yvals_core.h: reports unexpected CUDA version when being compiled with clang++ #2359
Description
Describe the bug
I fetched the latest sources from https://github.com/llvm/llvm-project.git, compiled clang and tried to run
> clang++.exe --cuda-path='C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.4' hello.cu
In file included from <built-in>:1:
In file included from C:\Programs\llvm-project\build\vs-2019-x86_64\Debug\lib\clang\14.0.0\include\__clang_cuda_runtime_wrapper.h:41:
In file included from C:\Program Files\Microsoft Visual Studio\2022\Preview\VC\Tools\MSVC\14.31.30818\include\cmath:9:
In file included from C:\Program Files\Microsoft Visual Studio\2022\Preview\VC\Tools\MSVC\14.31.30818\include\yvals.h:9:
C:\Program Files\Microsoft Visual Studio\2022\Preview\VC\Tools\MSVC\14.31.30818\include\yvals_core.h:582:2: error: STL1002: Unexpected compiler version,
expected CUDA 10.1 Update 2 or newer.
#error STL1002: Unexpected compiler version, expected CUDA 10.1 Update 2 or newer.
^
The problem is that the following code from https://github.com/microsoft/STL/blob/main/stl/inc/yvals_core.h#L576-L582 assumes that it's being compiled by nvcc and fails.
#ifndef _ALLOW_COMPILER_AND_STL_VERSION_MISMATCH
#ifdef __CUDACC__
#if __CUDACC_VER_MAJOR__ < 10 \
|| (__CUDACC_VER_MAJOR__ == 10 \
&& (__CUDACC_VER_MINOR__ < 1 || (__CUDACC_VER_MINOR__ == 1 && __CUDACC_VER_BUILD__ < 243)))
#error STL1002: Unexpected compiler version, expected CUDA 10.1 Update 2 or newer.
#endif // ^^^ old CUDA ^^^It doesn't take into account that the code might have been compiled with nvcc, see also see also https://www.llvm.org/docs/CompileCudaWithLLVM.html#detecting-clang-vs-nvcc-from-code
By running clang++.exe -x cuda /dev/null -dM -E I see that __CUDACC__ is defined, and the value of CUDA_VERSION is 11040, but __CUDACC_VER_MAJOR__ is missing.
For the time being I work around it by calling clang++.exe -D_ALLOW_COMPILER_AND_STL_VERSION_MISMATCH ..., but that doesn't sound correct. Could the above code be reworked a bit to account for clang as well?
See also #1544, @StephanTLavavej.
Command-line test case
C:\Temp> type hello.cu
__global__ void cuda_hello () {
const int a = 42;
}
int main() {
cuda_hello<<<1,1>>>();
return 0;
}
C:\Temp> clang++.exe --cuda-path='C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.4' hello.cu
In file included from <built-in>:1:
In file included from C:\Programs\llvm-project\build\vs-2019-x86_64\Debug\lib\clang\14.0.0\include\__clang_cuda_runtime_wrapper.h:41:
In file included from C:\Program Files\Microsoft Visual Studio\2022\Preview\VC\Tools\MSVC\14.31.30818\include\cmath:9:
In file included from C:\Program Files\Microsoft Visual Studio\2022\Preview\VC\Tools\MSVC\14.31.30818\include\yvals.h:9:
C:\Program Files\Microsoft Visual Studio\2022\Preview\VC\Tools\MSVC\14.31.30818\include\yvals_core.h:582:2: error: STL1002: Unexpected compiler version,
expected CUDA 10.1 Update 2 or newer.
#error STL1002: Unexpected compiler version, expected CUDA 10.1 Update 2 or newer.
^
1 error generated when compiling for sm_35.
Expected behavior
No error when compiling the sources.
STL version
Microsoft Visual Studio Community 2022 Preview
Version 17.1.0 Preview 1.1
(I'm generally using 2019 Preview, and I used that one for compiling clang as well, but for some reason clang automatically picks up the headers from 2022.)