[llvm] r269588 - ThinLTOCodeGenerator: fix race condition found using TSAN

Kostya Serebryany via llvm-commits llvm-commits at lists.llvm.org
Mon May 16 10:44:25 PDT 2016


On Mon, May 16, 2016 at 10:35 AM, Mehdi Amini <mehdi.amini at apple.com> wrote:

>
> On May 16, 2016, at 10:30 AM, Kostya Serebryany <kcc at google.com> wrote:
>
>
>
>
> On Mon, May 16, 2016 at 10:25 AM, Mehdi Amini <mehdi.amini at apple.com>
> wrote:
>
>>
>> On May 16, 2016, at 10:22 AM, Kostya Serebryany <kcc at google.com> wrote:
>>
>> Nice!
>> What exactly did you run under than?
>>
>>
>> Bootstrap clang with ThinLTO.
>>
>> Is this is something like "ninja check-foo" (check-thin-lto?), I'd like
>> to add this to our sanitizer bot.
>>
>>
>> I plan to add a bot that performs TSAN bootstrap with ThinLTO, probably
>> in a month.
>>
>
> Excellent!
> Are you going to set up your own bot?
>
>
> Yes, I plan to add it to our jobs here: http://lab.llvm.org:8080/green/
>
>
> We already have an asan/msan bootstrap
> bot: zorg/buildbot/builders/sanitizers/buildbot_bootstrap.sh
> Although adding tsan to this bot will make it too slow, we probably better
> have a separate one.
>
>
> I noticed that TSAN has some noticeable overhead (particularly memory
> overhead)
>

For sure. TSAN's shadow is 8-to-32 and there are lots of other things it
needs to track,  i.e. a 5x-10x overhead in RAM would be normal.
If you see more -- let us know.

--kcc




>
> --
> Mehdi
>
>
>
> --kcc
>
>
>> Right is it a bit challenging. Even though we could add a few llvm-lto
>> tests that would run with threads, this would require an artificially large
>> amount of input files to exhibit the races, not sure if there is a good way
>> for that.
>>
>
>
>
>
>>
>> --
>> Mehdi
>>
>>
>>
>>
>> On Sat, May 14, 2016 at 10:49 PM, Mehdi Amini via llvm-commits <
>> llvm-commits at lists.llvm.org> wrote:
>>
>>> Author: mehdi_amini
>>> Date: Sun May 15 00:49:47 2016
>>> New Revision: 269588
>>>
>>> URL: http://llvm.org/viewvc/llvm-project?rev=269588&view=rev
>>> Log:
>>> ThinLTOCodeGenerator: fix race condition found using TSAN
>>>
>>> From: Mehdi Amini <mehdi.amini at apple.com>
>>>
>>> Modified:
>>>     llvm/trunk/lib/LTO/ThinLTOCodeGenerator.cpp
>>>
>>> Modified: llvm/trunk/lib/LTO/ThinLTOCodeGenerator.cpp
>>> URL:
>>> http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/LTO/ThinLTOCodeGenerator.cpp?rev=269588&r1=269587&r2=269588&view=diff
>>>
>>> ==============================================================================
>>> --- llvm/trunk/lib/LTO/ThinLTOCodeGenerator.cpp (original)
>>> +++ llvm/trunk/lib/LTO/ThinLTOCodeGenerator.cpp Sun May 15 00:49:47 2016
>>> @@ -884,6 +884,11 @@ void ThinLTOCodeGenerator::run() {
>>>    auto GUIDPreservedSymbols =
>>>        computeGUIDPreservedSymbols(PreservedSymbols,
>>> TMBuilder.TheTriple);
>>>
>>> +  // Make sure that every module has an entry in the ExportLists to
>>> enable
>>> +  // threaded access to this map below
>>> +  for (auto &DefinedGVSummaries : ModuleToDefinedGVSummaries)
>>> +    ExportLists[DefinedGVSummaries.first()];
>>> +
>>>    // Parallel optimizer + codegen
>>>    {
>>>      ThreadPool Pool(ThreadCount);
>>>
>>>
>>> _______________________________________________
>>> llvm-commits mailing list
>>> llvm-commits at lists.llvm.org
>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>>>
>>
>>
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160516/4befc50e/attachment.html>


More information about the llvm-commits mailing list