ISO/IEC JTC1 SC22 WG21 N3493 = 2013-01-11
Jonathan Wakely, cxx@kayari.org
Introduction
Rationale and examples of use
Piecewise construction of std::pair
.
Convert array to tuple
N3466 more_perfect_forwarding_async
example
Solution
Type of integers in the sequence
Non-consecutive integer sequences
Members of integer sequences
Impact on standard
20.x Compile-time integer sequences [intseq]
20.x.1 In general [intseq.general]
20.x.2 Class template integer_seq
[intseq.intseq]
20.x.3 Alias template make_integer_seq
[intseq.make]
References
In a function template such as
template<class... T>
void f(T... t);
performing some operation using every member of the parameter pack is
as simple as expr(t...)
,
because t
can be used directly in a pack expansion.
When the parameter pack appears as part of another type such as
std::tuple
, the function parameter is not a
parameter pack so can't be expanded directly.
For example, in the function template
template<class... T>
void foo(std::tuple<T...> t);
performing an operation on every element of the tuple requires additional
code because t
is not a parameter pack and
expr(t...)
is not valid.
(By comparison, in Python unpacking tuples to pass to variadic functions
is built in to the language and is as convenient as func(*t)
[Python].)
For unpacking a tuple in C++ the brute force alternative to
expr(t...)
is to instantiate a class template with a
static member function, such as
DoExpr<0>::apply(t);
which calls
DoExpr<1>::apply(t, std::get<0>(t));
which calls
DoExpr<2>::apply(t, t0, std::get<1>(t));
and so on for each element until
DoExpr<sizeof...(T)>
is instantiated and can call expr(t0, t1, ..., tN)
.
This requires the user to write a recursive class template (and a
specialization to terminate the recursion) with a variadic member function
template and to use perfect forwarding (not shown here) to avoid
O(N2) copies.
To make a generic, reusable solution that can apply an arbitrary
function object to each element requires template partial
specialization and is probably beyond the abilities of many users of
std::tuple
, so the alternative is to write non-generic code
and then rewrite it slightly differently when a different operation on
tuples is needed. This is a fairly complicated task and the end result is
not even a very natural way to write the desired operation.
The call to
DoExpr<0>::apply(t);
is far less expressive than something like expr(t...)
.
A simpler and more natural solution is to introduce a new parameter pack,
I
, consisting of integers [0,N] and then use it in
a pack expansion such as expr(std::get<I>(t)...)
.
This requires no recursion and no class template specializations.
This proposal is to standardize a mechanism for creating a parameter
pack containing an integer sequence.
The proposal is to add a type like
template<int...> struct int_seq { };
so that an instantiation like int_seq<0, 1, 2, 3>
can be passed to a function template that deduces a parameter
pack containing the integer sequence 0, 1, 2, 3.
In order to pass an int_seq
to a function it must be
possible to construct the desired specialization either from the length of
sequence desired or from a parameter pack from which the length can be
deduced. This proposal offers both options, so that given the type
tuple<T...>
the sequence [0, sizeof...(T))
needed to expand the tuple can be created by using the alias
make_int_seq<sizeof...(T)>
(or equivalently
make_int_seq<tuple_size<tuple<T...>::value>
)
or by using the alias to_int_seq<T...>
.
With such a type in your toolbox it is easy to write the generic, reusable solution described above for expanding tuple elements as arguments to a function object:
template<typename F, typename Tuple, int... I>
auto
apply_(F&& f, Tuple&& args, int_seq<I...>) ->
decltype(std::forward<F>(f)(std::get<I>(std::forward<Tuple>(args))...))
{
return std::forward<F>(f)(std::get<I>(std::forward<Tuple>(args))...);
}
template<typename F, typename Tuple,
typename Indices = make_int_seq<std::tuple_size<Tuple>::value>>
auto
apply(F&& f, Tuple&& args) ->
decltype(apply_(std::forward<F>(f), std::forward<Tuple>(args), Indices()))
{
return apply_(std::forward<F>(f), std::forward<Tuple>(args), Indices());
}
This solution still uses recursion and template specialization, but
only in the implementation of make_int_seq
, where users don't
need to write or understand it.
Standardizing the components makes it trivial for users to expand
tuple-like types in any situation, including the examples below.
A type like int_seq
is a good candidate for standardization,
as similar types are already used in at least two major standard library
implementations and have been reinvented often.
Similar types feature in several answers to questions on sites such as
Stack Overflow
[Schaub]
[Wakely]
[Kühl],
where the number of questions about unpacking tuples
and the number of times types like int_seq
are used in answers
demonstrates the demand for a solution to the problem.
std::pair
.
Piecewise construction of std::pair
requires two
tuple
s of arguments to be expanded to form constructor
arguments for the pair members. One implementation technique is to
delegate to a private constructor for pair
that takes
two types like int_seq
, for example
template<class... Args1, class... Args2>
pair(piecewise_construct_t, tuple<Args1...> args1, tuple<Args2...> args2)
: pair(args1, args2, to_int_seq<Args1...>{}, to_int_seq<Args2...>{})
{ }
private:
template<class... A1, class... A2, int... I1, int... I2>
pair(tuple<A1...>& a1, tuple<A2...>& a2, int_seq<I1...>, int_seq<I2...>)
: first(std::get<I1>(std::move(a1))...),
second(std::get<I2>(std::move(a2))...)
{ }
A note in [tuple.creation] says that implementations might support using
std::tuple_cat
with other tuple-like type such as
std::array
, but it's not required.
Users who need to implement it themselves can do so using a recursive
template, using tuple_cat
to build up a tuple
element by element, but this creates lots of temporaries and instantiations.
An implementation is shown here for comparison with the
int_seq
solution:
template<typename T, std::size_t N, std::size_t I = 0>
struct a2t_
{
static_assert( I < N, "index within array bound" );
template<typename... Elts>
static auto
cat(std::tuple<Elts...>&& t, const std::array<T, N>& a)
-> decltype(a2t_<T, N, I+1>::cat(std::tuple_cat(t, std::make_tuple(a[I])), a))
{
return a2t_<T, N, I+1>::cat(std::tuple_cat(t, std::make_tuple(a[I])), a);
}
};
template<typename T, std::size_t N>
struct a2t_<T, N, N>
{
template<typename... Elts>
static std::tuple<Elts...>
cat(std::tuple<Elts...>&& t, const std::array<T, N>& a)
{
return t;
}
};
template<typename T, std::size_t N>
auto
a2t(const std::array<T, N>& a)
-> decltype(a2t_<T, N>::cat(std::tuple<>(), a))
{
return a2t_<T, N>::cat(std::tuple<>(), a);
}
Using int_seq
the implementation is almost trivial:
template<typename Array, int... I>
auto
a2t_(const Array& a, int_seq<I...>)
-> decltype(std::make_tuple(a[I]...))
{
return std::make_tuple(a[I]...);
}
template<typename T, std::size_t N, typename Indices = make_int_seq<N>>
auto
a2t(const std::array<T, N>& a)
-> decltype(a2t_(a, Indices()))
{
return a2t_(a, Indices());
}
more_perfect_forwarding_async
example
The more_perfect_forwarding_async
example in
N3466
refers to a "complicated template metafunction" [DRayX]
which can be replaced with the comparatively simple apply()
function template defined above.
While I hope there will be broad agreement that some form of integer sequence template should be standardized there are some parts of the solution which might be harder to agree on. The changes to the working paper proposed below are my preferred solution but the specific details are less important than getting some kind of template that make it easy to solve the problems described above.
It's not obvious to me what type of integer should be used in the
parameter pack of the class template. The natural type for indexing into
tuples and arrays is std::size_t
, but that type has a far
larger range of values than needed. I consider it unlikely that any
compiler will support parameter packs that can accept SIZE_MAX
arguments for the foreseeable future.
Another question is whether the integer type should be signed or unsigned.
Indices are naturally unsigned, but there could be other interesting uses
for integer sequences other than unpacking tuples and arrays which would
be more difficult if the values cannot be negative.
Conversely there might be valid use cases where the wrapping behaviour of
unsigned integers is preferable to worrying about overflow.
The proposed solution is to define a generic integer_seq
class template which can be instantiated as
integer_seq<int>
or
integer_seq<unsigned char>
or any other integer type.
By using alias templates to denote specializations for common types it is
just as convienent to use the generic template as it would be if it wasn't
parameterized by the integer type, for instance the examples above work
equally well whether int_seq<I...>
is a specialization
of a class template or an alias for integer_type<int,
I...>
.
There is no reason to restrict the values in an integer sequence to
consecutive integers starting from zero.
It is very simple to transform
int_seq<I...>
to int_seq<I+n...>
or int_seq<I*n>
.
Such sequences would be useful to access only the even elements of an array,
for example, but such transformations are not necessary for the main
purpose of unpacking tuples. If the standard provides the
integer_seq
type and simple metafunctions for generating
basic sequences users can more easily create custom sequences.
For most uses
template<int...> struct int_seq { };
is sufficient, but it's simple to define type
and
size
members describing the type and to provide
append
and next
members for extending
sequences, as shown in the proposed changes below.
These members are not essential to the proposal and the suggestion
might be due to my own bias as the next
member is used
by my implementation of the proposed changes below
[integer_seq]
i.e. in the unspecified details of how make_integer_seq
works.
This is a pure addition to the library.
It is expected that implementations would be able to replace existing
equivalent code with uses of integer_seq
if desired.
Add to the synopsis in [utility] paragraph 2, after the declaration of
tuple
:
// 20.x Compile-time integer sequences template<class T, T...> struct integer_seq; template<int... I> using int_seq = integer_seq<int, I...>; template<unsigned... I> using uint_seq = integer_seq<unsigned, I...>; template<size_t... I> using index_seq = integer_seq<size_t, I...>; template<class T, T N> using make_integer_seq = integer_seq<T, see below>; template<int N> using make_int_seq = make_integer_seq<int, N>; template<unsigned N> using make_uint_seq = make_integer_seq<unsigned, N>; template<size_t N> using make_index_seq = make_integer_seq<size_t, N>; template<class... T> using to_int_seq = make_int_seq<sizeof...(T)>; template<class... T> using to_uint_seq = make_uint_seq<sizeof...(T)>; template<class... T> using to_index_seq = make_index_seq<sizeof...(T)>;
Add a new subclause after [pairs]:
20.x Compile-time integer sequences [intseq]
20.x.1 In general [intseq.general]
The library provides a class template that can represent an integer sequence. When used as an argument to a function template the integer sequence can be deduced as a parameter pack and used in a pack expansion.
[Example:
template<class F, class Tuple, unsigned... I> auto apply_impl(F&& f, Tuple&& t, index_seq<I...>) -> decltype(std::forward<F>(f)(std::get<I>(std::forward<Tuple>(t))...) { return std::forward<F>(f)(std::get<I>(std::forward<Tuple>(t))...); } template<class F, class Tuple, class Indices = make_index_seq<std::tuple_size<Tuple>::value>> auto apply(F&& f, Tuple&& t) -> decltype(apply_impl(std::forward<F>(f), std::forward<Tuple>(t), Indices())) { return apply_impl(std::forward<F>(f), std::forward<Tuple>(t), Indices()); }
— end example ]
20.x.2 Class template
integer_seq
[intseq.intseq]namespace std { template<class T, T... I> struct integer_seq { typedef T type; static constexpr T size = sizeof...(I); template<T N> using append = integer_seq<T, I..., N>; using next = append<sizeof...(I)>; }; } // namespace std
T shall be an integer type.
20.x.3 Alias template
make_integer_seq
[intseq.make]template<class T, T N> using make_integer_seq = integer_seq<T, see below>;
The alias template
make_integer_seq
denotes a specialization ofinteger_seq
to ease construction of an integer sequence from the size of a template parameter pack. The typemake_integer_seq<T, N>
denotes the typeinteger_seq<T, 0, 1, ..., N-1>
.