[Need advice] Would we need a new universal protocol if we add flexibly-sized types?

Evolution Discussion
, ,
1

This thought came up while thinking on how fixed-sized arrays from my last proposal could be implemented.

A flexibly-sized type is:

  • A fixed-sized array with at least one extent deferred until run-time
  • A fixed-sized array that uses a flexibly-sized element type, even if the top-level array uses only compile-time extents
  • A tuple that has at least one member of a flexibly-sized type.

Said types can be used for:

  • top-level objects (code-block locals, globals, type-level stored properties)
  • function parameters
  • stored properties for classes

They cannot be used for:

  • function returns
  • stored properties for structs
  • computed properties
  • as part (or all) of an enum case's payload

So, what's supposed to happen when a struct with a stored property based off a generic parameter gets a flexibly-sized type:

struct MyStruct<T> {
    var hello: T
}

var world: MyStruct<[_ ; Int]>  // I guess this chokes?

We would have to either just let the user take their chances, or maybe allow some sort of protocol to distinguish flexibly-sized types from conventional ones so designers can ban flexibly-sized types from constructs that would choke on them.

Maybe they could be named like:

  • AnyFlexible
  • AnySizeLocked

But, if you think about it, inflexible types can be treated as flexibly-sized types, but not the other way. In other words, inflexible type are subtypes of flexibly-sized types. So either:

  • Make Any for both flexibly-sized and inflexible types, and add AnySizeLocked for inflexible types
  • Make Any for inflexible types and add AnyFlexible for flexibly-sized types.

I'm leaning towards the second. This breaks Any being the universal protocol (in favor of AnyFlexible), but being able to plug-in types into generic parameters that will end up choking in use in code is worse. Flexibly-sized type are coming in second, so using Any for inflexible types keeps backwards source compatibility. Note that the default for unadorned generic parameters would still be Any; AnyFlexible always has to explicitly listed, even though it's higher up the protocol hierarchy.

Would this be the right decision?

1 Like