Request for reviews (S): 6916062: assert(_inserts <= _insert_limit, "hash table overflow") in NodeHash::hash_insert

[Vladimir Kozlov]

I updated changes with bailout from IGVM optimization loop:

http://cr.openjdk.java.net/~kvn/6916062/webrev.03

Thanks,
Vladimir

On 9/29/10 12:14 PM, Vladimir Kozlov wrote:
> Thank you, Tom
>
> On 9/29/10 12:00 PM, Tom Rodriguez wrote:
>> That seems fine for a backport. Maybe we should consider some product level code to bailout if we do too many rounds
>> of IGVN?
>
> Yes, it is good idea to bailout. I will add it.
>
> Thanks,
> Vladimir
>
>>
>> tom
>>
>> On Sep 28, 2010, at 7:08 PM, Vladimir Kozlov wrote:
>>
>>> http://cr.openjdk.java.net/~kvn/6916062/webrev.02
>>>
>>> Fixed 6916062: assert(_inserts<= _insert_limit,"hash table overflow") in NodeHash::hash_insert
>>>
>>> Changes for 6711117 put a memory node back on IGVN
>>> worklist if address type does not match cached type
>>> in a hope that address nodes will be processed and
>>> the type will be updated. But it does not check that
>>> there are nodes on worklist which will trigger address
>>> processing. As result the memory node is pulled from
>>> and pushed on IGVN worklist infinitely.
>>> On each iteration NodeHash::_inserts value is incremented
>>> forcing hash table grow until integer value NodeHash::_max
>>> is overflow (became 0) which triggers bug's assert.
>>>
>>> I decided to go with simplest fix which is easy to backport
>>> and leave fixing address type inconsistency for later
>>> (I have bug 6715629).
>>> I also added asserts to catch such situations.
>>>
>>> Tested with failing case, CTW, JPRT.
>>