[clang] 23058f9 - [OPENMP]Do not use RTTI by default for NVPTX devices.

Wed Jan 15 15:47:24 PST 2020

On Wed, Jan 15, 2020 at 3:09 PM Alexey Bataev <a.bataev at outlook.com> wrote:

> And I disabled it only for device side, which is NVPTX, no? Can host side
> target class report that the target is NVPTX? If you look at the patch, it
> disable RTTI only if current triple is NVPTX. Can it be true for the host?
>
You are correct that the NVPTX as the triple is never used for the host
compilation.

However, I'm talking about host-side *code* as seen by compiler during
device-side *compilation*. Keep in mind that both host and device
compilations always see both sides of the source code. The host code seen
by the compiler during device-side compilation should be allowed to use
host features that are not available on GPU.

When you RTTI got disabled for NVPTX, it affected *all* code seen by the
compiler. When GPU-side compiler that targets NVPTX with RTTI disabled gets
to parse the host function hf() (https://godbolt.org/z/Gzsrof) it sees
typeid() and complains that RTTI is disabled, even though it will never
produce any code for hf(). It should've allowed RTTI use in the host-only
code.

--Artem

> -------------
> Best regards,
> Alexey Bataev
>
> 15.01.2020 1:06 PM, Artem Belevich пишет:
>
>
>
> On Wed, Jan 15, 2020 at 2:52 PM Alexey Bataev <a.bataev at outlook.com>
> wrote:
>
>> 1. The problem is that it does not produce errors,
>>
> ATM, it does produce errors when it's disabled.
>
>> it leads to the emission of some declaration that cannot be resolved by
>> the linker. This what I was trying to avoid.
>>
> I'm OK with disabling it on device, but it should be done in a way to keep
> RTTI in the host code working.
>
> E.g. this code should compile: https://godbolt.org/z/Gzsrof
> It still has your original change, so clang compilation fails. It does not
> with NVCC.
>
> 2. Yes, I did not disable it on the host side, just for the device side. I
>> disabled it only for NVPTX target, which is the device.
>>
> It's fine to disable it for code generated for the target. However, the
> front-end needs to be able to deal with the host-side code where it is OK
> to use RTTI.
>
> --Artem
>
>
>
>
>> -------------
>> Best regards,
>> Alexey Bataev
>>
>> 15.01.2020 12:49 PM, Artem Belevich пишет:
>>
>> Thank you.
>>
>> In general, RTTI should probably be treated similar to how we deal with
>> inline assembly and ignore errors if they are in the code that we're not
>> going to codegen during this side of compilation. E.g. during host-side
>> compilation we don't complain about GPU-side registers in inline assembly
>> that x86 target is not aware of.
>>
>> Disabling RTTI altogether on device side makes it impossible to use in
>> any host-side CUDA because RTTI code will be seen by the device-side
>> compiler, which will promptly fail, which is exactly what happened in this
>> case -- host-side template that relied on RTTI failed during device-side
>> compilation:
>>
>>
>> https://github.com/protocolbuffers/protobuf/blob/master/src/google/protobuf/arena.h#L536
>> google/protobuf/arena.h:536:15: error: use of typeid requires -frtti
>>     AllocHook(RTTI_TYPE_ID(T), n);
>> ...
>> 2 errors generated when compiling for sm_60.
>>
>> --Artem
>>
>> On Wed, Jan 15, 2020 at 2:36 PM Alexey Bataev <a.bataev at outlook.com>
>> wrote:
>>
>>> Sure, will revert it ASAP.
>>>
>>> -------------
>>> Best regards,
>>> Alexey Bataev
>>>
>>> 15.01.2020 12:26 PM, Artem Belevich пишет:
>>>
>>> Alexey,
>>>
>>> This breaks compilation of our cuda code which happens to transitively
>>> include protobuf headers.
>>> Can you, please, revert it for now until we figure out how RTTI should
>>> be handled?
>>>
>>> --Artem
>>>
>>> On Tue, Jan 14, 2020 at 3:15 PM Alexey Bataev via cfe-commits <
>>> cfe-commits at lists.llvm.org> wrote:
>>>
>>>>
>>>> Author: Alexey Bataev
>>>> Date: 2020-01-14T18:12:06-05:00
>>>> New Revision: 23058f9dd4d7e18239fd63b6da52549514b45fda
>>>>
>>>> URL:
>>>> https://github.com/llvm/llvm-project/commit/23058f9dd4d7e18239fd63b6da52549514b45fda
>>>> DIFF:
>>>> https://github.com/llvm/llvm-project/commit/23058f9dd4d7e18239fd63b6da52549514b45fda.diff
>>>>
>>>> LOG: [OPENMP]Do not use RTTI by default for NVPTX devices.
>>>>
>>>> NVPTX does not support RTTI, so disable it by default.
>>>>
>>>> Added:
>>>>     clang/test/Driver/openmp-offload-gpu.cpp
>>>>
>>>> Modified:
>>>>     clang/lib/Driver/ToolChain.cpp
>>>>
>>>> Removed:
>>>>
>>>>
>>>>
>>>>
>>>> ################################################################################
>>>> diff  --git a/clang/lib/Driver/ToolChain.cpp
>>>> b/clang/lib/Driver/ToolChain.cpp
>>>> index cab97b1a601a..3ebbd30195b3 100644
>>>> --- a/clang/lib/Driver/ToolChain.cpp
>>>> +++ b/clang/lib/Driver/ToolChain.cpp
>>>> @@ -68,7 +68,8 @@ static ToolChain::RTTIMode CalculateRTTIMode(const
>>>> ArgList &Args,
>>>>    }
>>>>
>>>>    // -frtti is default, except for the PS4 CPU.
>>>> -  return (Triple.isPS4CPU()) ? ToolChain::RM_Disabled :
>>>> ToolChain::RM_Enabled;
>>>> +  return (Triple.isPS4CPU() || Triple.isNVPTX()) ?
>>>> ToolChain::RM_Disabled
>>>> +                                                 :
>>>> ToolChain::RM_Enabled;
>>>>  }
>>>>
>>>>  ToolChain::ToolChain(const Driver &D, const llvm::Triple &T,
>>>>
>>>> diff  --git a/clang/test/Driver/openmp-offload-gpu.cpp
>>>> b/clang/test/Driver/openmp-offload-gpu.cpp
>>>> new file mode 100644
>>>> index 000000000000..9da7308506ae
>>>> --- /dev/null
>>>> +++ b/clang/test/Driver/openmp-offload-gpu.cpp
>>>> @@ -0,0 +1,20 @@
>>>> +///
>>>> +/// Perform several driver tests for OpenMP offloading
>>>> +///
>>>> +
>>>> +// REQUIRES: clang-driver
>>>> +// REQUIRES: x86-registered-target
>>>> +// REQUIRES: powerpc-registered-target
>>>> +// REQUIRES: nvptx-registered-target
>>>> +
>>>> +///
>>>> ###########################################################################
>>>> +
>>>> +/// PTXAS is passed -c flag by default when offloading to an NVIDIA
>>>> device using OpenMP
>>>> +/// Check that the flag is passed when -fopenmp-relocatable-target is
>>>> used.
>>>> +// RUN:   %clangxx -### -fopenmp=libomp
>>>> -fopenmp-targets=nvptx64-nvidia-cuda \
>>>> +// RUN:          -save-temps -no-canonical-prefixes %s -x c++ -c 2>&1 \
>>>> +// RUN:   | FileCheck -check-prefix=CHK-RTTI %s
>>>> +
>>>> +// CHK-RTTI: clang{{.*}}" "-triple" "nvptx64-nvidia-cuda"
>>>> +// CHK-RTTI-SAME: "-fno-rtti"
>>>> +
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> cfe-commits mailing list
>>>> cfe-commits at lists.llvm.org
>>>> https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
>>>>
>>>
>>>
>>> --
>>> --Artem Belevich
>>>
>>>
>>
>> --
>> --Artem Belevich
>>
>>
>
> --
> --Artem Belevich
>
>

-- 
--Artem Belevich
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20200115/dab9d6db/attachment-0001.html>