Makefile: let ccache cache objects built with the precompiled header - #2464

> Le 27 oct. 2026 à 08:49, Harald Nordgren via GitGitGadget <gitgitgadget@gmail.com> a écrit :
> 
> From: Harald Nordgren <haraldnordgren@gmail.com>
> 
> Since 8d3a673147 (Makefile: precompile "git-compat-util.h",
> 2026-09-15), rebuilding Git with ccache is almost as slow as building
> it without. ccache refuses to cache a compilation that uses a
> precompiled header unless its "pch_defines" and "time_macros"
> sloppiness is enabled, so it runs the compiler for nearly every object
> file, even after a "make clean" in a tree it has already built.
> 
> Both settings are safe for our build. The precompiled header is only
> pulled in with "-include" ahead of the source file, so no "#define" in
> a source file can come before it, and none of our sources use
> __DATE__, __TIME__ or __TIMESTAMP__.
> 
> Enable them for the precompiled header and the objects built with it.
> On my machine, "make clean && make -j8" with a warm cache goes from
> about 100 seconds to under 30.
> 
> With Clang, switching to a branch that changes one of the headers the
> precompiled header includes and back again still misses the cache,
> because Clang records the modification times of those headers in the
> precompiled header.
> 
> Signed-off-by: Harald Nordgren <haraldnordgren@gmail.com>
> ---
>    Makefile: let ccache cache objects built with the precompiled header
> 
>    With ccache, building after switching between two feature branches gets
>    about 6 times faster. The first build of each branch gets about 25%
>    slower while ccache fills its cache.
> 
> Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-git-2464%2FHaraldNordgren%2Fmakefile-ccache-pch-v1
> Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-git-2464/HaraldNordgren/makefile-ccache-pch-v1
> Pull-Request: https://github.com/git/git/pull/2464

I can’t comment on the above, not being a ccache user myself.

> Makefile | 2 ++
> 1 file changed, 2 insertions(+)
> 
> diff --git a/Makefile b/Makefile
> index a96be506b5..43ea5e0c40 100644
> --- a/Makefile
> +++ b/Makefile
> @@ -2946,6 +2946,8 @@ $(PRECOMPILED_HEADER_GCH): %.gch: % GIT-CFLAGS $(missing_dep_dirs) $(missing_com
> $(PRECOMPILED_HEADER_USERS): %.o: %.c $(PRECOMPILED_HEADER_GCH) GIT-CFLAGS $(missing_dep_dirs) $(missing_compdb_dir)
>    $(QUIET_CC)$(CC) -o $*.o -c -include $(PRECOMPILED_HEADER) -Winvalid-pch $(dep_args) $(compdb_args) $(ALL_CFLAGS) $(EXTRA_CPPFLAGS) $<
> 
> +$(PRECOMPILED_HEADER_GCH) $(PRECOMPILED_HEADER_USERS): export CCACHE_SLOPPINESS += pch_defines,time_macros
> +

This is a little curious to me. I’m familiar with the GNU Make-ism 
to set variables for particular targets; how does the export work?
Is it the same but with a propagation to the recipe processes?

Would it be better (for some definition…) to put the export in the 2
pre-compiled header recipes directly?

And, what about Meson?

> > @@ -2946,6 +2946,8 @@ $(PRECOMPILED_HEADER_GCH): %.gch: % GIT-CFLAGS $(missing_dep_dirs) $(missing_com
> > $(PRECOMPILED_HEADER_USERS): %.o: %.c $(PRECOMPILED_HEADER_GCH) GIT-CFLAGS $(missing_dep_dirs) $(missing_compdb_dir)
> >    $(QUIET_CC)$(CC) -o $*.o -c -include $(PRECOMPILED_HEADER) -Winvalid-pch $(dep_args) $(compdb_args) $(ALL_CFLAGS) $(EXTRA_CPPFLAGS) $<
> >
> > +$(PRECOMPILED_HEADER_GCH) $(PRECOMPILED_HEADER_USERS): export CCACHE_SLOPPINESS += pch_defines,time_macros
> > +
>
> This is a little curious to me. I’m familiar with the GNU Make-ism
> to set variables for particular targets; how does the export work?
> Is it the same but with a propagation to the recipe processes?
>
> Would it be better (for some definition…) to put the export in the 2
> pre-compiled header recipes directly?

Yeah, that could be a good idea.

> And, what about Meson?

I think it's not touched by this? My goal is just to speed up my local builds.


Harald

On Tue, Oct 27, 2026 at 4:00 PM Harald Nordgren
<haraldnordgren@gmail.com> wrote:
>
> > And, what about Meson?
>
> I think it's not touched by this? My goal is just to speed up my local builds.

Yeah, it'd generally be nice to keep the 2 on par though ;)

I don't know enough about our pch infra, though, so I'll bow out from
here. Just thought maybe we should make Meson do the same?

-- 
D. Ben Knoble

On Tue, Oct 27, 2026 at 05:57:38PM -0400, D. Ben Knoble wrote:
> On Tue, Oct 27, 2026 at 4:00 PM Harald Nordgren
> <haraldnordgren@gmail.com> wrote:
> >
> > > And, what about Meson?
> >
> > I think it's not touched by this? My goal is just to speed up my local builds.
> 
> Yeah, it'd generally be nice to keep the 2 on par though ;)
> 
> I don't know enough about our pch infra, though, so I'll bow out from
> here. Just thought maybe we should make Meson do the same?

As Meson auto-detects ccache I'd think that stuff like this should
rather be handled in Meson itself. I e.g. found [1], which feels like
the right direction to me.

But if somebody wants to implement this in our instructions meanwhile I
won't complain.

Patrick

[1]: https://github.com/mesonbuild/meson/issues/14662

> > > > And, what about Meson?
> > >
> > > I think it's not touched by this? My goal is just to speed up my local builds.
> >
> > Yeah, it'd generally be nice to keep the 2 on par though ;)
> >
> > I don't know enough about our pch infra, though, so I'll bow out from
> > here. Just thought maybe we should make Meson do the same?
>
> As Meson auto-detects ccache I'd think that stuff like this should
> rather be handled in Meson itself. I e.g. found [1], which feels like
> the right direction to me.
>
> But if somebody wants to implement this in our instructions meanwhile I
> won't complain.
>
> Patrick
>
> [1]: https://github.com/mesonbuild/meson/issues/14662

Ok, so I'll leave Meson out of this. Maybe I'll look into opening a PR
there instead.


Harald

> > > > > And, what about Meson?
> > > >
> > > > I think it's not touched by this? My goal is just to speed up my local builds.
> > >
> > > Yeah, it'd generally be nice to keep the 2 on par though ;)
> > >
> > > I don't know enough about our pch infra, though, so I'll bow out from
> > > here. Just thought maybe we should make Meson do the same?
> >
> > As Meson auto-detects ccache I'd think that stuff like this should
> > rather be handled in Meson itself. I e.g. found [1], which feels like
> > the right direction to me.
> >
> > But if somebody wants to implement this in our instructions meanwhile I
> > won't complain.
> >
> > Patrick
> >
> > [1]: https://github.com/mesonbuild/meson/issues/14662
>
> Ok, so I'll leave Meson out of this. Maybe I'll look into opening a PR
> there instead.

This has now been merged into Meson:
https://github.com/mesonbuild/meson/pull/16317


Harald