Run `GlobalSnapshotManager` on `trampolineDispatcher` by MatkovIvan · Pull Request #3175 · JetBrains/compose-multiplatform-core

Ivan Matkov (MatkovIvan) · 2026-07-01T18:16:13Z

CMP-10397 Compose drops frame on iOS when dragging
CMP-10411 Register GlobalSnapshotManager on the trampoline dispatcher

Supersedes #3171

It registers GlobalSnapshotManager on the trampoline dispatcher, so a scheduled global apply becomes just another trampoline task — and the flush drains in a loop that keeps polling until the queue is empty. That means an apply notification enqueued by running an earlier task (or a write made during pointer dispatch) is picked up in the same synchronous flush, instead of leaking to the next frame.

Release Notes

Fixes - iOS

(prerelease fix) Fix frame drops when dragging scrollable content

Andrei Salavei (ASalavei) · 2026-07-02T08:53:16Z

            "FrameRecomposer requires a ContinuationInterceptor in its coroutineContext"
        }
+        val hostDispatcher = coroutineContext[ContinuationInterceptor] as? CoroutineDispatcher
+        if (hostDispatcher != null && !hostDispatcher.isDispatchNeeded(hostDispatcher)) {


In what cases it's needed? Is it a temporary solution?
Could you please add comment there?

Not temporary. For things like Dispatchers.Unconfined (default in ImageComposeScene) or Dispatchers.Main.immediate,

Also, it's not CoroutineDispatcher in our test framework

Please add comment then... because it looks a bit weird (or maybe update constructor params to make the usage clear.

But it already has comments about that (on the next line), so I'm not sure what else to add here. Any particular suggestions?

For the "public" parameters - it doesn't seem that adjustment is required, it's an implementation detail, and previous behavior is still here

Reading the code, it looks like it just skipping the flush. But if we just use FlushCoroutineDispatcher, the flush shouldn't do anything, because the dispatcher should perform the task during the dispatch.

If I am correct, and behavior is the same, I would prefer the old code - it looks cleaner and understandable. If I am not correct - can you explain what difference does this code make?

If FlushCoroutineDispatcher doesn't work correctly with Dispatchers.Unconfined - it looks like a bug of FlushCoroutineDispatcher that we should fix

The issue here is that GSM doesn't work with Dispatchers.Unconfined. Register GSM on FlushCoroutineDispatcher that inlines the block leads to a deadlock. This particular behavior is explicitly marked as unsupported in compose:runtime, so it's about ensuring that

No GSM on FlushCoroutineDispatcher

No inlining in FlushCoroutineDispatcher if it reports isDispatchNeeded == true
Currently these checks and behaviors made explicit and in one place (not across 3 classes)

Creating FlushCoroutineDispatcher that has only one function - flush in cases where it does not work does not seem as readability improvement, but vice versa - it makes it quite tricky to understand what's going on there.

If we do not want such a cleanup and explicitness, it might be revered back, but do we?

Reserved most of it for now and replaced by delegation implementation of FlushCoroutineDispatcher.isDispatchNeeded

Igor Demin (igordmn) · 2026-07-02T10:58:06Z

            "FrameRecomposer requires a ContinuationInterceptor in its coroutineContext"
        }
+        val hostDispatcher = coroutineContext[ContinuationInterceptor] as? CoroutineDispatcher
+        if (hostDispatcher != null && !hostDispatcher.isDispatchNeeded(hostDispatcher)) {


Reading the code, it looks like it just skipping the flush. But if we just use FlushCoroutineDispatcher, the flush shouldn't do anything, because the dispatcher should perform the task during the dispatch.

If I am correct, and behavior is the same, I would prefer the old code - it looks cleaner and understandable. If I am not correct - can you explain what difference does this code make?

This reverts commit 96897e9.

…alidation count

…rceptor` in tests

Ivan Matkov (MatkovIvan) · 2026-07-02T15:45:58Z

rebased on jb-main + reverted #3171

Alexander Maryanovsky (m-sasha) · 2026-07-03T09:51:07Z

    private class Registration(val dispatcher: CoroutineDispatcher) {
        /** Number of live handles. Guarded by [GlobalSnapshotManager.lock]. */
        var refCount = 0
+        var threadId: Long? = null


Should probably be @Volatile.

Alexander Maryanovsky (m-sasha) · 2026-07-03T09:54:21Z

+                threadId = getCurrentThreadId()
+                warnIfMultipleThreads()


You could add this also inside the channel.consumeEach to catch multi-thread dispatchers (like Dispatchers.Default).

I'm avoiding this for a few reasons:

No guarantee that it will be printed once without more complex tracking

.distinct() is not so cheap to run it on each snapshot writing

Alexander Maryanovsky (m-sasha) · 2026-07-03T09:55:48Z

    }

+    private fun warnIfMultipleThreads() = synchronized(lock) {
+        if (registrations.values.mapNotNull { it.threadId }.distinct().size > 1) {


This could be much faster.
It's ok if it's only done on registration, but if you also call it on each value in channel, it should be rewritten.

I don't really want to store additional sets with complicated logic of clearing this. One time shot is enough for prinitting a warning purposes

Alexander Maryanovsky (m-sasha) · 2026-07-03T09:57:07Z

It would help to explain what is a/the "trampoline dispatcher".

Ivan Matkov (MatkovIvan) · 2026-07-03T10:01:26Z

It would help to explain what is a/the "trampoline dispatcher".

It's in KDoc to it:

compose-multiplatform-core/compose/ui/ui/src/skikoMain/kotlin/androidx/compose/ui/platform/FrameRecomposer.skiko.kt

Lines 66 to 73 in 6434d1e

    
               /** 
        
                * Trampoline queue (Android's `toRunTrampolined`): 
        
                *   - Coroutine dispatch 
        
                *   - Composition effects 
        
                *   - Scheduled apply notifications 
        
                * Rolled synchronously by [performTrampolineDispatch]. 
        
                */ 
        
               private val trampolineDispatcher = FlushCoroutineDispatcher(coroutineScope)

Ivan Matkov (MatkovIvan) requested review from Andrei Salavei (ASalavei) and Igor Demin (igordmn) July 1, 2026 18:16

Ivan Matkov (MatkovIvan) force-pushed the ivan.matkov/gsm-on-trampoline branch 3 times, most recently from 567510e to 41293b5 Compare July 1, 2026 22:55

Ivan Matkov (MatkovIvan) marked this pull request as ready for review July 1, 2026 22:56

Andrei Salavei (ASalavei) reviewed Jul 2, 2026

View reviewed changes

Comment thread compose/ui/ui/src/skikoTest/kotlin/androidx/compose/ui/scene/BaseComposeSceneTest.kt

Andrei Salavei (ASalavei) reviewed Jul 2, 2026

View reviewed changes

Ivan Matkov (MatkovIvan) requested a review from Alexander Maryanovsky (m-sasha) July 2, 2026 09:00

Andrei Salavei (ASalavei) reviewed Jul 2, 2026

View reviewed changes

Comment thread ...se/ui/ui/src/skikoMain/kotlin/androidx/compose/ui/platform/FlushCoroutineDispatcher.skiko.kt Outdated

Igor Demin (igordmn) reviewed Jul 2, 2026

View reviewed changes

Comment thread ...se/ui/ui/src/skikoMain/kotlin/androidx/compose/ui/platform/FlushCoroutineDispatcher.skiko.kt Outdated

Ivan Matkov (MatkovIvan) force-pushed the ivan.matkov/gsm-on-trampoline branch 2 times, most recently from a89f262 to 205f010 Compare July 2, 2026 11:26

Ivan Matkov (MatkovIvan) marked this pull request as draft July 2, 2026 11:29

Ivan Matkov (MatkovIvan) force-pushed the ivan.matkov/gsm-on-trampoline branch from 205f010 to 8bd66ac Compare July 2, 2026 11:46

Ivan Matkov (MatkovIvan) marked this pull request as ready for review July 2, 2026 11:52

Ivan Matkov (MatkovIvan) requested review from Andrei Salavei (ASalavei) and Igor Demin (igordmn) July 2, 2026 11:52

Igor Demin (igordmn) reviewed Jul 2, 2026

View reviewed changes

Comment thread compose/ui/ui/src/skikoMain/kotlin/androidx/compose/ui/platform/FrameRecomposer.skiko.kt Outdated

Igor Demin (igordmn) approved these changes Jul 2, 2026

View reviewed changes

Ivan Matkov (MatkovIvan) added 5 commits July 2, 2026 17:45

Revert "Fix frame drops when dragging scrollable content (#3171)"

07e95ab

This reverts commit 96897e9.

Run GlobalSnapshotManager on trampolineDispatcher

9426882

Add runComposeSceneTest helper and ComposeSceneTestScope to track inv…

5e149e2

…alidation count

Use real GlobalSnapshotManager instead of `ApplyingContinuationInte…

cc49564

…rceptor` in tests

Fix leaked ComposeContainer in tests (causes GSM warnings)

0c01547

Ivan Matkov (MatkovIvan) force-pushed the ivan.matkov/gsm-on-trampoline branch from 0a151ca to 0c01547 Compare July 2, 2026 15:45

Alexander Maryanovsky (m-sasha) reviewed Jul 3, 2026

View reviewed changes

Address feedback

4cd43b2

Uh oh!

Conversation

Ivan Matkov (MatkovIvan) commented Jul 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Release Notes

Fixes - iOS

Uh oh!

Uh oh!

Andrei Salavei (ASalavei) Jul 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Ivan Matkov (MatkovIvan) commented Jul 2, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Alexander Maryanovsky (m-sasha) commented Jul 3, 2026

Uh oh!

Ivan Matkov (MatkovIvan) commented Jul 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Ivan Matkov (MatkovIvan) commented Jul 1, 2026 •

edited

Loading

Andrei Salavei (ASalavei) Jul 2, 2026 •

edited

Loading