[10] (S) RFR 8191788: add jdk.internal.vm.compiler to --limit-modules if -Djvmci.Compiler=graal is in the command line

[David Holmes]

On 29/11/2017 7:05 PM, Alan Bateman wrote:
> On 28/11/2017 23:51, Vladimir Kozlov wrote:
>> http://cr.openjdk.java.net/~kvn/8191788/webrev.00/
>> https://bugs.openjdk.java.net/browse/JDK-8191788
>>
>> This is redo of JDK-8190975 [1] fix which added 
>> jdk.internal.vm.compiler to tests which have --limit-modules option. 
>> Unfortunately tests start failing on SPARC where Graal 
>> (jdk.internal.vm.compiler) is not available. JDK-8191653 [2] reversed 
>> 8190975 changes to fix that problem.
>>
>> This fix adds jdk.internal.vm.compiler to --limit-modules list in JVM 
>> only when Graal is used: when it is explicitly specified with 
>> -Djvmci.Compiler=graal or in default case and when UseJVMCICompiler is 
>> true.
>>
>> I tested the fix with failed tests from JDK-8190975 which are mostly 
>> AppCDS tests in test/runtime/appcds/jigsaw and also 
>> test/jdk/java/lang/String/concat/WithSecurityManager.java
>>
>> I think this fix may break upgradeable status of Graal (for Oracle 
>> Labs version of Graal).
>> But it should be fine since it is only used with --limit-modules which 
>> is not used by Labs.
> If jdk.internal.vm.compiler is not observable on SPARC then shouldn't 
> the tests have `@requires jdk.internal.vm.compiler` and jtreg will skip 
> the test on that platform?

I didn't know @requires supported modules that way! Even so we'd also 
have to reapply the original fix for JDK-8190975 to update --limit-modules.

> Just asking as augmenting the value passed to --limit-modules is very 
> strange. It's normal for XX options to augment the set of modules that 
> resolved (+EnableJVMCI implies `--add-modules jdk.internal.vm.ci` for 
> example) but doing this for --limit-modules suggests the VM is doing 
> something to mask an issue with the way that the tests are run.

The current approach basically prevents anyone using JVMCI from shooting 
themselves in the foot by setting --limit-modules in a way that excludes 
the jdk.internal.vm.compiler module. In essence we want to ensure that 
if using JVMCI then all the requisite pieces will be available. But 
perhaps we are doing this the wrong way: how do --limit-modules and 
--add-modules combine? We could add jdk.internal.vm.compiler rather than 
expanding the limit-set. But the end result would seem the same.

Thanks,
David

> 
> -Alan