Re: [PATCH] RISC-V: fix linker message when relaxation deletes bytes

[Nelson Chu <nelson@rivosinc.com>]

On Fri, Oct 14, 2022 at 3:24 PM Clément Chigot <chigot@adacore.com> wrote:
>
> On Fri, Oct 14, 2022 at 7:08 AM Nelson Chu <nelson@rivosinc.com> wrote:
> >
> > Hi,
> >
> > On Thu, Oct 13, 2022 at 8:06 PM Clément Chigot via Binutils
> > <binutils@sourceware.org> wrote:
> > >
> > > Gentle ping
> > >
> > > Thanks,
> > > Clément
> > >
> > > On Thu, Oct 6, 2022 at 1:46 PM Clément Chigot <chigot@adacore.com> wrote:
> > > >
> > > > The section relaxation can delete some bytes resulting in the symbols'
> > > > value being modified.
> > > > As the linker messages retrieve a symbol information using the
> > > > outsymbols field of abfd, it must be updated as well.
> >
> > Interesting, I never noticed this.  If this is necessary, then other
> > targets should also update the output symbols when relaxing (nds32 and
> > sh), but it seems like they don't do that as well.
>
> Yeah, I agree there are probably other targets having a similar issue.
> I don't know at what extend outsymbols is being used for. But the fact
> that is being filled with "untouched" symbols in object files looks
> weird to me. Maybe there is something to look into this direction.
> For now, this patch aims just to fix ld-undefined on Risc-V targets.
>
> > > > bfd/ChangeLog:
> > > >
> > > >         * elfnn-riscv.c (riscv_relax_delete_bytes): Update
> > > >         abfd->outsymbols.
> > > >
> > > > ld/ChangeLog:
> > > >
> > > >         * testsuite/ld-riscv-elf/ld-riscv-elf.exp: New test.
> > > >         * testsuite/ld-riscv-elf/undefined-align.d: New test.
> > > >         * testsuite/ld-riscv-elf/undefined-align.s: New test.
> > > > ---
> > > >  bfd/elfnn-riscv.c                           | 23 +++++++++++++++++++++
> > > >  ld/testsuite/ld-riscv-elf/ld-riscv-elf.exp  |  1 +
> > > >  ld/testsuite/ld-riscv-elf/undefined-align.d |  5 +++++
> > > >  ld/testsuite/ld-riscv-elf/undefined-align.s | 10 +++++++++
> > > >  4 files changed, 39 insertions(+)
> > > >  create mode 100644 ld/testsuite/ld-riscv-elf/undefined-align.d
> > > >  create mode 100644 ld/testsuite/ld-riscv-elf/undefined-align.s
> > > >
> > > > diff --git a/bfd/elfnn-riscv.c b/bfd/elfnn-riscv.c
> > > > index 3d2ddf4e651..7a6b66dcd8a 100644
> > > > --- a/bfd/elfnn-riscv.c
> > > > +++ b/bfd/elfnn-riscv.c
> > > > @@ -4060,6 +4060,7 @@ riscv_relax_delete_bytes (bfd *abfd,
> > > >    unsigned int sec_shndx = _bfd_elf_section_from_bfd_section (abfd, sec);
> > > >    struct bfd_elf_section_data *data = elf_section_data (sec);
> > > >    bfd_byte *contents = data->this_hdr.contents;
> > > > +  asymbol **outsyms = bfd_get_outsymbols (abfd);
> > > >
> > > >    /* Actually delete the bytes.  */
> > > >    sec->size -= count;
> > > > @@ -4158,6 +4159,28 @@ riscv_relax_delete_bytes (bfd *abfd,
> > > >         }
> > > >      }
> > > >
> > > > +  /* As linker messages are getting symbols through outsymbols field of abfd,
> > > > +     it must be adjusted too.  */
> > > > +  if (outsyms == NULL)
> > > > +    {
> > > > +      if (!bfd_generic_link_read_symbols (abfd))
> > > > +       link_info->callbacks->einfo (_("%F%P: %pB: could not read symbols: %E\n"), abfd);
> > > > +      outsyms = bfd_get_outsymbols (abfd);
> > > > +    }
> > > > +
> > > > +  for (i = 0; i < bfd_get_symcount (abfd); i++)
> > > > +    {
> > > > +      asymbol *sym = outsyms[i];
> > > > +
> > > > +      if (sym->section != sec)
> > > > +       continue;
> > > > +
> > > > +      /* If the symbol is in the range of memory we just moved, we
> > > > +        have to adjust its value.  */
> > > > +      if (sym->value > addr && sym->value <= toaddr)
> > > > +       sym->value -= count;
> > > > +    }
> > > > +
> > > >    return true;
> > > >  }
> > > >
> > > > diff --git a/ld/testsuite/ld-riscv-elf/ld-riscv-elf.exp b/ld/testsuite/ld-riscv-elf/ld-riscv-elf.exp
> > > > index df89e0ee68b..86f1d05bc08 100644
> > > > --- a/ld/testsuite/ld-riscv-elf/ld-riscv-elf.exp
> > > > +++ b/ld/testsuite/ld-riscv-elf/ld-riscv-elf.exp
> > > > @@ -176,6 +176,7 @@ if [istarget "riscv*-*-*"] {
> > > >         [list "Weak reference 64" "-T weakref.ld -m[riscv_choose_lp64_emul]" "" \
> > > >             "-march=rv64i -mabi=lp64" {weakref64.s} \
> > > >             {{objdump -d weakref64.d}} "weakref64"]]
> > > > +    run_dump_test "undefined-align"
> > > >
> > > >      # The following tests require shared library support.
> > > >      if ![check_shared_lib_support] {
> > > > diff --git a/ld/testsuite/ld-riscv-elf/undefined-align.d b/ld/testsuite/ld-riscv-elf/undefined-align.d
> > > > new file mode 100644
> > > > index 00000000000..c8cbb84ac7c
> > > > --- /dev/null
> > > > +++ b/ld/testsuite/ld-riscv-elf/undefined-align.d
> > > > @@ -0,0 +1,5 @@
> > > > +#name: undefined with alignment
> > > > +#source: undefined-align.s
> > > > +#as:
> > > > +#ld: --relax
> > > > +#error: \A[^\n]*\.o: in function `_start':\n\(\.text\+0x0\): undefined reference to `foo'\Z
> > > > diff --git a/ld/testsuite/ld-riscv-elf/undefined-align.s b/ld/testsuite/ld-riscv-elf/undefined-align.s
> > > > new file mode 100644
> > > > index 00000000000..577557c663a
> > > > --- /dev/null
> > > > +++ b/ld/testsuite/ld-riscv-elf/undefined-align.s
> > > > @@ -0,0 +1,10 @@
> > > > +# Make sure that the linker messages take into account the modification
> > > > +# of a symbol's value made during the section relaxation.
> > > > +# Here, "_start" will have an offset made by ".align 4" which will be
> > > > +# removed during the relaxation of R_RISCV_ALIGN.
> > > > +       .text
> > > > +       .align  4
> > > > +       .globl  _start
> > > > +       .type   _start, @function
> > > > +_start:
> > > > +       call    foo
> >
> > Tested with undefined-align.s, I get the following error message
> > before applying this patch,
> >
> > % riscv64-unknown-elf-ld tmp.o
> >
> > riscv64-unknown-elf-ld: tmp.o:(.text+0x0): undefined reference to `foo'
> >
> > And get the error after applying this patch,
> >
> > % riscv64-unknown-elf-ld tmp.o
> >
> > riscv64-unknown-elf-ld: tmp.o: in function `_start':
> >
> > (.text+0x0): undefined reference to `foo'
> >
> > It reports more detail, but I cannot see if the removed offset may
> > cause the wrong message?
>
> I'm not sure I understand what you mean.
> As in the object file, _start starts at 0xc, you mean that the error
> message should be something like:
>   |  riscv64-unknown-elf-ld: tmp.o: in function `_start':
>   |  (.text+0xc): undefined reference to `foo'
> ?

The error message doesn't seem wrong without your patch, just don't
report "in function `_start':".  The comment in your testcase said the
offset should be removed after handling alignment, but I get the same
zero offset with or without your patch.  It would be great if you can
provide other cases to show how the offset reported by error message
will be wrong.

Thanks
Nelson

> Thanks,
> Clément