Hi Aleksey,
On 8/01/2019 8:20 am, Aleksey Shipilev wrote:
> RFE:
> https://bugs.openjdk.java.net/browse/JDK-8216302
>
> Fix:
> http://cr.openjdk.java.net/~shade/8216302/webrev.01/
>
> There is already Class.name field that is used as the cache for Class.getName(). We can use that
> from inside the VM code when building the stack traces, to avoid converting Symbol*->String all the
> time. We have to take care that both Java (JVM_GetClassName) and internal paths yield the same
> cached value, so some code rearrangement and tests are in order.
It seems somewhat awkward to me to have two different code paths for
initializing the java.lang.Class name field. Can this be restructured a
little more (change Class.getName()) so that the VM always initializes
"name" and then JVM_GetClassName could just call java_lang_Class::name,
rather than duplicating the logic?
Specific comments:
src/hotspot/share/classfile/javaClasses.cpp
! Klass* k = as_Klass(java_class);
! assert(k->is_klass(), "just checking");
! name = k->external_name();
as_Klass already has the requisite assertions so there was no reason to
change this part of the code. I see that jvm.cpp already contains the
same redundant logic.
Copyright years need updating.
Thanks,
David
> This alleviates a part of stack trace performance degradation that happened 8->11, see the umbrella
> issue for test and discussion:
> https://bugs.openjdk.java.net/browse/JDK-8151751
>
> Linux x86_64 release:
>
> # 8u191
> StackTraceBench.test 1 avgt 15 10.851 ± 0.075 us/op
> StackTraceBench.test 10 avgt 15 15.325 ± 0.089 us/op
> StackTraceBench.test 100 avgt 15 59.717 ± 0.449 us/op
> StackTraceBench.test 1000 avgt 15 529.020 ± 3.654 us/op
>
> # jdk/jdk, baseline
> StackTraceBench.test 1 avgt 15 23.835 ± 0.188 us/op
> StackTraceBench.test 10 avgt 15 33.204 ± 0.191 us/op
> StackTraceBench.test 100 avgt 15 125.195 ± 0.694 us/op
> StackTraceBench.test 1000 avgt 15 1051.047 ± 9.779 us/op
>
> # jdk/jdk, patched
> StackTraceBench.test 1 avgt 15 14.450 ± 0.136 us/op
> StackTraceBench.test 10 avgt 15 20.182 ± 0.088 us/op
> StackTraceBench.test 100 avgt 15 77.107 ± 0.632 us/op
> StackTraceBench.test 1000 avgt 15 647.128 ± 6.159 us/op
>
> Testing: Linux x86_64 fastdebug {new test, hotspot tier1}, jdk-submit (running)
>
> Thanks,
> -Aleksey
>