[PATCH] D25560: [ELF] Emit relocation for GOT entry even if it is an absolute symbol

[Eugene Leviant via llvm-commits]

The problem is that such kind of bugs very quickly become "features" :)
Here is the real use case:

1) There is an undefined 'MySym' symbol in a shared library, which has GOT entry
2) You link this shared library using this script:

SECTIONS {
   /* .... */
   PROVIDE_HIDDEN(MySym = .);
   /* ... */
}

After library is loaded MySym will have incorrect value, because it is
actually an absolute symbol.
I think every GOT entry should be relocated, shouldn't it?



2016-10-27 17:32 GMT+03:00 Rafael Espíndola <rafael.espindola at gmail.com>:
> On 27 October 2016 at 04:42, Eugene Leviant <evgeny.leviant at gmail.com> wrote:
>>> The entry in the got should always contain 42, not 42 + load address, no?
>>
>> Looks like it shouldn't. Just try running the same test case
>> (abs-hidden.s) using ld.bfd and ld.gold.
>> Interesting fact is that ld.bfd emits two relative relocs: one for
>> .got and one for .text.
>
> So, where are you seeing this being an issue? My argument is that it
> looks like a bug in bfd/gold, no in lld.
>
> Gold produces a relocation to set the got entry to 42 + load_address.
> BFD does the same for the got, but also changes
>
> .quad foo
>
> to contain 42 + load_address at runtime. That really looks like a bug,
> if in a single assembly file we have
>
>         .quad foo
>         foo=42
>
> both gas and MC produce a file with no relocations and just the value
> 42. Why should splitting the asm file in two change it?
>
> Cheers,
> Rafael