Subject:
Rhino reflection patch
Date:
Wed, 28 Jul 1999 18:14:52 -0400
From:
Andrew Wason <aw@softcom.com>
To:
norris@netscape.com
CC:
mccabe@netscape.com, rogerl@netscape.com, Howard Lin <howard@softcom.com>
When JavaAdapter generates an adapter class, it does not take into account
the types of method parameters when wrapping the generated methods arguments.
This means that if a non-public class implements a public interface the
non-public class type will be wrapped instead of the declared public
interface - and methods cannot be invoked via the wrapper.
I have attached sample code (reflect-demo.zip) which shows this. The
JavaScript caller.js generates an adapter implementing the CallerInterface
interface. CallerInterface has a method (doSomething) which takes an
argument of type pkg.Interface. pkg.Target is a non-public class that
implements pkg.Interface. If an instance of pkg.Target is passed to the
CallerInterface adapter doSomething method, an Error is thrown because
pkg.Target.doSomething is called (instead of pkg.Interface.doSomething) and
pkg.Target is not public.
I have attached a patch to Context.java, ScriptRuntime.java and
JavaAdapter.java. I overloaded toObject in Context and ScriptRuntime to
take a 3rd argument which is the declared type of the object being
wrapped. This is passed to NativeJavaObject.wrap so that it generates the
correct wrapper. I changed JavaAdapter.generateOverride to generate
bytecode calling Context.toObject passing the declared Class type of the
argument.
Context.java also includes my previously submitted patch for dealing with
SecurityExceptions and the JavaAdapter property (because this patch has not
been checked into CVS yet).
Andrew
--
Andrew Wason
SoftCom, Inc.
aw@softcom.com
reflect-patch.txt
Name:
reflect-patch.txt
Type:
Plain Text (text/plain)
reflect-demo.zip
Name:
reflect-demo.zip
Type:
Zip Compressed Data (application/x-zip-compressed)
Encoding:
base64
2) add a test that acts as a sample for doing oberloaded methods on xpconnect wraped natives.
3) add a NOT_IMPLEMENTED method to wrapped native for getting the prototype JS object.
4) Set the global object of the JSContext as the wrapped native JSObject's parent when creating this JSObject. This makes JS code compiled against the wrapper actuall work!
5) fix the refcounting on factories in the tests/components module
Thanks go to Andrew Wason <aw@softcom.com> for finding this problem and providing the patch:
If we implement a Java interface in JavaScript, and a method in that interface returns a Java object, we get a java.lang.IncompatibleClassChangeError: org/mozilla/javascript/Wrapper exception.
We have attached a sample JavaScript file which duplicates the error.
The problem is JavaAdapter is generating an INVOKEVIRTUAL bytecode to call Wrapper.unwrap, but Wrapper is an interface and so INVOKEINTERFACE should be used instead. As a result of this change, the IFEQ bytecode generated needs to jump more bytes. We have attached a patch that fixes the problem.
Andrew
--
Andrew Wason
SoftCom, Inc.
aw@softcom.com
- normalized initial MPL comment to match the format of others in the tree, including an initial -*- Mode line.
- removed RCS $log$, etc. comments. We use CVS, and they just make spurious changes...
- js_DecompileValueGenerator had rusted due to bytecode/source-note changes,
or maybe parts of it never worked right. Anyway, it now does not induce a
crashing underflow in the decompiler. As part of this fix, it now takes a
checkStack flag telling whether to look for the jsval v argument on the JS
stack. The calls from ImportProperty, js_SetProperty, and js_DeleteProperty
pass in v a jsval for the property id, which should not be sought after on
the stack (it might happen to be there due to o['p'] = 2, but we want to
decompile o["p"], not "p").
- js_DecompileValueGenerator would load a generating pc even if the value v
did not match the pc's corresponding stack item! Oops. This lead to less
than idea diagnostics.
- js_DecompileValueGenerator was also not mapping JSOP_TRAP to the real op at
a sufficiently early and univeral point in its control flow.
- Fix PopOff to assert and check for stack underflow in the decompiler, and
beef up PushOff too (it asserted, but did not check).
- js_ReportIsNotFunction now avoids JS_InternString by indexing directly into
cx->runtime->atomState.typeAtoms with the result of JS_TypeOfValue.
- Removed unnecessary local GC root reserved by non-zero trailing member of
obj_eval's JSFunctionSpec initializer.
Add an #if defined(__arm) check to tell jsdtoa.c about the (somewhat bizzare) endian-ness of the ARM architecture.
Thanks go to Stefan Hanske <sh990154@mail.uni-greifswald.de> for this fix.
rogerl%netscape.com
mccabe%netscape.com
norris%netscape.com
jband%netscape.com
briano%netscape.com
brendan%mozilla.org
waldemar%netscape.com
cbegle%netscape.com
Jerry.Kirk%Nexwarecorp.com