views:

254

answers:

4

I'm using .NET 3.5. We have some complex third-party classes which are automatically generated and out of my control, but which we must work with for testing purposes. I see my team doing a lot of deeply-nested property getting/setting in our test code, and it's getting pretty cumbersome.

To remedy the problem, I'd like to make a fluent interface for setting properties on the various objects in the hierarchical tree. There are a large number of properties and classes in this third-party library, and it would be too tedious to map everything manually.

My initial thought was to just use object initializers. Red, Blue, and Green are properties, and Mix() is a method that sets a fourth property Color to the closest RGB-safe color with that mixed color. Paints must be homogenized with Stir() before they can be used.

Bucket b = new Bucket() {
  Paint = new Paint() {
    Red = 0.4;
    Blue = 0.2;
    Green = 0.1;
  }
};

That works to initialize the Paint, but I need to chain Mix() and other methods to it. Next attempt:

Create<Bucket>(Create<Paint>()
  .SetRed(0.4)
  .SetBlue(0.2)
  .SetGreen(0.1)
  .Mix().Stir()
)

But that doesn't scale well, because I'd have to define a method for each property I want to set, and there are hundreds of different properties in all the classes. Also, C# doesn't have a way to dynamically define methods prior to C# 4, so I don't think I can hook into things to do this automatically in some way.

Third attempt:

Create<Bucket>(Create<Paint>().Set(p => {
    p.Red = 0.4;
    p.Blue = 0.2;
    p.Green = 0.1;
  }).Mix().Stir()
)

That doesn't look too bad, and seems like it'd be feasible. Is this an advisable approach? Is it possible to write a Set method that works this way? Or should I be pursuing an alternate strategy?

+5  A: 

Does this work?

Bucket b = new Bucket() {
  Paint = new Paint() {
    Red = 0.4;
    Blue = 0.2;
    Green = 0.1;
  }.Mix().Stir()
};

Assuming Mix() and Stir() are defined to return a Paint object.

To call methods that return void, you can use an extension method that will allow you to perform additional initialization on the object you pass in:

public static T Init<T>(this T @this, Action<T> initAction) {
    if (initAction != null)
        initAction(@this);
    return @this;
}

Which could be used similar to Set() as described:

Bucket b = new Bucket() {
  Paint = new Paint() {
    Red = 0.4;
    Blue = 0.2;
    Green = 0.1;
  }.Init(p => {
    p.Mix().Stir();
  })
};
dahlbyk
In retrospect, this seems like an obvious approach. I must be missing the forest for the trees from staring at this for too long!
John Feminella
FYI for those puzzled by `@this`: the `@` means "treat the following token as a literal identifier" (not as a keyword). The effect is to name a parameter called `this`.
John Feminella
I'm not puzzled by it, but I don't see any reason to fight the compiler here when you can use an equivalent name like `self` or `source`. Otherwise, good answer and +1.
Aaronaught
+1  A: 

I would think of it this way:

You essentially want your last method in the chain to return a Bucket. In your case, I think you want that method to be Mix(), as you can Stir() the bucket afterwards

public class BucketBuilder
{
    private int _red = 0;
    private int _green = 0;
    private int _blue = 0;

    public Bucket Mix()
    {
        Bucket bucket = new Bucket(_paint);
        bucket.Mix();
        return bucket;
    }
}

So you need to set at least one colour before you call Mix(). Let's force that with some Syntax interfaces.

public interface IStillNeedsMixing : ICanAddColours
{
     Bucket Mix();
}

public interface ICanAddColours
{
     IStillNeedsMixing Red(int red);
     IStillNeedsMixing Green(int green);
     IStillNeedsMixing Blue(int blue);
}

And let's apply these to the BucketBuilder

public class BucketBuilder : IStillNeedsMixing, ICanAddColours
{
    private int _red = 0;
    private int _green = 0;
    private int _blue = 0;

    public IStillNeedsMixing Red(int red)
    {
         _red += red;
         return this;
    }

    public IStillNeedsMixing Green(int green)
    {
         _green += green;
         return this;
    }

    public IStillNeedsMixing Blue(int blue)
    {
         _blue += blue;
         return this;
    }

    public Bucket Mix()
    {
        Bucket bucket = new Bucket(new Paint(_red, _green, _blue));
        bucket.Mix();
        return bucket;
    }
}

Now you need an initial static property to kick off the chain

public static class CreateBucket
{
    public static ICanAddColours UsingPaint
    {
        return new BucketBuilder();
    }
}

And that's pretty much it, you now have a fluent interface with optional RGB parameters (as long as you enter at least one) as a bonus.

CreateBucket.UsingPaint.Red(0.4).Green(0.2).Mix().Stir();

The thing with Fluent Interfaces is that they're not that easy to put together, but they are easy for the developer to code against and they are very extensible. If you want to add a Matt/Gloss flag to this without changing all of your calling code, it's easy to do.

Also, if the provider of your API changes everything underneath you, you only have to rewrite this one piece of code; all the callin code can remain the same.

pdr
A: 

I would use the Init extension method because U can always play with the delegate. Hell You can always declare extension methods that take up expressions and even play up with the expresions (store them for later, modify, whatever) This way You can easily store default grups like:

Create<Paint>(() => new Paint{p.Red = 0.3, p.Blue = 0.2, p.Green = 0.1}).
Init(p => p.Mix().Stir())

This Way You can use all the actions (or funcs) and cache standard initializers as expression chains for later?

luckyluke
A: 

If you really want to be able to chain property settings without having to write a ton of code, one way to do this would be to use code generation (CodeDom). You can use Reflection to get a list of the mutable properties, the generate a fluent builder class with a final Build() method that returns the class you're actually trying to create.

I'm going to skip over all the boilerplate stuff about how to register the custom tool - that's fairly easy to find documentation on but still long-winded and I don't think I'd be adding much by including it. I will show you what I'm thinking of for the codegen though.

public static class PropertyBuilderGenerator
{
    public static CodeTypeDeclaration GenerateBuilder(Type destType)
    {
        if (destType == null)
            throw new ArgumentNullException("destType");
        CodeTypeDeclaration builderType = new
            CodeTypeDeclaration(destType.Name + "Builder");
        builderType.TypeAttributes = TypeAttributes.Public;
        CodeTypeReference destTypeRef = new CodeTypeReference(destType);
        CodeExpression resultExpr = AddResultField(builderType, destTypeRef);
        PropertyInfo[] builderProps = destType.GetProperties(
            BindingFlags.Instance | BindingFlags.Public);
        foreach (PropertyInfo prop in builderProps)
        {
            AddPropertyBuilder(builderType, resultExpr, prop);
        }
        AddBuildMethod(builderType, resultExpr, destTypeRef);
        return builderType;
    }

    private static void AddBuildMethod(CodeTypeDeclaration builderType,
        CodeExpression resultExpr, CodeTypeReference destTypeRef)
    {
        CodeMemberMethod method = new CodeMemberMethod();
        method.Attributes = MemberAttributes.Public | MemberAttributes.Final;
        method.Name = "Build";
        method.ReturnType = destTypeRef;
        method.Statements.Add(new MethodReturnStatement(resultExpr));
        builderType.Members.Add(method);
    }

    private static void AddPropertyBuilder(CodeTypeDeclaration builderType,
        CodeExpression resultExpr, PropertyInfo prop)
    {
        CodeMemberMethod method = new CodeMemberMethod();
        method.Attributes = MemberAttributes.Public | MemberAttributes.Final;
        method.Name = prop.Name;
        method.ReturnType = new CodeTypeReference(builderType.Name);
        method.Parameters.Add(new CodeParameterDeclarationExpression(prop.Type,
            "value"));
        method.Statements.Add(new CodeAssignStatement(
            new CodePropertyReferenceExpression(resultExpr, prop.Name),
            new CodeArgumentReferenceExpression("value")));
        method.Statements.Add(new MethodReturnStatement(
            new CodeThisExpression()));
        builderType.Members.Add(method);
    }

    private static CodeFieldReferenceExpression AddResultField(
        CodeTypeDeclaration builderType, CodeTypeReference destTypeRef)
    {
        const string fieldName = "_result";
        CodeMemberField resultField = new CodeMemberField(destTypeRef, fieldName);
        resultField.Attributes = MemberAttributes.Private;
        builderType.Members.Add(resultField);
        return new CodeFieldReferenceExpression(
            new CodeThisReferenceExpression(), fieldName);
    }
}

I think this should just about do it - it's obviously untested, but where you go from here is that you create a codegen (inheriting from BaseCodeGeneratorWithSite) that compiles a CodeCompileUnit populated with a list of types. That list comes from the file type you register with the tool - in this case I'd probably just make it a text file with a line-delimited list of types that you want to generate builder code for. Have the tool scan this, load the types (might have to load the assemblies first), and generate bytecode.

It's tough, but not as tough as it sounds, and when you're done you'll be able to write code like this:

Paint p = new PaintBuilder().Red(0.4).Blue(0.2).Green(0.1).Build().Mix.Stir();

Which I believe is almost exactly what you want. All you have to do to invoke the code generation is register the tool with a custom extension (let's say .buildertypes), put a file with that extension in your project, and put a list of types in it:

MyCompany.MyProject.Paint
MyCompany.MyProject.Foo
MyCompany.MyLibrary.Bar

And so on. When you save, it will automatically generate the code file you need that supports writing statements like the one above.

I've used this approach before for a highly convoluted messaging system with several hundred different message types. It was taking too long to always construct the message, set a bunch of properties, send it through the channel, receive from the channel, serialize the response, etc... using a codegen greatly simplified the work as it enabled me to generate a single messaging class that took all of the individual properties as arguments and spit back a response of the correct type. It's not something I would recommend to everyone, but when you're dealing with very large projects, sometimes you need to start inventing your own syntax!

Aaronaught