On Fri, Apr 09 2021, Jeff King wrote:
> On Fri, Apr 09, 2021 at 12:08:23AM +0200, Ævar Arnfjörð Bjarmason wrote:
>
>> > -config-list.h:
>> > +config-list.h: Documentation/*config.txt Documentation/config/*.txt
>> > $(QUIET_GEN)$(SHELL_PATH) ./generate-configlist.sh \
>> > >$@+ && mv $@+ $@
>> >
>> > command-list.h: generate-cmdlist.sh command-list.txt
>> >
>> > -command-list.h: $(wildcard Documentation/git*.txt) Documentation/*config.txt Documentation/config/*.txt
>> > +command-list.h: $(wildcard Documentation/git*.txt)
>> > $(QUIET_GEN)$(SHELL_PATH) ./generate-cmdlist.sh \
>> > $(patsubst %,--exclude-program %,$(EXCLUDED_PROGRAMS)) \
>> > command-list.txt >$@+ && mv $@+ $@
>>
>> This change makes sense.
>
> I agree it looks like it's moving in the right direction, but I am
> slightly puzzled by the existing code. Why do we need to use $(wildcard)
> for git*.txt, but not for the others?
>
>> I have a not-yet-submitted patch series where I added some more
>> config/*/*.txt that wouldn't be caught by this rule, I'd updated the
>> Documentation/Makefile, but missed this part in the top-level Makefile.
>>
>> So a relation question: Does anyone actually prefer this state of
>> affairs of having a Makefile, Documentation/Makefile, t/Makefile
>> t/perf/Makefile and template/Makefile?
>>
>> It seems to me with ever-closer coupling between them that it's getting
>> to be more of a hassle to manage state between them than it would be to
>> just move them all into one big Makefile.
>
> Yes, I'm generally a fan of avoiding recursive make when we can. I think
> the caveats are:
>
> - it would be nice to continue to have stub Makefiles in
> sub-directories that trigger the main one (so "cd t && make"
> continues to work, for example).
Yeah, we should definitely keep those in place. I also wonder if various
rules for local wildcards will be more complex when they need to reach
into subdirectories.
> - we may need some cleanup of parts of the top-level Makefile which
> are triggered without dependencies (e.g., I think we unconditionally
> run some scripts to compute GIT_VERSION in the top-level; this is
> already a bit wasteful, but may get even more so as we add more
> rules from sub-directories).
>
> Mostly my argument against it (and why I haven't purused it) would be:
> it sounds like a lot of work and risk of regression, and the current
> system seems pretty fine in practice.
One edge case I discovered the other day but didn't bother debugging
much was make at the top-level failing because "doc.dep" in
Documentation/Makefile uses this pattern:
rm x &&
script >x
Which would normally work in one Makefile, but in this case two rules in
the top-level called unrelated "make -C Documentation [...]", so both of
those sub-processes end up needing to generate the doc.dep, and they
race each other.
Another thing fixed (or, worked around) with a wider application of [1].
1. https://lore.kernel.org/git/patch-3.6-96e2338ed8e-20210329T161723Z-avarab@xxxxxxxxx/
