RFR: 8166002: Emulate client build on platforms with reduced virtual address space

Vladimir Kozlov vladimir.kozlov at oracle.com
Mon Jan 9 19:27:34 UTC 2017 

Hi, Jamsheed

Why you set flags related to segmented codecache? 
NonProfiledCodeHeapSize, ProfiledCodeHeapSize, NonNMethodCodeHeapSize

SegmentedCodeCache should be off in Client mode. SegmentedCodeCache 
setting is done in set_tiered_flags() which is called after you changed 
ReservedCodeCacheSize in set_client_compilation_mode() and will not be 
set to true.

Unfortunately we can do nothing for CodeCacheSegmentSize which is 128 
bytes in tiered VM since it is constant in product VM.

I also forgot to tell that you should use FLAG_SET_ERGO() macro for 
flags settings in set_client_compilation_mode().

Thanks,
Vladimir

On 1/3/17 9:56 AM, Vladimir Kozlov wrote:
> Good.
>
> Thanks,
> Vladimir
>
> On 1/3/17 8:17 AM, Jamsheed C m wrote:
>> Hi Vladimir,
>>
>> On 1/3/2017 11:15 AM, Vladimir Kozlov wrote:
>>> Hi Jamsheed
>>>
>>> Why you moved set_client_compilation_mode() to c1_globals and new
>>> class? Why not to keep it in compilerDefinitions?
>>>
>> Done.
>>> selected_with_other_flags setting and code in
>>> compilation_mode_selected() have the same checks - factor them into
>>> one function.
>> Done.
>>>
>>> I would suggest to keep comment in compilationPolicy.cpp.
>> Done.
>> revised webrev: http://cr.openjdk.java.net/~jcm/8166002/webrev.05/
>>
>> Best Regards,
>> Jamsheed
>>>
>>> I reviewed CCC request.
>>>
>>> Thanks,
>>> Vladimir
>>>
>>> On 1/2/17 8:25 PM, Jamsheed C m wrote:
>>>> Hi Vladimir, Igor,
>>>>
>>>> i changed implementation a little bit(i.e now more close to -client
>>>> behavior). Tests are modified to support this mode.
>>>>
>>>> http://cr.openjdk.java.net/~jcm/8166002/webrev.04/
>>>>
>>>>
>>>> A new command-line option -XX:CompilationMode=<server/client> will
>>>> be provided in JDK 9 to emulate a client build on
>>>> 32-bit Windows platforms using a standard tiered build.
>>>>
>>>> -XX:CompilationMode=server--selects to build time compilation mode
>>>> of server VM.
>>>> -XX:CompilationMode=client--selects non-tiered c1 compiler. (Serial
>>>> Garbage Collection in ON and Xshare is set to auto
>>>> in this mode).
>>>>
>>>> This option is available only on Tiered build (i.e. Server VM).
>>>> Client VM ignores this flag and continues with build time
>>>> compilation mode.
>>>>
>>>> This option will have no effect if any of TieredCompilation,
>>>> TieredStopAtLevel, UseAOT, EnableJVMCI, UseJVMCICompiler is
>>>> specified at the command-line.
>>>>
>>>> Platforms except 32-bit Windows runs in server compilation mode by
>>>> default. 32-bit Windows runs in client compilation
>>>> mode by default. The user should specify
>>>> -XX:CompilationMode=<server/client> to switch.
>>>>
>>>> -XX:+/-TieredCompilation or -XX:TieredStopAtLevel=<0-4> or
>>>> -XX:+/-UseAOT or -XX:+/-EnableJVMCI or
>>>> -XX:+/-UseJVMCICompiler always switch to build time compilation mode
>>>> plus corresponding option behavior irrespective of
>>>> CompilationMode option.
>>>>
>>>> I have updated the CCC request link in bug report, Please review.
>>>>
>>>> Best Regards,
>>>> Jamsheed
>>>>
>>>> On 12/6/2016 12:22 AM, Vladimir Kozlov wrote:
>>>>> On 12/5/16 6:21 AM, Jamsheed C m wrote:
>>>>>> Hi Vladimir,
>>>>>>
>>>>>> On 12/3/2016 8:02 AM, Vladimir Kozlov wrote:
>>>>>>> On 12/2/16 7:38 AM, Jamsheed C m wrote:
>>>>>>>> Hi Vladimir, Igor,
>>>>>>>>
>>>>>>>> I made a few more changes to include the optimization that was
>>>>>>>> available only on client.
>>>>>>>>
>>>>>>>> 1) a few c1 changes (JDK-7153771..)
>>>>>>>
>>>>>>> Which one?
>>>>>> This one.
>>>>>
>>>>> I see.
>>>>>
>>>>> Thanks,
>>>>> Vladimir
>>>>>
>>>>>> src/share/vm/c1/c1_Compilation.hpp
>>>>>>
>>>>>>   // will compilation make optimistic assumptions that might lead to
>>>>>>     // deoptimization and that the runtime will account for?
>>>>>>     bool is_optimistic() const                             {
>>>>>> - return !TieredCompilation &&
>>>>>> + return is_client_compilation_mode_vm() &&
>>>>>>         (RangeCheckElimination || UseLoopInvariantCodeMotion) &&
>>>>>> method()->method_data()->trap_count(Deoptimization::Reason_none)
>>>>>> == 0;
>>>>>>     }
>>>>>>
>>>>>> src/share/vm/c1/c1_Runtime1.cpp
>>>>>> - assert(!TieredCompilation, "incompatible with tiered compilation");
>>>>>> + assert(is_client_compilation_mode_vm(), "incompatible in non
>>>>>> client modes");
>>>>>>
>>>>>> mpegaudio was about 6% low without this.
>>>>>>
>>>>>>>>
>>>>>>>> 2) Made SerialGC as default GC.
>>>>>>>
>>>>>>> Should be only for client_compilation_mode
>>>>>> Ok.
>>>>>>>
>>>>>>>>
>>>>>>>> 3) Set CICompileCount=1 for client compilation mode.
>>>>>>>
>>>>>>> I think check should be is_client_compilation_mode_vm()
>>>>>> Ok.
>>>>>> Best Regards,
>>>>>> Jamsheed
>>>>>>>
>>>>>>>>
>>>>>>>> 4) Tuned thresholds to get better performance.
>>>>>>>
>>>>>>> ok
>>>>>>>
>>>>>>>>
>>>>>>>> 5) Added relevant flags in few tests, modified
>>>>>>>> TestSelectDefaultGC.java.
>>>>>>>
>>>>>>> ok
>>>>>>>
>>>>>>> Thanks,
>>>>>>> Vladimir
>>>>>>>
>>>>>>>>
>>>>>>>> revised webrev: http://cr.openjdk.java.net/~jcm/8166002/webrev.03/
>>>>>>>>
>>>>>>>> performance report is added in bug report.
>>>>>>>>
>>>>>>>> Best Regards,
>>>>>>>>
>>>>>>>> Jamsheed
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> On 11/1/2016 1:14 AM, Vladimir Kozlov wrote:
>>>>>>>>> webrev.02 looks good to me too.
>>>>>>>>>
>>>>>>>>> Thanks,
>>>>>>>>> Vladimri
>>>>>>>>>
>>>>>>>>> On 10/31/16 10:39 AM, Igor Veresov wrote:
>>>>>>>>>> Jamsheed explained to me that ReservedCodeCacheSize is now set
>>>>>>>>>> in the else clause (udiffs are not showing the
>>>>>>>>>> proper alignment and I missed it). The change looks good to me.
>>>>>>>>>>
>>>>>>>>>> igor
>>>>>>>>>>
>>>>>>>>>>> On Oct 31, 2016, at 9:42 AM, Igor Veresov
>>>>>>>>>>> <igor.veresov at oracle.com> wrote:
>>>>>>>>>>>
>>>>>>>>>>> Assuming it gets the performance/startup numbers close to the
>>>>>>>>>>> client VM it looks fine.
>>>>>>>>>>> But what about adjusting the code cache sizes? With tiered we
>>>>>>>>>>> get 240M ReservedCodeCacheSize, which also turns
>>>>>>>>>>> on SegmentedCodeCache. It seems like we won’t need the
>>>>>>>>>>> “profiled” segment of the code
>>>>>>>>>>> cache at all. It is also likely that we’d do fine with a
>>>>>>>>>>> smaller overall code cache.
>>>>>>>>>>>
>>>>>>>>>>> igor
>>>>>>>>>>>
>>>>>>>>>>>> On Oct 31, 2016, at 9:14 AM, Jamsheed C m
>>>>>>>>>>>> <jamsheed.c.m at oracle.com> wrote:
>>>>>>>>>>>>
>>>>>>>>>>>> Hi Vladimir, Igor,
>>>>>>>>>>>>
>>>>>>>>>>>> revised webrev:
>>>>>>>>>>>> http://cr.openjdk.java.net/~jcm/8166002/webrev.02/
>>>>>>>>>>>>
>>>>>>>>>>>> i took flags from bug comment.
>>>>>>>>>>>>
>>>>>>>>>>>> Igor V. suggested next flags setting to emulate Client VM
>>>>>>>>>>>> compile threshold with TieredStopAtLevel=1:
>>>>>>>>>>>>
>>>>>>>>>>>> Tier3BackEdgeThreshold=14000
>>>>>>>>>>>> Tier3CompileThreshold=1500
>>>>>>>>>>>> Tier3InvocationThreshold=1500
>>>>>>>>>>>> Tier3MinInvocationThreshold =1500
>>>>>>>>>>>>
>>>>>>>>>>>> Best Regards,
>>>>>>>>>>>>
>>>>>>>>>>>> Jamsheed
>>>>>>>>>>>>
>>>>>>>>>>>> On 10/28/2016 11:48 PM, Vladimir Kozlov wrote:
>>>>>>>>>>>>> Add comment what the code does and why. Move the code into
>>>>>>>>>>>>> separate function reaturning bool. And condition
>>>>>>>>>>>>> CodeCache setting based on result. It will reduce #ifdef mess.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Can you put && to the end of previous line? To get good
>>>>>>>>>>>>> alignment.
>>>>>>>>>>>>> Should we also change CompilationPolicyChoice?
>>>>>>>>>>>>>
>>>>>>>>>>>>> Ask Igor Veresov to verify settings.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Thanks,
>>>>>>>>>>>>> Vladimir
>>>>>>>>>>>>>
>>>>>>>>>>>>> On 10/28/16 10:59 AM, Jamsheed C m wrote:
>>>>>>>>>>>>>> Hi Vladimir,
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> revised webrev with just ergo settings for win32.
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> http://cr.openjdk.java.net/~jcm/8166002/webrev.01/
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> Best Regards
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> Jamsheed
>>>>>>>>>>>>>>
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> On 10/28/2016 12:25 PM, Vladimir Kozlov wrote:
>>>>>>>>>>>>>>> Hi Jamsheed,
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> Why you need changes in tests?
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> Why you can't use set_client_compilation_mode() in both
>>>>>>>>>>>>>>> places?
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> Looks fine otherwise.
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> Thanks,
>>>>>>>>>>>>>>> Vladimir
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> On 10/20/16 8:45 AM, Jamsheed C m wrote:
>>>>>>>>>>>>>>>> Hi All,
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> bug id:https://bugs.openjdk.java.net/browse/JDK-8166002
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> http://cr.openjdk.java.net/~jcm/8166002/webrev.00/
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> Fix: set NeverActAsServerClassMachine=true, Limit
>>>>>>>>>>>>>>>> TieredCompilation to C1 by default, for win32 platform.
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> A new flag CompilationMode=<client/server> is introduced
>>>>>>>>>>>>>>>> to select server/ client mode compilation. This
>>>>>>>>>>>>>>>> option is supported only in TIERED builds.
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> -XX:CompilationMode=server supports both
>>>>>>>>>>>>>>>> -XX:+/-TieredCompilation.
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> -XX:CompilationMode:client ignores TieredCompilation flag.
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> Please review,
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> Best Regards,
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> Jamsheed
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>
>>>>>>
>>>>
>>

More information about the hotspot-compiler-dev mailing list