RFR: 8193213 & 8182731: Make the UseAppCDS option obsolete

Jiangli Zhou jiangli.zhou at oracle.com
Thu Apr 26 00:24:45 UTC 2018

Hi David,

> On Apr 25, 2018, at 3:12 PM, David Holmes <david.holmes at oracle.com> wrote:
> Hi Jiangli,
> On 26/04/2018 3:39 AM, Jiangli Zhou wrote:
>> Hi David,
>> Thanks a lot for the review!
>>> On Apr 24, 2018, at 11:17 PM, David Holmes <david.holmes at oracle.com> wrote:
>>> cc'ing build-dev for the makefile change
>>> Hi Jiangli,
>>> On 25/04/2018 10:53 AM, Jiangli Zhou wrote:
>>>> Please review the following changes that streamline the support for application class data sharing by eliminating the requirement of using -XX:+UseAppCDS to enable the feature. The support for application class data sharing is now enabled transparently when application classes or platform classes present in the classlist or generated archive with -Xshare:dump/on/auto.
>>>> 	RFE: https://bugs.openjdk.java.net/browse/JDK-8193213
>>>> 	Bug: https://bugs.openjdk.java.net/browse/JDK-8182731
>>> These issues are not publicly visible, can you change them to be so?
>> Done. Thanks for noticing that!
>>>> 	webrev: http://cr.openjdk.java.net/~jiangli/8193213_8182731/webrev.00/
>>> Overall this seems fine. Thanks for explaining the logic changes below - much appreciated!
>>> One query: why was UseAppCDS not removed from the tests (or at least the tests you were modifying anyway) ? They will all need updating before 12 when the flag is expired.
>> The usages are left as backwards compatibility check. They also serve the testing purpose to make sure the presence of the flag does not cause any unexpected behavior. Those are the main reasons why I didn’t remove the flag usages in this webrev.
> This sounds like you are basically testing whether obsoleting the flag has worked correctly.

Right, that was part of the intention.

> You don't need to do that through formal testing. A simple run of "java -XX:+UseAppCDS" should show the obsoletion warning and that's that. We don't maintain an obsolete flag test the way we do for deprecated flags.

That sounds reasonable.

> So IMHO there's really no reason to keep the flag in any of the tests. As I said they will all have to be removed when the flag is expired in 12.

Ok. I’ll remove the flag from the tests.



> Thanks,
> David
>>> test/hotspot/jtreg/runtime/appcds/SharedArchiveFile.java
>>> Test 2 reference to UseAppCDS seems unnecessary now.
>>> Test 3 is not needed as you're not using any diagnostic flag now.
>>> Test 4 seems unnecessary now.
>> Will do.
>>> test/hotspot/jtreg/runtime/appcds/TestCommon.java
>>> makeCommandLineForAppCDS should just be removed (yes that will cause some churn in the other tests) - but this can be a follow up test cleanup.
>> Ok.
>>> test/hotspot/jtreg/runtime/appcds/UseAppCDS.java
>>> This test no longer makes much sense to me. The only tests should be that app classes get archived and used. It makes no sense to claim to be testing -XX:+UseAppCDS when that flag is obsolete. Likewise now that it is obsolete you don't need to test that -XX:-UseAppCDS has no affect.
>> Sounds reasonable. I’ll remove UseAppCDS.java. There are existing tests that check app classes get archived and used.
>> Thanks!
>> Jiangli
>>> Thanks,
>>> David
>>>> Highlights of the changes:
>>>> * UseAppCDS is obsolete in JDK 11
>>>> * Remove the +UnlockCommercialFeatures requirement when using application class data sharing in OracleJDK binary
>>>> * SharedArchiveFile is a product flag for all configurations in JDK 11
>>>> * DumpLoadedClassList produces a complete classlist including all loaded library and application classes
>>>> Thanks David Holmes for his guidance on CSR process.
>>>> Following are the details for the VM and test changes:
>>>> - Removed various ‘if (UseAppCDS)’ checks and UseAppCDS asserts.
>>>> - Removed some of the CDS/AppCDS specifics from general class loading code.
>>>> - Removed scattered checks for non-empty directory. Dump time non-empty directory check is done at one central place, FileMapInfo::check_nonempty_dir_in_shared_path_table() after loading all classes on the classlist. The behavior is backwards compatible:
>>>>   - If no app/platform classes are loaded, dump time only reports error if non-empty directory exists in -Xbootclasspath/a path
>>>>   - If app/platform classes are loaded, dump time reports error for any non-empty directory exists in -Xbootclasspath/a path, class path, or module path
>>>> - Removed unnecessary ‘if (!DumpSharedSpaces)’ from SharedClassUtil::initialize(). The code path is only executed when UseSharedSpaces is enabled.
>>>> - Return immediately in SystemDictionaryShared::find_or_load_shared_class() if the archive header indicates there is no app or platform classes in the shared archive. This reduces overhead in class loading if the shared archive only contains system classes. It also provides backwards compatibility in cases where shared application classes should be disabled. For example, when the default system class loader is replaced by user provided loader, archived application classes are inactive (not loaded from archive) at runtime without affecting the use of archived system classes.
>>>> - Changed GenerateLinkOptData.gmk to filter out HelloClasslist from the default classlist in JDK image, which is generated using DumpLoadedClassList option. Thanks for David and Ioi's input on this.
>>>> - Updated various cds/appcds tests to reflect the JVM changes. The use of -XX:+UseAppCDS in tests remains unchanged.
>>>>   - Removed -XX:+UnlockCommercialFeatures for -XX:+UseAppCDS in tests
>>>>   - Removed -XX:+UnlockDiagnosticVMOptions for -XX:SharedArchiveFile in tests
>>>>   - Updated NonBootLoaderClasses.java: ensure app/platform classes on classlist are archived without -XX:+UseAppCDS
>>>>   - Updated DirClasspathTest.java: reflect the change for non-empty directory handling
>>>>   - Updated MismatchedUseAppCDS.java:  -XX:-UseAppCDS has no effect
>>>> Tested with all cds and appcds tests locally with both fastdebug and release builds. Tested with hs-teir1, hs-tier2, jdk-tier1 and jdk-tier2. The hs-tier3, hs-tier4 and hs-tier4 are in progress.
>>>> Thanks,
>>>> Jiangli

More information about the hotspot-runtime-dev mailing list