RFR: 8323675: Race in jdk.javadoc-gendata [v2]

[Erik Joelsson]

On Thu, 18 Jan 2024 13:54:23 GMT, Erik Joelsson <erikj at openjdk.org> wrote:

>> @lahodaj Just to be absolutely clear: In `jdk.javadoc-gendata`, we're calling two tools: Not only `JavadocElementList` (which only requires source, not class files), but also `CreateSymbols build-javadoc-data`. Can you confirm that this too only requires sources?
>> 
>> (I take it that this usage is different from `CreateSymbols build-ctsym` as called in `jdk.compiler-gendata`.)
>
> Looking at `make/modules/jdk.javadoc/Gendata.gmk` the recipe never references `$(JDK_OUTPUTDIR)/modules/`, the output dir where the compiled classes are located. It only references the output of `$(call GetModuleSrcPath)` (and some static src dirs), so I'm pretty confident that this is correct.

Somewhat related to this, while investigating another bug and with this PR fresh in memory, I think we are missing another dependency. 

In `make/modules/jdk.javadoc/Gendata.gmk`, we are copying the generated files into `jdk.javadoc.interim` (part of interim langtools). This makes me believe that the generated files are needed when we use `NEW_JAVADOC` from interim langtools to build the API docs for the JDK. However, there is no dependency declared from any of the `docs-*-api-javadoc` targets to `jdk.javadoc-gendata`, so if this works today, it's just by luck.

I see two possible outcomes:

1. If the generated files from `jdk.javadoc-gendata` are expected to be present and used when generating the main API docs for the JDK, then we need to add dependencies for that.
2. If the generated files aren't actually needed, we should stop copying them to avoid unnecessary work and non deterministic behavior.

Not sure if we should hijack this PR for this problem, probably better to file a separate issue. @lahodaj what do you think about this?

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

PR Review Comment: https://git.openjdk.org/jdk/pull/17402#discussion_r1458067165