Compiled C# Lambda Expressions Performance

Question

Consider the following simple manipulation over a collection:

static List<int> x = new List<int>() { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 };
var result = x.Where(i => i % 2 == 0).Where(i => i > 5);

Now let's use Expressions. The following code is roughly equivalent:

static void UsingLambda() {
    Func<IEnumerable<int>, IEnumerable<int>> lambda = l => l.Where(i => i % 2 == 0).Where(i => i > 5);
    var t0 = DateTime.Now.Ticks;
    for (int j = 1; j < MAX; j++) 
        var sss = lambda(x).ToList();

    var tn = DateTime.Now.Ticks;
    Console.WriteLine("Using lambda: {0}", tn - t0);
}

But I want to build the expression on-the-fly, so here's a new test:

static void UsingCompiledExpression() {
    var f1 = (Expression<Func<IEnumerable<int>, IEnumerable<int>>>)(l => l.Where(i => i % 2 == 0));
    var f2 = (Expression<Func<IEnumerable<int>, IEnumerable<int>>>)(l => l.Where(i => i > 5));
    var argX = Expression.Parameter(typeof(IEnumerable<int>), "x");
    var f3 = Expression.Invoke(f2, Expression.Invoke(f1, argX));
    var f = Expression.Lambda<Func<IEnumerable<int>, IEnumerable<int>>>(f3, argX);

    var c3 = f.Compile();

    var t0 = DateTime.Now.Ticks;
    for (int j = 1; j < MAX; j++) 
        var sss = c3(x).ToList();

    var tn = DateTime.Now.Ticks;
    Console.WriteLine("Using lambda compiled: {0}", tn - t0);
}

Of course it isn't exactly like the above, so to be fair, I modify the first one slightly:

static void UsingLambdaCombined() {
    Func<IEnumerable<int>, IEnumerable<int>> f1 = l => l.Where(i => i % 2 == 0);
    Func<IEnumerable<int>, IEnumerable<int>> f2 = l => l.Where(i => i > 5);
    Func<IEnumerable<int>, IEnumerable<int>> lambdaCombined = l => f2(f1(l));
    var t0 = DateTime.Now.Ticks;
    for (int j = 1; j < MAX; j++) 
        var sss = lambdaCombined(x).ToList();

    var tn = DateTime.Now.Ticks;
    Console.WriteLine("Using lambda combined: {0}", tn - t0);
}

Now comes the results for MAX = 100000, VS2008, debugging ON:

Using lambda compiled: 23437500
Using lambda:           1250000
Using lambda combined:  1406250

And with debugging OFF:

Using lambda compiled: 21718750
Using lambda:            937500
Using lambda combined:  1093750

Surprise. The compiled expression is roughly 17x slower than the other alternatives. Now here comes the questions:

1. Am I comparing non-equivalent expressions?
2. Is there a mechanism to make .NET "optimize" the compiled expression?
3. How do I express the same chain call l.Where(i => i % 2 == 0).Where(i => i > 5); programatically?

---

Some more statistics. Visual Studio 2010, debugging ON, optimizations OFF:

Using lambda:           1093974
Using lambda compiled: 15315636
Using lambda combined:   781410

Debugging ON, optimizations ON:

Using lambda:            781305
Using lambda compiled: 15469839
Using lambda combined:   468783

Debugging OFF, optimizations ON:

Using lambda:            625020
Using lambda compiled: 14687970
Using lambda combined:   468765

New Surprise. Switching from VS2008 (C#3) to VS2010 (C#4), makes the UsingLambdaCombined faster than the native lambda.

---

Ok, I've found a way to improve the lambda compiled performance by more than an order of magnitude. Here's a tip; after running the profiler, 92% of the time is spent on:

System.Reflection.Emit.DynamicMethod.CreateDelegate(class System.Type, object)

Hmmmm... Why is it creating a new delegate in every iteration? I'm not sure, but the solution follows in a separate post.

Bonnibelle

Are these timings from running in Visual Studio? If so, repeat the timings using a Release mode build and run without debugging (i.e. Ctrl+F5 in Visual Studio, or from the command line). Also, consider using Stopwatch for timings rather than DateTime.Now. – Medovich

I don't know why it is slower, but your benchmark technique is not very good. First off, DateTime.Now is only accurate to 1/64 of a second, so your measurement rounding error is large. Use Stopwatch instead; it is accurate to a few nanoseconds. Second, you are measuring both the time to jit the code (the first call) and every subsequent call; that can throw off averages. (Though in this case a MAX of a hundred thousand is probably enough to average away the jit burden, still, it is a bad practice to include it in the average.) – Boycie

@Eric, rounding error can only be there if in each operation DateTime.Now.Ticks is used, before start and after end, the millisecond counts are high enough to show performance difference. – Inventive

if using stopwatch, i recommend following this article to ensure accurate results: codeproject.com/KB/testing/stopwatch-measure-precise.aspx – Uziel

@Eric, while I agree it is not the most precise measurement technique available, we are talking about an order of magnitude of difference. MAX is high enough to reduce significant deviations. – Bonnibelle

@Jim, done and added the results to the original post. – Bonnibelle

It sounds like you actually solved the problem. If so, you should add it as an answer to this question. – Hyperpyrexia

Answer 1

Could it be that the inner lambdas are not being compiled?!? Here's a proof of concept:

static void UsingCompiledExpressionWithMethodCall() {
        var where = typeof(Enumerable).GetMember("Where").First() as System.Reflection.MethodInfo;
        where = where.MakeGenericMethod(typeof(int));
        var l = Expression.Parameter(typeof(IEnumerable<int>), "l");
        var arg0 = Expression.Parameter(typeof(int), "i");
        var lambda0 = Expression.Lambda<Func<int, bool>>(
            Expression.Equal(Expression.Modulo(arg0, Expression.Constant(2)),
                             Expression.Constant(0)), arg0).Compile();
        var c1 = Expression.Call(where, l, Expression.Constant(lambda0));
        var arg1 = Expression.Parameter(typeof(int), "i");
        var lambda1 = Expression.Lambda<Func<int, bool>>(Expression.GreaterThan(arg1, Expression.Constant(5)), arg1).Compile();
        var c2 = Expression.Call(where, c1, Expression.Constant(lambda1));

        var f = Expression.Lambda<Func<IEnumerable<int>, IEnumerable<int>>>(c2, l);

        var c3 = f.Compile();

        var t0 = DateTime.Now.Ticks;
        for (int j = 1; j < MAX; j++)
        {
            var sss = c3(x).ToList();
        }

        var tn = DateTime.Now.Ticks;
        Console.WriteLine("Using lambda compiled with MethodCall: {0}", tn - t0);
    }

And now the timings are:

Using lambda:                            625020
Using lambda compiled:                 14687970
Using lambda combined:                   468765
Using lambda compiled with MethodCall:   468765

Woot! Not only it is fast, it is faster than the native lambda. (Scratch head).

---

Of course the above code is simply too painful to write. Let's do some simple magic:

static void UsingCompiledConstantExpressions() {
    var f1 = (Func<IEnumerable<int>, IEnumerable<int>>)(l => l.Where(i => i % 2 == 0));
    var f2 = (Func<IEnumerable<int>, IEnumerable<int>>)(l => l.Where(i => i > 5));
    var argX = Expression.Parameter(typeof(IEnumerable<int>), "x");
    var f3 = Expression.Invoke(Expression.Constant(f2), Expression.Invoke(Expression.Constant(f1), argX));
    var f = Expression.Lambda<Func<IEnumerable<int>, IEnumerable<int>>>(f3, argX);

    var c3 = f.Compile();

    var t0 = DateTime.Now.Ticks;
    for (int j = 1; j < MAX; j++) {
        var sss = c3(x).ToList();
    }

    var tn = DateTime.Now.Ticks;
    Console.WriteLine("Using lambda compiled constant: {0}", tn - t0);
}

And some timings, VS2010, Optimizations ON, Debugging OFF:

Using lambda:                            781260
Using lambda compiled:                 14687970
Using lambda combined:                   468756
Using lambda compiled with MethodCall:   468756
Using lambda compiled constant:          468756

Now you could argue that I'm not generating the whole expression dynamically; just the chaining invocations. But in the above example I generate the whole expression. And the timings match. This is just a shortcut to write less code.

---

From my understanding, what is going on is that the .Compile() method does not propagate the compilations to inner lambdas, and thus the constant invocation of CreateDelegate. But to truly understand this, I would love to have a .NET guru comment a little about the internal stuff going on.

And why, oh why is this now faster than a native lambda!?

Bonnibelle

I'm thinking in accepting my own answer, since it's the one with the most up votes. Should I wait a little longer? – Bonnibelle

About what happens with you getting code faster than the native lambda, you might want to take a look at this page about microbenchmarks (which has nothing really Java-specific, the name notwithstanding): code.google.com/p/caliper/wiki/JavaMicrobenchmarks – Falsify

As for why the dynamically compiled lambda is faster, I suspect that "using lambda", being run first, is being penalized with having to JIT some code. – Uncontrollable

I do not know what's happening, once when I tested compiled expression and createdelegate for setting and getting from fields and properties, createdelegate was way faster for properties, but compiled was very slightly faster for fields – Puett

Answer 2

Recently I asked an almost identical question:

Performance of compiled-to-delegate Expression

The solution for me was that I shouldn't call Compile on the Expression, but that I should call CompileToMethod on it and compile the Expression to a static method in a dynamic assembly.

Like so:

var assemblyBuilder = AppDomain.CurrentDomain.DefineDynamicAssembly(
  new AssemblyName("MyAssembly_" + Guid.NewGuid().ToString("N")), 
  AssemblyBuilderAccess.Run);

var moduleBuilder = assemblyBuilder.DefineDynamicModule("Module");

var typeBuilder = moduleBuilder.DefineType("MyType_" + Guid.NewGuid().ToString("N"), 
  TypeAttributes.Public));

var methodBuilder = typeBuilder.DefineMethod("MyMethod", 
  MethodAttributes.Public | MethodAttributes.Static);

expression.CompileToMethod(methodBuilder);

var resultingType = typeBuilder.CreateType();

var function = Delegate.CreateDelegate(expression.Type,
  resultingType.GetMethod("MyMethod"));

It's not ideal however. I'm not quite certain to which types this applies exactly, but I think that types that are taken as parameters by the delegate, or returned by the delegate have to be public and non-generic. It has to be non-generic because generic types apparently access System.__Canon which is an internal type used by .NET under the hood for generic types and this violates the "has to be a public type rule).

For those types, you can use the apparently slower Compile. I detect them in the following way:

private static bool IsPublicType(Type t)
{

  if ((!t.IsPublic && !t.IsNestedPublic) || t.IsGenericType)
  {
    return false;
  }

  int lastIndex = t.FullName.LastIndexOf('+');

  if (lastIndex > 0)
  {
    var containgTypeName = t.FullName.Substring(0, lastIndex);

    var containingType = Type.GetType(containgTypeName + "," + t.Assembly);

    if (containingType != null)
    {
      return containingType.IsPublic;
    }

    return false;
  }
  else
  {
    return t.IsPublic;
  }
}

But like I said, this isn't ideal and I would still like to know why compiling a method to a dynamic assembly is sometimes an order of magnitude faster. And I say sometimes because I've also seen cases where an Expression compiled with Compile is just as fast as a normal method. See my question for that.

Or if someone knows a way to bypass the "no non-public types" constraint with the dynamic assembly, that's welcome as well.

Biedermeier

Answer 3

Your expressions are not equivalent and thus you get skewed results. I wrote a test bench to test this. The tests include the regular lambda call, the equivalent compiled expression, a hand made equivalent compiled expression, as well as composed versions. These should be more accurate numbers. Interestingly, I'm not seeing much variation between the plain and composed versions. And the compiled expressions are slower naturally but only by very little. You need a large enough input and iteration count to get some good numbers. It makes a difference.

As for your second question, I don't know how you'd be able to get more performance out of this so I can't help you there. It looks as good as it's going to get.

You'll find my answer to your third question in the HandMadeLambdaExpression() method. Not the easiest expression to build due to the extension methods, but doable.

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

using System.Diagnostics;
using System.Linq.Expressions;

namespace ExpressionBench
{
    class Program
    {
        static void Main(string[] args)
        {
            var values = Enumerable.Range(0, 5000);
            var lambda = GetLambda();
            var lambdaExpression = GetLambdaExpression().Compile();
            var handMadeLambdaExpression = GetHandMadeLambdaExpression().Compile();
            var composed = GetComposed();
            var composedExpression = GetComposedExpression().Compile();
            var handMadeComposedExpression = GetHandMadeComposedExpression().Compile();

            DoTest("Lambda", values, lambda);
            DoTest("Lambda Expression", values, lambdaExpression);
            DoTest("Hand Made Lambda Expression", values, handMadeLambdaExpression);
            Console.WriteLine();
            DoTest("Composed", values, composed);
            DoTest("Composed Expression", values, composedExpression);
            DoTest("Hand Made Composed Expression", values, handMadeComposedExpression);
        }

        static void DoTest<TInput, TOutput>(string name, TInput sequence, Func<TInput, TOutput> operation, int count = 1000000)
        {
            for (int _ = 0; _ < 1000; _++)
                operation(sequence);
            var sw = Stopwatch.StartNew();
            for (int _ = 0; _ < count; _++)
                operation(sequence);
            sw.Stop();
            Console.WriteLine("{0}:", name);
            Console.WriteLine("  Elapsed: {0,10} {1,10} (ms)", sw.ElapsedTicks, sw.ElapsedMilliseconds);
            Console.WriteLine("  Average: {0,10} {1,10} (ms)", decimal.Divide(sw.ElapsedTicks, count), decimal.Divide(sw.ElapsedMilliseconds, count));
        }

        static Func<IEnumerable<int>, IList<int>> GetLambda()
        {
            return v => v.Where(i => i % 2 == 0).Where(i => i > 5).ToList();
        }

        static Expression<Func<IEnumerable<int>, IList<int>>> GetLambdaExpression()
        {
            return v => v.Where(i => i % 2 == 0).Where(i => i > 5).ToList();
        }

        static Expression<Func<IEnumerable<int>, IList<int>>> GetHandMadeLambdaExpression()
        {
            var enumerableMethods = typeof(Enumerable).GetMethods();
            var whereMethod = enumerableMethods
                .Where(m => m.Name == "Where")
                .Select(m => m.MakeGenericMethod(typeof(int)))
                .Where(m => m.GetParameters()[1].ParameterType == typeof(Func<int, bool>))
                .Single();
            var toListMethod = enumerableMethods
                .Where(m => m.Name == "ToList")
                .Select(m => m.MakeGenericMethod(typeof(int)))
                .Single();

            // helpers to create the static method call expressions
            Func<Expression, ParameterExpression, Func<ParameterExpression, Expression>, Expression> WhereExpression =
                (instance, param, body) => Expression.Call(whereMethod, instance, Expression.Lambda(body(param), param));
            Func<Expression, Expression> ToListExpression =
                instance => Expression.Call(toListMethod, instance);

            //return v => v.Where(i => i % 2 == 0).Where(i => i > 5).ToList();
            var exprParam = Expression.Parameter(typeof(IEnumerable<int>), "v");
            var expr0 = WhereExpression(exprParam,
                Expression.Parameter(typeof(int), "i"),
                i => Expression.Equal(Expression.Modulo(i, Expression.Constant(2)), Expression.Constant(0)));
            var expr1 = WhereExpression(expr0,
                Expression.Parameter(typeof(int), "i"),
                i => Expression.GreaterThan(i, Expression.Constant(5)));
            var exprBody = ToListExpression(expr1);
            return Expression.Lambda<Func<IEnumerable<int>, IList<int>>>(exprBody, exprParam);
        }

        static Func<IEnumerable<int>, IList<int>> GetComposed()
        {
            Func<IEnumerable<int>, IEnumerable<int>> composed0 =
                v => v.Where(i => i % 2 == 0);
            Func<IEnumerable<int>, IEnumerable<int>> composed1 =
                v => v.Where(i => i > 5);
            Func<IEnumerable<int>, IList<int>> composed2 =
                v => v.ToList();
            return v => composed2(composed1(composed0(v)));
        }

        static Expression<Func<IEnumerable<int>, IList<int>>> GetComposedExpression()
        {
            Expression<Func<IEnumerable<int>, IEnumerable<int>>> composed0 =
                v => v.Where(i => i % 2 == 0);
            Expression<Func<IEnumerable<int>, IEnumerable<int>>> composed1 =
                v => v.Where(i => i > 5);
            Expression<Func<IEnumerable<int>, IList<int>>> composed2 =
                v => v.ToList();
            var exprParam = Expression.Parameter(typeof(IEnumerable<int>), "v");
            var exprBody = Expression.Invoke(composed2, Expression.Invoke(composed1, Expression.Invoke(composed0, exprParam)));
            return Expression.Lambda<Func<IEnumerable<int>, IList<int>>>(exprBody, exprParam);
        }

        static Expression<Func<IEnumerable<int>, IList<int>>> GetHandMadeComposedExpression()
        {
            var enumerableMethods = typeof(Enumerable).GetMethods();
            var whereMethod = enumerableMethods
                .Where(m => m.Name == "Where")
                .Select(m => m.MakeGenericMethod(typeof(int)))
                .Where(m => m.GetParameters()[1].ParameterType == typeof(Func<int, bool>))
                .Single();
            var toListMethod = enumerableMethods
                .Where(m => m.Name == "ToList")
                .Select(m => m.MakeGenericMethod(typeof(int)))
                .Single();

            Func<ParameterExpression, Func<ParameterExpression, Expression>, Expression> LambdaExpression =
                (param, body) => Expression.Lambda(body(param), param);
            Func<Expression, ParameterExpression, Func<ParameterExpression, Expression>, Expression> WhereExpression =
                (instance, param, body) => Expression.Call(whereMethod, instance, Expression.Lambda(body(param), param));
            Func<Expression, Expression> ToListExpression =
                instance => Expression.Call(toListMethod, instance);

            var composed0 = LambdaExpression(Expression.Parameter(typeof(IEnumerable<int>), "v"),
                v => WhereExpression(
                    v,
                    Expression.Parameter(typeof(int), "i"),
                    i => Expression.Equal(Expression.Modulo(i, Expression.Constant(2)), Expression.Constant(0))));
            var composed1 = LambdaExpression(Expression.Parameter(typeof(IEnumerable<int>), "v"),
                v => WhereExpression(
                    v,
                    Expression.Parameter(typeof(int), "i"),
                    i => Expression.GreaterThan(i, Expression.Constant(5))));
            var composed2 = LambdaExpression(Expression.Parameter(typeof(IEnumerable<int>), "v"),
                v => ToListExpression(v));

            var exprParam = Expression.Parameter(typeof(IEnumerable<int>), "v");
            var exprBody = Expression.Invoke(composed2, Expression.Invoke(composed1, Expression.Invoke(composed0, exprParam)));
            return Expression.Lambda<Func<IEnumerable<int>, IList<int>>>(exprBody, exprParam);
        }
    }
}

And the results on my machine:

Lambda:
  Elapsed:  340971948     123230 (ms)
  Average: 340.971948    0.12323 (ms)
Lambda Expression:
  Elapsed:  357077202     129051 (ms)
  Average: 357.077202   0.129051 (ms)
Hand Made Lambda Expression:
  Elapsed:  345029281     124696 (ms)
  Average: 345.029281   0.124696 (ms)

Composed:
  Elapsed:  340409238     123027 (ms)
  Average: 340.409238   0.123027 (ms)
Composed Expression:
  Elapsed:  350800599     126782 (ms)
  Average: 350.800599   0.126782 (ms)
Hand Made Composed Expression:
  Elapsed:  352811359     127509 (ms)
  Average: 352.811359   0.127509 (ms)

Araby

Answer 4

Compiled lambda performance over delegates may be slower because Compiled code at runtime may not be optimized however the code you wrote manually and that compiled via C# compiler is optimized.

Second, multiple lambda expressions means multiple anonymous methods, and calling each of them takes little extra time over evaluating a straight method. For example, calling

Console.WriteLine(x);

and

Action x => Console.WriteLine(x);
x(); // this means two different calls..

are different, and with second one little more overhead is required as from compiler's perspective, its actually two different calls. First calling x itself and then within that calling x's statement.

So your combined Lambda will certainly have little slow performance over single lambda expression.

And this is independent of what is executing inside, because you are still evaluating correct logic, but you are adding additional steps for compiler to perform.

Even after expression tree is compiled, it will not have optimization, and it will still preserve its little complex structure, evaluating and calling it may have extra validation, null check etc which might be slowing down performance of compiled lambda expressions.

Inventive

If you look closely, the UsingLambdaCombined test is combining multiple lambda functions, and its performance is very close to UsingLambda. Regarding the optimizations, I was convinced that they were handled by the JIT engine, and thus runtime generated code (after compilation), would also be target of any JIT optimizations. – Bonnibelle

JIT optimization and compile time optimization are two different things you can turn off compile time optimization in project settings. Secondly, expression compilation will probably emit dynamic MSIL which again will be little slower as it's logic and sequence of operation will contain null checks and validity as per needs. You can look in reflector regarding how it is compiled. – Inventive

While your reasoning is sound, I have to disagree with you on this particular problem (i.e., the order of magnitude difference is not due to static compilation). First, because if you actually disable the compile-time optimizations, the difference is still considerable. Secondly, because I've already found a way to optimize the dynamic generation to be just marginally slower. Let me try to understand "why" and I'll post the results. – Bonnibelle