P2630: Kona (2022/11/11) LEWG feedback

submdspan Kona 2022/11/11 feedback

"The SliceSpecifier template argument(s)" in (2) should be SliceSpecifiers (typo).

Bikeshed: strided_index_range is obviously not a range in the sense of the "ranges" library.

Does src-indices need to be constexpr? In general, do these functions need to be constexpr?

submdspan_extents:

• Should the constraints be mandates instead?

• Typo: (2) says src.extents(), should be src_exts.extents()

• Why the difference between constraints and mandates? What's the goal? Note that the Mandates would subsume ... we could probably get rid of the Constraints there.

• Rather than requiring integral_constant, should we consider requiring it to be a "core constant literal type"? hana like? Expect some off-line comments to clarify this request.

submdspan_mapping

• Do we run into the problems that CPOs solve? Specifically making it hard to debug where exactly it landed? -- Users generally wouldn't call these functions, so users don't have to do the ADL dance. However, if this were a hidden friend, then you would only be able to call these by ADL. Some additional benefits. We could explore this.

• Use the state of the art for customization points; maybe tag_invoke if it is considered the state of the art.

Specializations of submdspan_mapping

(6) is missing a close angle bracket (array<SubExtents::index_type, SubExtents::rank()>).

sub_map was not introduced before (3.1) (the Mandates that starts with is_same_v and checks the extents with submdspan_extents result). (sub_map should be something like declval<decltype(submdspan_mapping(...))>.)

submdspan

(Effects rely on CTAD, but that's OK.)

Any design questions

1. Customization point design?
2. integral_constant?
3. Convertibility to tuple
4. stride design

IL: Niebloids are used in ranges, tag_invoke is used for sender/receivers. I would suggest looking at the paper by Barry (https://wg21.link/P2279) and applying a more "modern-style" CP mechanism.

Read the submdspan sections discussed in this issue, then review Barry's P2279 for the customization-point context. The work is complete when the listed wording errors are corrected and the open design questions about customization points, constexpr, constraints, integral_constant, tuple conversion, and strides have documented resolutions.