Undefined symbol _ZN3fmt2v86detail13error_handler8on_errorEPKc

[saraedum]

Upgrading fmt from 8.0.1 to 8.1.0 breaks a CI run here: https://github.com/flatsurf/flatsurf/runs/4698212878?check_suite_focus=true#step:6:554

In that CI run, we are trying to load a library, libintervalxt, which was built against fmt 8.0.1 with fmt 8.1.0 installed.

While the missing symbol is not part of the public API of fmt, it used to have FMT_API set in 8.0.1 but not anymore in 8.1.0. At least in the setup that the conda-forge distribution uses, this made that symbol disappear from the library. Since libraries such as libintervalxt build against the old ABI (without using this symbol explicitly in the code of course) this seems like a breaking ABI change was introduced in the 8.1.0 release.

Or maybe conda-forge's libintervalxt and fmt are using the wrong compiler flags somehow?

[saraedum]

So, I guess any inline (e.g. template) function in any of the header files that calls on_error produced a reference to the on_error symbol if it decided not to inline it. So it seems to me that there is nothing wrong with our build of fmt 8.0.1. Also libraries that built against fmt 8.0.1 will reference that symbol even if they don't call it directly. So, again, nothing seems to be wrong with our libintervalxt build here.

[zhsj]

Same issue for Debian, https://release.debian.org/britney/pseudo-excuses-experimental.html

fmtlib (8.0.1+ds1-4 to 8.1.0+ds1-1)
Migration status for fmtlib (8.0.1+ds1-4 to 8.1.0+ds1-1): BLOCKED: Rejected/violates migration policy/introduces a regression
Issues preventing migration:
autopkgtest for bear/3.0.17-1: amd64: Regression ♻ (reference ♻), arm64: Regression ♻ (reference ♻)
autopkgtest for ceph/16.2.7+ds-1: amd64: Regression ♻ (reference ♻), arm64: Regression ♻ (reference ♻)
autopkgtest for cryfs/0.11.0-2: amd64: Regression ♻ (reference ♻), arm64: Regression ♻ (reference ♻)
autopkgtest for spdlog/1:1.9.2+ds-0.2: amd64: Regression ♻ (reference ♻), arm64: Regression ♻ (reference ♻)
autopkgtest for therion/6.0.4-1: amd64: Regression ♻ (reference ♻), arm64: Regression ♻ (reference ♻)
Additional info:
Cannot be tested by piuparts (not a blocker) - (no link yet)
1 days old

spdlog seems to use this symbol as well.

/usr/bin/c++ -O2 CMakeFiles/spdlog-utests.dir/test_file_helper.cpp.o CMakeFiles/spdlog-utests.dir/test_file_logging.cpp.o CMakeFiles/spdlog-utests.dir/test_daily_logger.cpp.o CMakeFiles/spdlog-utests.dir/test_misc.cpp.o CMakeFiles/spdlog-utests.dir/test_eventlog.cpp.o CMakeFiles/spdlog-utests.dir/test_pattern_formatter.cpp.o CMakeFiles/spdlog-utests.dir/test_async.cpp.o CMakeFiles/spdlog-utests.dir/test_registry.cpp.o CMakeFiles/spdlog-utests.dir/test_macros.cpp.o CMakeFiles/spdlog-utests.dir/utils.cpp.o CMakeFiles/spdlog-utests.dir/main.cpp.o CMakeFiles/spdlog-utests.dir/test_mpmc_q.cpp.o CMakeFiles/spdlog-utests.dir/test_dup_filter.cpp.o CMakeFiles/spdlog-utests.dir/test_fmt_helper.cpp.o CMakeFiles/spdlog-utests.dir/test_stdout_api.cpp.o CMakeFiles/spdlog-utests.dir/test_backtrace.cpp.o CMakeFiles/spdlog-utests.dir/test_create_dir.cpp.o CMakeFiles/spdlog-utests.dir/test_cfg.cpp.o CMakeFiles/spdlog-utests.dir/test_time_point.cpp.o CMakeFiles/spdlog-utests.dir/test_stopwatch.cpp.o CMakeFiles/spdlog-utests.dir/test_errors.cpp.o -o spdlog-utests  /usr/lib/x86_64-linux-gnu/libspdlog.so.1.9.2 -lpthread /usr/lib/x86_64-linux-gnu/libfmt.so.8.1.0 -Wl,--as-needed 
/usr/bin/ld: /usr/lib/x86_64-linux-gnu/libspdlog.so.1.9.2: undefined reference to `fmt::v8::detail::error_handler::on_error(char const*)'

[zhsj]

I backported #2696 to Debian, but this symbol still doesn't get exported.

[vitaut]

Should be fixed in 98cbb6a. Thanks for reporting!