RFR (XXS): 8017571: JSR292: JVM crashing on assert "cast to instanceKlass" while producing MethodHandle for array methods with MethodHandle.findVirtual

[Christian Thalinger]

Thank you, Vladimir.  -- Chris

On Jul 2, 2013, at 4:48 PM, Vladimir Kozlov <vladimir.kozlov at oracle.com> wrote:

> Good.
> 
> Thanks,
> Vladimir
> 
> On 7/2/13 4:23 PM, Christian Thalinger wrote:
>> 
>> On Jun 26, 2013, at 11:25 AM, Vladimir Kozlov <vladimir.kozlov at oracle.com> wrote:
>> 
>>> On 6/26/13 11:19 AM, Christian Thalinger wrote:
>>>> 
>>>> On Jun 26, 2013, at 11:02 AM, Vladimir Kozlov <vladimir.kozlov at oracle.com> wrote:
>>>> 
>>>>> On 6/26/13 10:42 AM, Christian Thalinger wrote:
>>>>>> 
>>>>>> On Jun 26, 2013, at 10:02 AM, Vladimir Kozlov <vladimir.kozlov at oracle.com> wrote:
>>>>>> 
>>>>>>> You can code it simpler (you don't need to check all types variations):
>>>>>> 
>>>>>> I could but I wanted to keep the MemberName class around for debugging purposes.  I could get rid of the type array:
>>>>> 
>>>>> What debugging? C++ will optimize it anyway in product build. And with debug build you can call java_lang_invoke_MemberName::clazz(mname()) in debugger to get original klass.
>>>> 
>>>> Not always.  Sometimes the debugger just barfs on you (for whatever reason) and sometimes you can't call methods (e.g. when you are out of stack space).  I had all this recently.
>>> 
>>> You can simple have next for that:
>>> 
>>> !     Klass* reference_klass = java_lang_Class::as_Klass(java_lang_invoke_MemberName::clazz(mname()));
>>> +     DEBUG_ONLY( Klass* mname_klass = reference_klass; )
>> 
>> I've used your earlier suggestion and updated the webrev:
>> 
>> http://cr.openjdk.java.net/~twisti/8017571/webrev/
>> 
>> -- Chris
>> 
>>> 
>>> Vladimir
>>> 
>>>> 
>>>> -- Chris
>>>> 
>>>>> 
>>>>> Vladimir
>>>>> 
>>>>>> 
>>>>>> +           reference_klass = NULL;
>>>>>> 
>>>>>> though.  Sometimes I'm torn between simpler, better to read and less code.
>>>>>> 
>>>>>> -- Chris
>>>>>> 
>>>>>>> 
>>>>>>> !     Klass* reference_klass = java_lang_Class::as_Klass(java_lang_invoke_MemberName::clazz(mname()));
>>>>>>> +     if (reference_klass != NULL && reference_klass->oop_is_objArray()) {
>>>>>>> +       reference_klass = ObjArrayKlass::cast(reference_klass)->bottom_klass();
>>>>>>> +     }
>>>>>>> +
>>>>>>> +     // Reflection::verify_class_access can only handle instance classes.
>>>>>>> +     if (reference_klass != NULL && reference_klass->oop_is_instance()) {
>>>>>>>        // Emulate LinkResolver::check_klass_accessability.
>>>>>>> 
>>>>>>> Vladimir
>>>>>>> 
>>>>>>> On 6/26/13 9:15 AM, Christian Thalinger wrote:
>>>>>>>> 
>>>>>>>> On Jun 25, 2013, at 6:19 PM, Vladimir Kozlov <vladimir.kozlov at oracle.com> wrote:
>>>>>>>> 
>>>>>>>>> How you got ObjArrayKlass here? Also bottom_klass() could be TypeArrayKlass.
>>>>>>>> 
>>>>>>>> You are right; it doesn't work.  Reflection::verify_class_access can only handle instance classes.  Here is a new webrev that checks for instance classes before doing the check:
>>>>>>>> 
>>>>>>>> http://cr.openjdk.java.net/~twisti/8017571/webrev/
>>>>>>>> 
>>>>>>>> I noticed that is_public is a Klass method so we can remove the InstanceKlass cast.
>>>>>>>> 
>>>>>>>> -- Chris
>>>>>>>> 
>>>>>>>>> 
>>>>>>>>> Vladimir
>>>>>>>>> 
>>>>>>>>> On 6/25/13 6:03 PM, Christian Thalinger wrote:
>>>>>>>>>> http://cr.openjdk.java.net/~twisti/8017571/webrev/
>>>>>>>>>> 
>>>>>>>>>> 8017571: JSR292: JVM crashing on assert "cast to instanceKlass" while producing MethodHandle for array methods with MethodHandle.findVirtual
>>>>>>>>>> Reviewed-by:
>>>>>>>>>> 
>>>>>>>>>> Verification code in MHN_resolve_Mem calls Reflection::verify_class_access with MemberName.clazz as new_class.  If the class happens to be an array class we hit an assert.
>>>>>>>>>> 
>>>>>>>>>> The fix is to check for array types and use the bottom type for these.
>>>>>>>>>> 
>>>>>>>>>> src/share/vm/prims/methodHandles.cpp
>>>>>>>>>> 
>>>>>>>> 
>>>>>> 
>>>> 
>>