zero-argument getindex method in Base incorrectly assumes 1-based indexing

[yurivish]

It seems that the bounds check used in Base for zero-dimensional indexing hardcodes a comparison to 1, assuming one-based indexing.

This means that an array with nonstandard axes can successfully pass the bounds check and then access out-of-bounds memory when indexing. I encountered this in an issue on OffsetArrays and @jishnub tracked it down to this line in Base:

julia/base/abstractarray.jl

Line 642 in 69d2453

checkbounds_indices(::Type{Bool}, IA::Tuple, ::Tuple{}) = (@_inline_meta; all(x->unsafe_length(x)==1, IA))

[jishnub]

The issue here is that A[] gets mapped to A[1], and linear indexing for AbstractVectors should use their axes to obtain bounds on permissible indices. A 1-element OffsetVector may have an axis n:n for an arbitrary n, in which case 1 is not a valid index. Perhaps the best solution is to add a special method checkbounds_indices(::Type{Bool}, IA::Tuple{Any}, ::Tuple{}) that ensures that IA[1] begins at 1? This is the solution chosen in the fix in OffsetArrays.

[mbauman]

No, that's working correctly. That method gets hit if you index into a non-zero-dimensional array with zero indices. It asserts that all axes have length 1.

The problem is not in the bounds checking, but rather in the getindex.

[mbauman]

Here's the root problem:

julia/base/abstractarray.jl

Line 1196 in 69d2453

_to_linear_index(A::AbstractArray) = 1

Edit: This should probably be first(eachindex(A)) or somesuch.

[yurivish]

Thanks for debugging this further @mbauman. I think my original issue title is incorrect; I've changed it to what I think is an appropriate description. Feel free to change if there's a better or more correct title.

[jishnub]

This is great! I wonder if a similar fix might be applied to #37274?