October 31st, 2023

.NET 8 Performance Improvements in .NET MAUI

Jonathan Peppers
Principal Software Engineer

The major focus for .NET MAUI in the .NET 8 release is quality. As such, alot of our focus has been fixing bugs instead of chasing lofty performance goals. In .NET 8, we merged 1,559 pull requests that closed 596 total issues. These include changes from the .NET MAUI team as well as the .NET MAUI community. We are optimistic that this should result in a significant increase in quality in .NET 8.

However! We still have plenty of performance changes to showcase. Building upon the fundamental performance improvements in .NET 8 we discover “low-hanging” fruit constantly, and there were high-voted performance issues on GitHub we tried to tackle. Our goal is to continue to make .NET MAUI faster in each release, read on for details!

For a review of the performance improvements in past releases, see our posts for .NET 6 and 7. This also gives you an idea of the improvements you would see migrating from Xamarin.Forms to .NET MAUI:

Table Of Contents

New features

Build & Inner Loop Performance

Performance or App Size Improvements

Memory Leaks

Tooling and Documentation

New Features

AndroidStripILAfterAOT

Once Upon A Time™ we had a brilliant thought: if AOT pre-compiles C# methods, do we need the managed method anymore? Removing the C# method body would allow assemblies to be smaller. .NET iOS applications already do this, so why not Android as well?

While the idea is straightforward, implementation was not: iOS uses “Full” AOT, which AOT’s all methods into a form that doesn’t require a runtime JIT. This allowed iOS to run cil-strip, removing all method bodies from all managed types.

At the time, Xamarin.Android only supported “normal” AOT, and normal AOT requires a JIT for certain constructs such as generic types and generic methods. This meant that attempting to run cil-strip would result in runtime errors if a method body was removed that was actually required at runtime. This was particularly bad because cil-strip could only remove all method bodies!

We are re-intoducing IL stripping for .NET 8. Add a new $(AndroidStripILAfterAOT) MSBuild property. When true, the <MonoAOTCompiler/> task will track which method bodies were actually AOT’d, storing this information into %(_MonoAOTCompiledAssemblies.MethodTokenFile), and the new <ILStrip/> task will update the input assemblies, removing all method bodies that can be removed.

By default enabling $(AndroidStripILAfterAOT) will override the default $(AndroidEnableProfiledAot) setting, allowing all trimmable AOT’d methods to be removed. This choice was made because $(AndroidStripILAfterAOT) is most useful when AOT-compiling your entire application. Profiled AOT and IL stripping can be used together by explicitly setting both within the .csproj, but with the only benefit being a small .apk size improvement:

<PropertyGroup Condition=" '$(Configuration)' == 'Release' ">
    <AndroidStripILAfterAOT>true</AndroidStripILAfterAOT>
    <AndroidEnableProfiledAot>true</AndroidEnableProfiledAot>
</PropertyGroup>

.apk size results for a dotnet new android app:

$(AndroidStripILAfterAOT) $(AndroidEnableProfiledAot) .apk size
true true 7.7MB
true false 8.1MB
false true 7.7MB
false false 8.4MB

Note that AndroidStripILAfterAOT=false and AndroidEnableProfiledAot=true is the default Release configuration environment, for 7.7MB.

A project that only sets AndroidStripILAfterAOT=true implicitly sets AndroidEnableProfiledAot=false, resulting in an 8.1MB app.

See xamarin-android#8172 and dotnet/runtime#86722 for details about this feature.

AndroidEnableMarshalMethods

.NET 8 introduces a new experimental setting for Release configurations:

<PropertyGroup Condition=" '$(Configuration)' == 'Release' ">
    <AndroidEnableMarshalMethods>true</AndroidEnableMarshalMethods>
    <!-- Note that single-architecture apps will be most successful -->
    <RuntimeIdentifier>android-arm64</RuntimeIdentifier>
</PropertyGroup>

We hope to enable this feature by default in .NET 9, but for now we are providing the setting as an opt-in, experimental feature. Applications that only target one architecture, such as RuntimeIdentifier=android-arm64, will likely be able to enable this feature without issue.

Background on Marshal Methods

A JNI marshal method is a JNI-callable function pointer provided to JNIEnv::RegisterNatives(). Currently, JNI marshal methods are provided via the interaction between code we generate and JNINativeWrapper.CreateDelegate():

  • Our code-generator emits the “actual” JNI-callable method.

  • JNINativeWrapper.CreateDelegate() uses System.Reflection.Emit to wrap the method for exception marshaling.

JNI marshal methods are needed for all Java-to-C# transitions.

Consider the virtual Activity.OnCreate() method:

partial class Activity {
    static Delegate? cb_onCreate_Landroid_os_Bundle_;
    static Delegate GetOnCreate_Landroid_os_Bundle_Handler ()
    {
        if (cb_onCreate_Landroid_os_Bundle_ == null)
            cb_onCreate_Landroid_os_Bundle_ = JNINativeWrapper.CreateDelegate ((_JniMarshal_PPL_V) n_OnCreate_Landroid_os_Bundle_);
        return cb_onCreate_Landroid_os_Bundle_;
    }

    static void n_OnCreate_Landroid_os_Bundle_ (IntPtr jnienv, IntPtr native__this, IntPtr native_savedInstanceState)
    {
        var __this = global::Java.Lang.Object.GetObject<Android.App.Activity> (jnienv, native__this, JniHandleOwnership.DoNotTransfer)!;
        var savedInstanceState = global::Java.Lang.Object.GetObject<Android.OS.Bundle> (native_savedInstanceState, JniHandleOwnership.DoNotTransfer);
        __this.OnCreate (savedInstanceState);
    }

    // Metadata.xml XPath method reference: path="/api/package[@name='android.app']/class[@name='Activity']/method[@name='onCreate' and count(parameter)=1 and parameter[1][@type='android.os.Bundle']]"
    [Register ("onCreate", "(Landroid/os/Bundle;)V", "GetOnCreate_Landroid_os_Bundle_Handler")]
    protected virtual unsafe void OnCreate (Android.OS.Bundle? savedInstanceState) => ...
}

Activity.n_OnCreate_Landroid_os_Bundle_() is the JNI marshal method, responsible for marshaling parameters from JNI values into C# types, forwarding the method invocation to Activity.OnCreate(), and (if necessary) marshaling the return value back to JNI.

Activity.GetOnCreate_Landroid_os_Bundle_Handler() is part of the type registration infrastructure, providing a Delegate instance to RegisterNativeMembers .RegisterNativeMembers(), which is eventually passed to JNIEnv::RegisterNatives().

While this works, it’s not incredibly performant: unless using one of the optimized delegate types added in xamarin-android#6657, System.Reflection.Emit is used to create a wrapper around the marshal method, which is something we’ve wanted to avoid doing for years.

Thus, the idea: since we’re already bundling a native toolchain and using LLVM-IR to produce libxamarin-app.so, what if we emitted Java native method names and skipped all the done as part of Runtime.register() and JNIEnv.RegisterJniNatives()?

Given:

class MyActivity : Activity {
    protected override void OnCreate(Bundle? state) => ...
}

During the build, libxamarin-app.so would contain the function:

JNIEXPORT void JNICALL
Java_crc..._MyActivity_n_1onCreate (JNIEnv *env, jobject self, jobject state);

During App runtime, the Runtime.register() invocation present in Java Callable Wrappers would either be omitted or would be a no-op, and Android/JNI would instead resolve MyActivity.n_onCreate() as Java_crc..._MyActivity_n_1onCreate().

We call this effort “LLVM Marshal Methods”, which is currently experimental in .NET 8. Many of the specifics are still being investigated, and this feature will be spread across various areas.

See xamarin-android#7351 for details about this experimental feature.

NativeAOT on iOS

In .NET 7, we started an experiment to see what it would take to support NativeAOT on iOS. Going from prototype to an initial implementation: .NET 8 Preview 6 included NativeAOT as an experimental feature for iOS.

To opt into NativeAOT in a MAUI iOS project, use the following settings in your project file:

<PropertyGroup Condition="$([MSBuild]::GetTargetPlatformIdentifier('$(TargetFramework)')) == 'ios' and '$(Configuration)' == 'Release'">
    <!-- PublishAot=true indicates NativeAOT, while omitting this property would use Mono's AOT -->
    <PublishAot>true</PublishAot>
</PropertyGroup>

Then to build the application for an iOS device:

$ dotnet publish -f net8.0-ios -r ios-arm64
MSBuild version 17.8.0+6cdef4241 for .NET
...
Build succeeded.
    0 Error(s)

Note We may consider unifying and improving MSBuild property names for this feature in future .NET releases. To do a one-off build at the command-line you may also need to specify -p:PublishAotUsingRuntimePack=true in addition to -p:PublishAot=true.

One of the main culprits for the first release was how the iOS workload supports Objective-C interoperability. The problem was mainly related to the type registration system which is the key component for efficiently supporting iOS-like platforms (see docs for details). In its implementation, the type registration system depends on type metadata tokens which are not available with NativeAOT. Therefore, in order to leverage the benefits of highly efficient NativeAOT runtime, we had to adapt. dotnet/runtime#80912 includes the discussion around how to tackle this problem, and finally in xamarin-macios#18268 we implemented a new managed static registrar that works with NativeAOT. The new managed static registrar does not just benefit us with being compatible with NativeAOT, but is also much faster than the default one, and is available for all supported runtimes (see docs for details).

Along the way, we had a great help from our GH community and their contribution (code reviews, PRs) was essential to helps us move forward quickly and deliver this feature on time. A few from many PR’s that helped and unblocked us on our journey were:

and the list goes on…

As .NET 8 Preview 6 came along, we finally managed to release our first version of the NativeAOT on iOS which also supports MAUI. See the blog post on .NET 8 Preview 6 for details about what we were able to accomplish in the initial release.

In subsequent .NET 8 releases, results improved quite a bit, as we were identifying and resolving issues along the way. The graph below shows the .NET MAUI iOS template app size comparison throughout the preview releases:

.NET MAUI iOS app - size comparison

We had steady progress and estimated size savings reported, due to fixing the following issues:

  • dotnet/runtime#87924 – fixed major NativeAOT size issue with AOT-incompatible code paths in System.Linq.Expressions and also made fully NativeAOT compatible when targeting iOS

  • xamarin-macios#18332 – reduced the size of __LINKEDIT Export Info section in stripped binaries

Furthermore, in the latest RC 1 release the app size went even further down reaching -50% smaller apps for the template .NET MAUI iOS applications compared to Mono. Most impactful issues/PRs that contributed to this:

Even though app size was our primary metric to focus on, for the RC 1 release, we also measured startup time performance comparisons for a .NET MAUI iOS template app comparing NativeAOT and Mono where NativeAOT results with almost 2x faster startup time.

.NET MAUI iOS app - startup time

Key Takeaways

For NativeAOT scenarios on iOS, changing the default link mode to Full (xamarin-macios#18734) is probably the biggest improvement for application size. But at the same time, this change can also break applications which are not fully AOT and trim-compatible. In Full link mode, the trimmer might trim away AOT incompatible code paths (think about reflection usage) which are accessed dynamically at runtime. Full link mode is not the default configuration when using the Mono runtime, so it is possible that some applications are not fully AOT-compatible.

Supporting NativeAOT on iOS is an experimental feature and still a work-in-progress, and our plan is to address the potential issues with Full link mode incrementally:

  • As a first step, we enabled trim, AOT, and single-file warnings by default in xamarin-macios#18571. The enabled warnings should make our customers aware at build-time, whether a use of a certain framework or a library, or some C# constructs in their code, is incompatible with NativeAOT – and could crash at runtime. This information should guide our customers to write AOT-compatible code, but also to help us improve our frameworks and libraries with the same goal of fully utilising the benefits of AOT compilation.

  • The second step, was clearing up all the warnings coming from Microsoft.iOS and System.Private.CoreLib assemblies reported for a template iOS application with: xamarin-macios#18629 and dotnet/runtime#91520.

  • In future releases, we plan to address the warnings coming from the MAUI framework and further improve the overall user-experience. Our goal is to have fully AOT and trim-compatible frameworks.

.NET 8 will support targeting iOS platforms with NativeAOT as an opt-in feature and shows great potential by generating up to 50% smaller and 50% faster startup compared to Mono. Considering the great performance that NativeAOT promises, please help us on this journey and try out your applications with NativeAOT and report any potential issues. At the same time, let us know when NativeAOT “just works” out-of-the-box.

To follow future progress, see dotnet/runtime#80905. Last but not least, we would like to thank our GH contributors, who are helping us make NativeAOT on iOS possible.

Build & Inner Loop Performance

Filter Android ps -A output with grep

When profiling the Android inner loop for a .NET MAUI project with PerfView we found around 1.2% of CPU time was spent just trying to get the process ID of the running Android application.

When changing Tools > Options > Xamarin > Xamarin Diagnostics output verbosity to be Diagnostics, you could see:

-- Start GetProcessId - 12/02/2022 11:05:57 (96.9929ms) --
[INPUT] ps -A
[OUTPUT]
USER           PID  PPID     VSZ    RSS WCHAN            ADDR S NAME
root             1     0 10943736  4288 0                   0 S init
root             2     0       0      0 0                   0 S [kthreadd]
... Hundreds of more lines!
u0_a993      14500  1340 14910808 250404 0                  0 R com.companyname.mauiapp42
-- End GetProcessId --

The Xamarin/.NET MAUI extension in Visual Studio polls every second to see if the application has exited. This is useful for changing the play/stop button state if you force close the app, etc.

Testing on a Pixel 5, we could see the command is actually 762 lines of output!

> (adb shell ps -A).Count
762

What we could do instead is something like:

> adb shell "ps -A | grep -w -E 'PID|com.companyname.mauiapp42'"

Where we pipe the output of ps -A to the grep command on the Android device. Yes, Android has a subset of unix commands available! We filter on either a line containing PID or your application’s package name.

The result is now the IDE is only parsing 4 lines:

[INPUT] ps -A | grep -w -E 'PID|com.companyname.mauiapp42'
[OUTPUT]
USER           PID  PPID     VSZ    RSS WCHAN            ADDR S NAME
u0_a993      12856  1340 15020476 272724 0                  0 S com.companyname.mauiapp42

This not only improves memory used to split and parse this information in C#, but adb is also transmitting way less bytes across your USB cable or virtually from an emulator.

This feature shipped in recent versions of Visual Studio 2022, improving this scenario for all Xamarin and .NET MAUI customers.

Port WindowsAppSDK usage of vcmeta.dll to C#

We found that every incremental build of a .NET MAUI project running on Windows spent time in:

Top 10 most expensive tasks
CompileXaml = 3.972 s
... various tasks ...

This is the XAML compiler for WindowsAppSDK, that compiles the WinUI3 flavor of XAML (not .NET MAUI XAML). There is very little XAML of this type in .NET MAUI projects, in fact, the only file is Platforms/Windows/App.xaml in the project template.

Interestingly, if you installed the Desktop development with C++ workload in the Visual Studio installer, this time just completely went away!

Top 10 most expensive tasks
... various tasks ...
CompileXaml = 9 ms

The WindowsAppSDK XAML compiler p/invokes into a native library from the C++ workload, vcmeta.dll, to calculate a hash for .NET assembly files. This is used to make incremental builds fast — if the hash changes, compile the XAML again. If vcmeta.dll was not found on disk, the XAML compiler was effectively “recompiling everything” on every incremental build.

For an initial fix, we simply included a small part of the C++ workload as a dependency of .NET MAUI in Visual Studio. The slightly larger install size was a good tradeoff for saving upwards of 4 seconds in incremental build time.

Next, we implemented vcmeta.dll‘s hashing functionality in plain C# with System.Reflection.Metadata to compute indentical hash values as before. Not only was this implementation better, in that we could drop a dependency on the C++ workload, but it was also faster! The time to compute a single hash:

Method Mean Error StdDev
Native 217.31 us 1.704 us 1.594 us
Managed 86.43 us 1.700 us 2.210 us

Some of the reasons this was faster:

  • No p/invoke or COM-interfaces involved.

  • System.Reflection.Metadata has a fast struct-based API, perfect for iterating over types in a .NET assembly and computing a hash value.

The end result being that CompileXaml might actually be even faster than 9ms in incremental builds.

This feature shipped in WindowsAppSDK 1.3, which is now used by .NET MAUI in .NET 8. See WindowsAppSDK#3128 for details about this improvement.

Improvements to remote iOS builds on Windows

Comparing inner loop performance for iOS, there was a considerable gap between doing “remote iOS” development on Windows versus doing everything locally on macOS. Many small improvements were made, based on comparing inner-loop .binlog files recorded on macOS versus one recorded inside Visual Studio on Windows.

Some examples include:

  • maui#12747: don’t explicitly copy files to the build server
  • xamarin-macios#16752: do not copy files to build server for a Delete operation
  • xamarin-macios#16929: batch file deletion via DeleteFilesAsync
  • xamarin-macios#17033: cache AOT compiler path
  • Xamarin/MAUI Visual Studio extension: when running dotnet-install.sh on remote build hosts, set the explicit processor flag for M1 Macs.

We also made some improvements for all iOS & MacCatalyst projects, such as:

Improvements to Android inner-loop

We also made many small improvements to the “inner-loop” on Android — most of which were focused in a specific area.

Previously, Xamarin.Forms projects had the luxury of being organized into multiple projects, such as:

  • YourApp.Android.csproj: Xamarin.Android application project
  • YourApp.iOS.csproj: Xamarin.iOS application project
  • YourApp.csproj: netstandard2.0 class library

Where almost all of the logic for a Xamarin.Forms app was contained in the netstandard2.0 project. Nearly all the incremental builds would be changes to XAML or C# in the class library. This structure enabled the Xamarin.Android MSBuild targets to completely skip many Android-specific MSBuild steps. In .NET MAUI, the “single project” feature means that every incremental build has to run these Android-specific build steps.

In focusing specifically improving this area, we made many small changes, such as:

These changes should improve incremental builds in all .NET 8 Android project types.

XAML Compilation no longer uses LoadInSeparateAppDomain

Looking at the JITStats report in PerfView (for MSBuild.exe):

Name JitTime (ms)
Microsoft.Maui.Controls.Build.Tasks.dll 214.0
Mono.Cecil 119.0

It appears that Microsoft.Maui.Controls.Build.Tasks.dll was spending a lot of time in the JIT. What was confusing, is this was an incremental build where everything should already be loaded. The JIT’s work should be done already?

The cause appears to be usage of the [LoadInSeparateAppDomain] attribute defined by the <XamlCTask/> in .NET MAUI. This is an MSBuild feature that gives MSBuild tasks to run in an isolated AppDomain — with an obvious performance drawback. However, we couldn’t just remove it as there would be complications…

[LoadInSeparateAppDomain] also conveniently resets all static state when <XamlCTask/> runs again. Meaning that future incremental builds would potentially use old (garbage) values. There are several places that cache Mono.Cecil objects for performance reasons. Really weird bugs would result if we didn’t address this.

So, to actually make this change, we reworked all static state in the XAML compiler to be stored in instance fields & properties instead. This is a general software design improvement, in addition to giving us the ability to safely remove [LoadInSeparateAppDomain].

The results of this change, for an incremental build on a Windows PC:

Before:
XamlCTask = 743 ms
XamlCTask = 706 ms
XamlCTask = 692 ms
After:
XamlCTask = 128 ms
XamlCTask = 134 ms
XamlCTask = 117 ms

This saved about ~587ms on incremental builds on all platforms, an 82% improvement. This will help even more on large solutions with multiple .NET MAUI projects, where <XamlCTask/> runs multiple times.

See maui#11982 for further details about this improvement.

Performance or App Size Improvements

Structs and IEquatable in .NET MAUI

Using the Visual Studio’s .NET Object Allocation Tracking profiler on a customer .NET MAUI sample application, we saw:

Microsoft.Maui.WeakEventManager+Subscription
Allocations: 686,114
Bytes: 21,955,648

Screenshot of WeakEventManager+Subscription in the Visual Studio Profiler

This seemed like an exorbitant amount of memory to be used in a sample application’s startup!

Drilling in to see where these struct‘s were being created:

System.Collections.Generic.ObjectEqualityComparer<Microsoft.Maui.WeakEventManager+Subscription>.IndexOf()

The underlying problem was this struct didn’t implement IEquatable<T> and was being used as the key for a dictionary. The CA1815 code analysis rule was designed to catch this problem. This is not a rule that is enabled by default, so projects must opt into it.

To solve this:

  • Subscription is internal to .NET MAUI, and its usage made it possible to be a readonly struct. This was just an extra improvement.

  • We made CA1815 a build error across the entire dotnet/maui repository.

  • We implemented IEquatable<T> for all struct types.

After these changes, we could no longer found Microsoft.Maui.WeakEventManager+Subscription in memory snapshots at all. Which saved ~21 MB of allocations in this sample application. If your own projects have usage of struct, it seems quite worthwhile to make CA1815 a build error.

A smaller, targeted version of this change was backported to MAUI in .NET 7. See maui#13232 for details about this improvement.

Fix performance issue in {AppThemeBinding}

Profiling a .NET MAUI sample application from a customer, we noticed a lot of time spent in {AppThemeBinding} and WeakEventManager while scrolling:

2.08s (17%) microsoft.maui.controls!Microsoft.Maui.Controls.AppThemeBinding.Apply(object,Microsoft.Maui.Controls.BindableObject,Micr...
2.05s (16%) microsoft.maui.controls!Microsoft.Maui.Controls.AppThemeBinding.AttachEvents()
2.04s (16%) microsoft.maui!Microsoft.Maui.WeakEventManager.RemoveEventHandler(System.EventHandler`1<TEventArgs_REF>,string)

The following was happening in this application:

  • The standard .NET MAUI project template has lots of {AppThemeBinding} in the default Styles.xaml. This supports Light vs Dark theming.

  • {AppThemeBinding} subscribes to Application.RequestedThemeChanged

  • So, every MAUI view subscribe to this event — potentially multiple times.

  • Subscribers are a Dictionary<string, List<Subscriber>>, where there is a dictionary lookup followed by a O(N) search for unsubscribe operations.

There is potentially a usecase here to come up with a generalized “weak event” pattern for .NET. The implementation currently in .NET MAUI came over from Xamarin.Forms, but a generalized pattern could be useful for .NET developers using other UI frameworks.

To make this scenario fast, for now, in .NET 8:

Before:

  • For any {AppThemeBinding}, it calls both:
    • RequestedThemeChanged -= OnRequestedThemeChanged O(N) time
    • RequestedThemeChanged += OnRequestedThemeChanged constant time
  • Where the -= is notably slower, due to possibly 100s of subscribers.

After:

  • Create an _attached boolean, so we know know the “state” if it is attached or not.

  • New bindings only call +=, where -= will now only be called by {AppThemeBinding} in rare cases.

  • Most .NET MAUI apps do not “unapply” bindings, but -= would only be used in that case.

See the full details about this fix in maui#14625. See dotnet/runtime#61517 for how we could implement “weak events” in .NET in the future.

Address CA1307 and CA1309 for performance

Profiling a .NET MAUI sample application from a customer, we noticed time spent during “culture-aware” string operations:

77.22ms microsoft.maui!Microsoft.Maui.Graphics.MauiDrawable.SetDefaultBackgroundColor()
42.55ms System.Private.CoreLib!System.String.ToLower()

This case, we can improve by simply calling ToLowerInvariant() instead. In some cases you might even consider using string.Equals() with StringComparer.Ordinal. In this case, our code was further reviewed and optimized in Reduce Java interop in MauiDrawable on Android.

In .NET 7, we added CA1307 and CA1309 code analysis rules to catch cases like this, but it appears we missed some in Microsoft.Maui.Graphics.dll. These are likely useful rules to enable in your own .NET MAUI applications, as avoiding all culture-aware string operations can be quite impactful on mobile.

See maui#14627 for details about this improvement.

Address CA1311 for performance

After addressing the CA1307 and CA1309 code analysis rules, we took things further and addressed CA1311.

As mentioned in the turkish example, doing something like:

string text = something.ToUpper();
switch (text) { ... }

Can actually cause unexpected behavior in Turkish locales, because in Turkish, the character I (Unicode 0049) is considered the upper case version of a different character ý (Unicode 0131), and i (Unicode 0069) is considered the lower case version of yet another character Ý (Unicode 0130).

ToLowerInvariant() and ToUpperInvariant() are also better for performance as an invariant ToLower / ToUpper operation is slightly faster. Doing this also avoids loading the current culture, improving startup performance.

There are cases where you would want the current culture, such as in a CaseConverter type in .NET MAUI. To do this, you simply have to be explicit in which culture you want to use:

return ConvertToUpper ?
    v.ToUpper(CultureInfo.CurrentCulture) :
    v.ToLower(CultureInfo.CurrentCulture);

The goal of this CaseConverter is to display upper or lowercase text to a user. So it makes sense to use the CurrentCulture for this.

See maui#14773 for details about this improvement.

Remove unused ViewAttachedToWindow event on Android

Every Label in .NET MAUI was subscribing to:

public class MauiTextView : AppCompatTextView
{
    public MauiTextView(Context context) : base(context)
    {
        this.ViewAttachedToWindow += MauiTextView_ViewAttachedToWindow;
    }

    private void MauiTextView_ViewAttachedToWindow(object? sender, ViewAttachedToWindowEventArgs e)
    {
    }

    //...

This was leftover from refactoring, but appeared in dotnet-trace output as:

278.55ms (2.4%) mono.android!Android.Views.View.add_ViewAttachedToWindow(System.EventHandler`1<Android.Views.View/ViewAttachedToWindowEv
 30.55ms (0.26%) mono.android!Android.Views.View.IOnAttachStateChangeListenerInvoker.n_OnViewAttachedToWindow_Landroid_view_View__mm_wra

Where the first is the subscription, and the second is the event firing from Java to C# — only to run an empty managed method.

Simply removing this event subscription and empty method, resulted in only a few controls to subscribe to this event as needed:

2.76ms (0.02%) mono.android!Android.Views.View.add_ViewAttachedToWindow(System.EventHandler`1<Android.Views.View/ViewAttachedToWindowEv

See maui#14833 for details about this improvement.

Remove unneeded System.Reflection for {Binding}

All bindings in .NET MAUI commonly hit the code path:

if (property.CanWrite && property.SetMethod.IsPublic && !property.SetMethod.IsStatic)
{
    part.LastSetter = property.SetMethod;
    var lastSetterParameters = part.LastSetter.GetParameters();
    part.SetterType = lastSetterParameters[lastSetterParameters.Length - 1].ParameterType;
    //...

Where ~53% of the time spent applying a binding appeared in dotnet-trace in the MethodInfo.GetParameters() method:

core.benchmarks!Microsoft.Maui.Benchmarks.BindingBenchmarker.BindName()
...
microsoft.maui.controls!Microsoft.Maui.Controls.BindingExpression.SetupPart()
System.Private.CoreLib.il!System.Reflection.RuntimeMethodInfo.GetParameters()

The above C# is simply finding the property type. It is using a roundabout way of using the property setter’s first parameter, which can be simplified to:

part.SetterType = property.PropertyType;

We could see the results of this change in a BenchmarkDotNet benchmark:

Method Mean Error StdDev Gen0 Gen1 Allocated
–BindName 18.82 us 0.336 us 0.471 us 1.2817 1.2512 10.55 KB
++BindName 18.80 us 0.371 us 0.555 us 1.2512 1.2207 10.23 KB
–BindChild 27.47 us 0.542 us 0.827 us 2.0142 1.9836 16.56 KB
++BindChild 26.71 us 0.516 us 0.652 us 1.9226 1.8921 15.94 KB
–BindChildIndexer 58.39 us 1.113 us 1.143 us 3.1738 3.1128 26.17 KB
++BindChildIndexer 58.00 us 1.055 us 1.295 us 3.1128 3.0518 25.47 KB

Where ++ denotes the new changes.

See maui#14830 for further details about this improvement.

Use StringComparer.Ordinal for Dictionary and HashSet

Profiling a .NET MAUI sample application from a customer, we noticed 4% of the time while scrolling was spent doing dictionary lookups:

(4.0%) System.Private.CoreLib!System.Collections.Generic.Dictionary<TKey_REF,TValue_REF>.FindValue(TKey_REF)

Observing the call stack, some of these were coming from culture-aware string lookups in .NET MAUI:

  • microsoft.maui!Microsoft.Maui.PropertyMapper.GetProperty(string)
  • microsoft.maui!Microsoft.Maui.WeakEventManager.AddEventHandler(System.EventHandler<TEventArgs_REF>,string)
  • microsoft.maui!Microsoft.Maui.CommandMapper.GetCommand(string)

Which show up in dotnet-trace as a mixture of string comparers:

(0.98%) System.Private.CoreLib!System.Collections.Generic.NonRandomizedStringEqualityComparer.OrdinalComparer.GetHashCode(string)
(0.71%) System.Private.CoreLib!System.String.GetNonRandomizedHashCode()
(0.31%) System.Private.CoreLib!System.Collections.Generic.NonRandomizedStringEqualityComparer.OrdinalComparer.Equals(string,stri
(0.01%) System.Private.CoreLib!System.Collections.Generic.NonRandomizedStringEqualityComparer.GetStringComparer(object)

In cases of Dictionary<string, TValue> or HashSet<string>, we can use StringComparer.Ordinal in many cases to get faster dictionary lookups. This should slightly improve the performance of handlers & all .NET MAUI controls on all platforms.

See maui#14900 for details about this improvement.

Reduce Java interop in MauiDrawable on Android

Profiling a .NET MAUI customer sample while scrolling on a Pixel 5, we saw some interesting time being spent in:

(0.76%) microsoft.maui!Microsoft.Maui.Graphics.MauiDrawable.OnDraw(Android.Graphics.Drawables.Shapes.Shape,Android.Graphics.Canv
(0.54%) microsoft.maui!Microsoft.Maui.Graphics.MauiDrawable.SetDefaultBackgroundColor()

This sample has a <Border/> inside a <CollectionView/> and so you can see this work happening while scrolling.

Specifically, we reviewed code in .NET MAUI, such as:

_borderPaint.StrokeWidth = _strokeThickness;
_borderPaint.StrokeJoin = _strokeLineJoin;
_borderPaint.StrokeCap = _strokeLineCap;
_borderPaint.StrokeMiter = _strokeMiterLimit * 2;
if (_borderPathEffect != null)
    _borderPaint.SetPathEffect(_borderPathEffect);

This calls from C# to Java five times. Creating a new method in PlatformInterop.java allowed us to reduce it to a single time.

We also improved the following method, which would perform many calls from C# to Java:

// C#

void SetDefaultBackgroundColor()
{
    using (var background = new TypedValue())
    {
        if (_context == null || _context.Theme == null || _context.Resources == null)
            return;

        if (_context.Theme.ResolveAttribute(global::Android.Resource.Attribute.WindowBackground, background, true))
        {
            var resource = _context.Resources.GetResourceTypeName(background.ResourceId);
            var type = resource?.ToLowerInvariant();

            if (type == "color")
            {
                var color = new Android.Graphics.Color(ContextCompat.GetColor(_context, background.ResourceId));
                _backgroundColor = color;
            }
        }
    }
}

To be more succinctly implemented in Java as:

// Java

/**
 * Gets the value of android.R.attr.windowBackground from the given Context
 * @param context
 * @return the color or -1 if not found
 */
public static int getWindowBackgroundColor(Context context)
{
    TypedValue value = new TypedValue();
    if (!context.getTheme().resolveAttribute(android.R.attr.windowBackground, value, true) && isColorType(value)) {
        return value.data;
    } else {
        return -1;
    }
}

/**
 * Needed because TypedValue.isColorType() is only API Q+
 * https://github.com/aosp-mirror/platform_frameworks_base/blob/1d896eeeb8744a1498128d62c09a3aa0a2a29a16/core/java/android/util/TypedValue.java#L266-L268
 * @param value
 * @return true if the TypedValue is a Color
 */
private static boolean isColorType(TypedValue value)
{
    if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.Q) {
        return value.isColorType();
    } else {
        // Implementation from AOSP
        return (value.type >= TypedValue.TYPE_FIRST_COLOR_INT && value.type <= TypedValue.TYPE_LAST_COLOR_INT);
    }
}

Which reduces our new implementation on the C# side to be a single Java call and creation of an Android.Graphics.Color struct:

void SetDefaultBackgroundColor()
{
    var color = PlatformInterop.GetWindowBackgroundColor(_context);
    if (color != -1)
    {
        _backgroundColor = new Android.Graphics.Color(color);
    }
}

After these changes, we instead saw dotnet-trace output, such as:

(0.28%) microsoft.maui!Microsoft.Maui.Graphics.MauiDrawable.OnDraw(Android.Graphics.Drawables.Shapes.Shape,Android.Graphics.Canv
(0.04%) microsoft.maui!Microsoft.Maui.Graphics.MauiDrawable.SetDefaultBackgroundColor()

This improves the performance of any <Border/> (and other shapes) on Android, and drops about ~1% of the CPU usage while scrolling in this example.

See maui#14933 for further details about this improvement.

Improve layout performance of Label on Android

Testing various .NET MAUI sample applications on Android, we noticed around 5.1% of time spent in PrepareForTextViewArrange():

1.01s (5.1%) microsoft.maui!Microsoft.Maui.ViewHandlerExtensions.PrepareForTextViewArrange(Microsoft.Maui.IViewHandler,Microsoft.Maui
635.99ms (3.2%) mono.android!Android.Views.View.get_Context()

Most of the time is spent just calling Android.Views.View.Context to be able to then call into the extension method:

internal static int MakeMeasureSpecExact(this Context context, double size)
{
    // Convert to a native size to create the spec for measuring
    var deviceSize = (int)context!.ToPixels(size);
    return MeasureSpecMode.Exactly.MakeMeasureSpec(deviceSize);
}

Calling the Context property can be expensive due the interop from C# to Java. Java returns a handle to the instance, then we have to look up any existing, managed C# objects for the Context. If all this work can simply be avoided, it can improve performance dramatically.

In .NET 7, we made overloads to ToPixels() that allows you to get the same value with an Android.Views.View

So we can instead do:

internal static int MakeMeasureSpecExact(this PlatformView view, double size)
{
    // Convert to a native size to create the spec for measuring
    var deviceSize = (int)view.ToPixels(size);
    return MeasureSpecMode.Exactly.MakeMeasureSpec(deviceSize);
}

Not only did this change show improvements in dotnet-trace output, but we saw a noticeable difference in our LOLs per second test application from last year:

385.85 to 396.64 LOLs per second

See maui#14980 for details about this improvement.

Reduce Java interop calls for controls in .NET MAUI

Reviewing the beautiful .NET MAUI “Surfing App” sample by @jsuarezruiz:

Surfing app, made with .NET MAUI screenshot

We noticed that a lot of time is spent doing Java interop while scrolling:

1.76s (35%) Microsoft.Maui!Microsoft.Maui.Platform.WrapperView.DispatchDraw(Android.Graphics.Canvas)
1.76s (35%) Microsoft.Maui!Microsoft.Maui.Platform.ContentViewGroup.DispatchDraw(Android.Graphics.Canvas)

These methods were deeply nested doing interop from Java -> C# -> Java many levels deep. In this case, moving some code from C# to Java could make it where less interop would occur; and in some cases no interop at all!

So for example, previously DispatchDraw() was overridden in C# to implement clipping behavior:

// C#
// ContentViewGroup is used internally by many .NET MAUI Controls
class ContentViewGroup : Android.Views.ViewGroup
{
    protected override void DispatchDraw(Canvas? canvas)
    {
        if (Clip != null)
            ClipChild(canvas);

        base.DispatchDraw(canvas);
    }
}

By creating a PlatformContentViewGroup.java, we can do something like:

// Java

/**
 * Set by C#, determining if we need to call getClipPath()
 * @param hasClip
 */
protected final void setHasClip(boolean hasClip) {
    this.hasClip = hasClip;
    postInvalidate();
}

@Override
protected void dispatchDraw(Canvas canvas) {
    // Only call into C# if there is a Clip
    if (hasClip) {
        Path path = getClipPath(canvas.getWidth(), canvas.getHeight());
        if (path != null) {
            canvas.clipPath(path);
        }
    }
    super.dispatchDraw(canvas);
}

setHasClip() is called when clipping is enabled/disabled on any .NET MAUI control. This allowed the common path to not interop into C# at all, and only views that have opted into clipping would need to. This is very good because dispatchDraw() is called quite often during Android layout, scrolling, etc.

This same treatment was also done to a few other internal .NET MAUI types like WrapperView: improving the common case, making interop only occur when views have opted into clipping or drop shadows.

For testing the impact of these changes, we used Google’s FrameMetricsAggregator that can be setup in any .NET MAUI application’s Platforms/Android/MainActivity.cs:

// How often in ms you'd like to print the statistics to the console
const int Duration = 1000;
FrameMetricsAggregator aggregator;
Handler handler;

protected override void OnCreate(Bundle savedInstanceState)
{
    base.OnCreate(savedInstanceState);

    handler = new Handler(Looper.MainLooper);

    // We were interested in the "Total" time, other metrics also available
    aggregator = new FrameMetricsAggregator(FrameMetricsAggregator.TotalDuration);
    aggregator.Add(this);

    handler.PostDelayed(OnFrame, Duration);
}

void OnFrame()
{
    // We were interested in the "Total" time, other metrics also available
    var metrics = aggregator.GetMetrics()[FrameMetricsAggregator.TotalIndex];
    int size = metrics.Size();
    double sum = 0, count = 0, slow = 0;
    for (int i = 0; i < size; i++)
    {
        int value = metrics.Get(i);
        if (value != 0)
        {
            count += value;
            sum += i * value;
            if (i > 16)
                slow += value;
            Console.WriteLine($"Frame(s) that took ~{i}ms, count: {value}");
        }
    }
    if (sum > 0)
    {
        Console.WriteLine($"Average frame time: {sum / count:0.00}ms");
        Console.WriteLine($"No. of slow frames: {slow}");
        Console.WriteLine("-----");
    }
    handler.PostDelayed(OnFrame, Duration);
}

FrameMetricsAggregator‘s API is admittedly a bit odd, but the data we get out is quite useful. The result is basically a lookup table where the key is a duration in milliseconds, and the value is the number of “frames” that took that duration. The idea is any frame that takes longer than 16ms is considered “slow” or “janky” as the Android docs sometimes refer.

An example of the .NET MAUI “Surfing App” running on a Pixel 5:

Before:
Frame(s) that took ~4ms, count: 1
Frame(s) that took ~5ms, count: 6
Frame(s) that took ~6ms, count: 10
Frame(s) that took ~7ms, count: 12
Frame(s) that took ~8ms, count: 10
Frame(s) that took ~9ms, count: 6
Frame(s) that took ~10ms, count: 1
Frame(s) that took ~11ms, count: 2
Frame(s) that took ~12ms, count: 4
Frame(s) that took ~13ms, count: 2
Frame(s) that took ~15ms, count: 1
Frame(s) that took ~16ms, count: 1
Frame(s) that took ~18ms, count: 2
Frame(s) that took ~19ms, count: 1
Frame(s) that took ~20ms, count: 5
Frame(s) that took ~21ms, count: 2
Frame(s) that took ~22ms, count: 1
Frame(s) that took ~25ms, count: 1
Frame(s) that took ~32ms, count: 1
Frame(s) that took ~34ms, count: 1
Frame(s) that took ~60ms, count: 1
Frame(s) that took ~62ms, count: 1
Frame(s) that took ~63ms, count: 1
Frame(s) that took ~64ms, count: 2
Frame(s) that took ~66ms, count: 1
Frame(s) that took ~67ms, count: 1
Frame(s) that took ~68ms, count: 1
Frame(s) that took ~69ms, count: 2
Frame(s) that took ~70ms, count: 2
Frame(s) that took ~71ms, count: 2
Frame(s) that took ~72ms, count: 1
Frame(s) that took ~73ms, count: 2
Frame(s) that took ~74ms, count: 2
Frame(s) that took ~75ms, count: 1
Frame(s) that took ~76ms, count: 1
Frame(s) that took ~77ms, count: 2
Frame(s) that took ~78ms, count: 3
Frame(s) that took ~79ms, count: 1
Frame(s) that took ~80ms, count: 1
Frame(s) that took ~81ms, count: 1
Average frame time: 28.67ms
No. of slow frames: 43

After the changes to ContentViewGroup and WrapperView were in place, we got a very nice improvement! Even in an app making heavy usage of clipping and shadows:

After:
Frame(s) that took ~5ms, count: 3
Frame(s) that took ~6ms, count: 5
Frame(s) that took ~7ms, count: 7
Frame(s) that took ~8ms, count: 7
Frame(s) that took ~9ms, count: 4
Frame(s) that took ~10ms, count: 2
Frame(s) that took ~11ms, count: 6
Frame(s) that took ~12ms, count: 2
Frame(s) that took ~13ms, count: 3
Frame(s) that took ~14ms, count: 4
Frame(s) that took ~15ms, count: 1
Frame(s) that took ~16ms, count: 1
Frame(s) that took ~17ms, count: 1
Frame(s) that took ~18ms, count: 2
Frame(s) that took ~19ms, count: 1
Frame(s) that took ~20ms, count: 3
Frame(s) that took ~21ms, count: 2
Frame(s) that took ~22ms, count: 2
Frame(s) that took ~27ms, count: 2
Frame(s) that took ~29ms, count: 2
Frame(s) that took ~32ms, count: 1
Frame(s) that took ~34ms, count: 1
Frame(s) that took ~35ms, count: 1
Frame(s) that took ~64ms, count: 1
Frame(s) that took ~67ms, count: 1
Frame(s) that took ~68ms, count: 2
Frame(s) that took ~69ms, count: 1
Frame(s) that took ~72ms, count: 3
Frame(s) that took ~74ms, count: 3
Average frame time: 21.99ms
No. of slow frames: 29

See maui#14275 for further detail about these changes.

Improve performance of Entry.MaxLength on Android

Investigating a .NET MAUI customer sample:

  • Navigating from a Shell flyout.

  • To a new page with several Entry controls.

  • There was a noticeable performance delay.

When profiling on a Pixel 5, one “hot path” was Entry.MaxLength:

18.52ms (0.22%) microsoft.maui!Microsoft.Maui.Platform.EditTextExtensions.UpdateMaxLength(Android.Widget.EditText,Microsoft.Maui.IEntry)
16.03ms (0.19%) microsoft.maui!Microsoft.Maui.Platform.EditTextExtensions.UpdateMaxLength(Android.Widget.EditText,int)
12.16ms (0.14%) microsoft.maui!Microsoft.Maui.Platform.EditTextExtensions.SetLengthFilter(Android.Widget.EditText,int)
  • EditTextExtensions.UpdateMaxLength() calls
  • EditText.Text getter and setter
  • EditTextExtensions.SetLengthFilter() calls
  • EditText.Get/SetFilters()

What happens is we end up marshaling strings and IInputFilter[] back and forth between C# and Java for every Entry control. All Entry controls go through this code path (even ones with a default value for MaxLength), so it made sense to move some of this code from C# to Java instead.

Our C# code before:

// C#

public static void UpdateMaxLength(this EditText editText, int maxLength)
{
    editText.SetLengthFilter(maxLength);

    var newText = editText.Text.TrimToMaxLength(maxLength);
    if (editText.Text != newText)
        editText.Text = newText;
}

public static void SetLengthFilter(this EditText editText, int maxLength)
{
    if (maxLength == -1)
        maxLength = int.MaxValue;

    var currentFilters = new List<IInputFilter>(editText.GetFilters() ?? new IInputFilter[0]);
    var changed = false;

    for (var i = 0; i < currentFilters.Count; i++)
    {
        if (currentFilters[i] is InputFilterLengthFilter)
        {
            currentFilters.RemoveAt(i);
            changed = true;
            break;
        }
    }

    if (maxLength >= 0)
    {
        currentFilters.Add(new InputFilterLengthFilter(maxLength));
        changed = true;
    }

    if (changed)
        editText.SetFilters(currentFilters.ToArray());
}

Moved to Java (with identical behavior) instead:

// Java

 /**
 * Sets the maxLength of an EditText
 * @param editText
 * @param maxLength
 */
public static void updateMaxLength(@NonNull EditText editText, int maxLength)
{
    setLengthFilter(editText, maxLength);

    if (maxLength < 0)
        return;

    Editable currentText = editText.getText();
    if (currentText.length() > maxLength) {
        editText.setText(currentText.subSequence(0, maxLength));
    }
}

/**
 * Updates the InputFilter[] of an EditText. Used for Entry and SearchBar.
 * @param editText
 * @param maxLength
 */
public static void setLengthFilter(@NonNull EditText editText, int maxLength)
{
    if (maxLength == -1)
        maxLength = Integer.MAX_VALUE;

    List<InputFilter> currentFilters = new ArrayList<>(Arrays.asList(editText.getFilters()));
    boolean changed = false;
    for (int i = 0; i < currentFilters.size(); i++) {
        InputFilter filter = currentFilters.get(i);
        if (filter instanceof InputFilter.LengthFilter) {
            currentFilters.remove(i);
            changed = true;
            break;
        }
    }

    if (maxLength >= 0) {
        currentFilters.add(new InputFilter.LengthFilter(maxLength));
        changed = true;
    }
    if (changed) {
        InputFilter[] newFilter = new InputFilter[currentFilters.size()];
        editText.setFilters(currentFilters.toArray(newFilter));
    }
}

This avoids marshaling (copying!) string and array values back and forth from C# to Java. With these changes in place, the calls to EditTextExtensions.UpdateMaxLength() are now so fast they are missing completely from dotnet-trace output, saving ~19ms when navigating to the page in the customer sample.

See maui#15614 for details about this improvement.

Improve memory usage of CollectionView on Windows

We reviewed a .NET MAUI customer sample with a CollectionView of 150,000 data-bound rows. Debugging what happens at runtime, .NET MAUI was effectively doing:

_itemTemplateContexts = new List<ItemTemplateContext>(capacity: 150_000);
for (int n = 0; n < 150_000; n++)
{
    _itemTemplateContexts.Add(null);
}

And then each item is created as it is scrolled into view:

if (_itemTemplateContexts[index] == null)
{
    _itemTemplateContexts[index] = context = new ItemTemplateContext(...);
}
return _itemTemplateContexts[index];

This wasn’t the best approach, but to improve things:

  • use a Dictionary<int, T> instead, just let it size dynamically.

  • use TryGetValue(..., out var context), so each call accesses the indexer one less time than before.

  • use either the bound collection’s size or 64 (whichever is smaller) as a rough estimate of how many might fit on screen at a time

Our code changes to:

if (!_itemTemplateContexts.TryGetValue(index, out var context))
{
    _itemTemplateContexts[index] = context = new ItemTemplateContext(...);
}
return context;

With these changes in place, a memory snapshot of the app after startup:

Before:
Heap Size: 82,899.54 KB
After:
Heap Size: 81,768.76 KB

Which is saving about 1MB of memory on launch. In this case, it feels better to just let the Dictionary size itself with an estimate of what capacity will be.

See maui#16838 for details about this improvement.

Use UnmanagedCallersOnlyAttribute on Apple platforms

When unmanaged code calls into managed code, such as invoking a callback from Objective-C, the [MonoPInvokeCallbackAttribute] was previously used in Xamarin.iOS, Xamarin.Mac, and .NET 6+ for this purpose. The [UnmanagedCallersOnlyAttribute] attribute came along as a modern replacement for this Mono feature, which is implemented in a way with performance in mind.

Unfortunately, there are a few restrictions when using this new attribute:

  • Method must be marked static.
  • Must not be called from managed code.
  • Must only have blittable arguments.
  • Must not have generic type parameters or be contained within a generic class.

Not only did we have to refactor the “code generator” that produces many of the bindings for Apple APIs for AppKit, UIKit, etc., but we also had many manual bindings that would need the same treatment.

The end result is that most callbacks from Objective-C to C# should be faster in .NET 8 than before. See xamarin-macios#10470 and xamarin-macios#15783 for details about these improvements.

Faster Java interop for strings on Android

When binding members which have parameter types or return types which are java.lang.CharSequence, the member is “overloaded” to replace CharSequence with System.String, and the “original” member has a Formatted suffix.

For example, consider android.widget.TextView, which has getText() and setText() methods which have parameter types and return types which are java.lang.CharSequence:

// Java
class TextView extends View {
    public CharSequence getText();
    public final void setText(CharSequence text);
}

When bound, this results in two properties:

// C#
class TextView : View {
    public Java.Lang.ICharSequence? TextFormatted { get; set; }
    public string? Text { get; set; }
}

The “non-Formatted overload” works by creating a temporary String object to invoke the Formatted overload, so the actual implementation looks like:

partial class TextView {
    public string? Text {
        get => TextFormatted?.ToString ();
        set {
            var jls = value == null ? null : new Java.Lang.String (value);
            TextFormatted = jls;
            jls?.Dispose ();
        }
    }
}

TextView.Text is much easer to understand & simpler to consume for .NET developers than TextView.TextFormatted.

A problem with the this approach is performance: creating a new Java.Lang.String instance requires:

  1. Creating the managed peer (the Java.Lang.String instance),
  2. Creating the native peer (the java.lang.String instance),
  3. And registering the mapping between (1) and (2)

And then immediately use and dispose the value…

This is particularly noticeable with .NET MAUI apps. Consider a customer sample, which uses XAML to set data-bound Text values in a CollectionView, which eventually hit TextView.Text. Profiling shows:

653.69ms (6.3%) mono.android!Android.Widget.TextView.set_Text(string)
198.05ms (1.9%) mono.android!Java.Lang.String..ctor(string)
121.57ms (1.2%) mono.android!Java.Lang.Object.Dispose()

6.3% of scrolling time is spent in the TextView.Text property setter!

Partially optimize this case: if the *Formatted member is (1) a property, and (2) not virtual, then we can directly call the Java setter method. This avoids the need to create a managed peer and to register a mapping between the peers:

partial class TextView {
    public string? Text {
        get => TextFormatted?.ToString (); // unchanged
        set {
            const string __id               = "setText.(Ljava/lang/CharSequence;)V";
            JniObjectReference native_value = JniEnvironment.Strings.NewString (value);
            try {
                JniArgumentValue* __args    = stackalloc JniArgumentValue [1];
                __args [0] = new JniArgumentValue (native_value);
                _members.InstanceMethods.InvokeNonvirtualVoidMethod (__id, this, __args);
            } finally {
                JniObjectReference.Dispose (ref native_value);
            }
        }
    }
}

With the result being:

Method Mean Error StdDev Allocated
Before SetFinalText 6.632 us 0.0101 us 0.0079 us 112 B
After SetFinalText 1.361 us 0.0022 us 0.0019 us

The TextView.Text property setter invocation time is reduced to 20% of the previous average invocation time.

Note that the virtual case is problematic for other reasons, but luckily enough TextView.setText() is non-virtual and likely one of the more commonly used Android APIs.

See java-interop#1101 for details about this improvement.

Faster Java interop for C# events on Android

Profiling a .NET MAUI customer sample while scrolling on a Pixel 5, We saw ~2.2% of the time spent in the IOnFocusChangeListenerImplementor constructor, due to a subscription to the View.FocusChange event:

(2.2%) mono.android!Android.Views.View.IOnFocusChangeListenerImplementor..ctor()

MAUI subscribes to Android.Views.View.FocusChange for every view placed on the screen, which happens while scrolling in this sample.

Reviewing the generated code for the IOnFocusChangeListenerImplementor constructor, we see it still uses outdated JNIEnv APIs:

public IOnFocusChangeListenerImplementor () : base (
        Android.Runtime.JNIEnv.StartCreateInstance ("mono/android/view/View_OnFocusChangeListenerImplementor", "()V"),
        JniHandleOwnership.TransferLocalRef
    )
{
    Android.Runtime.JNIEnv.FinishCreateInstance (((Java.Lang.Object) this).Handle, "()V");
}

Which we can change to use the newer/faster Java.Interop APIs:

public unsafe IOnFocusChangeListenerImplementor ()
    : base (IntPtr.Zero, JniHandleOwnership.DoNotTransfer)
{
    const string __id = "()V";
    if (((Java.Lang.Object) this).Handle != IntPtr.Zero)
        return;
    var h = JniPeerMembers.InstanceMethods.StartCreateInstance (__id, ((object) this).GetType (), null);
    SetHandle (h.Handle, JniHandleOwnership.TransferLocalRef);
    JniPeerMembers.InstanceMethods.FinishCreateInstance (__id, this, null);
}

These are better because the equivalent call to JNIEnv.FindClass() is cached, among other things. This was just one of the cases that was accidentally missed when we implemented the new Java.Interop APIs in the Xamarin timeframe. We simply needed to update our code generator to emit a better C# binding for this case.

After these changes, we saw instead results in dotnet-trace:

(0.81%) mono.android!Android.Views.View.IOnFocusChangeListenerImplementor..ctor()

This should improve the performance of all C# events that wrap Java listeners, a design-pattern commonly used in Java and Android applications. This includes the FocusedChanged event used by all .NET MAUI views on Android.

See java-interop#1105 for details about this improvement.

Use Function Pointers for JNI

There is various machinery and generated code that makes Java interop possible from C#. Take, for example, the following instance method foo() in Java:

// Java
object foo(object bar) {
    // returns some value
}

A C# method named CallObjectMethod is responsible for calling Java’s Native Interface (JNI) that calls into the JVM to actually invoke the Java method:

public static unsafe JniObjectReference CallObjectMethod (JniObjectReference instance, JniMethodInfo method, JniArgumentValue* args)
{
    //...
    IntPtr thrown;
    var tmp = NativeMethods.java_interop_jnienv_call_object_method_a (JniEnvironment.EnvironmentPointer, out thrown, instance.Handle, method.ID, (IntPtr) args);

    Exception __e = JniEnvironment.GetExceptionForLastThrowable (thrown);
    if (__e != null)
        ExceptionDispatchInfo.Capture (__e).Throw ();

    JniEnvironment.LogCreateLocalRef (tmp);
    return new JniObjectReference (tmp, JniObjectReferenceType.Local);
}

In Xamarin.Android, .NET 6, and .NET 7 all calls into Java went through a java_interop_jnienv_call_object_method_a p/invoke, which signature looks like:

[DllImport (JavaInteropLib, CallingConvention = CallingConvention.Cdecl, CharSet = CharSet.Ansi)]
internal static extern unsafe jobject java_interop_jnienv_call_object_method_a (IntPtr jnienv, out IntPtr thrown, jobject instance, IntPtr method, IntPtr args);

Which is implemented in C as:

JI_API jobject
java_interop_jnienv_call_object_method_a (JNIEnv *env, jthrowable *_thrown, jobject instance, jmethodID method, jvalue* args)
{
    *_thrown = 0;
    jobject _r_ = (*env)->CallObjectMethodA (env, instance, method, args);
    *_thrown = (*env)->ExceptionOccurred (env);
    return _r_;
}

C# 9 introduced function pointers that allowed us a way to simplify things slightly — and make them faster as a result.

So instead of using p/invoke in .NET 8, we could instead call a new unsafe method named CallObjectMethodA:

// Before:
var tmp = NativeMethods.java_interop_jnienv_call_object_method_a (JniEnvironment.EnvironmentPointer, out thrown, instance.Handle, method.ID, (IntPtr) args);
// After:
var tmp = JniNativeMethods.CallObjectMethodA (JniEnvironment.EnvironmentPointer, instance.Handle, method.ID, (IntPtr) args);

Which calls a C# function pointer directly:

[System.Runtime.CompilerServices.MethodImpl (System.Runtime.CompilerServices.MethodImplOptions.AggressiveInlining)]
internal static unsafe jobject CallObjectMethodA (IntPtr env, jobject instance, IntPtr method, IntPtr args)
{
    return (*((JNIEnv**)env))->CallObjectMethodA (env, instance, method, args);
}

This function pointer declared using the new syntax introduced in C# 9:

public delegate* unmanaged <IntPtr, jobject, IntPtr, IntPtr, jobject> CallObjectMethodA;

Comparing the two implementations with a manual benchmark:

# JIPinvokeTiming timing: 00:00:01.6993644
#    Average Invocation: 0.00016993643999999998ms
# JIFunctionPointersTiming timing: 00:00:01.6561349
#    Average Invocation: 0.00016561349ms

With a Release build, the average invocation time for JIFunctionPointersTiming takes 97% of the time as JIPinvokeTiming, i.e. is 3% faster. Additionally, using C# 9 function pointers means we can get rid of all of the java_interop_jnienv_*() C functions, which shrinks libmonodroid.so by ~55KB for each architecture.

See xamarin-android#8234 and java-interop#938 for details about this improvement.

Removed Xamarin.AndroidX.Legacy.Support.V4

Reviewing .NET MAUI’s Android dependencies, we noticed a suspicious package:

Xamarin.AndroidX.Legacy.Support.V4

If you are familiar with the Android Support Libraries, these are a set of packages Google provides to “polyfill” APIs to past versions of Android. This gives them a way to bring new APIs to old OS versions, since the Android ecosystem (OEMs, etc.) are much slower to upgrade as compared to iOS, for example. This particular package, Legacy.Support.V4, is actually support for Android as far back as Android API 4! The minimum supported Android version in .NET is Android API 21, which was released in 2017.

It turns out this dependency was brought over from Xamarin.Forms and was not actually needed. As expected from this change, lots of Java code was removed from .NET MAUI apps. So much, in fact, that .NET 8 MAUI applications are now under the multi-dex limit — all Dalvik bytecode can fix into a single classes.dex file.

A detailed breakdown of the size changes using apkdiff:

> apkdiff -f com.companyname.maui_before-Signed.apk com.companyname.maui_after-Signed.apk
Size difference in bytes ([*1] apk1 only, [*2] apk2 only):
+   1,598,040 classes.dex
-           6 META-INF/androidx.asynclayoutinflater_asynclayoutinflater.version *1
-           6 META-INF/androidx.legacy_legacy-support-core-ui.version *1
-           6 META-INF/androidx.legacy_legacy-support-v4.version *1
-           6 META-INF/androidx.media_media.version *1
-         455 assemblies/assemblies.blob
-         564 res/layout/notification_media_action.xml *1
-         744 res/layout/notification_media_cancel_action.xml *1
-       1,292 res/layout/notification_template_media.xml *1
-       1,584 META-INF/BNDLTOOL.SF
-       1,584 META-INF/MANIFEST.MF
-       1,696 res/layout/notification_template_big_media.xml *1
-       1,824 res/layout/notification_template_big_media_narrow.xml *1
-       2,456 resources.arsc
-       2,756 res/layout/notification_template_media_custom.xml *1
-       2,872 res/layout/notification_template_lines_media.xml *1
-       3,044 res/layout/notification_template_big_media_custom.xml *1
-       3,216 res/layout/notification_template_big_media_narrow_custom.xml *1
-   2,030,636 classes2.dex
Summary:
-      24,111 Other entries -0.35% (of 6,880,759)
-     432,596 Dalvik executables -3.46% (of 12,515,440)
+           0 Shared libraries 0.00% (of 12,235,904)
-     169,179 Package size difference -1.12% (of 15,123,185)

See dotnet/maui#12232 for details about this improvement.

Deduplication of generics on iOS and macOS

In .NET 7, iOS applications experienced app size increases due to C# generics usage across multiple .NET assemblies. When the .NET 7 Mono AOT compiler encounters a generic instance that is not handled by generic sharing, it will emit code for the instance. If the same instance is encountered during AOT compilation in multiple assemblies, the code will be emitted multiple times, increasing code size.

In .NET 8, new dedup-skip and dedup-include command-line options are passed to the Mono AOT compiler. A new aot-instances.dll assembly is created for sharing this information in one place throughout the application.

The change was tested on MySingleView app and Monotouch tests in the xamarin/xamarin-macios codebase:

App Baseline size on disk .ipa (MB) Target size on disk .ipa (MB) Baseline size on disk .app (MB) Target size on disk .app (MB) Baseline build time (s) Target build time (s) .app diff (%)
MySingleView Release iOS 5.4 5.4 29.2 15.2 29.2 16.8 47.9
MySingleView Release iOSSimulator-arm64 N/A N/A 469.5 341.8 468.0 330.0 27.2
Monotouch Release llvm iOS 49.0 38.8 209.6 157.4 115.0 130.0 24.9

See xamarin-macios#17766 for details about this improvement.

Fix System.Linq.Expressions implementation on iOS-like platforms

In .NET 7, codepaths in System.Linq.Expressions were controlled by various flags such as:

  • CanCompileToIL
  • CanEmitObjectArrayDelegate
  • CanCreateArbitraryDelegates

These flags were controlling codepaths which are “AOT friendly” and those that are not. For desktop platforms, NativeAOT specifies the following configuration for AOT-compatible code:

<IlcArg Include="--feature:System.Linq.Expressions.CanCompileToIL=false" /> 
<IlcArg Include="--feature:System.Linq.Expressions.CanEmitObjectArrayDelegate=false" /> 
<IlcArg Include="--feature:System.Linq.Expressions.CanCreateArbitraryDelegates=false" /> 

When it comes to iOS-like platforms, System.Linq.Expressions library was built with constant propagation enabled and control variables were removed. This further caused above-listed NativeAOT feature switches not to have any effect (fail to trim during app build), potentially causing the AOT compilation to follow unsupported code paths on these platforms.

In .NET8, we have unified the build of System.Linq.Expressions.dll shipping the same assembly for all supported platforms and runtimes, and simplified these switches to respect IsDynamicCodeSupported so that the .NET trimmer can remove the appropriate IL in System.Linq.Expressions.dll at application build time.

See dotnet/runtime#87924 and dotnet/runtime#89308 for details about this improvement.

Set DynamicCodeSupport=false for iOS and Catalyst

In .NET 8, the feature switch $(DynamicCodeSupport) is set to false for platforms:

  • Where it is not possible to publish without the AOT compiler.

  • When interpreter is not enabled.

Which boils down to applications running on iOS, tvOS, MacCatalyst, etc.

DynamicCodeSupport=false enables the .NET trimmer to remove code paths depending on RuntimeFeature.IsDynamicCodeSupported such as this example in System.Linq.Expressions.

Estimated size savings are:

dotnet new maui (ios) old SLE.dll new SLE.dll + DynamicCodeSupported=false diff (%)
Size on disk (Mb) 40,53 38,78 -4,31%
.pkg (Mb) 14,83 14,20 -4,21%

When combined with the System.Linq.Expressions improvements on iOS-like platforms, this showed a nice overall improvement to application size:

Maui Mono vs NativeAOT

See xamarin-macios#18555 for details about this improvement.

Memory Leaks

Memory Leaks and Quality

Given that the major theme for .NET MAUI in .NET 8 is quality, memory-related issues became a focal point for this release. Some of the problems found existed even in the Xamarin.Forms codebase, so we are happy to work towards a framework that developers can rely on for their cross-platform .NET applications.

For full details on the work completed in .NET 8, we’ve various PRs and Issues related to memory issues at:

You can see that considerable progress was made in .NET 8 in this area.

If we compare .NET 7 MAUI versus .NET 8 MAUI in a sample application running on Windows, displaying the results of GC.GetTotalMemory() on screen:

Comparison on Windows: .NET 7 vs .NET 8

Then compare the sample application running on macOS, but with many more pages pushed onto the navigation stack:

Comparison on Mac: .NET 7 vs .NET 8

See the sample code for this project on GitHub for further details.

Diagnosing leaks in .NET MAUI

The symptom of a memory leak in a .NET MAUI application, could be something like:

  • Navigate from the landing page to a sub page.

  • Go back.

  • Navigate to the sub page again.

  • Repeat.

  • Memory grows consistently until the OS closes the application due to lack of memory.

In the case of Android, you may see log messages such as:

07-07 18:51:39.090 17079 17079 D Mono : GC_MAJOR: (user request) time 137.21ms, stw 140.60ms los size: 10984K in use: 3434K
07-07 18:51:39.090 17079 17079 D Mono : GC_MAJOR_SWEEP: major size: 116192K in use: 108493K
07-07 18:51:39.092 17079 17079 I monodroid-gc: 46204 outstanding GREFs. Performing a full GC!

In this example, a 116MB heap is quite large for a mobile application, as well as over 46,000 C# <-> Java wrapper objects!

To truly determine if the sub page is leaking, we can make a couple modifications to a .NET MAUI application:

  1. Add logging in a finalizer. For example:
~MyPage() => Console.WriteLine("Finalizer for ~MyPage()");

While navigating through your app, you can find out if entire pages are living forever if the log message is never displayed. This is a common symptom of a leak, because any View holds .Parent.Parent.Parent, etc. all the way up to the Page object.

  1. Call GC.Collect() somewhere in the app, such as the sub page’s constructor:
public MyPage()
{
    GC.Collect(); // For debugging purposes only, remove later
    InitializeComponent();
}

This makes the GC more deterministic, in that we are forcing it to run more frequently. Each time we navigate to the sub page, we are more likely causing the old sub page’s to go away. If things are working properly, we should see the log message from the finalizer.

Note GC.Collect() is for debugging purposes only. You should not need this in your app after investigation is complete, so be sure to remove it afterward.

  1. With these changes in place, test a Release build of your app.

On iOS, Android, macOS, etc. you can watch console output of your app to determine what is actually happening at runtime. adb logcat, for example, is a way to view these logs on Android.

If running on Windows, you can also use Debug > Windows > Diagnostic Tools inside Visual Studio to take memory snapshots inside Visual Studio. In the future, we would like Visual Studio’s diagnostic tooling to support .NET MAUI applications running on other platforms.

See our memory leaks wiki page for more information related to memory leaks in .NET MAUI applications.

Patterns that cause leaks: C# events

C# events, just like a field, property, etc. can create strong references between objects. Let’s look at a situation where things can go wrong.

Take for example, the cross-platform Grid.ColumnDefinitions property:

public class Grid : Layout, IGridLayout
{
    public static readonly BindableProperty ColumnDefinitionsProperty = BindableProperty.Create("ColumnDefinitions",
        typeof(ColumnDefinitionCollection), typeof(Grid), null, validateValue: (bindable, value) => value != null,
        propertyChanged: UpdateSizeChangedHandlers, defaultValueCreator: bindable =>
        {
            var colDef = new ColumnDefinitionCollection();
            colDef.ItemSizeChanged += ((Grid)bindable).DefinitionsChanged;
            return colDef;
        });

    public ColumnDefinitionCollection ColumnDefinitions
    {
        get { return (ColumnDefinitionCollection)GetValue(ColumnDefinitionsProperty); }
        set { SetValue(ColumnDefinitionsProperty, value); }
    }
  • Grid has a strong reference to its ColumnDefinitionCollection via the BindableProperty.

  • ColumnDefinitionCollection has a strong reference to Grid via the ItemSizeChanged event.

If you put a breakpoint on the line with ItemSizeChanged +=, you can see the event has an EventHandler object where the Target is a strong reference back to the Grid.

In some cases, circular references like this are completely OK. The .NET runtime(s)’ garbage collectors know how to collect cycles of objects that point each other. When there is no “root” object holding them both, they can both go away.

The problem comes in with object lifetimes: what happens if the ColumnDefinitionCollection lives for the life of the entire application?

Consider the following Style in Application.Resources or Resources/Styles/Styles.xaml:

<Style TargetType="Grid" x:Key="GridStyleWithColumnDefinitions">
    <Setter Property="ColumnDefinitions" Value="18,*"/>
</Style>

If you applied this Style to a Grid on a random Page:

  • Application‘s main ResourceDictionary holds the Style.
  • The Style holds a ColumnDefinitionCollection.
  • The ColumnDefinitionCollection holds the Grid.
  • Grid unfortunately holds the Page via .Parent.Parent.Parent, etc.

This situation could cause entire Page‘s to live forever!

Note The issue with Grid is fixed in maui#16145, but is an excellent example of illustrating how C# events can go wrong.

Circular references on Apple platforms

Even since the early days of Xamarin.iOS, there has existed an issue with “circular references” even in a garbage-collected runtime like .NET. C# objects co-exist with a reference-counted world on Apple platforms, and so a C# object that subclasses NSObject can run into situations where they can accidentally live forever — a memory leak. This is not a .NET-specific problem, as you can just as easily create the same situation in Objective-C or Swift. Note that this does not occur on Android or Windows platforms.

Take for example, the following circular reference:

class MyViewSubclass : UIView
{
    public UIView? Parent { get; set; }

    public void Add(MyViewSubclass subview)
    {
        subview.Parent = this;
        AddSubview(subview);
    }
}

//...

var parent = new MyViewSubclass();
var view = new MyViewSubclass();
parent.Add(view);

In this case:

  • parent -> view via Subviews
  • view -> parent via the Parent property
  • The reference count of both objects is non-zero.
  • Both objects live forever.

This problem isn’t limited to a field or property, you can create similar situations with C# events:

class MyView : UIView
{
    public MyView()
    {
        var picker = new UIDatePicker();
        AddSubview(picker);
        picker.ValueChanged += OnValueChanged;
    }

    void OnValueChanged(object? sender, EventArgs e) { }

    // Use this instead and it doesn't leak!
    //static void OnValueChanged(object? sender, EventArgs e) { }
}

In this case:

  • MyView -> UIDatePicker via Subviews
  • UIDatePicker -> MyView via ValueChanged and EventHandler.Target
  • Both objects live forever.

A solution for this example, is to make OnValueChanged method static, which would result in a null Target on the EventHandler instance.

Another solution, would be to put OnValueChanged in a non-NSObject subclass:

class MyView : UIView
{
    readonly Proxy _proxy = new();

    public MyView()
    {
        var picker = new UIDatePicker();
        AddSubview(picker);
        picker.ValueChanged += _proxy.OnValueChanged;
    }

    class Proxy
    {
        public void OnValueChanged(object? sender, EventArgs e) { }
    }
}

This is the pattern we’ve used in most .NET MAUI handlers and other UIView subclasses.

See the MemoryLeaksOniOS sample repo, if you would like to play with some of these scenarios in isolation in an iOS application without .NET MAUI.

Roslyn analyzer for Apple platforms

We also have an experimental Roslyn Analyzer that can detect these situations at build time. To add it to net7.0-ios, net8.0-ios, etc. projects, you can simply install a NuGet package:

<PackageReference Include="MemoryAnalyzers" Version="0.1.0-beta.3" PrivateAssets="all" />

Some examples of a warning would be:

public class MyView : UIView
{
    public event EventHandler MyEvent;
}
Event 'MyEvent' could could memory leaks in an NSObject subclass.
Remove the event or add the [UnconditionalSuppressMessage("Memory", "MA0001")]
attribute with a justification as to why the event will not leak.

Note that the analyzer can warns if there might be an issue, so it can be quite noisy to enable in a large, existing codebase. Inspecting memory at runtime is the best way to determine if there is truly a memory leak.

Tooling and Documentation

Simplified dotnet-trace and dotnet-dsrouter

In .NET 7, profiling a mobile application was a bit of a challenge. You had to run dotnet-dsrouter and dotnet-trace together and get all the settings right to be able to retrieve a .nettrace or speedscope file for performance investigations. There was also no built-in support for dotnet-gcdump to connect to dotnet-dsrouter to get memory snapshots of a running .NET MAUI application.

In .NET 8, we’ve streamlined this scenario by making new commands for dotnet-dsrouter that simplifies the workflow.

To verify you have the latest diagnostic tooling, you can install them via:

$ dotnet tool install -g dotnet-dsrouter
You can invoke the tool using the following command: dotnet-dsrouter
Tool 'dotnet-dsrouter' was successfully installed.
$ dotnet tool install -g dotnet-gcdump
You can invoke the tool using the following command: dotnet-gcdump
Tool 'dotnet-gcdump' was successfully installed.
$ dotnet tool install -g dotnet-trace
You can invoke the tool using the following command: dotnet-trace
Tool 'dotnet-trace' was successfully installed.

Verify you have at least 8.x versions of these tools:

$ dotnet tool list -g
Package Id                         Version                       Commands
--------------------------------------------------------------------------------------
dotnet-dsrouter                    8.0.452401                    dotnet-dsrouter
dotnet-gcdump                      8.0.452401                    dotnet-gcdump
dotnet-trace                       8.0.452401                    dotnet-trace

To profile an Android application on an Android emulator, first build and install your application in Release mode such as:

$ dotnet build -f net8.0-android -t:Install -c Release -p:AndroidEnableProfiler=true
Build SUCCEEDED.
    0 Warning(s)
    0 Error(s)

Next, open a terminal to run dotnet-dsrouter

$ dotnet-dsrouter android-emu
Start an application on android emulator with one of the following environment variables set:
DOTNET_DiagnosticPorts=10.0.2.2:9000,nosuspend,connect
DOTNET_DiagnosticPorts=10.0.2.2:9000,suspend,connect

Then in a second terminal window, we can set the debug.mono.profile Android system property, as the stand-in for $DOTNET_DiagnosticPorts:

$ adb shell setprop debug.mono.profile '10.0.2.2:9000,suspend,connect'
$ dotnet-trace ps
3248  dotnet-dsrouter
$ dotnet-trace collect -p 3248 --format speedscope
...
[00:00:00:09]   Recording trace 3.2522   (MB)
Press <Enter> or <Ctrl+C> to exit...

Note Android doesn’t have good support for environment variables like $DOTNET_DiagnosticPorts. You can create an AndroidEnvironment text file for setting environment variables, but Android system properties can be simpler as they would not require rebuilding the application to set them.

Upon launching the Android application, it should be able to connect to dotnet-dsrouter -> dotnet-trace and record performance profiling information for investigation. The --format argument is optional and it defaults to .nettrace. However, .nettrace files can be viewed only with Perfview on Windows, while the speedscope JSON files can be viewed “on” macOS or Linux by uploading them to https://speedscope.app.

Note When providing a process ID to dotnet-trace, it knows how to tell if a process ID is dotnet-dsrouter and connect through it appropriately.

dotnet-dsrouter has the following new commands to simplify the workflow:

  • dotnet-dsrouter android: Android devices
  • dotnet-dsrouter android-emu: Android emulators
  • dotnet-dsrouter ios: iOS devices
  • dotnet-dsrouter ios-sim: iOS simulators

See the .NET MAUI wiki for more information about profiling .NET MAUI applications on each platform.

dotnet-gcdump Support for Mobile

In .NET 7, we had a somewhat complex method (see wiki) for getting a memory snapshot of an application on the Mono runtime (such as iOS or Android). You had to use a Mono-specific event provider such as:

dotnet-trace collect --diagnostic-port /tmp/maui-app --providers Microsoft-DotNETRuntimeMonoProfiler:0xC900001:4

And then we relied on Filip Navara’s mono-gcdump tool (thanks Filip!) to convert the .nettrace file to .gcdump to be opened in Visual Studio or PerfView.

In .NET 8, we now have dotnet-gcdump support for mobile scenarios. If you want to get a memory snapshot of a running application, you can use dotnet-gcdump in a similar fashion as dotnet-trace:

$ dotnet-gcdump ps
3248  dotnet-dsrouter
$ dotnet-gcdump collect -p 3248
Writing gcdump to '20231018_115631_29880.gcdump'...

Note This requires the exact same setup as dotnet-trace, such as -p:AndroidEnableProfiler=true, dotnet-dsrouter, adb commands, etc.

This greatly streamlines our workflow for investigating memory leaks in .NET MAUI applications. See our memory leaks wiki page for more information.

Author

Jonathan Peppers
Principal Software Engineer

I used to build apps in C#. Now build *tools* to build apps in C#. Engineer on the .NET Android / .NET MAUI team.

8 comments

Discussion is closed. Login to edit/delete existing comments.

  • Karabin, Josh

    Is it safe to assume that 10.0.2.2 is an IP address for the host computer, reachable either via emulator or an Android device via wifi?

  • Ashish Khanal

    Great job. Thank you for your contributions.

  • Yunus Emre AYGÜN

    Turkey Test mentioned! 🇹🇷🇹🇷🐺🐺💪💪

    Jokes aside, keep up the great job! Every single time I see performance improvements in major .NET version, I am like "No way they managed to improve performance this much!" and you guys still show the same dedication!

    Born too late to become an assembly programmer, born too early to be a hippie Gen Z Python/JavaScript programmer, but born just in time to become a Microsoft .NET ecosystem fanboy!

    Read more
  • Eder Cardoso

    Amazing article and work. 👏

  • Syed Hoque

    Awesome work! Thanks a lot for the detective work and the solution!

  • Rafael Saccomani

    Thanks for the hard work in .net maui, we, from Brazil, loveit!!

  • Emmanuel Adebiyi

    Wheeew, that’s a lot of updates. A solid testament to the team’s hardwork and dedication!