Fix handling of array rvalues for ranges::cbegin and its friends

[frederick-vs-ja]

When the argument is an rvalue of an array, ranges::cbegin, ranges::cend, ranges::crbegin, and ranges::crend should reject it with substitution failures, despite of whether the element type is complete or not. I think such requirements haven't been changed by P2278R4 (i.e. unchanged between C++20 and C++23).

(On the other hand, if the argument is an lvalue of an array of an incomplete element type, the program is ill-formed, no diagnostic required.)

I originally discovered the bug of implementation in #3187 (comment), but hadn't determined how to fix it at that time.

---

Clang 15 is buggy - it fails to perform shortcut in concept substitution when using requires-clause. This is probably LLVM-55945 and fixed in Clang 16 (Godbolt link).

The current workaround is combining constraints into one concept and using the abbreviated syntax.

[strega-nil-ms]

I'm not sure about this; the standard says (for cbegin for example, but for the rest too):

If E is an rvalue and enable_­borrowed_­range<remove_­cv_­t<T>> is false, ranges​::​cbegin(E) is ill-formed. ([range.access.cbegin]/1.1)

They are also missing the very important Note 1 from ranges::begin (and also ranges::{end, rbegin, rend, size, empty}):

[Note 1: Diagnosable ill-formed cases above result in substitution failure when ranges​::begin(E) appears in the immediate context of a template instantiation. — end note] ([range.access.begin]/3)

What this says to me is that ranges::cbegin(T &&) (and the like) should be ill-formed (not a substitution failure) for all T that don't satisfy enable_borrowed_range<T>.

[StephanTLavavej]

We talked about this at the weekly maintainer meeting, and concluded that when WG21-P2278R4 updated the Standard, the Note that appears for ranges::begin et al. should have been applied to ranges::cbegin et al., but the Note is (by definition) informative (i.e. non-normative), so the mandated behavior is the same for all of them. Therefore the diagnosable ill-formed cases do indeed result in substitution failure.

@strega-nil-ms has volunteered to submit a editorial issue (or PR) to the C++ draft. Thanks Nicole!

[strega-nil-ms]

Submitted Draft-PR-6046.

[strega-nil-ms]

Thanks for the bug-fix!

[strega-nil-ms]

Thanks for the PR!

[StephanTLavavej]

No change requested: I observe that the noexcept expression differs from the return expression because _As_const_pointer is unconditionally noexcept - this is a reasonable shortcut for throughput.

[StephanTLavavej]

I'm mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.

[StephanTLavavej]

Thanks again for finding and fixing this, so the implementations are consistent! 😻 🔍 ✅