RFR: 8275509: ModuleDescriptor.hashCode isn't reproducible across builds [v9]

[Jaikiran Pai]

On 03/11/21 8:50 am, Jaikiran Pai wrote:
> Hello Alan,
>
> On 02/11/21 5:30 pm, Alan Bateman wrote:
>> On 02/11/2021 06:38, Jaikiran Pai wrote:
>>> :
>>>
>>> Perhaps run 1 writing the hash code of each of the boot modules' 
>>> descriptor into a file and then run 2 reading that same file to 
>>> compare the hash codes would be one way to do it. But I think that 
>>> would just make this test more complex, which I think is avoidable.
>>
>> I'm not sure that we really need this. There are several jlink tests 
>> that check reproducibility, we think one of them is failing 
>> intermittently with this issue already. So maybe we should leave 
>> build reproducibility to the jlink tests and expand them as needed.
>>
>> For ModuleDescriptor::hashCode then I was hoping we keep it simple 
>> and just test that the hashCode of the descriptors for modules in the 
>> boot layer matches the hashCode computed from re-parsing the 
>> module-info files, e.g. something simple like this:
>>
>>       for (Module module : ModuleLayer.boot().modules()) {
>>             System.out.println(module);
>>             ModuleDescriptor descriptor1 = module.getDescriptor();
>>             ModuleDescriptor descriptor2;
>>             try (InputStream in = 
>> module.getResourceAsStream("module-info.class")) {
>>                 descriptor2 = ModuleDescriptor.read(in);
>>             }
>>             assertEquals(descriptor1, descriptor2);
>>             assertEquals(descriptor1.hashCode(), 
>> descriptor2.hashCode());
>>             assertEquals(descriptor1.compareTo(descriptor2), 0);
>>         }
>>
>> It run can twice, with with the default, the other with -Xshare:off.
>
> Sounds reasonable. I've updated the PR to simplify the test and run it 
> once with default CDS and once with -Xshare:off. Given this 
> simplification, it now allows me to use testng for these comparisons, 
> so I've moved it from regular "main" to testng test case. I hope 
> that's OK.

Test continues to pass with these changes and fails (as expected) when 
the fix in the source code of ModuleDescriptor class is rolled back.

-Jaikiran