[AST/SIL] Expand SIL function type isolation to include `nonisolated(nonsending)` by xedin · Pull Request #88404 · swiftlang/swift

xedin · 2026-04-09T23:07:58Z

Explanation:

The primary driver for this change is the OptimizeHopToExecutor pass because it has to check whether a call is a suspension point or not and the most convenient way to rule out nonisolated(nonsending) would be to check a callee but since the nonisolated(nonsending) isolation wasn't represented before it had to go through actor isolation which is not always easier to figure out from SIL. This means that some hops were optimized away even though they shouldn't have been.

The change includes the following:
- Adding nonisolated(nonsending) to SILFunctionType;
- Expanding mangling to support that (uses N to represent the new isolation);
- Serialization of SILFunctionTypeLayout records has been expanded to store "isolation" instead of just "erased" bit;
- getSILFunctionType is updated to start producing lowering with nonisolated(nonsending) isolation;
- FullyApplySite::isCallerIsolationInheriting is updated to check callee instead of reaching of the isolation;
Resolves: [concurrency]: nonisolated(nonsending) can unexpectedly lose actor isolation #86083, rdar://169998321
Risk: Low to Medium. This is a representation change so it carries some risk of interactions with optimizations beyond specialization.
Testing: Added new test-cases to the suite, updated existing tests.

xedin · 2026-04-09T23:36:09Z

xedin · 2026-04-10T05:23:08Z

Currently there is only `erased` but this is soon to gain one more - `nonisolated(nonsending)`.

…ation Similarly to `@isolated(any)` it's sometimes necessary to check whether a call is to a `nonisolated(nonsending)` function value i.e. for hop elimination pass and without this isolation support it has to go trying to figure out the isolation from a callee which is not possible if it doesn't refer to a `SILFunction *`.

…on in SIL function types

Generalizes how isolation is stored for SILFunctionType and makes it possible to store not just "erased" but any future isolation we'd need to add as well.

…ation If a function type requires an implicit isolation parameter also set `nonisolated(nonsending)` isolation on it. This helps us track all of the references to `nonisolated(nonsending)` functions.

xedin · 2026-04-11T00:14:07Z

@swift-ci please test

jamieQ

A couple general interest questions for you if you have a moment @xedin:

Why is this approach preferable over a narrower fix to just address the issue in the executor hop optimization pass?
I assume the answer is "no", but is this expected to resolve the other apparent issue with that pass described in #88259?

jamieQ · 2026-04-11T15:04:38Z

lib/AST/ASTContext.cpp

+  // Make sure that the invariant that `nonisolated(nonsending)` function
+  // always has an implicit leading parameter is maintained.


Not sure if it should be done in this PR, but do you think the verifier should enforce that the actor isolation and implicit leading parameters correspond in the expected way? Something like <is_nonisolated_nonsending_isolation> <-> <has_implicit_leading_isolated_param>? Or are there expected cases in which that does not hold?

Yeah, I need to think about that some more that’s why I put an assert here for now.

jamieQ · 2026-04-11T15:11:28Z

lib/Demangling/Remangler.cpp

        Buffer << 'A';
        break;
+      case Node::Kind::ImplNonisolatedNonsendingIsolation:
+        Buffer << 'N';


Just out of curiosity, how are the letters for this encoding chosen?

To minimize possibility of clashes but really arbitrarily, I used the first letter of the attribute and checked where else we use it.

xedin · 2026-04-11T18:18:48Z

@swift-ci please test Windows platform

xedin · 2026-04-11T18:20:40Z

@jamieQ the reason why I did it this way is because there is no reliable way to make it narrow, we should track the isolation.

rjmccall

Generally looks good, except I don't think we want the mangling suppression flag.

include/swift/AST/ASTMangler.h

xedin · 2026-04-13T20:56:15Z

@swift-ci please test

lib/Demangling/NodePrinter.cpp

rjmccall

Approving because the patch seems overall fine, but I think we probably want to fix the printer.

…g)` isolation of SIL function type `nonisolated(nonsending)` is a modifier and it's really hard to work with in SIL because most of the things are represented as attributes there.

…allee Instead of trying to from isolation which is not always possible because that method replies on having a `SILFunction *` being called which is not the same for i.e. function values.

…implicit leading isolation parameter Doing so changes isolation because the invariant is that if the function is `@caller_isolated` it should have an implicit leading parameter.

…ted(nonsending)` mangling

xedin · 2026-04-13T21:30:24Z

@swift-ci please test

xedin · 2026-04-14T21:29:17Z

@swift-ci please smoke test macOS platform

xedin · 2026-04-15T00:54:35Z

@swift-ci please smoke test macOS platform

xedin · 2026-04-15T04:18:57Z

@swift-ci please smoke test macOS platform

xedin requested a review from rjmccall April 9, 2026 23:07

xedin requested a review from a team as a code owner April 9, 2026 23:07

xedin force-pushed the rdar-169998321 branch from 935a1d3 to db1d2ae Compare April 10, 2026 05:22

xedin added 5 commits April 10, 2026 09:35

[Mangling] Represent ImplFunctionType isolation as an enum

441c404

Currently there is only `erased` but this is soon to gain one more - `nonisolated(nonsending)`.

[Mangling/SIL] Mangling support for nonisolated(nonsending) isolati…

aa59c1a

…on in SIL function types

[Serialization] Store isolation in SILFunctionType records

5ded52d

Generalizes how isolation is stored for SILFunctionType and makes it possible to store not just "erased" but any future isolation we'd need to add as well.

[SIL] Teach getSILFunctionType about nonisolated(nonsending) isol…

25c339d

…ation If a function type requires an implicit isolation parameter also set `nonisolated(nonsending)` isolation on it. This helps us track all of the references to `nonisolated(nonsending)` functions.

xedin force-pushed the rdar-169998321 branch from db1d2ae to b2ba42a Compare April 10, 2026 22:23

jamieQ reviewed Apr 11, 2026

View reviewed changes

xedin mentioned this pull request Apr 12, 2026

[SILOptimizer]: fix an erroneous hop_to_executor removal #88285

Open

rjmccall requested changes Apr 13, 2026

View reviewed changes

include/swift/AST/ASTMangler.h Outdated Show resolved Hide resolved

rjmccall reviewed Apr 13, 2026

View reviewed changes

lib/Demangling/NodePrinter.cpp Outdated Show resolved Hide resolved

rjmccall approved these changes Apr 13, 2026

View reviewed changes

xedin added 5 commits April 13, 2026 14:18

[Frontend] Use @caller_isolated to represent `nonisolated(nonsendin…

5b6a29f

…g)` isolation of SIL function type `nonisolated(nonsending)` is a modifier and it's really hard to work with in SIL because most of the things are represented as attributes there.

[SIL] Adjust FullyApplySite::isCallerIsolationInheriting to check c…

6eafb85

…allee Instead of trying to from isolation which is not always possible because that method replies on having a `SILFunction *` being called which is not the same for i.e. function values.

[Tests] NFC: Add an interpreter test for swiftlang#86083

1cc3d53

[SILOptimizer] Prevent function signature optimization from removing …

1f8c8e1

…implicit leading isolation parameter Doing so changes isolation because the invariant is that if the function is `@caller_isolated` it should have an implicit leading parameter.

[ASTMangler] SILFunctionType: Remove the flag that disables `nonisola…

b7aa9f0

…ted(nonsending)` mangling

xedin force-pushed the rdar-169998321 branch from 377800f to b7aa9f0 Compare April 13, 2026 21:30

hborla approved these changes Apr 14, 2026

View reviewed changes

xedin mentioned this pull request Apr 14, 2026

[SILOptimizer] Make sure that cloning also transfers the actor isolation #88460

Open

		// Make sure that the invariant that `nonisolated(nonsending)` function
		// always has an implicit leading parameter is maintained.

Conversation

xedin commented Apr 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

xedin commented Apr 9, 2026

Uh oh!

xedin commented Apr 10, 2026

Uh oh!

xedin commented Apr 11, 2026

Uh oh!

jamieQ left a comment

Choose a reason for hiding this comment

Uh oh!

jamieQ Apr 11, 2026

Choose a reason for hiding this comment

Uh oh!

xedin Apr 11, 2026

Choose a reason for hiding this comment

Uh oh!

jamieQ Apr 11, 2026

Choose a reason for hiding this comment

Uh oh!

xedin Apr 11, 2026

Choose a reason for hiding this comment

Uh oh!

xedin commented Apr 11, 2026

Uh oh!

xedin commented Apr 11, 2026

Uh oh!

rjmccall left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

xedin commented Apr 13, 2026

Uh oh!

Uh oh!

rjmccall left a comment

Choose a reason for hiding this comment

Uh oh!

xedin commented Apr 13, 2026

Uh oh!

xedin commented Apr 14, 2026

Uh oh!

xedin commented Apr 15, 2026

Uh oh!

xedin commented Apr 15, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

xedin commented Apr 9, 2026 •

edited

Loading