Annotations on return-overridden methods also end up on bridge methods

[Charles Oliver Nutter]

Hey there!

Sorry if this has been asked before, but I've been trying to get JRuby
to compile properly on JDK8 javac and noticed a change.

JDK8 generates bridge methods in cases where a child class overrides a
superclass method with a more specific return type. This may have been
present in JDK7 and lower, I'm not sure.

What does seem to be different is that annotations attached to the
hand-written override are also being attached to the bridge method.

An example from JRuby...

RubyBasicObject defines op_equal:

public class RubyBasicObject ... {
...
    @Override
    public IRubyObject op_equal(ThreadContext context, IRubyObject obj) {
        return op_equal_19(context, obj);
    }
...
}

RubyMethod extends RubyObject which extends RubyBasicObject, and it
also overrides op_equal with a more specific return type:

public class RubyMethod extends RubyObject {
...
    @JRubyMethod(name = "==", required = 1)
    @Override
    public RubyBoolean op_equal(ThreadContext context, IRubyObject other) {
...
}

We use the JRubyMethod annotation to generate stubs for binding the
Ruby class's methods at boot time. These stubs are generated by
walking declared methods looking for the JRubyMethod annotation.

However, on JDK8, we get an error like this:

Exception in thread "main" java.lang.ClassFormatError: Duplicate
method name&signature in class file
org/jruby/RubyMethod$INVOKER$i$op_equal

This is because JDK8's generated IRubyObject-returning bridge method
also has the JRubyMethod annotation.

Here's the relevant entry in the constant pool:

system ~/projects/jruby $ javap -v -cp core/target/classes/ org.jruby.RubyMethod
Classfile /Users/headius/projects/jruby/core/target/classes/org/jruby/RubyMethod.class
...
  #122 = Utf8               RuntimeVisibleAnnotations
  #123 = Utf8               Lorg/jruby/anno/JRubyMethod;
  #124 = Utf8               name

Here's the hand-written method plus RuntimeVisibleAnnotations:

  public org.jruby.RubyBoolean
op_equal(org.jruby.runtime.ThreadContext,
org.jruby.runtime.builtin.IRubyObject);
    descriptor:
(Lorg/jruby/runtime/ThreadContext;Lorg/jruby/runtime/builtin/IRubyObject;)Lorg/jruby/RubyBoolean;
    ...
    RuntimeVisibleAnnotations:
      0: #123(#124=[s#149],#150=I#137)

And here's the bridge method, which also acquires this annotation:

          public org.jruby.runtime.builtin.IRubyObject
op_equal(org.jruby.runtime.ThreadContext,
org.jruby.runtime.builtin.IRubyObject);
            descriptor:
(Lorg/jruby/runtime/ThreadContext;Lorg/jruby/runtime/builtin/IRubyObject;)Lorg/jruby/runtime/builtin/IRubyObject;
            flags: ACC_PUBLIC, ACC_BRIDGE, ACC_SYNTHETIC
            ...
            RuntimeVisibleAnnotations:
              0: #123(#124=[s#149],#150=I#137)

I'm not arguing that this is correct or incorrect behavior...merely
that it is a change, and it may bite others the same way it has bitten
us. I have patched our stub generator to skip bridge methods, but it
feels kinda wrong.

- Charlie