"Spills" from adjacent views of ROCVector

[tkf]

With AMDGPU 0.2.5, I get

julia> function f(xs)
           fill!(xs, 0)
           a = @view xs[1:end÷2]
           b = @view xs[end÷2+1:end]
           fill!(a, 1)
           a[1], a[end], b[1], b[end], b[[1, end]]
       end
f (generic function with 1 method)

julia> f(ROCVector{Int}(undef, 2^11))
(1, 1, 1, 1, [0, 0])

julia> f(Vector{Int}(undef, 2^11))
(1, 1, 0, 0, [0, 0])

but I expect b[1] == b[end] == 0. It's also strange that b[[1, end]] is different from [b[1], b[end]].