[RFC FS-1031] Discussion: Allow implementing the same interface at different generic instantiations in the same type

[0x53A]

Discussion for https://github.com/fsharp/fslang-design/blob/master/FSharp-5.0/FS-1031-Allow%20implementing%20the%20same%20interface%20at%20different%20generic%20instantiations%20in%20the%20same%20type.md

[0x53A]

@dsyme Do I understand your comment correctly, that you explicitly don't want to enable the following scenario (multiple instantiations with type unknowns)?

letx={new IGet<_>withmemberx.Get()=1interface IGet<_>withmemberx.Get()="hello"interface IGet<_>withmemberx.Get()=1.}

I'm kind of torn, disallowing it would really simplify the implementation, but I think it would be nice to have.

[...] and we don't want to support this kind of inference, and equally don't want a non-orthogonality between object expressions and class definitions.

ref: dotnet/fsharp#18

[dsyme]

@dsyme Do I understand your comment correctly, that you explicitly don't want to enable the following scenario (multiple instantiations with type unknowns)?

Yes, my gut feeling was that we should require full type instantiations at these points (or else there will be trouble further down the line in processing the declarations) I may be wrong though, will look at your prototype and tests

[0x53A]

I have for now updated the RFC to explicitly disallow multiple type unknowns - should we figure out that the implementation would be trivial (which I doubt ...), it can be added again.

Besides, if the member implementations become slightly more complex than a single constant expression, it may be good design to explicitly state the type anyway.

[eiriktsarpalis]

I just noticed that the following code

typeIFoo<'T>=interfaceendletf<'Twhen'T:>IFoo<int>and'T:>IFoo<string>>()=()

fails with the type error

error FS0193: Type constraint mismatch. The type ''T' is not compatible with type
'IFoo<string>' 

The error feels somewhat related to the current generic interfaces restriction.

[0x53A]

It is currently possible to override an interface implementation from a base class:

typeIB<'a>=interfaceabstractX :unit->intendtypeCBase()=interface IB<int>withmemberx.X()=1typeC2()=inherit CBase()interface IB<int>withmemberx.X()=3

It is currently not possible to override a specific interface implementation with a generic implementation:

typeIB<'a>=interfaceabstractX :unit->intendtypeCBase()=interface IB<int>withmemberx.X()=1typeC2<'T>()=inherit CBase()interface IB<'T>withmemberx.X()=3

The equivalent is possible in C#:

interfaceI<T>{TGet();}publicclassBase:I<int>{publicintGet()=>1;}publicclassC<T>:Base,I<T>{publicTGet()=>default(T);}varc=newC<int>();>(casI<int>).Get()0

This issue is not directly related to this RFC, but I found it when testing the behavior.

@dsyme

I know that F# is not just a functional C#, so should it be considered "by-design", and a potential language suggestion, or should I move this to visualfsharp as a bugreport?

What about the behavior @eiriktsarpalis found? "By design", so a new language suggestion, or bug, so a new bugreport at visualfsharp?

[dsyme]

The behaviour mentioned by @eiriktsarpalis is by design. From the F# language spec (4.0)

A type variable cannot be constrained by two distinct instantiations of the same named type. If two
such constraints arise during constraint solving, the type instantiations are constrained to be equal.
For example, during type inference, if a type variable is constrained by both IA<int> and
IA<string>, an error occurs when the type instantiations are constrained to be equal. This
limitation is specifically necessary to simplify type inference, reduce the size of types shown to users,
and help ensure the reporting of useful error messages

[dsyme]

I know that F# is not just a functional C#, so should it be considered "by-design", and a potential language suggestion, or should I move this to visualfsharp as a bugreport?

I believe all the behaviours here are currently by-design. So the suggestion can stay open here.

[zpodlovics]

@0x53A Will it work with F# struct (value) types too? Eg.:

typeIA<'T>=abstractmemberGet :unit->obj[<Struct>]typeMyStruct=interface IA<int>withmemberx.Get()=upcast1interface IA<string>withmemberx.Get()=upcast2

How it supposed to work / interop with default interface methods (https://github.com/dotnet/csharplang/blob/master/proposals/default-interface-methods.md) ?

[Alxandr]

@zpodlovics this is a pure language feature. It will not affect anything wrt default interface methods (which can be a pure CLR feature, but probably need language support to not be super annoying). Also, I see no reason why it shouldn't work with structs.

[jackfoxy]

I just discovered this from the list of features that may make it into F# 5. I hope is does. I think this may resolve a type issue I have been struggling with. jackfoxy/DependentTypes#14

[seankearon]

This definitely gets a vote from me! I think it would be really useful to have as it increases the opportunities to use libraries written in other languages from F#.

As I understand it, there is no workaround for this other than implementing in C# and reuse from F#. This was quite a surprise as it is the first time in the 4 years or so that I have been working in F# where I have encountered an interop issue that stops implementation in F#.

My use case comes from is working with the Rebus .Net service bus. Anyone working with NServiceBus would also need this. Both of these libraries implement long-running, durable workflows for which I need to write something like this:

 type NewOrderSaga(b: IBus) =
inherit Saga<NewOrderSagaData>()
let bus = b
interface IAmInitiatedBy<OrderTaken> with
member this.Handle(m: OrderTaken) =
...
interface IHandleMessages<InvoiceGenerated> with
member this.Handle(m: InvoiceGenerated) =
...
interface IHandleMessages<OrderShipped> with
member this.Handle(m: OrderShipped) =
...