`@inbounds` isn't propagated for `CartesianIndices`

[Tokazama]

Recent discussion in ArrayInterface.jl demonstrated that @inbounds isn't fully propagated when using CartesianIndices to convert an integer to a CartesianIndex.
Examples across multiple versions and hardware are more fully demonstrated here JuliaArrays/ArrayInterface.jl#199).

I think the offending line is the missing @propagate_inbounds here:

julia/base/abstractarray.jl

Line 1254 in 027071f

function _getindex(::IndexCartesian, A::AbstractArray, I::Vararg{Int,M}) where M

This could probably be fixed here also:

julia/base/abstractarray.jl

Line 1242 in 027071f

function _getindex(::IndexLinear, A::AbstractArray, I::Vararg{Int,M}) where M

I think @timholy and @mbauman wrote the original code here.
Can we just put @propagate_inbounds in front of these?

[mbauman]

Those implementations look ok to me; @propagate_inbounds might possibly be needed on their caller function, but they will appropriately skip the boundscheck if called with @inbounds as written. Hard for me to quickly identify where the problem is in the linked issue. Note that @propagate_inbounds is only needed if you aren't using @inbounds in the implementation and want to rely upon the caller's inbounds state.

[Tokazama]

K. Trying with this inbounds_getindex(x, i) = @inbounds(x[i]) produces the following errors when something is out of bounds.

ERROR: BoundsError: attempt to access 2-element Base.OneTo{Int64} at index [3]
Stacktrace:
  [1] throw_boundserror(A::Base.OneTo{Int64}, I::Int64)
    @ Base ./abstractarray.jl:691
  [2] getindex
    @ ./range.jl:863 [inlined]
  [3] _broadcast_getindex_evalf
    @ ./broadcast.jl:670 [inlined]
  [4] _broadcast_getindex
    @ ./broadcast.jl:643 [inlined]
  [5] #29
    @ ./broadcast.jl:1119 [inlined]
  [6] ntuple
    @ ./ntuple.jl:49 [inlined]
  [7] copy
    @ ./broadcast.jl:1119 [inlined]
  [8] materialize
    @ ./broadcast.jl:904 [inlined]
  [9] getindex
    @ ./multidimensional.jl:357 [inlined]
 [10] inbounds_getindex(x::CartesianIndices{2, Tuple{Base.OneTo{Int64}, Base.OneTo{Int64}}}, i1::Int64, i2::Int64)
    @ Main ./REPL[8]:1
 [11] top-level scope
    @ REPL[10]:1

This doesn't happen with --check-bounds=no.

I'm assuming that once we get into the broadcasting machinery there's a big risk of inadvertently messing things up, so could we just avoid broadcasting here:

julia/base/multidimensional.jl

Line 361 in 027071f

CartesianIndex(getindex.(iter.indices, I))

and instead just call CartesianIndex(I)