Make .= work for arrays of arrays by returning a view from Base.dotview

[andreasnoack]

This was motivated by the example I've added as a test. This also changes the behavior of a[1] .= when the left-hand side is an array of arrays as can be seen from the tests I've commented out. I'm not sure about the behavior in this case. Part of me can see the point in the old behavior and part of me thinks it is wrong to let the broadcast behavior go one level down in the scalar case instead of considering the element a unit. One idea was I had was to restrict the definition I've now removed to

dotview{T<:AbstractArray}(A::AbstractArray{T}, args::Integer...) = getindex(A, args...)

but this wouldn't be correct for arrays with custom indices. Alternatively, I wondered if (a[1]) .= should parse differently such that the getindex is applied before the broadcast!.

[mbauman]

One possible option would be to re-introduce that special behavior in broadcast!(::AbstractArray{T,0} where T, …). Instead of using a shape of (), it'd use the shape of the element. This should generalize nicely for Array{Any,0}.

[stevengj]

We can't apply getindex before the broadcast, because for e.g. Array{Int} there will be no way to mutate the result of getindex.

What's wrong with dotview{T<:AbstractArray}(A::AbstractArray{T}, args::Integer...) = getindex(A, args...)? (If the indices aren't integers, then it isn't really an AbstractArray, and I don't think broadcast would work anyway.)

[andreasnoack]

With http://docs.julialang.org/en/release-0.5/devdocs/offset-arrays/#arrays-with-custom-indices, I didn't think that indices had to be Integers anymore. @timholy Am I wrong and should that be part of the documentation?

[stevengj]

Even if there are some weird non-integer-indexed arrays in the future, dotview{T<:AbstractArray}(A::AbstractArray{T}, args::Integer...) = getindex(A, args...) still seems worth adding now.

[andreasnoack]

I'm fine with that approach. We can revisit if it becomes a problem. I'll push the change in a moment.