nvcc cannot handle clang builtin

[ye-luo]

After merging #111516, I started to see

$ nvcc -allow-unsupported-compiler -ccbin=clang++ main.cu
/usr/lib/gcc/x86_64-linux-gnu/13/../../../../include/c++/13/type_traits(2967): error: type name is not allowed
          bool _Dangle = __reference_converts_from_temporary(_Tp, _Res_t)
                                                             ^

/usr/lib/gcc/x86_64-linux-gnu/13/../../../../include/c++/13/type_traits(2967): error: type name is not allowed
          bool _Dangle = __reference_converts_from_temporary(_Tp, _Res_t)
                                                                  ^

/usr/lib/gcc/x86_64-linux-gnu/13/../../../../include/c++/13/type_traits(2967): error: identifier "__reference_converts_from_temporary" is undefined
          bool _Dangle = __reference_converts_from_temporary(_Tp, _Res_t)
                         ^

3 errors detected in the compilation of "main.cu".

The error comes out of

#$ cudafe++ --c++17 --clang --clang_version=200000 --display_error_number --orig_src_file_name "main.cu" --orig_src_path_name "/home/yeluo/main.cu" --unicode_source_kind=UTF-8 --allow_managed  --m64 --parse_templates --gen_c_file_name "/tmp/tmpxft_000490f6_00000000-6_main.cudafe1.cpp" --stub_file_name "tmpxft_000490f6_00000000-6_main.cudafe1.stub.c" --gen_module_id_file --module_id_file_name "/tmp/tmpxft_000490f6_00000000-4_main.module_id" "/tmp/tmpxft_000490f6_00000000-5_main.cpp4.ii" 

main.cu

#include <type_traits>
int main() {}

My clang setting

$ clang++ -v
clang version 20.0.0git (https://github.com/llvm/llvm-project.git ed7868de03c7b93809f87ed1a01103b926564feb)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /soft/llvm/main-patched/bin
Build config: +assertions
Found candidate GCC installation: /usr/lib/gcc/x86_64-linux-gnu/11
Found candidate GCC installation: /usr/lib/gcc/x86_64-linux-gnu/13
Selected GCC installation: /usr/lib/gcc/x86_64-linux-gnu/13
Candidate multilib: .;@m64
Candidate multilib: 32;@m32
Candidate multilib: x32;@mx32
Selected multilib: .;@m64
Found CUDA installation: /usr/local/cuda-12.6, version 12.6
Found HIP installation: /opt/rocm-6.2.0, version 6.2.41133

failure requires gcc-13/14 toolchain.

I understand that nvcc needs to address this issue by supporting clang 19 and beyond but it takes time. I'm wondering if there is a way to workaround this issue and unblock me from using clang in development.

[ye-luo]

Tried -fno-builtin. No luck. Not sure if I used it correctly.

nvcc -fno-builtin-__reference_converts_from_temporary -forward-unknown-to-host-compiler -allow-unsupported-compiler -ccbin=clang++ main.cu

clang does receive -fno-builtin

#$ clang++ -D__NV_NO_HOST_COMPILER_CHECK=1 -D__CUDA_ARCH_LIST__=520 -D__NV_LEGACY_LAUNCH -E -x c++ -D__CUDACC__ -D__NVCC__  -fno-builtin-__reference_converts_from_temporary "-I/usr/local/cuda-12.6/bin/../targets/x86_64-linux/include"    -D__CUDACC_VER_MAJOR__=12 -D__CUDACC_VER_MINOR__=6 -D__CUDACC_VER_BUILD__=77 -D__CUDA_API_VER_MAJOR__=12 -D__CUDA_API_VER_MINOR__=6 -D__NVCC_DIAG_PRAGMA_SUPPORT__=1 -include "cuda_runtime.h" -m64 "main.cu" -o "/tmp/tmpxft_00049e5d_00000000-5_main.cpp4.ii" 

[tbaederr]

CC @cor3ntin

[ye-luo]

g++ 13 works

nvcc -allow-unsupported-compiler -ccbin=g++ main.cu

So it is a limitation in cudafe++

[erichkeane]

See the discussion on #111516: This MIGHT be an NVCC or config issue, but we can continue the conversation on there.

[ye-luo]

I think NVIDIA needs to update nvcc eventually to accommodate this change. I'm currently use gcc 11 toolchain as a workaround.

[erichkeane]

I think NVIDIA needs to update nvcc eventually to accommodate this change. I'm currently use gcc 11 toolchain as a workaround.

That is definitely my suspicion as well, but I've pinged a senior NVCC dev on the other thread to see if they can confirm.