P2897R0 (aligned_accessor): LEWG feedback 2023/10/10

[mhoemmen]

LEWG review of P2897R0 (aligned_accessor) on 2023/10/10

Feedback

• Add example in the wording that uses the static member function to check pointer precondition
• Add example in the wording that uses aligned_alloc to create an overaligned allocation, to show that aligned_accessor exists as part of a system
• Change gcd converting constructor requirement from Constraint to Mandate
• Explore mdspan construction safety in a later paper (post-Kona)

Please see also https://github.com/cplusplus/papers/issues/1568 .

Design alternatives

One design alternative for expressing the alignment precondition on the pointer would be to introduce an aligned_pointer_t type (not alias), as in the example currently in the mdspan repository. The type's constructor would take a pointer and impose an alignment precondition on the pointer. This would shift the burden for checking preconditions from the mdspan as a whole, to data_handle_type itself. That way, mdspan itself wouldn't need a system for checking preconditions, yet (from the user's perspective) preconditions would still be checked at mdspan creation time.