type-intersection / union missed optimization

[vtjnash]

Seeing this surprised me a bit, since in that first Union in the out, a === b, so it's trivially degenerate:

julia> typeintersect(Tuple{typeof(\),Factorization{T},Union{Array{Complex{T},2}, Array{Complex{T},1}}} where T<:Union{Float32, Float64}, Tuple{typeof(\),LQ{TA,S} where S<:AbstractMatrix{TA},StridedVecOrMat{TB}} where TB where TA)
Tuple{typeof(\),Union{LQ{T,S} where S<:AbstractMatrix{T}, LQ{T,S} where S<:AbstractMatrix{T}},Union{Array{Complex{T},2}, Array{Complex{T},1}}} where T<:Union{Float32, Float64}

julia> ans.body.parameters[2].a === ans.body.parameters[2].b
true

[martinholters]

Note that === is misleading here:

julia> ret = typeintersect(Tuple{typeof(\),Factorization{T},Union{Array{Complex{T},2}, Array{Complex{T},1}}} where T<:Union{Float32, Float64}, Tuple{typeof(\),LQ{TA,S} where S<:AbstractMatrix{TA},StridedVecOrMat{TB}} where TB where TA)
Tuple{typeof(\),Union{LQ{T,S} where S<:AbstractMatrix{T}, LQ{T,S} where S<:AbstractMatrix{T}},Union{Array{Complex{T},2}, Array{Complex{T},1}}} where T<:Union{Float32, Float64}

julia> ret.body.parameters[2].a === ret.body.parameters[2].b
true

julia> ret.body.parameters[2].a.body === ret.body.parameters[2].b.body
false

So the Union is not quite that trivially degenerate.

It is formed at:

julia/src/subtype.c

Line 3263 in 3d04d15

jl_svecset(vec, i, jl_type_union(ts, 2));

Here, the two entries in ts are not identical (see above), and have a free typevar (T), so the simplification here does not kick in:

julia/src/jltypes.c

Lines 500 to 505 in 3d04d15

	if (temp[i] == temp[j] || temp[i] == jl_bottom_type ||
	temp[j] == (jl_value_t*)jl_any_type ||
	(!has_free && !jl_has_free_typevars(temp[j]) &&
	jl_subtype(temp[i], temp[j]))) {
	temp[i] = NULL;
	}