RFR (XS): 8148044: Remove Enum[0] constants from EnumSet and EnumMap

[Claes Redestad]

On 2016-01-22 15:22, Chris Hegarty wrote:
> Claes,
>
> On 22/01/16 12:07, Claes Redestad wrote:
>> On 2016-01-22 12:56, Claes Redestad wrote:
>>> On 2016-01-22 12:53, Alan Bateman wrote:
>
>>>> On 22/01/2016 11:34, Claes Redestad wrote:
>>>>> :
>>>>> During testing I discovered that EnumSet was missing a
>>>>> serialVersionUID, which needs to be added.
>>>> I assume this is not needed, it already has:
>>>>
>>>> @SuppressWarnings("serial") // No serialVersionUID due to usage of
>>>> serial proxy pattern
>>>
>>> That's odd.. there's a jtreg test that fails if I don't explicitly add
>>> the serialVersionUID (that's where I got the actual UID from). I'll
>>> investigate.
>>
>> Right...  java/util/EnumSet/BogusEnumSet.java fails, which is an
>> explicit test to verify a corrupt EnumSet doesn't deserialize properly:
>
> Hmmm... this would indicate that the changes you have made has
> affected the serial version UID automatically generated. I'm not
> sure why removing a 'private static' field would do this. Is it
> possible that you ran into this during development, but it is
> not actually necessary with the final changes?

> This test is a little fragile, but should be ok.
>
>> It would seem the assessment that serialVersionUID is not needed when
>> using the serial proxy pattern might be misguided (or is this a bug in
>> ObjectInputStream)?
>
> It is debate-able whether this is a test bug or not. But I think
> we should not change the serialVersionUID for this class, whether
> truly Serializable or not. Even though in practice it is not all
> that useful.
>
> Can you please revert the serialization parts of the change, and
> rerun the test? To see if still passes or fails.

I backed out the serial version UID change and reran the test, same thing.

I dug deeper, as this is a bit vexxing...

While a static field does not in and of itself cause the generated UID 
to change,
the private ZERO_LENGTH_ENUM_ARRAY is being accessed from SerializationProxy
class, thus javac generates a package-private synthetic method 
(access$000) which
will be accounted for when generating serialVersionUID. That's unfortunate.

I ran a few experiments to confirm this:

- Adding a package-private static method to EnumSet makes the test fail
- Adding any static field and the test keep passing

So it seems our options here are limited.

Thanks!

/Claes