org.aspectj/docs/progGuideDB/pitfalls.xml

<chapter id="pitfalls" xreflabel="Pitfalls">
  <title>Pitfalls</title>

  <sect1 id="pitfalls-intro">
    <title>Introduction</title>

    <para>
      This chapter consists of a few AspectJ programs that may lead to
      surprising behavior and how to understand them.
    </para>

  </sect1>

  <sect1 id="pitfalls-infiniteLoops">
    <title>Infinite loops</title>

    <para>
      Here is a Java program with peculiar behavior
    </para>

<programlisting><![CDATA[
public class Main {
    public static void main(String[] args) {
        foo();
        System.out.println("done with call to foo");
    }

    static void foo() {
        try {
            foo();
        } finally {
            foo();
        }
    }
}
]]></programlisting>

    <para>
      This program will never reach the println call, but when it aborts
      may have no stack trace. 
    </para>

    <para>
      This silence is caused by multiple StackOverflowExceptions.  First
      the infinite loop in the body of the method generates one, which the
      finally clause tries to handle.  But this finally clause also
      generates an infinite loop which the current JVMs can't handle
      gracefully leading to the completely silent abort.
    </para>

    <para> 
      The following short aspect will also generate this behavior:
    </para>

<programlisting><![CDATA[
aspect A {
    before(): call(* *(..)) { System.out.println("before"); }
    after():  call(* *(..)) { System.out.println("after"); }
}
]]></programlisting>

    <para>
      Why?  Because the call to println is also a call matched by the
      pointcut <literal>call (* *(..))</literal>. We get no output because
      we used simple after() advice.  If the aspect were changed to
    </para>

<programlisting><![CDATA[
aspect A {
    before(): call(* *(..)) { System.out.println("before"); }
    after() returning:  call(* *(..)) { System.out.println("after"); }
}
]]></programlisting>

    <para>
      Then at least a StackOverflowException with a stack trace would be
      seen.  In both cases, though, the overall problem is advice applying
      within its own body.
    </para> 

    <para>
      There's a simple idiom to use if you ever have a worry that your
      advice might apply in this way.  Just restrict the advice from occurring in
      join points caused within the aspect.  So: 
    </para>

<programlisting><![CDATA[
aspect A {
    before(): call(* *(..)) && !within(A) { System.out.println("before"); }
    after() returning:  call(* *(..)) && !within(A) { System.out.println("after"); }
}
]]></programlisting>

    <para>
      Other solutions might be to more closely restrict the pointcut in
      other ways, for example:  
    </para>

<programlisting><![CDATA[
aspect A {
    before(): call(* MyObject.*(..))  { System.out.println("before"); }
    after() returning:  call(* MyObject.*(..))  { System.out.println("after"); }
}
]]></programlisting>

    <para>
      The moral of the story is that unrestricted generic pointcuts can
      pick out more join points than intended. 
    </para>

  </sect1>
</chapter>