How to instance a C# class in UNmanaged memory? (Possible?)

Question

UPDATE: There is now an accepted answer that "works". You should never, ever, ever, ever use it. Ever.

---

First let me preface my question by stating that I'm a game developer. There's a legitimate - if highly unusual - performance-related reason for wanting to do this.

---

Say I have a C# class like this:

class Foo
{
    public int a, b, c;
    public void MyMethod(int d) { a = d; b = d; c = a + b; }
}

Nothing fancy. Note that it is a reference type that contains only value types.

In managed code I'd like to have something like this:

Foo foo;
foo = Voodoo.NewInUnmanagedMemory<Foo>(); // <- ???
foo.MyMethod(1);

What would the function NewInUnmanagedMemory look like? If it can't be done in C#, could it be done in IL? (Or maybe C++/CLI?)

Basically: Is there a way - no matter how hacky - to turn some totally arbitrary pointer into an object reference. And - short of making the CLR explode - damn the consequences.

(Another way to put my question is: "I want to implement a custom allocator for C#")

This leads to the follow-up question: What does the garbage collector do (implementation-specific, if need be) when faced with a reference that points outside of managed memory?

And, related to that, what would happen if Foo had a reference as a member field? What if it pointed at managed memory? What if it only ever pointed at other objects allocated in unmanaged memory?

Finally, if this is impossible: Why?

---

Update: Here are the "missing pieces" so far:

#1: How to convert an IntPtr to an object reference? It might be possible though unverifiable IL (see comments). So far I've had no luck with this. The framework seems to be extremely careful to prevent this from happening.

(It would also be nice to be able to get the size and layout information for non-blittable managed types at runtime. Again, the framework tries to make this impossible.)

#2: Assuming problem one can be solved - what does the GC do when it encounters an object reference that points outside of the GC heap? Does it crash? Anton Tykhyy, in his answer, guesses that it will. Given how careful the framework is to prevent #1, it does seem likely. Something that confirms this would be nice.

(Alternatively the object reference could point to pinned memory inside the GC heap. Would that make a difference?)

Based on this, I'm inclined to think that this idea for a hack is impossible - or at least not worth the effort. But I'd be interested to get an answer that goes into the technical details of #1 or #2 or both.

Answer 1

I have been experimenting creating classes in unmanaged memory. It is possible but has a problem I am currently unable to solve - you can't assign objects to reference-type fields -see edit at the bottom-, so you can have only structure fields in your custom class. This is evil:

using System;
using System.Reflection.Emit;
using System.Runtime.InteropServices;

public class Voodoo<T> where T : class
{
    static readonly IntPtr tptr;
    static readonly int tsize;
    static readonly byte[] zero;

    public static T NewInUnmanagedMemory()
    {
        IntPtr handle = Marshal.AllocHGlobal(tsize);
        Marshal.Copy(zero, 0, handle, tsize);
        IntPtr ptr = handle+4;
        Marshal.WriteIntPtr(ptr, tptr);
        return GetO(ptr);
    }

    public static void FreeUnmanagedInstance(T obj)
    {
        IntPtr ptr = GetPtr(obj);
        IntPtr handle = ptr-4;
        Marshal.FreeHGlobal(handle);
    }

    delegate T GetO_d(IntPtr ptr);
    static readonly GetO_d GetO;
    delegate IntPtr GetPtr_d(T obj);
    static readonly GetPtr_d GetPtr;
    static Voodoo()
    {
        Type t = typeof(T);
        tptr = t.TypeHandle.Value;
        tsize = Marshal.ReadInt32(tptr, 4);
        zero = new byte[tsize];

        DynamicMethod m = new DynamicMethod("GetO", typeof(T), new[]{typeof(IntPtr)}, typeof(Voodoo<T>), true);
        var il = m.GetILGenerator();
        il.Emit(OpCodes.Ldarg_0);
        il.Emit(OpCodes.Ret);
        GetO = m.CreateDelegate(typeof(GetO_d)) as GetO_d;

        m = new DynamicMethod("GetPtr", typeof(IntPtr), new[]{typeof(T)}, typeof(Voodoo<T>), true);
        il = m.GetILGenerator();
        il.Emit(OpCodes.Ldarg_0);
        il.Emit(OpCodes.Ret);
        GetPtr = m.CreateDelegate(typeof(GetPtr_d)) as GetPtr_d;
    }
}

If you care about memory leak, you should always call FreeUnmanagedInstance when you are done with your class. If you want more complex solution, you can try this:

using System;
using System.Reflection.Emit;
using System.Runtime.InteropServices;


public class ObjectHandle<T> : IDisposable where T : class
{
    bool freed;
    readonly IntPtr handle;
    readonly T value;
    readonly IntPtr tptr;

    public ObjectHandle() : this(typeof(T))
    {

    }

    public ObjectHandle(Type t)
    {
        tptr = t.TypeHandle.Value;
        int size = Marshal.ReadInt32(tptr, 4);//base instance size
        handle = Marshal.AllocHGlobal(size);
        byte[] zero = new byte[size];
        Marshal.Copy(zero, 0, handle, size);//zero memory
        IntPtr ptr = handle+4;
        Marshal.WriteIntPtr(ptr, tptr);//write type ptr
        value = GetO(ptr);//convert to reference
    }

    public T Value{
        get{
            return value;
        }
    }

    public bool Valid{
        get{
            return Marshal.ReadIntPtr(handle, 4) == tptr;
        }
    }

    public void Dispose()
    {
        if(!freed)
        {
            Marshal.FreeHGlobal(handle);
            freed = true;
            GC.SuppressFinalize(this);
        }
    }

    ~ObjectHandle()
    {
        Dispose();
    }

    delegate T GetO_d(IntPtr ptr);
    static readonly GetO_d GetO;
    static ObjectHandle()
    {
        DynamicMethod m = new DynamicMethod("GetO", typeof(T), new[]{typeof(IntPtr)}, typeof(ObjectHandle<T>), true);
        var il = m.GetILGenerator();
        il.Emit(OpCodes.Ldarg_0);
        il.Emit(OpCodes.Ret);
        GetO = m.CreateDelegate(typeof(GetO_d)) as GetO_d;
    }
}

/*Usage*/
using(var handle = new ObjectHandle<MyClass>())
{
    //do some work
}

I hope it will help you on your path.

Edit: Found a solution to reference-type fields:

class MyClass
{
    private IntPtr a_ptr;
    public object a{
        get{
            return Voodoo<object>.GetO(a_ptr);
        }
        set{
            a_ptr = Voodoo<object>.GetPtr(value);
        }
    }
    public int b;
    public int c;
}

Edit: Even better solution. Just use ObjectContainer<object> instead of object and so on.

public struct ObjectContainer<T> where T : class
{
    private readonly T val;

    public ObjectContainer(T obj)
    {
        val = obj;
    }

    public T Value{
        get{
            return val;
        }
    }

    public static implicit operator T(ObjectContainer<T> @ref)
    {
        return @ref.val;
    }

    public static implicit operator ObjectContainer<T>(T obj)
    {
        return new ObjectContainer<T>(obj);
    }

    public override string ToString()
    {
        return val.ToString();
    }

    public override int GetHashCode()
    {
        return val.GetHashCode();
    }

    public override bool Equals(object obj)
    {
        return val.Equals(obj);
    }
}

Answer 2

"I want to implement a custom allocator for C#"

GC is at the core of the CLR. Only Microsoft (or the Mono team in case of Mono) can replace it, at a great cost in development effort. GC being at the core of the CLR, messing around with the GC or the managed heap will crash the CLR — quickly if you're very-very lucky.

What does the garbage collector do (implementation-specific, if need be) when faced with a reference that points outside of managed memory?

It crashes in an implementation-specific way ;)

Answer 3

Purely C# Approach

So, there are a few options. The easiest is to use new/delete in an unsafe context for structs. The second is to use built-in Marshalling services to deal with unmanaged memory (code for this is visible below). However, both of these deal with structs (though I think the latter method is very close to what you want). My code has a limitation in that you must stick to structures throughout and use IntPtrs for references (using ChunkAllocator.ConvertPointerToStructure to get the data and ChunkAllocator.StoreStructure to store the changed data). This is obviously cumbersome, so you'd better really want the performance if you use my approach. However, if you are dealing with only value-types, this approach is sufficient.

Detour: Classes in the CLR

Classes have a 8 byte "prefix" in their allocated memory. Four bytes are for the sync index for multithreading, and four bytes are for identifying their type (basically, virtual method table and run-time reflection). This makes it hard to deal with unamanaged memory since these are CLR specific and since the sync index can change during run-time. See here for details on run-time object creation and here for an overview of memory layout for a reference type. Also check out CLR via C# for a more in-depth explanation.

A Caveat

As usual, things are rarely so simple as yes/no. The real complexity of reference types has to do with how the garbage collector compacts allocated memory during a garbage collection. If you can somehow ensure that a garbage collection doesn't happen or that it won't affect the data in question (see the fixed keyword) then you can turn an arbitrary pointer into an object reference (just offset the pointer by 8 bytes, then interpret that data as a struct with the same fields and memory layout; perhaps use StructLayoutAttribute to be sure). I would experiment with non-virtual methods to see if they work; they should (especially if you put them on the struct) but virtual methods are no-go due to the virtual method table that you'd have to discard.

One Does Not Simply Walk Into Mordor

Simply put, this means that managed reference types (classes) cannot be allocated in unmanaged memory. You could use managed reference types in C++, but those would be subject to garbage collection... and the process and code is more painful than the struct-based approach. Where does that leave us? Back where we started, of course.

There is a Secret Way

We could brave Shelob's Lair memory allocation ourselves. Unfortunately, this is where our paths must part, because I am not that knowledgeable about it. I will provide you with a link or two - perhaps three or four in actuality. This is rather complicated and begs the question: Are there other optimizations you could try? Cache coherency and superior algorithms is one approach, as is judicious application of P/Invoke for performance-critical code. You could also apply the aforementioned structures-only memory allocation for key methods/classes.

Good luck, and let us know if you find a superior alternative.

Appendix: Source Code

ChunkAllocator.cs

using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Runtime.InteropServices;

namespace MemAllocLib
{
    public sealed class ChunkAllocator : IDisposable
    {
        IntPtr m_chunkStart;
        int m_offset;//offset from already allocated memory
        readonly int m_size;

        public ChunkAllocator(int memorySize = 1024)
        {
            if (memorySize < 1)
                throw new ArgumentOutOfRangeException("memorySize must be positive");

            m_size = memorySize;
            m_chunkStart = Marshal.AllocHGlobal(memorySize);
        }
        ~ChunkAllocator()
        {
            Dispose();
        }

        public IntPtr Allocate<T>() where T : struct
        {
            int reqBytes = Marshal.SizeOf(typeof(T));//not highly performant
            return Allocate<T>(reqBytes);
        }

        public IntPtr Allocate<T>(int reqBytes) where T : struct
        {
            if (m_chunkStart == IntPtr.Zero)
                throw new ObjectDisposedException("ChunkAllocator");
            if (m_offset + reqBytes > m_size)
                throw new OutOfMemoryException("Too many bytes allocated: " + reqBytes + " needed, but only " + (m_size - m_offset) + " bytes available");

            T created = default(T);
            Marshal.StructureToPtr(created, m_chunkStart + m_offset, false);
            m_offset += reqBytes;

            return m_chunkStart + (m_offset - reqBytes);
        }

        public void Dispose()
        {
            if (m_chunkStart != IntPtr.Zero)
            {
                Marshal.FreeHGlobal(m_chunkStart);
                m_offset = 0;
                m_chunkStart = IntPtr.Zero;
            }
        }

        public void ReleaseAllMemory()
        {
            m_offset = 0;
        }

        public int AllocatedMemory
        {
            get { return m_offset; }
        }

        public int AvailableMemory
        {
            get { return m_size - m_offset; }
        }

        public int TotalMemory
        {
            get { return m_size; }
        }

        public static T ConvertPointerToStruct<T>(IntPtr ptr) where T : struct
        {
            return (T)Marshal.PtrToStructure(ptr, typeof(T));
        }

        public static void StoreStructure<T>(IntPtr ptr, T data) where T : struct
        {
            Marshal.StructureToPtr(data, ptr, false);
        }
    }
}

Program.cs

using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;

namespace MemoryAllocation
{
    class Program
    {
        static void Main(string[] args)
        {
            using (MemAllocLib.ChunkAllocator chunk = new MemAllocLib.ChunkAllocator())
            {
                Console.WriteLine(">> Simple data test");
                SimpleDataTest(chunk);

                Console.WriteLine();

                Console.WriteLine(">> Complex data test");
                ComplexDataTest(chunk);
            }

            Console.ReadLine();
        }

        private static void SimpleDataTest(MemAllocLib.ChunkAllocator chunk)
        {
            IntPtr ptr = chunk.Allocate<System.Int32>();

            Console.WriteLine(MemAllocLib.ChunkAllocator.ConvertPointerToStruct<Int32>(ptr));
            System.Diagnostics.Debug.Assert(MemAllocLib.ChunkAllocator.ConvertPointerToStruct<Int32>(ptr) == 0, "Data not initialized properly");
            System.Diagnostics.Debug.Assert(chunk.AllocatedMemory == sizeof(Int32), "Data not allocated properly");

            int data = MemAllocLib.ChunkAllocator.ConvertPointerToStruct<Int32>(ptr);
            data = 10;
            MemAllocLib.ChunkAllocator.StoreStructure(ptr, data);

            Console.WriteLine(MemAllocLib.ChunkAllocator.ConvertPointerToStruct<Int32>(ptr));
            System.Diagnostics.Debug.Assert(MemAllocLib.ChunkAllocator.ConvertPointerToStruct<Int32>(ptr) == 10, "Data not set properly");

            Console.WriteLine("All tests passed");
        }

        private static void ComplexDataTest(MemAllocLib.ChunkAllocator chunk)
        {
            IntPtr ptr = chunk.Allocate<Person>();

            Console.WriteLine(MemAllocLib.ChunkAllocator.ConvertPointerToStruct<Person>(ptr));
            System.Diagnostics.Debug.Assert(MemAllocLib.ChunkAllocator.ConvertPointerToStruct<Person>(ptr).Age == 0, "Data age not initialized properly");
            System.Diagnostics.Debug.Assert(MemAllocLib.ChunkAllocator.ConvertPointerToStruct<Person>(ptr).Name == null, "Data name not initialized properly");
            System.Diagnostics.Debug.Assert(chunk.AllocatedMemory == System.Runtime.InteropServices.Marshal.SizeOf(typeof(Person)) + sizeof(Int32), "Data not allocated properly");

            Person data = MemAllocLib.ChunkAllocator.ConvertPointerToStruct<Person>(ptr);
            data.Name = "Bob";
            data.Age = 20;
            MemAllocLib.ChunkAllocator.StoreStructure(ptr, data);

            Console.WriteLine(MemAllocLib.ChunkAllocator.ConvertPointerToStruct<Person>(ptr));
            System.Diagnostics.Debug.Assert(MemAllocLib.ChunkAllocator.ConvertPointerToStruct<Person>(ptr).Age == 20, "Data age not set properly");
            System.Diagnostics.Debug.Assert(MemAllocLib.ChunkAllocator.ConvertPointerToStruct<Person>(ptr).Name == "Bob", "Data name not set properly");

            Console.WriteLine("All tests passed");
        }

        struct Person
        {
            public string Name;
            public int Age;

            public Person(string name, int age)
            {
                Name = name;
                Age = age;
            }

            public override string ToString()
            {
                if (string.IsNullOrWhiteSpace(Name))
                    return "Age is " + Age;
                return Name + " is " + Age + " years old";
            }
        }
    }
}

Answer 4

You can write code in C++ and call it from .NET using P/Invoke or you can you can write code in managed C++ that gives you full access to the native API from inside a .NET language. However, on the managed side you can only work with managed types so you will have to encapsulate your unmanaged objects.

To give a simple example: Marshal.AllocHGlobal allows you to allocate memory on the Windows heap. The handle returned is not of much use in .NET but can be required when calling a native Windows API requiring a buffer.

Answer 5

This is not possible.

However you can use a managed struct and create a pointer of this struct type. This pointer can point anywhere (including to unmanaged memory).

The question is, why would you want to have a class in unmanaged memory? You wouldn't get GC features anyway. You can just use a pointer-to-struct.

Answer 6

Nothing like that is possible. You can access managed memory in unsafe context, but said memory is still managed and subject to GC.

Why?

Simplicity and security.

But now that I think about it, I think you can mix managed and unmanaged with C++/CLI. But I'm not sure about that, because I never worked with C++/CLI.

Answer 7

I don't know a way to hold a C# class instance in the unmanaged heap, not even in C++/CLI.

Answer 8

It's possible to design a value-type allocator entirely within .net, without using any unmanaged code, which can allocate and free an arbitrary number of value-type instances without any significant GC pressure. The trick is to create a relatively small number of arrays (possibly one for each type) to hold the instances, and then pass around "instance reference" structs which hold the array indices of the index in question.

Suppose, for example, that I want to have a "creature" class which holds XYZ positions (float), XYZ velocity (also float), roll/pitch/yaw (ditto), damage (float), and kind (enumeration). An interface "ICreatureReference" would define getters and setters for all those properties. A typical implementation would be a struct CreatureReference with a single private field int _index, and property accessors like:

  float Position {
    get {return Creatures[_index].Position;} 
    set {Creatures[_index].Position = value;}
  };

The system would keep a list of which array slots are used and vacant (it could, if desired, use one of the fields within Creatures to form a linked list of vacant slots). The CreatureReference.Create method would allocate an item from the vacant-items list; the Dispose method of a CreatureReference instance would add its array slot to the vacant-items list.

This approach ends up requiring an annoying amount of boilerplate code, but it can be reasonably efficient and avoid GC pressure. The biggest problems with are probably that (1) it makes structs behave more like reference types than structs, and (2) it requires absolute discipline with calling IDispose, since non-disposed array slots will never get reclaimed. Another irksome quirk is that one will be unable to use property setters for read-only values of type CreatureReference, even though the property setters would not try to mutate any fields of the CreatureReference instance to which they are applied. Using an interface ICreatureReference may avoid this difficulty, but one must be careful to only declare storage locations of generic types constrained to ICreatureReference, rather than declaring storage locations of ICreatureReference.