RFR: 8293116: Incremental JDK build could be sped up [v3]

[Magnus Ihse Bursie]

On Wed, 7 Sep 2022 16:30:33 GMT, Erik Joelsson <erikj at openjdk.org> wrote:

>>> ... and also that we _must_ include the vardeps file, since any change there should trigger a complete rebuild.
>>> 
>>> The logic here seems correct, but apparently somewhat hard to fully understand correctly. Maybe a few lines of comments summarizing the important points raised here would be a good thing?
>> 
>> How about:
>> https://github.com/openjdk/jdk/pull/10104/commits/6c1a83b28ee72a99bf24e7f66c96d1294ad485dd
>> 
>> Thanks,
>>     Jan
>
> To clarify my last post. The Main compile rule has this prereq declaration:
> 
> 
> $$($1_COMPILE_TARGET): $$($1_SRCS) $$($1_FILELIST) $$($1_DEPENDS) \
>         $$($1_VARDEPS_FILE) $$($1_EXTRA_DEPS) $$($1_JAVAC_SERVER_CONFIG)
> 
> The `$$($1_DEPENDS)` and `$$($1_EXTRA_DEPS)` may contain prereqs/dependencies that we probably should always recompile everything for. The EXTRA_DEPS is where we get the API changes from other modules. The DEPENDS variable is set by the caller and can be anything. Both of these are (deliberately) missing from the FILELIST generation rule, but I think MODFILES generation needs them. Because of this, I think the new MODFILES logic needs to move to the main compilation rule (like in your original patch) but the FILELIST rule should still be left alone.
> 
> We need to verify that this still gives us the expected speedup in the common case though. I'm not sure in what situations other dependencies may be included in `$?` here.

Actually, this is not an entirely correct description. EXTRA_DEPS points to the `depend` javac plugin itself in buildtools. Once built, this is unlikely to change. Dependency on the API changes themselves happen on a higher level, in `CompileJavaModules.gmk`, where we add a dependency of the outputted `pubapi` file for modules we depend on. 

(I did have to look this up, though)

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

PR: https://git.openjdk.org/jdk/pull/10104