On Wed, Jan 19, 2022 at 1:31 PM Ingo Molnar <mingo@xxxxxxxxxx> wrote:
>
> * Arnd Bergmann <arnd@xxxxxxxx> wrote:
>
> > > I tried to avoid as many low level headers as possible from the main
> > > types headers - and the get_order() functionality also brings in bitops
> > > definitions, which I'm still hoping to be able to reduce from its
> > > current ~95% utilization in a distro kernel ...
> >
> > Agreed, I think reducing bitops.h and atomic.h usage is fairly important,
> > I think these are even bigger on arm64 than on x86.
>
> So what I'm using for 'header complexity metrics' is rather simple: passing
> -P -H to the preprocessor: stripping comments & not generating
> line-markers, and then counting linecount.
>
> Line-markers should *probably* remain, because the real build is generatinginclude/linux/mm_page_address.h
> them too - but I wanted to gain a crude & easily available metric to
> measure 'first-pass parsing complexity'. That's I think where most of the
> header bloat is concentrated: later passes don't really get any of the
> unused header definitions passed along. (But maybe this is an invalid
> assumption, because compiler warnings do get generated by later passes, and
> they are generated for mostly-unused header inlines too.)
>
> If we include comments & line-markers then the bloat goes up by another
> ~2x:
>
> kepler:~/mingo.tip.git> ./st include/linux/sched.h
> #include <linux/sched.h> | LOC: 2,186 | headers: 118
> kepler:~/mingo.tip.git> ./st include/linux/sched.h
> #include <linux/sched.h> | LOC: 4,092 | headers: 0
The metric I've been focusing on is bytes of the preprocessed header, which
is more sensitive to function definitions that get generated from macros,
and I multiply this by the number of inclusions (from scanning the
.file.o.cmd files). It probably helps to have a couple of metrics and look
at all of them occasionally to not miss something important.
In the meantime, I have made some progress on reducing the headers
for arm64, on top of your tree from Jan 8, but I have not looked at
later changes from your side, and I need to work on this a bit more
to ensure this doesn't break other architectures.
For an arm64 allmodconfig build, my additional improvements on top
of yours are significant but not as good as I had hoped for, this
can still improve I hope:
5.16-rc8-vanilla 32640 seconds user, 3286 seconds sys
5.16-rc8-mingo 22990 seconds user, 2304 seconds sys
5.16-rc8-arnd 19007 seconds user, 1853 seconds sys
As my tree builds any randconfig cleanly, I keep looking at
different configs and find that this has a big impact, some options
end up eliminating most of the benefits until I add further changes
to clean up certain files. This happened with kasan, kprobes,
and lse-atomics for instance. After eliminating all circular includes,
I was also able to revisit my old script to visualize the inclusions,
see[1] for the current arm64 defconfig output. This version uses
my arbitrary metric as font-size, and uses labels for the number
of inclusions.
Arnd
[1] https://drive.google.com/file/d/1wbs252I8LyswscBAeV3SpjBG2AGoBnB8/view?usp=sharing
