Dynamic Types

Goal

Interoperability with native JavaScript

Examples

Unbounded dynamic type:

fun jsFun(p: dynamic): dynamic

TODO

  • [ ] Dynamic functions?
    • [ ] what is the default return type?
    • [ ] Can we omit return expressions when the return type is dynamic?
    • [ ] Can we return Unit when return type is dynamic?
  • [ ] Dynamic classes/traits?
    • [ ] All members are implicitly dynamic
    • [ ] All types whose type constructors are marked dynamic are themselves dynamic types

Syntax

type
    : ...
    | "dynamic"
    ;

“dynamic” is a soft keyword:

  • if it occurs in a non-type context, it's an identifier
  • in a type context, when followed by a dot (except for a dot that separates a receiver type from a function/property name) or an angle bracket <, it's an identifier
  • on the left-hand-side of :: in a callable reference: dynamic::foo implies that dynamic there is a normal identifier

Typing rules

Internally, dynamic is represented as a flexible type Nothing..Any?, with the following capabilities:

  • makeNullable has no effect
  • All methods of JetType are delegated to the upper bounds, instead of lower bound

Rules:

  • dynamic is assignable to anything
  • everything is assignable to dynamic
  • dynamic variable may hold null
  • dynamic? is the same as dynamic, a warning is issued on usages of this syntactic form
  • safe calls and !! issue no warnings when called on dynamic expressions
  • lub(T, dynamic) = dynamic
  • glb(T, dynamic) = T
  • dynamic can‘t be substituted for reified parameters of function/constructor calls (this means that it’s not possible to create an array of dynamic)
  • dynamic types are forbidden on the right-hand side of is, !is, as and as? (but not as generic arguments, e.g. x is List<dynamic> is allowed)
  • dynamic can't be used as a supertype or upper bound for a type parameter
  • When it comes to overload resolution, dynamic is less specific than any other type

When there are two function available

fun foo(s: String)
fun foo(d: dynamic)

the first one is resolved whenever a matching argument is passed (because dynamic is less specific than String), i.e. both calls:

  • foo("")
  • foo(dyn) // dyn: dynamic

are resolved to the same function foo(String). This may seem counter-intuitive in the latter case, but there's no sane way around it.

Calls like foo(1) are resolved to foo(dynamic), because foo(String) does not fit the arguments.

To force the call of foo(dynamic) on any expression, one can up-cast the argument to a static type, e.g. foo(dyn as Any)

Resolution rules

If a receiver of a call is dynamic, the following resolution rules apply:

  • first, we are looking for matching members of the upper bound of the representing dynamic type (Any unless we implement bounded dynamics, see Appendix below).
  • next, we are looking for extensions declared for dynamic types (no extensions for static types are considered at this point)
  • lastly, we create a synthetic candidate that is bound to match the call (i.e. has the appropriate name, numbers of type- and value parameters, value parameters have appropriate names, if named arguments are used, etc.). Notes:
    • Augmented assignments on dynamic receivers (e.g. dyn += foo) are resolved to plusAssign() function, not plus, for generality: this permits calling them on vals (e.g. those holding collection-like objects)
    • The invoke convention is limited so that for calls like dyn.foo() we do not look for property foo that has invoke defined on it (same for other cases like +dyn etc)
  • dynamic candidates with no explicit dispatch receiver are discriminated against all other candidate, i.e. for a call foo(), we first try to match static candidates, and only then dynamic ones.

NOTE: we do not even try to resolve extensions declared for static types if the receiver is dynamic. As a workaround, one may use an upcast to a static type: (dyn as Foo).extensionForFoo().

Motivation: otherwise, any extension to any type that simply happens to be in scope and match the name and arguments will be bound for a call with a dynamic receiver, i.e. there's no way to force a call to be dynamic, and in the case of a *-import the code may change its semantics just because somebody added some extension in another file.

  • This means that an extension to a normal, non-dynamic type can not be called on a dynamic receiver without an upcast.

Type Argument Inference

When expected type of a call is dynamic, it does not automatically provide type arguments for nested calls. Example:

fun foo(d: dynamic) {...}

foo(listOf()) // can't determine T for listOf<T>()

Discussion:

  • we could tweak inference so that it takes dynamic as a bound for all type variables whose containing type has a dynamic bound, but it‘s hard to be sure it’s worth the while
  • one relevant case is passing lambdas to dynamic calls: we could make their arguments have dynamic types without declaration

Notes

  • dynamic types are not supported on the JVM back-end

Appendix. Prospect on bounded dynamic types

(not to be implemented now)

A bounded dynamic type dynamic B is represented as (Nothing .. B?).

Calls on such receivers are resolved statically against members of B, and dynamically against non-members of B (including extensions).

NOTE: this is an issue: some users would expect extensions to be bound statically, but we can't allow it, because otherwise a dynamic call with a name clashing with a name of an extension to B is impossible. Options:

  • bind extensions to B (i.e. extensions to Any for dynamic) statically, this leads to unexpected changes in semantics when a new extension is added in a *-imported package. Then, to make the dynamic calls possible, provide some sort of an intrinsic extension, e.g. dynamic) that takes a string for a name and a varargs of parameters of type dynamic. Thus, to call a recv.foo(a, b) as a dynamic call, we can always say recv.dynamic("foo", a, b).
  • never bind extensions statically on dynamic receivers, allow calling them passing the receiver as the first parameter, so that we can call foo(a) instead of a.foo(). This poses no risk of accidentally changing semantics of some calls from dynamic to static

Assignability rules:

  • any subtype of B? can be passed where dynamic B is expected
  • dynamic B can be passed where any supertype of B or subtype of B, but not a type unrelated to B is expected

Unbounded dynamic is the same as dynamic Any.