RFR(M): 8038297 Avoid placing CTI immediately following cbcond instruction on T4

[Igor Veresov]

Thanks, Chris!

igor

> On Apr 10, 2014, at 7:08 PM, Christian Thalinger <christian.thalinger at oracle.com> wrote:
> 
> Looks good.
> 
>> On Apr 10, 2014, at 3:48 PM, Igor Veresov <igor.veresov at oracle.com> wrote:
>> 
>> Vladimir, thanks for the review!
>> 
>> Here is the corrected webrev:  http://cr.openjdk.java.net/~iveresov/8038297/webrev.01/
>> 
>> Thanks!
>> igor
>> 
>>> On Apr 10, 2014, at 2:29 PM, Vladimir Kozlov <vladimir.kozlov at oracle.com> wrote:
>>> 
>>> Igor,
>>> 
>>> Please, clean up code style of if() in  MachEpilogNode methods you touched in sparc.ad. The same in output_h.cpp.
>>> 
>>> I think you should separate into 2 flags Flag_avoid_before and Flag_avoid_after:
>>> 
>>> Flag_avoid_back_to_back  = Flag_may_be_short_branch << 1, // Flag has two bits
>>> 
>>> And use them instead of new MachNode::AvoidBackToBackFlag. I don't like next expression:
>>> 
>>> ins_avoid_back_to_back() * Flag_avoid_back_to_back
>>> 
>>> thanks,
>>> Vladimir
>>> 
>>>> On 4/10/14 11:54 AM, Igor Veresov wrote:
>>>> We need to put nops between cbcond instructions and CTI instructions,
>>>> because otherwise it causes a pipeline flush.
>>>> This change extends the notion of the avoid_back_to_back instruction
>>>> attribute to hold a 2 bits flag, with bits being AVOID_BEFORE and
>>>> AVOID_AFTER.  The meaning of the bits is that an instruction with the
>>>> AVOID_BEFORE flag cannot immediately follow one with AVOID_AFTER flag.
>>>> The assembler has been modified accordingly to ensure that we don’t have
>>>> a CTI instruction immediately following a cbcond instruction.
>>>> 
>>>> Webrev: http://cr.openjdk.java.net/~iveresov/8038297/webrev.00/
>>>> 
>>>> Testing: CTW, hotspot regression tests, JPRT
>>>> 
>>>> There is a significant performance gain on the microbenchmark that
>>>> exposed the problem.
>>>> Before with -XX:+UseCBCond: 106.985ns
>>>> Before with -XX:-UseCBCond: 94.8967ns
>>>> After with -XX:+UseCBCond: 94.5262ns
>>>> 
>>>> 
>>>> Other performance results (some gains with jetstream and volano):
>>>> ==============================================================================
>>>> vanilla.logs:
>>>> Benchmark           Samples        Mean     Stdev             Geomean
>>>> Weight
>>>> reference_server         20    16136.35    107.24
>>>>   jetstream              20      153.30      1.63             0.10
>>>>   scimark                20      353.47      5.78             0.15
>>>>   specjbb2000            20   220984.11    638.71             0.15
>>>>   specjbb2005            20   287778.70   3987.52             0.25
>>>>   specjvm98              20      295.09      1.41             0.15
>>>>   volano25               20   224179.95   5213.65             0.20
>>>> ==============================================================================
>>>> hack1.logs:
>>>> Benchmark           Samples        Mean     Stdev   %Diff     P
>>>> Significant
>>>> reference_server         20    16357.48    111.46    1.37 0.000
>>>>  Yes
>>>>   jetstream              20      160.82      2.07    4.90 0.000
>>>>  Yes
>>>>   scimark                20      351.65      5.53   -0.51 0.316
>>>>    *
>>>>   specjbb2000            20   221142.43   1247.38    0.07 0.617
>>>>    *
>>>>   specjbb2005            20   288196.75   4598.83    0.15 0.760
>>>>    *
>>>>   specjvm98              20      298.11      0.89    1.02 0.000
>>>>  Yes
>>>>   volano25               20   232848.15   3952.61    3.87 0.000
>>>>  Yes
>>>> ==============================================================================
>