Re: Eclipse 3.0

[Eric Blake]

Robert Lougher wrote:
> OK, I'm yelling :)  Please don't use Mark's patch!
>
> It "fixes" it by stopping resolution finding the method on Object. 
> However, it is legal to have an invokeinterface on an Object method
> (as all interfaces subclass Object).
>
> Instead, please use the attached patch to the interpreter.  This
> converts an invokeinterface on an Object method into an invokevirtual
> (which is what 1.4 javac does anyway, but not jikes or 1.3 javac).

The JVMS is clear that invokeinterface is only to be used on interface 
methods.  JLS 13.1 also states that a method call of something that (also) 
exists in Object is converted into a method call directly on Object if and 
only if the method was not redeclared along the type hierarchy of the type of 
the method call's qualifying expression.  All versions of javac to date are 
buggy in some regard on this matter, but the latest jikes is emitting correct 
code.  So, a correct compiler produces:

class A {
  public String toString() { return ""; }
}
interface B {
  int hashCode();
}
class C extends A implements B {
  public boolean equals(Object o) { return false; }
  {
    toString(); // invokevirtual C.toString; javac 1.3 is wrong
    ((A) this).toString(); // invokevirtual A.toString
    ((B) this).toString(); // invokevirtual Object.toString
    hashCode(); // invokevirtual Object.hashCode
    ((A) this).hashCode(); // invokevirtual Object.hashCode
    ((B) this).hashCode(); // invokeinterface B.hashCode; javac 1.4 is wrong
    equals(null); // invokevirtual C.equals
    ((A) this).equals(null); // invokevirtual Object.equals
    ((B) this).equals(null); // invokevirtual Object.equals
  }
}

--
Someday, I might put a cute statement here.

Eric Blake             address@hidden