Enable cached interface dispatch by default on non-JIT platforms #123776

Copilot · 2026-01-29T21:57:46Z

Description

R2R virtual dispatch on iOS-like platforms was falling back to ExternalMethodFixupWorker on every call instead of patching callsites, causing 200x slowdown in interface-heavy code. Profile showed 18% of time in signature parsing alone.

Changes

Changed --enable-cached-interface-dispatch-support from bool to bool? to support platform-conditional defaults
Defaults to true when TargetAllowsRuntimeCodeGeneration is false (iOS, iOSSimulator, tvOS, tvOSSimulator, MacCatalyst, WASM)
Defaults to false on JIT-capable platforms (existing behavior)
Updated help text to document conditional default

Implementation

Crossgen2RootCommand.cs:

public Option<bool?> EnableCachedInterfaceDispatchSupport { get; } =
    new("--enable-cached-interface-dispatch-support", "--CID") { ... };

Program.cs:

nodeFactoryFlags.EnableCachedInterfaceDispatchSupport = 
    Get(_command.EnableCachedInterfaceDispatchSupport) ?? 
    !typeSystemContext.TargetAllowsRuntimeCodeGeneration;

User-specified values override the default on all platforms.

Original prompt

This section details on the original issue you should resolve

<issue_title>ValueTupleCompareCached microbenchmark slow in ios-like configuration</issue_title>
<issue_description>This benchmark is around 200x slower when running in ios-like configuration (composite r2r + interp fallback) on desktop, compared to default (jit).

Interestingly enough, after profiling, it turns out that most of the time spent is not in interpreting methods but rather in R2R machinery. In order to reproduce, the application was compiled with composite-r2r. The minimal runtime configuration diff that reproduces this issue is BrzVlad@3a4d118. At runtime, tiering has to be disabled as well (DOTNET_TieredCompilation=0).

Most of the time spent originates in ExternalMethodFixupWorker, which is called for every benchmark iteration, suggesting that a callsite from r2r is not properly patched. The runtime seems to attempt to resolve the call for each microbenchmark iteration, with excessive time spent related to signature decoding. Attached a profile showing the amount of time spent in the hottest methods.

Even when running with composite-r2r with jit fallback, the benchmark was still 50x slower than normal jit, with the above runtime configuration. Microbenchmark extracted from https://github.com/dotnet/performance/blob/main/src/benchmarks/micro/runtime/Devirtualization/DefaultEqualityComparerPerf.cs#L79
using System;
using System.Runtime.CompilerServices;
using System.Collections.Generic;
using System.Globalization;

public class EqualityComparerFixture<T> where T : IEquatable<T>
{
	IEqualityComparer<T> comparer;

	public EqualityComparerFixture()
	{
		comparer = EqualityComparer<T>.Default;
	}

	[MethodImpl(MethodImplOptions.NoInlining)]
	public bool CompareCached(ref T a, ref T b)
	{
		return comparer.Equals(a, b);
	}
}

public class Program {

	public enum E {
		RED = 1,
		BLUE = 2
	}

	private EqualityComparerFixture<ValueTuple<byte, E, int>> valueTupleFixture = new EqualityComparerFixture<ValueTuple<byte, E, int>>();
	private ValueTuple<byte, E, int> v0 = new ValueTuple<byte, E, int>(3, E.RED, 11);

	public bool ValueTupleCompareCached() => valueTupleFixture.CompareCached(ref v0, ref v0);

	public void BenchCompareTuple()
	{
		for (int j = 0; j < 100000000; j++)
			ValueTupleCompareCached();
		Console.WriteLine ("Benchmarked compare tuple");
	}

	public static void Main(string[] args)
	{
		Program p = new Program();
		p.BenchCompareTuple();
	}
}
```</issue_description>

## Comments on the Issue (you are @copilot in this section)

<comments>
<comment_new><author>@BrzVlad</author><body>
cc @jkotas @AndyAyersMS @davidwrighton  for any ideas. Not sure who would be the best person to look further into this.</body></comment_new>
<comment_new><author>@davidwrighton</author><body>
This is probably mine to investigate</body></comment_new>
<comment_new><author>@davidwrighton</author><body>
@BrzVlad  Ah, this is an issue that I had forgotten about, but already have a fix for. R2R dispatch to virtual methods using CachedInterfaceDispatch technically works without a special switch to the compiler, but it has to use a very expensive fallback. R2R compilation that is targeted at running in an environment which uses cached interface dispatch really needs to pass the "--enable-cached-interface-dispatch-support" command line argument. I believe we already pass a good --TargetOS switch to the compiler, so we could enable the cached interface dispatch support unconditionally for ios, iossimulator, tvos, tvossimulator, maccatalyst, and wasm.</body></comment_new>
<comment_new><author>@davidwrighton</author><body>
@copilot when implementing this, please use the TargetAllowsRuntimeCodeGeneration property to control the default value of the EnableCachedInterfaceDispatchSupport property set on the nodeFactoryFlags. This change should be in src/coreclr/tools/aot/crossgen2/Program.cs. In addition, please update the help text for the EnableCachedInterfaceDispatchSupport flag in src/coreclr/tools/aot/crossgen2/Properties/Resources.resx to describe that the default is true on platforms which do not support runtime code generation.</body></comment_new>
</comments>

Fixes ValueTupleCompareCached microbenchmark slow in ios-like configuration #123622

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.

Co-authored-by: davidwrighton <10779849+davidwrighton@users.noreply.github.com>

davidwrighton · 2026-01-30T00:35:51Z

@BrzVlad, this will require your scripts that call crossgen2 to use the right targetos switch. I believe @jkoritzinsky has made the behavior in the SDK work, but I don't know exactly what you're testing on now.

Copilot

Pull request overview

This PR addresses a significant performance issue (200x slowdown) on iOS-like platforms where R2R virtual dispatch was falling back to expensive signature parsing on every call. The fix enables cached interface dispatch by default on platforms that don't support runtime code generation (JIT).

Changes:

Changed the --enable-cached-interface-dispatch-support option from bool to bool? to support platform-conditional defaults
Implemented default logic that enables cached interface dispatch on non-JIT platforms (iOS, iOSSimulator, tvOS, tvOSSimulator, MacCatalyst, WASM/WASI) while maintaining the existing false default on JIT-capable platforms
Updated help text to document the conditional default behavior

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
Crossgen2RootCommand.cs	Changed EnableCachedInterfaceDispatchSupport option type from `bool` to `bool?` to allow tristate values (true/false/unspecified)
Program.cs	Added conditional default logic using null-coalescing operator to enable cached interface dispatch when not explicitly set and platform doesn't support JIT
Resources.resx	Updated help text to document that the default is true on platforms which do not support runtime code generation

jkoritzinsky · 2026-01-30T00:48:02Z

@BrzVlad, this will require your scripts that call crossgen2 to use the right targetos switch. I believe @jkoritzinsky has made the behavior in the SDK work, but I don't know exactly what you're testing on now.

Yes the SDK should be hooked up to pass the right target OS and arch.

BrzVlad

Thanks, I'll validate that it works as expected once I have the .net11 P1 working with ios bits.

Initial plan

fdff30b

Copilot AI assigned Copilot and davidwrighton Jan 29, 2026

Copilot started work on behalf of davidwrighton January 29, 2026 21:58 View session

Enable cached interface dispatch support by default on non-jit platforms

9394c9e

Co-authored-by: davidwrighton <10779849+davidwrighton@users.noreply.github.com>

Copilot AI changed the title ~~[WIP] Fix ValueTupleCompareCached benchmark performance in iOS-like config~~ Enable cached interface dispatch by default on non-JIT platforms Jan 29, 2026

Copilot AI requested a review from davidwrighton January 29, 2026 22:17

Copilot finished work on behalf of davidwrighton January 29, 2026 22:17

davidwrighton approved these changes Jan 30, 2026

View reviewed changes

davidwrighton requested a review from BrzVlad January 30, 2026 00:33

davidwrighton marked this pull request as ready for review January 30, 2026 00:33

Copilot AI review requested due to automatic review settings January 30, 2026 00:33

Copilot started reviewing on behalf of davidwrighton January 30, 2026 00:34 View session

Copilot AI reviewed Jan 30, 2026

View reviewed changes

jkoritzinsky approved these changes Jan 30, 2026

View reviewed changes

This was referenced Jan 30, 2026

The Operation will be canceled. The next steps may not contain expected logs. dotnet/dnceng#3008

Open

[android-arm64] The Operation will be canceled. The next steps may not contain expected logs. dotnet/dnceng#6408

Open

BrzVlad approved these changes Jan 30, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Enable cached interface dispatch by default on non-JIT platforms #123776

Enable cached interface dispatch by default on non-JIT platforms #123776

Copilot AI commented Jan 29, 2026 •

edited

Loading

Uh oh!

davidwrighton commented Jan 30, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

jkoritzinsky commented Jan 30, 2026

Uh oh!

BrzVlad left a comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Enable cached interface dispatch by default on non-JIT platforms #123776

Are you sure you want to change the base?

Enable cached interface dispatch by default on non-JIT platforms #123776

Conversation

Copilot AI commented Jan 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Changes

Implementation

Uh oh!

davidwrighton commented Jan 30, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

jkoritzinsky commented Jan 30, 2026

Uh oh!

BrzVlad left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Copilot AI commented Jan 29, 2026 •

edited

Loading