Request for reviews (M): 6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")

[Vladimir Kozlov]

http://cr.openjdk.java.net/~kvn/6910618/webrev.01

I added comment about return_oop in deoptimization.cpp.
And I added the regression test case. I finally made it fail
with current VM (have to force inlining StringBuilder::append).

Thanks,
Vladimir

Tom Rodriguez wrote:
> On Feb 8, 2010, at 1:44 PM, Vladimir Kozlov wrote:
> 
>> Thank you, Tom
>>
>> Bytecode_invoke_at_check() does not check for allocation bytecodes
>> and the bug case is deoptimization on the return from the call to
>> OptoRuntime::new_array_C(). So I am nervous about checking only
>> bytecode since I may miss something.
> 
> I hadn't even considered that case but your fix definitely seems right.  You might put a larger comment on return_oop indicating under what conditions it's used any what it can't be divined from the bytecodes themselves.
> 
> tom
> 
>> I will remove printing.
>>
>> Thanks,
>> Vladimir
>>
>> Tom Rodriguez wrote:
>>> Oh, right, wrong end of the stack.  So then why can't you get it from the bytecodes of the call site of the last scope?
>>> ScopeDesc* scope = chunk->at(chunk->length() - 1)->scope();
>>> Bytecode_invoke* invoke = Bytecode_invoke_at_check(scope->method(), scope->bci());
>>> if (invoke != NULL) {
>>>  save_oop_result = invoke->result_type() == T_OBJECT;
>>> }
>>> I guess it's possible that the site is being handled in some different way such that we don't actually have a real call there, particularly if reexecute were true.  So your fix seems more right.  What do you think about dropping the =true on the printing output since it seems redundant?
>>> tom
>>> On Feb 8, 2010, at 11:50 AM, Vladimir Kozlov wrote:
>>>> Because it will return deoptee (caller method) and not the method
>>>> from which we returned already (no frame on stack).
>>>>
>>>> Vladimir
>>>>
>>>> Tom Rodriguez wrote:
>>>>> Why do you need to record this in the PcDesc?  Why can't you just do:
>>>>> chunk->at(0)->scope()->method()->return_type() == T_OBJECT
>>>>> tom
>>>>> On Feb 8, 2010, at 11:12 AM, Vladimir Kozlov wrote:
>>>>>> http://cr.openjdk.java.net/~kvn/6910618/webrev.00
>>>>>>
>>>>>> Fixed 6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
>>>>>>
>>>>>> Problem:
>>>>>> If deoptimization happened on the return from a call which
>>>>>> which returns oop, the oop will be not updated during GC which
>>>>>> is triggered by scalar replaced objects reallocation.
>>>>>>
>>>>>> Solution:
>>>>>> Mark in PcDesc call sites which return oop (this is main part of changes)
>>>>>> and save the result oop across objects reallocation during deoptimization.
>>>>>>
>>>>>> Reviewed by:
>>>>>>
>>>>>> Fix verified (y/n): y, test
>>>>>>
>>>>>> Other testing:
>>>>>> JPRT
>>>>>>
>