Request for review(M): 6627983: G1: Bad oop deference during marking

[Igor Veresov]

On 3/1/11 12:23 PM, Vladimir Kozlov wrote:
> I have only comments about library_call.cpp:
>
> Why "false" in copy_to_clone()? It seems, you get double barriers in
> this method as well:
> 4085 /*dest_uninitialized*/false);

It doesn't matter here, because we're treating it all as T_LONGs, so 
there's going to be no barriers. But for consistency I suppose it could 
be "true" here.

>
>
> 4596 // so the the pre-barriers wouldn't peek into the old values. See
> CR 6627983.
> ^ 'the' 2 times ^ unitialized
Fixed. Thanks!

> 4597 const bool& dest_uninitialized = must_clear_dest;
> ^ why reference?
>

John wanted a separate variable to inform arraycopy of whether the 
target is initialized. This variable should in fact follow the state of 
the must_clear_dest, so I decided to make it a reference, so that when 
must_clear_dest changes so does dest_uninitialized. If you don't like 
that I can make it a regular variable.

Although I would rather just used must_clear_dest for everything and not 
replicate the state.


Thanks,
igor


> The rest changes seems fine.
>
> Thanks,
> Vladimir
>
> Igor Veresov wrote:
>> John, all,
>>
>> I've updated the change, making corrections that John recommended.
>> Also, now the decision about what to do with the barrier is deferred
>> to the barrier emitting procedure. Thus, if a new pre-barrier is added
>> that would need to do something even in case if the destination array
>> is uninitialized it can be handled properly.
>>
>> New webrev: http://cr.openjdk.java.net/~iveresov/6627983/webrev.01
>>
>> Thanks,
>> igor
>>
>>
>>
>> On 2/25/11 8:25 PM, Igor Veresov wrote:
>>> I think my fix is not good enough for the case when we'll need
>>> prebarriers that will not require the previous value and these could be
>>> a reality in the future. So, it is generally incorrect to elide
>>> prebarriers but is only ok for a specific flavor. I'm working to
>>> alleviate this problem a little bit and will post the updated version
>>> early next week. John, I will also address your recommendations.
>>>
>>> Thanks,
>>> igor
>>>
>>>
>>> On 2/24/11 6:28 PM, John Rose wrote:
>>>> On Feb 24, 2011, at 6:02 PM, Igor Veresov wrote:
>>>>
>>>>> On 2/24/11 5:27 PM, John Rose wrote:
>>>>>> On Feb 24, 2011, at 4:49 PM, Igor Veresov wrote:
>>>>>>
>>>>>> If it is simply foo(..., !must_clear_dest), it becomes hard for
>>>>>> maintainers to see what the argument means.
>>>>>
>>>>> I can see your point but I just didn't want to introduce additional
>>>>> clutter. For example generate_block_arraycopy() never requires
>>>>> barrier suppression, so why do we need to add an argument there?
>>>>
>>>> There's not a strong need. An extra "true" or "false" would be a
>>>> little clutter, but it would also hint to programmers what's going on.
>>>> It's a esthetic call mostly...
>>>>
>>>> (Here's some meandering on the subject. While reading the code, if I
>>>> am tracing the flow from caller definition to callee definition, when
>>>> arguments are explicit I can observe them directly. But if they are
>>>> defaulted, then I have to visit the callee's declaration as well. I
>>>> have to visit three locations in the source instead of two. I
>>>> personally find this awkward unless the meaning of the defaulted
>>>> argument is immediately and intuitively obvious. In this case the
>>>> meaning of the argument is subtly related to other parts of the
>>>> system.)
>>>>
>>>>> From the debugging standpoint it's way better to emit a barrier than
>>>>> not to. Missing barriers will be a nightmare to find. So, I'd rather
>>>>> have a barrier erroneously emitted in some path by default, which
>>>>> would make the marking crash almost immediately.
>>>>
>>>> Good point.
>>>>
>>>>> As for using !must_clear_dest as a gate, I thought it would more
>>>>> clearly convey to the reader the reason why the barrier is suppressed
>>>>> - because the dest array is not cleared and contains garbage, so the
>>>>> reader won't have to go back to see in which cases use_pre_barrier is
>>>>> set. Perhaps I should add more comments.
>>>>
>>>> The bug number would help a lot, I think. It's a subtle interaction
>>>> between (a) zeroing removal, (b) the arraycopy stubs, and (c) G1
>>>> invariants.
>>>>
>>>>>> In the stub generator the optional boolean is OK as is, although I
>>>>>> generally prefer optional booleans to default to 'false'. Given the
>>>>>> way the bug is structured, I keep wanting to call it
>>>>>> 'suppress_pre_barrier', meaning "just this once, trust me when I
>>>>>> tell you not to emit the pre-barrier, if there is one."
>>>>>>
>>>>>
>>>>> That's pretty much equivalent. I don't have strong feelings about
>>>>> this and will change it if you think it's more readable.
>>>>
>>>> I do think it would be more readable. It could be just an old Lisper
>>>> bias, though: Optional arguments default to NULL, which is the false
>>>> value in Lisp.
>>>>
>>>> -- John
>>>
>>