Description Andrew Clement 2008-06-05 16:08:11 EDT

Came in via an AspectJ bug - but seems to be a base JDT compiler issue (seen in 3.3 and 3.4rc1)

AspectJ Bug: https://bugs.eclipse.org/bugs/show_bug.cgi?id=235829

Two test files. Looks a bit contrived but it is from a real application:

a/Adapter.java
---
package a;

public interface Adapter<T> {
  interface Setter<V> {}

  public <V> Setter<V> makeSetter();
}
---


a/b/Adapter.java
---
package a.b;

public class Adapter<T> implements a.Adapter<T> {

  public <V> Adapter.Setter<V> makeSetter() {
    return new Adapter.Setter<V>() {};
  }

}
---

The signature for the type 'a.b.Adapter$1' is generated as:

Ljava/lang/Object;La/b/Adapter$Setter<TV>;

when it should be:

Ljava/lang/Object;La/Adapter$Setter<TV>; 

(which is what javac produces).

The bug was not in Eclipse 3.1 and has appeared due to the changes in 

ParameterizedTypeBinding.genericTypeSignature()

But I'm not convinced the bug is in there, it may have just surfaced because of changes in there.

Within that method the enclosingType() of Adapter.Setter is a.b.Adapter, rather than a.Adapter (where Setter is actually declared).  And so when it is used to generate the signature for Adapter.Setter, it gets it wrong.  Previously (in Eclipse 3.1) the enclosingType was only used to generate the signature if it was parameterized (not the case here).

If the references in the second test file are all changed to explicitly be a.Adapter then it produces the correct signature.

It looks to be an issue in ParameterizedQualifiedTypeReference when resolving the references to Adapter.Setter.  Proceeding through the internalResolveType() method in there we determine that qualifiedType is a.b.Adapter (the closest one) - then when we go looking for the member type Setter, we dont find it on a.b.Adapter, proceed up the hierarchy and find it on a.Adapter.  We then get back into internalResolveType() and reach this code:

ReferenceBinding currentType = (ReferenceBinding) this.resolvedType;
if (qualifiedType == null) {
    qualifiedType = currentType.enclosingType(); // if member type

Here the currentType is the right member type Setter but the qualifiedType reference is set incorrectly to the wrong containing type for the memberType. (maybe it isn't incorrect, but its not a valid qualifier for that member type)

Because qualifiedType isn't fixed up, we continue down further and call this:

ParameterizedTypeBinding parameterizedType = scope.environment().createParameterizedType((ReferenceBinding)currentType.erasure(), argTypes, qualifiedType);

where the qualifiedType is used incorrectly as the enclosing type for the currentType.  And so the incorrect ParameterizedType a.b.Adapter.Setter is created and eventually its' signature gets used in the output class file.

I changed:

ReferenceBinding currentType = (ReferenceBinding) this.resolvedType;
if (qualifiedType == null) {
    qualifiedType = currentType.enclosingType(); // if member type

to

ReferenceBinding currentType = (ReferenceBinding) this.resolvedType;
if (qualifiedType == null || (currentType.isMemberType() && qualifiedType!=currentType.enclosingType())) {
    qualifiedType = currentType.enclosingType(); // if member type

adding an extra condition to allow for the member type not quite being found where we first guessed.	

This may or may not be an appropriate fix, but it fixes the testcase and passes all the other generics tests I have (not many...)