SparseArrays: + has type-instability for `ArbNumerics.jl` type `ArbFloat` about sparsearrays.jl

[rayegun]

For some reason Base.Broadcast.combine_eltypes returns Any for your operands. Even dense:

julia> A = rand(ArbFloat, 2)
2-element Vector{ArbFloat}:
 0.2154487845270057641299827610966
 0.09626764182552727242704068481935

julia> B = rand(ArbFloat, 2)
2-element Vector{ArbFloat}:
 0.6275663792273822768657185851517
 0.06658759509104976444220998662436

julia> Base.Broadcast.combine_eltypes(+, (A, B))
Any

julia> @which Base.Broadcast.combine_eltypes(+, (A, B))
combine_eltypes(f, args::Tuple)
     @ Base.Broadcast broadcast.jl:757

julia> Base.Broadcast.eltypes((A, B))
Tuple{ArbFloat, ArbFloat}

julia> Base.promote_typejoin_union(Base._return_type(+, ans))
Any

That last one is particularly suspicious to me, although I may be doing something wrong.

[rayegun]

julia> Base.return_types(+, Tuple{ArbFloat, ArbFloat})
2-element Vector{Any}:
 ArbFloat
 Any

Yeah this seems to be some sort of type instability in addition of ArbFloats perhaps?

[rayegun]

ArbFloat is not concrete. You need to do spzeros(ArbFloat{128}, 2). While we don't specify that SparseMatrixCSC eltypes be concrete, it's not something we really support.

In this case at least you would have to define +(x::ArbFloat, y::ArbFloat) = ... rather than +(x::ArbFloat{P}, y::ArbFloat{P}) as is done now.