© Copyright 1998-2002 Palo Alto Research Center Incorporated, 2003-2004 Contributors. All rights reserved.

AspectJ Porting Notes

Porting pre-1.2 code to AspectJ 1.2

README-12.html contains a discussion of the changes between 1.1 and 1.2. The key points are:

The default AspectJ compiler compliance level is now 1.4 (whereas in previous releases the default compliance level was 1.3). This has a number of implications:

The AspectJ compiler can be restored to 1.3 compliance settings by specifying the "-1.3" option on the command-line.

The following example program illustrates the differences in join point matching with the call pointcut designator between 1.4 and 1.3 compliance levels.


01 class A {
02   public void doIt() {...};
03 }
04
05 class B extends A {
06   public void doThisToo() {...};
07 }
08
09
10 public class CallsAandB {
11 
12  public static void main(String[] args) {
13    B b = new B();
14    A bInDisguise = new B();
15   
16    b.doIt();               // AspectJ 1.2 matches here
17    bInDisguise.doIt();     // this is never matched
18  }
19
20 }
21
22 aspect CallPCDMatchingExample {
23
24   before() : call(* B.doIt(..)) {
25     System.out.println("About to call B.doIt(...)");
26   }
27
28 }

When this program is compiled with AspectJ 1.2 using the default compiler options, it will produce one line of output when it is executed:

About to call B.doIt(...)

The same program compiled under AspectJ 1.1 (or using AspectJ 1.2 with the -1.3 flag specified) does not produce any output when it is run.

The reason for the additional call pcd match is that prior to compliance level 1.4, Java compilers produced bytecodes that call A.doIt() (the defining type of the method), rather than B.doIt() (the declared type in the program text). The generated call to A.doIt() is not matched by the call pcd used in the before advice. At compliance level 1.4, the bytecodes retain the declared type of the receiver in the program source, generating a call to B.doIt(), which is matched by the call pcd.

This is a good example of why the recommended style is to use call(* doIt(..)) && target(B), which always matches based on the actual type of the receiver.

New warnings emitted by the compiler for unmatched call pcds. Because users have found the static type matching used for a type pattern specified in a call pcd confusing (as evidenced by the example above), AspectJ 1.2 has a new Xlint warning which is enable by default. The compiler will now produce a warning whenever a call pointcut designator does not match at a join point, and a user may have expected it to. Compiling the above program using AspectJ 1.2 produces the following compiler output:



CallsAandB.java:24 warning does not match because declaring type is A, if match desired use target(B) [Xlint:unmatchedSuperTypeInCall]
before() : call(* B.doIt(..)) {
           ^^^^^^^^^^^^^^^
	
	see also: CallsAandB.java:17


1 warning


The warning is telling us that the call pointcut associated with the before advice on line 24 of the source file does not match at a join point where the user may have expected it to. The source location corresponding to the unmatched join point is indicated by the "see also" line - in this case line 17 of the source file. At line 17 we find a call to bInDisguise.doIt(). Since the static type of bInDisguise is A, this call will never be matched. The warning also tells us a possible solution if we intended the pointcut to match at this join point: use call(* doIt(..) && target(B).

If you find warnings of this kind coming out when you use the AspectJ 1.2 compiler, the recommended fix is to switch to using the target designator in place of a type pattern in the call pointcut expression. Note that there is no loss of runtime efficiency here - runtime tests are only added in the cases where it cannot be determined at compile time whether the type of the receiver will match the type specified in the target expression. Note that target cannot be used in declare statements.

Use of non-statically determinable pointcut expressions in declare statements has always been forbidden, but prior to 1.2 the AspectJ compiler did not raise an error if they were used. The AspectJ Language Semantics appendix states that cflow, cflowbelow, this, target, args and if pointcut designators cannot be used directly or indirectly (through a user-defined pointcut) inside of a declare statment. When moving code from 1.1 to 1.2, additional errors may be raised due to the stricter policing of this rule. The solution is to recode the declare statement avoiding pointcut expressions that may require a run-time test.

Interface constructors no longer supported. Declaring a constructor on an interface is now (correctly) prohibited, and there will no longer be a constructor-execution join point for the interface. To initialize a field declared on an interface, use initialization, e.g.,

int I.i;
after(I i) returning: initialization(I) && this(i) { i.i = 2; }

To pick out the constructor-execution for any implementation of I, try

execution(I+.new(..))

For more information, see bug 49295.

Declaring a static method on an interface is now (correctly) prohibited. One workaround is to define a static method on the aspect instead. For more information, see bug 47754.

Watch for problems due to incompatible BCEL versions. AspectJ 1.2 includes a different version of BCEL than AspectJ 1.1. If you have the older version of BCEL available earlier on your classpath than the version included in the 1.2 aspectjtools.jar then you will see errors like:

C:\work\test\TestAspect.aj error Internal compiler error
java.lang.NoSuchMethodError: org.apache.bcel.generic.InstructionFactory.
createNewArray(Lorg/apache/bcel/generic/Type;S)Lorg/apache/bcel/generic/Instruction;
This typically happens because the old version of BCEL has been included as a standard extension in your JVM configuration. Ensure you have removed it from jre/lib/ext under your JDK installation.

For more information, see bugs including 60389, 59921.

Porting pre-1.1 code to AspectJ 1.1

README-11.html contains a discussion of the language changes from 1.0 to 1.1. The high points:

The call(..) pointcut designator is now implemented only at the call site; by contrast, the AspectJ 1.0 compiler could also implement it on the callee side. So in 1.0 if you compiled a pointcut using call(..) but only passed the compiler the code for the target of the call, the pointcut could be implemented. This is not true for 1.1. To fix this, use execution(..) in place of call(..), or include all calling clients in the compile. (more info)

Type-patterns are no longer permitted for the defining type of inter-type declarations. Replace the pattern with a type. In many cases, you can declare members on an interface type, and then declare that the types picked out by the type-pattern implement have the interface as their parent. (more info)

Type-patterns are no longer permitted when specifying declare soft. Replace the pattern with a literal type.

Wildcards patterns (foo..*) are no longer permitted for this(), target(), or args(). Replace the pattern with a literal type or with a subtype wildcard (Type+). (more info)

Conflicts will be reported for no-argument constructors generated by compilers when no constructor is defined for a class. That means the following code will compile in 1.0 but not in 1.1:

class C {}
aspect A {
   C.new() {}  // permitted in 1.0; conflict in 1.1
}
One fix is to declare a non-conflicting constructor by adding arguments (or defining a constructor in the target class); a better fix might be to do the work of the declared constructor in advice on the initialization join point for the object. (more info)

The pointcut designators within() and withincode() will not pick out code within the lexical extent of method-local and anonymous inner types (because these are not represented as such in bytecode form). Because within forms specify staticly-determinable pointcuts, they might be used in declare error or declare warning statements, which might produce different results. (more info)

The compiler will report an error that the form aspect {name} dominates {list}... is no longer supported. It has been replaced by a new declare statement:

declare precedence : {name} {list}...
(more info)

The field set join point now has a return type of void. Compiling programs using around advice on these join points might cause errors unless the return type of the around advice and the result of any proceed() call is Object or void. (more info)

The compiler cannot implement after or around advice for the handler PCD because the end of exception handlers is ambiguous in bytecode. Try to use before advice. (more info)

Porting pre-1.0.4 code

In versions of AspectJ prior to 1.0.4, the compiler was not correctly implementing the AspectJ-1.0 language design for some uses of after returning advice.

The main change that was made was of after returning advice for constructor execution join points. Previously, this advice was legal:

after() returning (Foo f): execution(Foo.new(..)) { ... }

However, it has always been a part of the 1.0 language design (and of Java's language design) that constructors themselves (as opposed to constructor calls) do not return the value of the new object. Rather, this is bound to the new object, and the constructor behaves like a void method. With that in mind, any code like the above should be conveted to the form.

after(Foo f) returning: this(f) && execution(Foo.new(..)) { ... }

In compilers prior to 1.0.4, the following advice could pick out join points

after() returning (String s): call(void foo()) { ... }

This is no longer picked out. This pattern was most commonly used in highly polymorphic contexts, such as

after() returning (String s): call(* foo()) { ... }

If you want to capture all calls, binding null objects for those that would otherwise have no value, you must use the Object type.

after() returning (Object o): call(* foo()) { ... }

Uses of both of these forms are highleted with compiler warnings in the 1.0.4 compiler.


Porting pre-1.0rc1 code

Aspects can no longer be declared to implement the Serializable or Cloneable interfaces. If you previously used serializable or cloneable aspects, you should refactor your code to keep the state you need to serialize or clone in objects associated with the aspects.


Porting pre-1.0beta1 code

The static modifier is no longer allowed on pointcut declarations anywhere. Porting is simple; just remove the static declarations when you find them.

Also, though the returns modifier on pointcuts has not been part of the language since 1.0alpha1, the compiler still accepted them until now. If you used this feature, now is the right time to remove the returns modifier when the compiler complains about it.


Porting pre-1.0alpha1 code

The release of AspectJ 1.0alpha1 involved sweeping cleanups of the language to bring it to 1.0 status.

Pointcuts

Removing the "s" from pointcuts

One of the most pervasive changes in porting code written before 1.0alpha1 is the change in some of the pointcut names from plural to singular, that is, they lose an "s". In one sense, making this change in your programs is easy: just go through and whever you see uses of the pointcuts

calls executions gets sets handlers initializations staticinitializations

Just take off the final "s", to make one of

call execution get set handler initialization staticinitialization

Often, there will be other changes you should make for each of these pointcuts, but as for the name, just take off the "s".

One risk you will have when doing this is creating name conflicts. If, for example, you named a parameter of a pointcut "set", you should (for your own sanity -- the compiler doesn't require it) rename it in the rewritten pointcut.

pointcut sort(Collection set): calls(void addAll(set));
==>
pointcut sort(Collection mySet): call(void addAll(mySet));

While converting to use singular nouns for the primitive pointcuts, you may also want to remove the "s" from your user-defined pointcuts.

pointcut publicCalls(): calls(public * *(..));
==>
pointcut publicCall(): call(public * *(..));

Of course, your naming conventions are your own, but throughout these porting notes we will be making these changes in our example ports.

Removing the receptions pointcut

Perhaps the largest semantic change in the 1.0 language is the removal of receptions join points. They have been merged with call join points in AspectJ 1.0, so now a call join point doesn't represent the "caller-side" of a call, but the call itself, both caller and receiver.

Changing code that used the receptions pointcut should be fairly straightforward, depending on whether the pointcut exposed state or not.

Not exposing state

Receptions pointcuts that did not expose state can simply be replaced by the new call and target pointcuts:

receptions(void Foo.m())
==>
target(Foo) && call(void m())
Exposing state

Some receptions pointcuts exposed the receiving object by replacing the receiving type with a pointcut formal. These PCDs should be rewritten to use the new target pointcut to expose the receiving object.

pointcut fooCallees(Foo f): receptions(void f.m());
==>
pointcut fooCallee(Foo f): target(f) && call(void m());

Like other pointcuts, receptions pointcuts that exposed one or more arguments should be rewritten to use the args pointcut:

pointcut intPassers(int i, int j): receptions(void Foo.m(i, j));
==> 
pointcut intPasser(int i, int j):
    args(i, j) && target(Foo) && call(void m(int, int));
Constructor receptions

There are two issues with constructor receptions in particular.

Like constructor calls, constructor receptions pointcuts had a dynamic character, in that receptions(C.new()) would capture constructions of not only C classes, but also of classes that extended C.

If you want this behaviour, then you need to use the new subtypes operator, +, on the type name in question. So,

receptions(C.new())  
==>
call(C+.new())

Also like constructor calls, constructor receptions allowed access to the constructed object in the same way as any other object. Since the only advice possible on constructor receptions join points was after returning advice, the object was always guaranteed to be there. But since constructor call join points allow all kinds of advice it may be that the object isn't constructed yet (say, in before or around advice). This is a benefit, in that it allows caching constructed objects

aspect Singleton {
    private C theC = null;

    C around(): call(C.new(..)) {
        if (c == null) theC = proceed();
        return theC;
    }
}

but it does require some rewriting. The new object can be accessed as the return value in after returning advice. So,

after(Point p) returning (): receptions(p.new(int, int)) { ... }
==>
after() returning (Point p): call(Point+.new(int, int)) { ... }

Fixing state access

In previous versions of AspectJ, state such as the currently executing object or a particular argument of a method call could be accessed from the signatures of many pointcuts, leading to difficult-to-read forms. In AspectJ 1.0, all state accesses now use only three pointcuts

args this target

which pick out argument values, the currently executing object, and the target object of a method call or field operation, respectively.

Using args

Any time you have a pointcut that has a signature where one of the arguments was a pointcut or advice formal, just replace that formal with its type and add an args pointcut.

pointcut intPassers(int i, int j): calls(void Foo.m(i, j));
==>
pointcut intPasser(int i, int j): args(i, j) && call(void Foo.m(int, int));
pointcut stringPassers(String s): receptions(void Foo.m(s, ..));
==>
pointcut stringPasser(String s): args(s, ..) && call(void Foo.m(String, ..));
Rewriting calls

If a calls pointcut exposed the the receiving object, such as

pointcut fooCallees(Foo f): calls(void f.m());

then the new version should use the target pointcut to get at that object

pointcut fooCallee(Foo f): target(f) && call(void Foo.m());

AspectJ's calls pointcut previously allowed the new object to be exposed, even though it may not have been constructed yet. AspectJ 1.0 no longer allows this; you can access the new instance only in after returning advice, when it is guaranteed that the object was successfully constructed. So instead of using the target pointcut to expose the value, you should use the normal after returning mechanism:

after(Point p) returning (): calls(p.new(int, int)) { ... }
==>
after() returning (Point p): call(Point+.new(int, int)) { ... }
Rewriting gets and sets

Exposing the target object of a gets or sets pointcut should be done the same way it was for calls pointcuts, with the new target pointcut.

before(Frame f): gets(Color f.color) { ... }
==>
before(Frame f): target(f) && get(Color Frame.color) { ... }
before(Frame f): sets(Color f.color) { ... }
==>
before(Frame f): target(f) && set(Color Frame.color) { ... }

In addition, the clumsy syntax for getting the old value of the field has been eliminated. For before advice, the port is simple; just access the field yourself in the body. Depending on the rest of your system, you may need to restrict the advice from the aspect body to eliminiate the circularity.

aspect A {
    before(Frame f, Color c): gets(Color f.color)[c] { ... }
}
==>
aspect A {
    before(Frame f):
            target(f) && get(Color Frame.color) && !within(A) {
        Color c = f.color;
        ...
    }
}

The same can be done for around advice. However, the only way to port after advice that needs the old value is to convert it to around advice.

aspect A {
    after(Frame f, Color c) returning (): gets(Color f.color)[c] { ... }
}
==>
aspect A {
    void around(Frame f):
            target(f) && get(Color Frame.color) && !within(A) {
        Color c = f.color;
        proceed(f);
        ...
    }
}

When porting sets pointcuts, the new value of a field is still available, but not the way it was previously. Instead of using the square bracket syntax, we use an args pointcut. All set join points are assumed to have exactly one argument, which holds the new value. So,

after(Color newColor): sets(Color Frame.color)[][newColor] { ... }
==>
after(Color newColor): args(newColor) && set(Color Frame.color) { ... }

Also, if the field was declared private, in order to get at its old value the aspect must be declared privileged.

Rewriting handlers

The value of the exception at an exception handler join point is now accessed through the args pointcut; all exception handler join points are treated as having exactly one argument, the exception value. So,

before(NotFoundException e): handlers(e) { ... }
==> 
before(NotFoundException e): args(e) && handler(NotFoundException) { ... }
Rewriting within

The within pointcut was not typically used to export context. Though it was accidentally possible to do so in versions of AspectJ before 1.0, it often didn't do what users expected it to. This loophole has now been closed, and within can only take type patterns, not pointcut or advice formals. A use of the this pointcut will capture what previous implementations did:

pointcut usesFoo(Foo f): within(f);
==>
pointcut usesFoo(Foo f): this(f) && within(Foo);

Understanding signatures

Now that we have this, target, and args pointcuts, all of our signatures are composed of just types, names, and wildcards; there are no more parameters.

Also, now that we have the + wildcard to pick out subtypes, we can make signature matching much more uniform.

Previously, some signatures matched based on subtypes, some based on instanceof, and some exactly. Now, we have made all signatures match exactly.

What does this mean for your program? Well, it means that you may have to add + to some of your signatures, depending on what you meant them to match.

For example, the pointcut

calls(void m(Object))

previously picked out all method calls to a method named m that took one argument, which was a subtype of Object. Now, however, it will only pick out method calls to methods that are defined to take exactly the type Object, which may be a lot fewer join points. If you want the old behaviour, simply convert to

call(void m(Object+))

Removing the instanceof pointcut

The intanceof pointcut has been split into two different pointcuts, this and target.

Typically, the instanceof pointcut would only exist in a compound pointcut, composed (with &&) with another pointcut. If the other pointcut was a receptions pointcut, then instanceof should be converted to target (and receptions converted to call). So,

pointcut stateChanges(Subject s): 
    instanceof(s) && receptions(void Button.click());
==>
pointcut stateChange(Subject s): 
    target(s) && call(void Button.click());

In all other cases, instanceof referred to the currently executing object, and so should be converted into this

before(Point p): instanceof(p) && executions(* makePolar(..)) { ... }
==>
before(Point p): this(p) && execution(* makePolar(..)) { ... }
pointcut setup(Client c): instanceof(c) && calls(Remote Naming.lookup(String));
==>
pointcut setup(Client c): this(c) && calls(Remote Naming.lookup(String));

Rewriting the initializations pointcut

Object initialization join points are now more complicated, and more true to Java's execution model. Now they bracket all of the initialization that a class can do, after the return of its super constructor call (before which no initialization can happen). Previous versions of AspectJ had object initialization join points that only included initialization that was made in dynamic initializers and fields.

The old behaviour can be recovered with a simple rewrite.

initializations(A)
==>
initialization(A.new(..)) && !execution(A.new(..))

Understanding constructor calls

Previously, constructor call join points were matched by subtypes, so calls(Foo.new()) would match both calls to create new Foo objects, and new SubFoo objects. The new call pointcut designator matches types exactly, so if you want the old behaviour, you should write call(Foo+.new()).

Similarly, constructor execution join points were matched by subtypes. So the old executions(Foo.new()) is now represented by execution(Foo+.new()).

In both of these cases, think before using the + operator; it may be that you didn't intend subtype matching in the first place.

Removing the hasaspect pointcut

The hasaspect pointcut is no longer defined, but you can get the same behaviour using the new if pointcut.

If the aspect whose presense you are checking for was defined of eachcflow, of eachcflowbelow, or, more unlikely, of eachJVM(), then the conversion is simple:

hasaspect(A)
==>
if(A.hasAspect())

If the aspect was defined of eachobject, then you will have to expose the current object in your pointcut or advice parameters:

pointcut cut(): hasaspect(A) ... ;
==>
pointcut cut(Object o): this(o) && if(A.hasAspect(o)) ... ;
or
pointcut cut(Object o): target(o) && if(A.hasAspect(o)) ... ;

If you were using the hasaspect pointcut to expose the state of the aspect, then you can get the same state by using A.aspectOf() in the body of the advice. For example, if the aspect A were defined of eachcflow, then

before(A myA): hasaspect(myA) {
    myA.checkStatus();
}
==>
before(): if(A.hasAspect()) {
    A myA = A.aspectOf();
    myA.checkStatus();
}

Removing the withinall pointcut

The withinall poinctut is no longer defined. You can use a combination of within and the new subtypes operator, +, instead. You'll save two characters and be using a simpler and more orthogonal language.

withinall(Foo)
==>
within(Foo+)

Removing returns modifier from pointcuts

The returns keyword is no longer necessary for user-defined pointcuts. Simply remove it when you find it.

pointcut publicIntCalls() returns int: calls(public int *(..));
==>
pointcut publicIntCall(): call(public int *(..));

Making some pointcuts static

In Java, only static members may be accessed by their declaring type name, like the static method Math.max() can be accessed.

Pointcuts now have that property too. Pointcuts may be declared to be static, in which case they can be accessed like MyAspect.move(), or they can be left non-static, in which case they can be overridden by a subaspect.

In addition, while pointcuts can still be defined in classes, only static pointcuts can be defined in classes.

Porting should be straightforward; just make all your pointcuts in classes static, and make any pointcut with a qualified reference static.

Type patterns

Understanding * and .. in type patterns

Previous versions of AspectJ treated * and .. too cleverly in type patterns, placing restrictions based on what is a package and what is a type, and basing their meanings on the definition of a package hierarchy.

In AspectJ 1.0, both of these wildcards are defined simply, and textually:

That's it.

This change won't affect most programs, but it will make understanding programs easier. There is one ugly idiom, however, that this change disposes of. If your program includes the type pattern *..*, which used to match all types, you can replace it with the much simpler *.

pointcut unaryVoidMethods(): call(void *(*..*));
==>
pointcut unaryVoidMethod(): call(void *(*));

Fixing subtypes in introduction

The new + operator is used to normalize the many places you want to use subtypes of some types.

In introduction forms, you will need to replace subtypes(TypePattern) type patterns with the new subtype operator, +. In the case where you wrote subtypes(Foo), i.e., the subtypes of a single type, simply replace this with Foo+. Otherwise, use the + operator as appropriate in TypePattern.

public void (subtypes(Target0 || Target1)).accept(Visitor v) {
    v.visit(this);
}
==>
public void (Target0+ || Target1+).accept(Visitor v) {
    v.visit(this);
}

Advice

Moving the return type of around

The returns keyword is no longer used for around advice. Instead, the return type is declared as it is for methods. So,

around(Point p) returns void: setters(p) { ... }
==>
void around(Point p): setter(p) { ... }

Adding a throws clause to around

Around advice must now declare the checked exceptions it throws with a throws clause, much like a method.

char around(char c) throws java.io.CharConversionException: converter(c) {
    char result;
    try { result = proceed(); }
    catch (Exception e) {
        throw new java.io.CharConversionException();
    }
    if (result == 0) throw new java.io.CharConversionException();
    return result;
}

Understanding advice precedence

In previous versions of AspectJ, advice precedence within an aspect was simple: if a piece of advice appeared before another piece, it was more precedent. This made perfect sense for before and around advice, but was the cause of confusion (even among the AspectJ designers, more than once) for after advice, as it seemed backward.

In addition, advice was ordered by kind, in that around advice always surrounded before and after advice.

AspectJ 1.0 has changed this; precedence for after advice is inverted, and advice is no longer ordered by kind.

This won't matter to you unless you write pieces of advice in the same aspect that apply to the same join point.

If you do, here's what to think about: If you're looking at two pieces of advice and want to know which has precedence, if either is after advice, then the second one has precedence. Otherwise, the first does.

This allows interesting advice interaction. In the following advice, for example, the after throwing advice will catch the exception thrown by the before advice

aspect A {
    before(): call(void main(..)) {
        throw new RuntimeException();
    }
    after() throwing(RuntimeException e): call(void main(..)) {
         System.err.println("caught you!");
    }
}

But reversing the order will give the before advice more precedence, making its exception uncatchable by the after throwing advice

aspect A {
    after() throwing(RuntimeException e): call(void main(..)) {
         System.err.println("missed you!");
    }
    before(): call(void main(..)) {
        throw new RuntimeException();
    }
}

Advice in different aspects is ordered by the normal aspect precedence rules of subtyping and the dominates modifier.

Fixing after returning

If you use after returning advice and do not need to expose the return value, you no longer need to write an empty set of parentheses to indicate that fact. So,

after(Formals) returning (): Pointcut { ... }
==>
after(Formals) returning: Pointcut { ... }

The same syntax is now available for after throwing advice, in case you do not care what Throwable is thrown.

after(Formals) throwing: Pointcut { ... }

Renaming thisStaticJoinPoint

thisStaticJoinPoint has been renamed thisJoinPointStaticPart, to reflect that it is now exactly the static part of thisJoinPoint: It will return the same object as thisJoinPoint.getStaticPart().

Converting access to thisJoinPoint

The JoinPoint object hierarchy has been folded into a single class, org.aspectj.lang.JoinPoint. A common pattern in logging, for example, was

before() executions(* myMethod()) {
    ExecutionJoinPoint jp = (ExecutionJoinPoint)thisJoinPoint;
    CodeSignature jp = (CodeSignature)jp.getSignature();
    System.err.println(jp.getParameters());
    System.err.println(jp.getParameterNames());
}    

While there is still a rich hierarchy for signatures, there is only one JoinPoint type, so this can be rewritten as:

before() executions(* myMethod()) {
    JoinPoint jp = thisJoinPoint;
    CodeSignature jp = (CodeSignature)jp.getSignature();
    System.err.println(jp.getArgs());
    System.err.println(jp.getParameterNames());
}    

Some of the method names of JoinPoint have been reorganized, as well.

Introduction and static crosscutting

Removing +implements and +extends

The keywords +implements and +extends no longer exist. Instead, AspectJ uses the declare form for exactly the same functionality.

Point +implements Serializable;
=> 
declare parents: Point implements Serializable;
MyButton +extends ButtonAdaptor;
=> 
declare parents: MyButton extends ButtonAdaptor;

Using declare soft

Around advice advice no longer effects the static exception checking of Java. This means that the following code previously compiled:

class C {
    void noExceptionDeclared() {
        exceptionDeclared();
    }
    void exceptionDeclared() throws IOException {}
}
aspect A {
    around(): call(void C.exceptionDeclared()) {
        try { proceed(); }
        catch (IOException e) {}
    }
}

even though the class C is not compilable on its own (because noExceptionDeclared actually throws an Exception).

AspectJ now firmly places everything that affects the type system of Java, including the declared-exception checking system, into the space of introduction and declare. So, in order to state that the call to exceptionDeclared() will not, actually, throw an exception, we now "soften" that exception, that is, take it out of the space of declared exceptions.

declare soft: ExceptionType: Pointcut;

The pointcuts allowed here are limited; you cannot use pointcuts that would require runtime information. But picking out method calls is just fine. So in order to make the above example work, one new declaration is needed:

declare soft: IOException:
    call(void C.exceptionDeclared()) &&
    withincode(void noExceptionDeclared());

Aspects

The syntax of "of each" modifiers has changed. For of eachcflow and of eachcflowbelow, you can simply replace "of each" with "per". So,

aspect A of eachcflow(...) { ... }
==>
aspect A percflow(...) { ... }

If you have any aspects defined of eachJVM(), then you should either remove that declaration entirely (because this is the default behaviour), or replace the of eachJVM() declaration with an issingleton declaration.

aspect of eachJVM() { ... }
==>
aspect A { ... }
or
aspect A issingleton { ... }

The of eachobject(Pointcut) modifier has been split into two different forms, of perthis(Pointcut) and of pertarget(Pointcut). Which one you replace with depends on the Pointcut you use.

If you use a pointcut that picked out reception join points, then use pertarget, and rewrite the pointcut to pick out call join points. So

aspect Shadow
        of eachobject(receptions(void Point.setX(int)) ||
                      receptions(void Point.setY(int))) {
    ...
}
==>
aspect Shadow pertarget(call(void Point.setX(int)) ||
                        call(void Point.setY(int))) {
    ...
}

Otherwise, in most cases, use perthis. When you convert, remember the meaning of each of these modifiers. perthis(Pointcut) indicates that an instance of the aspect should be associated with every object that is this at each of the join points picked out by Pointcut, while pertarget(Pointcut) associates with every object that is the target object at such join points.


Porting pre-0.8beta3 code

The following changes are only required when porting code written prior to the 0.8beta3 release of AspectJ.

Changing cflow terminology

Changing pre-0.8beta3 code that uses AspectJ's control-flow-based features only requires rewriting occurrences of eachcflowroot, cflow, and cflowtop. No editing of other aspect code is necessary.

eachcflowroot

The aspect modifier "of eachcflowroot(Pointcut)" should now be written more as "percflow(Pointcut)".

cflow

In previous versions of AspectJ, the pointcut cflow(Pointcut) picked out all join points in the cflow below the join points of Pointcut. That is, it did not include the join points of Pointcut, only the join points in their control flow.

As of version 0.8beta3, cflowbelow(Pointcut) has that behavior. cflow(Pointcut) includes the join points of Pointcut.

In many cases, you may not care whether the points of Pointcut are included or not, and so can safely leave cflow(Pointcut) pointcut designators alone. However, if you use the idiom

Pointcut && ! cflow(Pointcut)

to capture the non-recursive entries to a particular pointcut, you will definitely want to rewrite that as

Pointcut && ! cflowbelow(Pointcut)

cflowtop

The primitive pointcut designator cflowtop(Pointcut) has been removed from the language, as it is expressible with cflow or cflowbelow. All uses of cflowtop(Pointcut) can be rewritten as:

cflowbelow(Pointcut && ! cflowbelow(Pointcut))

Though in most cases the following is sufficient

cflow(Pointcut && ! cflowbelow(Pointcut))

Overriding abstract pointcuts

In previous versions of AspectJ, a concrete aspect would implicitly override all of its abstract pointcuts with an empty pointcut. AspectJ 0.8beta3 enforces the restriction that a concrete aspect may not have any abstract pointcuts. Thus the following extension:

abstract aspect A {
    abstract pointcut pc();
}

aspect B {}

will no longer compile.

Adding the new empty pointcut designator

pointcut Id();

in the declaration of the concrete aspect fixes this problem.

abstract aspect A {
    abstract pointcut pc();
}

aspect B {
    pointcut pc();
}

Limiting recursive advice

Previously, the compiler silently refrained from applying a piece of advice to join points within its own advice body. So, for example, in

class C {
    static int i;
}

aspect A {
    before(): gets(int C.i) {
        System.err.println("C.i was " + C.i)
    }
}

The advice would trace all references of the static field C.i except those in the body of the before.

The compiler has now removed this special case, and so running the above example will now cause a StackOverflowException to be thrown.

Most cases of this error can be fixed by correctly specifying the desired pointcut: In the above example, the intention is clearly not to trace all references of C.i, just those outside the aspect.

class C {
    static int i;
}

aspect A {
    before(): get(int C.i) && ! within(A) {
        System.err.println("C.i was " + C.i)
    }
}

In a very few cases, you may want the advice to be applicable to other code in the aspect, but not in the particular piece of advice. In such cases, you can pull the body of the advice into a method and restrict away from that method (and away from calls to that method):

class C {
    static int i;
}

aspect A {
    public static int getCi() {
        return C.i;                          // will be traced
    }

    before(): get(int C.i) &&
              ! withincode(void A.traceCi())
              ! call(void A.traceCi())      {
        traceCi();
    }
    private void traceCi() {
        System.err.println("C.i was " + C.i) // will not be traced
    }
}

Porting pre-0.8beta1 code

The following changes are only required when porting code written prior to the 0.8beta1 release of AspectJ.

Rewriting introductions

Syntax

The syntax of introduction has changed. Porting most programs should require some simple editing. Anywhere you have an introduction block

introduction GTN {
    ...
}

simply move the GTN down into the introduction declarations and remove the block.

For method introduction, place the GTN in front of the method name, For field introduction, place the GTN in front of the field name, and for constructor introduction, place the GTN in front of the new identifier.

introduction Foo {
    public void doStuff() { this.doStuffLater(); }
    public int calorieCount = 3;
    public new(int x) { super(); calorieCount = x; }
}

==>

public void Foo.doStuff() { this.doStuffLater(); }
public int Foo.calorieCount= 3;
public Foo.new(int x) { super(); calorieCount = x; }

For implements and extends introduction, move the GTN in front of the new identifiers implements or extends, and place that in a declare parents form.

introduction Foo {
    implements Comparable;
    extends Goo;
}

==>

declare parents: Foo implements Comparable;
declare parents: Foo extends Goo;

In all cases, if the GTN is just a type name, it can be moved down on its own. However, if the GTN uses any of &&, ||, and !, it must be parenthesized.

introduction subtypes(Foo) && !Goo {
    int x;
}

==>

int (Foo+ && !Goo).x;

Access

If you had an introduction that was referring to private or protected members of the target class, this will no longer work. You will either need to modify your code to avoid this accessibility issue, or you will need to use the privileged modifier on the aspect that contains the introduction.

class Counter {
    private int count = 2;
}

aspect ExposeCountersPrivates {
    introduction Counter {
        public int getCount() { return count; }
    }
}

==>
// in 0.8, only privileged aspects can expose a class's privates
privileged aspect ExposeCountersPrivates {
    public int Counter.getCount() { return count; }
}

If you have introduced private or package-protected members, you will probably have to re-write some code. Most previous uses of introducing privates can be improved by using private introduction instead.

class C {
}

aspect AddCounter {
    introduction C {
        private int count;
        public int getCount() { return count; }
    }
}

==>
aspect AddCounter {
    private int Counter.count;
    public int Counter.getCount() { return count; }
}

There is one case that we know of where the inability to perform the introduction of private members makes 0.7 code difficult to port to 0.8. If you were using the introduction of a private void writeObject(..) or a private void readObject(..) method to interact with Java's serialization API, you will need to come up with an alternative design. Using some combination of Externalizable, writeReplace(..) and/or readResolve(..) methods should allow you to port your code. If you find this isn't the case, we'd like to hear about it.

If you were introducing either a protected member or a package-private member onto a class in order to override a protected member that was inherited from a superclass, you will have to make this introduction public.

Removing static advice

Static advice has been removed from the language. Now, every piece of advice is non-static, meaning that it will run in the context of an aspect instance.

If you have an aspect that only contains static advice, has no "of" clause or is declared "of eachJVM()", and is not extended by another aspect, simply remove the keyword "static" from all pieces of advice, and make sure the aspect is not defined with the "abstract" modifier.

aspect Tracing {
    static before(): executions(* *(..)) {
        System.out.println("Got Here! " + thisJoinPoint);
    }
}

==>

aspect Tracing {
    before(): execution(* *(..)) {
        System.out.println("Got Here! " + thisJoinPoint);
    }
}

Otherwise, if you have an aspect contains both static and non-static advice, is extended, or is "of eachObject(...)" or "of eachcflowroot(...)", you should group your static advice together and put it in a new aspect, possibly even an inner aspect.

aspect ComplexTracing of eachobject(cflow(executions(void Main.main(..)))) {
    static before(): executions(* *(..)) {
        System.out.println("Got Here! " + thisJoinPoint);
    }
    static after(): executions(* *(..)) {
        System.out.println("Returned! " + thisJoinPoint);
    }

    // some other dynamic advice, fields, etc
}

==>

aspect ComplexTracing of eachobject(cflow(executions(void Main.main(..)))) {
    static aspect AlwaysTracing {
        before(): execution(* *(..)) {
            System.out.println("Got Here! " + thisJoinPoint);
        }
        after(): execution(* *(..)) {
            System.out.println("Returned! " + thisJoinPoint);
        }
    }

    // some other dynamic advice, fields, etc
}

Fixing aspect-aspect inheritance

Aspects can now only extend abstract aspects. This restriction may cause some redesign of aspect hierarchies. You will probably find that for the majority of your code the most serious change this requires is to add an explicit abstract modifier to a super-aspect that was already implicitly abstract.

aspect BaseTracing {
    abstract pointcut traced();
    before(): traced() {
        System.out.println("Got Here! " + thisJoinPoint);
    }
}

==>

// make this abstract aspect explicitly abstract
abstract aspect BaseTracing {
    ...
}

This change has also affected the getAspect static method. Now, getAspect is only defined on non-abstract aspects. Previously, you could call getAspect on an abstract superaspect and (sometimes) get an instance of a subaspect back.

This pattern was used in the Spacewar example in the AspectJ distribution. We had the class hierarchy

  SpaceObject (abstract)
    |- Ship
    |- Bullet
    |- EnergyPellet

And the aspect hierarchy

  SpaceObjectDA (abstract)
    |- ShipDA of eachobject(instanceof(Ship))
    |- BulletDA of eachobject(instanceof(Ship))
    |- EnergyPacketDA of eachobject(instanceof(Ship))

And we would call SpaceObjectDA.getAspect(SpaceObject) to access the aspect associated with a ship, bullet, or energy pellet. This pattern depended on the SpaceObjectDA aspect hierarchy exactly mirroring the SpaceObject hierarchy, and being maintained that way.

A better way to implement this kind of design aspect is to use private introduction, a new feature of AspectJ.

Using private introduction

A common pattern for AspectJ programs that need to associate some state with every object of a particular type has been to use aspects that are defined of eachobject(instanceof(...)). A prime example of this was the BoundPoint aspect of the bean example: which needed to associate each point with a PropertyChangeSupport object.

aspect BoundPoint of eachobject(instanceof(Point)) {

    java.beans.PropertyChangeSupport support = null;

    after() returning(Point p): receptions(p.new(..)){
        support = new PropertyChangeSupport(myPoint);
    }

    around(Point p) returns void: receptions(void p.set*(*)) {
        // code that uses support
    }
}

In the new version of AspectJ, a better way of accomplishing many of these state association is to use privately introduced fields. Instead of creating an aspect instance for every Point object, store the PropertyChagneSupport object in the Point objects themselves.

aspect BoundPoint {
    private PropertyChangeSupport Point.support = new PropertyChangeSupport(this);

    void around(Point p): setters(p) {
        // code that uses p.support
    }
}

Just as in the past, the PropertyChangeSupport object is not accessable to anyone but the aspect, but now less mechanism is needed.

There are times when changing aspects that are defined of eachobject(instanceof(...)) may not be reasonable. If the aspect instance is stored or passed to other methods, then having a real of eachobject(instanceof(...)), now written perthis(this(...)), association may capture the crosscutting concern best.


Porting pre-0.7beta11 code

The following changes are only required when porting code written prior to the 0.7beta11 release of AspectJ.

Removing two-argument calls

In AspectJ 0.7beta11, the two-argument calls primitive pointcut designator was deprecated. Removing these designators will require different cases depending on what the original pointcut did.

Calls to static methods

For pointcuts denoting calls to particular static methods, such as

calls(String, static String valueOf(int)) // deprecated

the transformation is easy. Simply make the desired signature explicit. Instead of catching all calls to any static method that happens to have the signature String valueOf(int), catch calls to that exact method defined in the String class.

call(static String String.valueOf(int))

Pointcuts denoting calls to classes of static methods can also be rewritten with these rules. For example,

calls(my.package.*, static * get*(..)) // deprecated

should now be written

call(static * my.package.*.get*(..))

Calls to non-static methods

Many pointcuts denoting calls to non-static methods can be fixed the same way that those pointcuts denoting calls to static methods are fixed. So,

calls(Thread, int getPriority()) // deprecated

which denotes all calls to nullary int methods named getPriority when the called object is an instance of the Thread type, can almost always be rewritten

call(int Thread.getPriority())

which denotes all calls to the nullary int Thread.getPriority() method.

Expanding the signature picks out slightly different join points than the original two-argument form. This won't matter for most programs, but in some cases the differences may be noticable. In particular, the expanded-signature form only picks out those calls where the called object is statically typed to Thread when its int getPriority() method is called. If you want to capture calls to the int Thread.getPriority() method, regardless of how the called object is statically typed, you shoud use the different translation:

call(int getPriority()) && target(Thread)

This will capture all call join points of methods with signature int Thread.getPriority().

It will also denote any join points if the Thread type does not define (possibly abstractly) some int getPriority() method, though.

Removing advice from Class declarations

The simplest way to remove an advice declaration from a class is to simply define the advice declaration in an inner aspect. So, instead of

class C {
    static before(): executions(C.new()) { ... } // deprecated
}

write

class C {
    static aspect ConstructionProtocol {
        static before(): executions(C.new()) { ... }
    }
}

If your advice doesn't refer to any inner classes or interfaces of C, you can move the inner aspect out of the class entirely.

class C { ... }

aspect ConstructionProtocol {
    static before(): execution(C.new()) { ... }
}

Your code will be clearer if you consider the purpose of each piece of advice when you make this change. It may be that some of the advice naturally belongs to another aspect, perhaps already existing. Or it may be that some pieces of advice in a class are associated to one concern and some to another; in which case more than aspect would be appropriate.


Porting pre-0.7beta10 code

The following changes are only required when porting code written prior to the 0.7beta10 release of AspectJ.

Changing access to thisJoinPoint

In AspectJ 0.7beta10, access to the reflective object thisJoinPoint substantially changed. The two parts of this change were the elimination of the runNext() static method, and the use of an interface hierarchy represent the join point object.

thisJoinPoint.runNext() to proceed()

The elimination of the runNext() static method requires almost no porting work. An automatic replacement of the string

thisJoinPoint.runNext

with the string

proceed

will do the job. However, if any around advice used the identifier "proceed" as a formal parameter or local variable, it must be renamed, and if any aspect used it as a field, then references to the field in around advice should be made explicit (prefixing the reference with the aspect name or "this", depending on whether the field is static or not).

Using thisJoinPoint

While access to reflective information through thisJoinPoint is more powerful and regular through its interface hierarchy, the previous uses must be rewritten. Changing your code will likely require manual editing, but in doing so your code should get simpler and cleaner.

Many existing uses of the fields on join points can be re-written to use one of:

For example:

System.out.println(thisJoinPoint.className + "." +
                   thisJoinPoint.methodName)

can be replaced with

System.out.println(thisJoinPoint)

or

System.out.println(thisJoinPoint.getSignature().toShortString())

with comparable behavior.

Accesses to the parameters field of join points should be changed as follows. A field access like:

thisJoinPoint.parameters

must be changed to:

Accesses to the methodName and className fields of join points that are not suitable for replacement with a toString method, should be changed as follows. Field accesses like:

must be changed to:

Accessses to the parameterNames and parameterTypes fields of join points, that are not suitable for conversion to one of the toString() methods should be changed as follows. Field access like:

must be changed to: