Performance of Expression.Compile vs Lambda, direc

Question

I'm curious how performant the Expression.Compile is versus lambda expression in the code and versus direct method usage, and also direct method calls vs virtual method calls (pseudo code):

var foo = new Foo();
var iFoo = (IFoo)foo;

foo.Bar();
iFoo.Bar();
(() => foo.Bar())();
(() => iFoo.Bar())();
Expression.Compile(foo, Foo.Bar)();
Expression.Compile(iFoo, IFoo.Bar)();
Expression.CompileToMethod(foo, Foo.Bar);
Expression.CompileToMethod(iFoo, IFoo.Bar);
MethodInfo.Invoke(foo, Foo.Bar);
MethodInfo.Invoke(iFoo, IFoo.Bar);

Answer 1

I didn't find any answer, so here is the performance test:

using System;
using System.Diagnostics;
using System.Linq.Expressions;
using System.Reflection;
using System.Reflection.Emit;

namespace ExpressionTest
{
    public interface IFoo
    {
        int Bar();
    }

    public sealed class FooImpl : IFoo
    {
        public int Bar()
        {
            return 0;
        }
    }

    class Program
    {
        static void Main(string[] args)
        {
            var foo = new FooImpl();
            var iFoo = (IFoo)foo;

            Func<int> directLambda = () => foo.Bar();
            Func<int> virtualLambda = () => iFoo.Bar();
            var compiledDirectCall = CompileBar(foo, asInterfaceCall: false);
            var compiledVirtualCall = CompileBar(foo, asInterfaceCall: true);
            var compiledArgDirectCall = CompileBar<FooImpl>();
            var compiledArgVirtualCall = CompileBar<IFoo>();
            var barMethodInfo = typeof(FooImpl).GetMethod(nameof(FooImpl.Bar));
            var iBarMethodInfo = typeof(IFoo).GetMethod(nameof(IFoo.Bar));
            var compiledToModuleDirect = CompileToModule<FooImpl>();
            var compiledToModuleVirtual = CompileToModule<IFoo>();

            var iterationCount = 200000000;
            Console.WriteLine($"Iteration count: {iterationCount:N0}");

            var sw = Stopwatch.StartNew();
            for (int i = 0; i < iterationCount; i++)
                compiledVirtualCall();
            var elapsedMs = sw.ElapsedMilliseconds;
            Console.WriteLine($"Virtual (Func<int>)Expression.Compile(): {elapsedMs} ms");

            sw.Restart();
            for (int i = 0; i < iterationCount; i++)
                compiledDirectCall();
            elapsedMs = sw.ElapsedMilliseconds;
            Console.WriteLine($"Direct (Func<int>)Expression.Compile(): {elapsedMs} ms");

            sw.Restart();
            for (int i = 0; i < iterationCount; i++)
                compiledArgVirtualCall(iFoo);
            elapsedMs = sw.ElapsedMilliseconds;
            Console.WriteLine($"Virtual (Func<IFoo, int>)Expression.Compile(): {elapsedMs} ms");

            sw.Restart();
            for (int i = 0; i < iterationCount; i++)
                compiledArgDirectCall(foo);
            elapsedMs = sw.ElapsedMilliseconds;
            Console.WriteLine($"Direct (Func<FooImpl, int>)Expression.Compile(): {elapsedMs} ms");

            sw.Restart();
            for (int i = 0; i < iterationCount; i++)
                compiledToModuleVirtual(iFoo);
            elapsedMs = sw.ElapsedMilliseconds;
            Console.WriteLine($"Virtual (Func<IFoo, int>)Expression.CompileToMethod(): {elapsedMs} ms");

            sw.Restart();
            for (int i = 0; i < iterationCount; i++)
                compiledToModuleDirect(foo);
            elapsedMs = sw.ElapsedMilliseconds;
            Console.WriteLine($"Direct (Func<FooImpl, int>)Expression.CompileToMethod(): {elapsedMs} ms");

            sw.Restart();
            for (int i = 0; i < iterationCount; i++)
                virtualLambda();
            elapsedMs = sw.ElapsedMilliseconds;
            Console.WriteLine($"Virtual () => IFoo.Bar(): {elapsedMs} ms");

            sw.Restart();
            for (int i = 0; i < iterationCount; i++)
                directLambda();
            elapsedMs = sw.ElapsedMilliseconds;
            Console.WriteLine($"Direct () => FooImpl.Bar(): {elapsedMs} ms");

            sw.Restart();
            for (int i = 0; i < iterationCount; i++)
                iFoo.Bar();
            elapsedMs = sw.ElapsedMilliseconds;
            Console.WriteLine($"Virtual IFoo.Bar(): {elapsedMs} ms");

            sw.Restart();
            for (int i = 0; i < iterationCount; i++)
                foo.Bar();
            elapsedMs = sw.ElapsedMilliseconds;
            Console.WriteLine($"Direct Foo.Bar(): {elapsedMs} ms");

            sw.Restart();
            for (int i = 0; i < iterationCount; i++) {
                int result = (int)iBarMethodInfo.Invoke(iFoo, null);
            }
            elapsedMs = sw.ElapsedMilliseconds;
            Console.WriteLine($"Virtual MethodInfo.Invoke(FooImpl, Bar): {elapsedMs} ms");

            sw.Restart();
            for (int i = 0; i < iterationCount; i++) {
                int result = (int)barMethodInfo.Invoke(foo, null);
            }
            elapsedMs = sw.ElapsedMilliseconds;
            Console.WriteLine($"Direct MethodInfo.Invoke(IFoo, Bar): {elapsedMs} ms");
        }

        static Func<int> CompileBar(IFoo foo, bool asInterfaceCall)
        {
            var fooType = asInterfaceCall ? typeof(IFoo) : foo.GetType();
            var methodInfo = fooType.GetMethod(nameof(IFoo.Bar));
            var instance = Expression.Constant(foo, fooType);
            var call = Expression.Call(instance, methodInfo);
            var lambda = Expression.Lambda(call);
            var compiledFunction = (Func<int>)lambda.Compile();
            return compiledFunction;
        }

        static Func<TInput, int> CompileBar<TInput>()
        {
            var fooType = typeof(TInput);
            var methodInfo = fooType.GetMethod(nameof(IFoo.Bar));
            var instance = Expression.Parameter(fooType, "foo");
            var call = Expression.Call(instance, methodInfo);
            var lambda = Expression.Lambda(call, instance);
            var compiledFunction = (Func<TInput, int>)lambda.Compile();
            return compiledFunction;
        }

        static Func<TInput, int> CompileToModule<TInput>()
        {
            var fooType = typeof(TInput);
            var methodInfo = fooType.GetMethod(nameof(IFoo.Bar));
            var instance = Expression.Parameter(fooType, "foo");
            var call = Expression.Call(instance, methodInfo);
            var lambda = Expression.Lambda(call, instance);

            var asmName = new AssemblyName(fooType.Name);
            var asmBuilder = AssemblyBuilder.DefineDynamicAssembly(asmName, AssemblyBuilderAccess.Run);
            var moduleBuilder = asmBuilder.DefineDynamicModule(fooType.Name);
            var typeBuilder = moduleBuilder.DefineType(fooType.Name, TypeAttributes.Public);
            var methodBuilder = typeBuilder.DefineMethod(nameof(IFoo.Bar), MethodAttributes.Static, typeof(int), new[] { fooType });
            Expression.Lambda<Action>(lambda).CompileToMethod(methodBuilder);
            var createdType = typeBuilder.CreateType();

            var mi = createdType.GetMethods(BindingFlags.NonPublic | BindingFlags.Static)[1];
            var func = Delegate.CreateDelegate(typeof(Func<TInput, int>), mi);
            return (Func<TInput, int>)func;
        }
    }
}

On my laptop (Release mode, 64 bit, .NET 4.5.2) it yields:

Iteration count: 200,000,000
Virtual MethodInfo.Invoke(FooImpl, Bar):               61811 ms
Direct MethodInfo.Invoke(IFoo, Bar):                   37078 ms
Virtual (Func<int>)Expression.Compile():                2894 ms
Direct (Func<int>)Expression.Compile():                 2242 ms
Virtual (Func<IFoo, int>)Expression.Compile():          2319 ms
Direct (Func<FooImpl, int>)Expression.Compile():        2051 ms
Virtual (Func<IFoo, int>)Expression.CompileToMethod():   996 ms
Direct (Func<FooImpl, int>)Expression.CompileToMethod(): 679 ms
Virtual () => IFoo.Bar():                                796 ms
Direct () => FooImpl.Bar():                              469 ms
Virtual IFoo.Bar():                                      531 ms
Direct Foo.Bar():                                         68 ms

Hope this helps.

Answer 2

Hint: in release mode no call made at all in "Direct Call" case. CPU going from 00B531BC (mov eax ...) to 00B531C8 (jl 00B531BC) only.

                for (int i = 0; i < iterationCount; i++)
00B531BA  xor         edx,edx  
                foo.Bar();
00B531BC  mov         eax,dword ptr [ebx+4]  // actual loop begin
00B531BF  cmp         byte ptr [eax],al  
            for (int i = 0; i < iterationCount; i++)
00B531C1  inc         edx  
00B531C2  cmp         edx,0BEBC200h // 0BEBC200h = 200000000
00B531C8  jl          00B531BC      // loop begin address

Answer 3

We can split a question to 2 cases:

• how bare .NET work with method calls itself (infrastructure question)?
• how optimizers assist to method calling?

ExpressionTest.exe in Release mode with optimization (default release settings) .NET 4.5.2:

Compiled Virtual Call: 4625 ms
Compiled Direct Call: 3361 ms
Lambda Virtual Call: 1096 ms
Lambda Direct Call: 576 ms
Virtual Call: 649 ms
Direct Call: 144 ms

We see that "Direct Call" in 4.5 times faster than "Virtual Call". But as we see above it's no call at all. Bar method was inlined.

ExpressionTest.exe in Release mode with no optimization .NET 4.5.2:

Compiled Virtual Call: 5394 ms
Compiled Direct Call: 4666 ms
Lambda Virtual Call: 1800 ms
Lambda Direct Call: 1683 ms
Virtual Call: 1154 ms
Direct Call: 1112 ms

So, "Direct Call" is about 3-4% faster than "Virtual Call".

Similar question: Performance of "direct" virtual call vs. interface call in C#