fixed_capacity_vector

A dynamically-resizable vector with fixed capacity and embedded storage (revision 1)

Document number: P0843r1.

Date: 2018-02-6.

Project: Programming Language C++, Library Working Group.

Audience: LEWG.

Reply-to: Gonzalo Brito Gadeschi <gonzalo.gadeschi at rwth-aachen dot de>.

Table of contents

1. Introduction

This paper proposes a modernized version of boost::container::static_vector<T,Capacity> [1]. That is, a dynamically-resizable vector with compile-time fixed capacity and contiguous embedded storage in which the elements are stored within the vector object itself.

Its API closely resembles that of std::vector<T, A>. It is a contiguous container with O(1) insertion and removal of elements at the end (non-amortized) and worst case O(size()) insertion and removal otherwise. Like std::vector, the elements are initialized on insertion and destroyed on removal. For trivial value_types, the vector is fully usable inside constexpr functions.

The name fixed_capacity_vector is a temporary placeholder. Better names are discussed in 4.6 Naming.

2. Motivation and Scope

The fixed_capacity_vector container is useful when:

  • memory allocation is not possible, e.g., embedded environments without a free store, where only a stack and the static memory segment are available,
  • memory allocation imposes an unacceptable performance penalty, e.g., with respect to latency,
  • allocation of objects with complex lifetimes in the static-memory segment is required,
  • std::array is not an option, e.g., if non-default constructible objects must be stored,
  • a dynamically-resizable array is required within constexpr functions,
  • the storage location of the fixed_capacity_vector elements is required to be within the fixed_capacity_vector object itself (e.g. to support memcopy for serialization purposes).

3. Existing practice

There are at least 3 widely used implementations of fixed_capacity_vector: Boost.Container [1], EASTL [2], and Folly [3]. The main difference between these is that Boost.Container implements fixed_capacity_vector as a standalone type with its own guarantees, while both EASTL and Folly implement it by adding an extra template parameter to their small_vector types.

A fixed_capacity_vector can also be poorly emulated by using a custom allocator, like for example Howard Hinnant's stack_alloc [4], on top of std::vector.

This proposal shares a similar purpose with P0494R0 [5] and P0597R0: std::constexpr_vector<T> [6]. The main difference is that this proposal closely follows boost::container::static_vector [1] and proposes to standardize existing practice. A prototype implementation of this proposal for standardization purposes is provided here: http://github.com/gnzlbg/fixed_capacity_vector.

4. Design Decisions

The most fundamental question that must be answered is:

Should fixed_capacity_vector be a standalone type or a special case of some other type?

The EASTL [2] and Folly [3] special case small_vector, e.g., using a 4th template parameter, to make it become a fixed_capacity_vector. The paper P0639R0: Changing attack vector of the constexpr_vector [7] proposes improving the Allocator concepts to allow fixed_capacity_vector, among others, to be implemented as a special case of std::vector with a custom allocator.

Both approaches run into the same fundamental issue: fixed_capacity_vector methods are identically-named to those of std::vector yet they have subtly different effects, exception-safety, iterator invalidation, and complexity guarantees.

This proposal follows boost::container::static_vector<T,Capacity> [1] closely and specifies the semantics that fixed_capacity_vector ought to have as a standalone type. As a library component this delivers immediate value.

I hope that having the concise semantics of this type specified will also be helpful for those that want to generalize the Allocator interface to allow implementing fixed_capacity_vector as a std::vector with a custom allocator.

4.1 Storage/Memory Layout

The container models ContiguousContainer. The elements of the fixed_capacity_vector are contiguously stored and properly aligned within the fixed_capacity_vector object itself. The exact location of the contiguous elements within the fixed_capacity_vector is not specified. If the Capacity is zero the container has zero size:

static_assert(is_empty_v<fixed_capacity_vector<T, 0>>); // for all T

This optimization is easily implementable, enables the EBO, and felt right.

4.2 Move semantics

The move semantics of fixed_capacity_vector<T, Capacity> are equal to those of std::array<T, Size>. That is, after

fixed_capacity_vector a(10);
fixed_capacity_vector b(std::move(a));

the elements of a have been moved element-wise into b, the elements of a are left in an initialized but unspecified state (have been moved from state), the size of a is not altered, and a.size() == b.size().

Note: this behavior differs from std::vector<T, Allocator>, in particular for the similar case in which std::allocator_traits<Allocator>::propagate_on_container_move_assignment is false. In this situation the state of std::vector is initialized but unspecified.

4.3 constexpr support

The API of fixed_capacity_vector<T, Capacity> is constexpr. If is_trivially_copyable_v<T> && is_default_constructible_v<T> is true, fixed_capacity_vectors can be seamlessly used from constexpr code. This allows using fixed_capacity_vector as a constexpr_vector to, e.g., implement other constexpr containers.

The implementation cost of this is small: the prototye implementation specializes the storage for trivial types to use a C array with value-initialized elements and a defaulted destructor.

This changes the algorithmic complexity of fixed_capacity_vector constructors for trivial-types from "Linear in N" to "Constant in Capacity. When the value-initialization takes place at run-time, this difference in behavior might be signficiant: fixed_capacity_vector<non_trivial_type, 38721943228473>(4) will only initialize 4 elements but fixed_capacity_vector<trivial_type, 38721943228473>(4) must value-initialize the 38721943228473 - 4 excess elements to be a valid constexpr constructor.

Very large fixed_capacity_vector's are not the target use case of this container class and will have, in general, worse performance than, e.g., std::vector (e.g. due to moves being O(N)).

Future improvements to constexpr (e.g. being able to properly use std::aligned_storage in constexpr contexts) allow improving the performance of fixed_capacity_vector in a backwards compatible way.

4.4 Exception Safety

The only operations that can actually fail within fixed_capacity_vector<value_type, Capacity> are:

  1. value_type's constructors/assignment/destructors/swap can potentially throw,

  2. Mutating operations exceeding the capacity (push_back, insert, emplace, fixed_capacity_vector(value_type, size), fixed_capacity_vector(begin, end)...).

  3. Out-of-bounds unchecked access:

    • 3.1 front/back/pop_back when empty, operator[] (unchecked random-access).
    • 3.2 at (checked random-access) which can throw out_of_range exception.

When value_type's operations are invoked, the exception safety guarantees of fixed_capacity_vector depend on whether these operations can throw. This is detected with noexcept.

Since its Capacity is fixed at compile-time, fixed_capacity_vector never dynamically allocates memory, the answer to the following question determines the exception safety for all other operations:

What should fixed_capacity_vector do when its Capacity is exceeded?

Two main answers were explored in the prototype implementation:

  1. Throw an exception.
  2. Make this a precondition violation.

Throwing an exception is appealing because it makes the interface slightly more similar to that of std::vector. However, which exception should be thrown? It cannot be std::bad_alloc, because nothing is being allocated. It could throw either std::out_of_bounds or std::logic_error but in any case the interface does not end up being equal to that of std::vector.

The alternative is to make not exceeding the capacity a precondition on the fixed_capacity_vector's methods. This approach allows implementations to provide good run-time diagnostics if they so desired, e.g., on debug builds by means of an assertion, and makes implementation that avoid run-time checks conforming as well. Since the mutating methods have a precondition, they have narrow contracts, and are not conditionally noexcept.

This proposal chooses this path and makes exceeding the fixed_capacity_vector's capacity a precondition violation that results in undefined behavior. Throwing checked_xxx methods can be provided in a backwards compatible way.

4.5 Iterator invalidation

The iterator invalidation rules are different than those for std::vector, since:

  • moving a fixed_capacity_vector invalidates all iterators,
  • swapping two fixed_capacity_vectors invalidates all iterators, and
  • inserting elements at the end of an fixed_capacity_vector never invalidates iterators.

The following functions can potentially invalidate the iterators of fixed_capacity_vectors: resize(n), resize(n, v), pop_back, erase, and swap.

4.6 Naming

The following names have been considered:

  • array_vector: a vector whose storage is backed up by a raw array.
  • bounded_vector: clearly indicates that the the size of the vector is bounded.
  • fixed_capacity_vector: clearly indicates that the capacity is fixed.
  • static_capacity_vector: clearly indicates that the capacity is fixed at compile time (static is overloaded).
  • static_vector (Boost.Container): due to "static" / compile-time allocation of the elements. The term static is, however, overloaded in C++ (e.g. static memory?).
  • embedded_vector<T, Capacity>: since the elements are "embedded" within the fixed_capacity_vector object itself. Sadly, the name embedded is overloaded, e.g., embedded systems.
  • inline_vector: the elements are stored "inline" within the fixed_capacity_vector object itself. The term inline is, however, already overloaded in C++ (e.g. inline functions => ODR, inlining, inline variables).
  • stack_vector: to denote that the elements can be stored on the stack. Is confusing since the elements can be on the stack, the heap, or the static memory segment. It also has a resemblance with std::stack.

@LEWG/LWG: Could you have a poll about these names and let me know which one you prefer?

4.7 Future extensions

The following extensions could be added in a backwards compatible way:

  • utilities for hiding the concrete type of vector-like containers (e.g. any_vector_ref<T>/any_vector<T>).

  • default-initialization of the vector elements (as opposed to value-initialization): e.g. by using a tagged constructor with a default_initialized_t tag.

  • tagged-constructor of the form fixed_capacity_vector(with_size_t, std::size_t N, T const& t = T()) to avoid the complexity introduced by initializer lists and braced initialization:

using vec_t = fixed_capacity_vector<std::size_t, Capacity>;
vec_t v0(2);  // two-elements: 0, 0
vec_t v1{2};  // one-element: 2
vec_t v2(2, 1);  // two-elements: 1, 1
vec_t v3{2, 1};  // two-elements: 2, 1

All these extensions are generally useful and not part of this proposal.

5. Technical specification

Note to editor: This enhancement is a pure header-only addition to the C++ standard library as the <fixed_capacity_vector> header. It belongs in the "Sequence containers" (\ref{sequences}) part of the "Containers library" (\ref{containers}) as "Class template fixed_capacity_vector".

5. Class template fixed_capacity_vector

5.1 Class template fixed_capacity_vector overview

    1. A fixed_capacity_vector is a contiguous container that supports constant time insert and erase operations at the end; insert and erase in the middle take linear time. Its capacity is part of its type and its elements are stored within the fixed_capacity_vector object itself, meaning that that if v is a fixed_capacity_vector<T, N> then it obeys the identity &v[n] == &v[0] + n for all 0 <= n <= v.size().
    1. A fixed_capacity_vector satisfies all of the requirements of a container and of a reversible container (given in two tables in \ref{container.requirements}), of a sequence container, including the optional sequence container requirements (\ref{sequence.reqmts}), and of a contiguous container (\ref{container.requirements.general}). The exceptions are the push_front, pop_front, and emplace_front member functions, which are not provided, and swap, which has linear complexity instead of constant complexity. Descriptions are provided here only for operations on fixed_capacity_vector that are not described in one of these tables or for operations where there is additional semantic information.

Note: An incomplete type T cannot be used to instantiate a fixed_capacity_vector.

namespace std {

template <typename T, size_t N>
class fixed_capacity_vector {
public:
// types:
using value_type = T;
using pointer = T*;
using const_pointer = const T*; 
using reference = value_type&;
using const_reference = const value_type&;
using size_type =  size_t;
using difference_type = make_signed_t<size_type>;
using iterator = implementation-defined;  // see [container.requirements]
using const_iterator = implementation-defined; // see [container.requirements]
using reverse_iterator = std::reverse_iterator<iterator>;
using const_reverse_iterator = std::reverse_iterator<const_iterator>;

// 5.2, copy/move construction:
constexpr fixed_capacity_vector() noexcept;
constexpr explicit fixed_capacity_vector(size_type n);
constexpr fixed_capacity_vector(size_type n, const value_type& value);
template <class InputIterator>
constexpr fixed_capacity_vector(InputIterator first, InputIterator last);
constexpr fixed_capacity_vector(const fixed_capacity_vector& other)
  noexcept(is_nothrow_copy_constructible_v<value_type>);
constexpr fixed_capacity_vector(fixed_capacity_vector&& other)
  noexcept(is_nothrow_move_constructible_v<value_type>);
constexpr fixed_capacity_vector(initializer_list<value_type> il);

// 5.3, copy/move assignment:
constexpr fixed_capacity_vector& operator=(const fixed_capacity_vector& other)
  noexcept(is_nothrow_copy_assignable_v<value_type>);
constexpr fixed_capacity_vector& operator=(fixed_capacity_vector&& other);
  noexcept(is_nothrow_move_assignable_v<value_type>);
template <class InputIterator>
constexpr void assign(InputIterator first, InputIterator last);
constexpr void assign(size_type n, const value_type& u);
constexpr void assign(initializer_list<value_type> il);

// 5.4, destruction
~fixed_capacity_vector();

// iterators
constexpr iterator               begin()         noexcept;
constexpr const_iterator         begin()   const noexcept;
constexpr iterator               end()           noexcept;
constexpr const_iterator         end()     const noexcept;
constexpr reverse_iterator       rbegin()        noexcept;
constexpr const_reverse_iterator rbegin()  const noexcept;
constexpr reverse_iterator       rend()          noexcept;
constexpr const_reverse_iterator rend()    const noexcept;
constexpr const_iterator         cbegin()        noexcept;
constexpr const_iterator         cend()    const noexcept;
constexpr const_reverse_iterator crbegin()       noexcept;
constexpr const_reverse_iterator crend()   const noexcept;

// 5.5, size/capacity:
constexpr bool empty() const noexcept;
constexpr size_type size() const noexcept;
static constexpr size_type max_size() noexcept;
static constexpr size_type capacity() noexcept;
constexpr void resize(size_type sz);
constexpr void resize(size_type sz, const value_type& c);

// 5.6, element and data access:
constexpr reference       operator[](size_type n); 
constexpr const_reference operator[](size_type n) const;
constexpr const_reference at(size_type n) const;
constexpr reference       at(size_type n);
constexpr reference       front();
constexpr const_reference front() const;
constexpr reference       back();
constexpr const_reference back() const;
constexpr       T* data()       noexcept;
constexpr const T* data() const noexcept;

// 5.7, modifiers:
constexpr iterator insert(const_iterator position, const value_type& x);
constexpr iterator insert(const_iterator position, value_type&& x);
constexpr iterator insert(const_iterator position, size_type n, const value_type& x);
template <class InputIterator>
  constexpr iterator insert(const_iterator position, InputIterator first, InputIterator last);
constexpr iterator insert(const_iterator position, initializer_list<value_type> il);

template <class... Args>
  constexpr iterator emplace(const_iterator position, Args&&... args);
template <class... Args>
  constexpr reference emplace_back(Args&&... args);
constexpr void push_back(const value_type& x);
constexpr void push_back(value_type&& x);

constexpr void pop_back();
constexpr iterator erase(const_iterator position);
constexpr iterator erase(const_iterator first, const_iterator last);

constexpr void clear() noexcept;

constexpr void swap(fixed_capacity_vector& x)
  noexcept(is_nothrow_swappable_v<value_type>);
};

template <typename T, size_t N>
constexpr bool operator==(const fixed_capacity_vector<T, N>& a, const fixed_capacity_vector<T, N>& b);
template <typename T, size_t N>
constexpr bool operator!=(const fixed_capacity_vector<T, N>& a, const fixed_capacity_vector<T, N>& b);
template <typename T, size_t N>
constexpr bool operator<(const fixed_capacity_vector<T, N>& a, const fixed_capacity_vector<T, N>& b);
template <typename T, size_t N>
constexpr bool operator<=(const fixed_capacity_vector<T, N>& a, const fixed_capacity_vector<T, N>& b);
template <typename T, size_t N>
constexpr bool operator>(const fixed_capacity_vector<T, N>& a, const fixed_capacity_vector<T, N>& b);
template <typename T, size_t N>
constexpr bool operator>=(const fixed_capacity_vector<T, N>& a, const fixed_capacity_vector<T, N>& b);

// 5.8, specialized algorithms:
template <typename T, size_t N>
constexpr void swap(fixed_capacity_vector<T, N>& x, fixed_capacity_vector<T, N>& y)
  noexcept(noexcept(x.swap(y)));
  
}  // namespace std

5.2 fixed_capacity_vector constructors

constexpr fixed_capacity_vector() noexcept;

  • Effects: constructs an empty fixed_capacity_vector.

  • Complexity: Constant.

constexpr explicit fixed_capacity_vector(size_type n);

  • Effects: constructs a fixed_capacity_vector with n default-inserted elements.

  • Requires: value_type shall be DefaultInsertable into *this and n <= capacity().

  • Complexity: Linear in n.

constexpr fixed_capacity_vector(size_type n, const value_type& value);

  • Effects: Constructs a fixed_capacity_vector with n copies of value.

  • Requires: value_type shall be CopyInsertable into *this and n <= capacity().

  • Complexity: Linear in n.

template <class InputIterator>
constexpr fixed_capacity_vector(InputIterator first, InputIterator last);

  • Effects: Constructs a fixed_capacity_vector equal to the range [first, last)

  • Requires: value_type shall be EmplaceConstructible into *this from *first and distance(first, last) <= capacity().

  • Complexity: Initializes distance(first, last) value_types.

5.3 Destruction

~fixed_capacity_vector();

Effects: Destroys the contents of the fixed_capacity_vector.

Remarks: This destructor shall be trivial if is_trivially_copyable_v<T> && is_default_constructible_v<T> is true.

5.4 Size and capacity

static constexpr size_type capacity() noexcept;
static constexpr size_type max_size() noexcept;
  • Effects: equivalent to return N;.

```c++
constexpr void resize(size_type sz);
constexpr void resize(size_type sz, const value_type& c);

  • Requires: sz shall be less than or equal to N. T shall be:

    • DefaultInsertable into *this for the first overload, or
    • CopyInsertable into *this for the second overload.

  • Effects: If sz < size(), erases the last size() - sz elements from the sequence. Otherwise, appends sz - size() elements to the sequence which are:

    • value-initialized for the first overload, or
    • copies of c for the second overload.

Remarks: These functions shall be constexpr if is_trivially_copyable_v<value_type> && is_default_constructible_v<value_type> is true.

5.5 Element and data access

constexpr       T* data()       noexcept;
constexpr const T* data() const noexcept;

  • Returns: A pointer such that [data(), data() + size()) is a valid range. For a non-empty fixed_capacity_vector, data() == addressof(front()).

  • Complexity: Constant time.

5.6 Modifiers

constexpr iterator insert(const_iterator position, const value_type& x);
constexpr iterator insert(const_iterator position, value_type&& x);
constexpr iterator insert(const_iterator position, size_type n, const value_type& x);
template <typename InputIterator>
  constexpr iterator insert(const_iterator position, InputIterator first, InputIterator last);
constexpr iterator insert(const_iterator position, initializer_list<value_type> il);

template <class... Args>
constexpr reference emplace_back(Args&&... args);
template <class... Args>
constexpr iterator emplace(const_iterator position, Args&&... args);
constexpr void push_back(const value_type& x);
constexpr void push_back(value_type&& x);

  • Requires: The number of elements to be inserted shall be at most C - size().

  • Remarks: All the iterators and references before the insertion point remain valid. If an exception is thrown other than by the copy constructor, move constructor, assignment operator, or move assignment operator of value_type or by any InputIterator operation there are no effects. If an exception is thrown while inserting a single element at the end and value_type is CopyInsertable or is_nothrow_move_constructible_v<value_type> is true, there are no effects. Otherwise, if an exception is thrown by the move constructor of a non-CopyInsertable value_type, the effects are unspecified.

  • Complexity: Linear in the number of elements inserted plus the distance from the insertion point to the end of the fixed_capacity_vector.

  • Throws: Any exception thrown by anassignment operator of value_type.

Note (not part of the specification): The insertion functions have as precondition new_size <= capacity(). Hence, they all have narrow contracts and are never noexcept(true).

constexpr void pop_back();
constexpr iterator erase(const_iterator position);
constexpr iterator erase(const_iterator first, const_iterator last);

  • Effects: Invalidates iterators and references at or after the point of the erase.

  • Complexity: The destructor of value_type is called the number of times equal to the number of the elements erased, but the assignment operator of value_type is called the number of times equal to the number of elements in the fixed_capacity_vector after the erased elements.

  • Throws: Any exception thrown by anassignment operator of value_type.

Note (not part of the specification): the erasure methods have as precondition new_size >= 0 (always satisfied) and new_size <= capacity(), hence they have narrow contracts.

constexpr void swap(fixed_capacity_vector x)
  noexcept(is_nothrow_swappable_v<value_type>);

  • Effects: Exchanges the contents of *this with x. All iterators pointing to the elements of *this and x are invalidated.

  • Complexity: Linear in the number of elements in *this and x.

5.7 fixed_capacity_vector specialized algorithms

template <typename T, size_t N>
constexpr void swap(fixed_capacity_vector<T, N>& x, 
                    fixed_capacity_vector<T, N>& y)
  noexcept(noexcept(x.swap(y)));

  • Remarks: This function shall not participate in overload resolution unless is_swappable_v<T> is true.

  • Effects: As if by x.swap(y).

  • Complexity: Linear in the number of elements in x and y.

6. Acknowledgments

The following people have significantly contributed to the development of this proposal. This proposal is based on Boost.Container's boost::container::static_vector and my extensive usage of this class over the years. As a consequence the authors of Boost.Container (Adam Wulkiewicz, Andrew Hundt, and Ion Gaztanaga) have had a very significant indirect impact on this proposal. The implementation of libc++ std::vector and the libc++ test-suite have been used extensively while prototyping this proposal, such that its author, Howard Hinnant, has had a significant indirect impact on the result of this proposal as well. The following people provided valuable feedback that influenced some aspects of this proposal: Walter Brown, Zach Laine, Rein Halbersma, and Andrzej Krzemieński. But I want to wholeheartedly acknowledge Casey Carter for taking the time to do a very detailed analysis of the whole proposal, which was invaluable and reshaped it in fundamental ways.

7. References