Issue 3369: span's deduction-guide for built-in arrays doesn't work

3369. `span`'s deduction-guide for built-in arrays doesn't work

Section: 23.7.2.2.1 [span.overview] Status: C++20 Submitter: Stephan T. Lavavej Opened: 2020-01-08 Last modified: 2021-02-25

Priority: 0

View all other issues in [span.overview].

View all issues with C++20 status.

Discussion:

N4842 22.7.3.1 [span.overview] depicts:

template<class T, size_t N>
span(T (&)[N]) -> span<T, N>;

This isn't constrained by 22.7.3.3 [span.deduct]. Then, 22.7.3.2 [span.cons]/10 specifies:

template<size_t N> constexpr span(element_type (&arr)[N]) noexcept;
template<size_t N> constexpr span(array<value_type, N>& arr) noexcept;
template<size_t N> constexpr span(const array<value_type, N>& arr) noexcept;
Constraints:

extent == dynamic_extent || N == extent is true, and

remove_pointer_t<decltype(data(arr))>(*)[] is convertible to ElementType(*)[].

Together, these cause CTAD to behave unexpectedly. Here's a minimal test case, reduced from libcxx's test suite:

C:\Temp>type span_ctad.cpp
#include <stddef.h>
#include <type_traits> 

inline constexpr size_t dynamic_extent = static_cast<size_t>(-1);

template <typename T, size_t Extent = dynamic_extent>
struct span {
  template <size_t Size>
  requires (Extent == dynamic_extent || Extent == Size)
#ifdef WORKAROUND_WITH_TYPE_IDENTITY_T
  span(std::type_identity_t<T> (&)[Size]) {}
#else
  span(T (&)[Size]) {}
#endif
};

template <typename T, size_t Extent>
#ifdef WORKAROUND_WITH_REQUIRES_TRUE
requires (true)
#endif
span(T (&)[Extent]) -> span<T, Extent>;

int main() {
  int arr[] = {1,2,3};
  span s{arr};
  static_assert(std::is_same_v<decltype(s), span<int, 3>>,
    "CTAD should deduce span<int, 3>.");
}

C:\Temp>cl /EHsc /nologo /W4 /std:c++latest span_ctad.cpp
span_ctad.cpp
span_ctad.cpp(26): error C2338: CTAD should deduce span<int, 3>.

C:\Temp>cl /EHsc /nologo /W4 /std:c++latest /DWORKAROUND_WITH_TYPE_IDENTITY_T span_ctad.cpp
span_ctad.cpp

C:\Temp>cl /EHsc /nologo /W4 /std:c++latest /DWORKAROUND_WITH_REQUIRES_TRUE span_ctad.cpp
span_ctad.cpp

C:\Temp>

(MSVC and GCC 10 demonstrate this behavior. Clang is currently affected by LLVM#44484.)

Usually, when there's an explicit deduction-guide, we can ignore any corresponding constructor, because the overload resolution tiebreaker 12.4.3 [over.match.best]/2.10 prefers deduction-guides. However, this is a mental shortcut only, and it's possible for guides generated from constructors to out-compete deduction-guides during CTAD. That's what's happening here.

Specifically, the constructor is constrained, while the deduction-guide is not constrained. This activates the "more specialized" tiebreaker first (12.4.3 [over.match.best]/2.5 is considered before /2.10 for deduction-guides). That goes through 13.7.6.2 [temp.func.order]/2 and 13.5.4 [temp.constr.order] to prefer the more constrained overload.

(In the test case, this results in span<int, dynamic_extent> being deduced. That's because the constructor allows T to be deduced to be int. The constructor's Size template parameter is deduced to be 3, but that's unrelated to the class's Extent parameter. Because Extent has a default argument of dynamic_extent, CTAD succeeds and deduces span<int, dynamic_extent>.)

There are at least two possible workarounds: we could alter the constructor to prevent it from participating in CTAD, or we could constrain the deduction-guide, as depicted in the test case. Either way, we should probably include a Note, following the precedent of 21.3.2.2 [string.cons]/12.

Note that there are also deduction-guides for span from std::array. However, the constructors take array<value_type, N> with using value_type = remove_cv_t<ElementType>; so that prevents the constructors from interfering with CTAD.

I'm currently proposing to alter the constructor from built-in arrays. An alternative resolution to constrain the deduction-guide would look like: "Constraints: true. [Note: This affects class template argument deduction. — end note]"

[2020-01-25 Status set to Tentatively Ready after seven positive votes on the reflector.]

Proposed resolution:

This wording is relative to N4842.

Modify 23.7.2.2.1 [span.overview], class template span synopsis, as indicated:

namespace std {
template<class ElementType, size_t Extent = dynamic_extent>
class span {
public:
  […]
  // 23.7.2.2.2 [span.cons], constructors, copy, and assignment
  constexpr span() noexcept;
  […]
  template<size_t N>
  constexpr span(type_identity_t<element_type> (&arr)[N]) noexcept;
  […]
};
[…]

Modify 23.7.2.2.2 [span.cons] as indicated:
```
template<size_t N> constexpr span(type_identity_t<element_type> (&arr)[N]) noexcept;
template<size_t N> constexpr span(array<value_type, N>& arr) noexcept;
template<size_t N> constexpr span(const array<value_type, N>& arr) noexcept;
```
-10- Constraints:
1. (10.1) — extent == dynamic_extent || N == extent is true, and
2. (10.2) — remove_pointer_t<decltype(data(arr))>(*)[] is convertible to ElementType(*)[].
-11- Effects: Constructs a span that is a view over the supplied array. [Note: type_identity_t affects class template argument deduction. — end note]
-12- Postconditions: size() == N && data() == data(arr).

3369. span's deduction-guide for built-in arrays doesn't work

3369. `span`'s deduction-guide for built-in arrays doesn't work