reduce with dims argument over empty arrays

[simonbyrne]

Currently this is a bit inconsistent:

julia> a = zeros(0,0)
0×0 Array{Float64,2}

julia> sum(a,dims=1)
1×0 Array{Float64,2}

julia> minimum(a,dims=1)
ERROR: ArgumentError: reducing over an empty collection is not allowed
Stacktrace:
 [1] _empty_reduce_error() at ./reduce.jl:216
 [2] reducedim_init(::Function, ::typeof(min), ::Array{Float64,2}, ::Int64) at ./reducedim.jl:135
 [3] _mapreduce_dim at ./reducedim.jl:313 [inlined]
 [4] #mapreduce#543 at ./reducedim.jl:304 [inlined]
 [5] #mapreduce at ./none:0 [inlined]
 [6] _minimum at ./reducedim.jl:675 [inlined]
 [7] _minimum at ./reducedim.jl:674 [inlined]
 [8] #minimum#551 at ./reducedim.jl:648 [inlined]
 [9] (::getfield(Base, Symbol("#kw##minimum")))(::NamedTuple{(:dims,),Tuple{Int64}}, ::typeof(minimum), ::Array{Float64,2}) at ./none:0
 [10] top-level scope at REPL[10]:1

Basically it comes down to whether or not Base.reduce_empty(op, eltype(a)) is defined. Is it possible to support this more generally?

[simonbyrne]

Maybe this isn't possible, since it we would need a value if the non-reduced dimension is non-zero:

julia> a = zeros(0,3)
0×3 Array{Float64,2}

julia> sum(a,dims=1)
1×3 Array{Float64,2}:
 0.0  0.0  0.0

[laborg]

The init keyword has been added in #36188 for this reason.