work with ArrayViews

[goretkin]

I am interested in having https://github.com/JuliaArrays/ArrayViews.jl work for MArrays, which I think would have just worked before #55.

Any suggestions for how to do this without defining specific methods for MArray inside of ArrayViews? ArrayViews works for any mutable array with a regular memory layout, so I thought this was exactly what DenseArray was for.

[andyferris]

The pointer stuff never worked out, partially because it interacted badly with Julia's GC, so yeah we dropped DenseArray.

But more importantly, that package was an old prototype for Julia v0.3 users. Doesn't view and SubArray (now in Base) work for you?

julia> using StaticArrays

julia> v = MVector(1,2,3)
3-element MVector{3,Int64}:
 1
 2
 3

julia> view(v, 1:2)
2-element SubArray{Int64,1,MVector{3,Int64},Tuple{UnitRange{Int64}},true}:
 1
 2

[andyferris]

Also, instead of worrying about memory layout, the focus shifted to "linear" vs "Cartesian" indexing, and linear indexing is supported by static arrays.

Additionally, in the future we may improve the design such the the indices of a static array are static unit ranges. In that sense, a SubArray of a StaticArray may become more-or-less just like any other StaticArray.

[goretkin]

I was trying (unreasonably) hard to get zero allocations in v0.6, and when I last checked, constructing a SubArray allocates (the SubArray object has a field referring to a heap-allocated object, so it itself is heap allocated, if I understand correctly). That's why I was trying to use ArrayViews, because it has an unsafe_aview, which doesn't keep a reference, but just a pointer.

[andyferris]

I see.

Out of curiosity, have you tried v0.7? They have eliminated a great many useless GC allocations for things like SubArray wrappers.

[c42f]

Looks like this issue is probably solved by using SubArray and view in Base. If allocations are still present there's probably not a lot we can do in this package though; we'll need to wait for further upstream compiler improvements.