { + // Move values from `other` one at a time when allocators are different. + copy_or_move_values_in_order(other); + } + } ~btree() { // Put static_asserts in destructor to avoid triggering them before the type @@ -1147,17 +1216,22 @@ class btree { return const_reverse_iterator(begin()); } - // Finds the first element whose key is not less than key. + // Finds the first element whose key is not less than `key`. template <typename K> iterator lower_bound(const K &key) { - return internal_end(internal_lower_bound(key)); + return internal_end(internal_lower_bound(key).value); } template <typename K> const_iterator lower_bound(const K &key) const { - return internal_end(internal_lower_bound(key)); + return internal_end(internal_lower_bound(key).value); } - // Finds the first element whose key is greater than key. + // Finds the first element whose key is not less than `key` and also returns + // whether that element is equal to `key`. + template <typename K> + std::pair<iterator, bool> lower_bound_equal(const K &key) const; + + // Finds the first element whose key is greater than `key`. template <typename K> iterator upper_bound(const K &key) { return internal_end(internal_upper_bound(key)); @@ -1239,18 +1313,8 @@ class btree { // to the element after the last erased element. std::pair<size_type, iterator> erase_range(iterator begin, iterator end); - // Erases the specified key from the btree. Returns 1 if an element was - // erased and 0 otherwise. - template <typename K> - size_type erase_unique(const K &key); - - // Erases all of the entries matching the specified key from the - // btree. Returns the number of elements erased. - template <typename K> - size_type erase_multi(const K &key); - - // Finds the iterator corresponding to a key or returns end() if the key is - // not present. + // Finds an element with key equivalent to `key` or returns `end()` if `key` + // is not present. template <typename K> iterator find(const K &key) { return internal_end(internal_find(key)); @@ -1260,23 +1324,6 @@ class btree { return internal_end(internal_find(key)); } - // Returns a count of the number of times the key appears in the btree. - template <typename K> - size_type count_unique(const K &key) const { - const iterator begin = internal_find(key); - if (begin.node == nullptr) { - // The key doesn't exist in the tree. - return 0; - } - return 1; - } - // Returns a count of the number of times the key appears in the btree. - template <typename K> - size_type count_multi(const K &key) const { - const auto range = equal_range(key); - return std::distance(range.first, range.second); - } - // Clear the btree, deleting all of the values it contains. void clear(); @@ -1339,12 +1386,14 @@ class btree { } } - // The average number of bytes used per value stored in the btree. + // The average number of bytes used per value stored in the btree assuming + // random insertion order. static double average_bytes_per_value() { - // Returns the number of bytes per value on a leaf node that is 75% - // full. Experimentally, this matches up nicely with the computed number of - // bytes per value in trees that had their values inserted in random order. - return node_type::LeafSize() / (kNodeValues * 0.75); + // The expected number of values per node with random insertion order is the + // average of the maximum and minimum numbers of values per node. + const double expected_values_per_node = + (kNodeSlots + kMinNodeValues) / 2.0; + return node_type::LeafSize() / expected_values_per_node; } // The fullness of the btree. Computed as the number of elements in the btree @@ -1354,7 +1403,7 @@ class btree { // Returns 0 for empty trees. double fullness() const { if (empty()) return 0.0; - return static_cast<double>(size()) / (nodes() * kNodeValues); + return static_cast<double>(size()) / (nodes() * kNodeSlots); } // The overhead of the btree structure in bytes per node. Computed as the // total number of bytes used by the btree minus the number of bytes used for @@ -1404,7 +1453,7 @@ class btree { } node_type *new_leaf_node(node_type *parent) { node_type *n = allocate(node_type::LeafSize()); - n->init_leaf(parent, kNodeValues); + n->init_leaf(parent, kNodeSlots); return n; } node_type *new_leaf_root_node(const int max_count) { @@ -1453,28 +1502,19 @@ class btree { static IterType internal_last(IterType iter); // Returns an iterator pointing to the leaf position at which key would - // reside in the tree. We provide 2 versions of internal_locate. The first - // version uses a less-than comparator and is incapable of distinguishing when - // there is an exact match. The second version is for the key-compare-to - // specialization and distinguishes exact matches. The key-compare-to - // specialization allows the caller to avoid a subsequent comparison to - // determine if an exact match was made, which is important for keys with - // expensive comparison, such as strings. + // reside in the tree, unless there is an exact match - in which case, the + // result may not be on a leaf. When there's a three-way comparator, we can + // return whether there was an exact match. This allows the caller to avoid a + // subsequent comparison to determine if an exact match was made, which is + // important for keys with expensive comparison, such as strings. template <typename K> SearchResult<iterator, is_key_compare_to::value> internal_locate( const K &key) const; - template <typename K> - SearchResult<iterator, false> internal_locate_impl( - const K &key, std::false_type /* IsCompareTo */) const; - - template <typename K> - SearchResult<iterator, true> internal_locate_impl( - const K &key, std::true_type /* IsCompareTo */) const; - // Internal routine which implements lower_bound(). template <typename K> - iterator internal_lower_bound(const K &key) const; + SearchResult<iterator, is_key_compare_to::value> internal_lower_bound( + const K &key) const; // Internal routine which implements upper_bound(). template <typename K> @@ -1503,13 +1543,6 @@ class btree { return res; } - public: - // Exposed only for tests. - static bool testonly_uses_linear_node_search() { - return node_type::testonly_uses_linear_node_search(); - } - - private: // We use compressed tuple in order to save space because key_compare and // allocator_type are usually empty. absl::container_internal::CompressedTuple<key_compare, allocator_type, @@ -1665,7 +1698,7 @@ template <typename P> void btree_node<P>::split(const int insert_position, btree_node *dest, allocator_type *alloc) { assert(dest->count() == 0); - assert(max_count() == kNodeValues); + assert(max_count() == kNodeSlots); // We bias the split based on the position being inserted. If we're // inserting at the beginning of the left node then bias the split to put @@ -1673,7 +1706,7 @@ void btree_node<P>::split(const int insert_position, btree_node *dest, // right node then bias the split to put more values on the left node. if (insert_position == start()) { dest->set_finish(dest->start() + finish() - 1); - } else if (insert_position == kNodeValues) { + } else if (insert_position == kNodeSlots) { dest->set_finish(dest->start()); } else { dest->set_finish(dest->start() + count() / 2); @@ -1744,7 +1777,7 @@ void btree_node<P>::clear_and_delete(btree_node *node, allocator_type *alloc) { // Navigate to the leftmost leaf under node, and then delete upwards. while (!node->leaf()) node = node->start_child(); - // Use `int` because `pos` needs to be able to hold `kNodeValues+1`, which + // Use `int` because `pos` needs to be able to hold `kNodeSlots+1`, which // isn't guaranteed to be a valid `field_type`. int pos = node->position(); btree_node *parent = node->parent(); @@ -1832,7 +1865,7 @@ void btree_iterator<N, R, P>::decrement_slow() { // btree methods template <typename P> template <typename Btree> -void btree<P>::copy_or_move_values_in_order(Btree *other) { +void btree<P>::copy_or_move_values_in_order(Btree &other) { static_assert(std::is_same<btree, Btree>::value || std::is_same<const btree, Btree>::value, "Btree type must be same or const."); @@ -1840,11 +1873,11 @@ void btree<P>::copy_or_move_values_in_order(Btree *other) { // We can avoid key comparisons because we know the order of the // values is the same order we'll store them in. - auto iter = other->begin(); - if (iter == other->end()) return; + auto iter = other.begin(); + if (iter == other.end()) return; insert_multi(maybe_move_from_iterator(iter)); ++iter; - for (; iter != other->end(); ++iter) { + for (; iter != other.end(); ++iter) { // If the btree is not empty, we can just insert the new value at the end // of the tree. internal_emplace(end(), maybe_move_from_iterator(iter)); @@ -1863,7 +1896,7 @@ constexpr bool btree<P>::static_assert_validation() { // Note: We assert that kTargetValues, which is computed from // Params::kTargetNodeSize, must fit the node_type::field_type. static_assert( - kNodeValues < (1 << (8 * sizeof(typename node_type::field_type))), + kNodeSlots < (1 << (8 * sizeof(typename node_type::field_type))), "target node size too large"); // Verify that key_compare returns an absl::{weak,strong}_ordering or bool. @@ -1883,31 +1916,29 @@ constexpr bool btree<P>::static_assert_validation() { } template <typename P> -btree<P>::btree(const key_compare &comp, const allocator_type &alloc) - : root_(comp, alloc, EmptyNode()), rightmost_(EmptyNode()), size_(0) {} - -template <typename P> -btree<P>::btree(const btree &other) - : btree(other.key_comp(), other.allocator()) { - copy_or_move_values_in_order(&other); +template <typename K> +auto btree<P>::lower_bound_equal(const K &key) const + -> std::pair<iterator, bool> { + const SearchResult<iterator, is_key_compare_to::value> res = + internal_lower_bound(key); + const iterator lower = iterator(internal_end(res.value)); + const bool equal = res.HasMatch() + ? res.IsEq() + : lower != end() && !compare_keys(key, lower.key()); + return {lower, equal}; } template <typename P> template <typename K> auto btree<P>::equal_range(const K &key) -> std::pair<iterator, iterator> { - const iterator lower = lower_bound(key); - // TODO(ezb): we should be able to avoid this comparison when there's a - // three-way comparator. - if (lower == end() || compare_keys(key, lower.key())) return {lower, lower}; + const std::pair<iterator, bool> lower_and_equal = lower_bound_equal(key); + const iterator lower = lower_and_equal.first; + if (!lower_and_equal.second) { + return {lower, lower}; + } const iterator next = std::next(lower); - // When the comparator is heterogeneous, we can't assume that comparison with - // non-`key_type` will be equivalent to `key_type` comparisons so there - // could be multiple equivalent keys even in a unique-container. But for - // heterogeneous comparisons from the default string adapted comparators, we - // don't need to worry about this. - if (!is_multi_container::value && - (std::is_same<K, key_type>::value || is_key_compare_adapted::value)) { + if (!params_type::template can_have_multiple_equivalent_keys<K>()) { // The next iterator after lower must point to a key greater than `key`. // Note: if this assert fails, then it may indicate that the comparator does // not meet the equivalence requirements for Compare @@ -1918,7 +1949,7 @@ auto btree<P>::equal_range(const K &key) -> std::pair<iterator, iterator> { // Try once more to avoid the call to upper_bound() if there's only one // equivalent key. This should prevent all calls to upper_bound() in cases of // unique-containers with heterogeneous comparators in which all comparison - // operators are equivalent. + // operators have the same equivalence classes. if (next == end() || compare_keys(key, next.key())) return {lower, next}; // In this case, we need to call upper_bound() to avoid worst case O(N) @@ -1934,8 +1965,8 @@ auto btree<P>::insert_unique(const K &key, Args &&... args) mutable_root() = rightmost_ = new_leaf_root_node(1); } - auto res = internal_locate(key); - iterator &iter = res.value; + SearchResult<iterator, is_key_compare_to::value> res = internal_locate(key); + iterator iter = res.value; if (res.HasMatch()) { if (res.IsEq()) { @@ -2049,7 +2080,7 @@ auto btree<P>::operator=(const btree &other) -> btree & { *mutable_allocator() = other.allocator(); } - copy_or_move_values_in_order(&other); + copy_or_move_values_in_order(other); } return *this; } @@ -2079,7 +2110,7 @@ auto btree<P>::operator=(btree &&other) noexcept -> btree & { // comparator while moving the values so we can't swap the key // comparators. *mutable_key_comp() = other.key_comp(); - copy_or_move_values_in_order(&other); + copy_or_move_values_in_order(other); } } } @@ -2203,31 +2234,6 @@ auto btree<P>::erase_range(iterator begin, iterator end) } template <typename P> -template <typename K> -auto btree<P>::erase_unique(const K &key) -> size_type { - const iterator iter = internal_find(key); - if (iter.node == nullptr) { - // The key doesn't exist in the tree, return nothing done. - return 0; - } - erase(iter); - return 1; -} - -template <typename P> -template <typename K> -auto btree<P>::erase_multi(const K &key) -> size_type { - const iterator begin = internal_lower_bound(key); - if (begin.node == nullptr) { - // The key doesn't exist in the tree, return nothing done. - return 0; - } - // Delete all of the keys between begin and upper_bound(key). - const iterator end = internal_end(internal_upper_bound(key)); - return erase_range(begin, end).first; -} - -template <typename P> void btree<P>::clear() { if (!empty()) { node_type::clear_and_delete(root(), mutable_allocator()); @@ -2271,7 +2277,7 @@ void btree<P>::rebalance_or_split(iterator *iter) { node_type *&node = iter->node; int &insert_position = iter->position; assert(node->count() == node->max_count()); - assert(kNodeValues == node->max_count()); + assert(kNodeSlots == node->max_count()); // First try to make room on the node by rebalancing. node_type *parent = node->parent(); @@ -2279,17 +2285,17 @@ void btree<P>::rebalance_or_split(iterator *iter) { if (node->position() > parent->start()) { // Try rebalancing with our left sibling. node_type *left = parent->child(node->position() - 1); - assert(left->max_count() == kNodeValues); - if (left->count() < kNodeValues) { + assert(left->max_count() == kNodeSlots); + if (left->count() < kNodeSlots) { // We bias rebalancing based on the position being inserted. If we're // inserting at the end of the right node then we bias rebalancing to // fill up the left node. - int to_move = (kNodeValues - left->count()) / - (1 + (insert_position < kNodeValues)); + int to_move = (kNodeSlots - left->count()) / + (1 + (insert_position < static_cast<int>(kNodeSlots))); to_move = (std::max)(1, to_move); if (insert_position - to_move >= node->start() || - left->count() + to_move < kNodeValues) { + left->count() + to_move < static_cast<int>(kNodeSlots)) { left->rebalance_right_to_left(to_move, node, mutable_allocator()); assert(node->max_count() - node->count() == to_move); @@ -2308,17 +2314,17 @@ void btree<P>::rebalance_or_split(iterator *iter) { if (node->position() < parent->finish()) { // Try rebalancing with our right sibling. node_type *right = parent->child(node->position() + 1); - assert(right->max_count() == kNodeValues); - if (right->count() < kNodeValues) { + assert(right->max_count() == kNodeSlots); + if (right->count() < kNodeSlots) { // We bias rebalancing based on the position being inserted. If we're // inserting at the beginning of the left node then we bias rebalancing // to fill up the right node. - int to_move = (kNodeValues - right->count()) / + int to_move = (static_cast<int>(kNodeSlots) - right->count()) / (1 + (insert_position > node->start())); to_move = (std::max)(1, to_move); if (insert_position <= node->finish() - to_move || - right->count() + to_move < kNodeValues) { + right->count() + to_move < static_cast<int>(kNodeSlots)) { node->rebalance_left_to_right(to_move, right, mutable_allocator()); if (insert_position > node->finish()) { @@ -2334,8 +2340,8 @@ void btree<P>::rebalance_or_split(iterator *iter) { // Rebalancing failed, make sure there is room on the parent node for a new // value. - assert(parent->max_count() == kNodeValues); - if (parent->count() == kNodeValues) { + assert(parent->max_count() == kNodeSlots); + if (parent->count() == kNodeSlots) { iterator parent_iter(node->parent(), node->position()); rebalance_or_split(&parent_iter); } @@ -2380,8 +2386,8 @@ bool btree<P>::try_merge_or_rebalance(iterator *iter) { if (iter->node->position() > parent->start()) { // Try merging with our left sibling. node_type *left = parent->child(iter->node->position() - 1); - assert(left->max_count() == kNodeValues); - if (1 + left->count() + iter->node->count() <= kNodeValues) { + assert(left->max_count() == kNodeSlots); + if (1U + left->count() + iter->node->count() <= kNodeSlots) { iter->position += 1 + left->count(); merge_nodes(left, iter->node); iter->node = left; @@ -2391,8 +2397,8 @@ bool btree<P>::try_merge_or_rebalance(iterator *iter) { if (iter->node->position() < parent->finish()) { // Try merging with our right sibling. node_type *right = parent->child(iter->node->position() + 1); - assert(right->max_count() == kNodeValues); - if (1 + iter->node->count() + right->count() <= kNodeValues) { + assert(right->max_count() == kNodeSlots); + if (1U + iter->node->count() + right->count() <= kNodeSlots) { merge_nodes(iter->node, right); return true; } @@ -2473,12 +2479,12 @@ inline auto btree<P>::internal_emplace(iterator iter, Args &&... args) allocator_type *alloc = mutable_allocator(); if (iter.node->count() == max_count) { // Make room in the leaf for the new item. - if (max_count < kNodeValues) { + if (max_count < kNodeSlots) { // Insertion into the root where the root is smaller than the full node // size. Simply grow the size of the root node. assert(iter.node == root()); iter.node = - new_leaf_root_node((std::min<int>)(kNodeValues, 2 * max_count)); + new_leaf_root_node((std::min<int>)(kNodeSlots, 2 * max_count)); // Transfer the values from the old root to the new root. node_type *old_root = root(); node_type *new_root = iter.node; @@ -2501,61 +2507,51 @@ template <typename P> template <typename K> inline auto btree<P>::internal_locate(const K &key) const -> SearchResult<iterator, is_key_compare_to::value> { - return internal_locate_impl(key, is_key_compare_to()); -} - -template <typename P> -template <typename K> -inline auto btree<P>::internal_locate_impl( - const K &key, std::false_type /* IsCompareTo */) const - -> SearchResult<iterator, false> { - iterator iter(const_cast<node_type *>(root())); - for (;;) { - iter.position = iter.node->lower_bound(key, key_comp()).value; - // NOTE: we don't need to walk all the way down the tree if the keys are - // equal, but determining equality would require doing an extra comparison - // on each node on the way down, and we will need to go all the way to the - // leaf node in the expected case. - if (iter.node->leaf()) { - break; - } - iter.node = iter.node->child(iter.position); - } - return {iter}; -} - -template <typename P> -template <typename K> -inline auto btree<P>::internal_locate_impl( - const K &key, std::true_type /* IsCompareTo */) const - -> SearchResult<iterator, true> { iterator iter(const_cast<node_type *>(root())); for (;;) { - SearchResult<int, true> res = iter.node->lower_bound(key, key_comp()); + SearchResult<int, is_key_compare_to::value> res = + iter.node->lower_bound(key, key_comp()); iter.position = res.value; - if (res.match == MatchKind::kEq) { + if (res.IsEq()) { return {iter, MatchKind::kEq}; } + // Note: in the non-key-compare-to case, we don't need to walk all the way + // down the tree if the keys are equal, but determining equality would + // require doing an extra comparison on each node on the way down, and we + // will need to go all the way to the leaf node in the expected case. if (iter.node->leaf()) { break; } iter.node = iter.node->child(iter.position); } + // Note: in the non-key-compare-to case, the key may actually be equivalent + // here (and the MatchKind::kNe is ignored). return {iter, MatchKind::kNe}; } template <typename P> template <typename K> -auto btree<P>::internal_lower_bound(const K &key) const -> iterator { +auto btree<P>::internal_lower_bound(const K &key) const + -> SearchResult<iterator, is_key_compare_to::value> { + if (!params_type::template can_have_multiple_equivalent_keys<K>()) { + SearchResult<iterator, is_key_compare_to::value> ret = internal_locate(key); + ret.value = internal_last(ret.value); + return ret; + } iterator iter(const_cast<node_type *>(root())); + SearchResult<int, is_key_compare_to::value> res; + bool seen_eq = false; for (;;) { - iter.position = iter.node->lower_bound(key, key_comp()).value; + res = iter.node->lower_bound(key, key_comp()); + iter.position = res.value; if (iter.node->leaf()) { break; } + seen_eq = seen_eq || res.IsEq(); iter.node = iter.node->child(iter.position); } - return internal_last(iter); + if (res.IsEq()) return {iter, MatchKind::kEq}; + return {internal_last(iter), seen_eq ? MatchKind::kEq : MatchKind::kNe}; } template <typename P> @@ -2575,7 +2571,7 @@ auto btree<P>::internal_upper_bound(const K &key) const -> iterator { template <typename P> template <typename K> auto btree<P>::internal_find(const K &key) const -> iterator { - auto res = internal_locate(key); + SearchResult<iterator, is_key_compare_to::value> res = internal_locate(key); if (res.HasMatch()) { if (res.IsEq()) { return res.value; |