NativeAOT TODO list

[jonathanpeppers]

What do we need to do for NativeAOT?

• dotnet new android && dotnet publish -p:PublishAot=true
  • [nativeaot] introduce Microsoft.Android.Runtime.NativeAOT.dll #9760
• dotnet new maui runs: [WIP] turn NativeAOT sample into a .NET MAUI app #9747
• Parse Android system properties like debug.mono.log, to disable lref/gref logging
  • [Microsoft.Android.Runtime.NativeAOT] Parse debug.mono.log #9824
• Setup an alternative to monodroid_log() for NativeAOT
• MSBuild warning for non-Mono runtimes, that says "experimental"
  • [Xamarin.Android.Build.Tasks] emit XA1040 for experimental runtimes #9834
• Default $(PublishAotUsingRuntimePack)=true, and see what xamarin/xamarin-macios does
  • [nativeaot] default to $(PublishAotUsingRuntimePack)=true #9815
• Get Mono.Android-Tests running on NativeAOT
• Mono-specific startup settings like Timezone notifications, unhandled exceptions, System.Threading.Tasks SyncContext.
• Solve all NativeAOT trimmer warnings
• statically link libc++_shared.so

[jonpryor]

Alternate "MVP 1.1" typemap idea, for posterity: instead of a Dictionary<string, Type>, have two fields:

1. Dictionary<string, RuntimeTypeHandle> or string[]+RuntimeTypeHandle[] pair, plus
2. Dictionary<string, Type>

The linker step would populate (1), which might be faster because RuntimeTypeHandle is a stuct.

At runtime, consult (2) first, and if a type mapping doesn't exist, use (1) to lookup the RuntimeTypeHandle + Type.GetTypeFromHandle(RuntimeTypeHandle), caching the result in (2).

Questions:

• Can we lookup the RuntimeTypeHandle value from Cecil?
• The caching will require locking. What kind of overhead will this imply?

[jonathanpeppers]

There must be a way to use a TypeHandle with Mono.Cecil, as the IL "typemap MVP 1.0" is writing is doing this:

android/src/Microsoft.Android.Sdk.ILLink/TypeMappingStep.cs

Lines 70 to 83 in 70bd636

	/*
	* IL_0000: ldarg.0
	* IL_0001: ldfld class [System.Runtime]System.Collections.Generic.IDictionary`2<string, class [System.Runtime]System.Type> Microsoft.Android.Runtime.NativeAotTypeManager::TypeMappings
	* IL_0006: ldstr "java/lang/Object"
	* IL_000b: ldtoken [Mono.Android]Java.Lang.Object
	* IL_0010: call class [System.Runtime]System.Type [System.Runtime]System.Type::GetTypeFromHandle(valuetype [System.Runtime]System.RuntimeTypeHandle)
	* IL_0015: callvirt instance void class [System.Runtime]System.Collections.Generic.IDictionary`2<string, class [System.Runtime]System.Type>::Add(!0, !1)
	*/
	il.Emit (Mono.Cecil.Cil.OpCodes.Ldarg_0);
	il.Emit (Mono.Cecil.Cil.OpCodes.Ldfld, field);
	il.Emit (Mono.Cecil.Cil.OpCodes.Ldstr, javaKey);
	il.Emit (Mono.Cecil.Cil.OpCodes.Ldtoken, module.ImportReference (typeDefinition));
	il.Emit (Mono.Cecil.Cil.OpCodes.Call, getTypeFromHandle);
	il.Emit (Mono.Cecil.Cil.OpCodes.Callvirt, addMethod);

[filipnavara]

Ldtoken returns RuntimeTypeHandle so you can use that in the dictionary. It's, however, not a pattern that can be represented in C# and ILSpy will not decode it correctly.

[simonrozsival]

I'm currently exploring this direction for the type map: https://gist.github.com/simonrozsival/8b3f2294d7fa1cf84ca32e8145984c98

• there's no startup initialization - the Java class name hashes are stored in a sorted array that's baked into the assembly
• the runtime type handles are stored in code and are accessed via a jumptable (switch) - this allows us to "serialize" RuntimeTypeHandles in a way which Native AOT understands

What do you think?

[jonathanpeppers]

@simonrozsival brought up a point that the RuntimeTypeHandle values might not match up at runtime in NativeAOT:

• We run ILLink
• At some point during/after ILLink, we generate a managed typemap, if we used RuntimeTypeHandle...
• ILC runs and uses its own set of integers (who knows?) instead of RuntimeTypeHandle

[simonrozsival]

ILC runs and uses its own set of integers (who knows?) instead of RuntimeTypeHandle

I don't think I expressed myself correctly. If we generate IL that builds the Dictionary<string, RuntimeTypeHandle> using il.Emit (Mono.Cecil.Cil.OpCodes.Ldtoken, module.ImportReference (typeDefinition));, then Native AOT will compile this correctly. That's not the problem.

I was trying to say that we cannot serialize the raw numeric values of RuntimeTypeHandle while we're processing the IL and use it for lookups via Type.GetTypeFromHandle at runtime.

We can of course create an array of RTH, but it will need to be initialized at startup:

static readonly RuntimeTypeHandle[] s_handles = [ typeof(X).TypeHandle, typeof(Y).TypeHandle ];

// is compiled to IL like this:
.method private hidebysig specialname rtspecialname static 
    void .cctor () cil managed 
{
    // Method begins at RVA 0x2058
    // Code size 56 (0x38)
    .maxstack 8

    IL_0000: ldc.i4.2
    IL_0001: newarr [System.Runtime]System.RuntimeTypeHandle
    IL_0006: dup
    IL_0007: ldc.i4.0
    IL_0008: ldtoken X
    IL_000d: call class [System.Runtime]System.Type [System.Runtime]System.Type::GetTypeFromHandle(valuetype [System.Runtime]System.RuntimeTypeHandle)
    IL_0012: callvirt instance valuetype [System.Runtime]System.RuntimeTypeHandle [System.Runtime]System.Type::get_TypeHandle()
    IL_0017: stelem [System.Runtime]System.RuntimeTypeHandle
    IL_001c: dup
    IL_001d: ldc.i4.1
    IL_001e: ldtoken Y
    IL_0023: call class [System.Runtime]System.Type [System.Runtime]System.Type::GetTypeFromHandle(valuetype [System.Runtime]System.RuntimeTypeHandle)
    IL_0028: callvirt instance valuetype [System.Runtime]System.RuntimeTypeHandle [System.Runtime]System.Type::get_TypeHandle()
    IL_002d: stelem [System.Runtime]System.RuntimeTypeHandle
    IL_0032: stsfld valuetype [System.Runtime]System.RuntimeTypeHandle[] X::s_handles
    IL_0037: ret
} // end of method X::.cctor

[filipnavara]

We can of course create an array of RTH, but it will need to be initialized at startup:

ILC can interpret some static constructors at compile time and produce preinitialized data in the executable image. It would not work for the Dictionary but it can probably work for a simple array... (or can be tweaked to work if it doesn't at the moment)