JIT: Optimize `AsyncHelpers` await helpers and enable inlining for them by jakobbotsch · Pull Request #128528 · dotnet/runtime

jakobbotsch · 2026-05-23T22:00:14Z

Allow inlining of async functions when all awaits are tail awaits (using AsyncHelpers.TailAwait)
Rewrite all await helpers in the tail await shape to allow inlining of them
Optimize all await paths to use hand rolled continuations with caching and that avoids additional async1 continuation allocations

ValueTaskPerf benchmarks from dotnet/performance:

Method	Toolchain	Mean	Error	StdDev	Median	Min	Max	Ratio
Await_FromResult	\main\corerun.exe	12.010 ns	0.2031 ns	0.1900 ns	12.024 ns	11.730 ns	12.353 ns	1.00
Await_FromResult	\pr\corerun.exe	8.318 ns	0.0989 ns	0.0925 ns	8.347 ns	8.180 ns	8.434 ns	0.69

Await_FromCompletedTask	\main\corerun.exe	16.347 ns	0.5397 ns	0.6215 ns	16.225 ns	15.546 ns	17.593 ns	1.00
Await_FromCompletedTask	\pr\corerun.exe	13.291 ns	0.2445 ns	0.2168 ns	13.230 ns	13.030 ns	13.715 ns	0.81

Await_FromCompletedValueTaskSource	\main\corerun.exe	26.584 ns	2.2930 ns	2.6407 ns	25.976 ns	23.479 ns	30.889 ns	1.00
Await_FromCompletedValueTaskSource	\pr\corerun.exe	22.867 ns	0.1866 ns	0.1654 ns	22.812 ns	22.623 ns	23.281 ns	0.87

CreateAndAwait_FromResult	\main\corerun.exe	11.605 ns	0.1786 ns	0.1670 ns	11.556 ns	11.342 ns	11.898 ns	1.00
CreateAndAwait_FromResult	\pr\corerun.exe	8.446 ns	0.1681 ns	0.1726 ns	8.421 ns	8.217 ns	8.719 ns	0.73

CreateAndAwait_FromResult_ConfigureAwait	\main\corerun.exe	15.849 ns	0.3147 ns	0.2944 ns	16.007 ns	15.381 ns	16.277 ns	1.00
CreateAndAwait_FromResult_ConfigureAwait	\pr\corerun.exe	8.364 ns	0.1429 ns	0.1337 ns	8.421 ns	8.098 ns	8.544 ns	0.53

CreateAndAwait_FromCompletedTask	\main\corerun.exe	11.933 ns	0.1488 ns	0.1392 ns	11.916 ns	11.704 ns	12.190 ns	1.00
CreateAndAwait_FromCompletedTask	\pr\corerun.exe	10.023 ns	0.0853 ns	0.0798 ns	10.030 ns	9.831 ns	10.156 ns	0.84

CreateAndAwait_FromCompletedTask_ConfigureAwait	\main\corerun.exe	14.678 ns	0.3420 ns	0.3938 ns	14.517 ns	14.262 ns	15.185 ns	1.00
CreateAndAwait_FromCompletedTask_ConfigureAwait	\pr\corerun.exe	10.986 ns	0.0637 ns	0.0564 ns	10.973 ns	10.904 ns	11.091 ns	0.75

CreateAndAwait_FromCompletedValueTaskSource	\main\corerun.exe	13.822 ns	0.0354 ns	0.0331 ns	13.812 ns	13.772 ns	13.884 ns	1.00
CreateAndAwait_FromCompletedValueTaskSource	\pr\corerun.exe	14.331 ns	0.0870 ns	0.0813 ns	14.313 ns	14.235 ns	14.465 ns	1.04

CreateAndAwait_FromYieldingAsyncMethod	\main\corerun.exe	261.416 ns	3.3161 ns	3.1019 ns	261.140 ns	257.977 ns	269.236 ns	1.00
CreateAndAwait_FromYieldingAsyncMethod	\pr\corerun.exe	249.452 ns	3.6599 ns	3.0562 ns	249.537 ns	244.468 ns	254.448 ns	0.95

CreateAndAwait_FromDelayedTCS	\main\corerun.exe	1,830.124 ns	89.6522 ns	103.2436 ns	1,801.707 ns	1,615.933 ns	2,016.293 ns	1.00
CreateAndAwait_FromDelayedTCS	\pr\corerun.exe	1,003.680 ns	44.8132 ns	51.6070 ns	976.779 ns	941.729 ns	1,099.581 ns	0.55

Copy_PassAsArgumentAndReturn_FromResult	\main\corerun.exe	4.759 ns	0.0122 ns	0.0114 ns	4.760 ns	4.741 ns	4.779 ns	1.00
Copy_PassAsArgumentAndReturn_FromResult	\pr\corerun.exe	4.777 ns	0.0197 ns	0.0184 ns	4.782 ns	4.745 ns	4.804 ns	1.00

Copy_PassAsArgumentAndReturn_FromTask	\main\corerun.exe	8.188 ns	0.0148 ns	0.0131 ns	8.191 ns	8.160 ns	8.213 ns	1.00
Copy_PassAsArgumentAndReturn_FromTask	\pr\corerun.exe	8.236 ns	0.1016 ns	0.0950 ns	8.199 ns	8.142 ns	8.423 ns	1.01

Copy_PassAsArgumentAndReturn_FromValueTaskSource	\main\corerun.exe	12.425 ns	0.1548 ns	0.1448 ns	12.410 ns	12.264 ns	12.664 ns	1.00
Copy_PassAsArgumentAndReturn_FromValueTaskSource	\pr\corerun.exe	12.650 ns	0.0285 ns	0.0253 ns	12.651 ns	12.612 ns	12.690 ns	1.02

CreateAndAwait_FromCompletedValueTaskSource_ConfigureAwait	\main\corerun.exe	16.194 ns	0.1986 ns	0.1858 ns	16.182 ns	15.943 ns	16.477 ns	1.00
CreateAndAwait_FromCompletedValueTaskSource_ConfigureAwait	\pr\corerun.exe	14.792 ns	0.1161 ns	0.1086 ns	14.767 ns	14.663 ns	15.080 ns	0.91

…inuation

More specifically this allows inlining of any async call with only tail awaits. In those cases inlining can be done just by inheriting context handling of the inlining call (and by removing the tail await).

…inuation

…line-await

Copilot

Pull request overview

This PR refactors the runtime-async await helper pipeline to enable more inlining (when all awaits are “tail awaits”) and to reduce allocations by using cached, hand-rolled continuations for Task / ValueTaskSource paths. It also updates async thunk IL generation (CoreCLR + tool) and adjusts the JIT importer to support inlining for tail-await-only async callees and to inherit async context args from the inliner when needed.

Changes:

Rewrites AsyncHelpers.Await* helpers into a tail-await shape and introduces cached TaskContinuation / ValueTaskSourceContinuation continuations to reduce continuation allocations.
Updates CoreCLR async thunk emission to tail-await TransparentAwait* helpers (including new generic TransparentAwaitOfT) and mirrors the change in the managed type-system stub generator.
Updates the JIT importer to (a) allow inlining async callees when awaits are marked tail-await, and (b) propagate async context args from the inliner into inlined async calls.

Reviewed changes

Copilot reviewed 16 out of 16 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
src/tests/async/valuetask-source/valuetask-source.cs	Adds debug output (currently noisy) while validating scheduling-context flags in a `ValueTaskSource` test.
src/libraries/System.Private.CoreLib/src/System/Runtime/CompilerServices/TaskAwaiter.cs	Makes `ConfiguredTaskAwaiter<TResult>.m_task` internal to enable fast-path access from `AsyncHelpers`.
src/libraries/System.Private.CoreLib/src/System/Runtime/CompilerServices/ConfiguredValueTaskAwaitable.cs	Makes configured `ValueTask` awaitable backing fields internal for fast-path access.
src/libraries/System.Private.CoreLib/src/System/Runtime/CompilerServices/AsyncHelpers.cs	Rewrites await helpers to tail-await and route through optimized `AwaitTask` / `AwaitValueTaskSource` paths.
src/coreclr/vm/corelib.h	Adds binder entry for `AsyncHelpers.TransparentAwaitOfT`.
src/coreclr/vm/asyncthunks.cpp	Emits thunk IL that tail-awaits `TransparentAwait*` and returns immediately (incl. generic `Task<T>` case).
src/coreclr/tools/Common/TypeSystem/IL/Stubs/AsyncThunks.cs	Mirrors thunk IL updates in the managed stub generator (adds `ret` after transparent await).
src/coreclr/System.Private.CoreLib/System.Private.CoreLib.csproj	Switches async helper continuation sources to new `TaskContinuation` and `ValueTaskSourceContinuation` files; minor formatting tweaks.
src/coreclr/System.Private.CoreLib/src/System/Runtime/CompilerServices/AsyncProfiler.CoreCLR.cs	Updates comment text to reflect new continuation type naming.
src/coreclr/System.Private.CoreLib/src/System/Runtime/CompilerServices/AsyncHelpers.ValueTaskSourceContinuation.cs	Renames/specializes continuation for `IValueTaskSource*` only and simplifies result handling accordingly.
src/coreclr/System.Private.CoreLib/src/System/Runtime/CompilerServices/AsyncHelpers.TaskContinuation.cs	Introduces cached continuation that can also perform context-queueing before dispatching the runtime-async task.
src/coreclr/System.Private.CoreLib/src/System/Runtime/CompilerServices/AsyncHelpers.CoreCLR.cs	Adds `AwaitTask*` helpers, splits `ValueTask` handling between `Task` vs `ValueTaskSource`, and updates dispatch logic to use new continuations/caches.
src/coreclr/nativeaot/System.Private.CoreLib/src/System.Private.CoreLib.csproj	Includes the new continuation source files for NativeAOT CoreLib build.
src/coreclr/jit/importercalls.cpp	Enables inlining for tail-await-only async callees; adds async-context inheritance for inlined async calls; renames ldvirtftn helper.
src/coreclr/jit/compiler.h	Updates declarations for renamed async-arg insertion helper and adds async-context inheritance helper.
src/coreclr/debug/di/rsstackwalk.cpp	Updates comment text to reflect new continuation type naming.

            // - in optimized (as in JitOptimizeAwait=1, which is the default) case
            //   we do not distinguish "false" config vs. having default/no scheduling context,
            //   and "None" is passed in either case.
+            System.Console.WriteLine(sUseContext.trace);


+//------------------------------------------------------------------------
+// impAddAsyncArgsToInlinedCall:
+//   Inherit async args from inlining call as part of a new async call.
+//
+// Arguments:
+//   call - The async call
+//
+// Remarks:
+//   Currently we only allow inlining of async calls when all awaits are tail
+//   awaits. In that case inlining is simplified as we can just inherit
+//   everything from the inlining call.


@@ -7050,7 +7109,7 @@ void Compiler::impSetupAsyncCall(GenTreeCall* call, OPCODE opcode, unsigned pref
 //   Should be called before the 'this' arg is inserted, but after other IL args
 //   have been inserted.
 //
-void Compiler::impInsertAsyncContinuationForLdvirtftnCall(GenTreeCall* call)
+void Compiler::impInsertAsyncArgsForLdvirtftnCall(GenTreeCall* call)


+                    Debug.Assert(continuationContext is TaskScheduler { });
+                    TaskScheduler sched = (TaskScheduler)continuationContext;
+
+                    // TODO: We do not need TaskSchedulerAwaitTaskContinuation here, just need to refactor its Run method...


+                    var taskSchedCont = new TaskSchedulerAwaitTaskContinuation(sched, (Action)RuntimeAsyncTask.m_action!, flowExecutionContext: false);
+                    taskSchedCont.Run(Task.CompletedTask, canInlineContinuationTask: true);


+        [MethodImpl(MethodImplOptions.Async | MethodImplOptions.AggressiveInlining)]
        [StackTraceHidden]
        public static T Await<T>(Task<T> task)
        {
-            TaskAwaiter<T> awaiter = task.GetAwaiter();
-            if (!awaiter.IsCompleted)
+            if (!task.IsCompleted)
            {
-                UnsafeAwaitAwaiter(awaiter);
+                TailAwait();
+                return AwaitTask(task, ConfigureAwaitOptions.ContinueOnCapturedContext);
            }


dotnet-policy-service · 2026-05-24T07:17:38Z

Tagging subscribers to this area: @dotnet/area-system-threading-tasks
See info in area-owners.md if you want to be subscribed.

jakobbotsch added 14 commits May 18, 2026 13:01

Cache Task continuation for runtime async callable task thunks

192f7ba

Merge branch 'main' of github.com:dotnet/runtime into cache-task-cont…

a84f1f9

…inuation

Fix after merge

b408bbb

Actually fix after merge

c8c51bd

Feedback

98fbcd6

JIT: Allow inlining of AsyncHelpers.Await

ec9ac48

More specifically this allows inlining of any async call with only tail awaits. In those cases inlining can be done just by inheriting context handling of the inlining call (and by removing the tail await).

WIP

8a699d2

Merge branch 'main' of github.com:dotnet/runtime into cache-task-cont…

8a56ec2

…inuation

Slight revise

c94a1e7

Merge remote-tracking branch 'origin/cache-task-continuation' into in…

f90af75

…line-await

WIP

26dfa36

WIP

b6c8df0

Fix

d1e9b6f

Run jit-format

6608440

Copilot AI review requested due to automatic review settings May 23, 2026 22:00

Copilot started reviewing on behalf of jakobbotsch May 23, 2026 22:00 View session

github-actions Bot added the area-System.Runtime.CompilerServices label May 23, 2026

dotnet-policy-service Bot assigned jakobbotsch May 23, 2026

Copilot AI reviewed May 23, 2026

View reviewed changes

jakobbotsch added area-System.Threading.Tasks runtime-async and removed area-System.Runtime.CompilerServices labels May 24, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

JIT: Optimize `AsyncHelpers` await helpers and enable inlining for them #128528

JIT: Optimize `AsyncHelpers` await helpers and enable inlining for them #128528
jakobbotsch wants to merge 14 commits into
dotnet:mainfrom
jakobbotsch:inline-await

jakobbotsch commented May 23, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

dotnet-policy-service Bot commented May 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

		var taskSchedCont = new TaskSchedulerAwaitTaskContinuation(sched, (Action)RuntimeAsyncTask.m_action!, flowExecutionContext: false);
		taskSchedCont.Run(Task.CompletedTask, canInlineContinuationTask: true);

Conversation

jakobbotsch commented May 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

dotnet-policy-service Bot commented May 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

jakobbotsch commented May 23, 2026 •

edited

Loading