convert: for arrays, copy or view?

[timholy]

The docstring for convert is wonderfully explicit about providing details about the underlying assumptions in many cases, but I've found one that isn't yet covered. convert(T, x) has the behavior that if x already has type T, we just return x. But for AbstractArrays, things get more complicated:

julia> A = rand(3,5)
3×5 Array{Float64,2}:
 0.324103  0.45017   0.469432  0.382722  0.241098 
 0.336333  0.481648  0.165303  0.616973  0.0452947
 0.567737  0.964775  0.82846   0.300232  0.140285 

julia> pointer(A)
Ptr{Float64} @0x00007fad96e33ba0

julia> pointer(convert(Array, A))
Ptr{Float64} @0x00007fad96e33ba0

julia> S = view(A, :, :)
3×5 SubArray{Float64,2,Array{Float64,2},Tuple{Colon,Colon},true}:
 0.324103  0.45017   0.469432  0.382722  0.241098 
 0.336333  0.481648  0.165303  0.616973  0.0452947
 0.567737  0.964775  0.82846   0.300232  0.140285 

julia> pointer(convert(Array, S))
Ptr{Float64} @0x00007fad970d8e20

In the first case we didn't copy, in the second we did. Often there is no choice but to copy, for example if we had defined S = view(A, 1:2, 1:3); however, in cases like view(A, :, :) it would be possible to return the original without making a copy (and it wouldn't violate type-stability).

An even more intermediate case comes from array types like

struct MyArray{T,N} <: AbstractArray{T,N}
    data::Array{T,N}
end

For this type, convert(Array, A::MyArray) could always return A.data directly. But is this advisable? Or should it be a copy any time typeof(A) != T?

[nalimilan]

Duplicate of #12441?

Although it could be made even more explicit, the current docstring mentions that aliasing may happen (introduced in #17519):

If T is a collection type and x a collection, the result of convert(T, x) may alias x.

So in general I think convert should avoid copies as much as possible. That means we need a way to convert with the guarantee of getting a copy (similar to copy_oftype discussed at #12441).

[timholy]

Indeed, thanks @nalimilan. With 223 open issues mentioning convert and 54 of those mentioning copy, I just missed that.