[PATCH] Makefile: fix u-boot-initial-env target if lto is enabled

[Max Krummenacher]

Hi all

Thanks for the feedback

On Tue, Oct 18, 2022 at 8:23 PM Tom Rini <trini at konsulko.com> wrote:
>
> On Tue, Oct 18, 2022 at 08:19:23PM +0200, Pali Rohár wrote:
> > On Tuesday 18 October 2022 14:17:23 Tom Rini wrote:
> > > On Tue, Oct 18, 2022 at 08:06:27PM +0200, Pali Rohár wrote:
> > > > On Tuesday 18 October 2022 14:04:46 Tom Rini wrote:
> > > > > On Tue, Oct 18, 2022 at 08:03:31PM +0200, Pali Rohár wrote:
> > > > > > On Tuesday 18 October 2022 19:48:27 Max Krummenacher wrote:
> > > > > > > From: Max Krummenacher <max.krummenacher at toradex.com>
> > > > > > >
> > > > > > > With LTO enabled the U-Boot initial environment is no longer stored
> > > > > > > in an easy accessible section in env/common.o. I.e. the section name
> > > > > > > changes from build to build, its content maybe compressed and it is
> > > > > > > annotated with additional data.
> > > > > > >
> > > > > > > For GCC adding the option '-ffat-lto-objects' when compiling common.o
> > > > > > > adds additionally the traditional sections in the object file and
> > > > > > > 'make u-boot-initial-env' would work also for the LTO enabled case.
> > > > > > > However clang doesn't have that option.
> > > > > > >
> > > > > > > Fix this by recompiling common.o into a object file only used for
> > > > > > > the creation of u-boot-initial-env if LTO is enabled.
> > > > > > >
> > > > > > > See also:
> > > > > > > https://lore.kernel.org/all/927b122e-1f62-e790-f5ca-30bae4332c77@foss.st.com/
> > > > > > >
> > > > > > > Signed-off-by: Max Krummenacher <max.krummenacher at toradex.com>
> > > > > > >
> > > > > > > ---
> > > > > > >
> > > > > > >  Makefile | 8 ++++++++
> > > > > > >  1 file changed, 8 insertions(+)
> > > > > > >
> > > > > > > diff --git a/Makefile b/Makefile
> > > > > > > index 3866cc62f9a..cd45c720d23 100644
> > > > > > > --- a/Makefile
> > > > > > > +++ b/Makefile
> > > > > > > @@ -2451,9 +2451,17 @@ endif
> > > > > > >     $(Q)$(MAKE) -f $(srctree)/scripts/Makefile.modpost
> > > > > > >
> > > > > > >  quiet_cmd_genenv = GENENV  $@
> > > > > > > +ifeq ($(LTO_ENABLE),y)
> > > > > > > +cmd_genenv = $(CC) $(filter-out $(LTO_CFLAGS),$(c_flags)) -c -o env/initial_env.o env/common.c; \
> > > > > > > +   $(OBJCOPY) --dump-section .rodata.default_environment=$@ env/initial_env.o; \
> > > > > > > +   sed --in-place -e 's/\x00/\x0A/g' $@; sed --in-place -e '/^\s*$$/d' $@; \
> > > > > > > +   sort --field-separator== -k1,1 --stable $@ -o $@; \
> > > > > > > +   rm -f env/initial_env.o env/initial_env.su
> > > > > > > +else
> > > > > > >  cmd_genenv = $(OBJCOPY) --dump-section .rodata.default_environment=$@ env/common.o; \
> > > > > >
> > > > > > This code is still broken because in some cases section name is not RO.
> > > > >
> > > > > Wait, when does that happen?
> > > >
> > > > E.g. for mvebu_espressobin-88f3720_defconfig
> > >
> > > Erm, ugh. I see 44be835d25ba ("arm: mvebu: Espressobin: Set default
> > > value for $ethNaddr env variable") and c4df0f6f315c ("arm: mvebu:
> > > Espressobin: Set default value for $fdtfile env variable") I guess we
> > > couldn't solve this any other way? The platforms aren't unique in
> > > needing / wanting to set MAC or fdtfile variables.
> >
> > Yes, we can solve it. Marek was working on solution for setting default
> > variables at runtime but seems it is not finished yet.
>
> OK, then lets not assume DEFAULT_ENV_IS_RW is something we need to worry
> about long term.

The patch was intended to fix u-boot-initial-env make target for cases
where it worked
without 'CONFIG_LTO' to also work if that config gets enabled.

Any use case where it already fails without 'CONFIG_LTO' set will not benefit.

I agree that a variant of tools/env/fw_printenv compiled for the build host may
solve additional issues present. However I will not be able to work on such a
solution in the near future.

Max

>
> --
> Tom