1.5: Surprising new pointer behaviour with adjoints and views

[dlfivefifty]

This seems wrong to me:

julia> A = randn(5,3);

julia> V = view(A',:,2:4)
3×3 view(::Adjoint{Float64,Array{Float64,2}}, :, 2:4) with eltype Float64:
 -0.6594    -0.124149   0.492452
  1.20338    0.224742  -0.0431509
 -0.425888   0.725104   0.976229

julia> Base.unsafe_load(pointer(V))
0.49245221105087544

I expected it to return the pointer to V[1,1], not V[1,end].

Related to #35929

@mbauman

[mbauman]

Ah, yes. Our SubArray is implicitly assuming strided parents are both contiguous and column-major:

julia/base/subarray.jl

Lines 401 to 402 in b594172

	unsafe_convert(::Type{Ptr{T}}, V::SubArray{T,N,P,<:Tuple{Vararg{RangeIndex}}}) where {T,N,P} =
	unsafe_convert(Ptr{T}, V.parent) + (first_index(V)-1)*sizeof(T)

That needs to be doing a strides computation.