Failure in mixed indexing with multidimensional logicals

[timholy]

Inspired by https://stackoverflow.com/questions/64549136/i-want-to-do-logical-masking-on-an-image-in-julia

julia> img = cat([1 2; 3 4], [1 5; 6 7]; dims=3)
2×2×2 Array{Int64, 3}:
[:, :, 1] =
 1  2
 3  4

[:, :, 2] =
 1  5
 6  7

julia> mask = img[:,:,1] .== img[:,:,2]
2×2 BitMatrix:
 1  0
 0  0

julia> img[mask,2] .= 0
ERROR: BoundsError: attempt to access 2×2×2 Array{Int64, 3} at index [Bool[1 0; 0 0], 2]
Stacktrace:
 [1] throw_boundserror(A::Array{Int64, 3}, I::Tuple{Base.LogicalIndex{CartesianIndex{2}, BitMatrix}, Int64})
   @ Base ./abstractarray.jl:601
 [2] checkbounds
   @ ./abstractarray.jl:566 [inlined]
 [3] view
   @ ./subarray.jl:158 [inlined]
 [4] maybeview
   @ ./views.jl:133 [inlined]
 [5] dotview(::Array{Int64, 3}, ::BitMatrix, ::Int64)
   @ Base.Broadcast ./broadcast.jl:1160
 [6] top-level scope
   @ REPL[3]:1

[mcabbott]

A possible downside is that this will make #35681 easier to hit: img[mask, end] won't work correctly (with say img = rand(1:5, (3,5,2)) so the sizes differ).

[timholy]

#35681 (comment)

[lxvm]

Here is a slight modification of the MWE that also errors

julia> img[2,mask] .= 0
ERROR: BoundsError: attempt to access 2×2×2 Array{Int64, 3} at index [2, 2×2 BitMatrix]
Stacktrace:
 [1] throw_boundserror(A::Array{Int64, 3}, I::Tuple{Int64, Base.LogicalIndex{Int64, BitMatrix}})
   @ Base ./abstractarray.jl:734
 [2] checkbounds
   @ Base ./abstractarray.jl:699 [inlined]
 [3] view
   @ Base ./subarray.jl:179 [inlined]
 [4] maybeview
   @ Base ./views.jl:148 [inlined]
 [5] dotview(::Array{Int64, 3}, ::Int64, ::BitMatrix)
   @ Base.Broadcast ./broadcast.jl:1244
 [6] top-level scope
   @ REPL[65]:1

I'll try opening a pr to fix it

[mbauman]

@lxvm this should be fixed by #45869:

julia> img = rand(Int, 2,2,2);

julia> img[2, trues(2,2)] .= 0
4-element view(::Array{Int64, 3}, 2, CartesianIndex{2}[CartesianIndex(1, 1), CartesianIndex(2, 1), CartesianIndex(1, 2), CartesianIndex(2, 2)]) with eltype Int64:
 0
 0
 0
 0

julia> img
2×2×2 Array{Int64, 3}:
[:, :, 1] =
 615342675984579180  -924825086713411107
                  0                    0

[:, :, 2] =
 -3776513274107704031  2722564769933431785
                    0                    0

[lxvm]

@mbauman Thanks! It looks like getindex is also fixed. I was concerned that the logic of _maybe_linear_logical_index in base/multidimensional.jl is incorrect. Does the following work for you?

julia> ones(2,2,2)[:,iseven.(ones(2,2))]
2×0 Matrix{Float64}

Edit: I rebuilt Julia and it is indeed fixed since _maybe_linear_logical_index is no longer called for trailing indices