More explicit representation of undef-ability of fields

[Keno]

I was griping to @JeffBezanson that the ability of non-isbits fields to be undef (i.e. throw an UndefRefError) didn't seem very first-class, especially since the definition of what isbits is changing these days. I was suggesting that we might want to have an explicit way to annotate that a field may be undef until defined, independent of whether that field is isbits or not. Possible syntax (names as always subject to bikeshed):

mutable struct foo{T}
    x::Union{Undef, T}
    foo{T}() = new()
end

# With String
@test_throws UndefRefError foo{String}().x # as before
# With Int
@test_throws UndefRefError foo{Int}().x # Instead of returning un-initialized memory as it does now

[vtjnash]

Rust has this. They use it instead of Union{} in type-parameters, since it doesn't suffer the same subtyping issues when applied to dispatch with TypeVars. Fortunately, it's also trivial to create and define:

mutable struct Never
    allow::Never
    let unconstructable end
end
convert(::Type{Union{Never, T}}, x) = convert(T, x) # and a few others, c.f. Some/Nothing

what isbits is changing these days

Not intentional. Will be fixed with my latest PR.

[Keno]

Fortunately, it's also trivial to create and define:

Not sure what you're trying to show with that example. The point would be to throw an UndefRefError on field access.

[vtjnash]

Yes, precisely.

julia> mutable struct foo{T}
           x::Union{Never, T}
           foo{T}() where {T} = new{T}()
       end

julia> foo{Int}().x
ERROR: UndefRefError: access to undefined reference

julia> foo{String}().x
ERROR: UndefRefError: access to undefined reference

[StefanKarpinski]

What about a ? on the field name?

struct Foo
    baz::Vector{Int}
    bar?::Int
end

For this baz would need to be supplied while bar would not. Of course, the other way to do this would be just to replace undef fields with Union{Nothing,Int} fields and defaulting to nothing, i.e.:

struct Foo
    baz::Vector{Int}
    bar::Union{Nothing,Int}=nothing
end

Of course, this isn't going to land in time for 1.0.

[quinnj]

I'm more in favor of @StefanKarpinski's 2nd idea of providing a syntax like Int? to mean Union{Int, Nothing} and getting rid of #undef entirely. One less kinda-form-of-nothingness! Plus we have better support for Unions all-around these days.

[nalimilan]

See also previous issue about using Union{T, Nothing} (at the time it was Union{T, Null}) at #23721.