[9] RFR (S) 8148639: Some MethodCounter fields can be excluded when not including C2

Vladimir Kozlov vladimir.kozlov at oracle.com
Tue Mar 22 00:02:08 UTC 2016

Chris can you also replace:







On 3/21/16 4:53 PM, Chris Plummer wrote:
> Hello,
> Please review the following for removing a couple of MethodCounter fields when not including C2 (or JVMCI) in the build.
> This helps reduce footprint for the minimal VM (and the client VM also).
> http://cr.openjdk.java.net/~cjplummer/8148639/webrev.03/webrev.hotspot/
> https://bugs.openjdk.java.net/browse/JDK-8148639
> There were a couple of changes the previously were in JDK-8147978 that I had to add to this webrev since JDK-8147978 was
> backed out. They include making ProfileInterpreter related options unsupported when not using C2, and also the new
> macros in macros.hpp.
> In order to make sure these counters really are not used when not using C2, I took a few safeguards. The first was to
> make no changes other than to assert that whenever these counters are fetched, they are equal to zero. I did quite a bit
> of testing with this and never hit the asserts.
> I would have liked to #ifdef out interpreter_invocation_count() and interpreter_throwout_count(), but there are too many
> places that call them, which meant too many #ifdef in IMHO. So instead of #ifdef'ing them out, I just make them return 0
> when not using C2. This is safe because of the assert testing I did above.
> I do completely #ifdef out the two increment methods. interpreter_throwout_increment() is only called by some
> ProfileInterpreter code in bytecodeInterpreter.cpp, so I #ifdef'd that code also. In interpreterRuntime.cpp I #ifdef'd
> out a call to interpreter_throwout_increment(). Although this code may have been executed when not using C2, the assert
> testing I did above showed that if the increment happened, the counter was never used later.
> There are quite a few #ifdefs in methodCounters.hpp. I could collpase 5 into one big #if/#else/#endif section for all
> the inline method implementations. It looks cleaner, be then also puts distance between the two different versions of
> the same method.
> Testing was done with jprt "-testset hotspot". I also did a lot of testing with various tools, svc, and compiler test
> lists, and also runThese. This was done on linux-x86 with "-client -Xcomp" and just "-client", and also on linux-x64
> with "-server -XX:+TieredCompilation" (although I think that was probably the default anyway).
> thanks,
> Chris

More information about the hotspot-compiler-dev mailing list