Boost.MultiIndex Release notesBoost.MultiIndex Release notes
indexed_by,
tag, nested typedefs index_specifier_type_list,
index_type_list, iterator_type_list and
const_iterator_type_list) are no longer based on
Boost.MPL but instead they are now
Boost.Mp11 lists.
As a result, Boost.MultiIndex doesn't depend anymore on Boost.MPL.
It is expected that most user code won't be impacted by this
change, but the previous behavior may be restored, however, by
globally defining the macro
BOOST_MULTI_INDEX_ENABLE_MPL_SUPPORT.
Legacy support for Boost.MPL may be eventually deprecated and removed in the future.
composite_key and associated
class templates (composite_key_equal_to, composite_key_compare,
composite_key_hash) have been made variadic (previously the maximum
number of template arguments was limited by BOOST_MULTI_INDEX_LIMIT_COMPOSITE_KEY_SIZE).
This change should be transparent to users, with the exception that
composite_key::key_extractors returns a std::tuple
instead of a boost::tuple (and similarly for the rest of affected
class templates).
insert in non-unique ordered indices to preserve insertion order
of equivalent elements
(issue #68).
std::size_t instead of unsigned long
to save and load the size of a multi_index_container
(unsigned long is smaller than std::size_t in LLP64 data models).
multi_index_container serialization
class version has been
bumped from 3 to 4 to reflect this change.
unsigned long instead of
collection_size_type,
and multi_index_container serialization
class version has been
bumped from 2 to 3 to reflect this change.
Reading old archives may fail for custom archive types giving special treatment to
collection_size_type values: if this is the case, globally define
the macro BOOST_MULTI_INDEX_ENABLE_SERIALIZATION_COMPATIBILITY_V2
to ensure backwards compatibility in reading operations.
count operations in ranked indices from
O(log(n) + count) to O(log(n)).
Contributed by Damian Sawicki.
contains to key-based indices
(issue #35).
merge operations to key-based indices. The functionality goes beyond
the standard specification for (unordered) associative containers in a number of ways,
most notably:
x.merge(y,first,last)
merges only the elements of y within [first,last).
splice for sequenced and random access indices
were destructive, i.e. elements were copy-inserted into the destination and then erased
from the source. Now, splice is based on node transfer much as merge
in key-based indices, and has been similarly extended to accept source indices of any type:
in fact, splice can be regarded as a frontend to the same functionality
provided by merge in key-based indices. For reasons of backwards compatibility,
the destructive behavior of splice has been retained in the case that the
source and destination containers have unequal allocators.
node_type
and the position of the index in its multi_index_container,
(except for hashed indices, where uniqueness/non-uniqueness is also a dependency). This has
implications on the validity of iterators to elements transferred by merge or
splice. This property is a variant of what has been called
SCARY iterators in the C++ standard mailing lists.
SCARYness is currently (August 2021) not mandated for standard containers.
multi_index_container is now
AllocatorAware.
KeyFromValue, Compare, Hash
and Pred objects now selects the appropriate swap
function between std::swap and ADL candidates,
in accordance with standard specifications in [swappable.requirements]
(issue #29).
const_mem_fun and mem_fun for
differently qualified member functions
(issue #24).
noexcept-specified member functions
(issue #24).
size_type and difference_type are now
defined as the allocator's same-named types. This will not make any
difference in the vast majority of cases, but allows for some degree
of control of resulting internal structures via user-defined allocator
types (see
issue #17
for motivation). For the moment being, this change is not
documented in the reference section (i.e., it has semi-official status).
multi_index_container's default constructor is no longer
explicit
(ticket #13518).
operator& are now fully accepted
(ticket #13307).
Thanks to Daniel Frey for his updating
Boost.Operators to help
fix this issue.
std::allocator members deprecated in C++17.
Contributed by Daniela Engert.
modify and modify_key more robust so that
the modified element is erased if the modifier throws or the rollback
functor does not properly restore the element (full discussion at
ticket #12542).
This is technically backwards incompatible; for instance, the following code:
keeps the containerc.modify(c.begin(),[](auto&){throw 0;});
c untouched in former versions of Boost whereas
now c.begin() is erased. Thanks to Jon Kalb for raising the issue.
foreach in this framework.
key_type. Consider for instance
the following code:
In previous versions of the library, thetypedef multi_index_container< std::string, indexed_by< ordered_unique<identity<std::string> > > > multi_t; ... multi_t m=...; m.find("boost"); // passed a const char*, not a std::string
find operation generates
several temporary std::strings (one every time an internal comparison is made).
In sufficiently advanced compilers, this is now avoided so that only one temporary is
created.
std::tuples are available, these can be used for lookup operations
in indices equipped with composite keys. boost::tuples are also supported
for backwards compatibility.
erase(iterator) member function of hashed indices
used to have poor performance under low load conditions due to the requirement
that an iterator to the next element must be returned (see ticket
#4264). In accordance with
the resolution of LWG
issue #579, this problem has been fixed while maintaining the interface of
erase, at the expense of using one more
word of memory per element. In fact, C++ complexity requirements on unordered
associative containers have been improved for hashed indices so that
O(n) but O(ndist),
where ndist is the number of non-equivalent elements in the index.
reserve.noexcept specifications
where appropriate (for compliant compilers).
multi_index_container.BOOST_BIND_NO_PLACEHOLDERS
(ticket #9798).
multi_index_container is now efficiently movable.std::unique_ptr<T>) supported. This includes
insertion of a range [first,last) where the iterators point to a type that is
convertible to that of the element: no copy construction happens in the process.
shrink_to_fit().
composite_key_compare
related with Boost.Ref.
modify_key with
rollback in ordered and
hashed indices when
Modifier and Rollback are different types
(ticket #4130).
save_construct_data and
load_construct_data.
multi_index_container serialization
class version has been
bumped from 1 to 2.
collection_size_type
type instead of the original std::size_t (ticket
#3365).
multi_index_container serialization
class version has been
bumped from 0 to 1.
nth_index_iterator<MultiIndexContainer,N>::type,multi_index_container<...>::nth_index_iterator<N>::type,nth_index_const_iterator<MultiIndexContainer,N>::type,multi_index_container<...>::nth_index_const_iterator<N>::type,index_iterator<MultiIndexContainer,Tag>::type,multi_index_container<...>::index_iterator<Tag>::type,index_const_iterator<MultiIndexContainer,Tag>::type,multi_index_container<...>::index_const_iterator<Tag>::type.nth_index<MultiIndexContainer,N>::type::iterator,multi_index_container<...>::nth_index<N>::type::iterator,nth_index<MultiIndexContainer,N>::type::const_iterator,multi_index_container<...>::nth_index<N>::type::const_iterator,index<MultiIndexContainer,Tag>::type::iterator,multi_index_container<...>::index<Tag>::type::iterator,index<MultiIndexContainer,Tag>::type::const_iterator,multi_index_container<...>::index<Tag>::type::const_iterator.
multi_index_container::multi_index_container(const allocator_type&)
to mimic the equivalent interface in STL sequence containers.
global_fun
predefined key extractor.
iterator_to
facility.
multi_index_containers placeable in shared memory.
modify and modify_key with
rollback, as described in the
tutorial.
cbegin, cend and,
when applicable, crbegin and crend
member functions, in accordance with the latest drafts of the next
revision of the C++ standard.
assign(f,l) and insert(p,f,l) (23.1.1/9): if
f and l are of the same integral type, the
iterator-based overloads of these member functions are avoided:
typedef multi_index_container< int,indexed_by<sequenced<> > > sequenced_container; std::list<int> l(...); sequenced_container c; // iterator-based overload of assign c.assign(l.begin(),l.end()); // The following is equivalent to // c.assign( // static_cast<sequenced_container::size_type>(10),100); // that is, "10" and "100" are not taken to be iterators as // in the previous expression. c.assign(10,100);
range and
equal_range has been improved.
T can handle objects of types
derived from or convertible to T (and
chained pointers
to those). Previously, only objects of the exact type specified (along with
reference_wrappers and chained pointers to them) were accepted.
composite_key_compare
and related classes accept operands not included in tuples as if they were passed
in a tuple of length 1; this allows the user to omit tuple enclosing in
lookup operations involving composite keys when only the first key is provided.
erase(it) and
erase(first,last) now return an iterator to the element
following those being deleted (previously nothing was returned), in
accordance with the C++ Standard Library
Defect
Report 130 and issue 6.19 of TR1
Issues
List.
multi_index_containers and clear
memfuns now perform faster.
Revised November 7th 2025
© Copyright 2003-2025 Joaquín M López Muñoz. Distributed under the Boost Software License, Version 1.0. (See accompanying file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)