[lld] r320817 - Handle a VersymIndex of 0 as an error.

[Rafael Avila de Espindola via llvm-commits]

I just tested with trunk binutils compiled with
-target=mipsel-unknown-linux and gold doesn't even put _gp_disp in the
dynamic symbol table.

So from a very quick look it seems to be a bfd bug. We can probably work
around the bug by ignoring the symbol, but we should still report the
bug to the bfd developers.

Cheers,
Rafael

Simon Atanasyan <simon at atanasyan.com> writes:

> On Fri, Dec 22, 2017 at 9:27 PM, Rafael Avila de Espindola
> <rafael.espindola at gmail.com> wrote:
>> Simon Atanasyan <simon at atanasyan.com> writes:
>>
>>> Hi Rafael,
>>>
>>> Starting from this commit I get the following error in attempt to
>>> build MIPS executable:
>>> [[
>>> error: corrupt input file: version definition index 0 for symbol
>>> _gp_disp is out of bounds
>>>>>> defined in ..../code-2017.10-mti/lib/gcc/mips-mti-linux-gnu/6.3.0/../../../../mips-mti-linux-gnu/lib/mipsel-r2-hard/lib/libgcc_s.so.1
>>> ]]
>>>
>>> The "_gp_disp" symbol defined in the "libgcc_s.so.1" as "SECTION
>>> GLOBAL DEFAULT ABS". The "libgcc_s.so.1" is a part of CodeScape
>>> toolchain and linked using GNU tools.
>>>
>>> Do you think that libgcc_s.so.1 has broken version info or global
>>> section symbol is a corner case and zero VersymIndex is a correct
>>> value?
>>
>> I can't think of a useful meaning for that value. Are current bfd linkers
>> still producing that value? If so, could you provide instructions on how
>> to create a .so with that issue?
>>
>> What behavior is expected of users of that .so? Should they ignore the
>> symbol?
>
> Short description of the problem: _gp_disp symbol in a DSO generated
> by the BFD linker gets zero version definition index. This symbol is
> global while '0' means unversioned local symbol. This triggers an
> assertion in the LLD code when it links this DSO with other code. As a
> result LLD cannot link MIPS code.
>
> The _gp_disp symbol is a special symbol resolved by a linker. It
> represents offset between start of a function and GOT. Linker does not
> need to write this symbol to an output file because _gp_disp is
> useless there. But GNU linkers do that. So in the output file produced
> by GNU linkers this symbol is defined as "SECTION GLOBAL DEFAULT ABS
> _gp_disp".
>
> I do not know is this behaviour specific to _gp_disp symbol only or
> all section global absolute symbol get zero version definition index.
>
> Matthew: What do you think - is it a bug in GNU tools?
>
> Reproduction script:
> $ cat gpdisp.s
>   .global foo
>   .text
> foo:
>   lui    $t0, %hi(_gp_disp)
>   addi   $t0, $t0, %lo(_gp_disp)
>
> $ cat gpdisp.ver
> TTT_7.0.0 {
>   global:
>     foo;
> };
>
> $ mips-mti-linux-gnu-as gpdisp.s -o gpdisp.o
> $ mips-mti-linux-gnu-ld -shared gpdisp.o -o libgpdisp.so
> --version-script gpdisp.ver
>
> $ cat main.s
>   .global __start
>   .text
> __start:
>   lw     $t0, %got(foo)($gp)
>
> $ mips-mti-linux-gnu-as main.s -o main.o
> $ ld.lld main.o libgpdisp.so
> ld.lld: error: corrupt input file: version definition index 0 for
> symbol _gp_disp is out of bounds
>>>> defined in libgpdisp.so
>
> --
> Simon Atanasyan