-
Notifications
You must be signed in to change notification settings - Fork 4.8k
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
Allow Marshal.GetFunctionPointerForDelegate() to work with generic types #32963
Comments
Marshal.GetFunctionPointerForDelegate for generic types is not straightforward to implement in CoreCLR, and it is even more difficult to implement a good AOT compilation scheme for it. The recommend way to do this for .NET 5 and beyond is going to use the My recommendation would be:
|
@jonpryor The discussion was had and decided that it wasn't the best solution due to the better approach described by @jkotas above. Does the recommendation above make sense to you? |
I'm not sure it does, likely because I'm not fully understanding the "function pointers" proposal. What we need to be able to do is call typedef struct {
char *name;
char *signature;
void *fnPtr;
} JNINativeMethod;
typedef const struct JNINativeInterface_ *JNIEnv
/* partial */ struct JNINativeInterface_ {
int (*RegisterNatives)(JNIEnv *env, class class, const JNINativeMethod *methods, int nMethods);
} What we currently do is emit code such as: partial class /* Java.Lang. */ Object {
static Delegate cb_toString;
static Delegate GetToStringHandler ()
{
if (cb_toString == null)
cb_toString = JNINativeWrapper.CreateDelegate ((Func<IntPtr, IntPtr, IntPtr>) n_ToString);
return cb_toString;
}
static IntPtr n_ToString (IntPtr jnienv, IntPtr native__this)
{
Java.Lang.Object __this = global::Java.Lang.Object.GetObject<Java.Lang.Object> (jnienv, native__this, JniHandleOwnership.DoNotTransfer);
return JNIEnv.NewString (__this.ToString ());
}
} We then lookup
We don't explicitly invoke I'm not immediately sure how to convert the above code into C#8 function pointers. I'd almost certainly need to drop the use of public struct JniNativeMethodRegistration {
public string Name;
public string Signature;
public IntPtr Marshaler;
public JniNativeMethodRegistration (string name, string signature, IntPtr marshaler)
{
Name = name ?? throw new ArgumentNullException (nameof (name));
Signature = signature ?? throw new ArgumentNullException (nameof (signature));
Marshaler = marshaler == IntPtr.Zero ? throw new ArgumentNullException (nameof (marshaler)) : marshaler;
}
} But I still need to get that
but that in no means I can pass
Additionally, note the I'm going to go out on a limb and guess that there's no way to use System.Reflection.Emit-generated delegate instances with I am not immediately convinced that Function Pointers are a path forward, and updating our generator to instead avoid |
If the method is marked with So you would produce this (no static fields, no delegate objects):
And then pass the address of
You should be able to emit method with |
@jonpryor See #33005 for the |
Fixes: #631 Context: dotnet/runtime#32963 Context: https://github.com/dotnet/csharplang/blob/master/proposals/function-pointers.md *Of `Delegate`s and JNI Callbacks…* ~~ Background ~~ In order for Java code to invoke Managed Code such as C#, several things must happen: 1. There must be a Java class which declares `native` methods. 2. The Java class' `native` methods must be [*resolvable*][0] Java `native` method resolution can be done by [C function name][1] *or* by using [`JNIEnv::RegisterNatives()`][2]: // C++ struct JNINativeMethod { const char *name; const char *signature; const void *fnPtr; }; /* partial */ struct JNIEnv { jint RegisterNatives(jclass clazz, const JNINativeMethod *methods, jint nMethods); }; `JNINativeMethods::fnPtr` is a pointer to a *C callable function* that accepts [JNI Native Method Arguments][3]. Java.Interop doesn't currently support resolution via C function name, and instead binds the `JNINativeMethod` struct as `JniNativeMethodRegistration`, and `JNIEnv::RegisterNatives()` as `Java.Interop.JniEnvironment.Types.RegisterNatives()`: // C# public partial struct JniNativeMethodRegistration { public string Name; public string Signature; public Delegate Marshaler; } public partial class JniEnvironment { public partial class Types { public static void RegisterNatives (JniObjectReference type, JniNativeMethodRegistration [] methods); } } Through the glory that is [Platform Invoke Delegate Marshaling][4] and/or [`Marshal.GetFunctionPointerForDelegate()`][5], managed code can provide a `Delegate` instance in `JniNativeMethodRegistration.Marshaler` and have JNI invoke that delegate when the corresponding Java `native` method is invoked. `tools/generator` is responsible for emitting this glue code, e.g. in order to support registering overrides of [`java.lang.Object.equals()`][6]: // C# emitted by `tools/generator`: namespace Java.Lang { partial class Object { static Delegate cb_equals_Ljava_lang_Object_; static Delegate GetEquals_Ljava_lang_Object_Handler () { if (cb_equals_Ljava_lang_Object_ == null) cb_equals_Ljava_lang_Object_ = JNINativeWrapper.CreateDelegate ((Func<IntPtr, IntPtr, IntPtr, bool>) n_Equals_Ljava_lang_Object_); return cb_equals_Ljava_lang_Object_; } static bool n_Equals_Ljava_lang_Object_ (IntPtr jnienv, IntPtr native__this, IntPtr native_obj) { var __this = global::Java.Lang.Object.GetObject<Java.Lang.Object> (jnienv, native__this, JniHandleOwnership.DoNotTransfer); var obj = global::Java.Lang.Object.GetObject<Java.Lang.Object> (native_obj, JniHandleOwnership.DoNotTransfer); bool __ret = __this.Equals (obj); return __ret; } } } `Object.n_Equals_Ljava_lang_Object()` is stored in a `Func<IntPtr, IntPtr, IntPtr, bool>` -- which conforms to JNI Native Method Arguments -- and is then provided to [`JNINativeWrapper.CreateDelegate()`][7], which uses `System.Reflection.Emit` to "wrap" `n_Equals_Ljava_lang_Object()` for exception propagation purposes. Eventually and ultimately, when a C# class overrides `Java.Lang.Object.Equals()`, `Object.GetEquals_Ljava_lang_Object_Handler()` will be invoked at runtime, and `Object.cb_equals_Ljava_lang_Object` will be stored into `JniNativeMethodRegistration.Marshaler`. ~~ `Action<…>` and `Func<…>` ~~ There is one problem with the above approach: its use of the `System.Action<…>` and `System.Func<…>` types used at the core of registering native methods with JNI. There are two problems with using these sets of types: 1. These delegate types only permit up to 16 parameters. Given that *two* parameters are always "eaten" by the `JNIEnv*` pointer and a `jobject` to Java's `this` or a `jclass` to the declaring class, that means that we can only bind methods taking up to 14 methods. Java methods which take more than 14 methods are skipped. 2. .NET Framework and CoreCLR don't support using generic types with the Platform Invoke marshaler and [`Marshal.GetFunctionPointerForDelegate()`][8]. (1) has been a longstanding problem, which we've been ignoring. (2) isn't *yet* a problem, and is something @jonpryor has been keen to address for awhile. ~~ C# Function Pointers? ~~ There is a proposal to [add Function Pointers to the C# language][9]. This would permit reduced overheads and improved efficiencies in obtaining a function pointer to pass into Java code. Unfortunately: 1. The proposal is still ongoing, with no known release date. 2. .NET Framework 4.x won't support them. 3. They can't be used within the current Xamarin.Android architecture. There doesn't appear to be a way to obtain a `Delegate` from a `delegate*`, which means `JNINativeWrapper.CreateDelegate()` cannot be used with Function Pointers. In order to use Function Pointers, we would likely need to *require* use of `tools/jnimarshalmethod-gen.exe` (176240d) so that appropriate JNI Native Method Argument-conforming methods with the `NativeCallableAttribute` can be generated at app build time, *avoiding* the current Reflection-heavy registration path which involves e.g. `Object.GetEquals_Ljava_lang_Object_Handler()`. Unfortunately, `jnimarshalmethod-gen.exe` isn't "done": it doesn't work on Windows, and it's use of `AppDomain`s and `System.Reflection.Emit` look to complicate a future .NET 5 port. ~~ Solution: Generate Delegates ~~ If `Action<…>` and `Func<…>` are to be avoided, and Function Pointers are out, how do we support more than 14 parameters? By updating `generator` to emit the required delegate types. When `Action<…>` or `Func<…>` would previously have been generated, instead emit *and record the name of* a delegate which follows the pattern: * Type name prefix: `_JniMarshal_PP` * Parameter types, using JNI encoding, e.g. `Z` for boolean, `I` for int, etc. *Reference types*, normally encoded as `L…;` and Arrays, encoded as `[`, are each encoded as `L`. Kotlin unsigned types are encoded as *lower-case* forms of the corresponding JNI types, e.g. `i` is an unsigned `I`. * Another `_`. * The above type encoding for the return type. For example, `Object.n_Equals_Ljava_lang_Object()` used `Func<IntPtr, IntPtr, IntPtr, bool>`. This would become `_JniMarshal_PPL_Z`. After the initial binding stage is complete and all required delegate types are recorded, the `_JniMarshal*` types are emitted into `__NamespaceMapping__.cs`: internal delegate bool _JniMarshal_PPL_Z (IntPtr jnienv, IntPtr klass, IntPtr a); The cost to declaring all these types is that a binding assembly contains more types. `Mono.Android.dll`, for example, grows ~20KB in size from all the required delegate declarations, pre-linking. ~~ Other ~~ Remove `tools/generator/generator.sln` and replace it with a `tools/generator/generator.slnf` solution filter file which makes it easier to work with `generator` in Visual Studio by only loading needed projects from `Java.Interop.sln`. [0]: https://docs.oracle.com/javase/8/docs/technotes/guides/jni/spec/design.html#compiling_loading_and_linking_native_methods [1]: https://docs.oracle.com/javase/8/docs/technotes/guides/jni/spec/design.html#resolving_native_method_names [2]: https://docs.oracle.com/javase/8/docs/technotes/guides/jni/spec/functions.html#RegisterNatives [3]: https://docs.oracle.com/javase/8/docs/technotes/guides/jni/spec/design.html#native_method_arguments [4]: https://docs.microsoft.com/en-us/dotnet/framework/interop/marshaling-a-delegate-as-a-callback-method [5]: https://docs.microsoft.com/en-us/dotnet/api/system.runtime.interopservices.marshal.getfunctionpointerfordelegate?view=netcore-3.1 [6]: https://docs.oracle.com/javase/7/docs/api/java/lang/Object.html#equals%28java.lang.Object%29 [7]: https://github.com/xamarin/xamarin-android/blob/42822e0488185cdf4bca7c0bd05b21ad03dfbd7e/src/Mono.Android/Android.Runtime/JNINativeWrapper.cs#L34-L97 [8]: dotnet/runtime#32963 [9]: https://github.com/dotnet/csharplang/blob/master/proposals/function-pointers.md
Fixes: #631 Context: dotnet/runtime#32963 Context: https://github.com/dotnet/csharplang/blob/master/proposals/function-pointers.md *Of `Delegate`s and JNI Callbacks…* ~~ Background ~~ In order for Java code to invoke Managed Code such as C#, several things must happen: 1. There must be a Java class which declares `native` methods. 2. The Java class' `native` methods must be [*resolvable*][0] Java `native` method resolution can be done by [C function name][1] *or* by using [`JNIEnv::RegisterNatives()`][2]: // C++ struct JNINativeMethod { const char *name; const char *signature; const void *fnPtr; }; /* partial */ struct JNIEnv { jint RegisterNatives(jclass clazz, const JNINativeMethod *methods, jint nMethods); }; `JNINativeMethods::fnPtr` is a pointer to a *C callable function* that accepts [JNI Native Method Arguments][3]. Java.Interop doesn't currently support resolution via C function name, and instead binds the `JNINativeMethod` struct as `JniNativeMethodRegistration`, and `JNIEnv::RegisterNatives()` as `Java.Interop.JniEnvironment.Types.RegisterNatives()`: // C# public partial struct JniNativeMethodRegistration { public string Name; public string Signature; public Delegate Marshaler; } public partial class JniEnvironment { public partial class Types { public static void RegisterNatives (JniObjectReference type, JniNativeMethodRegistration [] methods); } } Through the glory that is [Platform Invoke Delegate Marshaling][4] and/or [`Marshal.GetFunctionPointerForDelegate()`][5], managed code can provide a `Delegate` instance in `JniNativeMethodRegistration.Marshaler` and have JNI invoke that delegate when the corresponding Java `native` method is invoked. `tools/generator` is responsible for emitting this glue code, e.g. in order to support registering overrides of [`java.lang.Object.equals()`][6]: // C# emitted by `tools/generator`: namespace Java.Lang { partial class Object { static Delegate cb_equals_Ljava_lang_Object_; static Delegate GetEquals_Ljava_lang_Object_Handler () { if (cb_equals_Ljava_lang_Object_ == null) cb_equals_Ljava_lang_Object_ = JNINativeWrapper.CreateDelegate ((Func<IntPtr, IntPtr, IntPtr, bool>) n_Equals_Ljava_lang_Object_); return cb_equals_Ljava_lang_Object_; } static bool n_Equals_Ljava_lang_Object_ (IntPtr jnienv, IntPtr native__this, IntPtr native_obj) { var __this = global::Java.Lang.Object.GetObject<Java.Lang.Object> (jnienv, native__this, JniHandleOwnership.DoNotTransfer); var obj = global::Java.Lang.Object.GetObject<Java.Lang.Object> (native_obj, JniHandleOwnership.DoNotTransfer); bool __ret = __this.Equals (obj); return __ret; } } } `Object.n_Equals_Ljava_lang_Object()` is stored in a `Func<IntPtr, IntPtr, IntPtr, bool>` -- which conforms to JNI Native Method Arguments -- and is then provided to [`JNINativeWrapper.CreateDelegate()`][7], which uses `System.Reflection.Emit` to "wrap" `n_Equals_Ljava_lang_Object()` for exception propagation purposes. Eventually and ultimately, when a C# class overrides `Java.Lang.Object.Equals()`, `Object.GetEquals_Ljava_lang_Object_Handler()` will be invoked at runtime, and `Object.cb_equals_Ljava_lang_Object` will be stored into `JniNativeMethodRegistration.Marshaler`. ~~ `Action<…>` and `Func<…>` ~~ There is one problem with the above approach: its use of the `System.Action<…>` and `System.Func<…>` types used at the core of registering native methods with JNI. There are two problems with using these sets of types: 1. These delegate types only permit up to 16 parameters. Given that *two* parameters are always "eaten" by the `JNIEnv*` pointer and a `jobject` to Java's `this` or a `jclass` to the declaring class, that means that we can only bind methods taking up to 14 methods. Java methods which take more than 14 methods are skipped. 2. .NET Framework and CoreCLR don't support using generic types with the Platform Invoke marshaler and [`Marshal.GetFunctionPointerForDelegate()`][8]. (1) has been a longstanding problem, which we've been ignoring. (2) isn't *yet* a problem, and is something @jonpryor has been keen to address for awhile. ~~ C# Function Pointers? ~~ There is a proposal to [add Function Pointers to the C# language][9]. This would permit reduced overheads and improved efficiencies in obtaining a function pointer to pass into Java code. Unfortunately: 1. The proposal is still ongoing, with no known release date. 2. .NET Framework 4.x won't support them. 3. They can't be used within the current Xamarin.Android architecture. There doesn't appear to be a way to obtain a `Delegate` from a `delegate*`, which means `JNINativeWrapper.CreateDelegate()` cannot be used with Function Pointers. In order to use Function Pointers, we would likely need to *require* use of `tools/jnimarshalmethod-gen.exe` (176240d) so that appropriate JNI Native Method Argument-conforming methods with the `NativeCallableAttribute` can be generated at app build time, *avoiding* the current Reflection-heavy registration path which involves e.g. `Object.GetEquals_Ljava_lang_Object_Handler()`. Unfortunately, `jnimarshalmethod-gen.exe` isn't "done": it doesn't work on Windows, and it's use of `AppDomain`s and `System.Reflection.Emit` look to complicate a future .NET 5 port. ~~ Solution: Generate Delegates ~~ If `Action<…>` and `Func<…>` are to be avoided, and Function Pointers are out, how do we support more than 14 parameters? By updating `generator` to emit the required delegate types. When `Action<…>` or `Func<…>` would previously have been generated, instead emit *and record the name of* a delegate which follows the pattern: * Type name prefix: `_JniMarshal_PP` * Parameter types, using JNI encoding, e.g. `Z` for boolean, `I` for int, etc. *Reference types*, normally encoded as `L…;` and Arrays, encoded as `[`, are each encoded as `L`. Kotlin unsigned types are encoded as *lower-case* forms of the corresponding JNI types, e.g. `i` is an unsigned `I`. * Another `_`. * The above type encoding for the return type. For example, `Object.n_Equals_Ljava_lang_Object()` used `Func<IntPtr, IntPtr, IntPtr, bool>`. This would become `_JniMarshal_PPL_Z`. After the initial binding stage is complete and all required delegate types are recorded, the `_JniMarshal*` types are emitted into `__NamespaceMapping__.cs`: internal delegate bool _JniMarshal_PPL_Z (IntPtr jnienv, IntPtr klass, IntPtr a); The cost to declaring all these types is that a binding assembly contains more types. `Mono.Android.dll`, for example, grows ~20KB in size from all the required delegate declarations, pre-linking. ~~ Other ~~ Remove `tools/generator/generator.sln` and replace it with a `tools/generator/generator.slnf` solution filter file which makes it easier to work with `generator` in Visual Studio by only loading needed projects from `Java.Interop.sln`. [0]: https://docs.oracle.com/javase/8/docs/technotes/guides/jni/spec/design.html#compiling_loading_and_linking_native_methods [1]: https://docs.oracle.com/javase/8/docs/technotes/guides/jni/spec/design.html#resolving_native_method_names [2]: https://docs.oracle.com/javase/8/docs/technotes/guides/jni/spec/functions.html#RegisterNatives [3]: https://docs.oracle.com/javase/8/docs/technotes/guides/jni/spec/design.html#native_method_arguments [4]: https://docs.microsoft.com/en-us/dotnet/framework/interop/marshaling-a-delegate-as-a-callback-method [5]: https://docs.microsoft.com/en-us/dotnet/api/system.runtime.interopservices.marshal.getfunctionpointerfordelegate?view=netcore-3.1 [6]: https://docs.oracle.com/javase/7/docs/api/java/lang/Object.html#equals%28java.lang.Object%29 [7]: https://github.com/xamarin/xamarin-android/blob/42822e0488185cdf4bca7c0bd05b21ad03dfbd7e/src/Mono.Android/Android.Runtime/JNINativeWrapper.cs#L34-L97 [8]: dotnet/runtime#32963 [9]: https://github.com/dotnet/csharplang/blob/master/proposals/function-pointers.md
CoreCLR doesn't support it. Ref: dotnet/runtime#32963
CoreCLR doesn't support it. Ref: dotnet/runtime#32963
…delegate. * CoreCLR doesn't support generic Action delegates in reverse (P/Invokes), so we need to find a different solution. * The native CGPDFOperatorTable callback API is not very friendly to us, because we can't pass it callback-specific data, which means that the managed caller must conform to the FullAOT requirement of the managed callback (must be a static function; must have a MonoPInvokeCallback attribute). * We can leverage the new function pointer syntax in C# 9 to make these requirements enforced by the C# compiler (unmanaged function pointer + UnmanagedCallersOnly attribute). The resulting API is still clunky to use, but I don't see any way around that. Fixes this monotouch-test failure with CoreCLR: [FAIL] Tamarin : System.Runtime.InteropServices.MarshalDirectiveException : Cannot marshal 'parameter #3': Non-blittable generic types cannot be marshaled. at CoreGraphics.CGPDFOperatorTable.CGPDFOperatorTableSetCallback(IntPtr table, String name, Action`2 callback) at CoreGraphics.CGPDFOperatorTable.SetCallback(String name, Action`2 callback) at MonoTouchFixtures.CoreGraphics.PDFScannerTest.Tamarin() in xamarin-macios/tests/monotouch-test/CoreGraphics/PDFScannerTest.cs:line 102 Ref: dotnet/runtime#32963
…delegate. (#11560) * CoreCLR doesn't support generic Action delegates in reverse (P/Invokes), so we need to find a different solution. * The native CGPDFOperatorTable callback API is not very friendly to us, because we can't pass it callback-specific data, which means that the managed caller must conform to the FullAOT requirement of the managed callback (must be a static function; must have a MonoPInvokeCallback attribute). * We can leverage the new function pointer syntax in C# 9 to make these requirements enforced by the C# compiler (unmanaged function pointer + UnmanagedCallersOnly attribute). The resulting API is still clunky to use, but I don't see any way around that. Fixes this monotouch-test failure with CoreCLR: [FAIL] Tamarin : System.Runtime.InteropServices.MarshalDirectiveException : Cannot marshal 'parameter #3': Non-blittable generic types cannot be marshaled. at CoreGraphics.CGPDFOperatorTable.CGPDFOperatorTableSetCallback(IntPtr table, String name, Action`2 callback) at CoreGraphics.CGPDFOperatorTable.SetCallback(String name, Action`2 callback) at MonoTouchFixtures.CoreGraphics.PDFScannerTest.Tamarin() in xamarin-macios/tests/monotouch-test/CoreGraphics/PDFScannerTest.cs:line 102 Ref: dotnet/runtime#32963
This is a needed feature that will remove need for casting all |
There is a difference between Mono & .NET Framework/.NET Core regarding the use of
Marshal.GetFunctionPointerForDelegate()
and generic delegate types. Consider:On .NET Framework, this app fails:
This works on Mono.
.NET Core currently has the same check as .NET Framework:
runtime/src/coreclr/src/vm/comdelegate.cpp
Line 1228 in fcd862e
Rationale: Xamarin.Android currently uses
System.Action<...>
andSystem.Func<...>
withMarshal.GetFunctionPointerForDelegate()
to register function pointers with JNI. This has always worked on Mono, and Xamarin.Android uses mono, so it's been Fine.However, it would be nice to use Xamarin.Android's JNI infrastructure on .NET Core. At present, if this were to be done it would fail as soon as we hit
Marshal.GetFunctionPointerForDelegate()
for method registration.The Xamarin.Android team could instead alter their binding infrastructure so that
Action<...>
andFunc<...>
are not used. The Xamarin.Android team would like to know if this code generator change is required for eventual .NET Core support, or if we can instead forego this change.Thanks to the Similar issues window, #4547 was suggested. I'm not sure if this is entirely duplicative or not, but Issue #4547 is currently Closed, though it was also added to the .NET 5 milestone, so I'm not entirely sure if these are the same or not.
The text was updated successfully, but these errors were encountered: