Bug in subtyping resolution

[olynch]

This bug is succinctly described by the following code:

const NullableVec{T} = Union{AbstractVector{T}, AbstractVector{Union{T,Nothing}}}

struct NullableVecWrapper{T, V<:NullableVec{T}}
  v::V
end

abstract type P{A} end

NV = NullableVecWrapper{P, Vector{Union{P,Nothing}}}

NV <: NullableVecWrapper
> false

My julia version is 1.8.3, and I installed it through nix: https://github.com/olynch/scientific-fhs. Julia 1.7 and 1.6 do not have this bug.

[brenhinkeller]

Can reproduce on 1.8.3 and 1.9.0-alpha1 official binaries

[vtjnash]

You are allowed to create types that are not subtypes of the original, as long as they remain abstract. What remains interesting about this example though is that it remains reproducible after considerable simplification:

julia> abstract type P{A} end

julia> struct P2{A} end

julia> struct NullableVecWrapper{T, V<:Vector{Union{T,Nothing}}} end

julia> NullableVecWrapper{P} <: NullableVecWrapper # or AbstractRange, Ref, etc.
false

julia> NullableVecWrapper{P2} <: NullableVecWrapper # or Complex, Rational, Val, etc.
false

[N5N3]

The above looks quite similar to the example in #45874 (comment)

[jpfairbanks]

I just tested on 1.9.0-beta2 and it still reproduces.

[vtjnash]

@JeffBezanson This returns the right answer now. Can you close it?

[brenhinkeller]

Can confirm the fix on nightly!

julia> abstract type P{A} end

julia> struct P2{A} end

julia> struct NullableVecWrapper{T, V<:Vector{Union{T,Nothing}}} end

julia> NullableVecWrapper{P} <: NullableVecWrapper # or AbstractRange, Ref, etc.
true

julia> NullableVecWrapper{P2} <: NullableVecWrapper # or Complex, Rational, Val, etc.
true

[oscardssmith]

Should we add a test?