Replacing ICC C++14 with C++17

[rwgk]

Description

This PR reverts most of PR #3551. To work around the issue that started with Intel 2021.5.0.20211109, -Wno-conversion is added specifically for C++17 (as recommended here: #3570 (comment)).

Quality assurance: #3570 (comment)

Original PR description

For reporting to Intel.

The C++17 build was working with

• Intel 2021.4.0.20210910 (/opt/intel/oneapi/compiler/2021.4.0/linux/bin/intel64/icpc)

but started failing with

• Intel 2021.5.0.20211109 (/opt/intel/oneapi/compiler/2022.0.0/linux/bin/intel64/icpc)

See also: PR #3551

Suggested changelog entry:

- Fix: the latest release warned in ``#include <variant>``, which failed ``-DPYBIND11_WERROR=ON``

[ax3l]

@rscohn2@xmnboy would this be an oneAPI icc/icpc regression worth filing an internal ticket for?
We use pybind11 in our upcoming Python bindings for many of our C++ codes and C++17 builds recently broke with the latest oneAPI icpc release. Our codes are currently all using C++17.

[ 8%] Building CXX object tests/CMakeFiles/pybind11_tests.dir/test_buffers.cpp.o
cd /home/runner/work/pybind11/pybind11/build-17/tests && /opt/intel/oneapi/compiler/2022.0.1/linux/bin/intel64/icpc -Dpybind11_tests_EXPORTS -I/home/runner/work/pybind11/pybind11/include -isystem /usr/include/python3.8 -Os -DNDEBUG -fPIC -fvisibility=hidden -Wall -Wextra -Wconversion -Wcast-qual -Wdeprecated -Wundef -Wnon-virtual-dtor -Werror-all -diag-disable 11074,11076 -ipo -std=c++17 -MD -MT tests/CMakeFiles/pybind11_tests.dir/test_buffers.cpp.o -MF CMakeFiles/pybind11_tests.dir/test_buffers.cpp.o.d -o CMakeFiles/pybind11_tests.dir/test_buffers.cpp.o -c /home/runner/work/pybind11/pybind11/tests/test_buffers.cpp
In file included from /home/runner/work/pybind11/pybind11/include/pybind11/stl.h(31),
from /home/runner/work/pybind11/pybind11/tests/test_buffers.cpp(12):
/usr/include/c++/9/variant(123): error: integer conversion resulted in a change of sign
inline constexpr size_t variant_npos = -1;
^
In file included from /home/runner/work/pybind11/pybind11/include/pybind11/stl.h(31),
from /home/runner/work/pybind11/pybind11/tests/test_buffers.cpp(12):
/usr/include/c++/9/variant(1727): error: integer conversion resulted in a change of sign
constexpr size_t __magic_monostate_hash = -7777;
^

This looks like a general problem with std::variant/#include <variant>:

pybind11/include/pybind11/stl.h

Line 31 in 2cd32e5

# include<variant>

We see this on our GitHub actions CI, using the ubuntu-20.04 image (see this PR) and installing the oneAPI .deb packages for icc/icpc classic compilers.

[rscohn2]

I filed a ticket. I don't know if it will be fixed because there are workarounds (disabling warnings) and we are transitioning to icpx. Here is the reproducer:

rscohn1@rscohn1-mobl1:icx$ /opt/intel/oneapi/compiler/2022.0.1/linux/bin/intel64/icpc -Wconversion -std=c++17 -Werror-all -c variant\
.cpp
In file included from variant.cpp(1):
/usr/include/c++/9/variant(123): error: integer conversion resulted in a change of sign
inline constexpr size_t variant_npos = -1;
^
In file included from variant.cpp(1):
/usr/include/c++/9/variant(1727): error: integer conversion resulted in a change of sign
constexpr size_t __magic_monostate_hash = -7777;
^
compilation aborted for variant.cpp (code 2)
rscohn1@rscohn1-mobl1:icx$ cat variant.cpp
#include <variant>
rscohn1@rscohn1-mobl1:icx$ /opt/intel/oneapi/compiler/2021.4.0/linux/bin/intel64/icpc -Wconversion -std=c++17 -Werror-all -c variant\
.cpp
rscohn1@rscohn1-mobl1:icx$ /opt/intel/oneapi/compiler/2022.0.1/linux/bin/intel64/icpc --version
icpc (ICC) 2021.5.0 20211109
Copyright (C) 1985-2021 Intel Corporation. All rights reserved.
rscohn1@rscohn1-mobl1:icx$ /opt/intel/oneapi/compiler/2021.4.0/linux/bin/intel64/icpc --version
icpc (ICC) 2021.4.0 20210910
Copyright (C) 1985-2021 Intel Corporation. All rights reserved.
rscohn1@rscohn1-mobl1:icx$

[ax3l]

Thanks a lot @rscohn2!
I see, this is then just a diagnostic that went off for the shipped <variant> include.

Yes, fully understand with respect to the icx/icpx effort. We actually try to add coverage for it as well in #2769 :)

[rscohn2]

The changes you proposed are good workarounds, and I am glad to see you are adding icpx support.

[rwgk]

-DPYBIND11_WERROR=OFF works but -diag-disable:conversion did not:

icpc: warning #10034: Unrecognized keyword 'conversion' for option '-diag-disable'
In file included from /home/runner/work/pybind11/pybind11/include/pybind11/stl.h(31),
from /home/runner/work/pybind11/pybind11/tests/test_buffers.cpp(12):
/usr/include/c++/9/variant(123): error: integer conversion resulted in a change of sign
inline constexpr size_t variant_npos = -1;
^

@rscohn2 is there a way to enable building with -Werror-all while selectively disabling this diagnostic? I do not see a warning number that we could use with -diag-disable.

[ax3l]

Try maybe -Wno-conversion?

[rwgk]

Try maybe -Wno-conversion?

CI is running.

[rwgk]

For quality assurance, copying parts of the C++11 and C++17 compilation commands from the log:

-Wall -Wextra -Wconversion -Wcast-qual -Wdeprecated -Wundef -Wnon-virtual-dtor -Werror-all -diag-disable 11074,11076 -ipo -std=c++11 -MD -MT

-Wall -Wextra -Wconversion -Wcast-qual -Wdeprecated -Wundef -Wnon-virtual-dtor -Wno-conversion -Werror-all -diag-disable 11074,11076 -ipo -std=c++17 -MD -MT

[rwgk]

@henryiii@ax3l@Skylion007 I think this is good to go in.

[rwgk]

In view of the transition to icpx, I feel it's not worth the effort troubleshooting the icpc C++20 issue(s).

[henryiii]

Okay, irritating to have to turn off a warning, but it's not ours (and the other compilers should catch if we make a mistake)