RFR: 8216497: javadoc should auto-link to platform classes

[Hannes Wallnöfer]

On Tue, 15 Sep 2020 11:30:09 GMT, Jan Lahoda <jlahoda at openjdk.org> wrote:

>> This pull request is identical with the RFR previously sent for the Mercurial repository:
>> 
>> https://mail.openjdk.java.net/pipermail/javadoc-dev/2020-August/001796.html
>> 
>> I'm copy-pasting the comments from the original RFR below.
>> 
>> Most of the new code is added to the Extern class where it fits in quite nicely and can use the existing supporting
>> code for setting up external links.
>> The default behaviour is to generate links to docs.oracle.com for released versions and download.java.net for
>> prereleases. Platform documentation URLs can be configured using the new --link-platform-properties <url> option to
>> provide a properties file with URLs pointing to to alternative locations. See the CSR linked above for more details on
>> the new options.  The feature can be disabled using the --no-platform-link option, generating the same output as
>> previously.   One problem I had to solve was how to handle the transition from prerelease versions to final releases,
>> since the location of the documentation changes in this transition. For obvious reasons we don’t want to make that
>> switch via code change at a time shortly before release.  The way it is done is that we determine if the current
>> javadoc instance is a prerelease version as indicated by the Version returned by BaseConfiguration#getDocletVersion(),
>> and then check whether the release/source version of the current javadoc execution uses the same (latest) version. This
>> means that that only the latest version will ever generate prerelease links (e.g. running current javadoc 16 with
>> source version 15 will generate links to the final release documentation) but I think this is acceptable.  Another
>> issue I spent some time getting right was tests. New releases require a new element-list resource*), so tests have to
>> pick up new releases. On the other hand, we don’t want hundreds of tests to fail when a new release is added, ideally
>> there should be one test  with a clear message. Because of this, when a release is encountered for which no
>> element-list is available a warning is generated instead of an error, and the documentation is generated without
>> platform links as if running with --no-platform-link option. This allows most existing tests to pass and just the new
>> test to fail with a relatively clear message of what is wrong.
>> *) I also thought about generating the element-list for the current release at build time. It’s quite involved, and we
>>  still need to maintain element-lists for older versions, so I’m not sure it’s worth it.
>> 
>> For existing tests that check output affected by the new option I added  the --no-platform-link option to disable the
>> feature. Otherwise we’d have to update those tests with each new release (or else freeze the tests to use some static
>> release or source version, which we don’t want either).  I updated the CSR to the new code. It also needs to be
>> reviewed: https://bugs.openjdk.java.net/browse/JDK-8251181
>> 
>> Thanks,
>> Hannes
>
> I think it would be awesome if we could generate (most of) the {element,package}-list-VERSION.txt from the ct.sym
> historical data at build time. This would (hopefully) help with long-term maintenance. I've started to sketch that
> here: https://github.com/lahodaj/jdk/commit/36c1510587a4b050b148eda87ae7a7a89aff9690
> Some comments on the attempt:
> -in this PR, there is package-list-9.txt - should it be element-list-9.txt, and should it contain module names (dtto
>  element-list-10.txt)?
> -we may (for historical reasons) want to keep the lists for 7, 8, 9 and 10 (as the historical data in ct.sym do not
>  exactly match what is in the package/element lists). It would be better to generate everything, but having a fixed list
>  for some of the past versions would be better than having to create a new list for each new release.
> -the patch does not yet generate the list for the current release, but that should be doable.

Thanks for the suggestions and help, Jan!

> I think it would be awesome if we could generate (most of) the {element,package}-list-VERSION.txt from the ct.sym
> historical data at build time. This would (hopefully) help with long-term maintenance. I've started to sketch that
> here: [lahodaj at 36c1510](https://github.com/lahodaj/jdk/commit/36c1510587a4b050b148eda87ae7a7a89aff9690)

I agree files should be generated dynamically. I knew about the sym files but wasn't sure how to go about it. Thanks a
lot for stepping in and helping out, it's very much appreciated!

> Some comments on the attempt:
> -in this PR, there is package-list-9.txt - should it be element-list-9.txt, and should it contain module names (dtto
>  element-list-10.txt)?

Javadoc in 9 still uses the old package-centric layout (package-list and no module names in URL paths). It only became
fully module-aware in 10.

> -we may (for historical reasons) want to keep the lists for 7, 8, 9 and 10 (as the historical data in ct.sym do not
>  exactly match what is in the package/element lists). It would be better to generate everything, but having a fixed list
>  for some of the past versions would be better than having to create a new list for each new release.
> -the patch does not yet generate the list for the current release, but that should be doable.

I definitely agree. I'll work on a new version that generates as much of the lists as possible.

-------------

PR: https://git.openjdk.java.net/jdk/pull/171