[PATCH] Skip convert lui to GPREL_L for .LANCHOR

[Fāng-ruì Sòng]

On Fri, Mar 12, 2021 at 5:51 AM Nelson Chu <nelson.chu@sifive.com> wrote:
>
> Thanks!
>
> I have given my thoughts on the bugzilla, but I think it would be good
> to get more suggestions from other maintainers, including the experts
> from llvm and lld.
> What do you think?  Please feel free to share your thoughts if you are
> interested and have times :)
>
> Thank you very much
> Nelson
>
> On Fri, Mar 12, 2021 at 11:22 AM Lifang Xia
> <lifang_xia@linux.alibaba.com> wrote:
> >
> > Hi Nelson,
> >
> > I report a bug here: https://sourceware.org/bugzilla/show_bug.cgi?id=27566
> >
> > BRs,
> > Lifang
> >
> > 在 2021/3/12 10:38, Nelson Chu 写道:
> > > Hi Lifang,
> > >
> > > Could you add the ld testcase for this?  It's better to report the
> > > details somewhere, like sourceware bugzilla or riscv github.  I find a
> > > similar fix in PR25258,
> > > https://sourceware.org/bugzilla/show_bug.cgi?id=25258.  The patch
> > > works only for c.lui relaxation, so I think we might need a similar
> > > fix for the regular lui to gp relaxation, rather than just disable the
> > > relaxation.  Anyway, we need a testcase or at least a reduced case can
> > > be reproduced.
> > >
> > > Thanks
> > > Nelson
> > >
> > > On Thu, Mar 11, 2021 at 9:05 PM Lifang Xia via Binutils
> > > <binutils@sourceware.org> wrote:
> > >> If the .LANCHOR defined in .rodata, the gp may not reach the symbol
> > >> after lang_size_relro_segment while linking with "-z relro".
> > >>
> > >> bfd/
> > >>          * elfnn-riscv.c (_bfd_riscv_relax_lui): Skip that the symbol
> > >>            is defined in rodata and linking with "-z relro".
> > >> ---
> > >>   bfd/elfnn-riscv.c | 12 ++++++++----
> > >>   1 file changed, 8 insertions(+), 4 deletions(-)
> > >>
> > >> diff --git a/bfd/elfnn-riscv.c b/bfd/elfnn-riscv.c
> > >> index 364d67b..74dce3c 100644
> > >> --- a/bfd/elfnn-riscv.c
> > >> +++ b/bfd/elfnn-riscv.c
> > >> @@ -4251,10 +4251,14 @@ _bfd_riscv_relax_lui (bfd *abfd,
> > >>        Valid gp range conservatively because of alignment issue.  */
> > >>     if (undefined_weak
> > >>         || (VALID_ITYPE_IMM (symval)
> > >> -         || (symval >= gp
> > >> -             && VALID_ITYPE_IMM (symval - gp + max_alignment + reserve_size))
> > >> -         || (symval < gp
> > >> -             && VALID_ITYPE_IMM (symval - gp - max_alignment - reserve_size))))
> > >> +         /* Skip that the symbol defined in rodata by .set, such as .LANCHOR,
> > >> +            and link with -z relro. The symbol may not be accessed with gp.  */
> > >> +         || (!(sym_sec->flags & SEC_READONLY
> > >> +               && link_info->relro)
> > >> +             && ((symval >= gp
> > >> +                  && VALID_ITYPE_IMM (symval - gp + max_alignment + reserve_size))
> > >> +                 || (symval < gp
> > >> +                     && VALID_ITYPE_IMM (symval - gp - max_alignment - reserve_size))))))
> > >>       {
> > >>         unsigned sym = ELFNN_R_SYM (rel->r_info);
> > >>         switch (ELFNN_R_TYPE (rel->r_info))
> > >> --
> > >> 2.6.0-rc0
> > >>

I haven't looked into the details.
.text < DATA_SEGMENT_ALIGN < .sdata < DATA_SEGMENT_RELRO_END

If gp relaxation requires that ld estimates the distance between the
relocated location and .sdata, being conservative and assuming
DATA_SEGMENT_ALIGN+DATA_SEGMENT_RELRO_END may take
maxpagesize+commonpagesize bytes looks good to me.

LLD avoids the DATA_SEGMENT_ALIGN/DATA_SEGMENT_RELRO_END complexity by
having 2 RW segments and making the first RW overlap with PT_GNU_RELRO
(https://reviews.llvm.org/D58892).