- Proposal: SE-NNNN
- Authors: Sophia Poirier, Matt Massicotte, Konrad Malawski, John McCall
- Review Manager: TBD
- Status: Awaiting review
- Implementation: On
main
gated behind-enable-experimental-feature ClosureIsolation
- Previous Proposals: SE-0313, SE-0316
- Review: (pitch)
This proposal provides the ability to explicitly specify actor-isolation or non-isolation of a closure, as well as providing a parameter attribute to guarantee that a closure parameter inherits the isolation of the context. It makes the isolation inheritance rules more uniform, helps to better express intention at closure-creation time, and also makes integrating concurrency with non-Sendable types less restrictive.
- Introduction
- Motivation
- Proposed solution
- Detailed design
- Source compatibility
- ABI compatibility
- Implications on adoption
- Alternatives considered
- Future directions
- Acknowledgments
The formal isolation of a closure can be explicitly specified as global actor isolation:
Task { @MainActor in
print("global actor isolation")
}
Without a global actor isolation annotation, actor-isolation or non-isolation of a closure is inferred but cannot be explicitly specified. This proposal enables closures to be fully explicit about all three types of formal isolation:
nonisolated
- global actor
- specific actor value
Explicit annotation has the benefit of disabling inference rules and the potential that they lead to a formal isolation that is not preferred. For example, there are circumstances where it is beneficial to guarantee that a closure is nonisolated
therefore knowing that its execution will hop off the current actor. Explicit annotation also offers the ability to identify a mismatch of intention, such as a case where the developer expected nonisolated
but inference landed on actor-isolated, and the closure is mistakenly used in an isolated context. Using explicit annotation, the developer would receive a diagnostic about a nonisolated
closure being used in an actor-isolated context which helpfully identifies this mismatch of intention.
Additionally, there is a difference in how isolation inheritance behaves via the experimental attribute @_inheritActorContext
(as used by Task.init
) for isolated parameters vs actor isolation: global actor isolation is inherited by Task
's initializer closure argument, whereas an actor-isolated parameter is not inherited. This makes it challenging to build intuition around how isolation inheritance works. It also makes it impossible to allow a non-Sendable type to create a new Task that can access self.
class NonSendableType {
@MainActor
func globalActor() {
Task {
// accessing self okay
}
}
func isolatedParameter(_ actor: isolated any Actor) {
Task {
// not okay to access self
}
}
}
Enable explicit specification of non-isolation by allowing nonisolated
to be a modifier on a closure:
Task { nonisolated in
print("nonisolated")
}
Enable explicit specification of actor-isolation via an isolated parameter in a closure's capture list by using the isolated
specifier:
actor A {
nonisolated func isolate() {
Task { [isolated self] in
print("isolated to 'self'")
}
}
}
Provide a formal replacement of the experimental parameter attribute @_inheritActorContext
to resolve its ambiguity with closure isolation. Currently, @_inheritActorContext
actual context capture behavior is conditional on whether you capture an isolated parameter or isolated capture or actor-isolated function, but unconditional if the context is isolated to a global actor or nonisolated
. Its replacement @inheritsIsolation
changes the behavior so that it unconditionally and implicitly captures the isolation context.
class Old {
public init(@_inheritActorContext operation: () async -> Void)
}
class New {
public init(@inheritsIsolation operation: () async -> Void)
}
class C {
var value = 0
@MainActor
func staticIsolation() {
Old {
value = 1 // closure is MainActor-isolated and therefore okay to access self
}
New {
value = 2 // closure is MainActor-isolated and therefore okay to access self
}
}
func dynamicIsolation(_ actor: isolated any Actor) {
Old {
// not isolated to actor without explicit capture
}
New {
// isolated to actor through guaranteed implicit capture
}
}
}
An isolated parameter in a capture list must be of actor type, or conform to or imply an actor, potentially optional, and there can only be one isolated parameter captured, following the same rules described in SE-0313 for actor-isolated parameters.
The contexts in which an isolated parameter is permitted in the capture list of a synchronous closure are when the closure is:
- called immediately
- converted to an
async
function type - converted to an
@isolated(any)
function type - converted to a non-Sendable function type and has the correct isolation for the context that does the conversion
Due to the ambiguity between the nonisolated
modifier and a type-inferred closure parameter, most notably disambiguating { nonisolated parameter in ... }
as a modifier followed by a single parameter vs both as a bound pair of tokens, the use of parentheses for a parameter list is required when nonisolated
is specified.
{ nonisolated (parameter) in ... }
Opting out of @inheritsIsolation
can be achieved by explicitly annotating the closure argument as nonisolated
.
@_inheritActorContext
is currently used by the Task
initializer in the standard library which should be updated to use @inheritsIsolation
instead.
One further related clarification of isolation inheritence is that non-@Sendable
local functions should always inherit their enclosing isolation (unless explicitly nonisolated
or isolated some other way).
isolated
capture parameter works with distributed actors, however only statically "known to be local" distributed actors may be promoted to isolated
. Currently, this is achieved only through an isolated
distributed actor type, meaning that a task can only be made isolated to a distributed actor if the value already was isolated, like this:
import Distributed
distributed actor D {
func isolateSelf() {
// 'self' is isolated
Task { [isolated self] in print("OK") } // OK: self was isolated
}
nonisolated func bad() {
// 'self' is not isolated
Task { [isolated self] in print("BAD") } // error: self was not isolated, and may be remote
}
}
func isolate(d: isolated D) {
Task { [isolated d] in print("OK") } // OK: d was isolated, thus known-to-be-local
}
func isolate(d: D) {
Task { [isolated d] in print("OK") } // error: d was not isolated, and may be remote
}
While it is technically possible to enqueue work on a remote distributed actor reference, the enqueue on such an actor will always immediately crash. Because of that, we err on the side of disallowing such illegal code. Future directions discusses how this can be made more powerful when it is known that an actor is local. It is also worth noting the da.whenLocal { isolated da in ... }
API which allows dynamically recovering an isolated distributed actor reference after it has dynamically been checked for locality.
It is possible that existing code could have a closure that names a type-inferred parameter nonisolated
:
{ nonisolated in print(nonisolated) }
but with this proposed change, nonisolated
in this case would instead be interpreted as the contextual keyword specifying the formal isolation of the closure. Such code would then result in a compilation error when trying to use a parameter named nonisolated
.
The change to Task.init
in the standard library does have the potential to isolate some closures that previously were inferred to be nonisolated
. Prior behavior in those cases could be restored, if desired, by explicitly declaring the closure as nonisolated
.
It is worth noting that this does not affect the isolation semantics for actor-isolated types that make use of isolated parameters. It is currently impossible to access self in these cases, and even with this new inheritance rule that remains true.
actor MyActor {
var mutableState = 0
func isolatedParameter(_ actor: isolated any Actor) {
self.mutableState += 1 // invalid
Task {
self.mutableState += 1 // invalid
}
}
}
@MainActor
class MyClass {
var mutableState = 0
func isolatedParameter(_ actor: isolated any Actor) {
self.mutableState += 1 // invalid
Task {
self.mutableState += 1 // invalid
}
}
}
The language change does not add or affect ABI since formal isolation is already part of a closure's type regardless of whether it is explicitly specified. The Task.init
change does not impact ABI since the function is annotated with @_alwaysEmitIntoClient
and therefore has no ABI.
none
@nonisolated
in attribute form was considered to avert the potential for source breakage, but requires an unintuitive inconsistency in the language for when @
is required vs needs to be avoided.
One alternative to @inheritsIsolation
is to not use Task
in combination with non-Sendable types in this way, restructuring the code to avoid needing to rely on isolation inheritance in the first place.
class NonSendableType {
private var internalState = 0
func doSomeStuff(isolatedTo actor: isolated any Actor) async throws {
try await Task.sleep(for: .seconds(1))
print(self.internalState)
}
}
Despite this being a useful pattern, it does not address the underlying inheritance semantic differences.
There has also been discussion about the ability to make synchronous methods on actors. The scope of such a change is much larger than what is covered here and would still not address the underlying differences.
Explore support for explicitly isolated
closure captures to additionally be specified as weak
.
Distributed actors have a property that is currently not exposed in the type system that is "known to be local". If a distributed actor is known to be local, code may become isolated to it.
Once the locality of a type is expressed in the type system, the following would become possible:
let worker: local Worker
// silly example, showcasing isolating on a known-to-be-local distributed actor
func work(item: Item) async {
await Task { [isolated worker] in
worker.work(on: item)
}.value
}
Thank you to Franz Busch and Aron Lindberg for looking at the underlying problem so closely and suggesting alternatives. Thank you to Holly Borla for helping to clarify the current behavior, as well as suggesting a path forward that resulted in a much simpler and less-invasive change.