On Fri, 2017-03-31 at 15:54 +0300, Jani Nikula wrote: > > I'm sure the parameter name could be improved to capture what you > mean better; alas I don't have a suggestion. Yes, that's a fair point - perhaps "functions-not-linked" or something like that. > > Internally this works by collecting (per-file) those functions > > (and enums, structs, doc sections...) that are explicitly used, > > and invoking the kernel-doc script with "-nofunction" later. > > A quick thought that I don't have the time to check now, but should > be checked before merging: Is the order of directive extension > execution deterministic if the Sphinx run is parallelized (sphinx- > build -j)? Is it deterministic within an rst file? Surely it's not > deterministic when called from several rst files? The latter is, > perhaps, acceptable, but the former not. Interesting, TBH I never even considered this. How would I even run it that way? Presumably "make htmldocs" doesn't do this? Sphinx documentation (http://www.sphinx-doc.org/en/stable/extdev/) says this: The setup() function can return a dictionary. This is treated by Sphinx as metadata of the extension. Metadata keys currently recognized are: [...] 'parallel_read_safe': a boolean that specifies if parallel reading of source files can be used when the extension is loaded. It defaults to False, i.e. you have to explicitly specify your extension to be parallel-read-safe after checking that it is. We do set this right now, so I guess it'd only be guaranteed to work right within a single rst file, and then I should perhaps consider not making this state global but somehow linking it to the rst file being processed? johannes
