[llvm-dev] (Thin)LTO llvm build

Tue Oct 4 09:01:16 PDT 2016

Looks like there is a bug in LLVMgold plugin -- the visibility of the weak
symbol is set to LDPV_DEFAULT instead of LDPV_HIDDEN.

David

On Tue, Oct 4, 2016 at 8:22 AM, Xinliang David Li <xinliangli at gmail.com>
wrote:

> GCC LTO works ok for the test case with both bfd and gold linker.
>
> David
>
> On Tue, Oct 4, 2016 at 6:58 AM, Teresa Johnson <tejohnson at google.com>
> wrote:
>
>>
>>
>> On Mon, Oct 3, 2016 at 6:15 PM, Teresa Johnson <tejohnson at google.com>
>> wrote:
>>
>>>
>>>
>>> On Mon, Oct 3, 2016 at 5:24 PM, Xinliang David Li <xinliangli at gmail.com>
>>> wrote:
>>>
>>>> Small repro:
>>>>
>>>> __attribute__((weak)) int hello_world();
>>>>
>>>> int test() {
>>>>   if (hello_world)
>>>>       return hello_world();
>>>>   return 0;
>>>> }
>>>>
>>>> $ clang -fuse-ld=gold  -flto=thin -O2 -shared -fPIC -o libmore.so more.c
>>>> $ objdump -t libmore.so |grep hello
>>>> 0000000000000000  w      *UND* 0000000000000000
>>>>  hello_world
>>>>
>>>> $ clang -fuse-ld=bfd  -flto=thin -O2 -shared -fPIC -o libmore.so more.c
>>>> $ objdump -t libmore.so |grep hello
>>>> 0000000000000000       *UND* 0000000000000000              hello_world
>>>>
>>>
>>> Same issue reproduces with just -flto (not just -flto=thin). So this is
>>> a general issue with ld.bfd interactions with LLVMgold.so.
>>>
>>
>> Oddly, if I do -Wl,-plugin-opt,save-temps, all of the temps files saved
>> in both cases are exactly the same, including the native object file (which
>> has a weak hello_world). So it is not that we are getting different input
>> from ld.bfd, but rather that it is treating the returned object
>> differently. And if I use the same command as above but replace the IR
>> object with the saved native object, ld.bfd behaves correctly (the final
>> .so has a weak symbol).
>>
>> I'm going to start a different thread about gold-plugin and ld.bfd
>> interactions.
>>
>> Carsten, let me know if you are able to get ld.gold used for your build,
>> as mentioned that is the better tested linker for LLVMgold.so.
>>
>> Thanks,
>> Teresa
>>
>>
>>>
>>>>
>>>> On Mon, Oct 3, 2016 at 4:40 PM, Teresa Johnson <tejohnson at google.com>
>>>> wrote:
>>>>
>>>>>
>>>>>
>>>>> On Mon, Oct 3, 2016 at 3:53 PM, Xinliang David Li <
>>>>> xinliangli at gmail.com> wrote:
>>>>>
>>>>>> What is the linker command line buidling liblldb.so? is libgcc.a
>>>>>> passed in?
>>>>>>
>>>>>
>>>>> There is no difference in the linker command for liblldb.so or
>>>>> bin/lldb between the ld.bfd and ld.gold cases, and neither links libgcc.a
>>>>> that I can see.
>>>>>
>>>>> The difference appears to be that the __morestack symbol is weak in
>>>>> the ld.gold liblldb.so case (and simply doesn't appear in the resulting
>>>>> bin/lldb presumably because libgcc.a is not linked):
>>>>>
>>>>> $ nm lib/liblldb.so.3.9.1 | grep morestack
>>>>>                  w __morestack
>>>>>
>>>>> whereas it is an undef in the ld.bfd case:
>>>>>
>>>>> $ nm lib/liblldb.so | grep morestack
>>>>>                  U __morestack
>>>>>
>>>>> resulting in the failure linking bin/lldb in that case.
>>>>>
>>>>>
>>>>>> David
>>>>>>
>>>>>> On Mon, Oct 3, 2016 at 3:52 PM, Xinliang David Li <
>>>>>> xinliangli at gmail.com> wrote:
>>>>>>
>>>>>>> In  uint64_t
>>>>>>> RTDyldMemoryManager::getSymbolAddressInProcess(const std::string
>>>>>>> &Name)  {
>>>>>>>
>>>>>>> there is reference to morestack:
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> #if defined(__i386__) || defined(__x86_64__)
>>>>>>>   // __morestack lives in libgcc, a static library.
>>>>>>>   if (&__morestack && Name == "__morestack")
>>>>>>>     return (uint64_t)&__morestack;
>>>>>>> #endif
>>>>>>> #endif // __linux__ && __GLIBC__
>>>>>>>
>>>>>>>
>>>>>>> On Mon, Oct 3, 2016 at 3:32 PM, Teresa Johnson <tejohnson at google.com
>>>>>>> > wrote:
>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> On Mon, Oct 3, 2016 at 2:59 PM, Xinliang David Li <
>>>>>>>> xinliangli at gmail.com> wrote:
>>>>>>>>
>>>>>>>>> Is -fsplit-stack option used anywhere? My wild guess is that with
>>>>>>>>> ld.bfd, the thinLTO link for the DSO does not bring in morestack.o from
>>>>>>>>> libgcc.a, but the hidden symbol is defined in lldb binary.
>>>>>>>>>
>>>>>>>>
>>>>>>>> AFAICT, no - I had done "ninja -v" so I have all of the
>>>>>>>> intermediate build commands, and it doesn't show up in that.
>>>>>>>>
>>>>>>>> I'll have to do some more digging to figure out why it is
>>>>>>>> referenced from liblldb.so in the ld.bfd case and not when using ld.gold
>>>>>>>>
>>>>>>>> Teresa
>>>>>>>>
>>>>>>>> David
>>>>>>>>>
>>>>>>>>> On Mon, Oct 3, 2016 at 1:54 PM, Teresa Johnson via llvm-dev <
>>>>>>>>> llvm-dev at lists.llvm.org> wrote:
>>>>>>>>>
>>>>>>>>>> Aha - finally reproduced! The difference is using ld.bfd not
>>>>>>>>>> ld.gold. With that I get the same failure (using 3.9 to build 3.9 sources):
>>>>>>>>>>
>>>>>>>>>> /usr/local/google/home/tejohnson/binutils_build/install/bin/ld:
>>>>>>>>>> bin/lldb-3.9.1: hidden symbol `__morestack' in
>>>>>>>>>> /usr/lib/gcc/x86_64-linux-gnu/4.8/libgcc.a(morestack.o) is
>>>>>>>>>> referenced by DSO
>>>>>>>>>> /usr/local/google/home/tejohnson/binutils_build/install/bin/ld:
>>>>>>>>>> final link failed: Bad value
>>>>>>>>>> clang-3.9: error: linker command failed with exit code 1 (use -v
>>>>>>>>>> to see invocation)
>>>>>>>>>> ninja: build stopped: subcommand failed.
>>>>>>>>>>
>>>>>>>>>> I am not sure what the official support story is for LLVMgold.so
>>>>>>>>>> and ld.bfd. As mentioned earlier, the LLVM site indicates it should use the
>>>>>>>>>> gold linker. Can you use that while I try to figure out whether this is
>>>>>>>>>> something that should be supported/working?
>>>>>>>>>>
>>>>>>>>>> Thanks,
>>>>>>>>>> Teresa
>>>>>>>>>>
>>>>>>>>>> On Mon, Oct 3, 2016 at 9:56 AM, Carsten Mattner <
>>>>>>>>>> carstenmattner at gmail.com> wrote:
>>>>>>>>>>
>>>>>>>>>>> On Mon, Oct 3, 2016 at 3:50 PM, Teresa Johnson <
>>>>>>>>>>> tejohnson at google.com> wrote:
>>>>>>>>>>> >
>>>>>>>>>>> > On Sun, Oct 2, 2016 at 4:02 AM, Carsten Mattner <
>>>>>>>>>>> carstenmattner at gmail.com> wrote:
>>>>>>>>>>> >>
>>>>>>>>>>> >> On Sun, Oct 2, 2016 at 6:41 AM, Teresa Johnson <
>>>>>>>>>>> tejohnson at google.com> wrote:
>>>>>>>>>>> > >
>>>>>>>>>>> > > > I use trunk, but it depends on how close to the bleeding
>>>>>>>>>>> edge you
>>>>>>>>>>> > > > are comfortable with. But like I said, I also tried
>>>>>>>>>>> bootstrapping
>>>>>>>>>>> > > > with 3.9 (both trunk as well as 3.9 sources) and couldn't
>>>>>>>>>>> reproduce.
>>>>>>>>>>> >>
>>>>>>>>>>> >> Hmm, so you're saying neither fails for you, right?
>>>>>>>>>>> >
>>>>>>>>>>> > Yes
>>>>>>>>>>>
>>>>>>>>>>> Okay. Do you mind focusing on the 3.9 branch? It's less of a
>>>>>>>>>>> moving target
>>>>>>>>>>> and lends itself more to figuring out what's failing for me.
>>>>>>>>>>>
>>>>>>>>>>> As soon as I get around to it, I'll send you my full configure
>>>>>>>>>>> and build
>>>>>>>>>>> steps as a shell script. Need to streamline it.
>>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> --
>>>>>>>>>> Teresa Johnson |  Software Engineer |  tejohnson at google.com |
>>>>>>>>>> 408-460-2413
>>>>>>>>>>
>>>>>>>>>> _______________________________________________
>>>>>>>>>> LLVM Developers mailing list
>>>>>>>>>> llvm-dev at lists.llvm.org
>>>>>>>>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> --
>>>>>>>> Teresa Johnson |  Software Engineer |  tejohnson at google.com |
>>>>>>>> 408-460-2413
>>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>>
>>>>> --
>>>>> Teresa Johnson |  Software Engineer |  tejohnson at google.com |
>>>>> 408-460-2413
>>>>>
>>>>
>>>>
>>>
>>>
>>> --
>>> Teresa Johnson |  Software Engineer |  tejohnson at google.com |
>>> 408-460-2413
>>>
>>
>>
>>
>> --
>> Teresa Johnson |  Software Engineer |  tejohnson at google.com |
>> 408-460-2413
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20161004/a1cf0adc/attachment-0001.html>