Broadcasting breakage for non-1 indices

[timholy]

julia> using OffsetArrays

julia> a = OffsetArray(1:3, -1:1)
OffsetArrays.OffsetArray{Int64,1,UnitRange{Int64}} with indices -1:1:
 1
 2
 3

julia> Float64.(a)
ERROR: MethodError: Cannot `convert` an object of type Tuple{UnitRange{Int64}} to an object of type Array{Float64,N} where N
This may have arisen from a call to the constructor Array{Float64,N} where N(...),
since type constructors fall back to convert methods.
Stacktrace:
 [1] Array{Float64,N} where N(::Tuple{UnitRange{Int64}}) at ./sysimg.jl:24
 [2] similar(::Type{T} where T, ::Tuple{UnitRange{Int64}}) at ./abstractarray.jl:508
 [3] broadcast_t(::Type{T} where T, ::Type{T} where T, ::Tuple{UnitRange{Int64}}, ::CartesianRange{CartesianIndex{1}}, ::OffsetArrays.OffsetArray{Int64,1,UnitRange{Int64}}) at ./broadcast.jl:265
 [4] broadcast_c(::Type{T} where T, ::Type{Array}, ::OffsetArrays.OffsetArray{Int64,1,UnitRange{Int64}}) at ./broadcast.jl:293
 [5] broadcast(::Type{T} where T, ::OffsetArrays.OffsetArray{Int64,1,UnitRange{Int64}}) at ./broadcast.jl:406
 [6] eval(::Module, ::Any) at ./boot.jl:235

Works in Julia 0.5.

AFAICT broken in #18642, d16d994. CC @pabloferz.

[pabloferz]

Hi Tim, this is a problem with similar as much as it is a problem with broadcast.

On master similar(Array{Float64}, (-1:1,)) gives the error above, but on 0.5 it calls similar(f::Union{DataType,Function}, shape::Tuple{UnitRange,Vararg{UnitRange,N<:Any}}) form OffsetArrays which returns an OffsetArray.

I'm not sure if this has to be fixed in master or in OffsetArrays, though.

[timholy]

That's interesting. For reference, the Base method is

similar(f, shape::Tuple) = f(to_shape(shape))

and the one in OffsetArrays is

Base.similar{T}(f::Union{Function,DataType}, shape::Tuple{UnitRange,Vararg{UnitRange}}) = OffsetArray(f(map(length, shape)), map(indexoffset, shape))

To me it seems the one in OffsetArrays is unambiguously more specific. So why isn't it called?

I'm going to classify this as a type system bug.

[timholy]

I can confirm that splitting the Union fixes the problem:

Base.similar{T}(::Type{T}, shape::Tuple{UnitRange,Vararg{UnitRange}}) = _similar(T, shape)
Base.similar(f::Function, shape::Tuple{UnitRange,Vararg{UnitRange}}) = _similar(f, shape)
_similar(f, shape) = OffsetArray(f(map(length, shape)), map(indexoffset, shape))

[pabloferz]

The problem is that Array{Float64} is now a UnionAll (Array{Float64,N} where N) which is not a subtype of DataType. This could be fixed in OffsetArrays by changing the signature to ::Union{Function,Type}.

[pabloferz]

See JuliaArrays/OffsetArrays.jl#20