zero(u::StructArray) broken in Julia v1.8.0-rc1

[ranocha]

The generic fallback changed from

zero(x::AbstractArray{T}) where {T} = fill!(similar(x), zero(T))

in Julia v1.7.3 to

zero(x::AbstractArray{T}) where {T} = fill!(similar(x, typeof(zero(T))), zero(T))

in Julia v1.8.0-rc1. This means that we had

julia> using Pkg; Pkg.status(["StructArrays", "StaticArrays"])
      Status `~/.julia/environments/v1.7/Project.toml`
  [90137ffa] StaticArrays v1.4.4
  [09ab397b] StructArrays v0.6.8

julia> using StructArrays, StaticArrays

julia> u = StructArray([SVector(1.0)])
1-element StructArray(::Vector{Float64}) with eltype SVector{1, Float64}:
 [1.0]

julia> zero(u)
1-element StructArray(::Vector{Float64}) with eltype SVector{1, Float64}:
 [0.0]

in Julia v1.7 but get

julia> using Pkg; Pkg.status(["StructArrays", "StaticArrays"])
Status `~/.julia/environments/v1.8/Project.toml`
  [90137ffa] StaticArrays v1.4.6
  [09ab397b] StructArrays v0.6.8

julia> using StructArrays, StaticArrays

julia> u = StructArray([SVector(1.0)])
1-element StructArray(::Vector{Float64}) with eltype SVector{1, Float64}:
 [1.0]

julia> zero(u)
1-element Vector{SVector{1, Float64}}:
 [0.0]

in Julia v1.8.0-rc1. This breaks a lot of code, for example in Trixi.jl relying on OrdinaryDiffEq.jl, see https://github.com/trixi-framework/Trixi.jl/runs/6830499545?check_suite_focus=true#step:6:244.

To fix this, it would be great if #218 or #94 were finished. Is there anything I can do to help you with that?

[piever]

To be honest, #218 can probably be merged already. The minor issue that's holding me back is that IMO ideally reshape and similar with "offset indices" (i.e. AbstractUnitRange) applied to a StructArray should yield a StructArray of OffsetArrays, and not vice versa.

On the other hand, getting that right seems extremely brittle, as StructArrays should mimic the exact dispatches that OffsetArrays is using. But maybe it's not so bad.

D you know if one can rely on the exact dispatch chosen by OffsetArrays (ie Tuple{OffsetAxisKnownLength,Vararg{OffsetAxisKnownLength}} for similar and Tuple{OffsetAxis,Vararg{OffsetAxis}} for reshape)? In that case, I can probably just add the correct signatures. Otherwise it may be best to just merge the PR as is.

cc: @timholy (esp. on how to intercept the OffsetArrays fallback for similar with AbstractUnitRange)

[ranocha]

Thanks a lot! Sorry, I do not know a lot about the internals of OffsetArrays.jl.

[ranocha]

The minor issue that's holding me back is that IMO ideally reshape and similar with "offset indices" (i.e. AbstractUnitRange) applied to a StructArray should yield a StructArray of OffsetArrays, and not vice versa.

Is this a major blocking issue, @piever, or just something that would be nice to have and could also be changed/implemented later? If the latter is the case, I would be very happy to see #218 merged soon to let use test Julia v1.8.0-rc1 with Trixi.jl.

[piever]

Is this a major blocking issue, @piever, or just something that would be nice to have and could also be changed/implemented later?

Changing it is (very slightly) breaking, so I would like to avoid tweaking this too many times. That being said, #218 now implements it correctly, so if there is no further feedback in the next couple of days I'll go ahead and merge and tag a release.

[ranocha]

Thanks a lot, @piever!

[piever]

Fixed by #218

[ranocha]

Thanks a lot!