reinterpret of Arrays of static arrays failed

[Hong-Xiang]

reinterpret Arrays of static arrays failed.

steps to reproduce:

using StaticArrays
function svectors(x::Matrix{Float64})
    @assert size(x,1) == 3
    reinterpret(SVector{3,Float64}, x, (size(x,2),))
end
m = zeros(3, 10)
svectors(m)

outputs:

ERROR: MethodError: no method matching reinterpret(::Type{SArray{Tuple{3},Float64,1,3}}, ::Array{Float64,2}, ::Tuple{Int64})
Closest candidates are:
  reinterpret(::Type{T}, ::Base.ReshapedArray, ::Tuple{Vararg{Int64,N}} where N) where T at reshapedarray.jl:191
  reinterpret(::Type{T}, ::A<:AbstractArray{S,N}) where {T, N, S, A<:AbstractArray{S,N}} at reinterpretarray.jl:14
  reinterpret(::Type{T}, ::Any) where T at essentials.jl:370
  ...
Stacktrace:
 [1] svectors(::Array{Float64,2}) at ./REPL[2]:3
 [2] top-level scope at none:0

[chethega]

The reinterpret signature changed from 0.6 to 1.0. You can use the following:

julia> function svectormat(x::Matrix{Float64})
           @assert size(x,1) == 3
           reinterpret(SVector{3,Float64}, x)
       end
julia> function svectorvec(x::Matrix{Float64})
           @assert size(x,1) == 3
           reshape(reinterpret(SVector{3,Float64}, x), (size(x,2),))
       end
julia> svectormat(m)
1×10 reinterpret(SArray{Tuple{3},Float64,1,3}, ::Array{Float64,2}):
 [0.0, 0.0, 0.0]  [0.0, 0.0, 0.0]  …  [0.0, 0.0, 0.0]  [0.0, 0.0, 0.0]

julia> svectorvec(m)
10-element reshape(reinterpret(SArray{Tuple{3},Float64,1,3}, ::Array{Float64,2}), 10) with eltype SArray{Tuple{3},Float64,1,3}:
 [0.0, 0.0, 0.0]
 [0.0, 0.0, 0.0]
 [0.0, 0.0, 0.0]
 [0.0, 0.0, 0.0]
 [0.0, 0.0, 0.0]
 [0.0, 0.0, 0.0]
 [0.0, 0.0, 0.0]
 [0.0, 0.0, 0.0]
 [0.0, 0.0, 0.0]
 [0.0, 0.0, 0.0]

Also see #496 for a lengthy discussion on this topic. If you need this to be fast on 1.0 and don't mind relying on internals that will probably silently break your code in 1.1, you can also use the ccall-based unsafe variant from the PR. (there currently is a penalty for the reinterpret on every single access to the resulting object. If you don't rely on mutations and can afford the memory, then you should pull a copy of the reshape-reinterpret object).

[chethega]

@andyferris It appears that this is a documentation bug. Do we need a docfix before merging some variant of #496?

[andyferris]

Yes - perhaps clearly documenting the current behavior on Julia 1.0 would be the best first step.

[dmbates]

Bump. The old, failing, function svectors is still in the documentation as of v0.11.0