[9] RFR(S): 8162881: Effect of -XX:CICompilerCount depends on ordering of other flags

[Tobias Hartmann]

Hi,

RBT testing found some problems with the WB_IsIntrinsicAvailable method. With TieredStopAtLevel=1, we now don't create the C2 compiler threads anymore but WB_IsIntrinsicAvailable expects 'CompileBroker::compiler(compLevel)' to be non-null. I fixed the method to check for TieredStopAtLevel and also added the corresponding check to the tests.

New webrev:
http://cr.openjdk.java.net/~thartmann/8162881/webrev.01/

I verified that the failing tests pass and re-submitted RBT (still running).

Thanks,
Tobias

On 03.08.2016 08:58, Tobias Hartmann wrote:
> Thanks, Vladimir!
> 
> Best regards,
> Tobias
> 
> On 02.08.2016 23:46, Vladimir Kozlov wrote:
>> Thank you for fixing this, Tobias
>>
>> Changes looks good.
>>
>> Thanks,
>> Vladimir
>>
>> On 8/2/16 5:30 AM, Tobias Hartmann wrote:
>>> Hi,
>>>
>>> please review the following patch:
>>> https://bugs.openjdk.java.net/browse/JDK-8162881
>>> http://cr.openjdk.java.net/~thartmann/8162881/webrev.00/
>>>
>>> There are multiple problems and inconsistencies with the CICompilerCount flag:
>>>
>>> 1) We can not set the number of compiler threads to 1 if -TieredCompilation is specified after CICompilerCount=1:
>>>   $ java -XX:CICompilerCount=1 -XX:-TieredCompilation -version
>>>   CICompilerCount (1) must be at least 2
>>>   Improperly specified VM option 'CICompilerCount=1'
>>>   Error: Could not create the Java Virtual Machine.
>>>   Error: A fatal exception has occurred. Program will exit.
>>>
>>> 2) We can not set the number of compiler threads to 1 if TieredStopAtLevel=1 is specified after CICompilerCount=1:
>>>   $ java -XX:CICompilerCount=1 -XX:-TieredStopAtLevel=1 -XX:+PrintFlagsFinal -version
>>>   CICompilerCount (1) must be at least 2
>>>   Improperly specified VM option 'CICompilerCount=1'
>>>   Error: Could not create the Java Virtual Machine.
>>>   Error: A fatal exception has occurred. Program will exit.
>>>
>>> 3) The number of compiler counts can be set to 1 with TieredStopAtLevel=1 but has no effect:
>>>   $ java -XX:TieredStopAtLevel=1 -XX:CICompilerCount=1 -XX:+PrintFlagsFinal -version | grep CICompilerCount
>>>        intx CICompilerCount                          = 2                                        {product} {command line, ergonomic}
>>>
>>> 4) The value of CICompilerCount is listed as being "ergonomic" although it was set via command line:
>>>   $ java -XX:TieredStopAtLevel=1 -XX:CICompilerCount=5 -XX:+PrintFlagsFinal -version | grep CICompilerCount
>>>       intx CICompilerCount                          = 5                                        {product} {command line, ergonomic}
>>>
>>> To fix 1) and 2), I changed the ConstraintType of the CICompilerCount constraint to 'AfterErgo'. Like this, the value of CICompilerCount is only checked after the effect of -XX:-TieredCompilation took place.
>>>
>>> To fix 3), I added special cases to the initialization of SimpleThresholdPolicy and AdvancedThresholdPolicy to set the C2 compiler count to zero if TieredStopAtLevel < CompLevel_full_optimization.
>>>
>>> To fix 4), I changed the ergonomics code to only use FLAG_SET_ERGO if the value of CICompilerCount is updated. I also changed the computation of 'count' to always be >= 2 because with os::active_processor_count() == 1, count is ergonomically set to 1 but both c1_count and c2_count are set to 2. I added an assert to check for consistency.
>>>
>>> I added the corresponding test cases to CheckCICompilerCount.
>>>
>>> Tested with regression test, JPRT and RBT (running).
>>>
>>> Thanks,
>>> Tobias
>>>