Invalid caching behavior when dispatching malformed types

[JeffBezanson]

julia> f(::Type{T}) where T = T

julia> f(Vector.body)
ERROR: UndefVarError: `T` not defined in static parameter matching

julia> f(Vector)
ERROR: UndefVarError: `T` not defined in static parameter matching

The first error is correct, but the second call should work, and does in versions prior to 1.12.

[JeffBezanson]

Order matters:

julia> f(::Type{T}) where T = T

julia> f(Vector)
Vector (alias for Array{T, 1} where T)

julia> f(Vector.body)
Vector (alias for Array{T, 1} where T)

[JeffBezanson]

Maybe related: #52992

julia> f(::Type{<:Tuple{Vararg{A}}}) where {A} = A

julia> f(Tuple{Vararg{T}} where {T<:Int})
ERROR: UndefVarError: `A` not defined in static parameter matching

julia> f(Tuple{Vararg{Int}})
ERROR: UndefVarError: `A` not defined in static parameter matching

Does not error if the calls are reversed. Also happens on 1.11. Seems worse since there are no malformed types here.

[JeffBezanson]

BTW, we have historically had e.g. Vector == Vector.body, but I would rather that be false if we can pull it off, especially if it helps fix this issue.

[Keno]

We've had issues like this for a long time. @vtjnash had a proposal that types with free typevars be made well defined in the type system but as singletons with equality by egality and otherwise Union{} <: Vector.body <: Any. I think we should do that.

[JeffBezanson]

That sounds good to me, and IIUC indeed implies Vector != Vector.body?

[Keno]

That sounds good to me, and IIUC indeed implies Vector != Vector.body?

Yes

[nsajko]

If a newly-defined type, instead of Vector, is used for the reproducer, the issue reproduces down to Julia v0.7:

f(::Type{T}) where T = T
struct S{P} end
f(S.body)
f(S)