RFR: 8172751: OSR compilation at unreachable bci causes C1 crash

[Andreas Woess]

Cool, thanks, Tobias!

Best regards,
Andreas

On 2017-01-18 18:47, Tobias Hartmann wrote:
> Hi Andreas,
>
> I'll sponsor your change (I did some testing and the results look good).
>
> Best regards,
> Tobias
>
> On 16.01.2017 09:16, Tobias Hartmann wrote:
>> Hi Andreas,
>>
>> On 16.01.2017 02:53, Andreas Woess wrote:
>>> Hi Tobias,
>>>
>>> I've updated the test and added a comment to make it clearer.
>>>
>>> We enter the first loop, where we flip the outer if's condition to make
>>> the branch we're in unreachable, but continue in the first loop until
>>> it's OSR-compiled. C1 then skips the now seemingly unreachable first
>>> loop and doesn't visit the OSR entry point.
>> Thanks, it's clearer now. Looks good to me!
>>
>> Best regards,
>> Tobias
>>
>>> Thanks,
>>> Andreas
>>>
>>> On 2017-01-13 10:54, Tobias Hartmann wrote:
>>>> Hi Andreas,
>>>>
>>>> On 13.01.2017 05:04, Andreas Woess wrote:
>>>>> Hi all,
>>>>>
>>>>> please review the following patch:
>>>>> http://cr.openjdk.java.net/~aw/8172751/webrev/
>>>>> https://bugs.openjdk.java.net/browse/JDK-8172751
>>>>>
>>>>> This is a small fix for a extremely unlikely but consistently
>>>>> reproducible crash (see included test) in the C1 GraphBuilder with OSR.
>>>>> AFAICT, the problem is that we may never visit the osr entry block if
>>>>> it's found to be unreachable during parsing due to constant folding.
>>>>>
>>>>> The patch simply changes the existing assertion to a bailout. We'd crash
>>>>> immediately after it trying to get the osr block's state (in a product
>>>>> build; in a debug build the assertion fails).Observed on JDK 8u* and 9.
>>>> This looks reasonable to me!
>>>>
>>>> Could you please provide some more information on how the test case works? I guess we trigger OSR compilation in the first loop of executeLoop() and then make the branch unreachable during compilation by calling flipSwitch(), right? Maybe add some comments to the test.
>>>>
>>>> Thanks,
>>>> Tobias
>>>>
>>>>> Thanks,
>>>>> Andreas
>>>>>