On Tue, May 16, 2023 at 7:54 AM Sami Tolvanen <samitolvanen@xxxxxxxxxx> wrote:
>
> On Mon, May 15, 2023 at 2:39 PM Nicolas Pitre <nico@xxxxxxxxxxx> wrote:
> >
> > On Mon, 15 May 2023, Masahiro Yamada wrote:
> >
> > > When CONFIG_TRIM_UNUSED_KSYMS is enabled, Kbuild recursively traverses
> > > the directory tree to determine which EXPORT_SYMBOL to trim. If an
> > > EXPORT_SYMBOL turns out to be unused by anyone, Kbuild begins the
> > > second traverse, where some source files are recompiled with their
> > > EXPORT_SYMBOL() tuned into a no-op.
> > >
> > > Linus stated negative opinions about this slowness in commits:
> > >
> > > - 5cf0fd591f2e ("Kbuild: disable TRIM_UNUSED_KSYMS option")
> > > - a555bdd0c58c ("Kbuild: enable TRIM_UNUSED_KSYMS again, with some guarding")
> > >
> > > We can do this better now. The final data structures of EXPORT_SYMBOL
> > > are generated by the modpost stage, so modpost can selectively emit
> > > KSYMTAB entries that are really used by modules.
> > >
> > > Commit 2cce989f8461 ("kbuild: unify two modpost invocations") is another
> > > ground-work to do this in a one-pass algorithm. With the list of modules,
> > > modpost sets sym->used if it is used by a module. modpost emits KSYMTAB
> > > only for symbols with sym->used==true.
> > >
> > > BTW, Nicolas explained why the trimming was implemented with recursion:
> > >
> > > https://lore.kernel.org/all/2o2rpn97-79nq-p7s2-nq5-8p83391473r@xxxxxxxxxxx/
> > >
> > > Actually, we never achieved that level of optimization where the chain
> > > reaction of trimming comes into play because:
> > >
> > > - CONFIG_LTO_CLANG cannot remove any unused symbols
> > > - CONFIG_LD_DEAD_CODE_DATA_ELIMINATION is enabled only for vmlinux,
> > > but not modules
> >
> > I did achieve it using LTO with gcc back then. See the section called
> > "The tree that hides the forest" of https://lwn.net/Articles/746780/ for
> > example results.
>
> Clang can do similar optimizations, but not in relocatable links where
> the linker must obviously preserve all the globals.
Yeah, the issue is not in the compiler itself
but in the way CONFIG_LTO_CLANG was implemented.
If it had been implemented in the final link stage,
it would have required LTO running three times
with CONFIG_KALLSYMS=y.
But scripts/generate_initcall_order.pl would
have been unneeded. And, maybe we would get slightly
better vmlinux.
I think the help message of CONFIG_LTO_CLANG_FULL is
a misleading advertisement.
We did not achieve such deeper trimming
that is described in this link:
https://llvm.org/docs/LinkTimeOptimization.html
If I remember correctly, GCC LTO was implemented
in the final link stage. So, trimming was depper
but it ran three times.
> A while ago there
> was a suggestion of adding an option to LLD that allows one to pass a
> list of symbols to preserve in relocatable LTO links, which would
> allow us to better optimize vmlinux.o. However, I haven't had a chance
> to look into this deeper than this proof of concept:
>
> https://reviews.llvm.org/D142163
Interesting.
But, scripts/generate_initcall_order.pl is still needed, right?
--lto-export-symbol-list is a list of symbols,
but it does not specify the correct order?
Nocolas explained the chain reaction of
compiling modules with LTO, but I suspect it
because modules are always relocatable ELF.
The LWN article (https://lwn.net/Articles/746780/) is awesome
but I think the benefit of LTO is for vmlinux,
not for modules.
>
> > > If deeper trimming is required, we need to revisit this, but I guess
> > > that is unlikely to happen.
> >
> > Would have been nicer to keep this possibility as an option. The code is
> > already there and working as intended. The build cost is intrinsic to
> > the approach of course. The actual bug is to impose that cost onto
> > people who didn't explicitly ask for it.
> >
> > But I'm no longer fighting this battle.
>
> I agree, this looks like a reasonable solution for now.
>
> Sami
--
Best Regards
Masahiro Yamada
