llvm dev - Sep 2014 - [LLVMdev] [lld] A primitive program linked by lld core dumps

Hi,

I've encountered a newly introduced bug in lld today.

I am building my program with clang and lld on an X86-64 Ubuntu Linux 
machine.
Command line:
     clang main.c

The compiled program invokes the "main" function multiple times until
it
closes with a segmentation fault.
This affects even the most primitive programs like "int main() { return 
0; }".

The problem was introduced at the revision 218259 - [ELF] Fix linking 
when a regular object defines a symbol that is used in a DSO.

Would you be able to take a look at this please?

Thank you.

Kind regards,
Oleg

Thanks for reporting, I'll look into this.

Could you send me in a private message the binary attached?

Thanks,
Rafael

On Fri, Sep 26, 2014 at 10:24 AM, Oleg Ranevskyy <llvm.mail.list at
gmail.com>
wrote:
> Hi,
>
> I've encountered a newly introduced bug in lld today.
>
> I am building my program with clang and lld on an X86-64 Ubuntu Linux
> machine.
> Command line:
>     clang main.c
>
> The compiled program invokes the "main" function multiple times
until it
> closes with a segmentation fault.
> This affects even the most primitive programs like "int main() {
return 0;
> }".
>
> The problem was introduced at the revision 218259 - [ELF] Fix linking when
> a regular object defines a symbol that is used in a DSO.
>
> Would you be able to take a look at this please?
>
> Thank you.
>
> Kind regards,
> Oleg
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20140926/8c588b31/attachment.html>

Hi Oleg,

The problem is that lld incorrectly exports a dummy definition of
__tls_get_addr saying that this TLS runtime function resides at the first
address of the text section (which coincides with your main function). Your
libc (and mine doesn't, that's why this went unnoticed) requests this
symbol at link time, and the linker promptly puts this into the dynamic
symbol table.

At runtime, your program runs normally, but afterwards, when libc is
running its finalization hooks, it uses TLS storage, a mechanism that needs
the __tls_get_addr function to be correctly set up. Since it has been
incorrectly linked to your main function, it loops calling your main
function several times until it crashes.

Therefore, while the patch introduced in r218259 is correct, the problem is
that lld is incorrectly exporting __tls_get_addr. I spent this week working
on this issue (TLS relocations) and I will upload today a patch that fixes
the __tls_get_addr symbol on ELF targets, please wait a little more until
this new patch gets reviewed. Thanks for your report.

Rafael Auler

On Fri, Sep 26, 2014 at 10:32 AM, Rafael Auler <rafaelauler at gmail.com>
wrote:
> Thanks for reporting, I'll look into this.
>
> Could you send me in a private message the binary attached?
>
> Thanks,
> Rafael
>
> On Fri, Sep 26, 2014 at 10:24 AM, Oleg Ranevskyy <llvm.mail.list at
gmail.com
> > wrote:
>
>> Hi,
>>
>> I've encountered a newly introduced bug in lld today.
>>
>> I am building my program with clang and lld on an X86-64 Ubuntu Linux
>> machine.
>> Command line:
>>     clang main.c
>>
>> The compiled program invokes the "main" function multiple
times until it
>> closes with a segmentation fault.
>> This affects even the most primitive programs like "int main() {
return
>> 0; }".
>>
>> The problem was introduced at the revision 218259 - [ELF] Fix linking
>> when a regular object defines a symbol that is used in a DSO.
>>
>> Would you be able to take a look at this please?
>>
>> Thank you.
>>
>> Kind regards,
>> Oleg
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20140926/a490fce5/attachment.html>