One of the bugbears I have with Entity Framework in .NET is that when you use it behind a method in the business layer (think AddAuthor(Author author)) you need to manually wipe non-scalar properties when adding entities. I define non-scalar properties as:

  • EntityCollections: the "many" ends of relationships (think Author.Books)
  • The single end of relationships (think Book.Author)
  • Properties that participate in the Entity Key (primary key properties) (think Author.ID). These need to be wiped as the database will automatically fill them (identity columns in SQL Server)

If you don't wipe the relationship properties when you add the entity object you get an UpdateException when you try to SaveChanges on the ObjectContext.

You can't assume that someone calling your method knows about this Entity Framework behaviour, so I consider it good practice to wipe the properties for them so they don't see unexpected UpdateExceptions floating up from the business layer. Obviously, having to write

author.ID = default(int);

is no fun (especially on entities that have more properties than this!). So I wrote a method that you can pass an EntityObject to and it will wipe all non-scalar properties for you using reflection. This means that it doesn't need to know about your specific entity type until you call it, so you can use whatever entity classes you have for your project with it. However, calling methods dynamically is really slow. There will probably be lots faster ways of doing it in .NET 4.0 because of the DLR, but for us still back here in 3.5-land it is still slow.

So how can we enjoy the benefits of reflection but without the performance cost? Code generation at runtime, that's how! Instead of using reflection and dynamically calling methods and setting properties, we can instead use reflection once, create a new class at runtime that is able to wipe the non-scalar properties and then reuse this class over and over. How do we do this? With the System.CodeDom API, that's how (maybe in .NET 4.0 you could use the expanded Expression Trees functionality). The CodeDom API allows you to literally write code at runtime and have it compiled and loaded for you.

The following code I will go through is available as a part of DigitallyCreated Utilities open source libraries (see the DigitallyCreated.Utilities.Linq.EfUtil.ClearNonScalarProperties() method).

What I have done is create an interface called IEntityPropertyClearer that has a method that takes an object and wipes the non-scalar properties on it. Classes that implement this interface are able to wipe a single type of entity (EntityType).

public interface IEntityPropertyClearer
    Type EntityType { get; }

    void Clear(object entity);

I then have a helper abstract class that makes this interface easy to implement by the code generation logic by providing a type-safe method (ClearEntity) to implement and by doing some boilerplate generic magic for the EntityType property and casting in Clear:

public abstract class AbstractEntityPropertyClearer<TEntity> : IEntityPropertyClearer
    where TEntity : EntityObject
    public Type EntityType
        get { return typeof(TEntity); }
    public void Clear(object entity)
        if (entity is TEntity)

    protected abstract void ClearEntity(TEntity entity);

So the aim is to create a class at runtime that inherits from AbstractEntityPropertyClearer and implements the ClearEntity method so that it can clear a particular entity type. I have a method that we will now implement:

private static IEntityPropertyClearer GeneratePropertyClearer(EntityObject entity)
    Type entityType = entity.GetType();

So the first thing we need to do is to create a "CodeCompileUnit" to put all the code we are generating into:

CodeCompileUnit compileUnit = new CodeCompileUnit();
compileUnit.ReferencedAssemblies.Add(typeof(System.ComponentModel.INotifyPropertyChanging).Assembly.Location); //System.dll

Note that we get the path to the assemblies we want to reference from the actual types that we need. I think this is a good approach as it means we don't need to hardcode the paths to the assemblies (maybe they will move in the future?).

We then need to create the class that will inherit from AbstractEntityPropertyClearer and the namespace in which this class will reside:

//Create the namespace
string namespaceName = typeof(EfUtil).Namespace + ".CodeGen";
CodeNamespace codeGenNamespace = new CodeNamespace(namespaceName);

//Create the class
string genTypeName = entityType.FullName.Replace('.', '_') + "PropertyClearer";
CodeTypeDeclaration genClass = new CodeTypeDeclaration(genTypeName);
genClass.IsClass = true;

Type baseType = typeof(AbstractEntityPropertyClearer<>).MakeGenericType(entityType);
genClass.BaseTypes.Add(new CodeTypeReference(baseType));

The namespace we create is the namespace of the utility class + ".CodeGen". The class's name is the entity's full name (including namespace) where all "."s are replaced with "_"s and PropertyClearer appended to it (this will stop name collision). The class that the generated class will inherit from is AbstractEntityPropertyClearer but with the generic type set to be the type of entity we are dealing with (ie if the method was called with an Author, the type would be AbstractEntityPropertyClearer<Author>).

We now need to create the ClearEntity method that will go inside this class:

CodeMemberMethod clearEntityMethod = new CodeMemberMethod();
clearEntityMethod.Name = "ClearEntity";
clearEntityMethod.ReturnType = new CodeTypeReference(typeof(void));
clearEntityMethod.Parameters.Add(new CodeParameterDeclarationExpression(entityType, "entity"));
clearEntityMethod.Attributes = MemberAttributes.Override | MemberAttributes.Family;

Counterintuitively (for C# developers, anyway), "protected" scope is called MemberAttributes.Family in CodeDom.

We now need to find all EntityCollection properties on our entity type so that we can generate code to wipe them. We can do that with a smattering of LINQ against the reflection API:

IEnumerable<PropertyInfo> entityCollections = 
        from property in entityType.GetProperties()
          property.PropertyType.IsGenericType &&
          property.PropertyType.IsGenericTypeDefinition == false &&
          property.PropertyType.GetGenericTypeDefinition() ==
        select property;

We now need to generate statements inside our generated method to call Clear() on each of these properties:

foreach (PropertyInfo propertyInfo in entityCollections)
    CodePropertyReferenceExpression propertyReferenceExpression = new CodePropertyReferenceExpression(new CodeArgumentReferenceExpression("entity"), propertyInfo.Name);
    clearEntityMethod.Statements.Add(new CodeMethodInvokeExpression(propertyReferenceExpression, "Clear"));

On line 3 above we create a CodePropertyReferenceExpression that refers to the property on the "entity" variable which is an argument that we defined for the generated method. We then add to the method an expression that invokes the Clear method on this property reference (line 4). This will give us statements like entity.Books.Clear() (where entity is an Author).

We now need to find all the single-end-of-a-relationship properties (like book.Author) and entity key properties (like author.ID). Again, we use some LINQ to achieve this:

//Find all single multiplicity relation ends
IEnumerable<PropertyInfo> relationSingleEndProperties = 
        from property in entityType.GetProperties()
        from attribute in property.GetCustomAttributes(typeof(EdmRelationshipNavigationPropertyAttribute), true).Cast<EdmRelationshipNavigationPropertyAttribute>()
        select property;

//Find all entity key properties
IEnumerable<PropertyInfo> idProperties = 
        from property in entityType.GetProperties()
        from attribute in property.GetCustomAttributes(typeof(EdmScalarPropertyAttribute), true).Cast<EdmScalarPropertyAttribute>()
        where attribute.EntityKeyProperty
        select property;

We then can iterate over both these sets in the one foreach loop (by using .Concat) and for each property we will generate a statement that will set it to its default value (using a default(MyType) expression):

//Emit assignments that set the properties to their default value
foreach (PropertyInfo propertyInfo in relationSingleEndProperties.Concat(idProperties))
    CodeExpression defaultExpression = new CodeDefaultValueExpression(new CodeTypeReference(propertyInfo.PropertyType));

    CodePropertyReferenceExpression propertyReferenceExpression = new CodePropertyReferenceExpression(new CodeArgumentReferenceExpression("entity"), propertyInfo.Name);
    clearEntityMethod.Statements.Add(new CodeAssignStatement(propertyReferenceExpression, defaultExpression));

This will generate statements like author.ID = default(int) and book.Author = default(Author) which are added to the generated method.

Now that we've generated code to wipe out the non-scalar properties on an entity, we need to compile this code. We do this by passing our CodeCompileUnit to a CSharpCodeProvider for compilation:

CSharpCodeProvider provider = new CSharpCodeProvider();
CompilerParameters parameters = new CompilerParameters();
parameters.GenerateInMemory = true;
CompilerResults results = provider.CompileAssemblyFromDom(parameters, compileUnit);

We set GenerateInMemory to true, as we just want these types to be available as long as our App Domain exists. GenerateInMemory causes the types that we generated to be automatically loaded, so we now need to instantiate our new class and return it:

Type type = results.CompiledAssembly.GetType(namespaceName + "." + genTypeName);
return (IEntityPropertyClearer)Activator.CreateInstance(type);

What DigitallyCreated Utilities does is keep a Dictionary of instances of these types in memory. If you call on it with an entity type that it hasn't seen before it will generate a new IEntityPropertyClearer for that type using the method we just created and save it in the dictionary. This means that we can reuse these instances as well as keep track of which entity types we've seen before so we don't try to regenerate the clearer class. The method that does this is the method that you call to get your entities wiped:

public static void ClearNonScalarProperties(EntityObject entity)
    IEntityPropertyClearer propertyClearer;

    //_PropertyClearers is a private static readonly IDictionary<Type, IEntityPropertyClearer>
    lock (_PropertyClearers)
        if (_PropertyClearers.ContainsKey(entity.GetType()) == false)
            propertyClearer = GeneratePropertyClearer(entity);
            _PropertyClearers.Add(entity.GetType(), propertyClearer);
            propertyClearer = _PropertyClearers[entity.GetType()];


Note that this is done inside a lock so that this method is thread-safe.

So now the question is: how much faster is doing this using code generation instead of just using reflection and dynamically calling Clear() etc?

I created a small benchmark that creates 10,000 Author objects (Authors have an ID, a Name and an EntityCollection of Books) and then wipes them with ClearNonScalarProperties (which uses the code generation). It does this 100 times. It then does the same thing but this time it uses reflection only. Here are the results:

  Average Standard Deviation
Using Code Generation 466.0ms 39.9ms
Using Reflection Only 1817.5ms 11.6ms

As you can see, when using code generation, the code runs nearly four times as fast compared to when using reflection only. I assume that when an entity has more non-scalar properties than the paltry two that Author has this speed benefit would be even more pronounced.

Even though this code generation logic is harder to write than just doing dynamic method calls using reflection, the results are really worth it in terms of performance. The code for this is up as a part of DigitallyCreated Utilities (it's in trunk and not part of an official release at the time of writing), so if you're interested in seeing it in action go check it out there.