Move TypeReference.ToObject to managed (CoreCLR)

[Sergio0694]

Follow up to #69940

Overview

This PR moves TypedReference.ToObject to managed in CoreCLR.
I've only added an FCall to handle a slow path that's only ever taken when T is a function pointer.
This is to avoid having to port a whole lot of additional VM code to C# as well.
All other paths, both for reference types and value types (using the new RuntimeHelpers.Box API) are managed now.

Benchmarks

Method	Branch	Mean	Error	StdDev	Ratio
ToObject_String	PR	1.944 ns	0.0268 ns	0.0250 ns	0.48
ToObject_String	main	4.087 ns	0.0571 ns	0.0534 ns	1.00
ToObject_Int	PR	41.241 ns	0.2446 ns	0.2042 ns	0.93
ToObject_Int	main	44.195 ns	0.4417 ns	0.3916 ns	1.00
ToObject_FnPointer	PR	43.501 ns	0.3535 ns	0.3133 ns	0.92
ToObject_FnPointer	main	47.152 ns	0.5428 ns	0.5077 ns	1.00

Benchmark code (click to expand):

using System.Reflection;
using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Configs;
using BenchmarkDotNet.Running;

BenchmarkSwitcher.FromAssembly(Assembly.GetExecutingAssembly()).Run(args);

[DisassemblyDiagnoser]
[GroupBenchmarksBy(BenchmarkLogicalGroupRule.ByCategory)]
public unsafe class TypedReferenceBenchmark
{
    public string text = "Hello world";
    public int number = 42;
    public delegate*<int, float, string> pointer = (delegate*<int, float, string>)(void*)(nuint)0xDEADBEEF;

    [Benchmark]
    [BenchmarkCategory("REF_TYPE")]
    public object? ToObject_String()
    {
        return TypedReference.ToObject(__makeref(text));
    }

    [Benchmark]
    [BenchmarkCategory("VALUE_TYPE")]
    public object? ToObject_Int()
    {
        return TypedReference.ToObject(__makeref(number));
    }

    [Benchmark]
    [BenchmarkCategory("VALUE_TYPE")]
    public object? ToObject_FnPointer()
    {
        return TypedReference.ToObject(__makeref(pointer));
    }
}

[ghost]

Tagging subscribers to this area: @dotnet/area-system-reflection
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Follow up to #69940

Overview

This PR moves TypedReference.ToObject to managed in CoreCLR.
I've only added an FCall to handle a slow path that's only ever taken when T is a function pointer.
This is to avoid having to port a whole lot of additional VM code to C# as well.
All other paths, both for reference types and value types (using the new RuntimeHelpers.Box API) are managed now.

API additions (CoreCLR-only)

namespace System
{
    public unsafe struct RuntimeTypeHandle
    {
        [MethodImpl(MethodImplOptions.InternalCall)]
        internal static extern unsafe MethodTable* GetElementTypeMethodTable(CorElementType elementType);
    }
}

namespace System.Runtime.CompilerServices
{
    internal unsafe struct TypeHandle
    {
        public bool IsNull { get; }

        public TypeDesc* AsTypeDesc();
        public MethodTable* GetMethodTable();
    }

    internal unsafe struct TypeDesc
    {
        private readonly uint TypeAndFlags;

        public MethodTable* GetMethodTable();
        public CorElementType GetInternalCorElementType();
    }

    internal unsafe struct ParamTypeDesc
    {
        public readonly uint TypeAndFlags;
        public readonly MethodTable* m_TemplateMT;
        public readonly TypeHandle m_Arg;
        public readonly void* m_hExposedClassObject;

        public MethodTable* GetTemplateMethodTableInternal();
    }
}

Benchmarks

Method	Branch	Mean	Error	StdDev	Ratio
ToObject_String	main	3.8621 ns	0.1293 ns	0.1770 ns	1.00
ToObject_String	PR	0.9941 ns	0.0115 ns	0.0107 ns	0.25
ToObject_Int	main	44.6062 ns	0.2902 ns	0.4432 ns	1.00
ToObject_Int	PR	41.7305 ns	0.4657 ns	0.3889 ns	0.94

Benchmark code (click to expand):

using System.Reflection;
using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Configs;
using BenchmarkDotNet.Running;

BenchmarkSwitcher.FromAssembly(Assembly.GetExecutingAssembly()).Run(args);

[DisassemblyDiagnoser]
[GroupBenchmarksBy(BenchmarkLogicalGroupRule.ByCategory)]
public class ArrayGetValueBenchmark
{
    private string text = "Hello world";
    private int number = 42;

    [Benchmark]
    [BenchmarkCategory("REF_TYPE")]
    public object? ToObject_String()
    {
        return TypedReference.ToObject(__makeref(text));
    }

    [Benchmark]
    [BenchmarkCategory("VALUE_TYPE")]
    public object? ToObject_Int()
    {
        return TypedReference.ToObject(__makeref(number));
    }
}

Author:	Sergio0694
Assignees:	-
Labels:	area-System.Reflection
Milestone:	-

[jkotas]

The new tests are failing on Mono. Open an issue on it and disable the tests against it.

[Sergio0694]

Looks like the tests are not failing, but just can't even compile on Mono 🤔

"/tmp/helix/working/A9980937/p/build/apple/AppleApp.targets(104,5): error : Precompiling failed for /tmp/helix/working/A9980937/w/B32B097C/e/publish/System.Runtime.Tests.dll. [/private/tmp/helix/working/A9980937/w/B32B097C/e/publish/ProxyProjectForAOTOnHelix.proj]
/tmp/helix/working/A9980937/p/build/apple/AppleApp.targets(104,5): error : Mono Ahead of Time compiler - compiling assembly /private/tmp/helix/working/A9980937/w/B32B097C/e/publish/System.Runtime.Tests.dll [/private/tmp/helix/working/A9980937/w/B32B097C/e/publish/ProxyProjectForAOTOnHelix.proj]
/tmp/helix/working/A9980937/p/build/apple/AppleApp.targets(104,5): error : AOTID 2E205542-C4E2-0761-A6DA-F8F7672DD714 [/private/tmp/helix/working/A9980937/w/B32B097C/e/publish/ProxyProjectForAOTOnHelix.proj]
/tmp/helix/working/A9980937/p/build/apple/AppleApp.targets(104,5): error : * Assertion at /Users/runner/work/1/s/src/mono/mono/mini/aot-compiler.c:3406, condition `m_class_get_rank (klass) > 0' not met [/private/tmp/helix/working/A9980937/w/B32B097C/e/publish/ProxyProjectForAOTOnHelix.proj]"

So [SkipOnMono] didn't do the trick. I'm not seeing eg. CORECLR defined in the tests, is there a recommended way or a specific compile constant to just remove some tests entirely on a specific runtime so that they won't even be compiled?

[jkotas]

[ActiveIssue("https://github.com/dotnet/runtime/issues/...", TestRuntimes.Mono)]

[jkotas]

If it is causing the Mono AOT compiler the crash, we do not have a good way to deal with it today. I guess you can comment out the test for now.

[Sergio0694]

Ah, gotcha. Will do that, thank you! 🙂

[jkotas]

Failures are known issues #69125 and #69192

[jkotas]

Thank you!