Package: nvidia-cuda-toolkit
Version: 11.4.3-2
Severity: serious
Control: block 1003037 with -1
nvcc fails to compile bits/std_function.h from g++ 11.2:
$ echo '#include <functional>' | nvcc -ccbin g++-11 -x cu -c -
/usr/include/c++/11/bits/std_function.h:435:145: error: parameter packs not
expanded with ‘...’:
435 | function(_Functor&& __f)
|
^
/usr/include/c++/11/bits/std_function.h:435:145: note: ‘_ArgTypes’
/usr/include/c++/11/bits/std_function.h:530:146: error: parameter packs not
expanded with ‘...’:
530 | operator=(_Functor&& __f)
|
^
/usr/include/c++/11/bits/std_function.h:530:146: note: ‘_ArgTypes’
This is a regression from the header shipped with g++ 11.1
(11.2 includes a fix for an STL defect:
"2774. std::function construction vs assignment")
The offending code can be reduced to
===== nvcc-gcc112-failure.cu =====
template < typename >
class function ;
template < typename >
class _Function_handler ;
template < typename _Res , typename ... _ArgTypes >
class function < _Res ( _ArgTypes... ) >
{
template < typename = void >
using _Handler = _Function_handler < _Res ( _ArgTypes... ) > ;
function
() noexcept ( _Handler < > :: template _S_nothrow_init < > ) ;
} ;
=====
nvcc -v -x cu -c nvcc-gcc112-failure.cu yields these commands
g++-11 -D__CUDA_ARCH__=520 -E -x c++ -DCUDA_DOUBLE_MATH_FUNCTIONS -D__CUDACC__
-D__NVCC__ -D__CUDACC_VER_MAJOR__=11 -D__CUDACC_VER_MINOR__=4
-D__CUDACC_VER_BUILD__=152 -D__CUDA_API_VER_MAJOR__=11
-D__CUDA_API_VER_MINOR__=4 -include "cuda_runtime.h" -m64
"nvcc-gcc112-failure.cu" -o
"/tmp/tmpxft_00007ddc_00000000-7_nvcc-gcc112-failure.cpp1.ii"
cicc --c++17 --gnu_version=110200 --orig_src_file_name "nvcc-gcc112-failure.cu"
--allow_managed -arch compute_52 -m64 --no-version-ident -ftz=0 -prec_div=1
-prec_sqrt=1 -fmad=1 --include_file_name
"tmpxft_00007ddc_00000000-3_nvcc-gcc112-failure.fatbin.c" -tused
--gen_module_id_file --module_id_file_name
"/tmp/tmpxft_00007ddc_00000000-4_nvcc-gcc112-failure.module_id"
--gen_c_file_name
"/tmp/tmpxft_00007ddc_00000000-6_nvcc-gcc112-failure.cudafe1.c"
--stub_file_name
"/tmp/tmpxft_00007ddc_00000000-6_nvcc-gcc112-failure.cudafe1.stub.c"
--gen_device_file_name
"/tmp/tmpxft_00007ddc_00000000-6_nvcc-gcc112-failure.cudafe1.gpu"
"/tmp/tmpxft_00007ddc_00000000-7_nvcc-gcc112-failure.cpp1.ii" -o
"/tmp/tmpxft_00007ddc_00000000-6_nvcc-gcc112-failure.ptx"
ptxas -arch=sm_52 -m64
"/tmp/tmpxft_00007ddc_00000000-6_nvcc-gcc112-failure.ptx" -o
"/tmp/tmpxft_00007ddc_00000000-8_nvcc-gcc112-failure.sm_52.cubin"
fatbinary -64 --cicc-cmdline="-ftz=0 -prec_div=1 -prec_sqrt=1 -fmad=1 "
"--image3=kind=elf,sm=52,file=/tmp/tmpxft_00007ddc_00000000-8_nvcc-gcc112-failure.sm_52.cubin"
"--image3=kind=ptx,sm=52,file=/tmp/tmpxft_00007ddc_00000000-6_nvcc-gcc112-failure.ptx"
--embedded-fatbin="/tmp/tmpxft_00007ddc_00000000-3_nvcc-gcc112-failure.fatbin.c"
rm -f /tmp/tmpxft_00007ddc_00000000-3_nvcc-gcc112-failure.fatbin
g++-11 -E -x c++ -D__CUDACC__ -D__NVCC__ -D__CUDACC_VER_MAJOR__=11
-D__CUDACC_VER_MINOR__=4 -D__CUDACC_VER_BUILD__=152 -D__CUDA_API_VER_MAJOR__=11
-D__CUDA_API_VER_MINOR__=4 -include "cuda_runtime.h" -m64
"nvcc-gcc112-failure.cu" -o
"/tmp/tmpxft_00007ddc_00000000-5_nvcc-gcc112-failure.cpp4.ii"
cudafe++ --c++17 --gnu_version=110200 --orig_src_file_name
"nvcc-gcc112-failure.cu" --allow_managed --m64 --parse_templates
--gen_c_file_name
"/tmp/tmpxft_00007ddc_00000000-6_nvcc-gcc112-failure.cudafe1.cpp"
--stub_file_name
"tmpxft_00007ddc_00000000-6_nvcc-gcc112-failure.cudafe1.stub.c"
--module_id_file_name
"/tmp/tmpxft_00007ddc_00000000-4_nvcc-gcc112-failure.module_id"
"/tmp/tmpxft_00007ddc_00000000-5_nvcc-gcc112-failure.cpp4.ii"
g++-11 -D__CUDA_ARCH__=520 -c -x c++ -DCUDA_DOUBLE_MATH_FUNCTIONS -m64
"/tmp/tmpxft_00007ddc_00000000-6_nvcc-gcc112-failure.cudafe1.cpp" -o
"nvcc-gcc112-failure.o"
the last g++-11 call is the failing one:
nvcc-gcc112-failure.cu:10:28: error: parameter packs not expanded with ‘...’:
10 | function
| ^
nvcc-gcc112-failure.cu:10:28: note: ‘_ArgTypes’
Checking /tmp/tmpxft_00007ddc_00000000-6_nvcc-gcc112-failure.cudafe1.cpp
after rerunning the commands manually to preserve the temporary files,
the code has been rewritten by cudafe++ to
=====
template< class > class function;
template< class > class _Function_handler;
template< class _Res, class ..._ArgTypes>
class function< _Res (_ArgTypes ...)> {
template< class = void> using _Handler = _Function_handler< _Res (_ArgTypes
...)> ;
function() noexcept(&_Function_handler< _Res (_ArgTypes)> ::template
_S_nothrow_init<>);
};
=====
The actual rewriting change is
-noexcept(_Handler<> ::template _S_nothrow_init<>)
+noexcept(&_Function_handler<_Res(_ArgTypes)>::template _S_nothrow_init<>)
which correctly makes the code fail to compile.
nvidia-cuda-toolkit 11.5.1 seems to have the same issue.
Andreas
