`sum` with dims kwarg is incorrect on Any array when missings are present

[cstjean]

On Julia 1.10.4

julia> sum(Any[0.0 1; 0.0 missing], dims=2)
2×1 Matrix{Missing}:
 missing
 missing

whereas

julia> sum([0.0 1; 0.0 missing], dims=2)
2×1 Matrix{Union{Missing, Float64}}:
 1.0
  missing

[adienes]

yikes

[adienes]

related:

julia> sum(Union{Missing, Int64}[0 1; 0 missing]; dims=2)
2×1 Matrix{Union{Missing, Int64}}:
 1
  missing

julia> sum(Union{Missing, Int32}[0 1; 0 missing]; dims=2)
ERROR: MethodError: Cannot `convert` an object of type Missing to an object of type Int64

it seems there is a more general (and older) issue on this topic here #45566

[ararslan]

In the case of Any, it appears to stem from this:

julia/base/reducedim.jl

Lines 69 to 86 in 4387b8d

	function reducedim_init(f, op::Union{typeof(+),typeof(add_sum)}, A::AbstractArray, region)
	_reducedim_init(f, op, zero, sum, A, region)
	end
	function reducedim_init(f, op::Union{typeof(*),typeof(mul_prod)}, A::AbstractArray, region)
	_reducedim_init(f, op, one, prod, A, region)
	end
	function _reducedim_init(f, op, fv, fop, A, region)
	T = _realtype(f, promote_union(eltype(A)))
	if T !== Any && applicable(zero, T)
	x = f(zero(T))
	z = op(fv(x), fv(x))
	Tr = z isa T ? T : typeof(z)
	else
	z = fv(fop(f, A))
	Tr = typeof(z)
	end
	return reducedim_initarray(A, region, z, Tr)
	end

Note in particular line 82. We're summing the entire array to compute the neutral element but the missing values poison the sum and zero(missing) is missing, so the output array gets allocated as an array with eltype Missing.

[adienes]

yup. but also the other branch is wrong too; if T !== Any && applicable(zero, T)

since we are "lucky" that zero(Union{Missing, Float64}) == 0.0 instead of missing. but if that were some other custom number type with different behavior it would also poison the initialization.