RFR: CODETOOLS-7902799: perfasm still handles event modifiers incorrectly [v2]

[Aleksey Shipilev]

On Tue, 8 Dec 2020 09:23:00 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:

>> Unfortunately, this regresses `perfasm` on systems that reply `cycles:u:` in default configuration:
>> 
>> 21831844.118759: cycles:u:  ffff000010081478 [unknown] ([unknown])
>> 
>> ...yields:
>> 
>> PrintAssembly processed: 202041 total address lines.
>> Perf output processed (skipped 12.981 seconds):
>>  Column 1: cycles (0 events)
>
> What's worse, if I do:
> 
> $ java -jar benchmarks.jar JMHSample_09_Blackholes.measureRight_2 -f 1 -w 1s -r 1s -prof "perfasm:savePerf=true;events=cycles:ppp
> 
> ...
> Perf output processed (skipped 13.276 seconds):
>  Column 1: cycles:ppp (0 events)
> 
> $ head ./org.openjdk.jmh.samples.JMHSample_09_Blackholes.measureRight_2-AverageTime.perf
> 21832047.816581: cycles:pppu:      ffffa34b1080 _start (/usr/lib64/ld-2.28.so)
> 
> So perf liberally adds `u` to the event tags. I guess we should go back at fixing the parsing code to match events without tags?

Please try this patch -- it works with both `cycles:ppp` and default `cycles:u:` on some systems:
 https://cr.openjdk.java.net/~shade/jmh/7902799-perf-fixes-1.patch

If it works, please replace your PR contents with it :)

-------------

PR: https://git.openjdk.java.net/jmh/pull/9