[cfe-users] Compiling CUDA does not define __CUDACC__

[Edward Diener via cfe-users]

According to 
https://llvm.org/docs/CompileCudaWithLLVM.html#detecting-clang-vs-nvcc-from-code 
if I compile a CUDA file then __CUDACC__ should be defined. I have this 
program as 'clang_cuda.cu':

#if !defined(__clang__)

#error Clang is not defined

#elif !defined(__CUDACC__)

#error __CUDACC__ is not defined

#elif !defined(__CUDA__)

#error __CUDA__ is not defined

#endif

int main(void) {
     return 0;
}

which I compile with clang-9.0 targeting gcc on Windows with:

clang++ -c -m64 -O0 -fno-inline -Wall -g -pthread -target 
x86_64-pc-windows-gnu -nocudainc -nocudalib -x cuda -o clang_cuda.obj 
clang_cuda.cu

the program outputs the preprocessor error: __CUDACC__ is not defined

Why is this ?