Require a method for specific subtypes only?

[BatyLeo]

I have a situation where I require a method only for some subtypes of the interface.

Would it be possible to support something like this for example?

abstract type MyType{T} end

function my_func end
function my_other_func end

@required MyType my_func(::MyType)

@required MyType{<:AbstractVector} my_other_func(::MyType)

[Seelengrab]

This can get complicated very quickly. Should my_other_func error when it encounters a MyType{Int}, for example? This is not an issue in the current implementation, because @required doesn't handle UnionAll types right now. I'm inclined to think that they fall under the "optional interface" umbrella that IMO should be a distinct interface type, but I can also see how that is unnecessary duplication.

[BatyLeo]

This can get complicated very quickly. Should my_other_func error when it encounters a MyType{Int}, for example?

In my case it shouldn't error, my_other_func is optional for MyType{Int}, it can exist but is not needed for using all implemented features of MyType. For instance, there can be a method my_feature in the package that uses my_other_func only when dispatched on a MyType{<:AbstractVector}.

Having the possibility to specify optional methods would be nice indeed.

Here is an extended example:

using RequiredInterfaces
using Test
const RI = RequiredInterfaces

abstract type MyType{T} end

function my_func end
function my_other_func end

@required MyType begin
    my_func(::MyType)
    my_other_func(::MyType)
end

function my_feature(t::MyType)
    return my_func(t)
end

function my_feature(t::MyType{<:AbstractVector})
    return my_func(t) * " " * my_other_func(t)
end

struct MySubType <: MyType{Int} end

my_func(t::MySubType) = "Hello"

struct MyOtherSubType <: MyType{Vector{Int}} end

my_func(t::MyOtherSubType) = "Hello"
my_other_func(t::MyOtherSubType) = "world"

struct MyWrongOtherSubType <: MyType{Vector{Int}} end

my_func(t::MyWrongOtherSubType) = "Hello"

The my_feature methods works as I want it to, which is nice:

my_feature(MySubType())   # returns "Hello"
my_feature(MyOtherSubType())   # returns "Hello world"
my_feature(MyWrongOtherSubType())   # raises a not implemented error, and suggests to implement my_other_func

However, check_interface_implemented does not work as I would want it to because my_other_func is considered mandatory, even for MySubType:

@test RI.check_interface_implemented(MyType, MyOtherSubType)   # passes
@test RI.check_interface_implemented(MyType, MyWrongOtherSubType)   # fails
@test RI.check_interface_implemented(MyType, MySubType)   # fails, and ideally should pass

[Seelengrab]

Yes, that is intentional - an interface cannot really have optional methods. In your example, I'd be open to add something like

abstract type MyType{T} end

@required MyType begin
    my_func(::MyType)
end

@required MyType{<:AbstractArray} begin
    my_other_func(::MyType{<:AbstractArray})    
end

or, with some syntax sugar:

@required T=MyType{<:AbstractArray} begin
    my_other_func(::T)    
end

being equivalent to

abstract type MyType end
# note how the UnionAll is resolved here by hand, and
# `MyTypeAbstractArray` inherits the interface of `MyType`
abstract type MyTypeAbstractArray <: MyType

@required MyType begin
    my_func(::MyType)
end

@required MyTypeAbstractArray begin
    my_other_func(::MyTypeAbstractArray)
end

but I wouldn't make that happen automatically and/or add my_other_func as an "optional" method to the MyType interface in general. It's not optional at all for MyType at all - it is required, and that's exactly the case when the type is <: MyType{<:AbstractArray}.

In other words, I'd say that no user of your package should ever expect or think of my_other_func as being "optional". Having that available on a type means it implements an entirely different (larger) type than can be assumed from MyType alone.

[BatyLeo]

I didn't think of using a MyTypeAbstractArray inheriting from MyType, this works quite well in my case I think! I also like your other suggestion, but it might not be needed in my case.

Thanks a lot for your help!

[Seelengrab]

Great! I'll leave this open, since support for UnionAll-style child interfaces seems intriguing.