[R-package] [ci] remove unnecessary include in linear_tree_learner (fixes #6264)

[jameslamb]

Fixes #6264.

Please see the linked issue for details. In short, CRAN has begun testing projects using clang-18 (which is not yet released) targeting libc++-18 instead of the system C++ stdlib. Those checks resulted in this compiler warning:

/usr/local/clang-trunk/bin/../include/c++/v1/__fwd/string_view.h:22:41: warning: 'char_traitsfmt::detail::char8_type' is deprecated: char_traits for T not equal to char, wchar_t, char8_t, char16_t or char32_t is non-standard and is provided for a temporary period. It will be removed in LLVM 19, so please migrate off of it. [-Wdeprecated-declarations]

If we don't submit a fix before January 24, {lightgbm} will be archived on CRAN

This PR proposes the following:

• adds a CI job that reproduces the warning (build link)
• fixes the underlying issue

Notes for Reviewers

It appears that the underlying issue was "linear_tree_learning.h includes <string> unnecessarily". I guess that leads clang to expand a template in fmt/core.h to many unnecessary types, some of which are considered deprecated in v18 of libc++.

I think?

I'm not really sure, but I do know that:

• the CI job(s) added in this PR reproduce the failure from CRAN
• removing that unnecessary include causes them to pass

[jameslamb]

🎉 I was able to reproduce the warning from CRAN in CI!

https://github.com/microsoft/LightGBM/actions/runs/7497815594/job/20411976864?pr=6265

Now we can push potential fixes here and use that to test.

[jameslamb]

To start, I tried just upgrading to the latest fmt (v10.2.1) (following the steps in #6074).

Nevermind, I decided to revert this. The less we change in the next release, the lower the risk of CRAN finding an issue and archiving the package.

[jameslamb]

@guolinke @shiyu1994 could you please help with a review?

I'd like to merge this and then begin a 4.3.0 release as soon as possible, to minimize the risk of the package being archived on CRAN.

[jameslamb]

Without this, compiler invocations and most warnings are not visible in CI logs because R CMD check redirects them to this file instead of printing them.

I think this will be generically useful for CI in the future, and in fact we already do that for most other R CI jobs:

LightGBM/.ci/test_r_package.sh

Lines 226 to 232 in 2e3543c

	echo "R CMD check build logs:"
	BUILD_LOG_FILE=lightgbm.Rcheck/00install.out
	cat ${BUILD_LOG_FILE}
	if [[ $check_succeeded == "no" ]]; then
	exit -1
	fi

[fabsig]

FWIW: I am the main author of GPBoost which includes a fork of LightGBM. I got the same warning from CRAN and adapted it as suggested in this commit (simply dropping #include <string> from linear_tree_learner.h). Unfortunately, this does not fix the issue. I am still getting the following response from CRAN:

Thanks, we now see with clang 18:

• checking whether package ‘gpboost’ can be installed ... [28m/70m] WARNING
  Found the following significant warnings:

/usr/local/clang-trunk/bin/../include/c++/v1/__fwd/string_view.h:22:41:
warning: 'char_traitsfmt::detail::char8_type' is deprecated:
char_traits for T not equal to char, wchar_t, char8_t, char16_t or
char32_t is non-standard and is provided for a temporary period. It will
be removed in LLVM 19, so please migrate off of it.
[-Wdeprecated-declarations]

from

/usr/local/clang-trunk/bin/clang++ -std=gnu++17
-I"/data/gannet/ripley/R/R-clang
-trunk/include" -DNDEBUG -I./include -DEIGEN_MPL2_ONLY -DMM_PREFETCH=1
-DMM_MALL
OC=1 -DUSE_SOCKET -DLGB_R_BUILD -isystem /usr/local/clang-trunk/include
-I/usr/
local/clang/include -fopenmp -pthread -fpic -O3 -Wall -pedantic
-frtti -Wp,-D
_FORTIFY_SOURCE=3 -c treelearner/tree_learner.cpp -o
treelearner/tree_learner.o
/usr/local/clang-trunk/bin/clang++ -std=gnu++17
-I"/data/gannet/ripley/R/R-clang
-trunk/include" -DNDEBUG -I./include -DEIGEN_MPL2_ONLY -DMM_PREFETCH=1
-DMM_MALL
OC=1 -DUSE_SOCKET -DLGB_R_BUILD -isystem /usr/local/clang-trunk/include
-I/usr/
local/clang/include -fopenmp -pthread -fpic -O3 -Wall -pedantic
-frtti -Wp,-D
_FORTIFY_SOURCE=3 -c treelearner/voting_parallel_tree_learner.cpp -o
treelearne
r/voting_parallel_tree_learner.o
In file included from treelearner/voting_parallel_tree_learner.cpp:5:
In file included from ./include/LightGBM/utils/common.h:12:
In file included from ./include/LightGBM/utils/log.h:14:
In file included from
/usr/local/clang-trunk/bin/../include/c++/v1/iostream:43:
In file included from /usr/local/clang-trunk/bin/../include/c++/v1/ios:223:
In file included from
/usr/local/clang-trunk/bin/../include/c++/v1/__locale:23:
In file included from
/usr/local/clang-trunk/bin/../include/c++/v1/string:625:
In file included from
/usr/local/clang-trunk/bin/../include/c++/v1/string_view:2
13:
/usr/local/clang-trunk/bin/../include/c++/v1/__fwd/string_view.h:22:41:
warning: 'char_traitsfmt::detail::char8_type' is deprecated:
char_traits for T not equal to char, wchar_t, char8_t, char16_t or
char32_t is non-standard and is provided for a temporary period. It will
be removed in LLVM 19, so please migrate off of it.
[-Wdeprecated-declarations]
22 | template <class _CharT, class _Traits = char_traits<_CharT> >
| ^
./include/LightGBM/fmt/core.h:391:50: note: in instantiation of template
type alias 'std_string_view' requested here
391 | S,
detail::std_string_view>::value)>
| ^
./include/LightGBM/fmt/format.h:542:63: note: in instantiation of
template class 'fmt::basic_string_viewfmt::detail::char8_type'
requested here
542 | inline size_t count_code_points(basic_string_view<char8_type> s) {
| ^

Please fix and resubmit.

[jameslamb]

Thanks for posting @fabsig !

I still believe we should try to submit a version of LightGBM with this PR on it to CRAN. I'm confident in the change because the new CI jobs added here raised the same compilation warning + R CMD check failure as CRAN did, and now don't.

Maybe there are other sources of this warning in GPBoost... different paths or order of includes, different uses of fmt, different version of fmt, etc.

@guolinke @shiyu1994 could you please help with a review today? I'd like to try to submit to CRAN tonight. We only have 7 more days to fix this before the package is archived there.

[jameslamb]

thanks @guolinke !

[fabsig]

@jameslamb: I did indeed have a different version of fmt and now it also works for GPBoost. Thanks a lot for your comment! This saved my day 👍

[jameslamb]

Glad it helped @fabsig !

I submitted a version of {lightgbm} with this PR's fixes a few days ago, and it's passing CRAN's checks so far: https://cran.r-project.org/web/checks/check_results_lightgbm.html.

Including, most notably, that the clang18 check issue is no longer shown there.