alias template error with Intel 2016.0.3 compilers

[loriab]

Issue description

In trying to update some projects from 2.0.0 to 2.2.1, I've found that they consistently won't compile with the icpc (ICC) 16.0.3 2016041 line of compilers. Fine with GCC 5.2 and Intel icpc (ICC) 17.0.4 20170411. An error is below:

In file included from /home/psilocaluser/gits/cmake_example/pybind11/include/pybind11/attr.h(13),
                 from /home/psilocaluser/gits/cmake_example/pybind11/include/pybind11/pybind11.h(43),
                 from /home/psilocaluser/gits/cmake_example/src/main.cpp(1):
/home/psilocaluser/gits/cmake_example/pybind11/include/pybind11/cast.h(837): error: alias template "pybind11::detail::type_caster_base<type>::cast_op_type" may not be used in the type-id of the alias-declaration
      template <typename T> using cast_op_type = cast_op_type<T>;
                                                 ^

In file included from /home/psilocaluser/gits/cmake_example/pybind11/include/pybind11/attr.h(13),
                 from /home/psilocaluser/gits/cmake_example/pybind11/include/pybind11/pybind11.h(43),
                 from /home/psilocaluser/gits/cmake_example/src/main.cpp(1):
/home/psilocaluser/gits/cmake_example/pybind11/include/pybind11/cast.h(1814): warning #3373: nonstandard use of "auto" to both deduce the type from an initializer and to announce a trailing return type
      static constexpr auto args_pos = constexpr_first<argument_is_args, Args...>() - (int) sizeof...(Args),
                       ^

compilation aborted for /home/psilocaluser/gits/cmake_example/src/main.cpp (code 2)
gmake[2]: *** [CMakeFiles/cmake_example.dir/src/main.cpp.o] Error 2

Reproducible example code

The error can be seen with @dean0x7d's sample cmake+pybind11 repo, https://github.com/pybind/cmake_example (admittedly pinned at 2.2.0). When forced to switch compiler families as below, it gives the error above.

>>> icpc --version
icpc (ICC) 16.0.3 20160415
Copyright (C) 1985-2016 Intel Corporation.  All rights reserved.
>>> mkdir build/temp.linux-x86_64-3.5
>>> cd build/temp.linux-x86_64-3.5 &&  cmake /path/to/cmake_example -DCMAKE_LIBRARY_OUTPUT_DIRECTORY=/path/to/cmake_example/build/lib.linux-x86_64-3.5 -DPYTHON_EXECUTABLE=/miniconda/envs/idp35p4/bin/python -DCMAKE_BUILD_TYPE=Release -DCMAKE_C_COMPILER=icc -DCMAKE_CXX_COMPILER=icpc
>>> cmake --build . --config Release -- -j2

I'm sorry to not further narrow the error source. I'm glad to try out any changes since this particular compiler flavor may not be widespread.

[klaw9]

This issue also affects nvcc (cuda 8). For now I made it work by adding a namespace qualifier to the type-id of the alias statement.

template <typename T> using cast_op_type = detail::cast_op_type<T>;

[loriab]

Thanks, @klaw9, I confirm that fixes the stated problem.

Next problem is below, if anyone's curious.

/home/psilocaluser/gits/hrw-matt/objdir_aggh2016/stage/usr/local/psi4/include/pybind11/detail/descr.h(59): error: expression must have a constant value
      static constexpr auto digits = descr<sizeof...(Digits)>(('0' + Digits)...);

[henryiii]

This error is a regression and makes pybind11 master and 2.x unusable with CUDA. Adding the word detail:: fixes the issue

[jagerman]

The original issue is fixed by 32ef69a which I just pushed. I suspect @loriab's issue is going to be more annoying to track down and fix (I think it was probably caused initially by #934).

[henryiii]

Did you push it to the 2.2 branch as well?

[jagerman]

No, but I'll set the milestone to remind us to include it in the 2.2.2 release.

[jagerman]

@loriab - is that issue under master, or the 2.2 release? I don't think the #934 changes made it into 2.2.

[jagerman]

(I only have ICC 17 available to me, where both of these compile okay).

[henryiii]

Okay, I'd like to have it in soon, because I've transitioned to 2.2.1 using a normal compiler, but it's now broken under CUDA, so switching the checkout to the 2.2 branch would be a temporary fix. I closed my PR since I assume it's best to cherry-pick the same commit.

By the way, just for general information, you can grab NVIDIA's docker containers and build with CUDA there in normal Docker, just (obviously) can't run.

(CUDA seems to be happy with the Digits line)

[loriab]

Thanks for the help. I just rebuilt with pybind11 master. Confirm that the original cast_op_type fixed for me in master and the next issue Digits remains.

[jagerman]

Try playing around with that line to see if you can make ICC 16 happy (as I said, 17 is fine with it).

Does changing ('0' + Digits)... to ('0' + static_cast<char>(Digits))... help? (I'm basically just guessing right now).

[jagerman]

Another possible solution is replace ('0' + Digits)... with digit_to_str(Digits)... and add:

inline constexpr char digit_to_str(size_t digit) { return static_cast<char>('0' + digit); }

just before the templated struct int_to_str.