Skip to main content
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] [List Home]
Re: [aspectj-users] Using AOP to implement double dispatching technique 
Hi Rudi,

I think you are mixing up 2 different techniques in your implementation:
- static dispatching (like usual Java-code, but now using introduced methods) - "dynamic" dispatching (using dynamic pointcutcut designators, advice, etc.) 

Within your code your have both implemented:
- the after advice rather looks like a dynamic dispatcher (by the way: you should use an around advice) - the introduced methods "displayOn" are the "real" static double-dispatchers
Since you seem to be concerned about performance and contamination of classes, your should skip the advice/pointcut and move everything belonging to the double dispatching to an aspect: 

- introduce "display(GraphicalElement ge)" to class Screen.
- introduce a "displayOn (Screen screen)" to Point and Line.

So, the code should look like:

aspect ScreenGraphicalElementDDs {
  public void Screen.aoDisplay(GraphicalElement ge) {
    ge.aoDisplayOn(this);
  }
  public void (Point || Line).aoDisplayOn(Screen s) {
    s.display(this);
  }
  declare parents: (GraphicalElement) extends AODisplay;
}
interface AODisplay {
  public void aoDisplayOn(Screen s);
}

Some remarks:
1. you could put the code for performance measurement to an aspect
2. from my point of view GraphicalElement should not contain the double dispatch method: instead this should be shifted to an aspect 3. I am still accustomed to AspectJ 1.06 and i am not familiar with how introductions are now specified. Maybe the introductions like (Point || Line) are no longer valid in 1.1 and you have to add 2 different introductions to it. 

Regards,
Stefan


Vankeirsbilck, Rudi. (BE - Merelbeke) wrote:

Hi all,
First some introduction on double dispatching, my actual question is below. Very frequently, I use the double dispatching technique to solve the following problem: Consider 4 classes, A, B, C and T where A is the abstract superclass of B and C, T is a third party (i.e. with respect to A, B and C). The latter needs to process A's and has different processing for B and C types of A. You can image that there will be a method like 
public void process (A anA)
defined on the T class.
Most people will implement this method as: 
public void process (A anA) {
    if (anA instanceof B) {
        B aB = (B) anA;
        /* include code to process B objects
      }
    else {
        C aC = (C) anA;
        /* include code to process C objects
      }
}
some may even introduce separate methods to actually process B and C. In addition, I typically introduce a method on A, B and C such as the following: 
public abstract void process (T aT);
Classes B and C implement this abstract method to call the two methods on T that deal with processing Bs and Cs. The implementation of the process (A anA) method on the T class then simply calls the process method on A, passing this as an argument. In doing that, I am actually using polomorphism or Java's method dispatcher to implement the if-then-else. The primary advantage of the above technique is that when new subclasses of A are introduced there is no need to modify the process method (maybe the if-then-else appears in more locations that in the process method), al we need to do is add a new process method on the T class that accepts the new subclass of A. 
A disadvantage is that I am actually contaminating A.
I know that most of you will know the above technique but that was just as an introduction. While I was looking into AOP (finally), I will come as no surprise that AOP seemed like an excellent way to implement the double dispatching technique without contaminating A. HERE IS MY ACTUAL QUESTION: I succeeded in implementing it using AOP but I would like some confirmation of more experience guys like you all that there is no better way to do it in AOP. You will find my code attached in ddusingaop.zip Also, I discovered that there is a performance penalty for AOP. The AOPPerformanceTester class prints out the following: Testing display of points using immediate method dispatching:10 

Testing display of points using double dispatching:20

Testing display of graphical elements using instanceof and casting:30

Testing display of graphical elements using AOP:110
The first 3 are in line with my expectations, but the last one is a bit of a surprise. I looked at the generated code (using a decompiler because I could not find an obvious way to look at the generated .java file) and I found that the display method (which is post-processed by the aspect) is actually called twice. Is this related to my implementation or is this a problem of ajc? 

Many thanks for having a look at my question,

Regards,

Rudi.

Back to the top