RFR(S) 8240205: Avoid PackageEntry lookup when loading shared classes
calvin.cheung at oracle.com
Fri Apr 3 17:23:27 UTC 2020
Thanks for your review!
On 4/3/20 9:46 AM, Lois Foltan wrote:
> On 4/2/2020 8:58 PM, Calvin Cheung wrote:
>> RFE: https://bugs.openjdk.java.net/browse/JDK-8240205
>> webrev: http://cr.openjdk.java.net/~ccheung/jdk15/8240205/webrev.00/
>> During loading of a shared class from a CDS archive, the PackageEntry
>> is being looked up in several places within the same call stack. This
>> RFE is to do the PackageEntry lookup at one place and then pass it
>> along to the subsequently called functions which need it.
>> Perf improvement on loading approx. 2000 classes on linux:
>> instr delta = -17930620 -0.6155%
>> time delta = -6.311 ms -0.7516%
>> Passed tiers1 - 4 testing.
> Hi Calvin,
> This looks good & promising on the numbers! I have a slight concern
> that you've introduced yet another "get package entry from name" sort
> of function in SystemDictionaryShared. classfile/modules.cpp already
> contains a static method get_package_entry_by_name() but it operates
> on a package name not a class name. Please consider moving
> SystemDictionaryShared::get_package_entry_from_class_name into
> ClassLoader near ClassLoader::get_package_from_class_name(). That
> seems like a more consistent place to add it that various areas of the
> JVM can take advantage of.
The SystemDictionaryShared::get_package_entry_from_class_name function
is currently being used only within the SharedDictionaryShared class.
That's why it was declared private and I think that's the right place
There also another SystemDictionaryShared::get_package_entry() which
operates on a package name and is very similar to the one in
ClassLoader. I can file another RFE to consolidate those 2 functions.
More information about the hotspot-runtime-dev