Segmentation fault after deep-copying SharedArray nested in a type

[DanielArndt]

Copied largely from a topic I started on julia-users: https://groups.google.com/forum/#!topic/julia-users/Cg6K3-mpAig

I've created a small stripped down version of the problem I am encountering.

type A
    sa::SharedArray{Float64,1}
end


function A(n::Int)
    sa = SharedArray(Float64, n, init = S -> S[Base.localindexes(S)] = 0)
    return A(sa)
end

function foo(a::A)
    for i in 1:length(a.sa)
        a.sa[i] += 1
    end
end

Executing the following on this type will break it:

julia> instance = A(5)
[0.0,0.0,0.0,0.0,0.0]

julia> dc = deepcopy(instance); # Semi-colon necessary on REPL to prevent show being called

julia> foo(instance) # Works fine on the original

julia> for i in 1:length(dc.sa)
           dc.sa[i] += 1
       end

julia> dc.sa # The above worked, and everything prints out fine on the REPL
5-element Array{Float64,1}:
 1.0
 1.0
 1.0
 1.0
 1.0

julia> foo(dc) # Seg fault

Turns out, the deepcopy is creating the SharedArray as an Array which is causing the issue:

julia> instance = A(5)
[0.0,0.0,0.0,0.0,0.0]

julia> dc = deepcopy(instance);

julia> typeof(instance)
A

julia> typeof(instance.sa)
SharedArray{Float64,1}

julia> typeof(dc)
A

julia> typeof(dc.sa)
Array{Float64,1}

This all seems to be caused by the similar function, which is returning an array. I have successfully prevented the seg fault by introducing this function:

Base.similar(S::SharedArray) = SharedArray(eltype(S), size(S), pids=procs(S))

For which I will create a pull request. I would love to hear comments, alternative approaches, or suggestions that might make the PR more robust, since this will be my first attempt dabbling in the base language.

[DanielArndt]

I guess it makes more sense to change similar(S::SharedArray, T, dims::Dims) and modify the others to use that...

[tkelman]

ref #12964, which I was not a fan of.

[DanielArndt]

Great, thanks for the context. I'll hold off working on this then since I imagine some debate needs to be had.

[KristofferC]

Fixed with the change to SharedArray indexing I presume.