Constraint on layout_{left|right|stride}::stride is not implemented

[youyu3]

Per http://eel.is/c++draft/mdspan.layout.left.obs#5, should have

constexpr index_type stride(rank_type i) const;
Constraints: extents_­type​::​rank() > 0 is true

But is not implemented, for example: https://github.com/kokkos/mdspan/blob/stable/include/experimental/__p0009_bits/layout_left.hpp#L206

stride(0) returns 1 for a 0-D mapping (stdex::extents<size_t>).

[youyu3]

One possible fix is to have the following

    MDSPAN_TEMPLATE_REQUIRES(
      class E = Extents,
      /* requires */ (
        E::rank() > 0
      )
    )
    MDSPAN_INLINE_FUNCTION
    constexpr index_type stride(rank_type i) const noexcept {
      ...
    }

[youyu3]

stride isn't listed on http://eel.is/c++draft/mdspan.layout.stride.obs for layout_stride (it is on http://eel.is/c++draft/mdspan.layout.stride.overview), but I suppose it should have the same constraint?

[mhoemmen]

@youyu3 We defined layout_stride::mapping::stride using the exposition-only array<index_type, rank_> strides_. I suppose that one could be constrained too, but it doesn't need any constraints. It gets its precondition -- that the input is in [0, rank()) -- indirectly from the operator[] sequence container requirements.

[youyu3]

So is the behavior here https://godbolt.org/z/qWsnjeebT expected, i.e., stride(0) doesn't lead to a compilation error when the rank of the extent is 0?

[mhoemmen]

That is not the correct behavior per http://eel.is/c++draft/views#mdspan.layout.left.obs .

[crtrott]

@youyu3 can you issue that PR against kokkos/mdspan too?