Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[Xamarin.Android.Build.Tasks] Invokers after abstracts in typemaps #9545

Merged
merged 3 commits into from
Nov 22, 2024

Conversation

jonpryor
Copy link
Member

@jonpryor jonpryor commented Nov 21, 2024

Fixes: #9535

Context: a017561
Context: 35f41dc

Create a new .NET for Android project (dotnet new android), add
a @(PackageReference) to Xamarin.AndroidX.Media3.ExoPlayer,
and add the following code:

var builder = new AndroidX.Media3.ExoPlayer.ExoPlayerBuilder(this);
var player = builder.Build();

Build and run in Debug configuration, and it crashes:

System.ArgumentException: Could not determine Java type corresponding to `AndroidX.Media3.ExoPlayer.IExoPlayerInvoker, Xamarin.AndroidX.Media3.ExoPlayer, Version=1.0.0.0, Culture=neutral, PublicKeyToken=null`. (Parameter 'targetType')
   at Java.Interop.TypeManager.CreateInstance(IntPtr handle, JniHandleOwnership transfer, Type targetType) in /Users/runner/work/1/s/xamarin-android/src/Mono.Android/Java.Interop/TypeManager.cs:line 323
   at Java.Lang.Object.GetObject(IntPtr handle, JniHandleOwnership transfer, Type type) in /Users/runner/work/1/s/xamarin-android/src/Mono.Android/Java.Lang/Object.cs:line 302
   at Java.Lang.Object._GetObject[IExoPlayer](IntPtr handle, JniHandleOwnership transfer) in /Users/runner/work/1/s/xamarin-android/src/Mono.Android/Java.Lang/Object.cs:line 288
   at Java.Lang.Object.GetObject[IExoPlayer](IntPtr handle, JniHandleOwnership transfer) in /Users/runner/work/1/s/xamarin-android/src/Mono.Android/Java.Lang/Object.cs:line 280
   at AndroidX.Media3.ExoPlayer.ExoPlayerBuilder.Build() in D:\a\_work\1\s\generated\androidx.media3.media3-exoplayer\obj\Release\net8.0-android\generated\src\AndroidX.Media3.ExoPlayer.IExoPlayer.cs:line 749
   at media3.MainActivity.OnCreate(Bundle savedInstanceState) in /Users/moljac/Downloads/1036/media3/MainActivity.cs:line 16
   at Android.App.Activity.n_OnCreate_Landroid_os_Bundle_(IntPtr jnienv, IntPtr native__this, IntPtr native_savedInstanceState) in /Users/runner/work/1/s/xamarin-android/src/Mono.Android/obj/Release/net9.0/android-35/mcw/Android.App.Activity.cs:line 3196
   at Android.Runtime.JNINativeWrapper.Wrap_JniMarshal_PPL_V(_JniMarshal_PPL_V callback, IntPtr jnienv, IntPtr klazz, IntPtr p0) in /Users/runner/work/1/s/xamarin-android/src/Mono.Android/Android.Runtime/JNINativeWrapper.g.cs:line 121

☹️

The problem, as often appears to be the case these days, is an
unexpected interaction between different things.

Firstly, the immediate cause of the crash is 35f41dc, which
updated TypeManager.CreateInstance() to attempt to lookup the Java
type that corresponds to a possible *Invoker type:

// `type` is `typeof(AndroidX.Media3.ExoPlayer.IExoPlayerInvoker)`
var typeSig  = JNIEnvInit.androidRuntime?.TypeManager.GetTypeSignature (type) ?? default;
if (!typeSig.IsValid || typeSig.SimpleReference == null) {
        throw new ArgumentException ($"Could not determine Java type corresponding to `{type.AssemblyQualifiedName}`.", nameof (targetType));
}

The underlying cause of the crash, and what 35f41dc uncovered, is
that there is no type mapping for IExoPlayerInvoker. This is
(not quite) readily seen by examining the generated file:

% grep IExoPlayerInvoker obj/Debug/net*android*/android/typemaps.arm64-v8a.ll
# no match

This is also true on .NET 8: there is no type mapping for
IExoPlayerInvoker.

There are type mappings for other interface types, e.g.
IIntSupplier:

% grep IIntSupplier obj/Debug/net*android*/android/typemaps.arm64-v8a.ll                    
@.TypeMapEntry.16871_from = private unnamed_addr constant [47 x i8] c"Java.Util.Functions.IIntSupplier, Mono.Android\00", align 1
@.TypeMapEntry.16873_from = private unnamed_addr constant [54 x i8] c"Java.Util.Functions.IIntSupplierInvoker, Mono.Android\00", align 1

A bit more digging around, and the "problem" is as follows: type map
generation had a type ordering requirement, and needed the
Invoker types to follow their corresponding abstract types.

# Mono.Android.dll: IIntSupplier is before IIntSupplierInvoker
% monodis --typedef obj/Debug/net9.0-android/android/assets/arm64-v8a/Mono.Android.dll | grep IIntSupplier
7676: Java.Util.Functions.IIntSupplier (flist=57465, mlist=135759, flags=0x1000a1, extends=0x0)
7677: Java.Util.Functions.IIntSupplierInvoker (flist=57465, mlist=135760, flags=0x100000, extends=0x7c28)

# Xamarin.AndroidX.Media3.ExoPlayer.dll: IExoPlayerInvoker is *before* IExoPlayer
% monodis --typedef obj/Debug/net9.0-android/android/assets/arm64-v8a/Xamarin.AndroidX.Media3.ExoPlayer.dll| grep IExoPlayer
139: AndroidX.Media3.ExoPlayer.IExoPlayerInvoker (flist=20, mlist=567, flags=0x100000, extends=0x85)
…
166: AndroidX.Media3.ExoPlayer.IExoPlayer (flist=598, mlist=1671, flags=0x1000a1, extends=0x0)

Reasonable aside: why is IExoPlayerInvoker before IExoPlayer?
Because it's a "real" checked-in copy of generated code, not
generated code that follows the IExoPlayer generated declaration.

The reason why ordering matters comes from a017561:

Update typemap generation code in Xamarin.Android.Build.Tasks.dll so
that all the duplicate Java type names will point to the same managed
type name. Additionally, make sure we select the managed type in the
same fashion the old typemap generator in Java.Interop worked: prefer
base types to the derived ones if the type is an interface or an
abstract class. The effect of this change is that no matter which
entry EmbeddedAssemblies::binary_search() ends up selecting it will
always return the same managed type name for all aliased Java types.

The TypeMapGenerator.HandleDebugDuplicates() method from a017561
implicitly preferred the first TypeDefinition encountered for a given
Java type. When the *Invoker came first, it attempted to treat it
as a duplicate, but wound up removing it entirely:

oldEntry.TypeDefinition = td;
oldEntry.ManagedName = GetManagedTypeName (td);

oldEntry.TypeDefinition would refer to IExoPlayerInvoker, and
setting it to td would instead set it to IExoPlayer.
IExoPlayerInvoker is lost.

Fix this by updating HandleDebugDuplicates() to be clearer about
intent:

// oldEntry == typeof(IExoPlayerInvoker)
// td == typeof(IExoPlayer)
if ((td.IsAbstract || td.IsInterface) &&
		!oldEntry.TypeDefinition.IsAbstract &&
		!oldEntry.TypeDefinition.IsInterface &&
		td.IsAssignableFrom (oldEntry.TypeDefinition, cache)) {
	// We found the `Invoker` type *before* the declared type
	// Fix things up so the abstract type is first, and the `Invoker` is considered a duplicate.
	duplicates.Insert (0, entry);
	oldEntry.SkipInJavaToManaged = false;
}

Now, when the "new" TypeDefinition is:

  1. An abstract class or interface type, and
  2. The original entry is neither an abstract class nor interface, &
  3. The old entry can be assigned to the new entry,

then we assume that this is a *Invoker scenario, ensure things are
consistent with the "expected" { IExoPlayer, IExoPlayerInvoker } order:

  1. Inserting the new TypeDefinition first in duplicates.

    This ensures expected ordering.

  2. Setting oldEntry.SkipInJavaToManaged to false.

    This ensures that IExoPlayerInvoker is emitted in the typemaps.

WORKAROUND: Building the app in Release configuration avoids
the crash. (Then cry about inner dev loop performance.)

"But what about tests?" Originally, @jonpryor thought that
JavaObjectExtensionsTests.JavaCast_InterfaceCast() was a sufficient
test. This was wrong, because it didn't test what it thought it did:

IntPtr g;
using (var n = new Java.Lang.Integer (42)) {
	g = JNIEnv.NewGlobalRef (n.Handle);
}
// We want a Java.Lang.Object so that we create an IComparableInvoker
// instead of just getting back the original instance.
using (var o = Java.Lang.Object.GetObject<Java.Lang.Object> (g, JniHandleOwnership.TransferGlobalRef)) {
	var c = JavaObjectExtensions.JavaCast<Java.Lang.IComparable> (o);
	c.Dispose ();
}

The comment is not quite right: while it says we want a
Java.Lang.Object, that's not the runtime type o.GetType().
The runtime type o.GetType() would be Java.Lang.Integer (!),
because the first thing that TypeManager.CreateInstance() tries
to do is see if we have an existing type mapping for the runtime type
of the JNI handle. As the runtime type is java.lang.Integer, we
do have such a binding, and that bound type is implicitly
convertible to Java.Lang.Object.

Meaning this test never actually tested an interface cast!

So for starters, we need to fix this to something that ensures we
get an unbound type in the base class hierarchy, so that we can
actually test interface checking behavior.

Enter @(AndroidJavaSource), ValueProvider.java, and
Example.java. Example.getValueProvider() returns an anonymous
inner class that implements ValueProvider, ensuring that the only
bound type that will work is Java.Lang.Object. We then can use
JavaObjectExtensions.JavaCast<T>() as originally intended.

We could have used any interface type instead of introducing a new
ValueProvider interface, but introducing a new type is the only way
to attempt to reproduce the TypeDefinition ordering issue, by
adding a new partial class for IValueProviderInvoker.

Locally, we get the intended "wrong" ordering of Invoker before
abstract type:

% monodis --typedef tests/Mono.Android-Tests/Runtime-Microsoft.Android.Sdk/obj/Debug/net9.0-android/Mono.Android.NET-Tests.dll | grep IValueProvider
5: Net.Dot.Android.Test.IValueProviderInvoker (flist=10, mlist=20, flags=0x100000, extends=0x89)
7: Net.Dot.Android.Test.IValueProvider (flist=13, mlist=37, flags=0x1000a1, extends=0x0)

and the test passes as desired.

@jonpryor jonpryor requested a review from grendello as a code owner November 21, 2024 23:18
Fixes: #9535

Context: a017561
Context: 35f41dc

Create a new .NET for Android project (`dotnet new android`), add
a `@(PackageReference)` to [Xamarin.AndroidX.Media3.ExoPlayer][0],
and add the following code:

	var builder = new AndroidX.Media3.ExoPlayer.ExoPlayerBuilder(this);
	var player = builder.Build();

Build and run in *Debug* configuration, and it crashes:

	System.ArgumentException: Could not determine Java type corresponding to `AndroidX.Media3.ExoPlayer.IExoPlayerInvoker, Xamarin.AndroidX.Media3.ExoPlayer, Version=1.0.0.0, Culture=neutral, PublicKeyToken=null`. (Parameter 'targetType')
	   at Java.Interop.TypeManager.CreateInstance(IntPtr handle, JniHandleOwnership transfer, Type targetType) in /Users/runner/work/1/s/xamarin-android/src/Mono.Android/Java.Interop/TypeManager.cs:line 323
	   at Java.Lang.Object.GetObject(IntPtr handle, JniHandleOwnership transfer, Type type) in /Users/runner/work/1/s/xamarin-android/src/Mono.Android/Java.Lang/Object.cs:line 302
	   at Java.Lang.Object._GetObject[IExoPlayer](IntPtr handle, JniHandleOwnership transfer) in /Users/runner/work/1/s/xamarin-android/src/Mono.Android/Java.Lang/Object.cs:line 288
	   at Java.Lang.Object.GetObject[IExoPlayer](IntPtr handle, JniHandleOwnership transfer) in /Users/runner/work/1/s/xamarin-android/src/Mono.Android/Java.Lang/Object.cs:line 280
	   at AndroidX.Media3.ExoPlayer.ExoPlayerBuilder.Build() in D:\a\_work\1\s\generated\androidx.media3.media3-exoplayer\obj\Release\net8.0-android\generated\src\AndroidX.Media3.ExoPlayer.IExoPlayer.cs:line 749
	   at media3.MainActivity.OnCreate(Bundle savedInstanceState) in /Users/moljac/Downloads/1036/media3/MainActivity.cs:line 16
	   at Android.App.Activity.n_OnCreate_Landroid_os_Bundle_(IntPtr jnienv, IntPtr native__this, IntPtr native_savedInstanceState) in /Users/runner/work/1/s/xamarin-android/src/Mono.Android/obj/Release/net9.0/android-35/mcw/Android.App.Activity.cs:line 3196
	   at Android.Runtime.JNINativeWrapper.Wrap_JniMarshal_PPL_V(_JniMarshal_PPL_V callback, IntPtr jnienv, IntPtr klazz, IntPtr p0) in /Users/runner/work/1/s/xamarin-android/src/Mono.Android/Android.Runtime/JNINativeWrapper.g.cs:line 121

☹️

The problem, as often appears to be the case these days, is an
unexpected interaction between different things.

Firstly, the *immediate* cause of the crash is 35f41dc, which
updated `TypeManager.CreateInstance()` to attempt to lookup the Java
type that corresponds to a possible `*Invoker` type:

	// `type` is `typeof(AndroidX.Media3.ExoPlayer.IExoPlayerInvoker)`
	var typeSig  = JNIEnvInit.androidRuntime?.TypeManager.GetTypeSignature (type) ?? default;
	if (!typeSig.IsValid || typeSig.SimpleReference == null) {
	        throw new ArgumentException ($"Could not determine Java type corresponding to `{type.AssemblyQualifiedName}`.", nameof (targetType));
	}

The underlying cause of the crash, and what 35f41dc uncovered, is
that *there is no type mapping for `IExoPlayerInvoker`*.  This is
(not quite) readily seen by examining the generated file:

	% grep IExoPlayerInvoker obj/Debug/net*android*/android/typemaps.arm64-v8a.ll
	# no match

This is also true on .NET 8: there is no type mapping for
`IExoPlayerInvoker`.

There *are* type mappings for *other* interface types, e.g.
`IIntSupplier`:

	% grep IIntSupplier obj/Debug/net*android*/android/typemaps.arm64-v8a.ll
	@.TypeMapEntry.16871_from = private unnamed_addr constant [47 x i8] c"Java.Util.Functions.IIntSupplier, Mono.Android\00", align 1
	@.TypeMapEntry.16873_from = private unnamed_addr constant [54 x i8] c"Java.Util.Functions.IIntSupplierInvoker, Mono.Android\00", align 1

A bit more digging around, and the "problem" is as follows: type map
generation had a *type ordering* requirement, and needed the
`Invoker` types to *follow* their corresponding abstract types.

	# Mono.Android.dll: IIntSupplier is before IIntSupplierInvoker
	% monodis --typedef obj/Debug/net9.0-android/android/assets/arm64-v8a/Mono.Android.dll | grep IIntSupplier
	7676: Java.Util.Functions.IIntSupplier (flist=57465, mlist=135759, flags=0x1000a1, extends=0x0)
	7677: Java.Util.Functions.IIntSupplierInvoker (flist=57465, mlist=135760, flags=0x100000, extends=0x7c28)

	# Xamarin.AndroidX.Media3.ExoPlayer.dll: IExoPlayerInvoker is *before* IExoPlayer
	% monodis --typedef obj/Debug/net9.0-android/android/assets/arm64-v8a/Xamarin.AndroidX.Media3.ExoPlayer.dll| grep IExoPlayer
	139: AndroidX.Media3.ExoPlayer.IExoPlayerInvoker (flist=20, mlist=567, flags=0x100000, extends=0x85)
	…
	166: AndroidX.Media3.ExoPlayer.IExoPlayer (flist=598, mlist=1671, flags=0x1000a1, extends=0x0)

Reasonable aside: *why* is `IExoPlayerInvoker` before `IExoPlayer`?
Because it's a ["real" checked-in copy of generated code][1], not
generated code that follows the `IExoPlayer` generated declaration.

The reason why ordering matters comes from a017561:

> Update typemap generation code in `Xamarin.Android.Build.Tasks.dll` so
> that all the duplicate Java type names will point to the same managed
> type name.  Additionally, make sure we select the managed type in the
> same fashion the old typemap generator in `Java.Interop` worked: prefer
> base types to the derived ones if the type is an interface or an
> abstract class.  The effect of this change is that no matter which
> entry `EmbeddedAssemblies::binary_search()` ends up selecting it will
> always return the same managed type name for all aliased Java types.

The `TypeMapGenerator.HandleDebugDuplicates()` method from a017561
implicitly preferred the first TypeDefinition encountered for a given
Java type.  When the `*Invoker` came first, it attempted to treat it
as a duplicate, but wound up *removing* it entirely:

	oldEntry.TypeDefinition = td;
	oldEntry.ManagedName = GetManagedTypeName (td);

`oldEntry.TypeDefinition` would refer to `IExoPlayerInvoker`, and
setting it to `td` would instead set it to `IExoPlayer`.
`IExoPlayerInvoker` is lost.

Fix this by updating `HandleDebugDuplicates()` to be clearer about
intent:

	// oldEntry == typeof(IExoPlayerInvoker)
	// td == typeof(IExoPlayer)
	if ((td.IsAbstract || td.IsInterface) &&
			!oldEntry.TypeDefinition.IsAbstract &&
			!oldEntry.TypeDefinition.IsInterface &&
			td.IsAssignableFrom (oldEntry.TypeDefinition, cache)) {
		// We found the `Invoker` type *before* the declared type
		// Fix things up so the abstract type is first, and the `Invoker` is considered a duplicate.
		duplicates.Insert (0, entry);
		oldEntry.SkipInJavaToManaged = false;
	}

Now, when the "new" TypeDefinition is:

 1. An abstract class or interface type, and
 2. The original entry is *neither* an abstract class nor interface, &
 3. The old entry can be assigned to the new entry,

then we assume that this is a `*Invoker` scenario, ensure things are
consistent with the "expected" { IExoPlayer, IExoPlayerInvoker } order:

 1. Inserting the new TypeDefinition *first* in `duplicates`.

    This ensures expected ordering.

 2. Setting `oldEntry.SkipInJavaToManaged` to `false`.

    This ensures that `IExoPlayerInvoker` is emitted in the typemaps.

***WORKAROUND***: Building the app in Release configuration avoids
the crash.  (Then cry about inner dev loop performance.)

"But what about tests?"  Originally, @jonpryor thought that
`JavaObjectExtensionsTests.JavaCast_InterfaceCast()` was a sufficient
test.  This was wrong, because it didn't test what it thought it did:

	IntPtr g;
	using (var n = new Java.Lang.Integer (42)) {
		g = JNIEnv.NewGlobalRef (n.Handle);
	}
	// We want a Java.Lang.Object so that we create an IComparableInvoker
	// instead of just getting back the original instance.
	using (var o = Java.Lang.Object.GetObject<Java.Lang.Object> (g, JniHandleOwnership.TransferGlobalRef)) {
		var c = JavaObjectExtensions.JavaCast<Java.Lang.IComparable> (o);
		c.Dispose ();
	}

The comment is not quite right: while it *says* we want a
`Java.Lang.Object`, that's not the runtime type `o.GetType()`.
The runtime type `o.GetType()` would be `Java.Lang.Integer` (!),
*because* the first thing that `TypeManager.CreateInstance()` tries
to do is see if we have an existing type mapping for the runtime type
of the JNI handle.  As the runtime type is `java.lang.Integer`, we
*do* have such a binding, *and* that bound type is implicitly
convertible to `Java.Lang.Object`.

Meaning this test never actually tested an interface cast!

So for starters, we need to fix this to something that *ensures* we
get an unbound type in the base class hierarchy, so that we can
actually test interface checking behavior.

Enter `@(AndroidJavaSource)`, `ValueProvider.java`, and
`Example.java`.  `Example.getValueProvider()` returns an anonymous
inner class that implements `ValueProvider`, ensuring that the only
bound type that will work is `Java.Lang.Object`.  We then can use
`JavaObjectExtensions.JavaCast<T>()` as originally intended.

We could have used any interface type instead of introducing a new
`ValueProvider` interface, but introducing a new type is the only way
to *attempt* to reproduce the TypeDefinition ordering issue, by
adding a new `partial` class for `IValueProviderInvoker`.

*Locally*, we get the intended "wrong" ordering of Invoker before
abstract type:

	% monodis --typedef tests/Mono.Android-Tests/Runtime-Microsoft.Android.Sdk/obj/Debug/net9.0-android/Mono.Android.NET-Tests.dll | grep IValueProvider
	5: Net.Dot.Android.Test.IValueProviderInvoker (flist=10, mlist=20, flags=0x100000, extends=0x89)
	7: Net.Dot.Android.Test.IValueProvider (flist=13, mlist=37, flags=0x1000a1, extends=0x0)

and the test passes as desired.

[0]: https://www.nuget.org/packages/Xamarin.AndroidX.Media3.ExoPlayer/1.4.1
[1]: https://github.com/dotnet/android-libraries/blob/3c21f4643c5dfe74e64fec54e3679baf4dc8d067/source/androidx.media3/media3-exoplayer/Additions/AndroidX.Media3.ExoPlayer.IExoPlayer.cs#L10
@jonpryor jonpryor changed the title [Xamarin.Android.Build.Tasks] Allow invokers before abstracts [Xamarin.Android.Build.Tasks] Invokers after abstracts in typemaps Nov 22, 2024
@jonpryor jonpryor added this to the .NET 9 Servicing milestone Nov 22, 2024
@jonpryor jonpryor merged commit 7009522 into main Nov 22, 2024
58 checks passed
@jonpryor jonpryor deleted the dev/jonp/jonp-fix-android-libraries-1036 branch November 22, 2024 10:32
jonpryor added a commit that referenced this pull request Nov 22, 2024
…9545)

Fixes: #9535

Context: a017561
Context: 35f41dc

Create a new .NET for Android project (`dotnet new android`), add
a `@(PackageReference)` to [Xamarin.AndroidX.Media3.ExoPlayer][0],
and add the following code:

	var builder = new AndroidX.Media3.ExoPlayer.ExoPlayerBuilder(this);
	var player = builder.Build();

Build and run in *Debug* configuration, and it crashes:

	System.ArgumentException: Could not determine Java type corresponding to `AndroidX.Media3.ExoPlayer.IExoPlayerInvoker, Xamarin.AndroidX.Media3.ExoPlayer, Version=1.0.0.0, Culture=neutral, PublicKeyToken=null`. (Parameter 'targetType')
	   at Java.Interop.TypeManager.CreateInstance(IntPtr handle, JniHandleOwnership transfer, Type targetType) in /Users/runner/work/1/s/xamarin-android/src/Mono.Android/Java.Interop/TypeManager.cs:line 323
	   at Java.Lang.Object.GetObject(IntPtr handle, JniHandleOwnership transfer, Type type) in /Users/runner/work/1/s/xamarin-android/src/Mono.Android/Java.Lang/Object.cs:line 302
	   at Java.Lang.Object._GetObject[IExoPlayer](IntPtr handle, JniHandleOwnership transfer) in /Users/runner/work/1/s/xamarin-android/src/Mono.Android/Java.Lang/Object.cs:line 288
	   at Java.Lang.Object.GetObject[IExoPlayer](IntPtr handle, JniHandleOwnership transfer) in /Users/runner/work/1/s/xamarin-android/src/Mono.Android/Java.Lang/Object.cs:line 280
	   at AndroidX.Media3.ExoPlayer.ExoPlayerBuilder.Build() in D:\a\_work\1\s\generated\androidx.media3.media3-exoplayer\obj\Release\net8.0-android\generated\src\AndroidX.Media3.ExoPlayer.IExoPlayer.cs:line 749
	   at media3.MainActivity.OnCreate(Bundle savedInstanceState) in /Users/moljac/Downloads/1036/media3/MainActivity.cs:line 16
	   at Android.App.Activity.n_OnCreate_Landroid_os_Bundle_(IntPtr jnienv, IntPtr native__this, IntPtr native_savedInstanceState) in /Users/runner/work/1/s/xamarin-android/src/Mono.Android/obj/Release/net9.0/android-35/mcw/Android.App.Activity.cs:line 3196
	   at Android.Runtime.JNINativeWrapper.Wrap_JniMarshal_PPL_V(_JniMarshal_PPL_V callback, IntPtr jnienv, IntPtr klazz, IntPtr p0) in /Users/runner/work/1/s/xamarin-android/src/Mono.Android/Android.Runtime/JNINativeWrapper.g.cs:line 121

☹️

The problem, as often appears to be the case these days, is an
unexpected interaction between different things.

Firstly, the *immediate* cause of the crash is 35f41dc, which
updated `TypeManager.CreateInstance()` to attempt to lookup the Java
type that corresponds to a possible `*Invoker` type:

	// `type` is `typeof(AndroidX.Media3.ExoPlayer.IExoPlayerInvoker)`
	var typeSig  = JNIEnvInit.androidRuntime?.TypeManager.GetTypeSignature (type) ?? default;
	if (!typeSig.IsValid || typeSig.SimpleReference == null) {
	        throw new ArgumentException ($"Could not determine Java type corresponding to `{type.AssemblyQualifiedName}`.", nameof (targetType));
	}

The underlying cause of the crash, and what 35f41dc uncovered, is
that *there is no type mapping for `IExoPlayerInvoker`*.  This is
(not quite) readily seen by examining the generated file:

	% grep IExoPlayerInvoker obj/Debug/net*android*/android/typemaps.arm64-v8a.ll
	# no match

This is also true on .NET 8: there is no type mapping for
`IExoPlayerInvoker`.

There *are* type mappings for *other* interface types, e.g.
`IIntSupplier`:

	% grep IIntSupplier obj/Debug/net*android*/android/typemaps.arm64-v8a.ll
	@.TypeMapEntry.16871_from = private unnamed_addr constant [47 x i8] c"Java.Util.Functions.IIntSupplier, Mono.Android\00", align 1
	@.TypeMapEntry.16873_from = private unnamed_addr constant [54 x i8] c"Java.Util.Functions.IIntSupplierInvoker, Mono.Android\00", align 1

A bit more digging around, and the "problem" is as follows: type map
generation had a *type ordering* requirement, and needed the
`Invoker` types to *follow* their corresponding abstract types.

	# Mono.Android.dll: IIntSupplier is before IIntSupplierInvoker
	% monodis --typedef obj/Debug/net9.0-android/android/assets/arm64-v8a/Mono.Android.dll | grep IIntSupplier
	7676: Java.Util.Functions.IIntSupplier (flist=57465, mlist=135759, flags=0x1000a1, extends=0x0)
	7677: Java.Util.Functions.IIntSupplierInvoker (flist=57465, mlist=135760, flags=0x100000, extends=0x7c28)

	# Xamarin.AndroidX.Media3.ExoPlayer.dll: IExoPlayerInvoker is *before* IExoPlayer
	% monodis --typedef obj/Debug/net9.0-android/android/assets/arm64-v8a/Xamarin.AndroidX.Media3.ExoPlayer.dll| grep IExoPlayer
	139: AndroidX.Media3.ExoPlayer.IExoPlayerInvoker (flist=20, mlist=567, flags=0x100000, extends=0x85)
	…
	166: AndroidX.Media3.ExoPlayer.IExoPlayer (flist=598, mlist=1671, flags=0x1000a1, extends=0x0)

Reasonable aside: *why* is `IExoPlayerInvoker` before `IExoPlayer`?
Because it's a ["real" checked-in copy of generated code][1], not
generated code that follows the `IExoPlayer` generated declaration.

The reason why ordering matters comes from a017561:

> Update typemap generation code in `Xamarin.Android.Build.Tasks.dll` so
> that all the duplicate Java type names will point to the same managed
> type name.  Additionally, make sure we select the managed type in the
> same fashion the old typemap generator in `Java.Interop` worked: prefer
> base types to the derived ones if the type is an interface or an
> abstract class.  The effect of this change is that no matter which
> entry `EmbeddedAssemblies::binary_search()` ends up selecting it will
> always return the same managed type name for all aliased Java types.

The `TypeMapGenerator.HandleDebugDuplicates()` method from a017561
implicitly preferred the first TypeDefinition encountered for a given
Java type.  When the `*Invoker` came first, it attempted to treat it
as a duplicate, but wound up *removing* it entirely:

	oldEntry.TypeDefinition = td;
	oldEntry.ManagedName = GetManagedTypeName (td);

`oldEntry.TypeDefinition` would refer to `IExoPlayerInvoker`, and
setting it to `td` would instead set it to `IExoPlayer`.
`IExoPlayerInvoker` is lost.

Fix this by updating `HandleDebugDuplicates()` to be clearer about
intent:

	// oldEntry == typeof(IExoPlayerInvoker)
	// td == typeof(IExoPlayer)
	if ((td.IsAbstract || td.IsInterface) &&
			!oldEntry.TypeDefinition.IsAbstract &&
			!oldEntry.TypeDefinition.IsInterface &&
			td.IsAssignableFrom (oldEntry.TypeDefinition, cache)) {
		// We found the `Invoker` type *before* the declared type
		// Fix things up so the abstract type is first, and the `Invoker` is considered a duplicate.
		duplicates.Insert (0, entry);
		oldEntry.SkipInJavaToManaged = false;
	}

Now, when the "new" TypeDefinition is:

 1. An abstract class or interface type, and
 2. The original entry is *neither* an abstract class nor interface, &
 3. The old entry can be assigned to the new entry,

then we assume that this is a `*Invoker` scenario, ensure things are
consistent with the "expected" { IExoPlayer, IExoPlayerInvoker } order:

 1. Inserting the new TypeDefinition *first* in `duplicates`.

    This ensures expected ordering.

 2. Setting `oldEntry.SkipInJavaToManaged` to `false`.

    This ensures that `IExoPlayerInvoker` is emitted in the typemaps.

***WORKAROUND***: Building the app in Release configuration avoids
the crash.  (Then cry about inner dev loop performance.)

"But what about tests?"  Originally, @jonpryor thought that
`JavaObjectExtensionsTests.JavaCast_InterfaceCast()` was a sufficient
test.  This was wrong, because it didn't test what it thought it did:

	IntPtr g;
	using (var n = new Java.Lang.Integer (42)) {
		g = JNIEnv.NewGlobalRef (n.Handle);
	}
	// We want a Java.Lang.Object so that we create an IComparableInvoker
	// instead of just getting back the original instance.
	using (var o = Java.Lang.Object.GetObject<Java.Lang.Object> (g, JniHandleOwnership.TransferGlobalRef)) {
		var c = JavaObjectExtensions.JavaCast<Java.Lang.IComparable> (o);
		c.Dispose ();
	}

The comment is not quite right: while it *says* we want a
`Java.Lang.Object`, that's not the runtime type `o.GetType()`.
The runtime type `o.GetType()` would be `Java.Lang.Integer` (!),
*because* the first thing that `TypeManager.CreateInstance()` tries
to do is see if we have an existing type mapping for the runtime type
of the JNI handle.  As the runtime type is `java.lang.Integer`, we
*do* have such a binding, *and* that bound type is implicitly
convertible to `Java.Lang.Object`.

Meaning this test never actually tested an interface cast!

So for starters, we need to fix this to something that *ensures* we
get an unbound type in the base class hierarchy, so that we can
actually test interface checking behavior.

Enter `@(AndroidJavaSource)`, `ValueProvider.java`, and
`Example.java`.  `Example.getValueProvider()` returns an anonymous
inner class that implements `ValueProvider`, ensuring that the only
bound type that will work is `Java.Lang.Object`.  We then can use
`JavaObjectExtensions.JavaCast<T>()` as originally intended.

We could have used any interface type instead of introducing a new
`ValueProvider` interface, but introducing a new type is the only way
to *attempt* to reproduce the TypeDefinition ordering issue, by
adding a new `partial` class for `IValueProviderInvoker`.

*Locally*, we get the intended "wrong" ordering of Invoker before
abstract type:

	% monodis --typedef tests/Mono.Android-Tests/Runtime-Microsoft.Android.Sdk/obj/Debug/net9.0-android/Mono.Android.NET-Tests.dll | grep IValueProvider
	5: Net.Dot.Android.Test.IValueProviderInvoker (flist=10, mlist=20, flags=0x100000, extends=0x89)
	7: Net.Dot.Android.Test.IValueProvider (flist=13, mlist=37, flags=0x1000a1, extends=0x0)

and the test passes as desired.

[0]: https://www.nuget.org/packages/Xamarin.AndroidX.Media3.ExoPlayer/1.4.1
[1]: https://github.com/dotnet/android-libraries/blob/3c21f4643c5dfe74e64fec54e3679baf4dc8d067/source/androidx.media3/media3-exoplayer/Additions/AndroidX.Media3.ExoPlayer.IExoPlayer.cs#L10
@thisisthekap
Copy link

@jonpryor Any news on when a new service release is going to be published? We are depending on this fix to unblock us from upgrading to .net 9

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

[Media3] Attempting to create ExoPlayer instance crashes
3 participants