UnsafeArrays doesn't correctly handle dimensions of size zero

[MasonProtter]

Discovered in MasonProtter/Bumper.jl#43, here's a MWE:

Compare

julia> let A = Array{Float64}(undef, 0, 10)
           A[:, 1] .= 1
       end
0-element view(::Matrix{Float64}, :, 1) with eltype Float64

against

julia> let A = Array{Float64}(undef, 0, 10)
           UA = UnsafeArray(pointer(A), (0, 10))
           UA[:, 1] .= 1
       end
ERROR: BoundsError: attempt to access 0×10 LinearIndices{2, Tuple{Base.OneTo{Int64}, Base.OneTo{Int64}}} at index [1, 1]
Stacktrace:
  [1] throw_boundserror(A::LinearIndices{2, Tuple{Base.OneTo{Int64}, Base.OneTo{Int64}}}, I::Tuple{Int64, Int64})
    @ Base ./abstractarray.jl:737
  [2] checkbounds
    @ ./abstractarray.jl:702 [inlined]
  [3] _getindex
    @ ./abstractarray.jl:1323 [inlined]
  [4] getindex
    @ ./abstractarray.jl:1291 [inlined]
  [5] _unsafe_view_impl
    @ ~/Dropbox/Julia/UnsafeArrays/src/unsafe_array.jl:98 [inlined]
  [6] unsafe_view
    @ ~/Dropbox/Julia/UnsafeArrays/src/unsafe_array.jl:71 [inlined]
  [7] view
    @ ./subarray.jl:186 [inlined]
  [8] maybeview
    @ ./views.jl:148 [inlined]
  [9] dotview(::UnsafeArray{Float64, 2}, ::Function, ::Int64)
    @ Base.Broadcast ./broadcast.jl:1244
 [10] top-level scope
    @ REPL[13]:3

Offending code is https://github.com/JuliaArrays/UnsafeArrays.jl/blob/main/src/unsafe_array.jl#L89-L97

I guess the easiest would be to do something like

@inline function _unsafe_view_impl(IFwd::NTuple{M,Base.Slice}, A::UnsafeArray{T,N}, i::DenseIdx, I::Integer...) where {T,M,N}
    @assert IFwd == ntuple(i -> axes(A)[i], Val{M}())
    I_all = (IFwd..., i, I...)
    @boundscheck checkbounds(A, I_all...)
    if length(A) == 0
        p = pointer(A)
    else
        startidxs = map(first, (IFwd..., i, I...))
        sub_s = _sub_size(I_all...)
        p = pointer(A, LinearIndices(size(A))[startidxs...])
    end
    UnsafeArray(p, sub_s)
end

but I wonder if that branch would have negative performance impacts in code that's taking a lot of views (e.g. blocking SIMD), so I wonder if there's a non-branching way to do this.

[oschulz]

Thanks for reporting this! I hope #22 fixes this in a branch-free fashion.