Re: [RFC] Don't create sframes during build

[Indu Bhagat]

On 1/29/26 2:23 PM, Nathan Chancellor wrote:
> On Thu, Jan 29, 2026 at 10:13:03AM +0100, Jens Remus wrote:
> > Instead of dropping .sframe for kernel during final link it would be
> > better not to generate it to save some CPU cycles and disk space.
> ...
> > What about:
> >
> > diff --git a/Makefile b/Makefile
> > --- a/Makefile
> > +++ b/Makefile
> > @@ -906,6 +906,11 @@ KBUILD_CFLAGS	+= $(call cc-option,--param=allow-store-data-races=0)
> >   KBUILD_CFLAGS	+= $(call cc-option,-fno-allow-store-data-races)
> >   endif
> >
> > +# No .sframe generation for kernel if enabled in assembler by default
> > +CC_FLAGS_SFRAME	:= $(call as-option,-Wa$(comma)--gsframe=no)
> > +KBUILD_CFLAGS	+= $(CC_FLAGS_SFRAME)
> > +KBUILD_AFLAGS	+= $(CC_FLAGS_SFRAME)
> > +
> >   ifdef CONFIG_READABLE_ASM
> >   # Disable optimizations that make assembler listings hard to read.
> >   # reorder blocks reorders the control in the function
>
> This seems like a good start for the main kernel build. I still have an
> issue with bleeding edge binutils with mismatched .sframe input
> sections, which does not appear to disappear with -Wa,--gsframe=no.
>
>    $ cat test.c
>    void atexit();
>    void main() { atexit(); }
>
>    $ gcc --version | head -1
>    gcc (GCC) 15.2.1 20260103
>
>    $ ld --version | head -1
>    GNU ld (GNU Binutils) 2.45.50.20260128
>
>    $ gcc -o /dev/null test.c
>    .../binutils/2.45.50-2026-01-27_18-02-52-e8108cc5e6fe1748dc6033a297dd2c1c6234de78/bin/ld: input SFrame sections with different format versions prevent .sframe generation
>    .../binutils/2.45.50-2026-01-27_18-02-52-e8108cc5e6fe1748dc6033a297dd2c1c6234de78/bin/ld: final link failed
>    collect2: error: ld returned 1 exit status
>
>    $ gcc -o /dev/null -Wa,--gsframe=no test.c
>    .../binutils/2.45.50-2026-01-27_18-02-52-e8108cc5e6fe1748dc6033a297dd2c1c6234de78/bin/ld: input SFrame sections with different format versions prevent .sframe generation
>    .../binutils/2.45.50-2026-01-27_18-02-52-e8108cc5e6fe1748dc6033a297dd2c1c6234de78/bin/ld: final link failed
>    collect2: error: ld returned 1 exit status
>
>    $ gcc -o /dev/null -Wl,--discard-sframe test.c
>
> This was extracted from an error I see while building libiberty in
> binutils due to the atexit() configure test failing then the fallback to
> on_exit() resulting in an error but I also notice it when building
> tools/objtool, which is a host tool:
>
>    $ make -skj"$(nproc)" ARCH=x86_64 mrproper defconfig prepare
>    .../binutils/2.45.50-2026-01-27_18-02-52-e8108cc5e6fe1748dc6033a297dd2c1c6234de78/bin/ld: input SFrame sections with different format versions prevent .sframe generation
>    .../binutils/2.45.50-2026-01-27_18-02-52-e8108cc5e6fe1748dc6033a297dd2c1c6234de78/bin/ld: final link failed
>    collect2: error: ld returned 1 exit status
>    make[5]: *** [Makefile:127: tools/objtool/objtool] Error 1
>    ...
>
> It looks like several distribution ELF objects and libraries on my
> system (Arch Linux) have .sframe sections in them, so maybe we still
> need -Wl,--discard-sframe? They all appear to be SFrame v2 objects so if
> no objects in the build have SFrame v3 sections from -Wa,--gsframe=no,
> it seems odd that there is an error but maybe there is something else
> going on?

As to whats going on:

Some Arch Linux releases seem to have glibc configured with 
--enable-sframe, which is likely the cause of the SFrame V2 objects on 
your system.  Frederick (from ArchLinux is in CC) is planning to roll 
out a release with a glibc built without --enable-sframe.

Next, GNU ld creates linker-generated .sframe for the (linker-generated) 
.plt sections (on x86_64 and s390x), if GNU ld sees any input objects 
have .sframe in them[*], unless --discard-sframe is specified.  These 
are now SFrame V3 (after Jan 15 when SFrame V3 was merged).  The choice 
to drop the support for V2/V3 object mixing in the GNU ld was made to 
save implementation complexity in linkers and (WIP) stacktracers 
supporting SFrame, by adopting SFrame V3 and above.

So, --discard-sframe is the remedy if you have V2 sections lying around.

[*] The bug in the GNU Binutils snapshots between Jan 15 and Jan 24 was 
that these linker-generated .sframe for .plt were enthusiastically 
created even when _none_ of the inputs had .sframe in them (for x86_64 
and s390x), unless --discard-sframe was specified.