Named local symbols in the ELF dynamic symbol table

Named local symbols in the ELF dynamic symbol table

[Florian Weimer]

The symbol table of /lib/ld-linux-ia64.so.2 starts like this:

Symbol table [ 3] '.dynsym' contains 65 entries:
 33 local symbols  String table: [ 4] '.dynstr'
  Num:            Value   Size Type    Bind   Vis          Ndx Name
    0: 0000000000000000      0 NOTYPE  LOCAL  DEFAULT    UNDEF 
    1: 00000000000318c0    160 FUNC    LOCAL  DEFAULT       10 _dl_error_free
    2: 00000000000234c0     16 FUNC    LOCAL  DEFAULT       10 __GI__dl_debug_state
    3: 0000000000029f80    368 FUNC    LOCAL  DEFAULT       10 _dl_tls_get_addr_soft
    4: 000000000002be00   1728 FUNC    LOCAL  DEFAULT       10 _dl_open
    5: 0000000000001550     64 FUNC    LOCAL  DEFAULT       10 _start
    6: 0000000000031540    784 FUNC    LOCAL  DEFAULT       10 _dl_runtime_profile
    7: 0000000000031440    208 FUNC    LOCAL  DEFAULT       10 _dl_runtime_resolve
    8: 00000000000312c0    384 FUNC    LOCAL  DEFAULT       10 _dl_close
    9: 0000000000018240  10720 FUNC    LOCAL  DEFAULT       10 _dl_lookup_symbol_x
   10: 000000000003ab40     80 FUNC    LOCAL  DEFAULT       10 _rtld_catch_error

The interesting aspect is that these symbols are named, but are not
covered by the GNU_HASH table (because they are local).

Is there a way to get the same effect on other targets?  I want to
preserve the names of the functions that IFUNC resolvers return,
eventually improving diagnostics around IFUNC resolution.  It would best
if I wouldn't have to bloat the entire symbol table for that.

Thanks,
Florian

Re: Named local symbols in the ELF dynamic symbol table

[Nick Clifton]

Hi Florian,

> Is there a way to get the same effect on other targets?

How are the loaders for other targets linked ?  If they include the
-X or --discard-locals option on the command line, then removing that
option might achieve the result that you desire.

> It would best
> if I wouldn't have to bloat the entire symbol table for that.

I don't think that you have symbol names without having entries in
the symbol table.  Although as you say, if they are local symbols
then they will not also take up room in the GNU_HASH table.

Cheers
   Nick

Re: Named local symbols in the ELF dynamic symbol table

[Florian Weimer]

* Nick Clifton:

> Hi Florian,
>
>> Is there a way to get the same effect on other targets?
>
> How are the loaders for other targets linked ?  If they include the
> -X or --discard-locals option on the command line, then removing that
> option might achieve the result that you desire.

I'm using:

  gcc -nostdlib -nostartfiles -shared -o …/build/elf/ld.so.new \
          -Wl,-z,relro -Wl,-z,defs \
          …/build/elf/librtld.os -Wl,--version-script=…/build/ld.map \
          -Wl,-soname=ld-linux-x86-64.so.2

gcc -v tells me:

/usr/libexec/gcc/x86_64-redhat-linux/11/collect2 \
  -plugin /usr/libexec/gcc/x86_64-redhat-linux/11/liblto_plugin.so \
  -plugin-opt=/usr/libexec/gcc/x86_64-redhat-linux/11/lto-wrapper \
  -plugin-opt=-fresolution=/tmp/ccHaMg3U.res --build-id --no-add-needed \
  --eh-frame-hdr --hash-style=gnu -m elf_x86_64 -shared \
  -o …/build/elf/ld.so.new -L/usr/lib/gcc/x86_64-redhat-linux/11 \
  -L/usr/lib/gcc/x86_64-redhat-linux/11/../../../../lib64 -L/lib/../lib64 \
  -L/usr/lib/../lib64 -L/usr/lib/gcc/x86_64-redhat-linux/11/../../.. \
  -z relro -z defs …/build/elf/librtld.os --version-script=…/build/ld.map \
  -soname=ld-linux-x86-64.so.2

So it seems that -X or --discard-locals is the default?  How do I
override that, preferably for specific symbols only?

>> It would best
>> if I wouldn't have to bloat the entire symbol table for that.
>
> I don't think that you have symbol names without having entries in
> the symbol table.  Although as you say, if they are local symbols
> then they will not also take up room in the GNU_HASH table.

Yeah, should have said symbol hash table.  I really want them in .dynsym
actually.

Thanks,
Florian

Re: Named local symbols in the ELF dynamic symbol table

[Nick Clifton]

Hi Florian,

> So it seems that -X or --discard-locals is the default?

It is not clear. -{  The sources do not appear to initialise
the fields concerned with this in the link_info structure.

>  How do I
> override that, preferably for specific symbols only?

Please try -Wl,--no-strip-discarded and/or -Wl,--discard-none

Not sure what will happen....

Cheers
   Nick

Re: Named local symbols in the ELF dynamic symbol table

[Fangrui Song]

I have only seen one reason where STB_LOCAL symbols in .dynsym are useful:
unaligned absolute relocations for ppc32/ppc64
https://sourceware.org/bugzilla/show_bug.cgi?id=24896 .

If the ia64 .dynsym STB_LOCAL symbols are not used by special
relocations, it's likely a minor issue/missing improvement in ld's ia64
port.

On 2022-12-09, Florian Weimer via Binutils wrote:
>* Nick Clifton:
>
>> Hi Florian,
>>
>>> Is there a way to get the same effect on other targets?
>>
>> How are the loaders for other targets linked ?  If they include the
>> -X or --discard-locals option on the command line, then removing that
>> option might achieve the result that you desire.
>
>I'm using:
>
>  gcc -nostdlib -nostartfiles -shared -o …/build/elf/ld.so.new \
>          -Wl,-z,relro -Wl,-z,defs \
>          …/build/elf/librtld.os -Wl,--version-script=…/build/ld.map \
>          -Wl,-soname=ld-linux-x86-64.so.2
>
>gcc -v tells me:
>
>/usr/libexec/gcc/x86_64-redhat-linux/11/collect2 \
>  -plugin /usr/libexec/gcc/x86_64-redhat-linux/11/liblto_plugin.so \
>  -plugin-opt=/usr/libexec/gcc/x86_64-redhat-linux/11/lto-wrapper \
>  -plugin-opt=-fresolution=/tmp/ccHaMg3U.res --build-id --no-add-needed \
>  --eh-frame-hdr --hash-style=gnu -m elf_x86_64 -shared \
>  -o …/build/elf/ld.so.new -L/usr/lib/gcc/x86_64-redhat-linux/11 \
>  -L/usr/lib/gcc/x86_64-redhat-linux/11/../../../../lib64 -L/lib/../lib64 \
>  -L/usr/lib/../lib64 -L/usr/lib/gcc/x86_64-redhat-linux/11/../../.. \
>  -z relro -z defs …/build/elf/librtld.os --version-script=…/build/ld.map \
>  -soname=ld-linux-x86-64.so.2
>
>So it seems that -X or --discard-locals is the default?  How do I
>override that, preferably for specific symbols only?

-X should be irrelevant for .dynsym entries.

-X is not the default for almost all ports, but ld/emultempl/riscvelf.em
has a hack to default to --discard-locals. That was copied by loongarch
[Facepalm]. The proper way is to encode the logic in compiler drivers
(I did this for Clang in https://reviews.llvm.org/D127826).

>>> It would best
>>> if I wouldn't have to bloat the entire symbol table for that.
>>
>> I don't think that you have symbol names without having entries in
>> the symbol table.  Although as you say, if they are local symbols
>> then they will not also take up room in the GNU_HASH table.
>
>Yeah, should have said symbol hash table.  I really want them in .dynsym
>actually.
>
>Thanks,
>Florian
>

Re: Named local symbols in the ELF dynamic symbol table

[H.J. Lu]

On Tue, Dec 6, 2022 at 9:11 AM Florian Weimer via Binutils
<binutils@sourceware.org> wrote:
>
> The symbol table of /lib/ld-linux-ia64.so.2 starts like this:
>
> Symbol table [ 3] '.dynsym' contains 65 entries:
>  33 local symbols  String table: [ 4] '.dynstr'
>   Num:            Value   Size Type    Bind   Vis          Ndx Name
>     0: 0000000000000000      0 NOTYPE  LOCAL  DEFAULT    UNDEF
>     1: 00000000000318c0    160 FUNC    LOCAL  DEFAULT       10 _dl_error_free
>     2: 00000000000234c0     16 FUNC    LOCAL  DEFAULT       10 __GI__dl_debug_state
>     3: 0000000000029f80    368 FUNC    LOCAL  DEFAULT       10 _dl_tls_get_addr_soft
>     4: 000000000002be00   1728 FUNC    LOCAL  DEFAULT       10 _dl_open
>     5: 0000000000001550     64 FUNC    LOCAL  DEFAULT       10 _start
>     6: 0000000000031540    784 FUNC    LOCAL  DEFAULT       10 _dl_runtime_profile
>     7: 0000000000031440    208 FUNC    LOCAL  DEFAULT       10 _dl_runtime_resolve
>     8: 00000000000312c0    384 FUNC    LOCAL  DEFAULT       10 _dl_close
>     9: 0000000000018240  10720 FUNC    LOCAL  DEFAULT       10 _dl_lookup_symbol_x
>    10: 000000000003ab40     80 FUNC    LOCAL  DEFAULT       10 _rtld_catch_error
>
> The interesting aspect is that these symbols are named, but are not
> covered by the GNU_HASH table (because they are local).
>
> Is there a way to get the same effect on other targets?  I want to
> preserve the names of the functions that IFUNC resolvers return,
> eventually improving diagnostics around IFUNC resolution.  It would best
> if I wouldn't have to bloat the entire symbol table for that.
>

Are you interested in R_X86_64_IRELATIVE symbols or something
else?

-- 
H.J.

Re: Named local symbols in the ELF dynamic symbol table

[Florian Weimer]

* H. J. Lu:

> On Tue, Dec 6, 2022 at 9:11 AM Florian Weimer via Binutils
> <binutils@sourceware.org> wrote:
>>
>> The symbol table of /lib/ld-linux-ia64.so.2 starts like this:
>>
>> Symbol table [ 3] '.dynsym' contains 65 entries:
>>  33 local symbols  String table: [ 4] '.dynstr'
>>   Num:            Value   Size Type    Bind   Vis          Ndx Name
>>     0: 0000000000000000      0 NOTYPE  LOCAL  DEFAULT    UNDEF
>>     1: 00000000000318c0    160 FUNC    LOCAL  DEFAULT       10 _dl_error_free
>>     2: 00000000000234c0     16 FUNC    LOCAL  DEFAULT       10 __GI__dl_debug_state
>>     3: 0000000000029f80    368 FUNC    LOCAL  DEFAULT       10 _dl_tls_get_addr_soft
>>     4: 000000000002be00   1728 FUNC    LOCAL  DEFAULT       10 _dl_open
>>     5: 0000000000001550     64 FUNC    LOCAL  DEFAULT       10 _start
>>     6: 0000000000031540    784 FUNC    LOCAL  DEFAULT       10 _dl_runtime_profile
>>     7: 0000000000031440    208 FUNC    LOCAL  DEFAULT       10 _dl_runtime_resolve
>>     8: 00000000000312c0    384 FUNC    LOCAL  DEFAULT       10 _dl_close
>>     9: 0000000000018240  10720 FUNC    LOCAL  DEFAULT       10 _dl_lookup_symbol_x
>>    10: 000000000003ab40     80 FUNC    LOCAL  DEFAULT       10 _rtld_catch_error
>>
>> The interesting aspect is that these symbols are named, but are not
>> covered by the GNU_HASH table (because they are local).
>>
>> Is there a way to get the same effect on other targets?  I want to
>> preserve the names of the functions that IFUNC resolvers return,
>> eventually improving diagnostics around IFUNC resolution.  It would best
>> if I wouldn't have to bloat the entire symbol table for that.
>>
>
> Are you interested in R_X86_64_IRELATIVE symbols or something
> else?

Not particularly.  For ld.so --list-diagnostics, I want to iterate
through .dynsym, call all IFUNC resolvers, and report the results (for
libc, libm, libmvec at least).  To make the IFUNC resolver *result* more
meaningful, I want to map back the addresses to symbols like
__memmove_avx_unaligned_erms, but at the same time, I do not want to
make the symbol name __memmove_avx_unaligned_erms available for symbol
binding.  A local symbol in .dynsym would be a good fit here.

We can do this for IRELATIVE relocations, too, but it might be less
beneficial.

Thanks,
Florian