[temp.deduct.call] Deduction rules for array sizes are inconsistent with character array initialization

[eisenwave]

Reference (section label): [temp.deduct.call]

Issue description

Consider the following example:

struct T {
  template<int N>
  T(const char(&)[N]) {}
};

struct S {
  S(const char(&)[5]) {}
};

T t({"awoo"}); // error: initializer-string for char array is too long
S s({"awoo"}); // OK

There is a surprising and likely unintentional inconsistency here. Implementations diverge on whether the initialization of t is valid.

While [dcl.init.string] generally allows initialization of character arrays with a string-literal that is surrounded by braces, [temp.deduct.call] paragraph 1 states that N is deduced to the length of the initializer list, which is 1, resulting in a failed attempt at initializing const char(&)[1] with "awoo".

Suggested resolution

W.I.P.

[eisenwave]

To be fair, the wording is clear here, so maybe this is more EWG paper territory. I'm not really sure; I think it's possible to interpret this as a clear oversight, especially considering the implementation divergence.

[frederick-vs-ja]

Consider these modified examples:

struct W {
  template<class U, int N>
  W(const U(&)[N]) {}
};
W w({"awoo"}); // OK: `U` is `const char*` and `N` is `1`.

struct X {
  X(const char(&)[5]) {}        // #1
  X(const char* const(&)[1]) {} // #2
};
X x({"awoo"}); // error: ambiguous

Perhaps we want to allow consider argument {"awoo"} (more generally, a braced string literal) as both const char(&)[5] and const char* const(&)[1] (or some cvref-adjusted variant) in deduction. I think we want to leave the deduction result in W w({"awoo"}); unchanged.

Do we want to deduce template arguments in two ways in W w({"awoo"});, and decide that the existing way is better in overload resolution, or elsewhere? Likewise, do we want to resolve the ambiguity in X x({"awoo"}); by specifying #2 to be better? Or just retry in the deduction for a single function template and keep the ambiguity among more than one overloads?

[jensmaurer]

Non-obvious design choices appear to be required. The existing wording seems clear, although the difference between deduction and initialization is a bit disturbing. (We have those differences elsewhere, e.g. for access control, though.)

I think a short paper targeted at EWG suggesting a holistic approach for the examples given here would be appropriate.