tfuncs: Be more robust in the face of uninhabited types

[Keno]

Fixes #37943

[JeffBezanson]

What about a type like this:

struct Foo
    x
    y::Union{}
    Foo(x) = new(x)
end

[JeffBezanson]

Or similarly this, but this case should possibly be disallowed on construction:

struct Bar{T}
    x
    y::T
    Bar{T}(x) where {T} = new(x)
end
Bar{false}(1)

[Keno]

Is that legal? I guess there's nothing technically wrong with it, but boy. I guess I'll add a check to only bail if it's one of the initialized fields.

[JeffBezanson]

The first one is definitely legal. The second one is currently allowed but probably shouldn't be, since a constructable type needs a layout, and if a field type is a non-type it's impossible to say what its layout is.

[vtjnash]

There's also s.has_concrete_subtype which has already checked for conditions such as that?

It's not impossible to describe the layout, but impossible for that field to have a value. We could also treat it similar to apply_type failures, which returns the uninhabited type for the fieldtype and expand the check there (in inst_ftypes) instead of here:

julia> struct Bar{T}
           x
           y::Complex{T}
           Bar{T}(x) where {T} = new(x)
       end

julia> fieldtypes(Bar{false})
(Any, Union{})

[Keno]

I'm not entirely sure I understand what you're saying. Are you saying we should do this check when we construct the datatype and set any bad type to Union{}, in which case the existing tfunc would go through fine?

[vtjnash]

Yeah, there's no particular necessity in saying the field has type "false", when it's not even valid as a type

[Keno]

Updated as discussed.