From 2ab759a9b367399a8cc636e2bc561f1f51c41057 Mon Sep 17 00:00:00 2001 From: Yuan Date: Wed, 9 Feb 2022 19:43:52 +0800 Subject: [PATCH] [NSE-610] hashagg opt#1 (#715) * hashagg opt#1 instead of do agg per row, this patch changed the behavior to do one agg per column to achieve better cache efficency in my local test TPCH Q1 is improved 10% The optimize on keys and hashmap logic will be split into following patch Signed-off-by: Yuan Zhou * reserve space for keys in hashmap Signed-off-by: Yuan Zhou * s/sparsehash/parallelhashmap Signed-off-by: Yuan Zhou * using seperate code path for null/nonull batches Signed-off-by: Yuan Zhou * fix format Signed-off-by: Yuan Zhou * fix groupby with boolean Signed-off-by: Yuan Zhou --- .../ext/hash_aggregate_kernel.cc | 71 +- .../cpp/src/precompile/hash_map.cc | 4 +- .../cpp/src/precompile/unsafe_array.h | 2 +- .../src/third_party/parallel_hashmap/btree.h | 3914 +++++++++++++ .../third_party/parallel_hashmap/conanfile.py | 36 + .../third_party/parallel_hashmap/meminfo.h | 186 + .../src/third_party/parallel_hashmap/phmap.h | 4654 ++++++++++++++++ .../third_party/parallel_hashmap/phmap_base.h | 4956 +++++++++++++++++ .../third_party/parallel_hashmap/phmap_bits.h | 641 +++ .../parallel_hashmap/phmap_config.h | 760 +++ .../third_party/parallel_hashmap/phmap_dump.h | 256 + .../parallel_hashmap/phmap_fwd_decl.h | 157 + .../parallel_hashmap/phmap_utils.h | 347 ++ .../third_party/row_wise_memory/unsafe_row.h | 10 + .../third_party/sparsehash/sparse_hash_map.h | 39 +- 15 files changed, 15970 insertions(+), 63 deletions(-) create mode 100644 native-sql-engine/cpp/src/third_party/parallel_hashmap/btree.h create mode 100644 native-sql-engine/cpp/src/third_party/parallel_hashmap/conanfile.py create mode 100644 native-sql-engine/cpp/src/third_party/parallel_hashmap/meminfo.h create mode 100644 native-sql-engine/cpp/src/third_party/parallel_hashmap/phmap.h create mode 100644 native-sql-engine/cpp/src/third_party/parallel_hashmap/phmap_base.h create mode 100644 native-sql-engine/cpp/src/third_party/parallel_hashmap/phmap_bits.h create mode 100644 native-sql-engine/cpp/src/third_party/parallel_hashmap/phmap_config.h create mode 100644 native-sql-engine/cpp/src/third_party/parallel_hashmap/phmap_dump.h create mode 100644 native-sql-engine/cpp/src/third_party/parallel_hashmap/phmap_fwd_decl.h create mode 100644 native-sql-engine/cpp/src/third_party/parallel_hashmap/phmap_utils.h diff --git a/native-sql-engine/cpp/src/codegen/arrow_compute/ext/hash_aggregate_kernel.cc b/native-sql-engine/cpp/src/codegen/arrow_compute/ext/hash_aggregate_kernel.cc index 588ed9001..716b2191c 100644 --- a/native-sql-engine/cpp/src/codegen/arrow_compute/ext/hash_aggregate_kernel.cc +++ b/native-sql-engine/cpp/src/codegen/arrow_compute/ext/hash_aggregate_kernel.cc @@ -763,28 +763,29 @@ class HashAggregateKernel::Impl { std::vector indices; indices.resize(length, -1); - - for (int i = 0; i < length; i++) { - auto aggr_key = typed_key_in->GetView(i); - auto aggr_key_validity = - typed_key_in->null_count() == 0 ? true : !typed_key_in->IsNull(i); - - // 3. get key from hash_table - int memo_index = 0; - if (!aggr_key_validity) { - memo_index = aggr_hash_table_->GetOrInsertNull([](int) {}, [](int) {}); - } else { - aggr_hash_table_->GetOrInsert( - aggr_key, [](int) {}, [](int) {}, &memo_index); + if (typed_key_in->null_count() > 0) { + for (int i = 0; i < length; i++) { + auto aggr_key = typed_key_in->GetView(i); + auto aggr_key_validity = !typed_key_in->IsNull(i); + + if (aggr_key_validity) { + aggr_hash_table_->GetOrInsert( + aggr_key, [](int) {}, [](int) {}, &(indices[i])); + } else { + indices[i] = aggr_hash_table_->GetOrInsertNull([](int) {}, [](int) {}); + } } - - if (memo_index > max_group_id_) { - max_group_id_ = memo_index; + } else { + for (int i = 0; i < length; i++) { + auto aggr_key = typed_key_in->GetView(i); + aggr_hash_table_->GetOrInsert( + aggr_key, [](int) {}, [](int) {}, &(indices[i])); } - indices[i] = memo_index; } + max_group_id_ = aggr_hash_table_->size_ - 1; total_out_length_ = max_group_id_ + 1; + // 4. prepare action func and evaluate std::vector> eval_func_list; std::vector> eval_null_func_list; @@ -802,15 +803,9 @@ class HashAggregateKernel::Impl { eval_null_func_list.push_back(null_func); } - for (auto memo_index : indices) { - if (memo_index == -1) { - for (auto eval_func : eval_null_func_list) { - RETURN_NOT_OK(eval_func()); - } - } else { - for (auto eval_func : eval_func_list) { - RETURN_NOT_OK(eval_func(memo_index)); - } + for (auto eval_func : eval_func_list) { + for (auto memo_index : indices) { + RETURN_NOT_OK(eval_func(memo_index)); } } @@ -961,15 +956,9 @@ class HashAggregateKernel::Impl { eval_null_func_list.push_back(null_func); } - for (auto memo_index : indices) { - if (memo_index == -1) { - for (auto eval_func : eval_null_func_list) { - RETURN_NOT_OK(eval_func()); - } - } else { - for (auto eval_func : eval_func_list) { - RETURN_NOT_OK(eval_func(memo_index)); - } + for (auto eval_func : eval_func_list) { + for (auto memo_index : indices) { + RETURN_NOT_OK(eval_func(memo_index)); } } return arrow::Status::OK(); @@ -1115,15 +1104,9 @@ class HashAggregateKernel::Impl { eval_null_func_list.push_back(null_func); } - for (auto memo_index : indices) { - if (memo_index == -1) { - for (auto eval_func : eval_null_func_list) { - RETURN_NOT_OK(eval_func()); - } - } else { - for (auto eval_func : eval_func_list) { - RETURN_NOT_OK(eval_func(memo_index)); - } + for (auto eval_func : eval_func_list) { + for (auto memo_index : indices) { + RETURN_NOT_OK(eval_func(memo_index)); } } diff --git a/native-sql-engine/cpp/src/precompile/hash_map.cc b/native-sql-engine/cpp/src/precompile/hash_map.cc index 4b22bf09c..991c36b5c 100644 --- a/native-sql-engine/cpp/src/precompile/hash_map.cc +++ b/native-sql-engine/cpp/src/precompile/hash_map.cc @@ -53,7 +53,7 @@ namespace precompile { typename arrow::internal::HashTraits::MemoTableType; \ class HASHMAPNAME::Impl : public MEMOTABLETYPE { \ public: \ - Impl(arrow::MemoryPool* pool) : MEMOTABLETYPE(pool) {} \ + Impl(arrow::MemoryPool* pool) : MEMOTABLETYPE(pool, 128) {} \ }; \ \ HASHMAPNAME::HASHMAPNAME(arrow::MemoryPool* pool) { \ @@ -107,6 +107,6 @@ TYPED_ARROW_HASH_MAP_IMPL(StringHashMap, StringType, arrow::util::string_view, TYPED_ARROW_HASH_MAP_DECIMAL_IMPL(Decimal128HashMap, Decimal128Type, arrow::Decimal128, DecimalMemoTableType) #undef TYPED_ARROW_HASH_MAP_IMPL - +#undef TYPED_ARROW_HASH_MAP_DECIMAL_IMPL } // namespace precompile } // namespace sparkcolumnarplugin diff --git a/native-sql-engine/cpp/src/precompile/unsafe_array.h b/native-sql-engine/cpp/src/precompile/unsafe_array.h index 05b585854..42b0c2fe3 100644 --- a/native-sql-engine/cpp/src/precompile/unsafe_array.h +++ b/native-sql-engine/cpp/src/precompile/unsafe_array.h @@ -89,7 +89,7 @@ class TypedUnsafeArray> : public Unsaf if (!skip_null_check_ && typed_array_->IsNull(i)) { setNullAt((*unsafe_row).get(), idx_); } else { - auto v = typed_array_->GetString(i); + auto v = typed_array_->GetView(i); appendToUnsafeRow((*unsafe_row).get(), idx_, v); } return arrow::Status::OK(); diff --git a/native-sql-engine/cpp/src/third_party/parallel_hashmap/btree.h b/native-sql-engine/cpp/src/third_party/parallel_hashmap/btree.h new file mode 100644 index 000000000..b9b0d94da --- /dev/null +++ b/native-sql-engine/cpp/src/third_party/parallel_hashmap/btree.h @@ -0,0 +1,3914 @@ +// --------------------------------------------------------------------------- +// Copyright (c) 2019, Gregory Popovitch - greg7mdp@gmail.com +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// https://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +// +// Includes work from abseil-cpp (https://github.com/abseil/abseil-cpp) +// with modifications. +// +// Copyright 2018 The Abseil Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// https://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +// --------------------------------------------------------------------------- + +#ifndef PHMAP_BTREE_BTREE_CONTAINER_H_ +#define PHMAP_BTREE_BTREE_CONTAINER_H_ + +#ifdef _MSC_VER +#pragma warning(push) + +#pragma warning(disable : 4127) // conditional expression is constant +#pragma warning(disable : 4324) // structure was padded due to alignment specifier +#pragma warning(disable : 4355) // 'this': used in base member initializer list +#pragma warning(disable : 4365) // conversion from 'int' to 'const unsigned __int64', + // signed/unsigned mismatch +#pragma warning(disable : 4514) // unreferenced inline function has been removed +#pragma warning(disable : 4623) // default constructor was implicitly defined as deleted +#pragma warning(disable : 4625) // copy constructor was implicitly defined as deleted +#pragma warning(disable : 4626) // assignment operator was implicitly defined as deleted +#pragma warning(disable : 4710) // function not inlined +#pragma warning(disable : 4711) // selected for automatic inline expansion +#pragma warning(disable : 4820) // '6' bytes padding added after data member +#pragma warning(disable : 4868) // compiler may not enforce left-to-right evaluation + // order in braced initializer list +#pragma warning(disable : 5026) // move constructor was implicitly defined as deleted +#pragma warning( \ + disable : 5027) // move assignment operator was implicitly defined as deleted +#pragma warning(disable : 5045) // Compiler will insert Spectre mitigation for memory + // load if /Qspectre switch specified +#endif + +#include +#include +#include +#include +#include + +#include "phmap_base.h" +#include "phmap_fwd_decl.h" + +#if PHMAP_HAVE_STD_STRING_VIEW +#include +#endif + +// MSVC constructibility traits do not detect destructor properties and so our +// implementations should not use them as a source-of-truth. +#if defined(_MSC_VER) && !defined(__clang__) && !defined(__GNUC__) +#define PHMAP_META_INTERNAL_STD_CONSTRUCTION_TRAITS_DONT_CHECK_DESTRUCTION 1 +#endif + +namespace phmap { + +// Defined and documented later on in this file. +template +struct is_trivially_destructible; + +// Defined and documented later on in this file. +template +struct is_trivially_move_assignable; + +namespace type_traits_internal { + +// Silence MSVC warnings about the destructor being defined as deleted. +#if defined(_MSC_VER) && !defined(__GNUC__) +#pragma warning(push) +#pragma warning(disable : 4624) +#endif // defined(_MSC_VER) && !defined(__GNUC__) + +template +union SingleMemberUnion { + T t; +}; + +// Restore the state of the destructor warning that was silenced above. +#if defined(_MSC_VER) && !defined(__GNUC__) +#pragma warning(pop) +#endif // defined(_MSC_VER) && !defined(__GNUC__) + +template +struct IsTriviallyMoveConstructibleObject + : std::integral_constant< + bool, + std::is_move_constructible>::value && + phmap::is_trivially_destructible::value> {}; + +template +struct IsTriviallyCopyConstructibleObject + : std::integral_constant< + bool, + std::is_copy_constructible>::value && + phmap::is_trivially_destructible::value> {}; + +template +struct IsTriviallyMoveAssignableReference : std::false_type {}; + +template +struct IsTriviallyMoveAssignableReference + : phmap::is_trivially_move_assignable::type {}; + +template +struct IsTriviallyMoveAssignableReference + : phmap::is_trivially_move_assignable::type {}; + +} // namespace type_traits_internal + +template +using void_t = typename type_traits_internal::VoidTImpl::type; + +template +struct is_function + : std::integral_constant::value || + std::is_const::type>::value)> { +}; + +namespace type_traits_internal { + +template +class is_trivially_copyable_impl { + using ExtentsRemoved = typename std::remove_all_extents::type; + static constexpr bool kIsCopyOrMoveConstructible = + std::is_copy_constructible::value || + std::is_move_constructible::value; + static constexpr bool kIsCopyOrMoveAssignable = + phmap::is_copy_assignable::value || + phmap::is_move_assignable::value; + + public: + static constexpr bool kValue = + (__has_trivial_copy(ExtentsRemoved) || !kIsCopyOrMoveConstructible) && + (__has_trivial_assign(ExtentsRemoved) || !kIsCopyOrMoveAssignable) && + (kIsCopyOrMoveConstructible || kIsCopyOrMoveAssignable) && + is_trivially_destructible::value && + // We need to check for this explicitly because otherwise we'll say + // references are trivial copyable when compiled by MSVC. + !std::is_reference::value; +}; + +template +struct is_trivially_copyable + : std::integral_constant< + bool, type_traits_internal::is_trivially_copyable_impl::kValue> {}; +} // namespace type_traits_internal + +namespace swap_internal { + +// Necessary for the traits. +using std::swap; + +// This declaration prevents global `swap` and `phmap::swap` overloads from being +// considered unless ADL picks them up. +void swap(); + +template +using IsSwappableImpl = decltype(swap(std::declval(), std::declval())); + +// NOTE: This dance with the default template parameter is for MSVC. +template (), std::declval()))>> +using IsNothrowSwappableImpl = typename std::enable_if::type; + +template +struct IsSwappable : phmap::type_traits_internal::is_detected {}; + +template +struct IsNothrowSwappable + : phmap::type_traits_internal::is_detected {}; + +template ::value, int> = 0> +void Swap(T& lhs, T& rhs) noexcept(IsNothrowSwappable::value) { + swap(lhs, rhs); +} + +using StdSwapIsUnconstrained = IsSwappable; + +} // namespace swap_internal + +namespace type_traits_internal { + +// Make the swap-related traits/function accessible from this namespace. +using swap_internal::IsNothrowSwappable; +using swap_internal::IsSwappable; +using swap_internal::StdSwapIsUnconstrained; +using swap_internal::Swap; + +} // namespace type_traits_internal + +namespace compare_internal { + +using value_type = int8_t; + +template +struct Fail { + static_assert(sizeof(T) < 0, "Only literal `0` is allowed."); +}; + +template +struct OnlyLiteralZero { + constexpr OnlyLiteralZero(NullPtrT) noexcept {} // NOLINT + + template ::value || + (std::is_integral::value && + !std::is_same::value)>::type, + typename = typename Fail::type> + OnlyLiteralZero(T); // NOLINT +}; + +enum class eq : value_type { + equal = 0, + equivalent = equal, + nonequal = 1, + nonequivalent = nonequal, +}; + +enum class ord : value_type { less = -1, greater = 1 }; + +enum class ncmp : value_type { unordered = -127 }; + +#if defined(__cpp_inline_variables) && !defined(_MSC_VER) + +#define PHMAP_COMPARE_INLINE_BASECLASS_DECL(name) + +#define PHMAP_COMPARE_INLINE_SUBCLASS_DECL(type, name) static const type name; + +#define PHMAP_COMPARE_INLINE_INIT(type, name, init) inline constexpr type type::name(init) + +#else // __cpp_inline_variables + +#define PHMAP_COMPARE_INLINE_BASECLASS_DECL(name) static const T name; + +#define PHMAP_COMPARE_INLINE_SUBCLASS_DECL(type, name) + +#define PHMAP_COMPARE_INLINE_INIT(type, name, init) \ + template \ + const T compare_internal::type##_base::name(init) + +#endif // __cpp_inline_variables + +// These template base classes allow for defining the values of the constants +// in the header file (for performance) without using inline variables (which +// aren't available in C++11). +template +struct weak_equality_base { + PHMAP_COMPARE_INLINE_BASECLASS_DECL(equivalent) + PHMAP_COMPARE_INLINE_BASECLASS_DECL(nonequivalent) +}; + +template +struct strong_equality_base { + PHMAP_COMPARE_INLINE_BASECLASS_DECL(equal) + PHMAP_COMPARE_INLINE_BASECLASS_DECL(nonequal) + PHMAP_COMPARE_INLINE_BASECLASS_DECL(equivalent) + PHMAP_COMPARE_INLINE_BASECLASS_DECL(nonequivalent) +}; + +template +struct partial_ordering_base { + PHMAP_COMPARE_INLINE_BASECLASS_DECL(less) + PHMAP_COMPARE_INLINE_BASECLASS_DECL(equivalent) + PHMAP_COMPARE_INLINE_BASECLASS_DECL(greater) + PHMAP_COMPARE_INLINE_BASECLASS_DECL(unordered) +}; + +template +struct weak_ordering_base { + PHMAP_COMPARE_INLINE_BASECLASS_DECL(less) + PHMAP_COMPARE_INLINE_BASECLASS_DECL(equivalent) + PHMAP_COMPARE_INLINE_BASECLASS_DECL(greater) +}; + +template +struct strong_ordering_base { + PHMAP_COMPARE_INLINE_BASECLASS_DECL(less) + PHMAP_COMPARE_INLINE_BASECLASS_DECL(equal) + PHMAP_COMPARE_INLINE_BASECLASS_DECL(equivalent) + PHMAP_COMPARE_INLINE_BASECLASS_DECL(greater) +}; + +} // namespace compare_internal + +class weak_equality : public compare_internal::weak_equality_base { + explicit constexpr weak_equality(compare_internal::eq v) noexcept + : value_(static_cast(v)) {} + friend struct compare_internal::weak_equality_base; + + public: + PHMAP_COMPARE_INLINE_SUBCLASS_DECL(weak_equality, equivalent) + PHMAP_COMPARE_INLINE_SUBCLASS_DECL(weak_equality, nonequivalent) + + // Comparisons + friend constexpr bool operator==(weak_equality v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.value_ == 0; + } + friend constexpr bool operator!=(weak_equality v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.value_ != 0; + } + friend constexpr bool operator==(compare_internal::OnlyLiteralZero<>, + weak_equality v) noexcept { + return 0 == v.value_; + } + friend constexpr bool operator!=(compare_internal::OnlyLiteralZero<>, + weak_equality v) noexcept { + return 0 != v.value_; + } + + private: + compare_internal::value_type value_; +}; +PHMAP_COMPARE_INLINE_INIT(weak_equality, equivalent, compare_internal::eq::equivalent); +PHMAP_COMPARE_INLINE_INIT(weak_equality, nonequivalent, + compare_internal::eq::nonequivalent); + +class strong_equality : public compare_internal::strong_equality_base { + explicit constexpr strong_equality(compare_internal::eq v) noexcept + : value_(static_cast(v)) {} + friend struct compare_internal::strong_equality_base; + + public: + PHMAP_COMPARE_INLINE_SUBCLASS_DECL(strong_equality, equal) + PHMAP_COMPARE_INLINE_SUBCLASS_DECL(strong_equality, nonequal) + PHMAP_COMPARE_INLINE_SUBCLASS_DECL(strong_equality, equivalent) + PHMAP_COMPARE_INLINE_SUBCLASS_DECL(strong_equality, nonequivalent) + + // Conversion + constexpr operator weak_equality() const noexcept { // NOLINT + return value_ == 0 ? weak_equality::equivalent : weak_equality::nonequivalent; + } + // Comparisons + friend constexpr bool operator==(strong_equality v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.value_ == 0; + } + friend constexpr bool operator!=(strong_equality v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.value_ != 0; + } + friend constexpr bool operator==(compare_internal::OnlyLiteralZero<>, + strong_equality v) noexcept { + return 0 == v.value_; + } + friend constexpr bool operator!=(compare_internal::OnlyLiteralZero<>, + strong_equality v) noexcept { + return 0 != v.value_; + } + + private: + compare_internal::value_type value_; +}; + +PHMAP_COMPARE_INLINE_INIT(strong_equality, equal, compare_internal::eq::equal); +PHMAP_COMPARE_INLINE_INIT(strong_equality, nonequal, compare_internal::eq::nonequal); +PHMAP_COMPARE_INLINE_INIT(strong_equality, equivalent, compare_internal::eq::equivalent); +PHMAP_COMPARE_INLINE_INIT(strong_equality, nonequivalent, + compare_internal::eq::nonequivalent); + +class partial_ordering + : public compare_internal::partial_ordering_base { + explicit constexpr partial_ordering(compare_internal::eq v) noexcept + : value_(static_cast(v)) {} + explicit constexpr partial_ordering(compare_internal::ord v) noexcept + : value_(static_cast(v)) {} + explicit constexpr partial_ordering(compare_internal::ncmp v) noexcept + : value_(static_cast(v)) {} + friend struct compare_internal::partial_ordering_base; + + constexpr bool is_ordered() const noexcept { + return value_ != compare_internal::value_type(compare_internal::ncmp::unordered); + } + + public: + PHMAP_COMPARE_INLINE_SUBCLASS_DECL(partial_ordering, less) + PHMAP_COMPARE_INLINE_SUBCLASS_DECL(partial_ordering, equivalent) + PHMAP_COMPARE_INLINE_SUBCLASS_DECL(partial_ordering, greater) + PHMAP_COMPARE_INLINE_SUBCLASS_DECL(partial_ordering, unordered) + + // Conversion + constexpr operator weak_equality() const noexcept { // NOLINT + return value_ == 0 ? weak_equality::equivalent : weak_equality::nonequivalent; + } + // Comparisons + friend constexpr bool operator==(partial_ordering v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.is_ordered() && v.value_ == 0; + } + friend constexpr bool operator!=(partial_ordering v, + compare_internal::OnlyLiteralZero<>) noexcept { + return !v.is_ordered() || v.value_ != 0; + } + friend constexpr bool operator<(partial_ordering v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.is_ordered() && v.value_ < 0; + } + friend constexpr bool operator<=(partial_ordering v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.is_ordered() && v.value_ <= 0; + } + friend constexpr bool operator>(partial_ordering v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.is_ordered() && v.value_ > 0; + } + friend constexpr bool operator>=(partial_ordering v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.is_ordered() && v.value_ >= 0; + } + friend constexpr bool operator==(compare_internal::OnlyLiteralZero<>, + partial_ordering v) noexcept { + return v.is_ordered() && 0 == v.value_; + } + friend constexpr bool operator!=(compare_internal::OnlyLiteralZero<>, + partial_ordering v) noexcept { + return !v.is_ordered() || 0 != v.value_; + } + friend constexpr bool operator<(compare_internal::OnlyLiteralZero<>, + partial_ordering v) noexcept { + return v.is_ordered() && 0 < v.value_; + } + friend constexpr bool operator<=(compare_internal::OnlyLiteralZero<>, + partial_ordering v) noexcept { + return v.is_ordered() && 0 <= v.value_; + } + friend constexpr bool operator>(compare_internal::OnlyLiteralZero<>, + partial_ordering v) noexcept { + return v.is_ordered() && 0 > v.value_; + } + friend constexpr bool operator>=(compare_internal::OnlyLiteralZero<>, + partial_ordering v) noexcept { + return v.is_ordered() && 0 >= v.value_; + } + + private: + compare_internal::value_type value_; +}; + +PHMAP_COMPARE_INLINE_INIT(partial_ordering, less, compare_internal::ord::less); +PHMAP_COMPARE_INLINE_INIT(partial_ordering, equivalent, compare_internal::eq::equivalent); +PHMAP_COMPARE_INLINE_INIT(partial_ordering, greater, compare_internal::ord::greater); +PHMAP_COMPARE_INLINE_INIT(partial_ordering, unordered, compare_internal::ncmp::unordered); + +class weak_ordering : public compare_internal::weak_ordering_base { + explicit constexpr weak_ordering(compare_internal::eq v) noexcept + : value_(static_cast(v)) {} + explicit constexpr weak_ordering(compare_internal::ord v) noexcept + : value_(static_cast(v)) {} + friend struct compare_internal::weak_ordering_base; + + public: + PHMAP_COMPARE_INLINE_SUBCLASS_DECL(weak_ordering, less) + PHMAP_COMPARE_INLINE_SUBCLASS_DECL(weak_ordering, equivalent) + PHMAP_COMPARE_INLINE_SUBCLASS_DECL(weak_ordering, greater) + + // Conversions + constexpr operator weak_equality() const noexcept { // NOLINT + return value_ == 0 ? weak_equality::equivalent : weak_equality::nonequivalent; + } + constexpr operator partial_ordering() const noexcept { // NOLINT + return value_ == 0 + ? partial_ordering::equivalent + : (value_ < 0 ? partial_ordering::less : partial_ordering::greater); + } + // Comparisons + friend constexpr bool operator==(weak_ordering v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.value_ == 0; + } + friend constexpr bool operator!=(weak_ordering v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.value_ != 0; + } + friend constexpr bool operator<(weak_ordering v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.value_ < 0; + } + friend constexpr bool operator<=(weak_ordering v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.value_ <= 0; + } + friend constexpr bool operator>(weak_ordering v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.value_ > 0; + } + friend constexpr bool operator>=(weak_ordering v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.value_ >= 0; + } + friend constexpr bool operator==(compare_internal::OnlyLiteralZero<>, + weak_ordering v) noexcept { + return 0 == v.value_; + } + friend constexpr bool operator!=(compare_internal::OnlyLiteralZero<>, + weak_ordering v) noexcept { + return 0 != v.value_; + } + friend constexpr bool operator<(compare_internal::OnlyLiteralZero<>, + weak_ordering v) noexcept { + return 0 < v.value_; + } + friend constexpr bool operator<=(compare_internal::OnlyLiteralZero<>, + weak_ordering v) noexcept { + return 0 <= v.value_; + } + friend constexpr bool operator>(compare_internal::OnlyLiteralZero<>, + weak_ordering v) noexcept { + return 0 > v.value_; + } + friend constexpr bool operator>=(compare_internal::OnlyLiteralZero<>, + weak_ordering v) noexcept { + return 0 >= v.value_; + } + + private: + compare_internal::value_type value_; +}; + +PHMAP_COMPARE_INLINE_INIT(weak_ordering, less, compare_internal::ord::less); +PHMAP_COMPARE_INLINE_INIT(weak_ordering, equivalent, compare_internal::eq::equivalent); +PHMAP_COMPARE_INLINE_INIT(weak_ordering, greater, compare_internal::ord::greater); + +class strong_ordering : public compare_internal::strong_ordering_base { + explicit constexpr strong_ordering(compare_internal::eq v) noexcept + : value_(static_cast(v)) {} + explicit constexpr strong_ordering(compare_internal::ord v) noexcept + : value_(static_cast(v)) {} + friend struct compare_internal::strong_ordering_base; + + public: + PHMAP_COMPARE_INLINE_SUBCLASS_DECL(strong_ordering, less) + PHMAP_COMPARE_INLINE_SUBCLASS_DECL(strong_ordering, equal) + PHMAP_COMPARE_INLINE_SUBCLASS_DECL(strong_ordering, equivalent) + PHMAP_COMPARE_INLINE_SUBCLASS_DECL(strong_ordering, greater) + + // Conversions + constexpr operator weak_equality() const noexcept { // NOLINT + return value_ == 0 ? weak_equality::equivalent : weak_equality::nonequivalent; + } + constexpr operator strong_equality() const noexcept { // NOLINT + return value_ == 0 ? strong_equality::equal : strong_equality::nonequal; + } + constexpr operator partial_ordering() const noexcept { // NOLINT + return value_ == 0 + ? partial_ordering::equivalent + : (value_ < 0 ? partial_ordering::less : partial_ordering::greater); + } + constexpr operator weak_ordering() const noexcept { // NOLINT + return value_ == 0 ? weak_ordering::equivalent + : (value_ < 0 ? weak_ordering::less : weak_ordering::greater); + } + // Comparisons + friend constexpr bool operator==(strong_ordering v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.value_ == 0; + } + friend constexpr bool operator!=(strong_ordering v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.value_ != 0; + } + friend constexpr bool operator<(strong_ordering v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.value_ < 0; + } + friend constexpr bool operator<=(strong_ordering v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.value_ <= 0; + } + friend constexpr bool operator>(strong_ordering v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.value_ > 0; + } + friend constexpr bool operator>=(strong_ordering v, + compare_internal::OnlyLiteralZero<>) noexcept { + return v.value_ >= 0; + } + friend constexpr bool operator==(compare_internal::OnlyLiteralZero<>, + strong_ordering v) noexcept { + return 0 == v.value_; + } + friend constexpr bool operator!=(compare_internal::OnlyLiteralZero<>, + strong_ordering v) noexcept { + return 0 != v.value_; + } + friend constexpr bool operator<(compare_internal::OnlyLiteralZero<>, + strong_ordering v) noexcept { + return 0 < v.value_; + } + friend constexpr bool operator<=(compare_internal::OnlyLiteralZero<>, + strong_ordering v) noexcept { + return 0 <= v.value_; + } + friend constexpr bool operator>(compare_internal::OnlyLiteralZero<>, + strong_ordering v) noexcept { + return 0 > v.value_; + } + friend constexpr bool operator>=(compare_internal::OnlyLiteralZero<>, + strong_ordering v) noexcept { + return 0 >= v.value_; + } + + private: + compare_internal::value_type value_; +}; +PHMAP_COMPARE_INLINE_INIT(strong_ordering, less, compare_internal::ord::less); +PHMAP_COMPARE_INLINE_INIT(strong_ordering, equal, compare_internal::eq::equal); +PHMAP_COMPARE_INLINE_INIT(strong_ordering, equivalent, compare_internal::eq::equivalent); +PHMAP_COMPARE_INLINE_INIT(strong_ordering, greater, compare_internal::ord::greater); + +#undef PHMAP_COMPARE_INLINE_BASECLASS_DECL +#undef PHMAP_COMPARE_INLINE_SUBCLASS_DECL +#undef PHMAP_COMPARE_INLINE_INIT + +namespace compare_internal { +// We also provide these comparator adapter functions for internal phmap use. + +// Helper functions to do a boolean comparison of two keys given a boolean +// or three-way comparator. +// SFINAE prevents implicit conversions to bool (such as from int). +template ::value, int> = 0> +constexpr bool compare_result_as_less_than(const BoolType r) { + return r; +} +constexpr bool compare_result_as_less_than(const phmap::weak_ordering r) { return r < 0; } + +template +constexpr bool do_less_than_comparison(const Compare& compare, const K& x, const LK& y) { + return compare_result_as_less_than(compare(x, y)); +} + +// Helper functions to do a three-way comparison of two keys given a boolean or +// three-way comparator. +// SFINAE prevents implicit conversions to int (such as from bool). +template ::value, int> = 0> +constexpr phmap::weak_ordering compare_result_as_ordering(const Int c) { + return c < 0 + ? phmap::weak_ordering::less + : c == 0 ? phmap::weak_ordering::equivalent : phmap::weak_ordering::greater; +} +constexpr phmap::weak_ordering compare_result_as_ordering(const phmap::weak_ordering c) { + return c; +} + +template < + typename Compare, typename K, typename LK, + phmap::enable_if_t< + !std::is_same>::value, + int> = 0> +constexpr phmap::weak_ordering do_three_way_comparison(const Compare& compare, const K& x, + const LK& y) { + return compare_result_as_ordering(compare(x, y)); +} +template < + typename Compare, typename K, typename LK, + phmap::enable_if_t< + std::is_same>::value, + int> = 0> +constexpr phmap::weak_ordering do_three_way_comparison(const Compare& compare, const K& x, + const LK& y) { + return compare(x, y) ? phmap::weak_ordering::less + : compare(y, x) ? phmap::weak_ordering::greater + : phmap::weak_ordering::equivalent; +} + +} // namespace compare_internal +} // namespace phmap + +namespace phmap { + +namespace priv { + +// A helper class that indicates if the Compare parameter is a key-compare-to +// comparator. +template +using btree_is_key_compare_to = + std::is_convertible, + phmap::weak_ordering>; + +struct StringBtreeDefaultLess { + using is_transparent = void; + + StringBtreeDefaultLess() = default; + + // Compatibility constructor. + StringBtreeDefaultLess(std::less) {} // NOLINT +#if PHMAP_HAVE_STD_STRING_VIEW + StringBtreeDefaultLess(std::less) {} // NOLINT + StringBtreeDefaultLess(phmap::Less) {} // NOLINT + + phmap::weak_ordering operator()(std::string_view lhs, std::string_view rhs) const { + return compare_internal::compare_result_as_ordering(lhs.compare(rhs)); + } +#else + phmap::weak_ordering operator()(std::string lhs, std::string rhs) const { + return compare_internal::compare_result_as_ordering(lhs.compare(rhs)); + } +#endif +}; + +struct StringBtreeDefaultGreater { + using is_transparent = void; + + StringBtreeDefaultGreater() = default; + + StringBtreeDefaultGreater(std::greater) {} // NOLINT +#if PHMAP_HAVE_STD_STRING_VIEW + StringBtreeDefaultGreater(std::greater) {} // NOLINT + + phmap::weak_ordering operator()(std::string_view lhs, std::string_view rhs) const { + return compare_internal::compare_result_as_ordering(rhs.compare(lhs)); + } +#else + phmap::weak_ordering operator()(std::string lhs, std::string rhs) const { + return compare_internal::compare_result_as_ordering(rhs.compare(lhs)); + } +#endif +}; + +// A helper class to convert a boolean comparison into a three-way "compare-to" +// comparison that returns a negative value to indicate less-than, zero to +// indicate equality and a positive value to indicate greater-than. This helper +// class is specialized for less, greater, +// less, and greater. +// +// key_compare_to_adapter is provided so that btree users +// automatically get the more efficient compare-to code when using common +// google string types with common comparison functors. +// These string-like specializations also turn on heterogeneous lookup by +// default. +template +struct key_compare_to_adapter { + using type = Compare; +}; + +template <> +struct key_compare_to_adapter> { + using type = StringBtreeDefaultLess; +}; + +template <> +struct key_compare_to_adapter> { + using type = StringBtreeDefaultLess; +}; + +template <> +struct key_compare_to_adapter> { + using type = StringBtreeDefaultGreater; +}; + +#if PHMAP_HAVE_STD_STRING_VIEW +template <> +struct key_compare_to_adapter> { + using type = StringBtreeDefaultLess; +}; + +template <> +struct key_compare_to_adapter> { + using type = StringBtreeDefaultLess; +}; + +template <> +struct key_compare_to_adapter> { + using type = StringBtreeDefaultGreater; +}; +#endif + +template +struct common_params { + // If Compare is a common comparator for a std::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::type; + // 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; + + using allocator_type = Alloc; + using key_type = Key; + using size_type = std::size_t; + using difference_type = ptrdiff_t; + + // True if this is a multiset or multimap. + using is_multi_container = std::integral_constant; + + using slot_policy = SlotPolicy; + using slot_type = typename slot_policy::slot_type; + using value_type = typename slot_policy::value_type; + using init_type = typename slot_policy::mutable_value_type; + using pointer = value_type*; + using const_pointer = const value_type*; + using reference = value_type&; + using const_reference = const value_type&; + + enum { + kTargetNodeSize = TargetNodeSize, + + // Upper bound for the available space for values. This is largest for leaf + // nodes, which have overhead of at least a pointer + 4 bytes (for storing + // 3 field_types and an enum). + kNodeValueSpace = TargetNodeSize - /*minimum overhead=*/(sizeof(void*) + 4), + }; + + // This is an integral type large enough to hold as many + // ValueSize-values as will fit a node of TargetNodeSize bytes. + using node_count_type = phmap::conditional_t<(kNodeValueSpace / sizeof(value_type) > + (std::numeric_limits::max)()), + uint16_t, uint8_t>; // NOLINT + + // The following methods are necessary for passing this struct as PolicyTraits + // for node_handle and/or are used within btree. + static value_type& element(slot_type* slot) { return slot_policy::element(slot); } + static const value_type& element(const slot_type* slot) { + return slot_policy::element(slot); + } + template + static void construct(Alloc* alloc, slot_type* slot, Args&&... args) { + slot_policy::construct(alloc, slot, std::forward(args)...); + } + static void construct(Alloc* alloc, slot_type* slot, slot_type* other) { + slot_policy::construct(alloc, slot, other); + } + static void destroy(Alloc* alloc, slot_type* slot) { + slot_policy::destroy(alloc, slot); + } + static void transfer(Alloc* alloc, slot_type* new_slot, slot_type* old_slot) { + construct(alloc, new_slot, old_slot); + destroy(alloc, old_slot); + } + static void swap(Alloc* alloc, slot_type* a, slot_type* b) { + slot_policy::swap(alloc, a, b); + } + static void move(Alloc* alloc, slot_type* src, slot_type* dest) { + slot_policy::move(alloc, src, dest); + } + static void move(Alloc* alloc, slot_type* first, slot_type* last, slot_type* result) { + slot_policy::move(alloc, first, last, result); + } +}; + +// A parameters structure for holding the type parameters for a btree_map. +// Compare and Alloc should be nothrow copy-constructible. +template +struct map_params : common_params> { + using super_type = typename map_params::common_params; + using mapped_type = Data; + // This type allows us to move keys when it is safe to do so. It is safe + // for maps in which value_type and mutable_value_type are layout compatible. + using slot_policy = typename super_type::slot_policy; + using slot_type = typename super_type::slot_type; + using value_type = typename super_type::value_type; + using init_type = typename super_type::init_type; + + using key_compare = typename super_type::key_compare; + // Inherit from key_compare for empty base class optimization. + struct value_compare : private key_compare { + value_compare() = default; + explicit value_compare(const key_compare& cmp) : key_compare(cmp) {} + + template + auto operator()(const T& left, const U& right) const + -> decltype(std::declval()(left.first, right.first)) { + return key_compare::operator()(left.first, right.first); + } + }; + using is_map_container = std::true_type; + + static const Key& key(const value_type& x) { return x.first; } + static const Key& key(const init_type& x) { return x.first; } + static const Key& key(const slot_type* x) { return slot_policy::key(x); } + static mapped_type& value(value_type* value) { return value->second; } +}; + +// This type implements the necessary functions from the +// btree::priv::slot_type interface. +template +struct set_slot_policy { + using slot_type = Key; + using value_type = Key; + using mutable_value_type = Key; + + static value_type& element(slot_type* slot) { return *slot; } + static const value_type& element(const slot_type* slot) { return *slot; } + + template + static void construct(Alloc* alloc, slot_type* slot, Args&&... args) { + phmap::allocator_traits::construct(*alloc, slot, std::forward(args)...); + } + + template + static void construct(Alloc* alloc, slot_type* slot, slot_type* other) { + phmap::allocator_traits::construct(*alloc, slot, std::move(*other)); + } + + template + static void destroy(Alloc* alloc, slot_type* slot) { + phmap::allocator_traits::destroy(*alloc, slot); + } + + template + static void swap(Alloc* /*alloc*/, slot_type* a, slot_type* b) { + using std::swap; + swap(*a, *b); + } + + template + static void move(Alloc* /*alloc*/, slot_type* src, slot_type* dest) { + *dest = std::move(*src); + } + + template + static void move(Alloc* alloc, slot_type* first, slot_type* last, slot_type* result) { + for (slot_type *src = first, *dest = result; src != last; ++src, ++dest) + move(alloc, src, dest); + } +}; + +// A parameters structure for holding the type parameters for a btree_set. +// Compare and Alloc should be nothrow copy-constructible. +template +struct set_params + : common_params> { + using value_type = Key; + using slot_type = typename set_params::common_params::slot_type; + using value_compare = typename set_params::common_params::key_compare; + using is_map_container = std::false_type; + + static const Key& key(const value_type& x) { return x; } + static const Key& key(const slot_type* x) { return *x; } +}; + +// An adapter class that converts a lower-bound compare into an upper-bound +// compare. Note: there is no need to make a version of this adapter specialized +// for key-compare-to functors because the upper-bound (the first value greater +// than the input) is never an exact match. +template +struct upper_bound_adapter { + explicit upper_bound_adapter(const Compare& c) : comp(c) {} + template + bool operator()(const K& a, const LK& b) const { + // Returns true when a is not greater than b. + return !phmap::compare_internal::compare_result_as_less_than(comp(b, a)); + } + + private: + Compare comp; +}; + +enum class MatchKind : uint8_t { kEq, kNe }; + +template +struct SearchResult { + V value; + MatchKind match; + + static constexpr bool HasMatch() { return true; } + bool IsEq() const { return match == MatchKind::kEq; } +}; + +// When we don't use CompareTo, `match` is not present. +// This ensures that callers can't use it accidentally when it provides no +// useful information. +template +struct SearchResult { + V value; + + static constexpr bool HasMatch() { return false; } + static constexpr bool IsEq() { return false; } +}; + +// A node in the btree holding. The same node type is used for both internal +// and leaf nodes in the btree, though the nodes are allocated in such a way +// that the children array is only valid in internal nodes. +template +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; + + public: + using params_type = Params; + using key_type = typename Params::key_type; + using value_type = typename Params::value_type; + using pointer = typename Params::pointer; + using const_pointer = typename Params::const_pointer; + using reference = typename Params::reference; + using const_reference = typename Params::const_reference; + using key_compare = typename Params::key_compare; + using size_type = typename Params::size_type; + using difference_type = typename Params::difference_type; + + // Btree decides whether to use linear node search as follows: + // - 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::value && + (std::is_same, key_compare>::value || + std::is_same, key_compare>::value || + std::is_same, key_compare>::value)>; + + ~btree_node() = default; + btree_node(btree_node const&) = delete; + btree_node& operator=(btree_node const&) = delete; + + // Public for EmptyNodeType. + constexpr static size_type Alignment() { + static_assert(LeafLayout(1).Alignment() == InternalLayout().Alignment(), + "Alignment of all nodes must be equal."); + return (size_type)InternalLayout().Alignment(); + } + + protected: + btree_node() = default; + + private: + using layout_type = + phmap::priv::Layout; + constexpr static size_type SizeWithNValues(size_type n) { + return (size_type)layout_type(/*parent*/ 1, + /*position, start, count, max_count*/ 4, + /*values*/ (size_t)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 (size_type)(SizeWithNValues(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) { + return begin == end + ? begin + : SizeWithNValues((begin + end) / 2 + 1) > params_type::kTargetNodeSize + ? NodeTargetValues(begin, (begin + end) / 2) + : NodeTargetValues((begin + end) / 2 + 1, end); + } + + enum { + kTargetNodeSize = params_type::kTargetNodeSize, + kNodeTargetValues = NodeTargetValues(0, params_type::kTargetNodeSize), + + // We need a minimum of 3 values 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, + + // 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) { + return layout_type(/*parent*/ 1, + /*position, start, count, max_count*/ 4, + /*values*/ (size_t)max_values, + /*children*/ 0); + } + constexpr static layout_type InternalLayout() { + return layout_type(/*parent*/ 1, + /*position, start, count, max_count*/ 4, + /*values*/ kNodeValues, + /*children*/ kNodeValues + 1); + } + constexpr static size_type LeafSize(const int max_values = kNodeValues) { + return (size_type)LeafLayout(max_values).AllocSize(); + } + constexpr static size_type InternalSize() { + return (size_type)InternalLayout().AllocSize(); + } + + // N is the index of the type in the Layout definition. + // ElementType is the Nth type in the Layout definition. + template + inline typename layout_type::template ElementType* GetField() { + // We assert that we don't read from values that aren't there. + assert(N < 3 || !leaf()); + return InternalLayout().template Pointer(reinterpret_cast(this)); + } + + template + inline const typename layout_type::template ElementType* GetField() const { + assert(N < 3 || !leaf()); + return InternalLayout().template Pointer(reinterpret_cast(this)); + } + + void set_parent(btree_node* p) { *GetField<0>() = p; } + field_type& mutable_count() { return GetField<1>()[2]; } + slot_type* slot(size_type i) { return &GetField<2>()[i]; } + const slot_type* slot(size_type i) const { return &GetField<2>()[i]; } + void set_position(field_type v) { GetField<1>()[0] = v; } + void set_start(field_type v) { GetField<1>()[1] = v; } + void set_count(field_type v) { GetField<1>()[2] = v; } + void set_max_count(field_type v) { GetField<1>()[3] = v; } + + public: + // Whether this is a leaf node or not. This value doesn't change after the + // node is created. + bool leaf() const { return GetField<1>()[3] != kInternalNodeMaxCount; } + + // Getter for the position of this node in its parent. + field_type position() const { return GetField<1>()[0]; } + + // Getter for the offset of the first value in the `values` array. + field_type start() const { return GetField<1>()[1]; } + + // Getters for the number of values stored in this node. + field_type count() const { return GetField<1>()[2]; } + field_type max_count() const { + // Internal nodes have max_count==kInternalNodeMaxCount. + // Leaf nodes have max_count in [1, kNodeValues]. + const field_type max_cnt = GetField<1>()[3]; + return max_cnt == field_type{kInternalNodeMaxCount} ? field_type{kNodeValues} + : max_cnt; + } + + // Getter for the parent of this node. + btree_node* parent() const { return *GetField<0>(); } + // Getter for whether the node is the root of the tree. The parent of the + // root of the tree is the leftmost node in the tree which is guaranteed to + // be a leaf. + bool is_root() const { return parent()->leaf(); } + void make_root() { + assert(parent()->is_root()); + set_parent(parent()->parent()); + } + + // Getters for the key/value at position i in the node. + const key_type& key(size_type i) const { return params_type::key(slot(i)); } + reference value(size_type i) { return params_type::element(slot(i)); } + const_reference value(size_type i) const { return params_type::element(slot(i)); } + + // Getters/setter for the child at position i in the node. + btree_node* child(size_type i) const { return GetField<3>()[i]; } + btree_node*& mutable_child(size_type i) { return GetField<3>()[i]; } + void clear_child(size_type i) { phmap::priv::SanitizerPoisonObject(&mutable_child(i)); } + void set_child(size_type i, btree_node* c) { + phmap::priv::SanitizerUnpoisonObject(&mutable_child(i)); + mutable_child(i) = c; + c->set_position((field_type)i); + } + void init_child(int i, btree_node* c) { + set_child(i, c); + c->set_parent(this); + } + + // Returns the position of the first value whose key is not less than k. + template + SearchResult lower_bound(const K& k, + const key_compare& comp) const { + return use_linear_search::value ? linear_search(k, comp) : binary_search(k, comp); + } + // Returns the position of the first value whose key is greater than k. + template + int upper_bound(const K& k, const key_compare& comp) const { + auto upper_compare = upper_bound_adapter(comp); + return use_linear_search::value ? linear_search(k, upper_compare).value + : binary_search(k, upper_compare).value; + } + + template + SearchResult::value> linear_search( + const K& k, const Compare& comp) const { + return linear_search_impl(k, 0, count(), comp, + btree_is_key_compare_to()); + } + + template + SearchResult::value> binary_search( + const K& k, const Compare& comp) const { + return binary_search_impl(k, 0, count(), comp, + btree_is_key_compare_to()); + } + + // Returns the position of the first value whose key is not less than k using + // linear search performed using plain compare. + template + SearchResult linear_search_impl(const K& k, int s, const int e, + const Compare& comp, + std::false_type /* IsCompareTo */) const { + while (s < e) { + if (!comp(key(s), k)) { + break; + } + ++s; + } + return {s}; + } + + // Returns the position of the first value whose key is not less than k using + // linear search performed using compare-to. + template + SearchResult linear_search_impl(const K& k, int s, const int e, + const Compare& comp, + std::true_type /* IsCompareTo */) const { + while (s < e) { + const phmap::weak_ordering c = comp(key(s), k); + if (c == 0) { + return {s, MatchKind::kEq}; + } else if (c > 0) { + break; + } + ++s; + } + return {s, MatchKind::kNe}; + } + + // Returns the position of the first value whose key is not less than k using + // binary search performed using plain compare. + template + SearchResult binary_search_impl(const K& k, int s, int e, + const Compare& comp, + std::false_type /* IsCompareTo */) const { + while (s != e) { + const int mid = (s + e) >> 1; + if (comp(key(mid), k)) { + s = mid + 1; + } else { + e = mid; + } + } + return {s}; + } + + // Returns the position of the first value whose key is not less than k using + // binary search performed using compare-to. + template + SearchResult binary_search_impl(const K& k, int s, int e, + const CompareTo& comp, + std::true_type /* IsCompareTo */) const { + if (is_multi_container::value) { + MatchKind exact_match = MatchKind::kNe; + while (s != e) { + const int mid = (s + e) >> 1; + const phmap::weak_ordering c = comp(key(mid), k); + if (c < 0) { + s = mid + 1; + } else { + 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. + exact_match = MatchKind::kEq; + } + } + } + return {s, exact_match}; + } else { // Not a multi-container. + while (s != e) { + const int mid = (s + e) >> 1; + const phmap::weak_ordering c = comp(key(mid), k); + if (c < 0) { + s = mid + 1; + } else if (c > 0) { + e = mid; + } else { + return {mid, MatchKind::kEq}; + } + } + return {s, MatchKind::kNe}; + } + } + + // Emplaces a value at position i, shifting all existing values and + // children at positions >= i to the right by 1. + template + void emplace_value(size_type i, allocator_type* alloc, Args&&... args); + + // Removes the value at position i, shifting all existing values and children + // at positions > i to the left by 1. + void remove_value(int i, allocator_type* alloc); + + // Removes the values at positions [i, i + to_erase), shifting all values + // after that range to the left by to_erase. Does not change children at all. + void remove_values_ignore_children(int i, size_type to_erase, allocator_type* alloc); + + // Rebalances a node with its right sibling. + void rebalance_right_to_left(int to_move, btree_node* right, allocator_type* alloc); + void rebalance_left_to_right(int to_move, btree_node* right, allocator_type* alloc); + + // Splits a node, moving a portion of the node's values to its right sibling. + void split(int insert_position, btree_node* dest, allocator_type* alloc); + + // Merges a node with its right sibling, moving all of the values and the + // delimiting key in the parent node onto itself. + void merge(btree_node* sibling, allocator_type* alloc); + + // Swap the contents of "this" and "src". + void swap(btree_node* src, allocator_type* alloc); + + // Node allocation/deletion routines. + static btree_node* init_leaf(btree_node* n, btree_node* parent, int max_cnt) { + n->set_parent(parent); + n->set_position(0); + n->set_start(0); + n->set_count(0); + n->set_max_count((field_type)max_cnt); + phmap::priv::SanitizerPoisonMemoryRegion(n->slot(0), max_cnt * sizeof(slot_type)); + return n; + } + static btree_node* init_internal(btree_node* n, btree_node* parent) { + init_leaf(n, parent, kNodeValues); + // Set `max_count` to a sentinel value to indicate that this node is + // internal. + n->set_max_count(kInternalNodeMaxCount); + phmap::priv::SanitizerPoisonMemoryRegion(&n->mutable_child(0), + (kNodeValues + 1) * sizeof(btree_node*)); + return n; + } + void destroy(allocator_type* alloc) { + for (int i = 0; i < count(); ++i) { + value_destroy(i, alloc); + } + } + + public: + // Exposed only for tests. + static bool testonly_uses_linear_node_search() { return use_linear_search::value; } + + private: + template + void value_init(const size_type i, allocator_type* alloc, Args&&... args) { + phmap::priv::SanitizerUnpoisonObject(slot(i)); + params_type::construct(alloc, slot(i), std::forward(args)...); + } + void value_destroy(const size_type i, allocator_type* alloc) { + params_type::destroy(alloc, slot(i)); + phmap::priv::SanitizerPoisonObject(slot(i)); + } + + // Move n values starting at value i in this node into the values starting at + // value j in node x. + void uninitialized_move_n(const size_type n, const size_type i, const size_type j, + btree_node* x, allocator_type* alloc) { + phmap::priv::SanitizerUnpoisonMemoryRegion(x->slot(j), n * sizeof(slot_type)); + for (slot_type *src = slot(i), *end = src + n, *dest = x->slot(j); src != end; + ++src, ++dest) { + params_type::construct(alloc, dest, src); + } + } + + // Destroys a range of n values, starting at index i. + void value_destroy_n(const size_type i, const size_type n, allocator_type* alloc) { + for (size_type j = 0; j < n; ++j) { + value_destroy(i + j, alloc); + } + } + + template + friend class btree; + template + friend struct btree_iterator; + friend class BtreeNodePeer; +}; + +template +struct btree_iterator { + private: + using key_type = typename Node::key_type; + using size_type = typename Node::size_type; + using params_type = typename Node::params_type; + + using node_type = Node; + using normal_node = typename std::remove_const::type; + using const_node = const Node; + using normal_pointer = typename params_type::pointer; + using normal_reference = typename params_type::reference; + using const_pointer = typename params_type::const_pointer; + using const_reference = typename params_type::const_reference; + using slot_type = typename params_type::slot_type; + + using iterator = btree_iterator; + using const_iterator = btree_iterator; + + public: + // These aliases are public for std::iterator_traits. + using difference_type = typename Node::difference_type; + using value_type = typename params_type::value_type; + using pointer = Pointer; + using reference = Reference; + using iterator_category = std::bidirectional_iterator_tag; + + btree_iterator() : node(nullptr), position(-1) {} + 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. + template , iterator>::value && + std::is_same::value, + int> = 0> + btree_iterator(const btree_iterator& x) // NOLINT + : node(x.node), position(x.position) {} + + private: + // This SFINAE allows explicit conversions from const_iterator to + // iterator, but also avoids defining a 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, + phmap::enable_if_t, const_iterator>::value && + std::is_same::value, + int> = 0> + explicit btree_iterator(const btree_iterator& x) + : node(const_cast(x.node)), position(x.position) {} + + // Increment/decrement the iterator. + void increment() { + if (node->leaf() && ++position < node->count()) { + return; + } + increment_slow(); + } + void increment_slow(); + + void decrement() { + if (node->leaf() && --position >= 0) { + return; + } + decrement_slow(); + } + void decrement_slow(); + + public: + bool operator==(const const_iterator& x) const { + return node == x.node && position == x.position; + } + bool operator!=(const const_iterator& x) const { + return node != x.node || position != x.position; + } + + // Accessors for the key/value the iterator is pointing at. + reference operator*() const { return node->value(position); } + pointer operator->() const { return &node->value(position); } + + btree_iterator& operator++() { + increment(); + return *this; + } + btree_iterator& operator--() { + decrement(); + return *this; + } + btree_iterator operator++(int) { + btree_iterator tmp = *this; + ++*this; + return tmp; + } + btree_iterator operator--(int) { + btree_iterator tmp = *this; + --*this; + return tmp; + } + + private: + template + friend class btree; + template + friend class btree_container; + template + friend class btree_set_container; + template + friend class btree_map_container; + template + friend class btree_multiset_container; + template + friend struct btree_iterator; + template + friend class base_checker; + + const key_type& key() const { return node->key(position); } + slot_type* slot() { return node->slot(position); } + + // The node in the tree the iterator is pointing at. + Node* node; + // The position within the node of the tree the iterator is pointing at. + // TODO(ezb): make this a field_type + int position; +}; + +template +class btree { + using node_type = btree_node; + using is_key_compare_to = typename Params::is_key_compare_to; + + // We use a static empty node for the root/leftmost/rightmost of empty btrees + // in order to avoid branching in begin()/end(). + struct alignas(node_type::Alignment()) EmptyNodeType : node_type { + using field_type = typename node_type::field_type; + node_type* parent; + field_type position = 0; + field_type start = 0; + field_type count = 0; + // max_count must be != kInternalNodeMaxCount (so that this node is regarded + // as a leaf node). max_count() is never called when the tree is empty. + field_type max_count = node_type::kInternalNodeMaxCount + 1; + +#ifdef _MSC_VER + // MSVC has constexpr code generations bugs here. + EmptyNodeType() : parent(this) {} +#else + constexpr EmptyNodeType(node_type* p) : parent(p) {} +#endif + }; + + static node_type* EmptyNode() { +#ifdef _MSC_VER + static EmptyNodeType empty_node; + // This assert fails on some other construction methods. + assert(empty_node.parent == &empty_node); + return &empty_node; +#else + static constexpr EmptyNodeType empty_node(const_cast(&empty_node)); + return const_cast(&empty_node); +#endif + } + + enum { + kNodeValues = node_type::kNodeValues, + kMinNodeValues = kNodeValues / 2, + }; + + struct node_stats { + using size_type = typename Params::size_type; + + node_stats(size_type l, size_type i) : leaf_nodes(l), internal_nodes(i) {} + + node_stats& operator+=(const node_stats& x) { + leaf_nodes += x.leaf_nodes; + internal_nodes += x.internal_nodes; + return *this; + } + + size_type leaf_nodes; + size_type internal_nodes; + }; + + public: + using key_type = typename Params::key_type; + using value_type = typename Params::value_type; + using size_type = typename Params::size_type; + using difference_type = typename Params::difference_type; + using key_compare = typename Params::key_compare; + using value_compare = typename Params::value_compare; + using allocator_type = typename Params::allocator_type; + using reference = typename Params::reference; + using const_reference = typename Params::const_reference; + using pointer = typename Params::pointer; + using const_pointer = typename Params::const_pointer; + using iterator = btree_iterator; + using const_iterator = typename iterator::const_iterator; + using reverse_iterator = std::reverse_iterator; + using const_reverse_iterator = std::reverse_iterator; + using node_handle_type = node_handle; + + // Internal types made public for use by btree_container types. + using params_type = Params; + using slot_type = typename Params::slot_type; + + private: + // For use in copy_or_move_values_in_order. + const value_type& maybe_move_from_iterator(const_iterator x) { return *x; } + value_type&& maybe_move_from_iterator(iterator x) { return std::move(*x); } + + // Copies or moves (depending on the template parameter) the values in + // x into this btree in their order in x. This btree must be empty before this + // method is called. This method is used in copy construction, copy + // assignment, and move assignment. + template + void copy_or_move_values_in_order(Btree* x); + + // 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 btree& x); + btree(btree&& x) noexcept + : root_(std::move(x.root_)), + rightmost_(phmap::exchange(x.rightmost_, EmptyNode())), + size_(phmap::exchange(x.size_, 0)) { + x.mutable_root() = EmptyNode(); + } + + ~btree() { + // Put static_asserts in destructor to avoid triggering them before the type + // is complete. + static_assert(static_assert_validation(), "This call must be elided."); + clear(); + } + + // Assign the contents of x to *this. + btree& operator=(const btree& x); + btree& operator=(btree&& x) noexcept; + + iterator begin() { return iterator(leftmost(), 0); } + const_iterator begin() const { return const_iterator(leftmost(), 0); } + iterator end() { return iterator(rightmost_, rightmost_->count()); } + const_iterator end() const { return const_iterator(rightmost_, rightmost_->count()); } + reverse_iterator rbegin() { return reverse_iterator(end()); } + const_reverse_iterator rbegin() const { return const_reverse_iterator(end()); } + reverse_iterator rend() { return reverse_iterator(begin()); } + const_reverse_iterator rend() const { return const_reverse_iterator(begin()); } + + // Finds the first element whose key is not less than key. + template + iterator lower_bound(const K& key) { + return internal_end(internal_lower_bound(key)); + } + template + const_iterator lower_bound(const K& key) const { + return internal_end(internal_lower_bound(key)); + } + + // Finds the first element whose key is greater than key. + template + iterator upper_bound(const K& key) { + return internal_end(internal_upper_bound(key)); + } + template + const_iterator upper_bound(const K& key) const { + return internal_end(internal_upper_bound(key)); + } + + // Finds the range of values which compare equal to key. The first member of + // the returned pair is equal to lower_bound(key). The second member pair of + // the pair is equal to upper_bound(key). + template + std::pair equal_range(const K& key) { + return {lower_bound(key), upper_bound(key)}; + } + template + std::pair equal_range(const K& key) const { + return {lower_bound(key), upper_bound(key)}; + } + + // Inserts a value into the btree only if it does not already exist. The + // boolean return value indicates whether insertion succeeded or failed. + // Requirement: if `key` already exists in the btree, does not consume `args`. + // Requirement: `key` is never referenced after consuming `args`. + template + std::pair insert_unique(const key_type& key, Args&&... args); + + // Inserts with hint. Checks to see if the value should be placed immediately + // before `position` in the tree. If so, then the insertion will take + // amortized constant time. If not, the insertion will take amortized + // logarithmic time as if a call to insert_unique() were made. + // Requirement: if `key` already exists in the btree, does not consume `args`. + // Requirement: `key` is never referenced after consuming `args`. + template + std::pair insert_hint_unique(iterator position, const key_type& key, + Args&&... args); + + // Insert a range of values into the btree. + template + void insert_iterator_unique(InputIterator b, InputIterator e); + + // Inserts a value into the btree. + template + iterator insert_multi(const key_type& key, ValueType&& v); + + // Inserts a value into the btree. + template + iterator insert_multi(ValueType&& v) { + return insert_multi(params_type::key(v), std::forward(v)); + } + + // Insert with hint. Check to see if the value should be placed immediately + // before position in the tree. If it does, then the insertion will take + // amortized constant time. If not, the insertion will take amortized + // logarithmic time as if a call to insert_multi(v) were made. + template + iterator insert_hint_multi(iterator position, ValueType&& v); + + // Insert a range of values into the btree. + template + void insert_iterator_multi(InputIterator b, InputIterator e); + + // Erase the specified iterator from the btree. The iterator must be valid + // (i.e. not equal to end()). Return an iterator pointing to the node after + // the one that was erased (or end() if none exists). + // Requirement: does not read the value at `*iter`. + iterator erase(iterator iter); + + // Erases range. Returns the number of keys erased and an iterator pointing + // to the element after the last erased element. + std::pair erase(iterator begin, iterator end); + + // Erases the specified key from the btree. Returns 1 if an element was + // erased and 0 otherwise. + template + 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 + size_type erase_multi(const K& key); + + // Finds the iterator corresponding to a key or returns end() if the key is + // not present. + template + iterator find(const K& key) { + return internal_end(internal_find(key)); + } + template + const_iterator find(const K& key) const { + return internal_end(internal_find(key)); + } + + // Returns a count of the number of times the key appears in the btree. + template + size_type count_unique(const K& key) const { + const iterator beg = internal_find(key); + if (beg.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 + 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(); + + // Swap the contents of *this and x. + void swap(btree& x); + + const key_compare& key_comp() const noexcept { return root_.template get<0>(); } + template + bool compare_keys(const K& x, const LK& y) const { + return compare_internal::compare_result_as_less_than(key_comp()(x, y)); + } + + value_compare value_comp() const { return value_compare(key_comp()); } + + // Verifies the structure of the btree. + void verify() const; + + // Size routines. + size_type size() const { return size_; } + size_type max_size() const { return (std::numeric_limits::max)(); } + bool empty() const { return size_ == 0; } + + // The height of the btree. An empty tree will have height 0. + size_type height() const { + size_type h = 0; + if (!empty()) { + // Count the length of the chain from the leftmost node up to the + // root. We actually count from the root back around to the level below + // the root, but the calculation is the same because of the circularity + // of that traversal. + const node_type* n = root(); + do { + ++h; + n = n->parent(); + } while (n != root()); + } + return h; + } + + // The number of internal, leaf and total nodes used by the btree. + size_type leaf_nodes() const { return internal_stats(root()).leaf_nodes; } + size_type internal_nodes() const { return internal_stats(root()).internal_nodes; } + size_type nodes() const { + node_stats stats = internal_stats(root()); + return stats.leaf_nodes + stats.internal_nodes; + } + + // The total number of bytes used by the btree. + size_type bytes_used() const { + node_stats stats = internal_stats(root()); + if (stats.leaf_nodes == 1 && stats.internal_nodes == 0) { + return sizeof(*this) + node_type::LeafSize(root()->max_count()); + } else { + return sizeof(*this) + stats.leaf_nodes * node_type::LeafSize() + + stats.internal_nodes * node_type::InternalSize(); + } + } + + // The average number of bytes used per value stored in the btree. + 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 fullness of the btree. Computed as the number of elements in the btree + // divided by the maximum number of elements a tree with the current number + // of nodes could hold. A value of 1 indicates perfect space + // utilization. Smaller values indicate space wastage. + // Returns 0 for empty trees. + double fullness() const { + if (empty()) return 0.0; + return static_cast(size()) / (nodes() * kNodeValues); + } + // 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 + // storing elements divided by the number of elements. + // Returns 0 for empty trees. + double overhead() const { + if (empty()) return 0.0; + return (bytes_used() - size() * sizeof(value_type)) / static_cast(size()); + } + + // The allocator used by the btree. + allocator_type get_allocator() const { return allocator(); } + + private: + // Internal accessor routines. + node_type* root() { return root_.template get<2>(); } + const node_type* root() const { return root_.template get<2>(); } + node_type*& mutable_root() noexcept { return root_.template get<2>(); } + key_compare* mutable_key_comp() noexcept { return &root_.template get<0>(); } + + // The leftmost node is stored as the parent of the root node. + node_type* leftmost() { return root()->parent(); } + const node_type* leftmost() const { return root()->parent(); } + + // Allocator routines. + allocator_type* mutable_allocator() noexcept { return &root_.template get<1>(); } + const allocator_type& allocator() const noexcept { return root_.template get<1>(); } + + // Allocates a correctly aligned node of at least size bytes using the + // allocator. + node_type* allocate(const size_type sz) { + return reinterpret_cast( + phmap::priv::Allocate(mutable_allocator(), (size_t)sz)); + } + + // Node creation/deletion routines. + node_type* new_internal_node(node_type* parent) { + node_type* p = allocate(node_type::InternalSize()); + return node_type::init_internal(p, parent); + } + node_type* new_leaf_node(node_type* parent) { + node_type* p = allocate(node_type::LeafSize()); + return node_type::init_leaf(p, parent, kNodeValues); + } + node_type* new_leaf_root_node(const int max_count) { + node_type* p = allocate(node_type::LeafSize(max_count)); + return node_type::init_leaf(p, p, max_count); + } + + // Deletion helper routines. + void erase_same_node(iterator begin, iterator end); + iterator erase_from_leaf_node(iterator begin, size_type to_erase); + iterator rebalance_after_delete(iterator iter); + + // Deallocates a node of a certain size in bytes using the allocator. + void deallocate(const size_type sz, node_type* node) { + phmap::priv::Deallocate(mutable_allocator(), node, + (size_t)sz); + } + + void delete_internal_node(node_type* node) { + node->destroy(mutable_allocator()); + deallocate(node_type::InternalSize(), node); + } + void delete_leaf_node(node_type* node) { + node->destroy(mutable_allocator()); + deallocate(node_type::LeafSize(node->max_count()), node); + } + + // Rebalances or splits the node iter points to. + void rebalance_or_split(iterator* iter); + + // Merges the values of left, right and the delimiting key on their parent + // onto left, removing the delimiting key and deleting right. + void merge_nodes(node_type* left, node_type* right); + + // Tries to merge node with its left or right sibling, and failing that, + // rebalance with its left or right sibling. Returns true if a merge + // occurred, at which point it is no longer valid to access node. Returns + // false if no merging took place. + bool try_merge_or_rebalance(iterator* iter); + + // Tries to shrink the height of the tree by 1. + void try_shrink(); + + iterator internal_end(iterator iter) { return iter.node != nullptr ? iter : end(); } + const_iterator internal_end(const_iterator iter) const { + return iter.node != nullptr ? iter : end(); + } + + // Emplaces a value into the btree immediately before iter. Requires that + // key(v) <= iter.key() and (--iter).key() <= key(v). + template + iterator internal_emplace(iterator iter, Args&&... args); + + // Returns an iterator pointing to the first value >= the value "iter" is + // pointing at. Note that "iter" might be pointing to an invalid location as + // iter.position == iter.node->count(). This routine simply moves iter up in + // the tree to a valid location. + // Requires: iter.node is non-null. + template + 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. + template + SearchResult internal_locate(const K& key) const; + + template + SearchResult internal_locate_impl( + const K& key, std::false_type /* IsCompareTo */) const; + + template + SearchResult internal_locate_impl( + const K& key, std::true_type /* IsCompareTo */) const; + + // Internal routine which implements lower_bound(). + template + iterator internal_lower_bound(const K& key) const; + + // Internal routine which implements upper_bound(). + template + iterator internal_upper_bound(const K& key) const; + + // Internal routine which implements find(). + template + iterator internal_find(const K& key) const; + + // Deletes a node and all of its children. + void internal_clear(node_type* node); + + // Verifies the tree structure of node. + int internal_verify(const node_type* node, const key_type* lo, + const key_type* hi) const; + + node_stats internal_stats(const node_type* node) const { + // The root can be a static empty node. + if (node == nullptr || (node == root() && empty())) { + return node_stats(0, 0); + } + if (node->leaf()) { + return node_stats(1, 0); + } + node_stats res(0, 1); + for (int i = 0; i <= node->count(); ++i) { + res += internal_stats(node->child(i)); + } + 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. + phmap::priv::CompressedTuple root_; + + // A pointer to the rightmost node. Note that the leftmost node is stored as + // the root's parent. + node_type* rightmost_; + + // Number of values. + size_type size_; +}; + +//// +// btree_node methods +template +template +inline void btree_node

::emplace_value(const size_type i, allocator_type* alloc, + Args&&... args) { + assert(i <= count()); + // Shift old values to create space for new value and then construct it in + // place. + if (i < count()) { + value_init(count(), alloc, slot(count() - 1)); + for (size_type j = count() - 1; j > i; --j) + params_type::move(alloc, slot(j - 1), slot(j)); + value_destroy(i, alloc); + } + value_init(i, alloc, std::forward(args)...); + set_count((field_type)(count() + 1)); + + if (!leaf() && count() > i + 1) { + for (int j = count(); j > (int)(i + 1); --j) { + set_child(j, child(j - 1)); + } + clear_child(i + 1); + } +} + +template +inline void btree_node

::remove_value(const int i, allocator_type* alloc) { + if (!leaf() && count() > i + 1) { + assert(child(i + 1)->count() == 0); + for (size_type j = i + 1; j < count(); ++j) { + set_child(j, child(j + 1)); + } + clear_child(count()); + } + + remove_values_ignore_children(i, /*to_erase=*/1, alloc); +} + +template +inline void btree_node

::remove_values_ignore_children(int i, size_type to_erase, + allocator_type* alloc) { + params_type::move(alloc, slot(i + to_erase), slot(count()), slot(i)); + value_destroy_n(count() - to_erase, to_erase, alloc); + set_count((field_type)(count() - to_erase)); +} + +template +void btree_node

::rebalance_right_to_left(const int to_move, btree_node* right, + allocator_type* alloc) { + assert(parent() == right->parent()); + assert(position() + 1 == right->position()); + assert(right->count() >= count()); + assert(to_move >= 1); + assert(to_move <= right->count()); + + // 1) Move the delimiting value in the parent to the left node. + value_init(count(), alloc, parent()->slot(position())); + + // 2) Move the (to_move - 1) values from the right node to the left node. + right->uninitialized_move_n(to_move - 1, 0, count() + 1, this, alloc); + + // 3) Move the new delimiting value to the parent from the right node. + params_type::move(alloc, right->slot(to_move - 1), parent()->slot(position())); + + // 4) Shift the values in the right node to their correct position. + params_type::move(alloc, right->slot(to_move), right->slot(right->count()), + right->slot(0)); + + // 5) Destroy the now-empty to_move entries in the right node. + right->value_destroy_n(right->count() - to_move, to_move, alloc); + + if (!leaf()) { + // Move the child pointers from the right to the left node. + for (int i = 0; i < to_move; ++i) { + init_child(count() + i + 1, right->child(i)); + } + for (int i = 0; i <= right->count() - to_move; ++i) { + assert(i + to_move <= right->max_count()); + right->init_child(i, right->child(i + to_move)); + right->clear_child(i + to_move); + } + } + + // Fixup the counts on the left and right nodes. + set_count((field_type)(count() + to_move)); + right->set_count((field_type)(right->count() - to_move)); +} + +template +void btree_node

::rebalance_left_to_right(const int to_move, btree_node* right, + allocator_type* alloc) { + assert(parent() == right->parent()); + assert(position() + 1 == right->position()); + assert(count() >= right->count()); + assert(to_move >= 1); + assert(to_move <= count()); + + // Values in the right node are shifted to the right to make room for the + // new to_move values. Then, the delimiting value in the parent and the + // other (to_move - 1) values in the left node are moved into the right node. + // Lastly, a new delimiting value is moved from the left node into the + // parent, and the remaining empty left node entries are destroyed. + + if (right->count() >= to_move) { + // The original location of the right->count() values are sufficient to hold + // the new to_move entries from the parent and left node. + + // 1) Shift existing values in the right node to their correct positions. + right->uninitialized_move_n(to_move, right->count() - to_move, right->count(), right, + alloc); + for (slot_type *src = right->slot(right->count() - to_move - 1), + *dest = right->slot(right->count() - 1), *end = right->slot(0); + src >= end; --src, --dest) { + params_type::move(alloc, src, dest); + } + + // 2) Move the delimiting value in the parent to the right node. + params_type::move(alloc, parent()->slot(position()), right->slot(to_move - 1)); + + // 3) Move the (to_move - 1) values from the left node to the right node. + params_type::move(alloc, slot(count() - (to_move - 1)), slot(count()), + right->slot(0)); + } else { + // The right node does not have enough initialized space to hold the new + // to_move entries, so part of them will move to uninitialized space. + + // 1) Shift existing values in the right node to their correct positions. + right->uninitialized_move_n(right->count(), 0, to_move, right, alloc); + + // 2) Move the delimiting value in the parent to the right node. + right->value_init(to_move - 1, alloc, parent()->slot(position())); + + // 3) Move the (to_move - 1) values from the left node to the right node. + const size_type uninitialized_remaining = to_move - right->count() - 1; + uninitialized_move_n(uninitialized_remaining, count() - uninitialized_remaining, + right->count(), right, alloc); + params_type::move(alloc, slot(count() - (to_move - 1)), + slot(count() - uninitialized_remaining), right->slot(0)); + } + + // 4) Move the new delimiting value to the parent from the left node. + params_type::move(alloc, slot(count() - to_move), parent()->slot(position())); + + // 5) Destroy the now-empty to_move entries in the left node. + value_destroy_n(count() - to_move, to_move, alloc); + + if (!leaf()) { + // Move the child pointers from the left to the right node. + for (int i = right->count(); i >= 0; --i) { + right->init_child(i + to_move, right->child(i)); + right->clear_child(i); + } + for (int i = 1; i <= to_move; ++i) { + right->init_child(i - 1, child(count() - to_move + i)); + clear_child(count() - to_move + i); + } + } + + // Fixup the counts on the left and right nodes. + set_count((field_type)(count() - to_move)); + right->set_count((field_type)(right->count() + to_move)); +} + +template +void btree_node

::split(const int insert_position, btree_node* dest, + allocator_type* alloc) { + assert(dest->count() == 0); + assert(max_count() == kNodeValues); + + // 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 + // more values on the right node. If we're inserting at the end of the + // right node then bias the split to put more values on the left node. + if (insert_position == 0) { + dest->set_count((field_type)(count() - 1)); + } else if (insert_position == kNodeValues) { + dest->set_count(0); + } else { + dest->set_count((field_type)(count() / 2)); + } + set_count((field_type)(count() - dest->count())); + assert(count() >= 1); + + // Move values from the left sibling to the right sibling. + uninitialized_move_n(dest->count(), count(), 0, dest, alloc); + + // Destroy the now-empty entries in the left node. + value_destroy_n(count(), dest->count(), alloc); + + // The split key is the largest value in the left sibling. + set_count((field_type)(count() - 1)); + parent()->emplace_value(position(), alloc, slot(count())); + value_destroy(count(), alloc); + parent()->init_child(position() + 1, dest); + + if (!leaf()) { + for (int i = 0; i <= dest->count(); ++i) { + assert(child(count() + i + 1) != nullptr); + dest->init_child(i, child(count() + i + 1)); + clear_child(count() + i + 1); + } + } +} + +template +void btree_node

::merge(btree_node* src, allocator_type* alloc) { + assert(parent() == src->parent()); + assert(position() + 1 == src->position()); + + // Move the delimiting value to the left node. + value_init(count(), alloc, parent()->slot(position())); + + // Move the values from the right to the left node. + src->uninitialized_move_n(src->count(), 0, count() + 1, this, alloc); + + // Destroy the now-empty entries in the right node. + src->value_destroy_n(0, src->count(), alloc); + + if (!leaf()) { + // Move the child pointers from the right to the left node. + for (int i = 0; i <= src->count(); ++i) { + init_child(count() + i + 1, src->child(i)); + src->clear_child(i); + } + } + + // Fixup the counts on the src and dest nodes. + set_count((field_type)(1 + count() + src->count())); + src->set_count(0); + + // Remove the value on the parent node. + parent()->remove_value(position(), alloc); +} + +template +void btree_node

::swap(btree_node* x, allocator_type* alloc) { + using std::swap; + assert(leaf() == x->leaf()); + + // Determine which is the smaller/larger node. + btree_node *smaller = this, *larger = x; + if (smaller->count() > larger->count()) { + swap(smaller, larger); + } + + // Swap the values. + for (slot_type *a = smaller->slot(0), *b = larger->slot(0), *end = a + smaller->count(); + a != end; ++a, ++b) { + params_type::swap(alloc, a, b); + } + + // Move values that can't be swapped. + const size_type to_move = larger->count() - smaller->count(); + larger->uninitialized_move_n(to_move, smaller->count(), smaller->count(), smaller, + alloc); + larger->value_destroy_n(smaller->count(), to_move, alloc); + + if (!leaf()) { + // Swap the child pointers. + std::swap_ranges(&smaller->mutable_child(0), + &smaller->mutable_child(smaller->count() + 1), + &larger->mutable_child(0)); + // Update swapped children's parent pointers. + int i = 0; + for (; i <= smaller->count(); ++i) { + smaller->child(i)->set_parent(smaller); + larger->child(i)->set_parent(larger); + } + // Move the child pointers that couldn't be swapped. + for (; i <= larger->count(); ++i) { + smaller->init_child(i, larger->child(i)); + larger->clear_child(i); + } + } + + // Swap the counts. + swap(mutable_count(), x->mutable_count()); +} + +//// +// btree_iterator methods +template +void btree_iterator::increment_slow() { + if (node->leaf()) { + assert(position >= node->count()); + btree_iterator save(*this); + while (position == node->count() && !node->is_root()) { + assert(node->parent()->child(node->position()) == node); + position = node->position(); + node = node->parent(); + } + if (position == node->count()) { + *this = save; + } + } else { + assert(position < node->count()); + node = node->child(position + 1); + while (!node->leaf()) { + node = node->child(0); + } + position = 0; + } +} + +template +void btree_iterator::decrement_slow() { + if (node->leaf()) { + assert(position <= -1); + btree_iterator save(*this); + while (position < 0 && !node->is_root()) { + assert(node->parent()->child(node->position()) == node); + position = node->position() - 1; + node = node->parent(); + } + if (position < 0) { + *this = save; + } + } else { + assert(position >= 0); + node = node->child(position); + while (!node->leaf()) { + node = node->child(node->count()); + } + position = node->count() - 1; + } +} + +//// +// btree methods +template +template +void btree

::copy_or_move_values_in_order(Btree* x) { + static_assert( + std::is_same::value || std::is_same::value, + "Btree type must be same or const."); + assert(empty()); + + // We can avoid key comparisons because we know the order of the + // values is the same order we'll store them in. + auto iter = x->begin(); + if (iter == x->end()) return; + insert_multi(maybe_move_from_iterator(iter)); + ++iter; + for (; iter != x->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)); + } +} + +template +constexpr bool btree

::static_assert_validation() { + static_assert(std::is_nothrow_copy_constructible::value, + "Key comparison must be nothrow copy constructible"); + static_assert(std::is_nothrow_copy_constructible::value, + "Allocator must be nothrow copy constructible"); + static_assert(type_traits_internal::is_trivially_copyable::value, + "iterator not trivially copyable."); + + // 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))), + "target node size too large"); + + // Verify that key_compare returns an phmap::{weak,strong}_ordering or bool. + using compare_result_type = phmap::invoke_result_t; + static_assert(std::is_same::value || + std::is_convertible::value, + "key comparison function must return phmap::{weak,strong}_ordering or " + "bool."); + + // Test the assumption made in setting kNodeValueSpace. + static_assert(node_type::MinimumOverhead() >= sizeof(void*) + 4, + "node space assumption incorrect"); + + return true; +} + +template +btree

::btree(const key_compare& comp, const allocator_type& alloc) + : root_(comp, alloc, EmptyNode()), rightmost_(EmptyNode()), size_(0) {} + +template +btree

::btree(const btree& x) : btree(x.key_comp(), x.allocator()) { + copy_or_move_values_in_order(&x); +} + +template +template +auto btree

::insert_unique(const key_type& key, Args&&... args) + -> std::pair { + if (empty()) { + mutable_root() = rightmost_ = new_leaf_root_node(1); + } + + auto res = internal_locate(key); + iterator& iter = res.value; + + if (res.HasMatch()) { + if (res.IsEq()) { + // The key already exists in the tree, do nothing. + return {iter, false}; + } + } else { + iterator last = internal_last(iter); + if (last.node && !compare_keys(key, last.key())) { + // The key already exists in the tree, do nothing. + return {last, false}; + } + } + return {internal_emplace(iter, std::forward(args)...), true}; +} + +template +template +inline auto btree

::insert_hint_unique(iterator position, const key_type& key, + Args&&... args) -> std::pair { + if (!empty()) { + if (position == end() || compare_keys(key, position.key())) { + iterator prev = position; + if (position == begin() || compare_keys((--prev).key(), key)) { + // prev.key() < key < position.key() + return {internal_emplace(position, std::forward(args)...), true}; + } + } else if (compare_keys(position.key(), key)) { + ++position; + if (position == end() || compare_keys(key, position.key())) { + // {original `position`}.key() < key < {current `position`}.key() + return {internal_emplace(position, std::forward(args)...), true}; + } + } else { + // position.key() == key + return {position, false}; + } + } + return insert_unique(key, std::forward(args)...); +} + +template +template +void btree

::insert_iterator_unique(InputIterator b, InputIterator e) { + for (; b != e; ++b) { + insert_hint_unique(end(), params_type::key(*b), *b); + } +} + +template +template +auto btree

::insert_multi(const key_type& key, ValueType&& v) -> iterator { + if (empty()) { + mutable_root() = rightmost_ = new_leaf_root_node(1); + } + + iterator iter = internal_upper_bound(key); + if (iter.node == nullptr) { + iter = end(); + } + return internal_emplace(iter, std::forward(v)); +} + +template +template +auto btree

::insert_hint_multi(iterator position, ValueType&& v) -> iterator { + if (!empty()) { + const key_type& key = params_type::key(v); + if (position == end() || !compare_keys(position.key(), key)) { + iterator prev = position; + if (position == begin() || !compare_keys(key, (--prev).key())) { + // prev.key() <= key <= position.key() + return internal_emplace(position, std::forward(v)); + } + } else { + iterator next = position; + ++next; + if (next == end() || !compare_keys(next.key(), key)) { + // position.key() < key <= next.key() + return internal_emplace(next, std::forward(v)); + } + } + } + return insert_multi(std::forward(v)); +} + +template +template +void btree

::insert_iterator_multi(InputIterator b, InputIterator e) { + for (; b != e; ++b) { + insert_hint_multi(end(), *b); + } +} + +template +auto btree

::operator=(const btree& x) -> btree& { + if (this != &x) { + clear(); + + *mutable_key_comp() = x.key_comp(); + if (phmap::allocator_traits< + allocator_type>::propagate_on_container_copy_assignment::value) { + *mutable_allocator() = x.allocator(); + } + + copy_or_move_values_in_order(&x); + } + return *this; +} + +template +auto btree

::operator=(btree&& x) noexcept -> btree& { + if (this != &x) { + clear(); + + using std::swap; + if (phmap::allocator_traits< + allocator_type>::propagate_on_container_copy_assignment::value) { + // Note: `root_` also contains the allocator and the key comparator. + swap(root_, x.root_); + swap(rightmost_, x.rightmost_); + swap(size_, x.size_); + } else { + if (allocator() == x.allocator()) { + swap(mutable_root(), x.mutable_root()); + swap(*mutable_key_comp(), *x.mutable_key_comp()); + swap(rightmost_, x.rightmost_); + swap(size_, x.size_); + } else { + // We aren't allowed to propagate the allocator and the allocator is + // different so we can't take over its memory. We must move each element + // individually. We need both `x` and `this` to have `x`s key comparator + // while moving the values so we can't swap the key comparators. + *mutable_key_comp() = x.key_comp(); + copy_or_move_values_in_order(&x); + } + } + } + return *this; +} + +template +auto btree

::erase(iterator iter) -> iterator { + bool internal_delete = false; + if (!iter.node->leaf()) { + // Deletion of a value on an internal node. First, move the largest value + // from our left child here, then delete that position (in remove_value() + // below). We can get to the largest value from our left child by + // decrementing iter. + iterator internal_iter(iter); + --iter; + assert(iter.node->leaf()); + params_type::move(mutable_allocator(), iter.node->slot(iter.position), + internal_iter.node->slot(internal_iter.position)); + internal_delete = true; + } + + // Delete the key from the leaf. + iter.node->remove_value(iter.position, mutable_allocator()); + --size_; + + // We want to return the next value after the one we just erased. If we + // erased from an internal node (internal_delete == true), then the next + // value is ++(++iter). If we erased from a leaf node (internal_delete == + // false) then the next value is ++iter. Note that ++iter may point to an + // internal node and the value in the internal node may move to a leaf node + // (iter.node) when rebalancing is performed at the leaf level. + + iterator res = rebalance_after_delete(iter); + + // If we erased from an internal node, advance the iterator. + if (internal_delete) { + ++res; + } + return res; +} + +template +auto btree

::rebalance_after_delete(iterator iter) -> iterator { + // Merge/rebalance as we walk back up the tree. + iterator res(iter); + bool first_iteration = true; + for (;;) { + if (iter.node == root()) { + try_shrink(); + if (empty()) { + return end(); + } + break; + } + if (iter.node->count() >= kMinNodeValues) { + break; + } + bool merged = try_merge_or_rebalance(&iter); + // On the first iteration, we should update `res` with `iter` because `res` + // may have been invalidated. + if (first_iteration) { + res = iter; + first_iteration = false; + } + if (!merged) { + break; + } + iter.position = iter.node->position(); + iter.node = iter.node->parent(); + } + + // Adjust our return value. If we're pointing at the end of a node, advance + // the iterator. + if (res.position == res.node->count()) { + res.position = res.node->count() - 1; + ++res; + } + + return res; +} + +template +auto btree

::erase(iterator _begin, iterator _end) -> std::pair { + difference_type count = std::distance(_begin, _end); + assert(count >= 0); + + if (count == 0) { + return {0, _begin}; + } + + if (count == (difference_type)size_) { + clear(); + return {count, this->end()}; + } + + if (_begin.node == _end.node) { + erase_same_node(_begin, _end); + size_ -= count; + return {count, rebalance_after_delete(_begin)}; + } + + const size_type target_size = size_ - count; + while (size_ > target_size) { + if (_begin.node->leaf()) { + const size_type remaining_to_erase = size_ - target_size; + const size_type remaining_in_node = _begin.node->count() - _begin.position; + _begin = + erase_from_leaf_node(_begin, (std::min)(remaining_to_erase, remaining_in_node)); + } else { + _begin = erase(_begin); + } + } + return {count, _begin}; +} + +template +void btree

::erase_same_node(iterator _begin, iterator _end) { + assert(_begin.node == _end.node); + assert(_end.position > _begin.position); + + node_type* node = _begin.node; + size_type to_erase = _end.position - _begin.position; + if (!node->leaf()) { + // Delete all children between _begin and _end. + for (size_type i = 0; i < to_erase; ++i) { + internal_clear(node->child(_begin.position + i + 1)); + } + // Rotate children after _end into new positions. + for (size_type i = _begin.position + to_erase + 1; i <= node->count(); ++i) { + node->set_child(i - to_erase, node->child(i)); + node->clear_child(i); + } + } + node->remove_values_ignore_children(_begin.position, to_erase, mutable_allocator()); + + // Do not need to update rightmost_, because + // * either _end == this->end(), and therefore node == rightmost_, and still + // exists + // * or _end != this->end(), and therefore rightmost_ hasn't been erased, since + // it wasn't covered in [_begin, _end) +} + +template +auto btree

::erase_from_leaf_node(iterator _begin, size_type to_erase) -> iterator { + node_type* node = _begin.node; + assert(node->leaf()); + assert(node->count() > _begin.position); + assert(_begin.position + to_erase <= node->count()); + + node->remove_values_ignore_children(_begin.position, to_erase, mutable_allocator()); + + size_ -= to_erase; + + return rebalance_after_delete(_begin); +} + +template +template +auto btree

::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 +template +auto btree

::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(_begin, _end).first; +} + +template +void btree

::clear() { + if (!empty()) { + internal_clear(root()); + } + mutable_root() = EmptyNode(); + rightmost_ = EmptyNode(); + size_ = 0; +} + +template +void btree

::swap(btree& x) { + using std::swap; + if (phmap::allocator_traits::propagate_on_container_swap::value) { + // Note: `root_` also contains the allocator and the key comparator. + swap(root_, x.root_); + } else { + // It's undefined behavior if the allocators are unequal here. + assert(allocator() == x.allocator()); + swap(mutable_root(), x.mutable_root()); + swap(*mutable_key_comp(), *x.mutable_key_comp()); + } + swap(rightmost_, x.rightmost_); + swap(size_, x.size_); +} + +template +void btree

::verify() const { + assert(root() != nullptr); + assert(leftmost() != nullptr); + assert(rightmost_ != nullptr); + assert(empty() || size() == internal_verify(root(), nullptr, nullptr)); + assert(leftmost() == (++const_iterator(root(), -1)).node); + assert(rightmost_ == (--const_iterator(root(), root()->count())).node); + assert(leftmost()->leaf()); + assert(rightmost_->leaf()); +} + +template +void btree

::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()); + + // First try to make room on the node by rebalancing. + node_type* parent = node->parent(); + if (node != root()) { + if (node->position() > 0) { + // Try rebalancing with our left sibling. + node_type* left = parent->child(node->position() - 1); + assert(left->max_count() == kNodeValues); + if (left->count() < kNodeValues) { + // 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)); + to_move = (std::max)(1, to_move); + + if (((insert_position - to_move) >= 0) || + ((left->count() + to_move) < kNodeValues)) { + left->rebalance_right_to_left(to_move, node, mutable_allocator()); + + assert(node->max_count() - node->count() == to_move); + insert_position = insert_position - to_move; + if (insert_position < 0) { + insert_position = insert_position + left->count() + 1; + node = left; + } + + assert(node->count() < node->max_count()); + return; + } + } + } + + if (node->position() < parent->count()) { + // Try rebalancing with our right sibling. + node_type* right = parent->child(node->position() + 1); + assert(right->max_count() == kNodeValues); + if (right->count() < kNodeValues) { + // 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()) / (1 + (insert_position > 0)); + to_move = (std::max)(1, to_move); + + if ((insert_position <= (node->count() - to_move)) || + ((right->count() + to_move) < kNodeValues)) { + node->rebalance_left_to_right(to_move, right, mutable_allocator()); + + if (insert_position > node->count()) { + insert_position = insert_position - node->count() - 1; + node = right; + } + + assert(node->count() < node->max_count()); + return; + } + } + } + + // Rebalancing failed, make sure there is room on the parent node for a new + // value. + assert(parent->max_count() == kNodeValues); + if (parent->count() == kNodeValues) { + iterator parent_iter(node->parent(), node->position()); + rebalance_or_split(&parent_iter); + } + } else { + // Rebalancing not possible because this is the root node. + // Create a new root node and set the current root node as the child of the + // new root. + parent = new_internal_node(parent); + parent->init_child(0, root()); + mutable_root() = parent; + // If the former root was a leaf node, then it's now the rightmost node. + assert(!parent->child(0)->leaf() || parent->child(0) == rightmost_); + } + + // Split the node. + node_type* split_node; + if (node->leaf()) { + split_node = new_leaf_node(parent); + node->split(insert_position, split_node, mutable_allocator()); + if (rightmost_ == node) rightmost_ = split_node; + } else { + split_node = new_internal_node(parent); + node->split(insert_position, split_node, mutable_allocator()); + } + + if (insert_position > node->count()) { + insert_position = insert_position - node->count() - 1; + node = split_node; + } +} + +template +void btree

::merge_nodes(node_type* left, node_type* right) { + left->merge(right, mutable_allocator()); + if (right->leaf()) { + if (rightmost_ == right) rightmost_ = left; + delete_leaf_node(right); + } else { + delete_internal_node(right); + } +} + +template +bool btree

::try_merge_or_rebalance(iterator* iter) { + node_type* parent = iter->node->parent(); + if (iter->node->position() > 0) { + // 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) { + iter->position += 1 + left->count(); + merge_nodes(left, iter->node); + iter->node = left; + return true; + } + } + if (iter->node->position() < parent->count()) { + // 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) { + merge_nodes(iter->node, right); + return true; + } + // Try rebalancing with our right sibling. We don't perform rebalancing if + // we deleted the first element from iter->node and the node is not + // empty. This is a small optimization for the common pattern of deleting + // from the front of the tree. + if ((right->count() > kMinNodeValues) && + ((iter->node->count() == 0) || (iter->position > 0))) { + int to_move = (right->count() - iter->node->count()) / 2; + to_move = (std::min)(to_move, right->count() - 1); + iter->node->rebalance_right_to_left(to_move, right, mutable_allocator()); + return false; + } + } + if (iter->node->position() > 0) { + // Try rebalancing with our left sibling. We don't perform rebalancing if + // we deleted the last element from iter->node and the node is not + // empty. This is a small optimization for the common pattern of deleting + // from the back of the tree. + node_type* left = parent->child(iter->node->position() - 1); + if ((left->count() > kMinNodeValues) && + ((iter->node->count() == 0) || (iter->position < iter->node->count()))) { + int to_move = (left->count() - iter->node->count()) / 2; + to_move = (std::min)(to_move, left->count() - 1); + left->rebalance_left_to_right(to_move, iter->node, mutable_allocator()); + iter->position += to_move; + return false; + } + } + return false; +} + +template +void btree

::try_shrink() { + if (root()->count() > 0) { + return; + } + // Deleted the last item on the root node, shrink the height of the tree. + if (root()->leaf()) { + assert(size() == 0); + delete_leaf_node(root()); + mutable_root() = EmptyNode(); + rightmost_ = EmptyNode(); + } else { + node_type* child = root()->child(0); + child->make_root(); + delete_internal_node(root()); + mutable_root() = child; + } +} + +template +template +inline IterType btree

::internal_last(IterType iter) { + assert(iter.node != nullptr); + while (iter.position == iter.node->count()) { + iter.position = iter.node->position(); + iter.node = iter.node->parent(); + if (iter.node->leaf()) { + iter.node = nullptr; + break; + } + } + return iter; +} + +template +template +inline auto btree

::internal_emplace(iterator iter, Args&&... args) -> iterator { + if (!iter.node->leaf()) { + // We can't insert on an internal node. Instead, we'll insert after the + // previous value which is guaranteed to be on a leaf node. + --iter; + ++iter.position; + } + const int max_count = iter.node->max_count(); + if (iter.node->count() == max_count) { + // Make room in the leaf for the new item. + if (max_count < kNodeValues) { + // 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)(kNodeValues, 2 * max_count)); + iter.node->swap(root(), mutable_allocator()); + delete_leaf_node(root()); + mutable_root() = iter.node; + rightmost_ = iter.node; + } else { + rebalance_or_split(&iter); + } + } + iter.node->emplace_value(iter.position, mutable_allocator(), + std::forward(args)...); + ++size_; + return iter; +} + +template +template +inline auto btree

::internal_locate(const K& key) const + -> SearchResult { + return internal_locate_impl(key, is_key_compare_to()); +} + +template +template +inline auto btree

::internal_locate_impl(const K& key, + std::false_type /* IsCompareTo */) const + -> SearchResult { + iterator iter(const_cast(root()), 0); + 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 +template +inline auto btree

::internal_locate_impl(const K& key, + std::true_type /* IsCompareTo */) const + -> SearchResult { + iterator iter(const_cast(root()), 0); + for (;;) { + SearchResult res = iter.node->lower_bound(key, key_comp()); + iter.position = res.value; + if (res.match == MatchKind::kEq) { + return {iter, MatchKind::kEq}; + } + if (iter.node->leaf()) { + break; + } + iter.node = iter.node->child(iter.position); + } + return {iter, MatchKind::kNe}; +} + +template +template +auto btree

::internal_lower_bound(const K& key) const -> iterator { + iterator iter(const_cast(root()), 0); + for (;;) { + iter.position = iter.node->lower_bound(key, key_comp()).value; + if (iter.node->leaf()) { + break; + } + iter.node = iter.node->child(iter.position); + } + return internal_last(iter); +} + +template +template +auto btree

::internal_upper_bound(const K& key) const -> iterator { + iterator iter(const_cast(root()), 0); + for (;;) { + iter.position = iter.node->upper_bound(key, key_comp()); + if (iter.node->leaf()) { + break; + } + iter.node = iter.node->child(iter.position); + } + return internal_last(iter); +} + +template +template +auto btree

::internal_find(const K& key) const -> iterator { + auto res = internal_locate(key); + if (res.HasMatch()) { + if (res.IsEq()) { + return res.value; + } + } else { + const iterator iter = internal_last(res.value); + if (iter.node != nullptr && !compare_keys(key, iter.key())) { + return iter; + } + } + return {nullptr, 0}; +} + +template +void btree

::internal_clear(node_type* node) { + if (!node->leaf()) { + for (int i = 0; i <= node->count(); ++i) { + internal_clear(node->child(i)); + } + delete_internal_node(node); + } else { + delete_leaf_node(node); + } +} + +template +int btree

::internal_verify(const node_type* node, const key_type* lo, + const key_type* hi) const { + assert(node->count() > 0); + assert(node->count() <= node->max_count()); + if (lo) { + assert(!compare_keys(node->key(0), *lo)); + } + if (hi) { + assert(!compare_keys(*hi, node->key(node->count() - 1))); + } + for (int i = 1; i < node->count(); ++i) { + assert(!compare_keys(node->key(i), node->key(i - 1))); + } + int count = node->count(); + if (!node->leaf()) { + for (int i = 0; i <= node->count(); ++i) { + assert(node->child(i) != nullptr); + assert(node->child(i)->parent() == node); + assert(node->child(i)->position() == i); + count += internal_verify(node->child(i), (i == 0) ? lo : &node->key(i - 1), + (i == node->count()) ? hi : &node->key(i)); + } + } + return count; +} + +// A common base class for btree_set, btree_map, btree_multiset, and btree_multimap. +// --------------------------------------------------------------------------------- +template +class btree_container { + using params_type = typename Tree::params_type; + + protected: + // Alias used for heterogeneous lookup functions. + // `key_arg` evaluates to `K` when the functors are transparent and to + // `key_type` otherwise. It permits template argument deduction on `K` for the + // transparent case. + template + using key_arg = + typename KeyArg::value>::template type< + K, typename Tree::key_type>; + + public: + using key_type = typename Tree::key_type; + using value_type = typename Tree::value_type; + using size_type = typename Tree::size_type; + using difference_type = typename Tree::difference_type; + using key_compare = typename Tree::key_compare; + using value_compare = typename Tree::value_compare; + using allocator_type = typename Tree::allocator_type; + using reference = typename Tree::reference; + using const_reference = typename Tree::const_reference; + using pointer = typename Tree::pointer; + using const_pointer = typename Tree::const_pointer; + using iterator = typename Tree::iterator; + using const_iterator = typename Tree::const_iterator; + using reverse_iterator = typename Tree::reverse_iterator; + using const_reverse_iterator = typename Tree::const_reverse_iterator; + using node_type = typename Tree::node_handle_type; + + // Constructors/assignments. + btree_container() : tree_(key_compare(), allocator_type()) {} + explicit btree_container(const key_compare& comp, + const allocator_type& alloc = allocator_type()) + : tree_(comp, alloc) {} + btree_container(const btree_container& x) = default; + btree_container(btree_container&& x) noexcept = default; + btree_container& operator=(const btree_container& x) = default; + btree_container& operator=(btree_container&& x) noexcept( + std::is_nothrow_move_assignable::value) = default; + + // Iterator routines. + iterator begin() { return tree_.begin(); } + const_iterator begin() const { return tree_.begin(); } + const_iterator cbegin() const { return tree_.begin(); } + iterator end() { return tree_.end(); } + const_iterator end() const { return tree_.end(); } + const_iterator cend() const { return tree_.end(); } + reverse_iterator rbegin() { return tree_.rbegin(); } + const_reverse_iterator rbegin() const { return tree_.rbegin(); } + const_reverse_iterator crbegin() const { return tree_.rbegin(); } + reverse_iterator rend() { return tree_.rend(); } + const_reverse_iterator rend() const { return tree_.rend(); } + const_reverse_iterator crend() const { return tree_.rend(); } + + // Lookup routines. + template + iterator find(const key_arg& key) { + return tree_.find(key); + } + template + const_iterator find(const key_arg& key) const { + return tree_.find(key); + } + + template + bool contains(const key_arg& key) const { + return find(key) != end(); + } + + template + iterator lower_bound(const key_arg& key) { + return tree_.lower_bound(key); + } + + template + const_iterator lower_bound(const key_arg& key) const { + return tree_.lower_bound(key); + } + + template + iterator upper_bound(const key_arg& key) { + return tree_.upper_bound(key); + } + + template + const_iterator upper_bound(const key_arg& key) const { + return tree_.upper_bound(key); + } + + template + std::pair equal_range(const key_arg& key) { + return tree_.equal_range(key); + } + + template + std::pair equal_range(const key_arg& key) const { + return tree_.equal_range(key); + } + + iterator erase(const_iterator iter) { return tree_.erase(iterator(iter)); } + iterator erase(iterator iter) { return tree_.erase(iter); } + iterator erase(const_iterator first, const_iterator last) { + return tree_.erase(iterator(first), iterator(last)).second; + } + + node_type extract(iterator position) { + // Use Move instead of Transfer, because the rebalancing code expects to + // have a valid object to scribble metadata bits on top of. + auto node = CommonAccess::Move(get_allocator(), position.slot()); + erase(position); + return node; + } + + node_type extract(const_iterator position) { return extract(iterator(position)); } + + public: + void clear() { tree_.clear(); } + void swap(btree_container& x) { tree_.swap(x.tree_); } + void verify() const { tree_.verify(); } + + size_type size() const { return tree_.size(); } + size_type max_size() const { return tree_.max_size(); } + bool empty() const { return tree_.empty(); } + + friend bool operator==(const btree_container& x, const btree_container& y) { + if (x.size() != y.size()) return false; + return std::equal(x.begin(), x.end(), y.begin()); + } + + friend bool operator!=(const btree_container& x, const btree_container& y) { + return !(x == y); + } + + friend bool operator<(const btree_container& x, const btree_container& y) { + return std::lexicographical_compare(x.begin(), x.end(), y.begin(), y.end()); + } + + friend bool operator>(const btree_container& x, const btree_container& y) { + return y < x; + } + + friend bool operator<=(const btree_container& x, const btree_container& y) { + return !(y < x); + } + + friend bool operator>=(const btree_container& x, const btree_container& y) { + return !(x < y); + } + + // The allocator used by the btree. + allocator_type get_allocator() const { return tree_.get_allocator(); } + + // The key comparator used by the btree. + key_compare key_comp() const { return tree_.key_comp(); } + value_compare value_comp() const { return tree_.value_comp(); } + + // Support absl::Hash. + template + friend State AbslHashValue(State h, const btree_container& b) { + for (const auto& v : b) { + h = State::combine(std::move(h), v); + } + return State::combine(std::move(h), b.size()); + } + + protected: + Tree tree_; +}; + +// A common base class for btree_set and btree_map. +// ----------------------------------------------- +template +class btree_set_container : public btree_container { + using super_type = btree_container; + using params_type = typename Tree::params_type; + using init_type = typename params_type::init_type; + using is_key_compare_to = typename params_type::is_key_compare_to; + friend class BtreeNodePeer; + + protected: + template + using key_arg = typename super_type::template key_arg; + + public: + using key_type = typename Tree::key_type; + using value_type = typename Tree::value_type; + using size_type = typename Tree::size_type; + using key_compare = typename Tree::key_compare; + using allocator_type = typename Tree::allocator_type; + using iterator = typename Tree::iterator; + using const_iterator = typename Tree::const_iterator; + using node_type = typename super_type::node_type; + using insert_return_type = InsertReturnType; + using super_type::super_type; + btree_set_container() {} + + template + btree_set_container(InputIterator b, InputIterator e, + const key_compare& comp = key_compare(), + const allocator_type& alloc = allocator_type()) + : super_type(comp, alloc) { + insert(b, e); + } + + btree_set_container(std::initializer_list init, + const key_compare& comp = key_compare(), + const allocator_type& alloc = allocator_type()) + : btree_set_container(init.begin(), init.end(), comp, alloc) {} + + // Lookup routines. + template + size_type count(const key_arg& key) const { + return this->tree_.count_unique(key); + } + + // Insertion routines. + std::pair insert(const value_type& x) { + return this->tree_.insert_unique(params_type::key(x), x); + } + std::pair insert(value_type&& x) { + return this->tree_.insert_unique(params_type::key(x), std::move(x)); + } + template + std::pair emplace(Args&&... args) { + init_type v(std::forward(args)...); + return this->tree_.insert_unique(params_type::key(v), std::move(v)); + } + iterator insert(const_iterator position, const value_type& x) { + return this->tree_.insert_hint_unique(iterator(position), params_type::key(x), x) + .first; + } + iterator insert(const_iterator position, value_type&& x) { + return this->tree_ + .insert_hint_unique(iterator(position), params_type::key(x), std::move(x)) + .first; + } + + template + iterator emplace_hint(const_iterator position, Args&&... args) { + init_type v(std::forward(args)...); + return this->tree_ + .insert_hint_unique(iterator(position), params_type::key(v), std::move(v)) + .first; + } + + template + void insert(InputIterator b, InputIterator e) { + this->tree_.insert_iterator_unique(b, e); + } + + void insert(std::initializer_list init) { + this->tree_.insert_iterator_unique(init.begin(), init.end()); + } + + insert_return_type insert(node_type&& node) { + if (!node) return {this->end(), false, node_type()}; + std::pair res = this->tree_.insert_unique( + params_type::key(CommonAccess::GetSlot(node)), CommonAccess::GetSlot(node)); + if (res.second) { + CommonAccess::Destroy(&node); + return {res.first, true, node_type()}; + } else { + return {res.first, false, std::move(node)}; + } + } + + iterator insert(const_iterator hint, node_type&& node) { + if (!node) return this->end(); + std::pair res = this->tree_.insert_hint_unique( + iterator(hint), params_type::key(CommonAccess::GetSlot(node)), + CommonAccess::GetSlot(node)); + if (res.second) CommonAccess::Destroy(&node); + return res.first; + } + + template + size_type erase(const key_arg& key) { + return this->tree_.erase_unique(key); + } + using super_type::erase; + + template + node_type extract(const key_arg& key) { + auto it = this->find(key); + return it == this->end() ? node_type() : extract(it); + } + + using super_type::extract; + + // Merge routines. + // Moves elements from `src` into `this`. If the element already exists in + // `this`, it is left unmodified in `src`. + template , + std::is_same, + std::is_same>::value, + int> = 0> + void merge(btree_container& src) { // NOLINT + for (auto src_it = src.begin(); src_it != src.end();) { + if (insert(std::move(*src_it)).second) { + src_it = src.erase(src_it); + } else { + ++src_it; + } + } + } + + template , + std::is_same, + std::is_same>::value, + int> = 0> + void merge(btree_container&& src) { + merge(src); + } +}; + +// Base class for btree_map. +// ------------------------- +template +class btree_map_container : public btree_set_container { + using super_type = btree_set_container; + using params_type = typename Tree::params_type; + + protected: + template + using key_arg = typename super_type::template key_arg; + + public: + using key_type = typename Tree::key_type; + using mapped_type = typename params_type::mapped_type; + using value_type = typename Tree::value_type; + using key_compare = typename Tree::key_compare; + using allocator_type = typename Tree::allocator_type; + using iterator = typename Tree::iterator; + using const_iterator = typename Tree::const_iterator; + + // Inherit constructors. + using super_type::super_type; + btree_map_container() {} + + // Insertion routines. + template + std::pair try_emplace(const key_type& k, Args&&... args) { + return this->tree_.insert_unique(k, std::piecewise_construct, + std::forward_as_tuple(k), + std::forward_as_tuple(std::forward(args)...)); + } + template + std::pair try_emplace(key_type&& k, Args&&... args) { + // Note: `key_ref` exists to avoid a ClangTidy warning about moving from `k` + // and then using `k` unsequenced. This is safe because the move is into a + // forwarding reference and insert_unique guarantees that `key` is never + // referenced after consuming `args`. + const key_type& key_ref = k; + return this->tree_.insert_unique(key_ref, std::piecewise_construct, + std::forward_as_tuple(std::move(k)), + std::forward_as_tuple(std::forward(args)...)); + } + template + iterator try_emplace(const_iterator hint, const key_type& k, Args&&... args) { + return this->tree_ + .insert_hint_unique(iterator(hint), k, std::piecewise_construct, + std::forward_as_tuple(k), + std::forward_as_tuple(std::forward(args)...)) + .first; + } + template + iterator try_emplace(const_iterator hint, key_type&& k, Args&&... args) { + // Note: `key_ref` exists to avoid a ClangTidy warning about moving from `k` + // and then using `k` unsequenced. This is safe because the move is into a + // forwarding reference and insert_hint_unique guarantees that `key` is + // never referenced after consuming `args`. + const key_type& key_ref = k; + return this->tree_ + .insert_hint_unique(iterator(hint), key_ref, std::piecewise_construct, + std::forward_as_tuple(std::move(k)), + std::forward_as_tuple(std::forward(args)...)) + .first; + } + mapped_type& operator[](const key_type& k) { return try_emplace(k).first->second; } + mapped_type& operator[](key_type&& k) { + return try_emplace(std::move(k)).first->second; + } + + template + mapped_type& at(const key_arg& key) { + auto it = this->find(key); + if (it == this->end()) base_internal::ThrowStdOutOfRange("phmap::btree_map::at"); + return it->second; + } + template + const mapped_type& at(const key_arg& key) const { + auto it = this->find(key); + if (it == this->end()) base_internal::ThrowStdOutOfRange("phmap::btree_map::at"); + return it->second; + } +}; + +// A common base class for btree_multiset and btree_multimap. +template +class btree_multiset_container : public btree_container { + using super_type = btree_container; + using params_type = typename Tree::params_type; + using init_type = typename params_type::init_type; + using is_key_compare_to = typename params_type::is_key_compare_to; + + template + using key_arg = typename super_type::template key_arg; + + public: + using key_type = typename Tree::key_type; + using value_type = typename Tree::value_type; + using size_type = typename Tree::size_type; + using key_compare = typename Tree::key_compare; + using allocator_type = typename Tree::allocator_type; + using iterator = typename Tree::iterator; + using const_iterator = typename Tree::const_iterator; + using node_type = typename super_type::node_type; + + // Inherit constructors. + using super_type::super_type; + btree_multiset_container() {} + + // Range constructor. + template + btree_multiset_container(InputIterator b, InputIterator e, + const key_compare& comp = key_compare(), + const allocator_type& alloc = allocator_type()) + : super_type(comp, alloc) { + insert(b, e); + } + + // Initializer list constructor. + btree_multiset_container(std::initializer_list init, + const key_compare& comp = key_compare(), + const allocator_type& alloc = allocator_type()) + : btree_multiset_container(init.begin(), init.end(), comp, alloc) {} + + // Lookup routines. + template + size_type count(const key_arg& key) const { + return this->tree_.count_multi(key); + } + + // Insertion routines. + iterator insert(const value_type& x) { return this->tree_.insert_multi(x); } + iterator insert(value_type&& x) { return this->tree_.insert_multi(std::move(x)); } + iterator insert(const_iterator position, const value_type& x) { + return this->tree_.insert_hint_multi(iterator(position), x); + } + iterator insert(const_iterator position, value_type&& x) { + return this->tree_.insert_hint_multi(iterator(position), std::move(x)); + } + template + void insert(InputIterator b, InputIterator e) { + this->tree_.insert_iterator_multi(b, e); + } + void insert(std::initializer_list init) { + this->tree_.insert_iterator_multi(init.begin(), init.end()); + } + template + iterator emplace(Args&&... args) { + return this->tree_.insert_multi(init_type(std::forward(args)...)); + } + template + iterator emplace_hint(const_iterator position, Args&&... args) { + return this->tree_.insert_hint_multi(iterator(position), + init_type(std::forward(args)...)); + } + iterator insert(node_type&& node) { + if (!node) return this->end(); + iterator res = this->tree_.insert_multi(params_type::key(CommonAccess::GetSlot(node)), + CommonAccess::GetSlot(node)); + CommonAccess::Destroy(&node); + return res; + } + iterator insert(const_iterator hint, node_type&& node) { + if (!node) return this->end(); + iterator res = this->tree_.insert_hint_multi( + iterator(hint), std::move(params_type::element(CommonAccess::GetSlot(node)))); + CommonAccess::Destroy(&node); + return res; + } + + // Deletion routines. + template + size_type erase(const key_arg& key) { + return this->tree_.erase_multi(key); + } + using super_type::erase; + + // Node extraction routines. + template + node_type extract(const key_arg& key) { + auto it = this->find(key); + return it == this->end() ? node_type() : extract(it); + } + using super_type::extract; + + // Merge routines. + // Moves all elements from `src` into `this`. + template , + std::is_same, + std::is_same>::value, + int> = 0> + void merge(btree_container& src) { // NOLINT + insert(std::make_move_iterator(src.begin()), std::make_move_iterator(src.end())); + src.clear(); + } + + template , + std::is_same, + std::is_same>::value, + int> = 0> + void merge(btree_container&& src) { + merge(src); + } +}; + +// A base class for btree_multimap. +template +class btree_multimap_container : public btree_multiset_container { + using super_type = btree_multiset_container; + using params_type = typename Tree::params_type; + + public: + using mapped_type = typename params_type::mapped_type; + + // Inherit constructors. + using super_type::super_type; + btree_multimap_container() {} +}; + +} // namespace priv + +// ---------------------------------------------------------------------- +// btree_set - default values in phmap_fwd_decl.h +// ---------------------------------------------------------------------- +template +class btree_set : public priv::btree_set_container>> { + using Base = typename btree_set::btree_set_container; + + public: + btree_set() {} + using Base::Base; + using Base::begin; + using Base::cbegin; + using Base::cend; + using Base::clear; + using Base::contains; + using Base::count; + using Base::emplace; + using Base::emplace_hint; + using Base::empty; + using Base::end; + using Base::equal_range; + using Base::erase; + using Base::extract; + using Base::find; + using Base::get_allocator; + using Base::insert; + using Base::key_comp; + using Base::max_size; + using Base::merge; + using Base::size; + using Base::swap; + using Base::value_comp; +}; + +// Swaps the contents of two `phmap::btree_set` containers. +// ------------------------------------------------------- +template +void swap(btree_set& x, btree_set& y) { + return x.swap(y); +} + +// Erases all elements that satisfy the predicate pred from the container. +// ---------------------------------------------------------------------- +template +void erase_if(btree_set& set, Pred pred) { + for (auto it = set.begin(); it != set.end();) { + if (pred(*it)) { + it = set.erase(it); + } else { + ++it; + } + } +} + +// ---------------------------------------------------------------------- +// btree_multiset - default values in phmap_fwd_decl.h +// ---------------------------------------------------------------------- +template +class btree_multiset + : public priv::btree_multiset_container>> { + using Base = typename btree_multiset::btree_multiset_container; + + public: + btree_multiset() {} + using Base::Base; + using Base::begin; + using Base::cbegin; + using Base::cend; + using Base::clear; + using Base::contains; + using Base::count; + using Base::emplace; + using Base::emplace_hint; + using Base::empty; + using Base::end; + using Base::equal_range; + using Base::erase; + using Base::extract; + using Base::find; + using Base::get_allocator; + using Base::insert; + using Base::key_comp; + using Base::max_size; + using Base::merge; + using Base::size; + using Base::swap; + using Base::value_comp; +}; + +// Swaps the contents of two `phmap::btree_multiset` containers. +// ------------------------------------------------------------ +template +void swap(btree_multiset& x, btree_multiset& y) { + return x.swap(y); +} + +// Erases all elements that satisfy the predicate pred from the container. +// ---------------------------------------------------------------------- +template +void erase_if(btree_multiset& set, Pred pred) { + for (auto it = set.begin(); it != set.end();) { + if (pred(*it)) { + it = set.erase(it); + } else { + ++it; + } + } +} + +// ---------------------------------------------------------------------- +// btree_map - default values in phmap_fwd_decl.h +// ---------------------------------------------------------------------- +template +class btree_map + : public priv::btree_map_container>> { + using Base = typename btree_map::btree_map_container; + + public: + btree_map() {} + using Base::at; + using Base::Base; + using Base::begin; + using Base::cbegin; + using Base::cend; + using Base::clear; + using Base::contains; + using Base::count; + using Base::emplace; + using Base::emplace_hint; + using Base::empty; + using Base::end; + using Base::equal_range; + using Base::erase; + using Base::extract; + using Base::find; + using Base::insert; + using Base::max_size; + using Base::merge; + using Base::size; + using Base::swap; + using Base::try_emplace; + using Base::operator[]; + using Base::get_allocator; + using Base::key_comp; + using Base::value_comp; +}; + +// Swaps the contents of two `phmap::btree_map` containers. +// ------------------------------------------------------- +template +void swap(btree_map& x, btree_map& y) { + return x.swap(y); +} + +// ---------------------------------------------------------------------- +template +void erase_if(btree_map& map, Pred pred) { + for (auto it = map.begin(); it != map.end();) { + if (pred(*it)) { + it = map.erase(it); + } else { + ++it; + } + } +} + +// ---------------------------------------------------------------------- +// btree_multimap - default values in phmap_fwd_decl.h +// ---------------------------------------------------------------------- +template +class btree_multimap + : public priv::btree_multimap_container>> { + using Base = typename btree_multimap::btree_multimap_container; + + public: + btree_multimap() {} + using Base::Base; + using Base::begin; + using Base::cbegin; + using Base::cend; + using Base::clear; + using Base::contains; + using Base::count; + using Base::emplace; + using Base::emplace_hint; + using Base::empty; + using Base::end; + using Base::equal_range; + using Base::erase; + using Base::extract; + using Base::find; + using Base::get_allocator; + using Base::insert; + using Base::key_comp; + using Base::max_size; + using Base::merge; + using Base::size; + using Base::swap; + using Base::value_comp; +}; + +// Swaps the contents of two `phmap::btree_multimap` containers. +// ------------------------------------------------------------ +template +void swap(btree_multimap& x, btree_multimap& y) { + return x.swap(y); +} + +// Erases all elements that satisfy the predicate pred from the container. +// ---------------------------------------------------------------------- +template +void erase_if(btree_multimap& map, Pred pred) { + for (auto it = map.begin(); it != map.end();) { + if (pred(*it)) { + it = map.erase(it); + } else { + ++it; + } + } +} + +} // namespace phmap + +#ifdef _MSC_VER +#pragma warning(pop) +#endif + +#endif // PHMAP_BTREE_BTREE_CONTAINER_H_ diff --git a/native-sql-engine/cpp/src/third_party/parallel_hashmap/conanfile.py b/native-sql-engine/cpp/src/third_party/parallel_hashmap/conanfile.py new file mode 100644 index 000000000..c046377d1 --- /dev/null +++ b/native-sql-engine/cpp/src/third_party/parallel_hashmap/conanfile.py @@ -0,0 +1,36 @@ +#!/usr/bin/env python +# -*- coding: utf-8 -*- + +from conans import ConanFile, tools +import os + +class SparseppConan(ConanFile): + name = "parallel_hashmap" + version = "1.27" + description = "A header-only, very fast and memory-friendly hash map" + + # Indicates License type of the packaged library + license = "https://github.com/greg7mdp/parallel-hashmap/blob/master/LICENSE" + + # Packages the license for the conanfile.py + exports = ["LICENSE"] + + # Custom attributes for Bincrafters recipe conventions + source_subfolder = "source_subfolder" + + def source(self): + source_url = "https://github.com/greg7mdp/parallel-hashmap" + tools.get("{0}/archive/{1}.tar.gz".format(source_url, self.version)) + extracted_dir = self.name + "-" + self.version + + #Rename to "source_folder" is a convention to simplify later steps + os.rename(extracted_dir, self.source_subfolder) + + + def package(self): + include_folder = os.path.join(self.source_subfolder, "parallel_hashmap") + self.copy(pattern="LICENSE") + self.copy(pattern="*", dst="include/parallel_hashmap", src=include_folder) + + def package_id(self): + self.info.header_only() diff --git a/native-sql-engine/cpp/src/third_party/parallel_hashmap/meminfo.h b/native-sql-engine/cpp/src/third_party/parallel_hashmap/meminfo.h new file mode 100644 index 000000000..47feb26cd --- /dev/null +++ b/native-sql-engine/cpp/src/third_party/parallel_hashmap/meminfo.h @@ -0,0 +1,186 @@ +#if !defined(spp_memory_h_guard) +#define spp_memory_h_guard + +#include +#include +#include + +#if defined(_WIN32) || defined(__CYGWIN__) +#define SPP_WIN +#endif + +#ifdef SPP_WIN +#include +#include +#undef min +#undef max +#elif defined(__linux__) +#include +#include +#elif defined(__FreeBSD__) +#include +#include +#include +#include +#include +#include +#endif + +namespace spp { +uint64_t GetSystemMemory(); +uint64_t GetTotalMemoryUsed(); +uint64_t GetProcessMemoryUsed(); +uint64_t GetPhysicalMemory(); + +uint64_t GetSystemMemory() { +#ifdef SPP_WIN + MEMORYSTATUSEX memInfo; + memInfo.dwLength = sizeof(MEMORYSTATUSEX); + GlobalMemoryStatusEx(&memInfo); + return static_cast(memInfo.ullTotalPageFile); +#elif defined(__linux__) + struct sysinfo memInfo; + sysinfo(&memInfo); + auto totalVirtualMem = memInfo.totalram; + + totalVirtualMem += memInfo.totalswap; + totalVirtualMem *= memInfo.mem_unit; + return static_cast(totalVirtualMem); +#elif defined(__FreeBSD__) + kvm_t* kd; + u_int pageCnt; + size_t pageCntLen = sizeof(pageCnt); + u_int pageSize; + struct kvm_swap kswap; + uint64_t totalVirtualMem; + + pageSize = static_cast(getpagesize()); + + sysctlbyname("vm.stats.vm.v_page_count", &pageCnt, &pageCntLen, NULL, 0); + totalVirtualMem = pageCnt * pageSize; + + kd = kvm_open(NULL, _PATH_DEVNULL, NULL, O_RDONLY, "kvm_open"); + kvm_getswapinfo(kd, &kswap, 1, 0); + kvm_close(kd); + totalVirtualMem += kswap.ksw_total * pageSize; + + return totalVirtualMem; +#else + return 0; +#endif +} + +uint64_t GetTotalMemoryUsed() { +#ifdef SPP_WIN + MEMORYSTATUSEX memInfo; + memInfo.dwLength = sizeof(MEMORYSTATUSEX); + GlobalMemoryStatusEx(&memInfo); + return static_cast(memInfo.ullTotalPageFile - memInfo.ullAvailPageFile); +#elif defined(__linux__) + struct sysinfo memInfo; + sysinfo(&memInfo); + auto virtualMemUsed = memInfo.totalram - memInfo.freeram; + + virtualMemUsed += memInfo.totalswap - memInfo.freeswap; + virtualMemUsed *= memInfo.mem_unit; + + return static_cast(virtualMemUsed); +#elif defined(__FreeBSD__) + kvm_t* kd; + u_int pageSize; + u_int pageCnt, freeCnt; + size_t pageCntLen = sizeof(pageCnt); + size_t freeCntLen = sizeof(freeCnt); + struct kvm_swap kswap; + uint64_t virtualMemUsed; + + pageSize = static_cast(getpagesize()); + + sysctlbyname("vm.stats.vm.v_page_count", &pageCnt, &pageCntLen, NULL, 0); + sysctlbyname("vm.stats.vm.v_free_count", &freeCnt, &freeCntLen, NULL, 0); + virtualMemUsed = (pageCnt - freeCnt) * pageSize; + + kd = kvm_open(NULL, _PATH_DEVNULL, NULL, O_RDONLY, "kvm_open"); + kvm_getswapinfo(kd, &kswap, 1, 0); + kvm_close(kd); + virtualMemUsed += kswap.ksw_used * pageSize; + + return virtualMemUsed; +#else + return 0; +#endif +} + +uint64_t GetProcessMemoryUsed() { +#ifdef SPP_WIN + PROCESS_MEMORY_COUNTERS_EX pmc; + GetProcessMemoryInfo(GetCurrentProcess(), + reinterpret_cast(&pmc), sizeof(pmc)); + return static_cast(pmc.PrivateUsage); +#elif defined(__linux__) + auto parseLine = [](char* line) -> int { + auto i = strlen(line); + + while (*line < '0' || *line > '9') { + line++; + } + + line[i - 3] = '\0'; + i = atoi(line); + return i; + }; + + auto file = fopen("/proc/self/status", "r"); + auto result = -1; + char line[128]; + + while (fgets(line, 128, file) != nullptr) { + if (strncmp(line, "VmSize:", 7) == 0) { + result = parseLine(line); + break; + } + } + + fclose(file); + return static_cast(result) * 1024; +#elif defined(__FreeBSD__) + struct kinfo_proc info; + size_t infoLen = sizeof(info); + int mib[] = {CTL_KERN, KERN_PROC, KERN_PROC_PID, getpid()}; + + sysctl(mib, sizeof(mib) / sizeof(*mib), &info, &infoLen, NULL, 0); + return static_cast(info.ki_rssize * getpagesize()); +#else + return 0; +#endif +} + +uint64_t GetPhysicalMemory() { +#ifdef SPP_WIN + MEMORYSTATUSEX memInfo; + memInfo.dwLength = sizeof(MEMORYSTATUSEX); + GlobalMemoryStatusEx(&memInfo); + return static_cast(memInfo.ullTotalPhys); +#elif defined(__linux__) + struct sysinfo memInfo; + sysinfo(&memInfo); + + auto totalPhysMem = memInfo.totalram; + + totalPhysMem *= memInfo.mem_unit; + return static_cast(totalPhysMem); +#elif defined(__FreeBSD__) + u_long physMem; + size_t physMemLen = sizeof(physMem); + int mib[] = {CTL_HW, HW_PHYSMEM}; + + sysctl(mib, sizeof(mib) / sizeof(*mib), &physMem, &physMemLen, NULL, 0); + return physMem; +#else + return 0; +#endif +} + +} // namespace spp + +#endif // spp_memory_h_guard diff --git a/native-sql-engine/cpp/src/third_party/parallel_hashmap/phmap.h b/native-sql-engine/cpp/src/third_party/parallel_hashmap/phmap.h new file mode 100644 index 000000000..4628cca30 --- /dev/null +++ b/native-sql-engine/cpp/src/third_party/parallel_hashmap/phmap.h @@ -0,0 +1,4654 @@ +#if !defined(phmap_h_guard_) +#define phmap_h_guard_ + +// --------------------------------------------------------------------------- +// Copyright (c) 2019, Gregory Popovitch - greg7mdp@gmail.com +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// https://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +// +// Includes work from abseil-cpp (https://github.com/abseil/abseil-cpp) +// with modifications. +// +// Copyright 2018 The Abseil Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// https://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +// --------------------------------------------------------------------------- + +#ifdef _MSC_VER +#pragma warning(push) + +#pragma warning(disable : 4127) // conditional expression is constant +#pragma warning(disable : 4324) // structure was padded due to alignment specifier +#pragma warning(disable : 4514) // unreferenced inline function has been removed +#pragma warning(disable : 4623) // default constructor was implicitly defined as deleted +#pragma warning(disable : 4625) // copy constructor was implicitly defined as deleted +#pragma warning(disable : 4626) // assignment operator was implicitly defined as deleted +#pragma warning(disable : 4710) // function not inlined +#pragma warning(disable : 4711) // selected for automatic inline expansion +#pragma warning(disable : 4820) // '6' bytes padding added after data member +#pragma warning(disable : 4868) // compiler may not enforce left-to-right evaluation + // order in braced initializer list +#pragma warning( \ + disable : 5027) // move assignment operator was implicitly defined as deleted +#pragma warning(disable : 5045) // Compiler will insert Spectre mitigation for memory + // load if /Qspectre switch specified +#endif + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include "phmap_base.h" +#include "phmap_fwd_decl.h" +#include "phmap_utils.h" + +#if PHMAP_HAVE_STD_STRING_VIEW +#include +#endif + +namespace phmap { + +namespace priv { + +// -------------------------------------------------------------------------- +template +class probe_seq { + public: + probe_seq(size_t hashval, size_t mask) { + assert(((mask + 1) & mask) == 0 && "not a mask"); + mask_ = mask; + offset_ = hashval & mask_; + } + size_t offset() const { return offset_; } + size_t offset(size_t i) const { return (offset_ + i) & mask_; } + + void next() { + index_ += Width; + offset_ += index_; + offset_ &= mask_; + } + // 0-based probe index. The i-th probe in the probe sequence. + size_t getindex() const { return index_; } + + private: + size_t mask_; + size_t offset_; + size_t index_ = 0; +}; + +// -------------------------------------------------------------------------- +template +struct RequireUsableKey { + template + std::pair()(std::declval())), + decltype(std::declval()(std::declval(), + std::declval()))>* + operator()(const PassedKey&, const Args&...) const; +}; + +// -------------------------------------------------------------------------- +template +struct IsDecomposable : std::false_type {}; + +template +struct IsDecomposable< + phmap::void_t(), std::declval()...))>, + Policy, Hash, Eq, Ts...> : std::true_type {}; + +// TODO(alkis): Switch to std::is_nothrow_swappable when gcc/clang supports it. +// -------------------------------------------------------------------------- +template +constexpr bool IsNoThrowSwappable() { + using std::swap; + return noexcept(swap(std::declval(), std::declval())); +} + +// -------------------------------------------------------------------------- +template +int TrailingZeros(T x) { + PHMAP_IF_CONSTEXPR(sizeof(T) == 8) + return base_internal::CountTrailingZerosNonZero64(static_cast(x)); + else return base_internal::CountTrailingZerosNonZero32(static_cast(x)); +} + +// -------------------------------------------------------------------------- +template +int LeadingZeros(T x) { + PHMAP_IF_CONSTEXPR(sizeof(T) == 8) + return base_internal::CountLeadingZeros64(static_cast(x)); + else return base_internal::CountLeadingZeros32(static_cast(x)); +} + +// -------------------------------------------------------------------------- +// An abstraction over a bitmask. It provides an easy way to iterate through the +// indexes of the set bits of a bitmask. When Shift=0 (platforms with SSE), +// this is a true bitmask. On non-SSE, platforms the arithematic used to +// emulate the SSE behavior works in bytes (Shift=3) and leaves each bytes as +// either 0x00 or 0x80. +// +// For example: +// for (int i : BitMask(0x5)) -> yields 0, 2 +// for (int i : BitMask(0x0000000080800000)) -> yields 2, 3 +// -------------------------------------------------------------------------- +template +class BitMask { + static_assert(std::is_unsigned::value, ""); + static_assert(Shift == 0 || Shift == 3, ""); + + public: + // These are useful for unit tests (gunit). + using value_type = int; + using iterator = BitMask; + using const_iterator = BitMask; + + explicit BitMask(T mask) : mask_(mask) {} + BitMask& operator++() { + mask_ &= (mask_ - 1); + return *this; + } + explicit operator bool() const { return mask_ != 0; } + int operator*() const { return LowestBitSet(); } + int LowestBitSet() const { return priv::TrailingZeros(mask_) >> Shift; } + int HighestBitSet() const { + return (sizeof(T) * CHAR_BIT - priv::LeadingZeros(mask_) - 1) >> Shift; + } + + BitMask begin() const { return *this; } + BitMask end() const { return BitMask(0); } + + int TrailingZeros() const { return priv::TrailingZeros(mask_) >> Shift; } + + int LeadingZeros() const { + constexpr int total_significant_bits = SignificantBits << Shift; + constexpr int extra_bits = sizeof(T) * 8 - total_significant_bits; + return priv::LeadingZeros(mask_ << extra_bits) >> Shift; + } + + private: + friend bool operator==(const BitMask& a, const BitMask& b) { + return a.mask_ == b.mask_; + } + friend bool operator!=(const BitMask& a, const BitMask& b) { + return a.mask_ != b.mask_; + } + + T mask_; +}; + +// -------------------------------------------------------------------------- +using ctrl_t = signed char; +using h2_t = uint8_t; + +// -------------------------------------------------------------------------- +// The values here are selected for maximum performance. See the static asserts +// below for details. +// -------------------------------------------------------------------------- +enum Ctrl : ctrl_t { + kEmpty = -128, // 0b10000000 + kDeleted = -2, // 0b11111110 + kSentinel = -1, // 0b11111111 +}; + +static_assert(kEmpty & kDeleted & kSentinel & 0x80, + "Special markers need to have the MSB to make checking for them efficient"); +static_assert(kEmpty < kSentinel && kDeleted < kSentinel, + "kEmpty and kDeleted must be smaller than kSentinel to make the " + "SIMD test of IsEmptyOrDeleted() efficient"); +static_assert(kSentinel == -1, + "kSentinel must be -1 to elide loading it from memory into SIMD " + "registers (pcmpeqd xmm, xmm)"); +static_assert(kEmpty == -128, + "kEmpty must be -128 to make the SIMD check for its " + "existence efficient (psignb xmm, xmm)"); +static_assert(~kEmpty & ~kDeleted & kSentinel & 0x7F, + "kEmpty and kDeleted must share an unset bit that is not shared " + "by kSentinel to make the scalar test for MatchEmptyOrDeleted() " + "efficient"); +static_assert(kDeleted == -2, + "kDeleted must be -2 to make the implementation of " + "ConvertSpecialToEmptyAndFullToDeleted efficient"); + +// -------------------------------------------------------------------------- +// A single block of empty control bytes for tables without any slots allocated. +// This enables removing a branch in the hot path of find(). +// -------------------------------------------------------------------------- +inline ctrl_t* EmptyGroup() { + alignas(16) static constexpr ctrl_t empty_group[] = { + kSentinel, kEmpty, kEmpty, kEmpty, kEmpty, kEmpty, kEmpty, kEmpty, + kEmpty, kEmpty, kEmpty, kEmpty, kEmpty, kEmpty, kEmpty, kEmpty}; + return const_cast(empty_group); +} + +// -------------------------------------------------------------------------- +inline size_t HashSeed(const ctrl_t* ctrl) { + // The low bits of the pointer have little or no entropy because of + // alignment. We shift the pointer to try to use higher entropy bits. A + // good number seems to be 12 bits, because that aligns with page size. + return reinterpret_cast(ctrl) >> 12; +} + +#ifdef PHMAP_NON_DETERMINISTIC + +inline size_t H1(size_t hashval, const ctrl_t* ctrl) { + // use ctrl_ pointer to add entropy to ensure + // non-deterministic iteration order. + return (hashval >> 7) ^ HashSeed(ctrl); +} + +#else + +inline size_t H1(size_t hashval, const ctrl_t*) { return (hashval >> 7); } + +#endif + +inline ctrl_t H2(size_t hashval) { return (ctrl_t)(hashval & 0x7F); } + +inline bool IsEmpty(ctrl_t c) { return c == kEmpty; } +inline bool IsFull(ctrl_t c) { return c >= 0; } +inline bool IsDeleted(ctrl_t c) { return c == kDeleted; } +inline bool IsEmptyOrDeleted(ctrl_t c) { return c < kSentinel; } + +#if PHMAP_HAVE_SSE2 + +#ifdef _MSC_VER +#pragma warning(push) +#pragma warning(disable : 4365) // conversion from 'int' to 'T', signed/unsigned mismatch +#endif + +// -------------------------------------------------------------------------- +// https://github.com/abseil/abseil-cpp/issues/209 +// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87853 +// _mm_cmpgt_epi8 is broken under GCC with -funsigned-char +// Work around this by using the portable implementation of Group +// when using -funsigned-char under GCC. +// -------------------------------------------------------------------------- +inline __m128i _mm_cmpgt_epi8_fixed(__m128i a, __m128i b) { +#if defined(__GNUC__) && !defined(__clang__) +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Woverflow" + + if (std::is_unsigned::value) { + const __m128i mask = _mm_set1_epi8(static_cast(0x80)); + const __m128i diff = _mm_subs_epi8(b, a); + return _mm_cmpeq_epi8(_mm_and_si128(diff, mask), mask); + } + +#pragma GCC diagnostic pop +#endif + return _mm_cmpgt_epi8(a, b); +} + +// -------------------------------------------------------------------------- +// -------------------------------------------------------------------------- +struct GroupSse2Impl { + enum { kWidth = 16 }; // the number of slots per group + + explicit GroupSse2Impl(const ctrl_t* pos) { + ctrl = _mm_loadu_si128(reinterpret_cast(pos)); + } + + // Returns a bitmask representing the positions of slots that match hash. + // ---------------------------------------------------------------------- + BitMask Match(h2_t hash) const { + auto match = _mm_set1_epi8((char)hash); + return BitMask(_mm_movemask_epi8(_mm_cmpeq_epi8(match, ctrl))); + } + + // Returns a bitmask representing the positions of empty slots. + // ------------------------------------------------------------ + BitMask MatchEmpty() const { +#if PHMAP_HAVE_SSSE3 + // This only works because kEmpty is -128. + return BitMask(_mm_movemask_epi8(_mm_sign_epi8(ctrl, ctrl))); +#else + return Match(static_cast(kEmpty)); +#endif + } + + // Returns a bitmask representing the positions of empty or deleted slots. + // ----------------------------------------------------------------------- + BitMask MatchEmptyOrDeleted() const { + auto special = _mm_set1_epi8(kSentinel); + return BitMask( + _mm_movemask_epi8(_mm_cmpgt_epi8_fixed(special, ctrl))); + } + + // Returns the number of trailing empty or deleted elements in the group. + // ---------------------------------------------------------------------- + uint32_t CountLeadingEmptyOrDeleted() const { + auto special = _mm_set1_epi8(kSentinel); + return TrailingZeros(_mm_movemask_epi8(_mm_cmpgt_epi8_fixed(special, ctrl)) + 1); + } + + // ---------------------------------------------------------------------- + void ConvertSpecialToEmptyAndFullToDeleted(ctrl_t* dst) const { + auto msbs = _mm_set1_epi8(static_cast(-128)); + auto x126 = _mm_set1_epi8(126); +#if PHMAP_HAVE_SSSE3 + auto res = _mm_or_si128(_mm_shuffle_epi8(x126, ctrl), msbs); +#else + auto zero = _mm_setzero_si128(); + auto special_mask = _mm_cmpgt_epi8_fixed(zero, ctrl); + auto res = _mm_or_si128(msbs, _mm_andnot_si128(special_mask, x126)); +#endif + _mm_storeu_si128(reinterpret_cast<__m128i*>(dst), res); + } + + __m128i ctrl; +}; + +#ifdef _MSC_VER +#pragma warning(pop) +#endif + +#endif // PHMAP_HAVE_SSE2 + +// -------------------------------------------------------------------------- +// -------------------------------------------------------------------------- +struct GroupPortableImpl { + enum { kWidth = 8 }; + + explicit GroupPortableImpl(const ctrl_t* pos) : ctrl(little_endian::Load64(pos)) {} + + BitMask Match(h2_t hash) const { + // For the technique, see: + // http://graphics.stanford.edu/~seander/bithacks.html##ValueInWord + // (Determine if a word has a byte equal to n). + // + // Caveat: there are false positives but: + // - they only occur if there is a real match + // - they never occur on kEmpty, kDeleted, kSentinel + // - they will be handled gracefully by subsequent checks in code + // + // Example: + // v = 0x1716151413121110 + // hash = 0x12 + // retval = (v - lsbs) & ~v & msbs = 0x0000000080800000 + constexpr uint64_t msbs = 0x8080808080808080ULL; + constexpr uint64_t lsbs = 0x0101010101010101ULL; + auto x = ctrl ^ (lsbs * hash); + return BitMask((x - lsbs) & ~x & msbs); + } + + BitMask MatchEmpty() const { + constexpr uint64_t msbs = 0x8080808080808080ULL; + return BitMask((ctrl & (~ctrl << 6)) & msbs); + } + + BitMask MatchEmptyOrDeleted() const { + constexpr uint64_t msbs = 0x8080808080808080ULL; + return BitMask((ctrl & (~ctrl << 7)) & msbs); + } + + uint32_t CountLeadingEmptyOrDeleted() const { + constexpr uint64_t gaps = 0x00FEFEFEFEFEFEFEULL; + return (uint32_t)((TrailingZeros(((~ctrl & (ctrl >> 7)) | gaps) + 1) + 7) >> 3); + } + + void ConvertSpecialToEmptyAndFullToDeleted(ctrl_t* dst) const { + constexpr uint64_t msbs = 0x8080808080808080ULL; + constexpr uint64_t lsbs = 0x0101010101010101ULL; + auto x = ctrl & msbs; + auto res = (~x + (x >> 7)) & ~lsbs; + little_endian::Store64(dst, res); + } + + uint64_t ctrl; +}; + +#if PHMAP_HAVE_SSE2 +using Group = GroupSse2Impl; +#else +using Group = GroupPortableImpl; +#endif + +template +class raw_hash_set; + +inline bool IsValidCapacity(size_t n) { return ((n + 1) & n) == 0 && n > 0; } + +// -------------------------------------------------------------------------- +// PRECONDITION: +// IsValidCapacity(capacity) +// ctrl[capacity] == kSentinel +// ctrl[i] != kSentinel for all i < capacity +// Applies mapping for every byte in ctrl: +// DELETED -> EMPTY +// EMPTY -> EMPTY +// FULL -> DELETED +// -------------------------------------------------------------------------- +inline void ConvertDeletedToEmptyAndFullToDeleted(ctrl_t* ctrl, size_t capacity) { + assert(ctrl[capacity] == kSentinel); + assert(IsValidCapacity(capacity)); + for (ctrl_t* pos = ctrl; pos != ctrl + capacity + 1; pos += Group::kWidth) { + Group{pos}.ConvertSpecialToEmptyAndFullToDeleted(pos); + } + // Copy the cloned ctrl bytes. + std::memcpy(ctrl + capacity + 1, ctrl, Group::kWidth); + ctrl[capacity] = kSentinel; +} + +// -------------------------------------------------------------------------- +// Rounds up the capacity to the next power of 2 minus 1, with a minimum of 1. +// -------------------------------------------------------------------------- +inline size_t NormalizeCapacity(size_t n) { return n ? ~size_t{} >> LeadingZeros(n) : 1; } + +// -------------------------------------------------------------------------- +// We use 7/8th as maximum load factor. +// For 16-wide groups, that gives an average of two empty slots per group. +// -------------------------------------------------------------------------- +inline size_t CapacityToGrowth(size_t capacity) { + assert(IsValidCapacity(capacity)); + // `capacity*7/8` + PHMAP_IF_CONSTEXPR(Group::kWidth == 8) { + if (capacity == 7) { + // x-x/8 does not work when x==7. + return 6; + } + } + return capacity - capacity / 8; +} + +// -------------------------------------------------------------------------- +// From desired "growth" to a lowerbound of the necessary capacity. +// Might not be a valid one and required NormalizeCapacity(). +// -------------------------------------------------------------------------- +inline size_t GrowthToLowerboundCapacity(size_t growth) { + // `growth*8/7` + PHMAP_IF_CONSTEXPR(Group::kWidth == 8) { + if (growth == 7) { + // x+(x-1)/7 does not work when x==7. + return 8; + } + } + return growth + static_cast((static_cast(growth) - 1) / 7); +} + +namespace hashtable_debug_internal { + +// If it is a map, call get<0>(). +using std::get; +template +auto GetKey(const typename T::value_type& pair, int) -> decltype(get<0>(pair)) { + return get<0>(pair); +} + +// If it is not a map, return the value directly. +template +const typename T::key_type& GetKey(const typename T::key_type& key, char) { + return key; +} + +// -------------------------------------------------------------------------- +// Containers should specialize this to provide debug information for that +// container. +// -------------------------------------------------------------------------- +template +struct HashtableDebugAccess { + // Returns the number of probes required to find `key` in `c`. The "number of + // probes" is a concept that can vary by container. Implementations should + // return 0 when `key` was found in the minimum number of operations and + // should increment the result for each non-trivial operation required to find + // `key`. + // + // The default implementation uses the bucket api from the standard and thus + // works for `std::unordered_*` containers. + // -------------------------------------------------------------------------- + static size_t GetNumProbes(const Container& c, + const typename Container::key_type& key) { + if (!c.bucket_count()) return {}; + size_t num_probes = 0; + size_t bucket = c.bucket(key); + for (auto it = c.begin(bucket), e = c.end(bucket);; ++it, ++num_probes) { + if (it == e) return num_probes; + if (c.key_eq()(key, GetKey(*it, 0))) return num_probes; + } + } +}; + +} // namespace hashtable_debug_internal + +// ---------------------------------------------------------------------------- +// I N F O Z S T U B S +// ---------------------------------------------------------------------------- +struct HashtablezInfo { + void PrepareForSampling() {} +}; + +inline void RecordRehashSlow(HashtablezInfo*, size_t) {} + +static inline void RecordInsertSlow(HashtablezInfo*, size_t, size_t) {} + +static inline void RecordEraseSlow(HashtablezInfo*) {} + +static inline HashtablezInfo* SampleSlow(int64_t*) { return nullptr; } +static inline void UnsampleSlow(HashtablezInfo*) {} + +class HashtablezInfoHandle { + public: + inline void RecordStorageChanged(size_t, size_t) {} + inline void RecordRehash(size_t) {} + inline void RecordInsert(size_t, size_t) {} + inline void RecordErase() {} + friend inline void swap(HashtablezInfoHandle&, HashtablezInfoHandle&) noexcept {} +}; + +static inline HashtablezInfoHandle Sample() { return HashtablezInfoHandle(); } + +class HashtablezSampler { + public: + // Returns a global Sampler. + static HashtablezSampler& Global() { + static HashtablezSampler hzs; + return hzs; + } + HashtablezInfo* Register() { + static HashtablezInfo info; + return &info; + } + void Unregister(HashtablezInfo*) {} + + using DisposeCallback = void (*)(const HashtablezInfo&); + DisposeCallback SetDisposeCallback(DisposeCallback) { return nullptr; } + int64_t Iterate(const std::function&) { return 0; } +}; + +static inline void SetHashtablezEnabled(bool) {} +static inline void SetHashtablezSampleParameter(int32_t) {} +static inline void SetHashtablezMaxSamples(int32_t) {} + +namespace memory_internal { + +// Constructs T into uninitialized storage pointed by `ptr` using the args +// specified in the tuple. +// ---------------------------------------------------------------------------- +template +void ConstructFromTupleImpl(Alloc* alloc, T* ptr, Tuple&& t, + phmap::index_sequence) { + phmap::allocator_traits::construct(*alloc, ptr, + std::get(std::forward(t))...); +} + +template +struct WithConstructedImplF { + template + decltype(std::declval()(std::declval())) operator()(Args&&... args) const { + return std::forward(f)(T(std::forward(args)...)); + } + F&& f; +}; + +template +decltype(std::declval()(std::declval())) WithConstructedImpl( + Tuple&& t, phmap::index_sequence, F&& f) { + return WithConstructedImplF{std::forward(f)}( + std::get(std::forward(t))...); +} + +template +auto TupleRefImpl(T&& t, phmap::index_sequence) + -> decltype(std::forward_as_tuple(std::get(std::forward(t))...)) { + return std::forward_as_tuple(std::get(std::forward(t))...); +} + +// Returns a tuple of references to the elements of the input tuple. T must be a +// tuple. +// ---------------------------------------------------------------------------- +template +auto TupleRef(T&& t) -> decltype(TupleRefImpl( + std::forward(t), + phmap::make_index_sequence::type>::value>())) { + return TupleRefImpl( + std::forward(t), + phmap::make_index_sequence::type>::value>()); +} + +template +decltype(std::declval()(std::declval(), std::piecewise_construct, + std::declval>(), std::declval())) +DecomposePairImpl(F&& f, std::pair, V> p) { + const auto& key = std::get<0>(p.first); + return std::forward(f)(key, std::piecewise_construct, std::move(p.first), + std::move(p.second)); +} + +} // namespace memory_internal + +// ---------------------------------------------------------------------------- +// R A W _ H A S H _ S E T +// ---------------------------------------------------------------------------- +// An open-addressing +// hashtable with quadratic probing. +// +// This is a low level hashtable on top of which different interfaces can be +// implemented, like flat_hash_set, node_hash_set, string_hash_set, etc. +// +// The table interface is similar to that of std::unordered_set. Notable +// differences are that most member functions support heterogeneous keys when +// BOTH the hash and eq functions are marked as transparent. They do so by +// providing a typedef called `is_transparent`. +// +// When heterogeneous lookup is enabled, functions that take key_type act as if +// they have an overload set like: +// +// iterator find(const key_type& key); +// template +// iterator find(const K& key); +// +// size_type erase(const key_type& key); +// template +// size_type erase(const K& key); +// +// std::pair equal_range(const key_type& key); +// template +// std::pair equal_range(const K& key); +// +// When heterogeneous lookup is disabled, only the explicit `key_type` overloads +// exist. +// +// find() also supports passing the hash explicitly: +// +// iterator find(const key_type& key, size_t hash); +// template +// iterator find(const U& key, size_t hash); +// +// In addition the pointer to element and iterator stability guarantees are +// weaker: all iterators and pointers are invalidated after a new element is +// inserted. +// +// IMPLEMENTATION DETAILS +// +// The table stores elements inline in a slot array. In addition to the slot +// array the table maintains some control state per slot. The extra state is one +// byte per slot and stores empty or deleted marks, or alternatively 7 bits from +// the hash of an occupied slot. The table is split into logical groups of +// slots, like so: +// +// Group 1 Group 2 Group 3 +// +---------------+---------------+---------------+ +// | | | | | | | | | | | | | | | | | | | | | | | | | +// +---------------+---------------+---------------+ +// +// On lookup the hash is split into two parts: +// - H2: 7 bits (those stored in the control bytes) +// - H1: the rest of the bits +// The groups are probed using H1. For each group the slots are matched to H2 in +// parallel. Because H2 is 7 bits (128 states) and the number of slots per group +// is low (8 or 16) in almost all cases a match in H2 is also a lookup hit. +// +// On insert, once the right group is found (as in lookup), its slots are +// filled in order. +// +// On erase a slot is cleared. In case the group did not have any empty slots +// before the erase, the erased slot is marked as deleted. +// +// Groups without empty slots (but maybe with deleted slots) extend the probe +// sequence. The probing algorithm is quadratic. Given N the number of groups, +// the probing function for the i'th probe is: +// +// P(0) = H1 % N +// +// P(i) = (P(i - 1) + i) % N +// +// This probing function guarantees that after N probes, all the groups of the +// table will be probed exactly once. +// ---------------------------------------------------------------------------- +template +class raw_hash_set { + using PolicyTraits = hash_policy_traits; + using KeyArgImpl = KeyArg::value && IsTransparent::value>; + + public: + using init_type = typename PolicyTraits::init_type; + using key_type = typename PolicyTraits::key_type; + // TODO(sbenza): Hide slot_type as it is an implementation detail. Needs user + // code fixes! + using slot_type = typename PolicyTraits::slot_type; + using allocator_type = Alloc; + using size_type = size_t; + using difference_type = ptrdiff_t; + using hasher = Hash; + using key_equal = Eq; + using policy_type = Policy; + using value_type = typename PolicyTraits::value_type; + using reference = value_type&; + using const_reference = const value_type&; + using pointer = typename phmap::allocator_traits< + allocator_type>::template rebind_traits::pointer; + using const_pointer = typename phmap::allocator_traits< + allocator_type>::template rebind_traits::const_pointer; + + // Alias used for heterogeneous lookup functions. + // `key_arg` evaluates to `K` when the functors are transparent and to + // `key_type` otherwise. It permits template argument deduction on `K` for the + // transparent case. + template + using key_arg = typename KeyArgImpl::template type; + + private: + // Give an early error when key_type is not hashable/eq. + auto KeyTypeCanBeHashed(const Hash& h, const key_type& k) -> decltype(h(k)); + auto KeyTypeCanBeEq(const Eq& eq, const key_type& k) -> decltype(eq(k, k)); + + using Layout = phmap::priv::Layout; + + static Layout MakeLayout(size_t capacity) { + assert(IsValidCapacity(capacity)); + return Layout(capacity + Group::kWidth + 1, capacity); + } + + using AllocTraits = phmap::allocator_traits; + using SlotAlloc = + typename phmap::allocator_traits::template rebind_alloc; + using SlotAllocTraits = + typename phmap::allocator_traits::template rebind_traits; + + static_assert(std::is_lvalue_reference::value, + "Policy::element() must return a reference"); + + template + struct SameAsElementReference + : std::is_same< + typename std::remove_cv< + typename std::remove_reference::type>::type, + typename std::remove_cv::type>::type> {}; + + // An enabler for insert(T&&): T must be convertible to init_type or be the + // same as [cv] value_type [ref]. + // Note: we separate SameAsElementReference into its own type to avoid using + // reference unless we need to. MSVC doesn't seem to like it in some + // cases. + template + using RequiresInsertable = + typename std::enable_if, + SameAsElementReference>::value, + int>::type; + + // RequiresNotInit is a workaround for gcc prior to 7.1. + // See https://godbolt.org/g/Y4xsUh. + template + using RequiresNotInit = + typename std::enable_if::value, int>::type; + + template + using IsDecomposable = IsDecomposable; + + public: + static_assert(std::is_same::value, + "Allocators with custom pointer types are not supported"); + static_assert(std::is_same::value, + "Allocators with custom pointer types are not supported"); + + class iterator { + friend class raw_hash_set; + + public: + using iterator_category = std::forward_iterator_tag; + using value_type = typename raw_hash_set::value_type; + using reference = phmap::conditional_t; + using pointer = phmap::remove_reference_t*; + using difference_type = typename raw_hash_set::difference_type; + + iterator() {} + + // PRECONDITION: not an end() iterator. + reference operator*() const { return PolicyTraits::element(slot_); } + + // PRECONDITION: not an end() iterator. + pointer operator->() const { return &operator*(); } + + // PRECONDITION: not an end() iterator. + iterator& operator++() { + ++ctrl_; + ++slot_; + skip_empty_or_deleted(); + return *this; + } + // PRECONDITION: not an end() iterator. + iterator operator++(int) { + auto tmp = *this; + ++*this; + return tmp; + } + +#if PHMAP_BIDIRECTIONAL + // PRECONDITION: not a begin() iterator. + iterator& operator--() { + assert(ctrl_); + do { + --ctrl_; + --slot_; + } while (IsEmptyOrDeleted(*ctrl_)); + return *this; + } + + // PRECONDITION: not a begin() iterator. + iterator operator--(int) { + auto tmp = *this; + --*this; + return tmp; + } +#endif + + friend bool operator==(const iterator& a, const iterator& b) { + return a.ctrl_ == b.ctrl_; + } + friend bool operator!=(const iterator& a, const iterator& b) { return !(a == b); } + + private: + iterator(ctrl_t* ctrl) : ctrl_(ctrl) {} // for end() + iterator(ctrl_t* ctrl, slot_type* slot) : ctrl_(ctrl), slot_(slot) {} + + void skip_empty_or_deleted() { + while (IsEmptyOrDeleted(*ctrl_)) { + // ctrl is not necessarily aligned to Group::kWidth. It is also likely + // to read past the space for ctrl bytes and into slots. This is ok + // because ctrl has sizeof() == 1 and slot has sizeof() >= 1 so there + // is no way to read outside the combined slot array. + uint32_t shift = Group{ctrl_}.CountLeadingEmptyOrDeleted(); + ctrl_ += shift; + slot_ += shift; + } + } + + ctrl_t* ctrl_ = nullptr; + // To avoid uninitialized member warnings, put slot_ in an anonymous union. + // The member is not initialized on singleton and end iterators. + union { + slot_type* slot_; + }; + }; + + class const_iterator { + friend class raw_hash_set; + + public: + using iterator_category = typename iterator::iterator_category; + using value_type = typename raw_hash_set::value_type; + using reference = typename raw_hash_set::const_reference; + using pointer = typename raw_hash_set::const_pointer; + using difference_type = typename raw_hash_set::difference_type; + + const_iterator() {} + // Implicit construction from iterator. + const_iterator(iterator i) : inner_(std::move(i)) {} + + reference operator*() const { return *inner_; } + pointer operator->() const { return inner_.operator->(); } + + const_iterator& operator++() { + ++inner_; + return *this; + } + const_iterator operator++(int) { return inner_++; } + + friend bool operator==(const const_iterator& a, const const_iterator& b) { + return a.inner_ == b.inner_; + } + friend bool operator!=(const const_iterator& a, const const_iterator& b) { + return !(a == b); + } + + private: + const_iterator(const ctrl_t* ctrl, const slot_type* slot) + : inner_(const_cast(ctrl), const_cast(slot)) {} + + iterator inner_; + }; + + using node_type = node_handle, Alloc>; + using insert_return_type = InsertReturnType; + + raw_hash_set() noexcept( + std::is_nothrow_default_constructible::value&& + std::is_nothrow_default_constructible::value&& + std::is_nothrow_default_constructible::value) {} + + explicit raw_hash_set(size_t bucket_cnt, const hasher& hashfn = hasher(), + const key_equal& eq = key_equal(), + const allocator_type& alloc = allocator_type()) + : ctrl_(EmptyGroup()), settings_(0, hashfn, eq, alloc) { + if (bucket_cnt) { + size_t new_capacity = NormalizeCapacity(bucket_cnt); + reset_growth_left(new_capacity); + initialize_slots(new_capacity); + capacity_ = new_capacity; + } + } + + raw_hash_set(size_t bucket_cnt, const hasher& hashfn, const allocator_type& alloc) + : raw_hash_set(bucket_cnt, hashfn, key_equal(), alloc) {} + + raw_hash_set(size_t bucket_cnt, const allocator_type& alloc) + : raw_hash_set(bucket_cnt, hasher(), key_equal(), alloc) {} + + explicit raw_hash_set(const allocator_type& alloc) + : raw_hash_set(0, hasher(), key_equal(), alloc) {} + + template + raw_hash_set(InputIter first, InputIter last, size_t bucket_cnt = 0, + const hasher& hashfn = hasher(), const key_equal& eq = key_equal(), + const allocator_type& alloc = allocator_type()) + : raw_hash_set(bucket_cnt, hashfn, eq, alloc) { + insert(first, last); + } + + template + raw_hash_set(InputIter first, InputIter last, size_t bucket_cnt, const hasher& hashfn, + const allocator_type& alloc) + : raw_hash_set(first, last, bucket_cnt, hashfn, key_equal(), alloc) {} + + template + raw_hash_set(InputIter first, InputIter last, size_t bucket_cnt, + const allocator_type& alloc) + : raw_hash_set(first, last, bucket_cnt, hasher(), key_equal(), alloc) {} + + template + raw_hash_set(InputIter first, InputIter last, const allocator_type& alloc) + : raw_hash_set(first, last, 0, hasher(), key_equal(), alloc) {} + + // Instead of accepting std::initializer_list as the first + // argument like std::unordered_set does, we have two overloads + // that accept std::initializer_list and std::initializer_list. + // This is advantageous for performance. + // + // // Turns {"abc", "def"} into std::initializer_list, then + // // copies the strings into the set. + // std::unordered_set s = {"abc", "def"}; + // + // // Turns {"abc", "def"} into std::initializer_list, then + // // copies the strings into the set. + // phmap::flat_hash_set s = {"abc", "def"}; + // + // The same trick is used in insert(). + // + // The enabler is necessary to prevent this constructor from triggering where + // the copy constructor is meant to be called. + // + // phmap::flat_hash_set a, b{a}; + // + // RequiresNotInit is a workaround for gcc prior to 7.1. + template = 0, RequiresInsertable = 0> + raw_hash_set(std::initializer_list init, size_t bucket_cnt = 0, + const hasher& hashfn = hasher(), const key_equal& eq = key_equal(), + const allocator_type& alloc = allocator_type()) + : raw_hash_set(init.begin(), init.end(), bucket_cnt, hashfn, eq, alloc) {} + + raw_hash_set(std::initializer_list init, size_t bucket_cnt = 0, + const hasher& hashfn = hasher(), const key_equal& eq = key_equal(), + const allocator_type& alloc = allocator_type()) + : raw_hash_set(init.begin(), init.end(), bucket_cnt, hashfn, eq, alloc) {} + + template = 0, RequiresInsertable = 0> + raw_hash_set(std::initializer_list init, size_t bucket_cnt, const hasher& hashfn, + const allocator_type& alloc) + : raw_hash_set(init, bucket_cnt, hashfn, key_equal(), alloc) {} + + raw_hash_set(std::initializer_list init, size_t bucket_cnt, + const hasher& hashfn, const allocator_type& alloc) + : raw_hash_set(init, bucket_cnt, hashfn, key_equal(), alloc) {} + + template = 0, RequiresInsertable = 0> + raw_hash_set(std::initializer_list init, size_t bucket_cnt, + const allocator_type& alloc) + : raw_hash_set(init, bucket_cnt, hasher(), key_equal(), alloc) {} + + raw_hash_set(std::initializer_list init, size_t bucket_cnt, + const allocator_type& alloc) + : raw_hash_set(init, bucket_cnt, hasher(), key_equal(), alloc) {} + + template = 0, RequiresInsertable = 0> + raw_hash_set(std::initializer_list init, const allocator_type& alloc) + : raw_hash_set(init, 0, hasher(), key_equal(), alloc) {} + + raw_hash_set(std::initializer_list init, const allocator_type& alloc) + : raw_hash_set(init, 0, hasher(), key_equal(), alloc) {} + + raw_hash_set(const raw_hash_set& that) + : raw_hash_set( + that, AllocTraits::select_on_container_copy_construction(that.alloc_ref())) {} + + raw_hash_set(const raw_hash_set& that, const allocator_type& a) + : raw_hash_set(0, that.hash_ref(), that.eq_ref(), a) { + reserve(that.size()); + // Because the table is guaranteed to be empty, we can do something faster + // than a full `insert`. + for (const auto& v : that) { + const size_t hashval = PolicyTraits::apply(HashElement{hash_ref()}, v); + auto target = find_first_non_full(hashval); + set_ctrl(target.offset, H2(hashval)); + emplace_at(target.offset, v); + infoz_.RecordInsert(hashval, target.probe_length); + } + size_ = that.size(); + growth_left() -= that.size(); + } + + raw_hash_set(raw_hash_set&& that) noexcept( + std::is_nothrow_copy_constructible::value&& + std::is_nothrow_copy_constructible::value&& + std::is_nothrow_copy_constructible::value) + : ctrl_(phmap::exchange(that.ctrl_, EmptyGroup())), + slots_(phmap::exchange(that.slots_, nullptr)), + size_(phmap::exchange(that.size_, 0)), + capacity_(phmap::exchange(that.capacity_, 0)), + infoz_(phmap::exchange(that.infoz_, HashtablezInfoHandle())), + // Hash, equality and allocator are copied instead of moved because + // `that` must be left valid. If Hash is std::function, moving it + // would create a nullptr functor that cannot be called. + settings_(that.settings_) { + // growth_left was copied above, reset the one from `that`. + that.growth_left() = 0; + } + + raw_hash_set(raw_hash_set&& that, const allocator_type& a) + : ctrl_(EmptyGroup()), + slots_(nullptr), + size_(0), + capacity_(0), + settings_(0, that.hash_ref(), that.eq_ref(), a) { + if (a == that.alloc_ref()) { + std::swap(ctrl_, that.ctrl_); + std::swap(slots_, that.slots_); + std::swap(size_, that.size_); + std::swap(capacity_, that.capacity_); + std::swap(growth_left(), that.growth_left()); + std::swap(infoz_, that.infoz_); + } else { + reserve(that.size()); + // Note: this will copy elements of dense_set and unordered_set instead of + // moving them. This can be fixed if it ever becomes an issue. + for (auto& elem : that) insert(std::move(elem)); + } + } + + raw_hash_set& operator=(const raw_hash_set& that) { + raw_hash_set tmp(that, AllocTraits::propagate_on_container_copy_assignment::value + ? that.alloc_ref() + : alloc_ref()); + swap(tmp); + return *this; + } + + raw_hash_set& operator=(raw_hash_set&& that) noexcept( + phmap::allocator_traits::is_always_equal::value&& + std::is_nothrow_move_assignable::value&& + std::is_nothrow_move_assignable::value) { + // TODO(sbenza): We should only use the operations from the noexcept clause + // to make sure we actually adhere to that contract. + return move_assign(std::move(that), + typename AllocTraits::propagate_on_container_move_assignment()); + } + + ~raw_hash_set() { destroy_slots(); } + + iterator begin() { + auto it = iterator_at(0); + it.skip_empty_or_deleted(); + return it; + } + iterator end() { +#if PHMAP_BIDIRECTIONAL + return iterator_at(capacity_); +#else + return {ctrl_ + capacity_}; +#endif + } + + const_iterator begin() const { return const_cast(this)->begin(); } + const_iterator end() const { return const_cast(this)->end(); } + const_iterator cbegin() const { return begin(); } + const_iterator cend() const { return end(); } + + bool empty() const { return !size(); } + size_t size() const { return size_; } + size_t capacity() const { return capacity_; } + size_t max_size() const { return (std::numeric_limits::max)(); } + + PHMAP_ATTRIBUTE_REINITIALIZES void clear() { + // Iterating over this container is O(bucket_count()). When bucket_count() + // is much greater than size(), iteration becomes prohibitively expensive. + // For clear() it is more important to reuse the allocated array when the + // container is small because allocation takes comparatively long time + // compared to destruction of the elements of the container. So we pick the + // largest bucket_count() threshold for which iteration is still fast and + // past that we simply deallocate the array. + if (empty()) return; + if (capacity_ > 127) { + destroy_slots(); + } else if (capacity_) { + for (size_t i = 0; i != capacity_; ++i) { + if (IsFull(ctrl_[i])) { + PolicyTraits::destroy(&alloc_ref(), slots_ + i); + } + } + size_ = 0; + reset_ctrl(capacity_); + reset_growth_left(capacity_); + } + assert(empty()); + infoz_.RecordStorageChanged(0, capacity_); + } + + // This overload kicks in when the argument is an rvalue of insertable and + // decomposable type other than init_type. + // + // flat_hash_map m; + // m.insert(std::make_pair("abc", 42)); + template = 0, + typename std::enable_if::value, int>::type = 0, + T* = nullptr> + std::pair insert(T&& value) { + return emplace(std::forward(value)); + } + + // This overload kicks in when the argument is a bitfield or an lvalue of + // insertable and decomposable type. + // + // union { int n : 1; }; + // flat_hash_set s; + // s.insert(n); + // + // flat_hash_set s; + // const char* p = "hello"; + // s.insert(p); + // + // TODO(romanp): Once we stop supporting gcc 5.1 and below, replace + // RequiresInsertable with RequiresInsertable. + // We are hitting this bug: https://godbolt.org/g/1Vht4f. + template = 0, + typename std::enable_if::value, int>::type = 0> + std::pair insert(const T& value) { + return emplace(value); + } + + // This overload kicks in when the argument is an rvalue of init_type. Its + // purpose is to handle brace-init-list arguments. + // + // flat_hash_set s; + // s.insert({"abc", 42}); + std::pair insert(init_type&& value) { + return emplace(std::move(value)); + } + + template = 0, + typename std::enable_if::value, int>::type = 0, + T* = nullptr> + iterator insert(const_iterator, T&& value) { + return insert(std::forward(value)).first; + } + + // TODO(romanp): Once we stop supporting gcc 5.1 and below, replace + // RequiresInsertable with RequiresInsertable. + // We are hitting this bug: https://godbolt.org/g/1Vht4f. + template = 0, + typename std::enable_if::value, int>::type = 0> + iterator insert(const_iterator, const T& value) { + return insert(value).first; + } + + iterator insert(const_iterator, init_type&& value) { + return insert(std::move(value)).first; + } + + template + using IsRandomAccess = + std::is_same::iterator_category, + std::random_access_iterator_tag>; + + template + struct has_difference_operator { + private: + using yes = std::true_type; + using no = std::false_type; + + template + static auto test(int) -> decltype(std::declval() - std::declval() == 1, yes()); + template + static no test(...); + + public: + static constexpr bool value = std::is_same(0)), yes>::value; + }; + + template ::value, int> = 0> + void insert(InputIt first, InputIt last) { + this->reserve(this->size() + (last - first)); + for (; first != last; ++first) emplace(*first); + } + + template ::value, int> = 0> + void insert(InputIt first, InputIt last) { + for (; first != last; ++first) emplace(*first); + } + + template = 0, RequiresInsertable = 0> + void insert(std::initializer_list ilist) { + insert(ilist.begin(), ilist.end()); + } + + void insert(std::initializer_list ilist) { + insert(ilist.begin(), ilist.end()); + } + + insert_return_type insert(node_type&& node) { + if (!node) return {end(), false, node_type()}; + const auto& elem = PolicyTraits::element(CommonAccess::GetSlot(node)); + auto res = PolicyTraits::apply( + InsertSlot{*this, std::move(*CommonAccess::GetSlot(node))}, elem); + if (res.second) { + CommonAccess::Reset(&node); + return {res.first, true, node_type()}; + } else { + return {res.first, false, std::move(node)}; + } + } + + insert_return_type insert(node_type&& node, size_t hashval) { + if (!node) return {end(), false, node_type()}; + const auto& elem = PolicyTraits::element(CommonAccess::GetSlot(node)); + auto res = PolicyTraits::apply( + InsertSlotWithHash{*this, std::move(*CommonAccess::GetSlot(node)), + hashval}, + elem); + if (res.second) { + CommonAccess::Reset(&node); + return {res.first, true, node_type()}; + } else { + return {res.first, false, std::move(node)}; + } + } + + iterator insert(const_iterator, node_type&& node) { + return insert(std::move(node)).first; + } + + // This overload kicks in if we can deduce the key from args. This enables us + // to avoid constructing value_type if an entry with the same key already + // exists. + // + // For example: + // + // flat_hash_map m = {{"abc", "def"}}; + // // Creates no std::string copies and makes no heap allocations. + // m.emplace("abc", "xyz"); + template ::value, int>::type = 0> + std::pair emplace(Args&&... args) { + return PolicyTraits::apply(EmplaceDecomposable{*this}, std::forward(args)...); + } + + template ::value, int>::type = 0> + std::pair emplace_with_hash(size_t hashval, Args&&... args) { + return PolicyTraits::apply(EmplaceDecomposableHashval{*this, hashval}, + std::forward(args)...); + } + + // This overload kicks in if we cannot deduce the key from args. It constructs + // value_type unconditionally and then either moves it into the table or + // destroys. + template ::value, int>::type = 0> + std::pair emplace(Args&&... args) { + typename std::aligned_storage::type raw; + slot_type* slot = reinterpret_cast(&raw); + + PolicyTraits::construct(&alloc_ref(), slot, std::forward(args)...); + const auto& elem = PolicyTraits::element(slot); + return PolicyTraits::apply(InsertSlot{*this, std::move(*slot)}, elem); + } + + template ::value, int>::type = 0> + std::pair emplace_with_hash(size_t hashval, Args&&... args) { + typename std::aligned_storage::type raw; + slot_type* slot = reinterpret_cast(&raw); + + PolicyTraits::construct(&alloc_ref(), slot, std::forward(args)...); + const auto& elem = PolicyTraits::element(slot); + return PolicyTraits::apply(InsertSlotWithHash{*this, std::move(*slot), hashval}, + elem); + } + + template + iterator emplace_hint(const_iterator, Args&&... args) { + return emplace(std::forward(args)...).first; + } + + template + iterator emplace_hint_with_hash(size_t hashval, const_iterator, Args&&... args) { + return emplace_with_hash(hashval, std::forward(args)...).first; + } + + // Extension API: support for lazy emplace. + // + // Looks up key in the table. If found, returns the iterator to the element. + // Otherwise calls f with one argument of type raw_hash_set::constructor. f + // MUST call raw_hash_set::constructor with arguments as if a + // raw_hash_set::value_type is constructed, otherwise the behavior is + // undefined. + // + // For example: + // + // std::unordered_set s; + // // Makes ArenaStr even if "abc" is in the map. + // s.insert(ArenaString(&arena, "abc")); + // + // flat_hash_set s; + // // Makes ArenaStr only if "abc" is not in the map. + // s.lazy_emplace("abc", [&](const constructor& ctor) { + // ctor(&arena, "abc"); + // }); + // + // WARNING: This API is currently experimental. If there is a way to implement + // the same thing with the rest of the API, prefer that. + class constructor { + friend class raw_hash_set; + + public: + template + void operator()(Args&&... args) const { + assert(*slot_); + PolicyTraits::construct(alloc_, *slot_, std::forward(args)...); + *slot_ = nullptr; + } + + private: + constructor(allocator_type* a, slot_type** slot) : alloc_(a), slot_(slot) {} + + allocator_type* alloc_; + slot_type** slot_; + }; + + template + iterator lazy_emplace(const key_arg& key, F&& f) { + auto res = find_or_prepare_insert(key); + if (res.second) { + lazy_emplace_at(res.first, std::forward(f)); + } + return iterator_at(res.first); + } + + template + iterator lazy_emplace_with_hash(const key_arg& key, size_t& hashval, F&& f) { + auto res = find_or_prepare_insert(key, hashval); + if (res.second) { + lazy_emplace_at(res.first, std::forward(f)); + } + return iterator_at(res.first); + } + + template + void lazy_emplace_at(size_t& idx, F&& f) { + slot_type* slot = slots_ + idx; + std::forward(f)(constructor(&alloc_ref(), &slot)); + assert(!slot); + } + + // Extension API: support for heterogeneous keys. + // + // std::unordered_set s; + // // Turns "abc" into std::string. + // s.erase("abc"); + // + // flat_hash_set s; + // // Uses "abc" directly without copying it into std::string. + // s.erase("abc"); + template + size_type erase(const key_arg& key) { + auto it = find(key); + if (it == end()) return 0; + _erase(it); + return 1; + } + + iterator erase(const_iterator cit) { return erase(cit.inner_); } + + // Erases the element pointed to by `it`. Unlike `std::unordered_set::erase`, + // this method returns void to reduce algorithmic complexity to O(1). In + // order to erase while iterating across a map, use the following idiom (which + // also works for standard containers): + // + // for (auto it = m.begin(), end = m.end(); it != end;) { + // if () { + // m._erase(it++); + // } else { + // ++it; + // } + // } + void _erase(iterator it) { + assert(it != end()); + PolicyTraits::destroy(&alloc_ref(), it.slot_); + erase_meta_only(it); + } + void _erase(const_iterator cit) { _erase(cit.inner_); } + + // This overload is necessary because otherwise erase(const K&) would be + // a better match if non-const iterator is passed as an argument. + iterator erase(iterator it) { + auto res = it; + ++res; + _erase(it); + return res; + } + + iterator erase(const_iterator first, const_iterator last) { + while (first != last) { + _erase(first++); + } + return last.inner_; + } + + // Moves elements from `src` into `this`. + // If the element already exists in `this`, it is left unmodified in `src`. + template + void merge(raw_hash_set& src) { // NOLINT + assert(this != &src); + for (auto it = src.begin(), e = src.end(); it != e; ++it) { + if (PolicyTraits::apply(InsertSlot{*this, std::move(*it.slot_)}, + PolicyTraits::element(it.slot_)) + .second) { + src.erase_meta_only(it); + } + } + } + + template + void merge(raw_hash_set&& src) { + merge(src); + } + + node_type extract(const_iterator position) { + auto node = CommonAccess::Make(alloc_ref(), position.inner_.slot_); + erase_meta_only(position); + return node; + } + + template ::value, int>::type = 0> + node_type extract(const key_arg& key) { + auto it = find(key); + return it == end() ? node_type() : extract(const_iterator{it}); + } + + void swap(raw_hash_set& that) noexcept( + IsNoThrowSwappable() && IsNoThrowSwappable() && + (!AllocTraits::propagate_on_container_swap::value || + IsNoThrowSwappable())) { + using std::swap; + swap(ctrl_, that.ctrl_); + swap(slots_, that.slots_); + swap(size_, that.size_); + swap(capacity_, that.capacity_); + swap(growth_left(), that.growth_left()); + swap(hash_ref(), that.hash_ref()); + swap(eq_ref(), that.eq_ref()); + swap(infoz_, that.infoz_); + if (AllocTraits::propagate_on_container_swap::value) { + swap(alloc_ref(), that.alloc_ref()); + } else { + // If the allocators do not compare equal it is officially undefined + // behavior. We choose to do nothing. + } + } + +#if !defined(PHMAP_NON_DETERMINISTIC) + template + bool phmap_dump(OutputArchive&) const; + + template + bool phmap_load(InputArchive&); +#endif + + void rehash(size_t n) { + if (n == 0 && capacity_ == 0) return; + if (n == 0 && size_ == 0) { + destroy_slots(); + infoz_.RecordStorageChanged(0, 0); + return; + } + // bitor is a faster way of doing `max` here. We will round up to the next + // power-of-2-minus-1, so bitor is good enough. + auto m = NormalizeCapacity((std::max)(n, size())); + // n == 0 unconditionally rehashes as per the standard. + if (n == 0 || m > capacity_) { + resize(m); + } + } + + void reserve(size_t n) { rehash(GrowthToLowerboundCapacity(n)); } + + // Extension API: support for heterogeneous keys. + // + // std::unordered_set s; + // // Turns "abc" into std::string. + // s.count("abc"); + // + // ch_set s; + // // Uses "abc" directly without copying it into std::string. + // s.count("abc"); + template + size_t count(const key_arg& key) const { + return find(key) == end() ? size_t(0) : size_t(1); + } + + // Issues CPU prefetch instructions for the memory needed to find or insert + // a key. Like all lookup functions, this support heterogeneous keys. + // + // NOTE: This is a very low level operation and should not be used without + // specific benchmarks indicating its importance. + void prefetch_hash(size_t hashval) const { + (void)hashval; +#if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86)) + auto seq = probe(hashval); + _mm_prefetch((const char*)(ctrl_ + seq.offset()), _MM_HINT_NTA); + _mm_prefetch((const char*)(slots_ + seq.offset()), _MM_HINT_NTA); +#elif defined(__GNUC__) + auto seq = probe(hashval); + __builtin_prefetch(static_cast(ctrl_ + seq.offset())); + __builtin_prefetch(static_cast(slots_ + seq.offset())); +#endif // __GNUC__ + } + + template + void prefetch(const key_arg& key) const { + prefetch_hash(this->hash(key)); + } + + // The API of find() has two extensions. + // + // 1. The hash can be passed by the user. It must be equal to the hash of the + // key. + // + // 2. The type of the key argument doesn't have to be key_type. This is so + // called heterogeneous key support. + template + iterator find(const key_arg& key, size_t hashval) { + size_t offset; + if (find_impl(key, hashval, offset)) + return iterator_at(offset); + else + return end(); + } + + template + pointer find_ptr(const key_arg& key, size_t hashval) { + size_t offset; + if (find_impl(key, hashval, offset)) + return &PolicyTraits::element(slots_ + offset); + else + return nullptr; + } + + template + iterator find(const key_arg& key) { + return find(key, this->hash(key)); + } + + template + const_iterator find(const key_arg& key, size_t hashval) const { + return const_cast(this)->find(key, hashval); + } + template + const_iterator find(const key_arg& key) const { + return find(key, this->hash(key)); + } + + template + bool contains(const key_arg& key) const { + return find(key) != end(); + } + + template + bool contains(const key_arg& key, size_t hashval) const { + return find(key, hashval) != end(); + } + + template + std::pair equal_range(const key_arg& key) { + auto it = find(key); + if (it != end()) return {it, std::next(it)}; + return {it, it}; + } + template + std::pair equal_range(const key_arg& key) const { + auto it = find(key); + if (it != end()) return {it, std::next(it)}; + return {it, it}; + } + + size_t bucket_count() const { return capacity_; } + float load_factor() const { + return capacity_ ? static_cast(size()) / capacity_ : 0.0; + } + float max_load_factor() const { return 1.0f; } + void max_load_factor(float) { + // Does nothing. + } + + hasher hash_function() const { + return hash_ref(); + } // warning: doesn't match internal hash - use hash() member function + key_equal key_eq() const { return eq_ref(); } + allocator_type get_allocator() const { return alloc_ref(); } + + friend bool operator==(const raw_hash_set& a, const raw_hash_set& b) { + if (a.size() != b.size()) return false; + const raw_hash_set* outer = &a; + const raw_hash_set* inner = &b; + if (outer->capacity() > inner->capacity()) std::swap(outer, inner); + for (const value_type& elem : *outer) + if (!inner->has_element(elem)) return false; + return true; + } + + friend bool operator!=(const raw_hash_set& a, const raw_hash_set& b) { + return !(a == b); + } + + friend void swap(raw_hash_set& a, raw_hash_set& b) noexcept(noexcept(a.swap(b))) { + a.swap(b); + } + + template + size_t hash(const K& key) const { + return HashElement{hash_ref()}(key); + } + + private: + template + friend struct phmap::priv::hashtable_debug_internal::HashtableDebugAccess; + + template + bool find_impl(const key_arg& key, size_t hashval, size_t& offset) { + auto seq = probe(hashval); + while (true) { + Group g{ctrl_ + seq.offset()}; + for (int i : g.Match((h2_t)H2(hashval))) { + offset = seq.offset((size_t)i); + if (PHMAP_PREDICT_TRUE(PolicyTraits::apply( + EqualElement{key, eq_ref()}, PolicyTraits::element(slots_ + offset)))) + return true; + } + if (PHMAP_PREDICT_TRUE(g.MatchEmpty())) return false; + seq.next(); + } + } + + struct FindElement { + template + const_iterator operator()(const K& key, Args&&...) const { + return s.find(key); + } + const raw_hash_set& s; + }; + + struct HashElement { + template + size_t operator()(const K& key, Args&&...) const { + return phmap_mix()(h(key)); + } + const hasher& h; + }; + + template + struct EqualElement { + template + bool operator()(const K2& lhs, Args&&...) const { + return eq(lhs, rhs); + } + const K1& rhs; + const key_equal& eq; + }; + + template + std::pair emplace_decomposable(const K& key, size_t hashval, + Args&&... args) { + auto res = find_or_prepare_insert(key, hashval); + if (res.second) { + emplace_at(res.first, std::forward(args)...); + } + return {iterator_at(res.first), res.second}; + } + + struct EmplaceDecomposable { + template + std::pair operator()(const K& key, Args&&... args) const { + return s.emplace_decomposable(key, s.hash(key), std::forward(args)...); + } + raw_hash_set& s; + }; + + struct EmplaceDecomposableHashval { + template + std::pair operator()(const K& key, Args&&... args) const { + return s.emplace_decomposable(key, hashval, std::forward(args)...); + } + raw_hash_set& s; + size_t hashval; + }; + + template + struct InsertSlot { + template + std::pair operator()(const K& key, Args&&...) && { + auto res = s.find_or_prepare_insert(key); + if (res.second) { + PolicyTraits::transfer(&s.alloc_ref(), s.slots_ + res.first, &slot); + } else if (do_destroy) { + PolicyTraits::destroy(&s.alloc_ref(), &slot); + } + return {s.iterator_at(res.first), res.second}; + } + raw_hash_set& s; + // Constructed slot. Either moved into place or destroyed. + slot_type&& slot; + }; + + template + struct InsertSlotWithHash { + template + std::pair operator()(const K& key, Args&&...) && { + auto res = s.find_or_prepare_insert(key, hashval); + if (res.second) { + PolicyTraits::transfer(&s.alloc_ref(), s.slots_ + res.first, &slot); + } else if (do_destroy) { + PolicyTraits::destroy(&s.alloc_ref(), &slot); + } + return {s.iterator_at(res.first), res.second}; + } + raw_hash_set& s; + // Constructed slot. Either moved into place or destroyed. + slot_type&& slot; + size_t& hashval; + }; + + // "erases" the object from the container, except that it doesn't actually + // destroy the object. It only updates all the metadata of the class. + // This can be used in conjunction with Policy::transfer to move the object to + // another place. + void erase_meta_only(const_iterator it) { + assert(IsFull(*it.inner_.ctrl_) && "erasing a dangling iterator"); + --size_; + const size_t index = (size_t)(it.inner_.ctrl_ - ctrl_); + const size_t index_before = (index - Group::kWidth) & capacity_; + const auto empty_after = Group(it.inner_.ctrl_).MatchEmpty(); + const auto empty_before = Group(ctrl_ + index_before).MatchEmpty(); + + // We count how many consecutive non empties we have to the right and to the + // left of `it`. If the sum is >= kWidth then there is at least one probe + // window that might have seen a full group. + bool was_never_full = + empty_before && empty_after && + static_cast(empty_after.TrailingZeros() + empty_before.LeadingZeros()) < + Group::kWidth; + + set_ctrl(index, was_never_full ? kEmpty : kDeleted); + growth_left() += was_never_full; + infoz_.RecordErase(); + } + + void initialize_slots(size_t new_capacity) { + assert(new_capacity); + if (std::is_same>::value && slots_ == nullptr) { + infoz_ = Sample(); + } + + auto layout = MakeLayout(new_capacity); + char* mem = static_cast( + Allocate(&alloc_ref(), layout.AllocSize())); + ctrl_ = reinterpret_cast(layout.template Pointer<0>(mem)); + slots_ = layout.template Pointer<1>(mem); + reset_ctrl(new_capacity); + reset_growth_left(new_capacity); + infoz_.RecordStorageChanged(size_, new_capacity); + } + + void destroy_slots() { + if (!capacity_) return; + for (size_t i = 0; i != capacity_; ++i) { + if (IsFull(ctrl_[i])) { + PolicyTraits::destroy(&alloc_ref(), slots_ + i); + } + } + auto layout = MakeLayout(capacity_); + // Unpoison before returning the memory to the allocator. + SanitizerUnpoisonMemoryRegion(slots_, sizeof(slot_type) * capacity_); + Deallocate(&alloc_ref(), ctrl_, layout.AllocSize()); + ctrl_ = EmptyGroup(); + slots_ = nullptr; + size_ = 0; + capacity_ = 0; + growth_left() = 0; + } + + void resize(size_t new_capacity) { + assert(IsValidCapacity(new_capacity)); + auto* old_ctrl = ctrl_; + auto* old_slots = slots_; + const size_t old_capacity = capacity_; + initialize_slots(new_capacity); + capacity_ = new_capacity; + + for (size_t i = 0; i != old_capacity; ++i) { + if (IsFull(old_ctrl[i])) { + size_t hashval = PolicyTraits::apply(HashElement{hash_ref()}, + PolicyTraits::element(old_slots + i)); + auto target = find_first_non_full(hashval); + size_t new_i = target.offset; + set_ctrl(new_i, H2(hashval)); + PolicyTraits::transfer(&alloc_ref(), slots_ + new_i, old_slots + i); + } + } + if (old_capacity) { + SanitizerUnpoisonMemoryRegion(old_slots, sizeof(slot_type) * old_capacity); + auto layout = MakeLayout(old_capacity); + Deallocate(&alloc_ref(), old_ctrl, layout.AllocSize()); + } + } + + void drop_deletes_without_resize() PHMAP_ATTRIBUTE_NOINLINE { + assert(IsValidCapacity(capacity_)); + assert(!is_small()); + // Algorithm: + // - mark all DELETED slots as EMPTY + // - mark all FULL slots as DELETED + // - for each slot marked as DELETED + // hash = Hash(element) + // target = find_first_non_full(hash) + // if target is in the same group + // mark slot as FULL + // else if target is EMPTY + // transfer element to target + // mark slot as EMPTY + // mark target as FULL + // else if target is DELETED + // swap current element with target element + // mark target as FULL + // repeat procedure for current slot with moved from element (target) + ConvertDeletedToEmptyAndFullToDeleted(ctrl_, capacity_); + typename std::aligned_storage::type raw; + slot_type* slot = reinterpret_cast(&raw); + for (size_t i = 0; i != capacity_; ++i) { + if (!IsDeleted(ctrl_[i])) continue; + size_t hashval = + PolicyTraits::apply(HashElement{hash_ref()}, PolicyTraits::element(slots_ + i)); + auto target = find_first_non_full(hashval); + size_t new_i = target.offset; + + // Verify if the old and new i fall within the same group wrt the hashval. + // If they do, we don't need to move the object as it falls already in the + // best probe we can. + const auto probe_index = [&](size_t pos) { + return ((pos - probe(hashval).offset()) & capacity_) / Group::kWidth; + }; + + // Element doesn't move. + if (PHMAP_PREDICT_TRUE(probe_index(new_i) == probe_index(i))) { + set_ctrl(i, H2(hashval)); + continue; + } + if (IsEmpty(ctrl_[new_i])) { + // Transfer element to the empty spot. + // set_ctrl poisons/unpoisons the slots so we have to call it at the + // right time. + set_ctrl(new_i, H2(hashval)); + PolicyTraits::transfer(&alloc_ref(), slots_ + new_i, slots_ + i); + set_ctrl(i, kEmpty); + } else { + assert(IsDeleted(ctrl_[new_i])); + set_ctrl(new_i, H2(hashval)); + // Until we are done rehashing, DELETED marks previously FULL slots. + // Swap i and new_i elements. + PolicyTraits::transfer(&alloc_ref(), slot, slots_ + i); + PolicyTraits::transfer(&alloc_ref(), slots_ + i, slots_ + new_i); + PolicyTraits::transfer(&alloc_ref(), slots_ + new_i, slot); + --i; // repeat + } + } + reset_growth_left(capacity_); + } + + void rehash_and_grow_if_necessary() { + if (capacity_ == 0) { + resize(1); + } else if (size() <= CapacityToGrowth(capacity()) / 2) { + // Squash DELETED without growing if there is enough capacity. + drop_deletes_without_resize(); + } else { + // Otherwise grow the container. + resize(capacity_ * 2 + 1); + } + } + + bool has_element(const value_type& elem, size_t hashval) const { + auto seq = probe(hashval); + while (true) { + Group g{ctrl_ + seq.offset()}; + for (int i : g.Match((h2_t)H2(hashval))) { + if (PHMAP_PREDICT_TRUE(PolicyTraits::element(slots_ + seq.offset((size_t)i)) == + elem)) + return true; + } + if (PHMAP_PREDICT_TRUE(g.MatchEmpty())) return false; + seq.next(); + assert(seq.getindex() < capacity_ && "full table!"); + } + return false; + } + + bool has_element(const value_type& elem) const { + size_t hashval = PolicyTraits::apply(HashElement{hash_ref()}, elem); + return has_element(elem, hashval); + } + + // Probes the raw_hash_set with the probe sequence for hash and returns the + // pointer to the first empty or deleted slot. + // NOTE: this function must work with tables having both kEmpty and kDelete + // in one group. Such tables appears during drop_deletes_without_resize. + // + // This function is very useful when insertions happen and: + // - the input is already a set + // - there are enough slots + // - the element with the hash is not in the table + struct FindInfo { + size_t offset; + size_t probe_length; + }; + FindInfo find_first_non_full(size_t hashval) { + auto seq = probe(hashval); + while (true) { + Group g{ctrl_ + seq.offset()}; + auto mask = g.MatchEmptyOrDeleted(); + if (mask) { + return {seq.offset((size_t)mask.LowestBitSet()), seq.getindex()}; + } + assert(seq.getindex() < capacity_ && "full table!"); + seq.next(); + } + } + + // TODO(alkis): Optimize this assuming *this and that don't overlap. + raw_hash_set& move_assign(raw_hash_set&& that, std::true_type) { + raw_hash_set tmp(std::move(that)); + swap(tmp); + return *this; + } + raw_hash_set& move_assign(raw_hash_set&& that, std::false_type) { + raw_hash_set tmp(std::move(that), alloc_ref()); + swap(tmp); + return *this; + } + + protected: + template + std::pair find_or_prepare_insert(const K& key, size_t hashval) { + auto seq = probe(hashval); + while (true) { + Group g{ctrl_ + seq.offset()}; + for (int i : g.Match((h2_t)H2(hashval))) { + if (PHMAP_PREDICT_TRUE(PolicyTraits::apply( + EqualElement{key, eq_ref()}, + PolicyTraits::element(slots_ + seq.offset((size_t)i))))) + return {seq.offset((size_t)i), false}; + } + if (PHMAP_PREDICT_TRUE(g.MatchEmpty())) break; + seq.next(); + } + return {prepare_insert(hashval), true}; + } + + template + std::pair find_or_prepare_insert(const K& key) { + return find_or_prepare_insert(key, this->hash(key)); + } + + size_t prepare_insert(size_t hashval) PHMAP_ATTRIBUTE_NOINLINE { + auto target = find_first_non_full(hashval); + if (PHMAP_PREDICT_FALSE(growth_left() == 0 && !IsDeleted(ctrl_[target.offset]))) { + rehash_and_grow_if_necessary(); + target = find_first_non_full(hashval); + } + ++size_; + growth_left() -= IsEmpty(ctrl_[target.offset]); + set_ctrl(target.offset, H2(hashval)); + infoz_.RecordInsert(hashval, target.probe_length); + return target.offset; + } + + // Constructs the value in the space pointed by the iterator. This only works + // after an unsuccessful find_or_prepare_insert() and before any other + // modifications happen in the raw_hash_set. + // + // PRECONDITION: i is an index returned from find_or_prepare_insert(k), where + // k is the key decomposed from `forward(args)...`, and the bool + // returned by find_or_prepare_insert(k) was true. + // POSTCONDITION: *m.iterator_at(i) == value_type(forward(args)...). + template + void emplace_at(size_t i, Args&&... args) { + PolicyTraits::construct(&alloc_ref(), slots_ + i, std::forward(args)...); + + assert(PolicyTraits::apply(FindElement{*this}, *iterator_at(i)) == iterator_at(i) && + "constructed value does not match the lookup key"); + } + + iterator iterator_at(size_t i) { return {ctrl_ + i, slots_ + i}; } + const_iterator iterator_at(size_t i) const { return {ctrl_ + i, slots_ + i}; } + + private: + friend struct RawHashSetTestOnlyAccess; + + probe_seq probe(size_t hashval) const { + return probe_seq(H1(hashval, ctrl_), capacity_); + } + + // Reset all ctrl bytes back to kEmpty, except the sentinel. + void reset_ctrl(size_t capacity) { + std::memset(ctrl_, kEmpty, capacity + Group::kWidth); + ctrl_[capacity] = kSentinel; + SanitizerPoisonMemoryRegion(slots_, sizeof(slot_type) * capacity); + } + + void reset_growth_left(size_t capacity) { + growth_left() = CapacityToGrowth(capacity) - size_; + } + + // Sets the control byte, and if `i < Group::kWidth`, set the cloned byte at + // the end too. + void set_ctrl(size_t i, ctrl_t h) { + assert(i < capacity_); + + if (IsFull(h)) { + SanitizerUnpoisonObject(slots_ + i); + } else { + SanitizerPoisonObject(slots_ + i); + } + + ctrl_[i] = h; + ctrl_[((i - Group::kWidth) & capacity_) + 1 + ((Group::kWidth - 1) & capacity_)] = h; + } + + size_t& growth_left() { return settings_.template get<0>(); } + + template class RefSet, class M, + class P, class H, class E, class A> + friend class parallel_hash_set; + + template class RefSet, class M, + class P, class H, class E, class A> + friend class parallel_hash_map; + + // The representation of the object has two modes: + // - small: For capacities < kWidth-1 + // - large: For the rest. + // + // Differences: + // - In small mode we are able to use the whole capacity. The extra control + // bytes give us at least one "empty" control byte to stop the iteration. + // This is important to make 1 a valid capacity. + // + // - In small mode only the first `capacity()` control bytes after the + // sentinel are valid. The rest contain dummy kEmpty values that do not + // represent a real slot. This is important to take into account on + // find_first_non_full(), where we never try ShouldInsertBackwards() for + // small tables. + bool is_small() const { return capacity_ < Group::kWidth - 1; } + + hasher& hash_ref() { return settings_.template get<1>(); } + const hasher& hash_ref() const { return settings_.template get<1>(); } + key_equal& eq_ref() { return settings_.template get<2>(); } + const key_equal& eq_ref() const { return settings_.template get<2>(); } + allocator_type& alloc_ref() { return settings_.template get<3>(); } + const allocator_type& alloc_ref() const { return settings_.template get<3>(); } + + // TODO(alkis): Investigate removing some of these fields: + // - ctrl/slots can be derived from each other + // - size can be moved into the slot array + ctrl_t* ctrl_ = EmptyGroup(); // [(capacity + 1) * ctrl_t] + slot_type* slots_ = nullptr; // [capacity * slot_type] + size_t size_ = 0; // number of full slots + size_t capacity_ = 0; // total number of slots + HashtablezInfoHandle infoz_; + phmap::priv::CompressedTuple + settings_{0, hasher{}, key_equal{}, allocator_type{}}; +}; + +// -------------------------------------------------------------------------- +// -------------------------------------------------------------------------- +template +class raw_hash_map : public raw_hash_set { + // P is Policy. It's passed as a template argument to support maps that have + // incomplete types as values, as in unordered_map. + // MappedReference<> may be a non-reference type. + template + using MappedReference = decltype( + P::value(std::addressof(std::declval()))); + + // MappedConstReference<> may be a non-reference type. + template + using MappedConstReference = decltype( + P::value(std::addressof(std::declval()))); + + using KeyArgImpl = KeyArg::value && IsTransparent::value>; + + using Base = raw_hash_set; + + public: + using key_type = typename Policy::key_type; + using mapped_type = typename Policy::mapped_type; + template + using key_arg = typename KeyArgImpl::template type; + + static_assert(!std::is_reference::value, ""); + // TODO(alkis): remove this assertion and verify that reference mapped_type is + // supported. + static_assert(!std::is_reference::value, ""); + + using iterator = typename raw_hash_map::raw_hash_set::iterator; + using const_iterator = typename raw_hash_map::raw_hash_set::const_iterator; + + raw_hash_map() {} + using Base::raw_hash_set; // use raw_hash_set constructor + + // The last two template parameters ensure that both arguments are rvalues + // (lvalue arguments are handled by the overloads below). This is necessary + // for supporting bitfield arguments. + // + // union { int n : 1; }; + // flat_hash_map m; + // m.insert_or_assign(n, n); + template + std::pair insert_or_assign(key_arg&& k, V&& v) { + return insert_or_assign_impl(std::forward(k), std::forward(v)); + } + + template + std::pair insert_or_assign(key_arg&& k, const V& v) { + return insert_or_assign_impl(std::forward(k), v); + } + + template + std::pair insert_or_assign(const key_arg& k, V&& v) { + return insert_or_assign_impl(k, std::forward(v)); + } + + template + std::pair insert_or_assign(const key_arg& k, const V& v) { + return insert_or_assign_impl(k, v); + } + + template + iterator insert_or_assign(const_iterator, key_arg&& k, V&& v) { + return insert_or_assign(std::forward(k), std::forward(v)).first; + } + + template + iterator insert_or_assign(const_iterator, key_arg&& k, const V& v) { + return insert_or_assign(std::forward(k), v).first; + } + + template + iterator insert_or_assign(const_iterator, const key_arg& k, V&& v) { + return insert_or_assign(k, std::forward(v)).first; + } + + template + iterator insert_or_assign(const_iterator, const key_arg& k, const V& v) { + return insert_or_assign(k, v).first; + } + + template ::value, + int>::type = 0, + K* = nullptr> + std::pair try_emplace(key_arg&& k, Args&&... args) { + return try_emplace_impl(std::forward(k), std::forward(args)...); + } + + template ::value, + int>::type = 0> + std::pair try_emplace(const key_arg& k, Args&&... args) { + return try_emplace_impl(k, std::forward(args)...); + } + + template + iterator try_emplace(const_iterator, key_arg&& k, Args&&... args) { + return try_emplace(std::forward(k), std::forward(args)...).first; + } + + template + iterator try_emplace(const_iterator, const key_arg& k, Args&&... args) { + return try_emplace(k, std::forward(args)...).first; + } + + template + MappedReference

at(const key_arg& key) { + auto it = this->find(key); + if (it == this->end()) + phmap::base_internal::ThrowStdOutOfRange("phmap at(): lookup non-existent key"); + return Policy::value(&*it); + } + + template + MappedConstReference

at(const key_arg& key) const { + auto it = this->find(key); + if (it == this->end()) + phmap::base_internal::ThrowStdOutOfRange("phmap at(): lookup non-existent key"); + return Policy::value(&*it); + } + + template + MappedReference

operator[](key_arg&& key) { + return Policy::value(&*try_emplace(std::forward(key)).first); + } + + template + MappedReference

operator[](const key_arg& key) { + return Policy::value(&*try_emplace(key).first); + } + + private: + template + std::pair insert_or_assign_impl(K&& k, V&& v) { + auto res = this->find_or_prepare_insert(k); + if (res.second) + this->emplace_at(res.first, std::forward(k), std::forward(v)); + else + Policy::value(&*this->iterator_at(res.first)) = std::forward(v); + return {this->iterator_at(res.first), res.second}; + } + + template + std::pair try_emplace_impl(K&& k, Args&&... args) { + auto res = this->find_or_prepare_insert(k); + if (res.second) + this->emplace_at(res.first, std::piecewise_construct, + std::forward_as_tuple(std::forward(k)), + std::forward_as_tuple(std::forward(args)...)); + return {this->iterator_at(res.first), res.second}; + } +}; + +// ---------------------------------------------------------------------------- +// ---------------------------------------------------------------------------- +// Returns "random" seed. +inline size_t RandomSeed() { +#if PHMAP_HAVE_THREAD_LOCAL + static thread_local size_t counter = 0; + size_t value = ++counter; +#else // PHMAP_HAVE_THREAD_LOCAL + static std::atomic counter(0); + size_t value = counter.fetch_add(1, std::memory_order_relaxed); +#endif // PHMAP_HAVE_THREAD_LOCAL + return value ^ static_cast(reinterpret_cast(&counter)); +} + +// ---------------------------------------------------------------------------- +// ---------------------------------------------------------------------------- +template class RefSet, class Mtx_, + class Policy, class Hash, class Eq, class Alloc> +class parallel_hash_set { + using PolicyTraits = hash_policy_traits; + using KeyArgImpl = KeyArg::value && IsTransparent::value>; + + static_assert(N <= 12, "N = 12 means 4096 hash tables!"); + constexpr static size_t num_tables = 1 << N; + constexpr static size_t mask = num_tables - 1; + + public: + using EmbeddedSet = RefSet; + using EmbeddedIterator = typename EmbeddedSet::iterator; + using EmbeddedConstIterator = typename EmbeddedSet::const_iterator; + using constructor = typename EmbeddedSet::constructor; + using init_type = typename PolicyTraits::init_type; + using key_type = typename PolicyTraits::key_type; + using slot_type = typename PolicyTraits::slot_type; + using allocator_type = Alloc; + using size_type = size_t; + using difference_type = ptrdiff_t; + using hasher = Hash; + using key_equal = Eq; + using policy_type = Policy; + using value_type = typename PolicyTraits::value_type; + using reference = value_type&; + using const_reference = const value_type&; + using pointer = typename phmap::allocator_traits< + allocator_type>::template rebind_traits::pointer; + using const_pointer = typename phmap::allocator_traits< + allocator_type>::template rebind_traits::const_pointer; + + // Alias used for heterogeneous lookup functions. + // `key_arg` evaluates to `K` when the functors are transparent and to + // `key_type` otherwise. It permits template argument deduction on `K` for the + // transparent case. + // -------------------------------------------------------------------- + template + using key_arg = typename KeyArgImpl::template type; + + protected: + using Lockable = phmap::LockableImpl; + + // -------------------------------------------------------------------- + struct Inner : public Lockable { + bool operator==(const Inner& o) const { + typename Lockable::SharedLocks l(const_cast(*this), const_cast(o)); + return set_ == o.set_; + } + + EmbeddedSet set_; + }; + + private: + // Give an early error when key_type is not hashable/eq. + // -------------------------------------------------------------------- + auto KeyTypeCanBeHashed(const Hash& h, const key_type& k) -> decltype(h(k)); + auto KeyTypeCanBeEq(const Eq& eq, const key_type& k) -> decltype(eq(k, k)); + + using AllocTraits = phmap::allocator_traits; + + static_assert(std::is_lvalue_reference::value, + "Policy::element() must return a reference"); + + template + struct SameAsElementReference + : std::is_same< + typename std::remove_cv< + typename std::remove_reference::type>::type, + typename std::remove_cv::type>::type> {}; + + // An enabler for insert(T&&): T must be convertible to init_type or be the + // same as [cv] value_type [ref]. + // Note: we separate SameAsElementReference into its own type to avoid using + // reference unless we need to. MSVC doesn't seem to like it in some + // cases. + // -------------------------------------------------------------------- + template + using RequiresInsertable = + typename std::enable_if, + SameAsElementReference>::value, + int>::type; + + // RequiresNotInit is a workaround for gcc prior to 7.1. + // See https://godbolt.org/g/Y4xsUh. + template + using RequiresNotInit = + typename std::enable_if::value, int>::type; + + template + using IsDecomposable = IsDecomposable; + + public: + static_assert(std::is_same::value, + "Allocators with custom pointer types are not supported"); + static_assert(std::is_same::value, + "Allocators with custom pointer types are not supported"); + + // --------------------- i t e r a t o r ------------------------------ + class iterator { + friend class parallel_hash_set; + + public: + using iterator_category = std::forward_iterator_tag; + using value_type = typename parallel_hash_set::value_type; + using reference = phmap::conditional_t; + using pointer = phmap::remove_reference_t*; + using difference_type = typename parallel_hash_set::difference_type; + using Inner = typename parallel_hash_set::Inner; + using EmbeddedSet = typename parallel_hash_set::EmbeddedSet; + using EmbeddedIterator = typename EmbeddedSet::iterator; + + iterator() {} + + reference operator*() const { return *it_; } + pointer operator->() const { return &operator*(); } + + iterator& operator++() { + assert(inner_); // null inner means we are already at the end + ++it_; + skip_empty(); + return *this; + } + + iterator operator++(int) { + assert(inner_); // null inner means we are already at the end + auto tmp = *this; + ++*this; + return tmp; + } + + friend bool operator==(const iterator& a, const iterator& b) { + return a.inner_ == b.inner_ && (!a.inner_ || a.it_ == b.it_); + } + + friend bool operator!=(const iterator& a, const iterator& b) { return !(a == b); } + + private: + iterator(Inner* inner, Inner* inner_end, const EmbeddedIterator& it) + : inner_(inner), inner_end_(inner_end), it_(it) { // for begin() and end() + if (inner) it_end_ = inner->set_.end(); + } + + void skip_empty() { + while (it_ == it_end_) { + ++inner_; + if (inner_ == inner_end_) { + inner_ = nullptr; // marks end() + break; + } else { + it_ = inner_->set_.begin(); + it_end_ = inner_->set_.end(); + } + } + } + + Inner* inner_ = nullptr; + Inner* inner_end_ = nullptr; + EmbeddedIterator it_, it_end_; + }; + + // --------------------- c o n s t i t e r a t o r ----------------- + class const_iterator { + friend class parallel_hash_set; + + public: + using iterator_category = typename iterator::iterator_category; + using value_type = typename parallel_hash_set::value_type; + using reference = typename parallel_hash_set::const_reference; + using pointer = typename parallel_hash_set::const_pointer; + using difference_type = typename parallel_hash_set::difference_type; + using Inner = typename parallel_hash_set::Inner; + + const_iterator() {} + // Implicit construction from iterator. + const_iterator(iterator i) : iter_(std::move(i)) {} + + reference operator*() const { return *(iter_); } + pointer operator->() const { return iter_.operator->(); } + + const_iterator& operator++() { + ++iter_; + return *this; + } + const_iterator operator++(int) { return iter_++; } + + friend bool operator==(const const_iterator& a, const const_iterator& b) { + return a.iter_ == b.iter_; + } + friend bool operator!=(const const_iterator& a, const const_iterator& b) { + return !(a == b); + } + + private: + const_iterator(const Inner* inner, const Inner* inner_end, const EmbeddedIterator& it) + : iter_(const_cast(inner), const_cast(inner_end), + const_cast(it)) {} + + iterator iter_; + }; + + using node_type = node_handle, Alloc>; + using insert_return_type = InsertReturnType; + + // ------------------------- c o n s t r u c t o r s ------------------ + + parallel_hash_set() noexcept( + std::is_nothrow_default_constructible::value&& + std::is_nothrow_default_constructible::value&& + std::is_nothrow_default_constructible::value) {} + + explicit parallel_hash_set(size_t bucket_cnt, const hasher& hash_param = hasher(), + const key_equal& eq = key_equal(), + const allocator_type& alloc = allocator_type()) { + for (auto& inner : sets_) + inner.set_ = EmbeddedSet(bucket_cnt / N, hash_param, eq, alloc); + } + + parallel_hash_set(size_t bucket_cnt, const hasher& hash_param, + const allocator_type& alloc) + : parallel_hash_set(bucket_cnt, hash_param, key_equal(), alloc) {} + + parallel_hash_set(size_t bucket_cnt, const allocator_type& alloc) + : parallel_hash_set(bucket_cnt, hasher(), key_equal(), alloc) {} + + explicit parallel_hash_set(const allocator_type& alloc) + : parallel_hash_set(0, hasher(), key_equal(), alloc) {} + + template + parallel_hash_set(InputIter first, InputIter last, size_t bucket_cnt = 0, + const hasher& hash_param = hasher(), + const key_equal& eq = key_equal(), + const allocator_type& alloc = allocator_type()) + : parallel_hash_set(bucket_cnt, hash_param, eq, alloc) { + insert(first, last); + } + + template + parallel_hash_set(InputIter first, InputIter last, size_t bucket_cnt, + const hasher& hash_param, const allocator_type& alloc) + : parallel_hash_set(first, last, bucket_cnt, hash_param, key_equal(), alloc) {} + + template + parallel_hash_set(InputIter first, InputIter last, size_t bucket_cnt, + const allocator_type& alloc) + : parallel_hash_set(first, last, bucket_cnt, hasher(), key_equal(), alloc) {} + + template + parallel_hash_set(InputIter first, InputIter last, const allocator_type& alloc) + : parallel_hash_set(first, last, 0, hasher(), key_equal(), alloc) {} + + // Instead of accepting std::initializer_list as the first + // argument like std::unordered_set does, we have two overloads + // that accept std::initializer_list and std::initializer_list. + // This is advantageous for performance. + // + // // Turns {"abc", "def"} into std::initializer_list, then copies + // // the strings into the set. + // std::unordered_set s = {"abc", "def"}; + // + // // Turns {"abc", "def"} into std::initializer_list, then + // // copies the strings into the set. + // phmap::flat_hash_set s = {"abc", "def"}; + // + // The same trick is used in insert(). + // + // The enabler is necessary to prevent this constructor from triggering where + // the copy constructor is meant to be called. + // + // phmap::flat_hash_set a, b{a}; + // + // RequiresNotInit is a workaround for gcc prior to 7.1. + // -------------------------------------------------------------------- + template = 0, RequiresInsertable = 0> + parallel_hash_set(std::initializer_list init, size_t bucket_cnt = 0, + const hasher& hash_param = hasher(), + const key_equal& eq = key_equal(), + const allocator_type& alloc = allocator_type()) + : parallel_hash_set(init.begin(), init.end(), bucket_cnt, hash_param, eq, alloc) {} + + parallel_hash_set(std::initializer_list init, size_t bucket_cnt = 0, + const hasher& hash_param = hasher(), + const key_equal& eq = key_equal(), + const allocator_type& alloc = allocator_type()) + : parallel_hash_set(init.begin(), init.end(), bucket_cnt, hash_param, eq, alloc) {} + + template = 0, RequiresInsertable = 0> + parallel_hash_set(std::initializer_list init, size_t bucket_cnt, + const hasher& hash_param, const allocator_type& alloc) + : parallel_hash_set(init, bucket_cnt, hash_param, key_equal(), alloc) {} + + parallel_hash_set(std::initializer_list init, size_t bucket_cnt, + const hasher& hash_param, const allocator_type& alloc) + : parallel_hash_set(init, bucket_cnt, hash_param, key_equal(), alloc) {} + + template = 0, RequiresInsertable = 0> + parallel_hash_set(std::initializer_list init, size_t bucket_cnt, + const allocator_type& alloc) + : parallel_hash_set(init, bucket_cnt, hasher(), key_equal(), alloc) {} + + parallel_hash_set(std::initializer_list init, size_t bucket_cnt, + const allocator_type& alloc) + : parallel_hash_set(init, bucket_cnt, hasher(), key_equal(), alloc) {} + + template = 0, RequiresInsertable = 0> + parallel_hash_set(std::initializer_list init, const allocator_type& alloc) + : parallel_hash_set(init, 0, hasher(), key_equal(), alloc) {} + + parallel_hash_set(std::initializer_list init, const allocator_type& alloc) + : parallel_hash_set(init, 0, hasher(), key_equal(), alloc) {} + + parallel_hash_set(const parallel_hash_set& that) + : parallel_hash_set( + that, AllocTraits::select_on_container_copy_construction(that.alloc_ref())) {} + + parallel_hash_set(const parallel_hash_set& that, const allocator_type& a) + : parallel_hash_set(0, that.hash_ref(), that.eq_ref(), a) { + for (size_t i = 0; i < num_tables; ++i) sets_[i].set_ = {that.sets_[i].set_, a}; + } + + parallel_hash_set(parallel_hash_set&& that) noexcept( + std::is_nothrow_copy_constructible::value&& + std::is_nothrow_copy_constructible::value&& + std::is_nothrow_copy_constructible::value) + : parallel_hash_set(std::move(that), that.alloc_ref()) {} + + parallel_hash_set(parallel_hash_set&& that, const allocator_type& a) { + for (size_t i = 0; i < num_tables; ++i) + sets_[i].set_ = {std::move(that.sets_[i]).set_, a}; + } + + parallel_hash_set& operator=(const parallel_hash_set& that) { + for (size_t i = 0; i < num_tables; ++i) sets_[i].set_ = that.sets_[i].set_; + return *this; + } + + parallel_hash_set& operator=(parallel_hash_set&& that) noexcept( + phmap::allocator_traits::is_always_equal::value&& + std::is_nothrow_move_assignable::value&& + std::is_nothrow_move_assignable::value) { + for (size_t i = 0; i < num_tables; ++i) sets_[i].set_ = std::move(that.sets_[i].set_); + return *this; + } + + ~parallel_hash_set() {} + + iterator begin() { + auto it = iterator(&sets_[0], &sets_[0] + num_tables, sets_[0].set_.begin()); + it.skip_empty(); + return it; + } + + iterator end() { return iterator(); } + const_iterator begin() const { return const_cast(this)->begin(); } + const_iterator end() const { return const_cast(this)->end(); } + const_iterator cbegin() const { return begin(); } + const_iterator cend() const { return end(); } + + bool empty() const { return !size(); } + + size_t size() const { + size_t sz = 0; + for (const auto& inner : sets_) sz += inner.set_.size(); + return sz; + } + + size_t capacity() const { + size_t c = 0; + for (const auto& inner : sets_) c += inner.set_.capacity(); + return c; + } + + size_t max_size() const { return (std::numeric_limits::max)(); } + + PHMAP_ATTRIBUTE_REINITIALIZES void clear() { + for (auto& inner : sets_) { + typename Lockable::UniqueLock m(inner); + inner.set_.clear(); + } + } + + // extension - clears only soecified submap + // ---------------------------------------- + void clear(std::size_t submap_index) { + Inner& inner = sets_[submap_index]; + typename Lockable::UniqueLock m(inner); + inner.set_.clear(); + } + + // This overload kicks in when the argument is an rvalue of insertable and + // decomposable type other than init_type. + // + // flat_hash_map m; + // m.insert(std::make_pair("abc", 42)); + // -------------------------------------------------------------------- + template = 0, + typename std::enable_if::value, int>::type = 0, + T* = nullptr> + std::pair insert(T&& value) { + return emplace(std::forward(value)); + } + + // This overload kicks in when the argument is a bitfield or an lvalue of + // insertable and decomposable type. + // + // union { int n : 1; }; + // flat_hash_set s; + // s.insert(n); + // + // flat_hash_set s; + // const char* p = "hello"; + // s.insert(p); + // + // TODO(romanp): Once we stop supporting gcc 5.1 and below, replace + // RequiresInsertable with RequiresInsertable. + // We are hitting this bug: https://godbolt.org/g/1Vht4f. + // -------------------------------------------------------------------- + template = 0, + typename std::enable_if::value, int>::type = 0> + std::pair insert(const T& value) { + return emplace(value); + } + + // This overload kicks in when the argument is an rvalue of init_type. Its + // purpose is to handle brace-init-list arguments. + // + // flat_hash_set> s; + // s.insert({"abc", 42}); + // -------------------------------------------------------------------- + std::pair insert(init_type&& value) { + return emplace(std::move(value)); + } + + template = 0, + typename std::enable_if::value, int>::type = 0, + T* = nullptr> + iterator insert(const_iterator, T&& value) { + return insert(std::forward(value)).first; + } + + // TODO(romanp): Once we stop supporting gcc 5.1 and below, replace + // RequiresInsertable with RequiresInsertable. + // We are hitting this bug: https://godbolt.org/g/1Vht4f. + // -------------------------------------------------------------------- + template = 0, + typename std::enable_if::value, int>::type = 0> + iterator insert(const_iterator, const T& value) { + return insert(value).first; + } + + iterator insert(const_iterator, init_type&& value) { + return insert(std::move(value)).first; + } + + template + void insert(InputIt first, InputIt last) { + for (; first != last; ++first) insert(*first); + } + + template = 0, RequiresInsertable = 0> + void insert(std::initializer_list ilist) { + insert(ilist.begin(), ilist.end()); + } + + void insert(std::initializer_list ilist) { + insert(ilist.begin(), ilist.end()); + } + + insert_return_type insert(node_type&& node) { + if (!node) return {end(), false, node_type()}; + auto& key = node.key(); + size_t hashval = this->hash(key); + Inner& inner = sets_[subidx(hashval)]; + auto& set = inner.set_; + + typename Lockable::UniqueLock m(inner); + auto res = set.insert(std::move(node), hashval); + return {make_iterator(&inner, res.position), res.inserted, + res.inserted ? node_type() : std::move(res.node)}; + } + + iterator insert(const_iterator, node_type&& node) { + return insert(std::move(node)).first; + } + + struct ReturnKey_ { + template + Key operator()(Key&& k, const Args&...) const { + return std::forward(k); + } + }; + + // -------------------------------------------------------------------- + // phmap expension: emplace_with_hash + // ---------------------------------- + // same as emplace, but hashval is provided + // -------------------------------------------------------------------- + template + std::pair emplace_decomposable_with_hash(const K& key, size_t hashval, + Args&&... args) { + Inner& inner = sets_[subidx(hashval)]; + auto& set = inner.set_; + typename Lockable::UniqueLock m(inner); + return make_rv(&inner, + set.emplace_decomposable(key, hashval, std::forward(args)...)); + } + + struct EmplaceDecomposableHashval { + template + std::pair operator()(const K& key, Args&&... args) const { + return s.emplace_decomposable_with_hash(key, hashval, std::forward(args)...); + } + parallel_hash_set& s; + size_t hashval; + }; + + // This overload kicks in if we can deduce the key from args. This enables us + // to avoid constructing value_type if an entry with the same key already + // exists. + // + // For example: + // + // flat_hash_map m = {{"abc", "def"}}; + // // Creates no std::string copies and makes no heap allocations. + // m.emplace("abc", "xyz"); + // -------------------------------------------------------------------- + template ::value, int>::type = 0> + std::pair emplace_with_hash(size_t hashval, Args&&... args) { + return PolicyTraits::apply(EmplaceDecomposableHashval{*this, hashval}, + std::forward(args)...); + } + + // This overload kicks in if we cannot deduce the key from args. It constructs + // value_type unconditionally and then either moves it into the table or + // destroys. + // -------------------------------------------------------------------- + template ::value, int>::type = 0> + std::pair emplace_with_hash(size_t hashval, Args&&... args) { + typename std::aligned_storage::type raw; + slot_type* slot = reinterpret_cast(&raw); + + PolicyTraits::construct(&alloc_ref(), slot, std::forward(args)...); + const auto& elem = PolicyTraits::element(slot); + Inner& inner = sets_[subidx(hashval)]; + auto& set = inner.set_; + typename Lockable::UniqueLock m(inner); + typename EmbeddedSet::template InsertSlotWithHash f{inner, std::move(*slot), + hashval}; + return make_rv(PolicyTraits::apply(f, elem)); + } + + template + iterator emplace_hint_with_hash(size_t hashval, const_iterator, Args&&... args) { + return emplace_with_hash(hashval, std::forward(args)...).first; + } + + template + iterator lazy_emplace_with_hash(size_t hashval, const key_arg& key, F&& f) { + Inner& inner = sets_[subidx(hashval)]; + auto& set = inner.set_; + typename Lockable::UniqueLock m(inner); + return make_iterator(&inner, + set.lazy_emplace_with_hash(key, hashval, std::forward(f))); + } + + // -------------------------------------------------------------------- + // end of phmap expension + // -------------------------------------------------------------------- + + template + std::pair emplace_decomposable(const K& key, Args&&... args) { + size_t hashval = this->hash(key); + Inner& inner = sets_[subidx(hashval)]; + auto& set = inner.set_; + typename Lockable::UniqueLock m(inner); + return make_rv(&inner, + set.emplace_decomposable(key, hashval, std::forward(args)...)); + } + + struct EmplaceDecomposable { + template + std::pair operator()(const K& key, Args&&... args) const { + return s.emplace_decomposable(key, std::forward(args)...); + } + parallel_hash_set& s; + }; + + // This overload kicks in if we can deduce the key from args. This enables us + // to avoid constructing value_type if an entry with the same key already + // exists. + // + // For example: + // + // flat_hash_map m = {{"abc", "def"}}; + // // Creates no std::string copies and makes no heap allocations. + // m.emplace("abc", "xyz"); + // -------------------------------------------------------------------- + template ::value, int>::type = 0> + std::pair emplace(Args&&... args) { + return PolicyTraits::apply(EmplaceDecomposable{*this}, std::forward(args)...); + } + + // This overload kicks in if we cannot deduce the key from args. It constructs + // value_type unconditionally and then either moves it into the table or + // destroys. + // -------------------------------------------------------------------- + template ::value, int>::type = 0> + std::pair emplace(Args&&... args) { + typename std::aligned_storage::type raw; + slot_type* slot = reinterpret_cast(&raw); + size_t hashval = this->hash(PolicyTraits::key(slot)); + + PolicyTraits::construct(&alloc_ref(), slot, std::forward(args)...); + const auto& elem = PolicyTraits::element(slot); + Inner& inner = sets_[subidx(hashval)]; + auto& set = inner.set_; + typename Lockable::UniqueLock m(inner); + typename EmbeddedSet::template InsertSlotWithHash f{inner, std::move(*slot), + hashval}; + return make_rv(PolicyTraits::apply(f, elem)); + } + + template + iterator emplace_hint(const_iterator, Args&&... args) { + return emplace(std::forward(args)...).first; + } + + iterator make_iterator(Inner* inner, const EmbeddedIterator it) { + if (it == inner->set_.end()) return iterator(); + return iterator(inner, &sets_[0] + num_tables, it); + } + + std::pair make_rv(Inner* inner, + const std::pair& res) { + return {iterator(inner, &sets_[0] + num_tables, res.first), res.second}; + } + + template + iterator lazy_emplace(const key_arg& key, F&& f) { + auto hashval = this->hash(key); + Inner& inner = sets_[subidx(hashval)]; + auto& set = inner.set_; + typename Lockable::UniqueLock m(inner); + return make_iterator(&inner, + set.lazy_emplace_with_hash(key, hashval, std::forward(f))); + } + + template + bool lazy_emplace_l(const key_arg& key, FExists&& fExists, FEmplace&& fEmplace) { + typename Lockable::UniqueLock m; + auto res = this->find_or_prepare_insert(key, m); + Inner* inner = std::get<0>(res); + if (std::get<2>(res)) + inner->set_.lazy_emplace_at(std::get<1>(res), std::forward(fEmplace)); + else { + auto it = this->iterator_at(inner, inner->set_.iterator_at(std::get<1>(res))); + std::forward(fExists)(Policy::value(&*it)); + } + return std::get<2>(res); + } + + // Extension API: support iterating over all values + // + // flat_hash_set s; + // s.insert(...); + // s.for_each([](auto const & key) { + // // Safely iterates over all the keys + // }); + template + void for_each(F&& fCallback) const { + for (auto const& inner : sets_) { + typename Lockable::SharedLock m(const_cast(inner)); + std::for_each(inner.set_.begin(), inner.set_.end(), fCallback); + } + } + + // this version allows to modify the values + void for_each_m(std::function&& fCallback) { + for (auto& inner : sets_) { + typename Lockable::UniqueLock m(const_cast(inner)); + std::for_each(inner.set_.begin(), inner.set_.end(), fCallback); + } + } + + // Extension API: support for heterogeneous keys. + // + // std::unordered_set s; + // // Turns "abc" into std::string. + // s.erase("abc"); + // + // flat_hash_set s; + // // Uses "abc" directly without copying it into std::string. + // s.erase("abc"); + // -------------------------------------------------------------------- + template + size_type erase(const key_arg& key) { + auto hashval = this->hash(key); + Inner& inner = sets_[subidx(hashval)]; + auto& set = inner.set_; + typename Lockable::UpgradeLock m(inner); + auto it = set.find(key, hashval); + if (it == set.end()) return 0; + + typename Lockable::UpgradeToUnique unique(m); + set._erase(it); + return 1; + } + + // -------------------------------------------------------------------- + iterator erase(const_iterator cit) { return erase(cit.iter_); } + + // Erases the element pointed to by `it`. Unlike `std::unordered_set::erase`, + // this method returns void to reduce algorithmic complexity to O(1). In + // order to erase while iterating across a map, use the following idiom (which + // also works for standard containers): + // + // for (auto it = m.begin(), end = m.end(); it != end;) { + // if () { + // m._erase(it++); + // } else { + // ++it; + // } + // } + // -------------------------------------------------------------------- + void _erase(iterator it) { + assert(it.inner_ != nullptr); + it.inner_->set_._erase(it.it_); + } + void _erase(const_iterator cit) { _erase(cit.iter_); } + + // This overload is necessary because otherwise erase(const K&) would be + // a better match if non-const iterator is passed as an argument. + // -------------------------------------------------------------------- + iterator erase(iterator it) { + _erase(it++); + return it; + } + + iterator erase(const_iterator first, const_iterator last) { + while (first != last) { + _erase(first++); + } + return last.iter_; + } + + // Moves elements from `src` into `this`. + // If the element already exists in `this`, it is left unmodified in `src`. + // -------------------------------------------------------------------- + template + void merge(parallel_hash_set& src) { // NOLINT + assert(this != &src); + if (this != &src) { + for (size_t i = 0; i < num_tables; ++i) { + typename Lockable::UniqueLocks l(sets_[i], src.sets_[i]); + sets_[i].set_.merge(src.sets_[i].set_); + } + } + } + + template + void merge(parallel_hash_set&& src) { + merge(src); + } + + node_type extract(const_iterator position) { + return position.iter_.inner_->set_.extract(EmbeddedConstIterator(position.iter_.it_)); + } + + template ::value, int>::type = 0> + node_type extract(const key_arg& key) { + auto it = find(key); + return it == end() ? node_type() : extract(const_iterator{it}); + } + + void swap(parallel_hash_set& that) noexcept( + IsNoThrowSwappable() && + (!AllocTraits::propagate_on_container_swap::value || + IsNoThrowSwappable())) { + using std::swap; + for (size_t i = 0; i < num_tables; ++i) { + typename Lockable::UniqueLocks l(sets_[i], that.sets_[i]); + swap(sets_[i].set_, that.sets_[i].set_); + } + } + + void rehash(size_t n) { + size_t nn = n / num_tables; + for (auto& inner : sets_) { + typename Lockable::UniqueLock m(inner); + inner.set_.rehash(nn); + } + } + + void reserve(size_t n) { + size_t target = GrowthToLowerboundCapacity(n); + size_t normalized = 16 * NormalizeCapacity(n / num_tables); + rehash(normalized > target ? normalized : target); + } + + // Extension API: support for heterogeneous keys. + // + // std::unordered_set s; + // // Turns "abc" into std::string. + // s.count("abc"); + // + // ch_set s; + // // Uses "abc" directly without copying it into std::string. + // s.count("abc"); + // -------------------------------------------------------------------- + template + size_t count(const key_arg& key) const { + return find(key) == end() ? 0 : 1; + } + + // Issues CPU prefetch instructions for the memory needed to find or insert + // a key. Like all lookup functions, this support heterogeneous keys. + // + // NOTE: This is a very low level operation and should not be used without + // specific benchmarks indicating its importance. + // -------------------------------------------------------------------- + void prefetch_hash(size_t hashval) const { + const Inner& inner = sets_[subidx(hashval)]; + const auto& set = inner.set_; + typename Lockable::SharedLock m(const_cast(inner)); + set.prefetch_hash(hashval); + } + + template + void prefetch(const key_arg& key) const { + prefetch_hash(this->hash(key)); + } + + // The API of find() has two extensions. + // + // 1. The hash can be passed by the user. It must be equal to the hash of the + // key. + // + // 2. The type of the key argument doesn't have to be key_type. This is so + // called heterogeneous key support. + // -------------------------------------------------------------------- + template + iterator find(const key_arg& key, size_t hashval) { + typename Lockable::SharedLock m; + return find(key, hashval, m); + } + + template + iterator find(const key_arg& key) { + return find(key, this->hash(key)); + } + + template + const_iterator find(const key_arg& key, size_t hashval) const { + return const_cast(this)->find(key, hashval); + } + + template + const_iterator find(const key_arg& key) const { + return find(key, this->hash(key)); + } + + template + bool contains(const key_arg& key) const { + return find(key) != end(); + } + + template + bool contains(const key_arg& key, size_t hashval) const { + return find(key, hashval) != end(); + } + + template + std::pair equal_range(const key_arg& key) { + auto it = find(key); + if (it != end()) return {it, std::next(it)}; + return {it, it}; + } + + template + std::pair equal_range(const key_arg& key) const { + auto it = find(key); + if (it != end()) return {it, std::next(it)}; + return {it, it}; + } + + size_t bucket_count() const { + size_t sz = 0; + for (const auto& inner : sets_) { + typename Lockable::SharedLock m(const_cast(inner)); + sz += inner.set_.bucket_count(); + } + return sz; + } + + float load_factor() const { + size_t _capacity = bucket_count(); + return _capacity ? static_cast(static_cast(size()) / _capacity) : 0; + } + + float max_load_factor() const { return 1.0f; } + void max_load_factor(float) { + // Does nothing. + } + + hasher hash_function() const { + return hash_ref(); + } // warning: doesn't match internal hash - use hash() member function + key_equal key_eq() const { return eq_ref(); } + allocator_type get_allocator() const { return alloc_ref(); } + + friend bool operator==(const parallel_hash_set& a, const parallel_hash_set& b) { + return std::equal(a.sets_.begin(), a.sets_.end(), b.sets_.begin()); + } + + friend bool operator!=(const parallel_hash_set& a, const parallel_hash_set& b) { + return !(a == b); + } + + friend void swap(parallel_hash_set& a, + parallel_hash_set& b) noexcept(noexcept(a.swap(b))) { + a.swap(b); + } + + template + size_t hash(const K& key) const { + return HashElement{hash_ref()}(key); + } + +#if !defined(PHMAP_NON_DETERMINISTIC) + template + bool phmap_dump(OutputArchive& ar) const; + + template + bool phmap_load(InputArchive& ar); +#endif + + private: + template + friend struct phmap::priv::hashtable_debug_internal::HashtableDebugAccess; + + struct FindElement { + template + const_iterator operator()(const K& key, Args&&...) const { + return s.find(key); + } + const parallel_hash_set& s; + }; + + struct HashElement { + template + size_t operator()(const K& key, Args&&...) const { + return phmap_mix()(h(key)); + } + const hasher& h; + }; + + template + struct EqualElement { + template + bool operator()(const K2& lhs, Args&&...) const { + return eq(lhs, rhs); + } + const K1& rhs; + const key_equal& eq; + }; + + // "erases" the object from the container, except that it doesn't actually + // destroy the object. It only updates all the metadata of the class. + // This can be used in conjunction with Policy::transfer to move the object to + // another place. + // -------------------------------------------------------------------- + void erase_meta_only(const_iterator cit) { + auto& it = cit.iter_; + assert(it.set_ != nullptr); + it.set_.erase_meta_only(const_iterator(it.it_)); + } + + void drop_deletes_without_resize() PHMAP_ATTRIBUTE_NOINLINE { + for (auto& inner : sets_) { + typename Lockable::UniqueLock m(inner); + inner.set_.drop_deletes_without_resize(); + } + } + + bool has_element(const value_type& elem) const { + size_t hashval = PolicyTraits::apply(HashElement{hash_ref()}, elem); + Inner& inner = sets_[subidx(hashval)]; + auto& set = inner.set_; + typename Lockable::SharedLock m(const_cast(inner)); + return set.has_element(elem, hashval); + } + + // TODO(alkis): Optimize this assuming *this and that don't overlap. + // -------------------------------------------------------------------- + parallel_hash_set& move_assign(parallel_hash_set&& that, std::true_type) { + parallel_hash_set tmp(std::move(that)); + swap(tmp); + return *this; + } + + parallel_hash_set& move_assign(parallel_hash_set&& that, std::false_type) { + parallel_hash_set tmp(std::move(that), alloc_ref()); + swap(tmp); + return *this; + } + + protected: + template + pointer find_ptr(const key_arg& key, size_t hashval, L& mutexlock) { + Inner& inner = sets_[subidx(hashval)]; + auto& set = inner.set_; + mutexlock = std::move(L(inner)); + return set.find_ptr(key, hashval); + } + + template + iterator find(const key_arg& key, size_t hashval, L& mutexlock) { + Inner& inner = sets_[subidx(hashval)]; + auto& set = inner.set_; + mutexlock = std::move(L(inner)); + return make_iterator(&inner, set.find(key, hashval)); + } + + template + std::tuple find_or_prepare_insert_with_hash( + size_t hashval, const K& key, typename Lockable::UniqueLock& mutexlock) { + Inner& inner = sets_[subidx(hashval)]; + auto& set = inner.set_; + mutexlock = std::move(typename Lockable::UniqueLock(inner)); + auto p = set.find_or_prepare_insert(key, hashval); // std::pair + return std::make_tuple(&inner, p.first, p.second); + } + + template + std::tuple find_or_prepare_insert( + const K& key, typename Lockable::UniqueLock& mutexlock) { + return find_or_prepare_insert_with_hash(this->hash(key), key, mutexlock); + } + + iterator iterator_at(Inner* inner, const EmbeddedIterator& it) { + return {inner, &sets_[0] + num_tables, it}; + } + const_iterator iterator_at(Inner* inner, const EmbeddedIterator& it) const { + return {inner, &sets_[0] + num_tables, it}; + } + + static size_t subidx(size_t hashval) { + return ((hashval >> 8) ^ (hashval >> 16) ^ (hashval >> 24)) & mask; + } + + static size_t subcnt() { return num_tables; } + + private: + friend struct RawHashSetTestOnlyAccess; + + size_t growth_left() { + size_t sz = 0; + for (const auto& set : sets_) sz += set.growth_left(); + return sz; + } + + hasher& hash_ref() { return sets_[0].set_.hash_ref(); } + const hasher& hash_ref() const { return sets_[0].set_.hash_ref(); } + key_equal& eq_ref() { return sets_[0].set_.eq_ref(); } + const key_equal& eq_ref() const { return sets_[0].set_.eq_ref(); } + allocator_type& alloc_ref() { return sets_[0].set_.alloc_ref(); } + const allocator_type& alloc_ref() const { return sets_[0].set_.alloc_ref(); } + + protected: // protected in case users want to derive fromm this + std::array sets_; +}; + +// -------------------------------------------------------------------------- +// -------------------------------------------------------------------------- +template class RefSet, class Mtx_, + class Policy, class Hash, class Eq, class Alloc> +class parallel_hash_map + : public parallel_hash_set { + // P is Policy. It's passed as a template argument to support maps that have + // incomplete types as values, as in unordered_map. + // MappedReference<> may be a non-reference type. + template + using MappedReference = decltype( + P::value(std::addressof(std::declval()))); + + // MappedConstReference<> may be a non-reference type. + template + using MappedConstReference = decltype(P::value( + std::addressof(std::declval()))); + + using KeyArgImpl = KeyArg::value && IsTransparent::value>; + + using Base = typename parallel_hash_map::parallel_hash_set; + using Lockable = phmap::LockableImpl; + + public: + using key_type = typename Policy::key_type; + using mapped_type = typename Policy::mapped_type; + template + using key_arg = typename KeyArgImpl::template type; + + static_assert(!std::is_reference::value, ""); + // TODO(alkis): remove this assertion and verify that reference mapped_type is + // supported. + static_assert(!std::is_reference::value, ""); + + using iterator = typename parallel_hash_map::parallel_hash_set::iterator; + using const_iterator = typename parallel_hash_map::parallel_hash_set::const_iterator; + + parallel_hash_map() {} + +#ifdef __INTEL_COMPILER + using Base::parallel_hash_set; +#else + using parallel_hash_map::parallel_hash_set::parallel_hash_set; +#endif + + // The last two template parameters ensure that both arguments are rvalues + // (lvalue arguments are handled by the overloads below). This is necessary + // for supporting bitfield arguments. + // + // union { int n : 1; }; + // flat_hash_map m; + // m.insert_or_assign(n, n); + template + std::pair insert_or_assign(key_arg&& k, V&& v) { + return insert_or_assign_impl(std::forward(k), std::forward(v)); + } + + template + std::pair insert_or_assign(key_arg&& k, const V& v) { + return insert_or_assign_impl(std::forward(k), v); + } + + template + std::pair insert_or_assign(const key_arg& k, V&& v) { + return insert_or_assign_impl(k, std::forward(v)); + } + + template + std::pair insert_or_assign(const key_arg& k, const V& v) { + return insert_or_assign_impl(k, v); + } + + template + iterator insert_or_assign(const_iterator, key_arg&& k, V&& v) { + return insert_or_assign(std::forward(k), std::forward(v)).first; + } + + template + iterator insert_or_assign(const_iterator, key_arg&& k, const V& v) { + return insert_or_assign(std::forward(k), v).first; + } + + template + iterator insert_or_assign(const_iterator, const key_arg& k, V&& v) { + return insert_or_assign(k, std::forward(v)).first; + } + + template + iterator insert_or_assign(const_iterator, const key_arg& k, const V& v) { + return insert_or_assign(k, v).first; + } + + template ::value, + int>::type = 0, + K* = nullptr> + std::pair try_emplace(key_arg&& k, Args&&... args) { + return try_emplace_impl(std::forward(k), std::forward(args)...); + } + + template ::value, + int>::type = 0> + std::pair try_emplace(const key_arg& k, Args&&... args) { + return try_emplace_impl(k, std::forward(args)...); + } + + template + iterator try_emplace(const_iterator, key_arg&& k, Args&&... args) { + return try_emplace(std::forward(k), std::forward(args)...).first; + } + + template + iterator try_emplace(const_iterator, const key_arg& k, Args&&... args) { + return try_emplace(k, std::forward(args)...).first; + } + + template + MappedReference

at(const key_arg& key) { + auto it = this->find(key); + if (it == this->end()) + phmap::base_internal::ThrowStdOutOfRange("phmap at(): lookup non-existent key"); + return Policy::value(&*it); + } + + template + MappedConstReference

at(const key_arg& key) const { + auto it = this->find(key); + if (it == this->end()) + phmap::base_internal::ThrowStdOutOfRange("phmap at(): lookup non-existent key"); + return Policy::value(&*it); + } + + // ----------- phmap extensions -------------------------- + + template ::value, + int>::type = 0, + K* = nullptr> + std::pair try_emplace_with_hash(size_t hashval, key_arg&& k, + Args&&... args) { + return try_emplace_impl_with_hash(hashval, std::forward(k), + std::forward(args)...); + } + + template ::value, + int>::type = 0> + std::pair try_emplace_with_hash(size_t hashval, const key_arg& k, + Args&&... args) { + return try_emplace_impl_with_hash(hashval, k, std::forward(args)...); + } + + template + iterator try_emplace_with_hash(size_t hashval, const_iterator, key_arg&& k, + Args&&... args) { + return try_emplace_with_hash(hashval, std::forward(k), std::forward(args)...) + .first; + } + + template + iterator try_emplace_with_hash(size_t hashval, const_iterator, const key_arg& k, + Args&&... args) { + return try_emplace_with_hash(hashval, k, std::forward(args)...).first; + } + + // if map contains key, lambda is called with the mapped value (under read lock + // protection), and if_contains returns true. This is a const API and lambda should not + // modify the value + // ----------------------------------------------------------------------------------------- + template + bool if_contains(const key_arg& key, F&& f) const { + return const_cast(this) + ->template modify_if_impl( + key, std::forward(f)); + } + + // if map contains key, lambda is called with the mapped value without read lock + // protection, and if_contains_unsafe returns true. This is a const API and lambda + // should not modify the value This should be used only if we know that no other thread + // may be mutating the map at the time. + // ----------------------------------------------------------------------------------------- + template + bool if_contains_unsafe(const key_arg& key, F&& f) const { + return const_cast(this) + ->template modify_if_impl::DoNothing>( + key, std::forward(f)); + } + + // if map contains key, lambda is called with the mapped value (under write lock + // protection), and modify_if returns true. This is a non-const API and lambda is + // allowed to modify the mapped value + // ---------------------------------------------------------------------------------------------------- + template + bool modify_if(const key_arg& key, F&& f) { + return modify_if_impl(key, std::forward(f)); + } + + // if map contains key, lambda is called with the mapped value (under write lock + // protection). If the lambda returns true, the key is subsequently erased from the map + // (the write lock is only released after erase). returns true if key was erased, false + // otherwise. + // ---------------------------------------------------------------------------------------------------- + template + bool erase_if(const key_arg& key, F&& f) { + return erase_if_impl(key, std::forward(f)); + } + + // if map does not contains key, it is inserted and the mapped value is + // value-constructed with the provided arguments (if any), as with try_emplace. if map + // already contains key, then the lambda is called with the mapped value (under write + // lock protection) and can update the mapped value. returns true if key was not already + // present, false otherwise. + // --------------------------------------------------------------------------------------- + template + bool try_emplace_l(K&& k, F&& f, Args&&... args) { + typename Lockable::UniqueLock m; + auto res = this->find_or_prepare_insert(k, m); + typename Base::Inner* inner = std::get<0>(res); + if (std::get<2>(res)) + inner->set_.emplace_at(std::get<1>(res), std::piecewise_construct, + std::forward_as_tuple(std::forward(k)), + std::forward_as_tuple(std::forward(args)...)); + else { + auto it = this->iterator_at(inner, inner->set_.iterator_at(std::get<1>(res))); + std::forward(f)(Policy::value(&*it)); + } + return std::get<2>(res); + } + + // ----------- end of phmap extensions -------------------------- + + template + MappedReference

operator[](key_arg&& key) { + return Policy::value(&*try_emplace(std::forward(key)).first); + } + + template + MappedReference

operator[](const key_arg& key) { + return Policy::value(&*try_emplace(key).first); + } + + private: + template + bool modify_if_impl(const key_arg& key, F&& f) { +#if __cplusplus >= 201703L + static_assert(std::is_invocable::value); +#endif + L m; + auto ptr = this->template find_ptr(key, this->hash(key), m); + if (ptr == nullptr) return false; + std::forward(f)(Policy::value(ptr)); + return true; + } + + template + bool erase_if_impl(const key_arg& key, F&& f) { +#if __cplusplus >= 201703L + static_assert(std::is_invocable::value); +#endif + L m; + auto it = this->template find(key, this->hash(key), m); + if (it == this->end()) return false; + if (std::forward(f)(Policy::value(&*it))) { + this->erase(it); + return true; + } + return false; + } + + template + std::pair insert_or_assign_impl(K&& k, V&& v) { + typename Lockable::UniqueLock m; + auto res = this->find_or_prepare_insert(k, m); + typename Base::Inner* inner = std::get<0>(res); + if (std::get<2>(res)) + inner->set_.emplace_at(std::get<1>(res), std::forward(k), std::forward(v)); + else + Policy::value(&*inner->set_.iterator_at(std::get<1>(res))) = std::forward(v); + return {this->iterator_at(inner, inner->set_.iterator_at(std::get<1>(res))), + std::get<2>(res)}; + } + + template + std::pair try_emplace_impl(K&& k, Args&&... args) { + typename Lockable::UniqueLock m; + auto res = this->find_or_prepare_insert(k, m); + typename Base::Inner* inner = std::get<0>(res); + if (std::get<2>(res)) + inner->set_.emplace_at(std::get<1>(res), std::piecewise_construct, + std::forward_as_tuple(std::forward(k)), + std::forward_as_tuple(std::forward(args)...)); + return {this->iterator_at(inner, inner->set_.iterator_at(std::get<1>(res))), + std::get<2>(res)}; + } + + template + std::pair try_emplace_impl_with_hash(size_t hashval, K&& k, + Args&&... args) { + typename Lockable::UniqueLock m; + auto res = this->find_or_prepare_insert_with_hash(hashval, k, m); + typename Base::Inner* inner = std::get<0>(res); + if (std::get<2>(res)) + inner->set_.emplace_at(std::get<1>(res), std::piecewise_construct, + std::forward_as_tuple(std::forward(k)), + std::forward_as_tuple(std::forward(args)...)); + return {this->iterator_at(inner, inner->set_.iterator_at(std::get<1>(res))), + std::get<2>(res)}; + } +}; + +// Constructs T into uninitialized storage pointed by `ptr` using the args +// specified in the tuple. +// ---------------------------------------------------------------------------- +template +void ConstructFromTuple(Alloc* alloc, T* ptr, Tuple&& t) { + memory_internal::ConstructFromTupleImpl( + alloc, ptr, std::forward(t), + phmap::make_index_sequence< + std::tuple_size::type>::value>()); +} + +// Constructs T using the args specified in the tuple and calls F with the +// constructed value. +// ---------------------------------------------------------------------------- +template +decltype(std::declval()(std::declval())) WithConstructed(Tuple&& t, F&& f) { + return memory_internal::WithConstructedImpl( + std::forward(t), + phmap::make_index_sequence< + std::tuple_size::type>::value>(), + std::forward(f)); +} + +// ---------------------------------------------------------------------------- +// Given arguments of an std::pair's consructor, PairArgs() returns a pair of +// tuples with references to the passed arguments. The tuples contain +// constructor arguments for the first and the second elements of the pair. +// +// The following two snippets are equivalent. +// +// 1. std::pair p(args...); +// +// 2. auto a = PairArgs(args...); +// std::pair p(std::piecewise_construct, +// std::move(p.first), std::move(p.second)); +// ---------------------------------------------------------------------------- +inline std::pair, std::tuple<>> PairArgs() { return {}; } + +template +std::pair, std::tuple> PairArgs(F&& f, S&& s) { + return {std::piecewise_construct, std::forward_as_tuple(std::forward(f)), + std::forward_as_tuple(std::forward(s))}; +} + +template +std::pair, std::tuple> PairArgs(const std::pair& p) { + return PairArgs(p.first, p.second); +} + +template +std::pair, std::tuple> PairArgs(std::pair&& p) { + return PairArgs(std::forward(p.first), std::forward(p.second)); +} + +template +auto PairArgs(std::piecewise_construct_t, F&& f, S&& s) + -> decltype(std::make_pair(memory_internal::TupleRef(std::forward(f)), + memory_internal::TupleRef(std::forward(s)))) { + return std::make_pair(memory_internal::TupleRef(std::forward(f)), + memory_internal::TupleRef(std::forward(s))); +} + +// A helper function for implementing apply() in map policies. +// ---------------------------------------------------------------------------- +template +auto DecomposePair(F&& f, Args&&... args) -> decltype(memory_internal::DecomposePairImpl( + std::forward(f), PairArgs(std::forward(args)...))) { + return memory_internal::DecomposePairImpl(std::forward(f), + PairArgs(std::forward(args)...)); +} + +// A helper function for implementing apply() in set policies. +// ---------------------------------------------------------------------------- +template +decltype(std::declval()(std::declval(), std::declval())) +DecomposeValue(F&& f, Arg&& arg) { + const auto& key = arg; + return std::forward(f)(key, std::forward(arg)); +} + +// -------------------------------------------------------------------------- +// Policy: a policy defines how to perform different operations on +// the slots of the hashtable (see hash_policy_traits.h for the full interface +// of policy). +// +// Hash: a (possibly polymorphic) functor that hashes keys of the hashtable. The +// functor should accept a key and return size_t as hash. For best performance +// it is important that the hash function provides high entropy across all bits +// of the hash. +// +// Eq: a (possibly polymorphic) functor that compares two keys for equality. It +// should accept two (of possibly different type) keys and return a bool: true +// if they are equal, false if they are not. If two keys compare equal, then +// their hash values as defined by Hash MUST be equal. +// +// Allocator: an Allocator [https://devdocs.io/cpp/concept/allocator] with which +// the storage of the hashtable will be allocated and the elements will be +// constructed and destroyed. +// -------------------------------------------------------------------------- +template +struct FlatHashSetPolicy { + using slot_type = T; + using key_type = T; + using init_type = T; + using constant_iterators = std::true_type; + + template + static void construct(Allocator* alloc, slot_type* slot, Args&&... args) { + phmap::allocator_traits::construct(*alloc, slot, + std::forward(args)...); + } + + template + static void destroy(Allocator* alloc, slot_type* slot) { + phmap::allocator_traits::destroy(*alloc, slot); + } + + template + static void transfer(Allocator* alloc, slot_type* new_slot, slot_type* old_slot) { + construct(alloc, new_slot, std::move(*old_slot)); + destroy(alloc, old_slot); + } + + static T& element(slot_type* slot) { return *slot; } + + template + static decltype(phmap::priv::DecomposeValue(std::declval(), std::declval()...)) + apply(F&& f, Args&&... args) { + return phmap::priv::DecomposeValue(std::forward(f), std::forward(args)...); + } + + static size_t space_used(const T*) { return 0; } +}; + +// -------------------------------------------------------------------------- +// -------------------------------------------------------------------------- +template +struct FlatHashMapPolicy { + using slot_policy = priv::map_slot_policy; + using slot_type = typename slot_policy::slot_type; + using key_type = K; + using mapped_type = V; + using init_type = std::pair; + + template + static void construct(Allocator* alloc, slot_type* slot, Args&&... args) { + slot_policy::construct(alloc, slot, std::forward(args)...); + } + + template + static void destroy(Allocator* alloc, slot_type* slot) { + slot_policy::destroy(alloc, slot); + } + + template + static void transfer(Allocator* alloc, slot_type* new_slot, slot_type* old_slot) { + slot_policy::transfer(alloc, new_slot, old_slot); + } + + template + static decltype(phmap::priv::DecomposePair(std::declval(), std::declval()...)) + apply(F&& f, Args&&... args) { + return phmap::priv::DecomposePair(std::forward(f), std::forward(args)...); + } + + static size_t space_used(const slot_type*) { return 0; } + + static std::pair& element(slot_type* slot) { return slot->value; } + + static V& value(std::pair* kv) { return kv->second; } + static const V& value(const std::pair* kv) { return kv->second; } +}; + +template +struct node_hash_policy { + static_assert(std::is_lvalue_reference::value, ""); + + using slot_type = + typename std::remove_cv::type>::type*; + + template + static void construct(Alloc* alloc, slot_type* slot, Args&&... args) { + *slot = Policy::new_element(alloc, std::forward(args)...); + } + + template + static void destroy(Alloc* alloc, slot_type* slot) { + Policy::delete_element(alloc, *slot); + } + + template + static void transfer(Alloc*, slot_type* new_slot, slot_type* old_slot) { + *new_slot = *old_slot; + } + + static size_t space_used(const slot_type* slot) { + if (slot == nullptr) return Policy::element_space_used(nullptr); + return Policy::element_space_used(*slot); + } + + static Reference element(slot_type* slot) { return **slot; } + + template + static auto value(T* elem) -> decltype(P::value(elem)) { + return P::value(elem); + } + + template + static auto apply(Ts&&... ts) -> decltype(P::apply(std::forward(ts)...)) { + return P::apply(std::forward(ts)...); + } +}; + +// -------------------------------------------------------------------------- +// -------------------------------------------------------------------------- +template +struct NodeHashSetPolicy : phmap::priv::node_hash_policy> { + using key_type = T; + using init_type = T; + using constant_iterators = std::true_type; + + template + static T* new_element(Allocator* alloc, Args&&... args) { + using ValueAlloc = + typename phmap::allocator_traits::template rebind_alloc; + ValueAlloc value_alloc(*alloc); + T* res = phmap::allocator_traits::allocate(value_alloc, 1); + phmap::allocator_traits::construct(value_alloc, res, + std::forward(args)...); + return res; + } + + template + static void delete_element(Allocator* alloc, T* elem) { + using ValueAlloc = + typename phmap::allocator_traits::template rebind_alloc; + ValueAlloc value_alloc(*alloc); + phmap::allocator_traits::destroy(value_alloc, elem); + phmap::allocator_traits::deallocate(value_alloc, elem, 1); + } + + template + static decltype(phmap::priv::DecomposeValue(std::declval(), std::declval()...)) + apply(F&& f, Args&&... args) { + return phmap::priv::DecomposeValue(std::forward(f), std::forward(args)...); + } + + static size_t element_space_used(const T*) { return sizeof(T); } +}; + +// -------------------------------------------------------------------------- +// -------------------------------------------------------------------------- +template +class NodeHashMapPolicy + : public phmap::priv::node_hash_policy&, + NodeHashMapPolicy> { + using value_type = std::pair; + + public: + using key_type = Key; + using mapped_type = Value; + using init_type = std::pair; + + template + static value_type* new_element(Allocator* alloc, Args&&... args) { + using PairAlloc = + typename phmap::allocator_traits::template rebind_alloc; + PairAlloc pair_alloc(*alloc); + value_type* res = phmap::allocator_traits::allocate(pair_alloc, 1); + phmap::allocator_traits::construct(pair_alloc, res, + std::forward(args)...); + return res; + } + + template + static void delete_element(Allocator* alloc, value_type* pair) { + using PairAlloc = + typename phmap::allocator_traits::template rebind_alloc; + PairAlloc pair_alloc(*alloc); + phmap::allocator_traits::destroy(pair_alloc, pair); + phmap::allocator_traits::deallocate(pair_alloc, pair, 1); + } + + template + static decltype(phmap::priv::DecomposePair(std::declval(), std::declval()...)) + apply(F&& f, Args&&... args) { + return phmap::priv::DecomposePair(std::forward(f), std::forward(args)...); + } + + static size_t element_space_used(const value_type*) { return sizeof(value_type); } + + static Value& value(value_type* elem) { return elem->second; } + static const Value& value(const value_type* elem) { return elem->second; } +}; + +// -------------------------------------------------------------------------- +// hash_default +// -------------------------------------------------------------------------- + +#if PHMAP_HAVE_STD_STRING_VIEW + +// support char16_t wchar_t .... +template +struct StringHashT { + using is_transparent = void; + + size_t operator()(std::basic_string_view v) const { + std::string_view bv{reinterpret_cast(v.data()), + v.size() * sizeof(CharT)}; + return std::hash()(bv); + } +}; + +// Supports heterogeneous lookup for basic_string-like elements. +template +struct StringHashEqT { + using Hash = StringHashT; + + struct Eq { + using is_transparent = void; + + bool operator()(std::basic_string_view lhs, + std::basic_string_view rhs) const { + return lhs == rhs; + } + }; +}; + +template <> +struct HashEq : StringHashEqT {}; + +template <> +struct HashEq : StringHashEqT {}; + +// char16_t +template <> +struct HashEq : StringHashEqT {}; + +template <> +struct HashEq : StringHashEqT {}; + +// wchar_t +template <> +struct HashEq : StringHashEqT {}; + +template <> +struct HashEq : StringHashEqT {}; + +#endif + +// Supports heterogeneous lookup for pointers and smart pointers. +// ------------------------------------------------------------- +template +struct HashEq { + struct Hash { + using is_transparent = void; + template + size_t operator()(const U& ptr) const { + return phmap::Hash{}(HashEq::ToPtr(ptr)); + } + }; + + struct Eq { + using is_transparent = void; + template + bool operator()(const A& a, const B& b) const { + return HashEq::ToPtr(a) == HashEq::ToPtr(b); + } + }; + + private: + static const T* ToPtr(const T* ptr) { return ptr; } + + template + static const T* ToPtr(const std::unique_ptr& ptr) { + return ptr.get(); + } + + template + static const T* ToPtr(const std::shared_ptr& ptr) { + return ptr.get(); + } +}; + +template +struct HashEq> : HashEq {}; + +template +struct HashEq> : HashEq {}; + +namespace hashtable_debug_internal { + +// -------------------------------------------------------------------------- +// -------------------------------------------------------------------------- +template +struct HashtableDebugAccess> { + using Traits = typename Set::PolicyTraits; + using Slot = typename Traits::slot_type; + + static size_t GetNumProbes(const Set& set, const typename Set::key_type& key) { + size_t num_probes = 0; + size_t hashval = set.hash(key); + auto seq = set.probe(hashval); + while (true) { + priv::Group g{set.ctrl_ + seq.offset()}; + for (int i : g.Match(priv::H2(hashval))) { + if (Traits::apply( + typename Set::template EqualElement{key, + set.eq_ref()}, + Traits::element(set.slots_ + seq.offset((size_t)i)))) + return num_probes; + ++num_probes; + } + if (g.MatchEmpty()) return num_probes; + seq.next(); + ++num_probes; + } + } + + static size_t AllocatedByteSize(const Set& c) { + size_t capacity = c.capacity_; + if (capacity == 0) return 0; + auto layout = Set::MakeLayout(capacity); + size_t m = layout.AllocSize(); + + size_t per_slot = Traits::space_used(static_cast(nullptr)); + if (per_slot != ~size_t{}) { + m += per_slot * c.size(); + } else { + for (size_t i = 0; i != capacity; ++i) { + if (priv::IsFull(c.ctrl_[i])) { + m += Traits::space_used(c.slots_ + i); + } + } + } + return m; + } + + static size_t LowerBoundAllocatedByteSize(size_t size) { + size_t capacity = GrowthToLowerboundCapacity(size); + if (capacity == 0) return 0; + auto layout = Set::MakeLayout(NormalizeCapacity(capacity)); + size_t m = layout.AllocSize(); + size_t per_slot = Traits::space_used(static_cast(nullptr)); + if (per_slot != ~size_t{}) { + m += per_slot * size; + } + return m; + } +}; + +} // namespace hashtable_debug_internal +} // namespace priv + +// ----------------------------------------------------------------------------- +// phmap::flat_hash_set +// ----------------------------------------------------------------------------- +// An `phmap::flat_hash_set` is an unordered associative container which has +// been optimized for both speed and memory footprint in most common use cases. +// Its interface is similar to that of `std::unordered_set` with the +// following notable differences: +// +// * Supports heterogeneous lookup, through `find()`, `operator[]()` and +// `insert()`, provided that the set is provided a compatible heterogeneous +// hashing function and equality operator. +// * Invalidates any references and pointers to elements within the table after +// `rehash()`. +// * Contains a `capacity()` member function indicating the number of element +// slots (open, deleted, and empty) within the hash set. +// * Returns `void` from the `_erase(iterator)` overload. +// ----------------------------------------------------------------------------- +template // default values in phmap_fwd_decl.h +class flat_hash_set : public phmap::priv::raw_hash_set, + Hash, Eq, Alloc> { + using Base = typename flat_hash_set::raw_hash_set; + + public: + flat_hash_set() {} +#ifdef __INTEL_COMPILER + using Base::raw_hash_set; +#else + using Base::Base; +#endif + using Base::begin; + using Base::bucket_count; + using Base::capacity; + using Base::cbegin; + using Base::cend; + using Base::clear; // may shrink - To avoid shrinking `erase(begin(), end())` + using Base::contains; + using Base::count; + using Base::emplace; + using Base::emplace_hint; + using Base::empty; + using Base::end; + using Base::equal_range; + using Base::erase; + using Base::extract; + using Base::find; + using Base::get_allocator; + using Base::hash; + using Base::hash_function; + using Base::insert; + using Base::key_eq; + using Base::load_factor; + using Base::max_load_factor; + using Base::max_size; + using Base::merge; + using Base::rehash; + using Base::reserve; + using Base::size; + using Base::swap; +}; + +// ----------------------------------------------------------------------------- +// phmap::flat_hash_map +// ----------------------------------------------------------------------------- +// +// An `phmap::flat_hash_map` is an unordered associative container which +// has been optimized for both speed and memory footprint in most common use +// cases. Its interface is similar to that of `std::unordered_map` with +// the following notable differences: +// +// * Supports heterogeneous lookup, through `find()`, `operator[]()` and +// `insert()`, provided that the map is provided a compatible heterogeneous +// hashing function and equality operator. +// * Invalidates any references and pointers to elements within the table after +// `rehash()`. +// * Contains a `capacity()` member function indicating the number of element +// slots (open, deleted, and empty) within the hash map. +// * Returns `void` from the `_erase(iterator)` overload. +// ----------------------------------------------------------------------------- +template // default values in phmap_fwd_decl.h +class flat_hash_map + : public phmap::priv::raw_hash_map, Hash, Eq, + Alloc> { + using Base = typename flat_hash_map::raw_hash_map; + + public: + flat_hash_map() {} +#ifdef __INTEL_COMPILER + using Base::raw_hash_map; +#else + using Base::Base; +#endif + using Base::at; + using Base::begin; + using Base::capacity; + using Base::cbegin; + using Base::cend; + using Base::clear; + using Base::contains; + using Base::count; + using Base::emplace; + using Base::emplace_hint; + using Base::empty; + using Base::end; + using Base::equal_range; + using Base::erase; + using Base::extract; + using Base::find; + using Base::insert; + using Base::insert_or_assign; + using Base::max_size; + using Base::merge; + using Base::rehash; + using Base::reserve; + using Base::size; + using Base::swap; + using Base::try_emplace; + using Base::operator[]; + using Base::bucket_count; + using Base::get_allocator; + using Base::hash; + using Base::hash_function; + using Base::key_eq; + using Base::load_factor; + using Base::max_load_factor; +}; + +// ----------------------------------------------------------------------------- +// phmap::node_hash_set +// ----------------------------------------------------------------------------- +// An `phmap::node_hash_set` is an unordered associative container which +// has been optimized for both speed and memory footprint in most common use +// cases. Its interface is similar to that of `std::unordered_set` with the +// following notable differences: +// +// * Supports heterogeneous lookup, through `find()`, `operator[]()` and +// `insert()`, provided that the map is provided a compatible heterogeneous +// hashing function and equality operator. +// * Contains a `capacity()` member function indicating the number of element +// slots (open, deleted, and empty) within the hash set. +// * Returns `void` from the `erase(iterator)` overload. +// ----------------------------------------------------------------------------- +template // default values in phmap_fwd_decl.h +class node_hash_set : public phmap::priv::raw_hash_set, + Hash, Eq, Alloc> { + using Base = typename node_hash_set::raw_hash_set; + + public: + node_hash_set() {} +#ifdef __INTEL_COMPILER + using Base::raw_hash_set; +#else + using Base::Base; +#endif + using Base::begin; + using Base::bucket_count; + using Base::capacity; + using Base::cbegin; + using Base::cend; + using Base::clear; + using Base::contains; + using Base::count; + using Base::emplace; + using Base::emplace_hint; + using Base::emplace_hint_with_hash; + using Base::emplace_with_hash; + using Base::empty; + using Base::end; + using Base::equal_range; + using Base::erase; + using Base::extract; + using Base::find; + using Base::get_allocator; + using Base::hash; + using Base::hash_function; + using Base::insert; + using Base::key_eq; + using Base::load_factor; + using Base::max_load_factor; + using Base::max_size; + using Base::merge; + using Base::rehash; + using Base::reserve; + using Base::size; + using Base::swap; + typename Base::hasher hash_funct() { return this->hash_function(); } + void resize(typename Base::size_type hint) { this->rehash(hint); } +}; + +// ----------------------------------------------------------------------------- +// phmap::node_hash_map +// ----------------------------------------------------------------------------- +// +// An `phmap::node_hash_map` is an unordered associative container which +// has been optimized for both speed and memory footprint in most common use +// cases. Its interface is similar to that of `std::unordered_map` with +// the following notable differences: +// +// * Supports heterogeneous lookup, through `find()`, `operator[]()` and +// `insert()`, provided that the map is provided a compatible heterogeneous +// hashing function and equality operator. +// * Contains a `capacity()` member function indicating the number of element +// slots (open, deleted, and empty) within the hash map. +// * Returns `void` from the `erase(iterator)` overload. +// ----------------------------------------------------------------------------- +template // default values in phmap_fwd_decl.h +class node_hash_map + : public phmap::priv::raw_hash_map, Hash, + Eq, Alloc> { + using Base = typename node_hash_map::raw_hash_map; + + public: + node_hash_map() {} +#ifdef __INTEL_COMPILER + using Base::raw_hash_map; +#else + using Base::Base; +#endif + using Base::at; + using Base::begin; + using Base::capacity; + using Base::cbegin; + using Base::cend; + using Base::clear; + using Base::contains; + using Base::count; + using Base::emplace; + using Base::emplace_hint; + using Base::empty; + using Base::end; + using Base::equal_range; + using Base::erase; + using Base::extract; + using Base::find; + using Base::insert; + using Base::insert_or_assign; + using Base::max_size; + using Base::merge; + using Base::rehash; + using Base::reserve; + using Base::size; + using Base::swap; + using Base::try_emplace; + using Base::operator[]; + using Base::bucket_count; + using Base::get_allocator; + using Base::hash; + using Base::hash_function; + using Base::key_eq; + using Base::load_factor; + using Base::max_load_factor; + typename Base::hasher hash_funct() { return this->hash_function(); } + void resize(typename Base::size_type hint) { this->rehash(hint); } +}; + +// ----------------------------------------------------------------------------- +// phmap::parallel_flat_hash_set +// ----------------------------------------------------------------------------- +template // default values in phmap_fwd_decl.h +class parallel_flat_hash_set + : public phmap::priv::parallel_hash_set, Hash, Eq, + Alloc> { + using Base = typename parallel_flat_hash_set::parallel_hash_set; + + public: + parallel_flat_hash_set() {} +#ifdef __INTEL_COMPILER + using Base::parallel_hash_set; +#else + using Base::Base; +#endif + using Base::begin; + using Base::bucket_count; + using Base::capacity; + using Base::cbegin; + using Base::cend; + using Base::clear; + using Base::contains; + using Base::count; + using Base::emplace; + using Base::emplace_hint; + using Base::emplace_hint_with_hash; + using Base::emplace_with_hash; + using Base::empty; + using Base::end; + using Base::equal_range; + using Base::erase; + using Base::extract; + using Base::find; + using Base::get_allocator; + using Base::hash; + using Base::hash_function; + using Base::insert; + using Base::key_eq; + using Base::load_factor; + using Base::max_load_factor; + using Base::max_size; + using Base::merge; + using Base::rehash; + using Base::reserve; + using Base::size; + using Base::subcnt; + using Base::subidx; + using Base::swap; +}; + +// ----------------------------------------------------------------------------- +// phmap::parallel_flat_hash_map - default values in phmap_fwd_decl.h +// ----------------------------------------------------------------------------- +template +class parallel_flat_hash_map + : public phmap::priv::parallel_hash_map, Hash, + Eq, Alloc> { + using Base = typename parallel_flat_hash_map::parallel_hash_map; + + public: + parallel_flat_hash_map() {} +#ifdef __INTEL_COMPILER + using Base::parallel_hash_map; +#else + using Base::Base; +#endif + using Base::at; + using Base::begin; + using Base::capacity; + using Base::cbegin; + using Base::cend; + using Base::clear; + using Base::contains; + using Base::count; + using Base::emplace; + using Base::emplace_hint; + using Base::emplace_hint_with_hash; + using Base::emplace_with_hash; + using Base::empty; + using Base::end; + using Base::equal_range; + using Base::erase; + using Base::extract; + using Base::find; + using Base::hash; + using Base::insert; + using Base::insert_or_assign; + using Base::max_size; + using Base::merge; + using Base::rehash; + using Base::reserve; + using Base::size; + using Base::subcnt; + using Base::subidx; + using Base::swap; + using Base::try_emplace; + using Base::try_emplace_with_hash; + using Base::operator[]; + using Base::bucket_count; + using Base::get_allocator; + using Base::hash_function; + using Base::key_eq; + using Base::load_factor; + using Base::max_load_factor; +}; + +// ----------------------------------------------------------------------------- +// phmap::parallel_node_hash_set +// ----------------------------------------------------------------------------- +template +class parallel_node_hash_set + : public phmap::priv::parallel_hash_set, Hash, Eq, + Alloc> { + using Base = typename parallel_node_hash_set::parallel_hash_set; + + public: + parallel_node_hash_set() {} +#ifdef __INTEL_COMPILER + using Base::parallel_hash_set; +#else + using Base::Base; +#endif + using Base::begin; + using Base::bucket_count; + using Base::capacity; + using Base::cbegin; + using Base::cend; + using Base::clear; + using Base::contains; + using Base::count; + using Base::emplace; + using Base::emplace_hint; + using Base::emplace_hint_with_hash; + using Base::emplace_with_hash; + using Base::empty; + using Base::end; + using Base::equal_range; + using Base::erase; + using Base::extract; + using Base::find; + using Base::get_allocator; + using Base::hash; + using Base::hash_function; + using Base::insert; + using Base::key_eq; + using Base::load_factor; + using Base::max_load_factor; + using Base::max_size; + using Base::merge; + using Base::rehash; + using Base::reserve; + using Base::size; + using Base::subcnt; + using Base::subidx; + using Base::swap; + typename Base::hasher hash_funct() { return this->hash_function(); } + void resize(typename Base::size_type hint) { this->rehash(hint); } +}; + +// ----------------------------------------------------------------------------- +// phmap::parallel_node_hash_map +// ----------------------------------------------------------------------------- +template +class parallel_node_hash_map + : public phmap::priv::parallel_hash_map, + Hash, Eq, Alloc> { + using Base = typename parallel_node_hash_map::parallel_hash_map; + + public: + parallel_node_hash_map() {} +#ifdef __INTEL_COMPILER + using Base::parallel_hash_map; +#else + using Base::Base; +#endif + using Base::at; + using Base::begin; + using Base::capacity; + using Base::cbegin; + using Base::cend; + using Base::clear; + using Base::contains; + using Base::count; + using Base::emplace; + using Base::emplace_hint; + using Base::emplace_hint_with_hash; + using Base::emplace_with_hash; + using Base::empty; + using Base::end; + using Base::equal_range; + using Base::erase; + using Base::extract; + using Base::find; + using Base::hash; + using Base::insert; + using Base::insert_or_assign; + using Base::max_size; + using Base::merge; + using Base::rehash; + using Base::reserve; + using Base::size; + using Base::subcnt; + using Base::subidx; + using Base::swap; + using Base::try_emplace; + using Base::try_emplace_with_hash; + using Base::operator[]; + using Base::bucket_count; + using Base::get_allocator; + using Base::hash_function; + using Base::key_eq; + using Base::load_factor; + using Base::max_load_factor; + typename Base::hasher hash_funct() { return this->hash_function(); } + void resize(typename Base::size_type hint) { this->rehash(hint); } +}; + +} // namespace phmap + +#ifdef _MSC_VER +#pragma warning(pop) +#endif + +#endif // phmap_h_guard_ diff --git a/native-sql-engine/cpp/src/third_party/parallel_hashmap/phmap_base.h b/native-sql-engine/cpp/src/third_party/parallel_hashmap/phmap_base.h new file mode 100644 index 000000000..3b3b6b120 --- /dev/null +++ b/native-sql-engine/cpp/src/third_party/parallel_hashmap/phmap_base.h @@ -0,0 +1,4956 @@ +#if !defined(phmap_base_h_guard_) +#define phmap_base_h_guard_ + +// --------------------------------------------------------------------------- +// Copyright (c) 2019, Gregory Popovitch - greg7mdp@gmail.com +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// https://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +// +// Includes work from abseil-cpp (https://github.com/abseil/abseil-cpp) +// with modifications. +// +// Copyright 2018 The Abseil Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// https://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +// --------------------------------------------------------------------------- + +#include +#include +#include +#include +#include +#include +#include +#include // for std::lock +#include +#include +#include +#include + +#include "phmap_config.h" + +#ifdef PHMAP_HAVE_SHARED_MUTEX +#include // after "phmap_config.h" +#endif + +#ifdef _MSC_VER +#pragma warning(push) +#pragma warning(disable : 4514) // unreferenced inline function has been removed +#pragma warning(disable : 4582) // constructor is not implicitly called +#pragma warning(disable : 4625) // copy constructor was implicitly defined as deleted +#pragma warning(disable : 4626) // assignment operator was implicitly defined as deleted +#pragma warning(disable : 4710) // function not inlined +#pragma warning(disable : 4711) // selected for automatic inline expansion +#pragma warning(disable : 4820) // '6' bytes padding added after data member +#endif // _MSC_VER + +namespace phmap { + +template +using Allocator = typename std::allocator; + +template +using Pair = typename std::pair; + +template +struct EqualTo { + inline bool operator()(const T& a, const T& b) const { + return std::equal_to()(a, b); + } +}; + +template +struct Less { + inline bool operator()(const T& a, const T& b) const { return std::less()(a, b); } +}; + +namespace type_traits_internal { + +template +struct VoidTImpl { + using type = void; +}; + +// This trick to retrieve a default alignment is necessary for our +// implementation of aligned_storage_t to be consistent with any implementation +// of std::aligned_storage. +// --------------------------------------------------------------------------- +template > +struct default_alignment_of_aligned_storage; + +template +struct default_alignment_of_aligned_storage> { + static constexpr size_t value = Align; +}; + +// NOTE: The `is_detected` family of templates here differ from the library +// fundamentals specification in that for library fundamentals, `Op` is +// evaluated as soon as the type `is_detected` undergoes +// substitution, regardless of whether or not the `::value` is accessed. That +// is inconsistent with all other standard traits and prevents lazy evaluation +// in larger contexts (such as if the `is_detected` check is a trailing argument +// of a `conjunction`. This implementation opts to instead be lazy in the same +// way that the standard traits are (this "defect" of the detection idiom +// specifications has been reported). +// --------------------------------------------------------------------------- + +template class Op, class... Args> +struct is_detected_impl { + using type = std::false_type; +}; + +template