harmony-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Tim Ellison <t.p.elli...@gmail.com>
Subject Re: Understanding JAPI tools
Date Mon, 06 Mar 2006 16:35:23 GMT
Dalibor Topic wrote:
> On Mon, Mar 06, 2006 at 02:12:16PM +0800, Paulex Yang wrote:
>> Stuart,
>> Stuart Ballard wrote:
>>> Tim Ellison <t.p.ellison <at> gmail.com> writes:
>>>> Stuart: is there some way to avoid JAPI taking inherited API signatures
>>>> into account when doing the diff?  kinda like the javadoc tool.
>>> This is something of a Japi FAQ (I should probably put together a Japi FAQ 
>>> page
>>> somewhere now that it's no longer the case that pretty much everyone using 
>>> Japi
>>> is on the Classpath list ;) ).
>>> The answer is no, it doesn't do this, and I'm not going to implement it. 
>>> The
>>> reason is that whether a method is overridden or not is *not* part of the 
>>> public
>>> API. Javadoc documents it, sure, but that's not really sufficient to say 
>>> whether
>>> it's part of the API or not - it's quite likely it was simply easier to 
>>> write
>>> Javadoc that way (so it didn't have to go and find the doc comments in the
>>> superclass's source file, etc).
>>> The rules of binary compatibility say that moving a method up or down the 
>>> class
>>> hierarchy is a completely compatible change.
>>> A method may be overridden, or not, purely for performance reasons and the
>>> behavior could be completely identical. For example, LinkedList.subList is 
>>> not
>>> documented as being overridden, but if it's not, that guarantees that the
>>> iterator() method will always be at least O(N) where N is the startIndex 
>>> that
>>> the sublist was created with. An overridden LinkedList.subList method 
>>> could hang
>>> on to the start and end nodes of the list and provide an O(1) 
>>> implementation of
>>> iterator(). As long as some effort was taken to ensure that serialization 
>>> stayed
>>> compatible, there's no reason why a Free runtime shouldn't do that 
>>> optimization
>>> even if Sun doesn't.
>>> I've also heard of cases where Sun documents a method as being overridden 
>>> even
>>> though as far as any testing can determine, all it does is call super. 
>>> There
>>> should be no obligation for any Free runtime to duplicate this 
>>> inefficiency.
>>> It's also possible to implement a method that needs changed behavior in a
>>> subclass by making the superclass delegate some parts of its behavior to
>>> non-public methods, so the superclass method is unchanged but the subclass 
>>> just
>>> overrides the non-public methods (obviously, this only works if the 
>>> subclass is
>>> in the same package, but that's pretty common).
>>> Finally, it's possible (albeit unlikely) that even if there *was* an 
>>> apparent
>>> need to override a method in a subclass because it has different behavior, 
>>> there
>>> might be some weird optimization-related reason why somebody might want to
>>> actually write it as "if (this instanceof Whatever)" in the superclass. I 
>>> can't
>>> see why that would happen, but if someone did that and got all the behavior
>>> right, they shouldn't be penalized for it in Japi.
>> Most of above is true and has good reason,  but as a practicer trying to 
>> implement a RI compatible class library, I still feel strongly that the 
>> method override information has its value, maybe a most common example 
>> is equals(Object), most classes don't override this method and only call 
>> j.l.Object's, but others DO has their own criteria, it's very reasonable 
>> to just remind programmer to implement this method, otherwise it will 
>> introduce serious compatibility issue.
> It won't introduce a binary compatibility issue. That's what really
> matters.

Of course, we need to be more than simply binary compatible with the
reference implementation, and while compliance with the spec gives some
flexibility in the implementation approach, being binary compatible is
not enough.

AIUI the JCK doesn't only test binary compatibility with the API, so if
we were to introduce a new method that is not in the spec, that would be
binary compatible but would fail the JCK.

> You can't get semantics like who overloads what right via a tool,
> you need to read the specs to see if you should override equals/hashCode or 
> not, and how.

That is the point.  We have to be careful not to use JAPI tools to say
Harmony is "X%-complete" -- as we all agree, that will not take into
account the semantics of inherited methods / fields  into account (e.g.
equals() or abstract methods etc.).  I believe that it is useful to
compare the Harmony implementation against the spec even if it does
produce some 'false positives' in order to guide developers to areas
that need enhancing.

>> Just as a proposal, if you still feel that override difference is 
>> another story with common difference, how about to separate them into 
>> another category, similar with Compiler's warning, which at least can be 
>> as reference to programmer, and this information even can be switched on 
>> only if some user want to check the difference more *strictly*.
>>> Basically, Japi's business is to test the coverage of the API, and it's a
>>> fundamental design decision that Japi will *not* be testing implementation
>>> details. I've done quite a lot of work fixing bugs where implementation 
>>> details
>>> inadvertently get exposed in the Japi results and if I find any more I'll 
>>> fix
>>> them too.
>>> I think this is *extremely* important because, among other things, it's the
>>> *only* reason I feel legally secure in distributing Japi files for the 
>>> JDK. As
>>> long as they only contain functional information necessary for 
>>> interoperability,
>>> they're uncopyrightable. If they start exposing copyrightable 
>>> implementation
>>> details, they become derived works of the JDK, and illegal to distribute 
>>> (or,
>>> technically, to even create in the first place). And potentially so do the
>>> results of the comparisons. And if you then use those comparisons to make
>>> decisions on what to implement in Harmony, it jeopardizes Harmony's 
>>> cleanroom
>>> status too. (IANAL; the people who actually decide whether legal risks are
>>> acceptable might think this isn't actually a problem for Harmony, but *I* 
>>> don't
>>> want to be responsible for it ;) )
>>> The right way to ensure that you've overridden all the right methods is to 
>>> test
>>> the *behavior* of those methods. Obviously this is a lot more work than a 
>>> simple
>>> automated run via Japi or equivalent, but it's the only way to get 
>>> meaningful
>>> results.
>> It is definitely the right way, but maybe not the most practical one. 
>> Still the example of equals(), do we need a testEquals for every single 
>> class?

The short answer is 'yes' we do to be sure that equals works for every
object, and to ensure that our inherited implementation of equals()
works the same way as the reference implementation; however, when we
read the javadoc specification we expect every departure from the more
general spec to be documented in the subtype.  Where the choice is made
to implement in the supertype there still needs to be a subtype test case.

What we seem to be missing is a mechanism for stating in Harmony that a
supertype's implementation satisfies a subtypes' specification
requirement.  Again, I am prepared to believe that the set of those
implementation/specification departures are small enough to deal with,
and that the Harmony implementation to javadoc (not RI code) comparison
is worthwhile.

>> or do we still need to check the JavaDoc one by one to find
>> difference on this method? why not the tool just give some warning to 
>> some possible incompatible?
> Because a tool can't know if an implementation detail outside of the
> binary compatibility spec is relevant for an implementation of the J2SE
> API spec or not. AI hasn't caught up with that goal yet. :)
> For example, different compatible implementations 
> of the spec may or may not chose to override equals for a class: as long
> as they are binary compatible (trivial if we are not overriding), and fulfil
> both the contract of the spec for equals in their class, and the
> contract for equals in the superclasses, they are OK. 
> You could for
> example, implement Object.equals as a giant instanceof if-else cascade
> for all the defined types in J2SE, and save yourself all the wasted
> bytes of storage going into having many little trivial methods in many
> classes. Such an implementation would be convenient for embedded uses,
> as it shaves off unnecessary bytecodes very nicely, and still be fully 
> compatible with the specs. 
> You could also implement a generic,
> reflection based equals method in Object that does the right thing in
> 99% of the cases, and just override the few cases where that's not
> good enough. Or use AOP, or something else, as long as your code is 
> binary compatible.
> So yes, you need to wade through the specs one by one to implement them
> faithfully. On the other hand, the specs also give you a lot of leeway
> to do the right thing efficiently, without having to explicitely mimic
> the RI. :)

Any suggestions about how you record (in the code) that you have waded
through the spec and it is implemented right?  or do you think this is a
completeness measure that has to be kept elsewhere?


> cheers,
> dalibor topic
>>> Obviously since Japi is Free Software I can't prevent others from making a
>>> version that does this, but I hope I've shown why I think it would be a 
>>> bad idea.
>> There is another issue of JAPI, pls. correct me if I'm wrong, seems the 
>> *final* information is also ignored. I also think this difference of 
>> method signature is a compatibility issue which could not be overlooked, 
>> because the spec writer might have consider very carefully about which 
>> methods can be overridden and which should be  immutable. Comments?
>>> Stuart.
>> -- 
>> Paulex Yang
>> China Software Development Lab
>> IBM


Tim Ellison (t.p.ellison@gmail.com)
IBM Java technology centre, UK.

View raw message