[Proposal] More Powerful Constraints for Associated Types

(Thorsten Seitz) #1

To me it reads as a constraint on R. Otherwise we would have to write `protocol R where … : Q where …, S where … { … }` which would not only be confusing but would not make much sense IMHO because I want to be able to write constraints which combine associated types from R, Q and S, i.e. from all sources contributing to R.

My early post was poorly worded. It reads as a constraint on the associated types R inherits (from Q in this example). You would be able to relate any inherited associated types to each other here if desired. Any new associated types introduced by R should be constrained where they are introduced.

Ah, yes, you are right, of course!

New associated types introduced by R do have their own where clause, so that would be the proper place to constrain them.

Answering Doug’s question: I’d like the name lookup to start in R and to be able to refer to an associated type newly defined in R.

This seems confusing to me. The associated type has not been introduced until inside the body of R. It doesn't make sense to allow them to be constrained prior to introduction and it offers no additional functionality. Separating the constraints on inherited associated types from the constraints on newly introduced associated types adds clarity and readability IMO. Lookup should only consider inherited protocols.

I totally agree.




Am 05. Mai 2016 um 14:59 schrieb Matthew Johnson <matthew@anandabits.com>:
Sent from my iPad
On May 5, 2016, at 7:30 AM, Thorsten Seitz <tseitz42@icloud.com> wrote:

Am 03.05.2016 um 14:03 schrieb Matthew Johnson via swift-evolution <swift-evolution@swift.org>:

Sent from my iPad

On May 3, 2016, at 3:37 AM, Douglas Gregor via swift-evolution <swift-evolution@swift.org> wrote:

Sent from my iPhone

On May 2, 2016, at 3:50 PM, David Hart <david@hartbit.com> wrote:

Hi Doug,

In the latest version of the proposal, which is now linked to a pull request, I mentioned in the Detail Design section that the following syntax be valid:

protocol R : Q where AssocType : P {
// …

Can you read through that part of the proposal and let me know if it is descriptive enough?

I think you need to clarify the intended name lookup semantics. Does name lookup for "AssocType" start its lookup in R or in Q? If the former, can one refer to an associated type newly defined in R?

To me this syntax reads as a constraint on Q only. If we need a constraint on associated types defined in R the constraint should be attached to the definition.

- Doug


On 26 Apr 2016, at 05:28, Douglas Gregor <dgregor@apple.com> wrote:

On Apr 24, 2016, at 1:34 PM, David Hart via swift-evolution <swift-evolution@swift.org> wrote:

I wrote the proposal which was discussed to introduce generic constraints for associated types. I’d like to get some feedback on it and get it ready before submitting it:

More Powerful Constraints for Associated Types
Proposal: SE-XXXX

Author(s): David Hart

Status: TBD
Review manager: TBD

This proposal seeks to introduce a where expression to associated types declarations to bring the same expressive power as generic type constraints.

This proposal was discussed on the Swift Evolution list in the [swift-evolution] [Completing Generics] Arbitrary requirements in protocols thread.

Believe it or not, I support this direction…


Currently, associated type declarations can only express simple inheritance constraints and not the more sophisticated constraints available to generic types with the where expression. Some designs, including many in the Standard Library, require more powerful constraints for associated types to be truly elegant. For example, the SequenceType protocol can be declared as follows:

protocol Sequence {
    associatedtype Iterator : IteratorProtocol
    associatedtype SubSequence : Sequence where SubSequence.Iterator.Element == Iterator.Element

Detail Design

With this proposal, the grammar for protocols associated types would be modified to:

protocol-associated-type-declaration → attributesopt access-level-modifieropt associatedtype typealias-name ­type-inheritance-clause­opt­ typealias-assignment­opt requirement-clauseopt

The new requirement-clause is then used by the compiler to validate the associated types of conforming types.

The only thing that bothers me about this syntax is that I have to introduce an associated type to add requirements. For example, what if I want my inheriting protocol to add a requirement to an existing associated type?

protocol P { }

protocol Q {
typealias AssocType

protocol R : Q {
// I want to just add “AssocType : P”, but I have to redeclare AssocType to do so
typealias AssocType where AssocType : P

Did you consider an alternate syntax that puts the where clause outside the braces, e.g.,

protocol R : Q where AssocType : P {
// …

There are two things I like about this. First, it breaks the unnecessary link between an associated type and a (possibly unrelated) where clause, eliminating the need to redeclare associated types in inheriting protocols. Second, it’s effectively the same syntax as constrained extensions, which have a similar feel.

Note that, if we do the above, I’d love to make it an error to define a new associated type with the same name as an associated type in an inherited protocol. It’s odd that we do so, and IIRC the only use case for it is to add requirement to an “existing” associated type.

- Doug

swift-evolution mailing list

swift-evolution mailing list