[Bug ld/21464] relocation truncated to fit: R_OR1K_GOT16 on OpenRISC, when linking libQtGui.so.4.8.7

[giulio.benetti at micronovasrl dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=21464

--- Comment #10 from Giulio Benetti <giulio.benetti at micronovasrl dot com> ---
Il 21/03/2021 22:29, shorne at gmail dot com ha scritto:
> https://sourceware.org/bugzilla/show_bug.cgi?id=21464
> 
> --- Comment #9 from Stafford Horne <shorne at gmail dot com> ---
> On Sun, Mar 21, 2021 at 06:09:09PM +0000, giulio.benetti at micronovasrl dot
> com wrote:
>> https://sourceware.org/bugzilla/show_bug.cgi?id=21464
>>
>> --- Comment #8 from Giulio Benetti <giulio.benetti at micronovasrl dot com> 
>> ---
>> Hi Stafford,
>>
>> Il 21/03/2021 04:00, shorne at gmail dot com ha scritto:
>>> https://sourceware.org/bugzilla/show_bug.cgi?id=21464
>>>
>>> --- Comment #7 from Stafford Horne <shorne at gmail dot com> ---
>>> On Wed, Mar 17, 2021 at 10:10:13PM +0000, shorne at gmail dot com wrote:
>>>> https://sourceware.org/bugzilla/show_bug.cgi?id=21464
>>>>
>>>> --- Comment #6 from Stafford Horne <shorne at gmail dot com> ---
>>>> On Wed, Mar 17, 2021 at 09:31:59PM +0000, giulio.benetti at micronovasrl 
>>>> dot
>>>> com wrote:
>>>>> https://sourceware.org/bugzilla/show_bug.cgi?id=21464
>>>>>
>>>>> --- Comment #5 from Giulio Benetti <giulio.benetti at micronovasrl dot 
>>>>> com> ---
>>>>> Hi Nick,
>>>>>
>>>>> this bug still shows up with binutils 2.36.1 like:
>>>>> /home/giuliobenetti/git/upstream/or1k-binutils-2.36.1/host/lib/gcc/or1k-buildroot-linux-uclibc/9.3.0/../../../../or1k-buildroot-linux-uclibc/bin/ld:
>>>>> BFD (GNU Binutils) 2.36.1 assertion fail elf32-or1k.c:2377
>>>>> /home/giuliobenetti/git/upstream/or1k-binutils-2.36.1/host/lib/gcc/or1k-buildroot-linux-uclibc/9.3.0/../../../../or1k-buildroot-linux-uclibc/bin/ld:
>>>>> BFD (GNU Binutils) 2.36.1 assertion fail elf32-or1k.c:2377
>>>>> /home/giuliobenetti/git/upstream/or1k-binutils-2.36.1/host/lib/gcc/or1k-buildroot-linux-uclibc/9.3.0/../../../../or1k-buildroot-linux-uclibc/bin/ld:
>>>>> BFD (GNU Binutils) 2.36.1 assertion fail elf32-or1k.c:2377
>>>>> collect2: error: ld returned 1 exit status
>>>>>
>>>
>>> Hello,
>>>
>>> I was able to reproduce the "assertion fail elf32-or1k.c:2377" issue with my
>>> latest glibc toolchain versions, and I also did some debugging of LD to try 
>>> to
>>> get an idea of what is going on.
>>>
>>> Reproduced with:
>>>     - Just building protobuf on its own, not using buildroot, that was 
>>> giving me
>>>       some exceptions about not being able to find a thread librarty.
>>>
>>> Notes
>>>
>>> 1. The "relocation truncated to fit: R_OR1K_GOT16" error seems not be 
>>> outputted
>>>      in the latest LD.  I maybe have removed it incorrectly with some recent
>>> changes?  The
>>>      issue is not fixed.  The R_OR1K_GOT16 relocation limit means OpenRISC 
>>> fails
>>> to
>>>      run some PLT entries at run time when there are too many symbols.  I am
>>> planning
>>>      to fix this.
>>
>> It seems to be the same bug since it fails on assert(), but maybe I'm wrong.
>>
>>>      I have two ideas:
>>>       1. When we get to big plt entries over the limit will use 2 
>>> relocations to
>>>          compose the GOT offset
>>>       2. Use an extra PLT trampoline entry to add the extra offset i.e. 16k
>>>          When we go over 32k we have another trampoline, that adds 16k and 
>>> jumps
>>> to
>>>         the previous trampoline.
>>>
>>> 2. The "assertion fail elf32-or1k.c:2377" seems to be something different to
>>> me,
>>>      and not related to the count of GOT entries, maybe it should be a 
>>> different
>>> bug issue?
>>
>> Yes, it seems since...
>>
>>> 3. When debugging the "assertion fail elf32-or1k.c:2377" I see, the line we 
>>> are
>>> failing on is:
>>>
>>>          Breakpoint 1, or1k_elf_finish_dynamic_symbol (output_bfd=0x568460,
>>> info=0x543320 <link_info>, h=0x81c0f8, sym=0x7fffffffb260) at
>>> /home/shorne/work/gnu-toolchain/binutils-gdb/bfd/elf32-or1k.c:2377
>>>          2377          BFD_ASSERT (h->dynindx != -1);
>>>
>>>      So, it's asserting that the dynindx is defined, in
>>> or1k_elf_finish_dynamic_symbol before
>>>      we write out the PLT entry.
>>>
>>>      In the failure cases we see:
>>>
>>>          h->root.type bfd_link_hash_defweak AND h->forced_local
>>>
>>>      These failing symbols have been forced_local but we are still trying to
>>> write the PLT entry.
>>>      That doesn't seem to make much sense.
>>>
>>>      The forced_local has been set in _bfd_elf_link_hash_hide_symbol(), the
>>> generic
>>>      elflink code:
>>>
>>>         #0  _bfd_elf_link_hash_hide_symbol (info=0x543320 <link_info>,
>>> h=0x2bcbeb8, force_local=1) at
>>> /home/shorne/work/gnu-toolchain/binutils-gdb/bfd/elflink.c:7756
>>>         #1  0x000000000045e61e in _bfd_elf_link_assign_sym_version 
>>> (h=0x2bcbeb8,
>>> data=0x7fffffffb2f0) at
>>> /home/shorne/work/gnu-toolchain/binutils-gdb/bfd/elflink.c:2483
>>>         #2  0x000000000043055d in bfd_link_hash_traverse (htab=0x56a790,
>>> func=func@entry=0x45e33d <_bfd_elf_link_assign_sym_version>,
>>> info=info@entry=0x7fffffffb2f0)
>>>
>>>      I tried to look at other architectures and I see that most do the same
>>> assert, so
>>>      so would they also fail?
>>>
>>>      In riscv (a relatively modern implementation) they seem to just detect 
>>> this
>>> case and return FALSE.
>>>      I will try to do that and send a patch, but it will be hard to test.
>>
>> I've tried to imitate what riscv does:
>> '''
>> diff --git a/bfd/elf32-or1k.c b/bfd/elf32-or1k.c
>> index 65938e5137..74cff4bf01 100644
>> --- a/bfd/elf32-or1k.c
>> +++ b/bfd/elf32-or1k.c
>> @@ -2372,15 +2372,22 @@ or1k_elf_finish_dynamic_symbol (bfd *output_bfd,
>>          bfd_vma got_addr;
>>          Elf_Internal_Rela rela;
>>
>> -      /* This symbol has an entry in the procedure linkage table.  Set
>> -        it up.  */
>> -      BFD_ASSERT (h->dynindx != -1);
>> -
>>          splt = htab->root.splt;
>>          sgot = htab->root.sgotplt;
>>          srela = htab->root.srelplt;
>>          BFD_ASSERT (splt != NULL && sgot != NULL && srela != NULL);
>>
>> +      /* This symbol has an entry in the procedure linkage table.  Set
>> +        it up.  */
>> +      if ((h->dynindx == -1
>> +           && !((h->forced_local || bfd_link_executable (info))
>> +                && h->def_regular
>> +                && h->type == STT_GNU_IFUNC))
>> +          || splt == NULL
>> +          || sgot == NULL
>> +          || srela == NULL)
>> +        return FALSE;
>> +
>>          plt_base_addr = splt->output_section->vma + splt->output_offset;
>>          got_base_addr = sgot->output_section->vma + sgot->output_offset;
>>
>> --
>> '''
>>
>> but it still fails and without emitting anything, only:
>> collect2: error: ld returned 1 exit status
>>
>>>      Any suggestion for a test case for this?
>>>
> 
> I did some more investigation about why we are actually calling the dynamic
> finishup code for these symbols.
> 
> I found some other code that might help...
> 
> diff --git a/bfd/elf32-or1k.c b/bfd/elf32-or1k.c
> index 38406eda3d6..f1cfed230a1 100644
> --- a/bfd/elf32-or1k.c
> +++ b/bfd/elf32-or1k.c
> @@ -2566,11 +2566,10 @@ or1k_elf_adjust_dynamic_symbol (struct bfd_link_info
> *info,
>     if (h->type == STT_FUNC
>         || h->needs_plt)
>       {
> -      if (! bfd_link_pic (info)
> -         && !h->def_dynamic
> -         && !h->ref_dynamic
> -         && h->root.type != bfd_link_hash_undefweak
> -         && h->root.type != bfd_link_hash_undefined)
> +      if (h->plt.refcount <= 0
> +         || (SYMBOL_CALLS_LOCAL (info, h)
> +         || (ELF_ST_VISIBILITY (h->other) != STV_DEFAULT
> +             && h->root.type == bfd_link_hash_undefweak)))
>          {
>            /* This case can occur if we saw a PLT reloc in an input
>               file, but the symbol was never referred to by a dynamic
> 
> --
> 
> With this patch we skip the PLT calls once we determine they have been "forced
> local".  It seems to work.
> 
> -Stafford
> 

Hi Stafford,

thank you for providing such fix, it works with binutils:
- 2.32
- 2.34
- 2.35.2
- 2.36.1

while building these packages:
- protobuf
- libtheora
- zeromq

that were previously failing to link.

Thanks a lot!
Best regards
Giulio Benetti

-- 
You are receiving this mail because:
You are on the CC list for the bug.