SubArray: avoid invalid elimination of singleton indices

[N5N3]

close #53209

[vadim-z]

What if the last index is not singleton but, for example, 3? I fear it would be incorrectly removed from the list J'.

[N5N3]

Your concern would be took care by bounds checking. See

julia/base/abstractarray.jl

Line 725 in d453af8

return checkindex(Bool, get(inds, 1, OneTo(1)), I[1])::Bool &

L159 only eliminates trailing singleton indices after all parent's dimensions have been consumed.
It's safe as all of them == 1.
In fact, all of these trailing indices could be replaced with something like fill(CartesianIndex(), index_shape(trailing_indices...)) if length(index_shape(trailing_indices...)) > 0
But at present we don't have lazy fill in Base.

[vadim-z]

Sorry, but because of CartesianIndex() bound checking will pass, and this is correct behavior.

Example:

julia> A = rand(2,3)
2×3 Matrix{Float64}:
 0.0221974  0.820445  0.861314
 0.555607   0.457571  0.875773

julia> A[[CartesianIndex()], :, 3]
1×2 Matrix{Float64}:
 0.861314  0.875773

julia> J = map(i->Base.unalias(A,i), to_indices(A, ([CartesianIndex()], :, 3)))
([CartesianIndex()], Base.Slice(Base.OneTo(2)), 3)

julia> @boundscheck checkbounds(A, J...)

julia>

If some extra axes are prepended, the trailing indices are not really "trailing", as they refer to real dimensions of a parent array.

[N5N3]

Yes, but CartesianIndex() also consumes no parent's dimension.
Thus rm_singleton_indices won't eliminate 3 from J = ([CartesianIndex()], :, 3))
I guess the added test

@test A[CartesianIndices(()), :, 1] == @inferred(view(A, CartesianIndices(()), :, 1))

is enough?

[vadim-z]

Yes, you are right.

Yes, but CartesianIndex() also consumes no parent's dimension.

I've missed the point.