RFR(S) 8007270: Make IsMethodCompilable test work with tiered

[Nils Eliasson]

New:

http://cr.openjdk.java.net/~neliasso/8007270/webrev.05/

Answers inline.

On 2014-02-12 22:41, Vladimir Kozlov wrote:
> Indention is slightly off, otherwise new code is good. Can you shift 
> code a little?:
>
> +        for (long attempts = 0, successes = 0;
> +             (successes < PER_METHOD_RECOMPILATION_CUTOFF) &&
> +             (attempts  < PER_METHOD_RECOMPILATION_CUTOFF*2) &&
> +             isCompilable(COMP_LEVEL_FULL_OPTIMIZATION);
> +             attempts++) {
> +          if (compileAndDeoptimize() == COMP_LEVEL_FULL_OPTIMIZATION) {
> +              successes++;
>            }
>         }
>
> Do we need to assert (throw exception) after the loop when (successes 
> != PER_METHOD_RECOMPILATION_CUTOFF)?

No, that will be caught be the other tests. First that it is still 
compilable, and then after one more compilation and deoptimization, that 
it is no longer compilable.

> You left debugging output code in compileAndDeoptimize(). Do you want 
> to keep it?

It was supposed to be removed.

Thanks,
Nils

>
> Thanks,
> Vladimir
>
> On 2/12/14 1:46 AM, Nils Eliasson wrote:
>> I rewrote the code. It should be much cleared now.
>>
>> http://cr.openjdk.java.net/~neliasso/8007270/webrev.04/
>>
>> Thanks,
>> //N
>>
>>
>> On 2014-02-11 23:10, Vladimir Kozlov wrote:
>>> Sorry, I missed that. That code is very messy. I think you should
>>> decrement 'i' instead and change comment:
>>>
>>> +        long cutoff = PER_METHOD_RECOMPILATION_CUTOFF;
>>> +        // deoptimize 'PerMethodRecompilationCutoff' times
>>> +        for (long i = 0L; (i < cutoff) &&
>>> isCompilable(COMP_LEVEL_FULL_OPTIMIZATION); ++i) {
>>> +          int level = compileAndDeoptimize();
>>> +          if (level < COMP_LEVEL_FULL_OPTIMIZATION) {
>>> +            // Add one iteration if compiled with a lower tier
>>> +            --i;
>>>            }
>>>          }
>>>
>>> Thanks,
>>> Vladimir
>>>
>>> On 2/11/14 1:34 PM, Nils Eliasson wrote:
>>>> cutoff may increase if the method was compiled for a lower level 
>>>> and we
>>>> want another attempt for full opt, the second check is to limit the
>>>> amount of attempts.
>>>>
>>>> //N
>>>>
>>>> On 2014-02-11 20:06, Vladimir Kozlov wrote:
>>>>> (i <c) should be always true since (i < cutoff):
>>>>>
>>>>> +        for (long i = 0L; (i < cutoff) &&
>>>>> +              (i < PER_METHOD_RECOMPILATION_CUTOFF*2) &&
>>>>>
>>>>> Thanks,
>>>>> Vladimir
>>>>>
>>>>> On 2/11/14 5:15 AM, Nils Eliasson wrote:
>>>>>> Hi,
>>>>>>
>>>>>> Thanks for your contribution Igor.
>>>>>>
>>>>>> After some additional testing I noticed some inconsistencies and
>>>>>> filed https://bugs.openjdk.java.net/browse/JDK-8034188
>>>>>>
>>>>>> This test now runs faster, and targets c2-compiles only, in server
>>>>>> and with tiered.
>>>>>>
>>>>>> Webrev: http://cr.openjdk.java.net/~neliasso/8007270/webrev.02/
>>>>>>
>>>>>> Thanks,
>>>>>> //Nils
>>>>>>
>>>>>>
>>>>>> On 2014-01-29 11:11, Igor Ignatyev wrote:
>>>>>>> Hi Nils,
>>>>>>>
>>>>>>> you can skip running on client by yourself, see attached diff.
>>>>>>> Igor
>>>>>>>
>>>>>>> On 01/28/2014 07:02 PM, Nils Eliasson wrote:
>>>>>>>> Hi all,
>>>>>>>>
>>>>>>>> I need a review for this change.
>>>>>>>>
>>>>>>>> http://cr.openjdk.java.net/~neliasso/8007270/webrev.01/
>>>>>>>>
>>>>>>>> This test was disabled since it didn't work very well with tiered
>>>>>>>> (or
>>>>>>>> client). It tests the PerMethodRecompilationCutoff that was
>>>>>>>> introduced
>>>>>>>> to disable c2-compilations of a method when it has been 
>>>>>>>> deoptimized
>>>>>>>> too
>>>>>>>> many times.  The bug report suggested we should disable c1
>>>>>>>> compilations
>>>>>>>> as well but I don't think that was the intent of the cutoff 
>>>>>>>> feature.
>>>>>>>>
>>>>>>>> I have changed the following in the test
>>>>>>>> * skip test when running client only (not supported by jtreg at 
>>>>>>>> the
>>>>>>>> moment)
>>>>>>>> * check what compilation level was used when compiling so that it
>>>>>>>> can
>>>>>>>> keep track of the number of c2 compiles (and deopts) correctly in
>>>>>>>> tiered
>>>>>>>> mode
>>>>>>>> * compile and deopt up to the cutoff limit only once
>>>>>>>> * added PerMethodRecompilationCutoff=4 flag to commandline to 
>>>>>>>> reduce
>>>>>>>> wasted time in test (default 400)
>>>>>>>>
>>>>>>>> Now the test works and the running time has been reduced to 
>>>>>>>> seconds
>>>>>>>> instead of minutes.
>>>>>>>>
>>>>>>>> Kind regards,
>>>>>>>> Nils Eliasson
>>>>>>>>
>>>>>>>>
>>>>>>
>>>>
>>