make_array, revision 2

Changes since N4031

• make_array<D> renamed to array_of<D>.
• Fixed cv-reference_wrapper<T> detection.

Changes since N3824

• Editorial wording updates.
• Comments updates.

Motivation

We have make_tuple, make_pair, but not make_array, while std::array creation can also benefit from this “semiofficial” tuple-like interface to deduce both element type and array bound.

Scope

LWG 851 intended to provide a replacement syntax to

array<T, N> a = { E1, E2, ... };

, so the following

auto a = make_array(42u, 3.14);

is well-formed (with additional static_casts applied inside) because

array<double, 2> = { 42u, 3.14 };

is well-formed.

This paper intends to provide a set of std::array creation interfaces which are comprehensive from both tuple’s point of view and array’s point of view, so narrowing is just naturally banned. See more details driven by this direction in Design Decisions.

Examples

auto a1 = make_array(2, 3L);        // array<long, 2>
auto ax = make_array(2, 3U);        // error: narrowing

auto a2 = array_of<long>(2, 3U);      // explicit destination type
auto ax = array_of<unsigned>(2, 3U);  // error: narrowing

auto a3 = make_array("foo");        // array<char const*, 1>, decayed
auto a4 = to_array("foo");          // array<char, 4>

Design Decisions

• Provide both make_tuple-like, type-deduced interface and raw array style bound-deduced interface.

• Ban reference_wrapper in the make_tuple-like interface. make_tuple and make_pair have special handling of reference_wrapper, then user might expect that the expression

also results in a tuple-like object storing T&. However, std::array does not store “real” references, and any attempts to workaround this break the interfaces in different ways. Note that “doing nothing” is not an option since, for example, common_type_t<reference_wrapper<int>, reference_wrapper<long>> is long, not reference or reference_wrapper.

• A saperated interface to perform constructing from raw array instead of array-to-pointer conversion. make_tuple and make_pair unconditionally decay, but such a behavior, when being applied to make_array,

is inexplicable. However, to keep the interfaces consistent, I decide to name a new utility differently instead of to ban this conversion.

Wording

This wording is relative to N3936.

Add to 23.3.1/2 [sequences.general], <array> synopsis:

namespace std {
  template <class T, size_t N > struct array;

…

  template <class T, size_t N >
    void swap(array<T,N>& x, array<T,N>& y) noexcept(noexcept(x.swap(y)));

…

}

New section 23.3.2.9 [array.creation] (between [array.zero] and [array.tuple], which was 23.3.2.9):

23.3.2.9 Array creation functions [array.creation]

template <class... Types>
  constexpr array<CT, sizeof...(Types)> make_array(Types&&...);

Let Ui be decay_t<Ti> for each Ti in Types.

Remarks: This function shall not participate in overload resolution unless Ui is not reference_wrapper<Ti> for all i.

[Author’s note: We allow users to detect and handle this case. –end note]

Returns: array<CT, sizeof...(Types)>{ std::forward<Types>(t))... }, where CT is common_type_t<Types...>.

[Example:

    int i = 1; int& ri = i;
    make_array(i, ri, 42L)

creates an array of type

    array<long, 3>

–end example]

template <class D, class... Types>
  constexpr array<D, sizeof...(Types)> array_of(Types&&...);

Returns: array<D, sizeof...(Types)>{ std::forward<Types>(t))... }.

template <class T, size_t N>
  constexpr array<V, N> to_array(T (&a)[N]);

Returns: An array<V, N> such that each element is copy-initialized with the corresponding element of a, where V is remove_cv_t<T>.

[Author’s note: If std::array is extended to support multidimensional array (N3794), this function might get extended as well, but we don’t let users to roll their own. On the other hand, if users understand multidimensional std::array as array of arrays, it might be more convenient and clear to write

make_array(make_array(1, 2, 3), make_array(4, 5, 6)...)

instead of to adapt a raw array. –end note]

Alternative Design

The past two revisions of this paper adds 2 names (make_array, to_array) rather than 3 (with the standalone array_of). If you prefer keeping the number of symbols minimal, or believe that the type&bound-deduced interface and the bound-deduced-only interface should be named the same, here is an alternative wording for such a design.

Sample Implementation

A sample implementation is available at https://gist.github.com/lichray/6034753.

Acknowledgments

Jonathan Wakely, who showed me how index_sequence helps initializing std::array from a raw array.

Daniel Krügler, who explained why an explicit destination type is essential.

Ville Voutilainen and other people who reviewed this paper.

Stephan T. Lavavej, who pointed out the ambiguity issue of the two make_array overloads, and came with with the naming ”*_of”.