RE: [cp-patches] RFC: RMI Class loader fix

[Andrew Haley]

Jeroen Frijters writes:
 > Andrew Haley wrote:
 > > Tom Tromey writes:
 > >  > >>>>> "Jeroen" == Jeroen Frijters <address@hidden> writes:
 > >  > 
 > >  > Jeroen>         * java/rmi/server/RMIClassLoader.java
 > >  > Jeroen>         (loadClass(String, String)): Use 
 > > Class.forName() instead of
 > >  > Jeroen>         directly calling ClassLoader.loadClass(), to 
 > > handle array
 > >  > Jeroen>         types correctly.
 > >  > 
 > >  > Andrew, didn't we run into code that expected ClassLoader.loadClass
 > >  > to handle array signatures?
 > > 
 > > Yes, we did.  This is what the other runtimes I tried do, and it
 > > should be fixed in ClassLoader.loadClass() in order to be compatible.
 > 
 > I investigated this in detail (before I discovered the RMI "bug") and I
 > found that Sun does not, in general, support loading array classes via
 > ClassLoader.loadClass, however they become visible thru
 > ClassLoader.loadClass() (actually, ClassLoader.findLoadedClass()) under
 > specific circumstances:
 > 
 > JDK 1.4: After a class has been loaded, all arrays of the class are also
 > available thru ClassLoader.findLoadedClass().
 > 
 > JDK 1.5: After a Class.forName() has been done on an array type, the
 > specific array type becomes available thru
 > ClassLoader.findLoadedClass().
 > 
 > So this means that, in general, you cannot load an array type thru
 > ClassLoader.loadClass().

All of this is fairly ill-specified.  ClassLoader.loadClass() can to
cope with "[I", for example, and there is extant application code that
requires it.  That's how I discovered the same incompatibility that
you've found.  This may be a bug, or perhaps just a side-effect of a
particular implementation.

Is there a strong reason to prefer fixing all the clients of
loadClass() ?  There doesn't seem to be any strong specification
reason to do so.

Andrew.