Styjun

I'm trying to compile a very simple cuda file:

#include <tuple>

__global__
void hello() 
 return;


int main() 
 dim3 dimBlock(16, 1);
 dim3 dimGrid(1, 1);
 hello<<<dimGrid, dimBlock>>();

 return 0;

I am using CUDA 9.0 and clang 3.9.1 with its accompanying libc++. Clag is not
installed in an standard location and I'm trying to compile the file with the
following command:

PATH=/path/to/clang/bin:$PATH 
path/to/nvcc 
 --verbose 
 -D_FORCE_INLINES 
 -gencode=arch=compute_35,"code=sm_35,compute_35" 
 -gencode=arch=compute_37,"code=sm_37,compute_37" 
 -gencode=arch=compute_52,"code=sm_52,compute_52" 
 -DNDEBUG 
 -std=c++11 
 --compiler-options " 
 -nostdinc 
 -nostdinc++ 
 -isystem /path/to/clang/include/c++/v1 
 -isystem /usr/local/include 
 -isystem /path/to/clang/lib/include 
 -isystem /path/to/clang/lib/clang/3.9.1/include 
 -isystem /usr/include/x86_64-linux-gnu 
 -isystem /usr/include 
 -isystem /path/to/cuda/includes 
 -iquote . 
 -g0 
 -no-canonical-prefixes 
 -fPIC 
 " 
 --compiler-bindir=/path/to/clang/bin/clang++ 
 -I . 
 -x cu 
 -O3 
 -c main.cu 
 -o main.cu.o

That compilation fails with the following error:

#$ cicc --c++11 --clang --clang_version=30901 --unicode_source_kind=UTF-8 --allow_managed -arch compute_52 -m64 -ftz=0 -prec_div=1 -prec_sqrt=1 -fmad=1 --include_file_name "tmpxft_00004816_00000000-2_main.cu.fatbin.c" -tused -nvvmir-library "path/to/cuda9/bin/../nvvm/libdevice/libdevice.10.bc" --gen_module_id_file --module_id_file_name "/tmp/tmpxft_00004816_00000000-3_main.cu.module_id" --orig_src_file_name "main.cu" --gen_c_file_name "/tmp/tmpxft_00004816_00000000-5_main.cu.compute_52.cudafe1.c" --stub_file_name "/tmp/tmpxft_00004816_00000000-5_main.cu.compute_52.cudafe1.stub.c" --gen_device_file_name "/tmp/tmpxft_00004816_00000000-5_main.cu.compute_52.cudafe1.gpu" "/tmp/tmpxft_00004816_00000000-8_main.cu.compute_52.cpp1.ii" -o "/tmp/tmpxft_00004816_00000000-5_main.cu.compute_52.ptx"
/path/to/clang/include/c++/v1/__tuple(265): error: __type_pack_element is not a template

/path/to/clang/include/c++/v1/__tuple(332): error: identifier "__type_pack_element" is undefined

/path/to/clang/include/c++/v1/__tuple(332): error: parameter pack "_Idx" was referenced but not expanded

/path/to/clang/include/c++/v1/__tuple(332): error: expected a ";"

4 errors detected in the compilation of "/tmp/tmpxft_00004816_00000000-8_main.cu.compute_52.cpp1.ii".
# --error 0x1 --

From what I figured out __type_pack_element is checked in case it is a clang
built-in, if it isn't it is defined. In the first pass of nvcc, it calls the
clang preprocessor which clearly knows __type_pack_element is a built in so
the preprocessor skips the section defining the type, but the cicc pass fails
because it doesn't have __type_pack_element as a built int.

Removing the tuple include causes the compilation to succeed. Using any other
stl include file succeeds.

The used clang was downloaded from the llvm page