lastindex and firstindex return type

[nalimilan]

As noted at #26775, we currently allow custom lastindex and firstindex implementations to return non-Int values. For example, Date ranges use an Int64 index, and their last index (and length) cannot always be represented using an Int on 32-bit.

In PR #26775, this required defining lastindex and firstindex in terms of eachindex(IndexLinear(), x) rather than of LinearIndices(x). While it's not really an issue in practice, the fact that LinearIndices is not always able to represent the indices of ranges could be considered as a problem.

Is this really the desired behavior, or should we require that indices be Int?

[mbauman]

This is a little unfortunate, but non-Int indices have always been a bit of a poorly-supported case. We should be able to slowly improve things, but it's not currently possible to reshape large Date ranges on an Int32 machine. Or, equivalently:

julia> 1:(typemax(Int128)÷2+1)
1:85070591730234615865843651857942052864

julia> length(1:(typemax(Int128)÷2+1))
85070591730234615865843651857942052864

julia> reshape(1:(typemax(Int128)÷2+1), :, 2)
ERROR: MethodError: no method matching _reshape(::UnitRange{Int128}, ::Tuple{Int128,Int64})

We could allow LinearIndices (and CartesianIndices) to be parameterized by the resulting index type, but even then they'll have difficulty doing the ind2sub work they're meant to do because the abstract infrastructure does the index computations in terms of Int. I suppose we could similarly re-implement those methods.