[PATCH] D15363: [UBSan] Implement runtime suppressions (PR25066).

[Alexey Samsonov via cfe-commits]

Thank you for fixing this!

On Fri, Jan 22, 2016 at 12:30 PM, Nico Weber <thakis at chromium.org> wrote:

> I implemented Reid's suggestion in 258545. I tested that the test runs and
> passes when the DIA SDK is around (this is usually the case), and doesn't
> run when it isn't.
>
> On Thu, Jan 21, 2016 at 2:09 PM, Alexey Samsonov <vonosmas at gmail.com>
> wrote:
>
>> On Thu, Jan 21, 2016 at 10:58 AM, Reid Kleckner <rnk at google.com> wrote:
>>
>>> On Thu, Jan 21, 2016 at 10:47 AM, Nico Weber <thakis at chromium.org>
>>> wrote:
>>>
>>>> Oh, ok, let's not depend on lld then. So, we need to either determine
>>>>> if DIA SDK is available at compiler-rt configure time, or teach
>>>>> llvm-symbolizer.exe to tell us that (as we support standalone compiler-rt
>>>>> build)...
>>>>>
>>>>
>>>> Maybe the lit config file could compile a small program with clang-cl
>>>> and then run it through llvm-symbolizer and see if a symbol is returned?
>>>>
>>>
>>> I'd rather duplicate this check in compiler-rt than add auto-conf style
>>> tests to check-asan startup:
>>>   # See if the DIA SDK is available and usable.
>>>   set(MSVC_DIA_SDK_DIR "$ENV{VSINSTALLDIR}DIA SDK")
>>>   if (IS_DIRECTORY ${MSVC_DIA_SDK_DIR})
>>>     set(HAVE_DIA_SDK 1)
>>>   else()
>>>     set(HAVE_DIA_SDK 0)
>>>   endif()
>>>
>>
>> Yes. It's not 100% correct, but probably fine to assume that if we failed
>> to find DIA SDK at configure time, so did llvm-symbolizer when it was built.
>>
>>
>>>
>>> It is also possible to change LLVMSymbolizer::SymbolizePC to return
>>> false when llvm-symbolizer returns no useful information.
>>>
>>
>> Do we want LLVMSymbolizer::SymbolizePC to succeed if we fetched file/line
>> info, but not function name?
>> Also, there may well be another issue that will be hard to debug. It
>> makes sense to store information about what features of llvm-symbolizer we
>> expect to work explicitly.
>>
>>
>>> This will cause the sanitizers to try calling into dbghelp, which will
>>> probably work. It doesn't return PDB column info or work with dwarf, which
>>> is why we try llvm-symbolizer first. I think I made this change at one
>>> point, but I was asked to revert it or there were minor issues that didn't
>>> seem worth fixing.
>>>
>>
>>
>> --
>> Alexey Samsonov
>> vonosmas at gmail.com
>>
>
>


-- 
Alexey Samsonov
vonosmas at gmail.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20160122/95303b6a/attachment.html>