Move instead of deep-copy ordered_json values when an object grows

ordered_map keeps its elements in a std::vector<std::pair<const Key, T>>.
With a std::string key, that pair is not nothrow move constructible (the
const key has to be copied), so std::vector copies every element when it
reallocates. For ordered_json, this deep-copies every member value an
object already holds, including whole nested subtrees, on each growth
step.

Grow the storage in ordered_map instead, copying the keys and moving the
values. This happens in two phases, so the strong exception guarantee is
kept without try/catch. The first phase may throw, but only touches a
temporary buffer: it copies the keys, value-initializes the values, and
constructs the new element. The second phase moves the values (noexcept)
and swaps the buffers. Because the new element is constructed before any
value is moved, arguments that refer to elements of the container stay
valid, as with std::vector. Types that cannot take this path keep the
std::vector behavior.

Parsing into ordered_json (ParseStringOrdered, Apple M1 Max, clang -O3):
twitter 3.20 -> 1.70 ms, citm_catalog 7.73 -> 3.67 ms, jeopardy 219 ->
177 ms, canada unchanged. The number of allocations for twitter and
citm_catalog drops by two thirds.

Also add ParseStringOrdered rows to the benchmarks, and document the
growth behavior and the exception safety of ordered_map.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
This commit is contained in:
Niels Lohmann
2026-09-28 18:48:05 +02:00
parent fc03b9912e
commit 4b2d2244c3
6 changed files with 540 additions and 10 deletions
+65 -5
View File
@@ -8,13 +8,15 @@
#pragma once
#include <algorithm> // max, min
#include <functional> // equal_to, less
#include <initializer_list> // initializer_list
#include <iterator> // input_iterator_tag, iterator_traits
#include <memory> // allocator
#include <stdexcept> // for out_of_range
#include <type_traits> // enable_if, is_convertible
#include <utility> // pair
#include <tuple> // forward_as_tuple
#include <type_traits> // enable_if, integral_constant, is_convertible, is_nothrow_move_constructible
#include <utility> // forward, move, pair, piecewise_construct
#include <vector> // vector
#include <nlohmann/detail/macro_scope.hpp>
@@ -79,7 +81,7 @@ template <class Key, class T, class IgnoredLess = std::less<Key>,
return {it, false};
}
}
Container::emplace_back(key, std::forward<T>(t));
append(key, std::forward<T>(t));
return {std::prev(this->end()), true};
}
@@ -94,7 +96,7 @@ template <class Key, class T, class IgnoredLess = std::less<Key>,
return {it, false};
}
}
Container::emplace_back(std::forward<KeyType>(key), std::forward<T>(t));
append(std::forward<KeyType>(key), std::forward<T>(t));
return {std::prev(this->end()), true};
}
@@ -368,7 +370,7 @@ template <class Key, class T, class IgnoredLess = std::less<Key>,
return {it, false};
}
}
Container::push_back(value);
append(value);
return {--this->end(), true};
}
@@ -386,6 +388,64 @@ template <class Key, class T, class IgnoredLess = std::less<Key>,
}
private:
/*!
@brief add an element whose key is not yet contained at the end
A std::vector copies all elements when it grows, because their const keys
make them not nothrow move constructible. For ordered_json, this is a deep
copy of every value. Where the strong exception guarantee can be kept, grow
the storage here instead, copying only the keys and moving the values.
*/
template<typename... Args>
void append(Args&& ... args)
{
// evaluated here rather than at class scope, because T is still
// incomplete when basic_json instantiates its object_t
using move_values = std::integral_constant<bool, detail::conjunction<
detail::negation<std::is_nothrow_move_constructible<value_type>>,
std::is_copy_constructible<key_type>,
detail::is_default_constructible<mapped_type>,
std::is_nothrow_move_assignable<mapped_type>>::value>;
append_impl(move_values{}, std::forward<Args>(args)...);
}
template<typename... Args>
void append_impl(std::true_type /*unused*/, Args&& ... args)
{
if (this->size() < this->capacity())
{
Container::emplace_back(std::forward<Args>(args)...);
return;
}
// 1. May throw, but only changes tmp: copy the keys, value-initialize
// the values, and add the new element. The arguments may refer to
// elements of this container, so they are used before any value is
// moved out of it.
Container tmp(this->get_allocator()); // equal allocators, so swap() is valid
tmp.reserve((std::min)(this->max_size(), (std::max)(size_type{1}, 2 * this->size())));
for (const auto& element : *this)
{
tmp.emplace_back(std::piecewise_construct, std::forward_as_tuple(element.first), std::forward_as_tuple());
}
tmp.emplace_back(std::forward<Args>(args)...);
// 2. Cannot throw: move the values over and adopt the new storage.
auto it = tmp.begin();
for (auto& element : *this)
{
it->second = std::move(element.second);
++it;
}
Container::swap(tmp);
}
template<typename... Args>
void append_impl(std::false_type /*unused*/, Args&& ... args)
{
Container::emplace_back(std::forward<Args>(args)...);
}
JSON_NO_UNIQUE_ADDRESS key_compare m_compare = key_compare();
};