C# non-boxing conversion of generic enum to int?

2019-01-16 03:32发布

站内文章 / C#

67 0

狗以群分

女 | 书童

私信

可以将文章内容翻译成中文,广告屏蔽插件可能会导致该功能失效(如失效，请关闭广告屏蔽插件后再试):

问题:

Given a generic parameter TEnum which always will be an enum type, is there any way to cast from TEnum to int without boxing/unboxing?

See this example code. This will box/unbox the value unnecessarily.

private int Foo<TEnum>(TEnum value)
    where TEnum : struct  // C# does not allow enum constraint
{
    return (int) (ValueType) value;
}

The above C# is release-mode compiled to the following IL (note boxing and unboxing opcodes):

.method public hidebysig instance int32  Foo<valuetype 
    .ctor ([mscorlib]System.ValueType) TEnum>(!!TEnum 'value') cil managed
{
  .maxstack  8
  IL_0000:  ldarg.1
  IL_0001:  box        !!TEnum
  IL_0006:  unbox.any  [mscorlib]System.Int32
  IL_000b:  ret
}

Enum conversion has been treated extensively on SO, but I could not find a discussion addressing this specific case.

回答1:

I'm not sure that this is possible in C# without using Reflection.Emit. If you use Reflection.Emit, you could load the value of the enum onto the stack and then treat it as though it's an int.

You have to write quite a lot of code though, so you'd want to check whether you'll really gain any performance in doing this.

I believe the equivalent IL would be:

.method public hidebysig instance int32  Foo<valuetype 
    .ctor ([mscorlib]System.ValueType) TEnum>(!!TEnum 'value') cil managed
{
  .maxstack  8
  IL_0000:  ldarg.1
  IL_000b:  ret
}

Note that this would fail if your enum derived from long (a 64 bit integer.)

EDIT

Another thought on this approach. Reflection.Emit can create the method above, but the only way you'd have of binding to it would be via a virtual call (i.e. it implements a compile-time known interface/abstract that you could call) or an indirect call (i.e. via a delegate invocation). I imagine that both of these scenarios would be slower than the overhead of boxing/unboxing anyway.

Also, don't forget that the JIT is not dumb and may take care of this for you. (EDIT see Eric Lippert's comment on the original question -- he says the jitter does not currently perform this optimisation.)

As with all performance related issues: measure, measure, measure!

回答2:

This is similar to answers posted here, but uses expression trees to emit il to cast between types. Expression.Convert does the trick. The compiled delegate (caster) is cached by an inner static class. Since source object can be inferred from the argument, I guess it offers cleaner call. For e.g. a generic context:

static int Generic<T>(T t)
{
    int variable = -1;

    // may be a type check - if(...
    variable = CastTo<int>.From(t);

    return variable;
}

The class:

/// <summary>
/// Class to cast to type <see cref="T"/>
/// </summary>
/// <typeparam name="T">Target type</typeparam>
public static class CastTo<T>
{
    /// <summary>
    /// Casts <see cref="S"/> to <see cref="T"/>.
    /// This does not cause boxing for value types.
    /// Useful in generic methods.
    /// </summary>
    /// <typeparam name="S">Source type to cast from. Usually a generic type.</typeparam>
    public static T From<S>(S s)
    {
        return Cache<S>.caster(s);
    }    

    private static class Cache<S>
    {
        public static readonly Func<S, T> caster = Get();

        private static Func<S, T> Get()
        {
            var p = Expression.Parameter(typeof(S));
            var c = Expression.ConvertChecked(p, typeof(T));
            return Expression.Lambda<Func<S, T>>(c, p).Compile();
        }
    }
}

You can replace the caster func with other implementations. I will compare performance of a few:

direct object casting, ie, (T)(object)S

caster1 = (Func<T, T>)(x => x) as Func<S, T>;

caster2 = Delegate.CreateDelegate(typeof(Func<S, T>), ((Func<T, T>)(x => x)).Method) as Func<S, T>;

caster3 = my implementation above

caster4 = EmitConverter();
static Func<S, T> EmitConverter()
{
    var method = new DynamicMethod(string.Empty, typeof(T), new[] { typeof(S) });
    var il = method.GetILGenerator();

    il.Emit(OpCodes.Ldarg_0);
    if (typeof(S) != typeof(T))
    {
        il.Emit(OpCodes.Conv_R8);
    }
    il.Emit(OpCodes.Ret);

    return (Func<S, T>)method.CreateDelegate(typeof(Func<S, T>));
}

Boxed casts:

int to int

object casting -> 42 ms
caster1 -> 102 ms
caster2 -> 102 ms
caster3 -> 90 ms
caster4 -> 101 ms
int to int?

object casting -> 651 ms
caster1 -> fail
caster2 -> fail
caster3 -> 109 ms
caster4 -> fail
int? to int

object casting -> 1957 ms
caster1 -> fail
caster2 -> fail
caster3 -> 124 ms
caster4 -> fail
enum to int

object casting -> 405 ms
caster1 -> fail
caster2 -> 102 ms
caster3 -> 78 ms
caster4 -> fail
int to enum

object casting -> 370 ms
caster1 -> fail
caster2 -> 93 ms
caster3 -> 87 ms
caster4 -> fail
int? to enum

object casting -> 2340 ms
caster1 -> fail
caster2 -> fail
caster3 -> 258 ms
caster4 -> fail
enum? to int

object casting -> 2776 ms
caster1 -> fail
caster2 -> fail
caster3 -> 131 ms
caster4 -> fail

Expression.Convert puts a direct cast from source type to target type, so it can work out explicit and implicit casts (not to mention reference casts). So this gives way for handling casting which is otherwise possible only when non-boxed (ie, in a generic method if you do (TTarget)(object)(TSource) it will explode if it is not identity conversion (as in previous section) or reference conversion (as shown in later section)). So I will include them in tests.

Non-boxed casts:

int to double

object casting -> fail
caster1 -> fail
caster2 -> fail
caster3 -> 109 ms
caster4 -> 118 ms
enum to int?

object casting -> fail
caster1 -> fail
caster2 -> fail
caster3 -> 93 ms
caster4 -> fail
int to enum?

object casting -> fail
caster1 -> fail
caster2 -> fail
caster3 -> 93 ms
caster4 -> fail
enum? to int?

object casting -> fail
caster1 -> fail
caster2 -> fail
caster3 -> 121 ms
caster4 -> fail
int? to enum?

object casting -> fail
caster1 -> fail
caster2 -> fail
caster3 -> 120 ms
caster4 -> fail

For the fun of it, I tested a few reference type conversions:

PrintStringProperty to string (representation changing)

object casting -> fail (quite obvious, since it is not cast back to original type)
caster1 -> fail
caster2 -> fail
caster3 -> 315 ms
caster4 -> fail
string to object (representation preserving reference conversion)

object casting -> 78 ms
caster1 -> fail
caster2 -> fail
caster3 -> 322 ms
caster4 -> fail

Tested like this:

static void TestMethod<T>(T t)
{
    CastTo<int>.From(t); //computes delegate once and stored in a static variable

    int value = 0;
    var watch = Stopwatch.StartNew();
    for (int i = 0; i < 10000000; i++) 
    {
        value = (int)(object)t; 

        // similarly value = CastTo<int>.From(t);

        // etc
    }
    watch.Stop();
    Console.WriteLine(watch.Elapsed.TotalMilliseconds);
}

Note:

My estimate is that unless you run this at least a hundred thousand times, it's not worth it, and you have almost nothing to worry about boxing. Mind you caching delegates has a hit on memory. But beyond that limit, the speed improvement is significant, especially when it comes to casting involving nullables.
But the real advantage of the CastTo<T> class is when it allows casts that are possible non-boxed, like (int)double in a generic context. As such (int)(object)double fails in these scenarios.
I have used Expression.ConvertChecked instead of Expression.Convert so that arithmetic overflows and underflows are checked (ie results in exception). Since il is generated during run time, and checked settings are a compile time thing, there is no way you can know the checked context of calling code. This is something you have to decide yourself. Choose one, or provide overload for both (better).
If a cast doesn't exist from TSource to TTarget, exception is thrown while the delegate is compiled. If you want a different behaviour, like get a default value of TTarget, you can check type compatibility using reflection before compiling delegate. You have the full control of the code being generated. Its going to be extremely tricky though, you have to check for reference compatibility (IsSubClassOf, IsAssignableFrom), conversion operator existence (going to be hacky), and even for some built in type convertibility between primitive types. Going to be extremely hacky. Easier is to catch exception and return default value delegate based on ConstantExpression. Just stating a possibility that you can mimic behaviour of as keyword which doesnt throw. Its better to stay away from it and stick to convention.

回答3:

I know I'm way late to the party, but if you just need to do a safe cast like this you can use the following using Delegate.CreateDelegate:

public static int Identity(int x){return x;}
// later on..
Func<int,int> identity = Identity;
Delegate.CreateDelegate(typeof(Func<int,TEnum>),identity.Method) as Func<int,TEnum>

now without writing Reflection.Emit or expression trees you have a method that will convert int to enum without boxing or unboxing. Note that TEnum here must have an underlying type of int or this will throw an exception saying it cannot be bound.

Edit: Another method that works too and might be a little less to write...

Func<TEnum,int> converter = EqualityComparer<TEnum>.Default.GetHashCode;

This works to convert your 32bit or less enum from a TEnum to an int. Not the other way around. In .Net 3.5+, the EnumEqualityComparer is optimized to basically turn this into a return (int)value;

You are paying the overhead of using a delegate, but it certainly will be better than boxing.

回答4:

...I'm even 'later' : )

but just to extend on the previous post (Michael B), which did all the interesting work

and got me interested into making a wrapper for a generic case (if you want to cast generic to enum actually)

...and optimized a bit... (note: the main point is to use 'as' on Func<>/delegates instead - as Enum, value types do not allow it)

public static class Identity<TEnum, T>
{
    public static readonly Func<T, TEnum> Cast = (Func<TEnum, TEnum>)((x) => x) as Func<T, TEnum>;
}

...and you can use it like this...

enum FamilyRelation { None, Father, Mother, Brother, Sister, };
class FamilyMember
{
    public FamilyRelation Relation { get; set; }
    public FamilyMember(FamilyRelation relation)
    {
        this.Relation = relation;
    }
}
class Program
{
    static void Main(string[] args)
    {
        FamilyMember member = Create<FamilyMember, FamilyRelation>(FamilyRelation.Sister);
    }
    static T Create<T, P>(P value)
    {
        if (typeof(T).Equals(typeof(FamilyMember)) && typeof(P).Equals(typeof(FamilyRelation)))
        {
            FamilyRelation rel = Identity<FamilyRelation, P>.Cast(value);
            return (T)(object)new FamilyMember(rel);
        }
        throw new NotImplementedException();
    }
}

...for (int) - just (int)rel

回答5:

I guess you can always use System.Reflection.Emit to create a dynamic method and emit the instructions that do this without boxing, although it might be unverifiable.

回答6:

Here is a simplest and fastest way.
(with a little restriction. :-) )

public class BitConvert
{
    [StructLayout(LayoutKind.Explicit)]
    struct EnumUnion32<T> where T : struct {
        [FieldOffset(0)]
        public T Enum;

        [FieldOffset(0)]
        public int Int;
    }

    public static int Enum32ToInt<T>(T e) where T : struct {
        var u = default(EnumUnion32<T>);
        u.Enum = e;
        return u.Int;
    }

    public static T IntToEnum32<T>(int value) where T : struct {
        var u = default(EnumUnion32<T>);
        u.Int = value;
        return u.Enum;
    }
}

Restriction:
This works in Mono. (ex. Unity3D)

More information about Unity3D:
ErikE's CastTo class is a really neat way to solve this problem.
BUT it can't be used as is in Unity3D

First, it have to be fixed like below.
(because that the mono compiler can't compile the original code)

public class CastTo {
    protected static class Cache<TTo, TFrom> {
        public static readonly Func<TFrom, TTo> Caster = Get();

        static Func<TFrom, TTo> Get() {
            var p = Expression.Parameter(typeof(TFrom), "from");
            var c = Expression.ConvertChecked(p, typeof(TTo));
            return Expression.Lambda<Func<TFrom, TTo>>(c, p).Compile();
        }
    }
}

public class ValueCastTo<TTo> : ValueCastTo {
    public static TTo From<TFrom>(TFrom from) {
        return Cache<TTo, TFrom>.Caster(from);
    }
}

Second, ErikE's code can't be used in AOT platform.
So, my code is the best solution for Mono.

To commenter 'Kristof':
I am sorry that I didn't write all the details.

回答7:

Here's a very straight forward solution with C# 7.3's unmanaged generic type constraint:

using System;
public static class EnumExtensions<TEnum> where TEnum : unmanaged, Enum
{
    /// <summary>
    /// Will fail if <see cref="TResult"/>'s type is smaller than <see cref="TEnum"/>'s underlying type
    /// </summary>
    public static TResult To<TResult>( TEnum value ) where TResult : unmanaged
    {
        unsafe
        {
            TResult outVal = default;
            Buffer.MemoryCopy( &value, &outVal, sizeof(TResult), sizeof(TEnum) );
            return outVal;
        }
    }

    public static TEnum From<TSource>( TSource value ) where TSource : unmanaged
    {
        unsafe
        {
            TEnum outVal = default;
            long size = sizeof(TEnum) < sizeof(TSource) ? sizeof(TEnum) : sizeof(TSource);
            Buffer.MemoryCopy( &value, &outVal, sizeof(TEnum), size );
            return outVal;
        }
    }
}

Requires unsafe toggle in your project configuration.

Usage:

int intValue = EnumExtensions<YourEnumType>.To<int>( yourEnumValue );

回答8:

I hope that I'm not too late...

I think that you should consider to solve your problem with a different approach instead of using Enums try to creating a class with a public static readonly properties.

if you will use that approach you will have an object that "feels" like an Enum but you will have all the flexibility of a class which means that you can override any of the operators.

there are other advantages like making that class a partial which will enable you to define the same enum in more then one file/dll which makes it possible to add values to a common dll without recompiling it.

I couldn't find any good reason not to take that approach (this class will be located in the heap and not on the stack which is slower but it's worth it)

please let me know what you think.