[PATCH v2 1/2] elf: Do not over-align __FRAME_END__ in sofini.c
Adrian Ratiu
adrian.ratiu@collabora.com
Mon Nov 2 11:17:54 GMT 2026
On Mon, 02 Nov 2026, Andreas Schwab <schwab@suse.de> wrote:
> On Okt 30 2026, Florian Weimer wrote:
>
>> * Adrian Ratiu:
>>
>>> __FRAME_END__ has no explicit alignment, so on targets where GCC's
>>> DATA_ALIGNMENT raises the alignment of arrays (e.g. on aarch64,
>>> loongarch64 and mips64) it gets 8-byte alignment.
>>>
>>> The .eh_frame section of sofini.os then has sh_addralign 8 while the
>>> one of crtendS.o has 4, and GNU ld may insert 4 bytes of zero padding
>>> before the terminator. The padding reads as a zero length so the result
>>> still works, but it should not be relied upon.
>>>
>>> So we set the alignment explicitly, as libgcc's crtstuff.c does (see
>>> GCC commit 51084e13c01, "Don't overalign if DECL_USER_ALIGN").
>>>
>>> Suggested-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
>>> Signed-off-by: Adrian Ratiu <adrian.ratiu@collabora.com>
>>> ---
>>> elf/sofini.c | 8 +++++---
>>> 1 file changed, 5 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/elf/sofini.c b/elf/sofini.c
>>> index 1c526fd603..1dae5ccc67 100644
>>> --- a/elf/sofini.c
>>> +++ b/elf/sofini.c
>>> @@ -1,7 +1,9 @@
>>> /* Terminate the frame unwind info section with a 4byte 0 as a sentinel;
>>> - this would be the 'length' field in a real FDE. */
>>> -
>>> + this would be the 'length' field in a real FDE. PT_GNU_EH_FRAME only
>>> + provides the start of .eh_frame, so unwinders that scan it linearly
>>> + need the sentinel even when there is an .eh_frame_hdr. Set the alignment
>>> + explicitly so that the compiler does not over-align the array. */
>>> typedef unsigned int ui32 __attribute__ ((mode (SI)));
>>> static const ui32 __FRAME_END__[1]
>>> - __attribute__ ((used, section (".eh_frame")))
>>> + __attribute__ ((used, section (".eh_frame"), aligned (__alignof__ (ui32))))
>>> = { 0 };
>>
>> Please write 4 if you want 4-byte alignment. SImode alignment on m68k
>> is 2, I think.
>
> I don't think underalignment would be a problem.
I agree with Andreas, a lower alignment doesn't add any padding
(over-alignment padding is the problem) and GCC has been doing this for
decades without any issues.
More information about the Libc-alpha
mailing list