Hi,
Thank you for working on this!
On Sat, 9 Aug 2025 at 03:06, Andres Freund <andres@anarazel.de> wrote:
>
> Hi,
>
> As $subject says, I'd like to introduce stamp files for generated headers. The
> immediate motivation is that more than half of build.ninja just consists out
> of repetitions of the list of generated headers (listed as order-only
> dependencies). That makes invocations of ninja a tad slower and makes meson
> take a bit longer to generate build.ninja.
>
> Without the change:
>
> andres@awork3:/srv/dev/build/postgres/m-master-assert$ wc build.ninja
> 12713 260280 7754708 build.ninja
>
> andres@awork3:/srv/dev/build/postgres/m-master-assert$ numactl --physcpubind 11 time ninja reconfigure
> ...
> 3.48user 0.29system 0:03.84elapsed 97%CPU (0avgtext+0avgdata 87640maxresident)k
> 0inputs+41520outputs (0major+71323minor)pagefaults 0swaps
>
> andres@awork3:/srv/dev/build/postgres/m-master-assert$ numactl --physcpubind 11 time ninja
> ninja: no work to do.
> 0.06user 0.03system 0:00.10elapsed 98%CPU (0avgtext+0avgdata 22496maxresident)k
> 0inputs+0outputs (0major+5216minor)pagefaults 0swaps
>
>
> With the change:
>
> andres@awork3:/srv/dev/build/postgres/m-dev-assert$ wc build.ninja
> 12733 143627 3376899 build.ninja
>
> andres@awork3:/srv/dev/build/postgres/m-dev-assert$ numactl --physcpubind 11 time ninja reconfigure
> ...
> 3.14user 0.30system 0:03.53elapsed 97%CPU (0avgtext+0avgdata 74392maxresident)k
> 0inputs+34576outputs (0major+66026minor)pagefaults 0swaps
>
> andres@awork3:/srv/dev/build/postgres/m-dev-assert$ numactl --physcpubind 11 time ninja
> ninja: no work to do.
> 0.04user 0.02system 0:00.06elapsed 100%CPU (0avgtext+0avgdata 19332maxresident)k
> 0inputs+0outputs (0major+4389minor)pagefaults 0swaps
>
I got similar results.
> My main motivation however is not those, in absolute terms, relatively small
> speedups, but that I have some patches (LLVM bitcode generation, headerscheck
> support, compile_commands.json targets for headers) that add more targets
> depending on all those headers, and that the slowdown becomes more noticeable
> with those. The improvements with just this patch seems a sufficient
> improvement on its own, though - hence this thread.
>
>
> It's a bit sad to use python to create the stamp files, it's more than 3x
> slower than using 'touch' - but touch doesn't work on windows. It's also very
> hard to imagine a case where creating stamp files is a significant bottleneck,
> given that e.g. genbki.pl is somewhat expensive.
I agree.
My only concern was losing this header dependency information in the
build.ninja file but build commands of these stamp files are in the
build.ninja and they include the list of headers, so it can be easily
found.
--
Regards,
Nazir Bilal Yavuz
Microsoft