Skip to content

UT C++ — gtest libstdc++ ABI Mismatch

TL;DR

If cmake --build tests/ut/cpp/build fails with undefined reference to testing::internal::EqFailure(... std::string const& ...), the system libgtest.a that find_library picked up was hand-compiled with -D_GLIBCXX_USE_CXX11_ABI=0 (the old pre-cxx11 ABI). The tests now build with the compiler default (cxx11) ABI, so symbol mangling does not match. Either rebuild gtest without the ABI flag, or force FetchContent by hiding the system copy:

cmake -B tests/ut/cpp/build -S tests/ut/cpp \
      -DGTEST_LIB=GTEST_LIB-NOTFOUND -DGTEST_MAIN_LIB=GTEST_MAIN_LIB-NOTFOUND

This doc exists so the next person who hits the symptom does not re-investigate the same rabbit hole.

Symptom

/usr/bin/ld: CMakeFiles/test_tensormap.dir/hierarchical/test_tensormap.cpp.o:
in function testing::internal::CmpHelperEQFailure<int,int>(...):
undefined reference to `testing::internal::EqFailure(char const*, char const*,
                       std::string const&, std::string const&, bool)`
collect2: error: ld returned 1 exit status

The mangled form of the missing symbol is _ZN7testing8internal9EqFailureEPKcS2_RKSsS4_b — RKSs is const std::string& in the old ABI (the new ABI mangles std::string as NSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE).

Equivalently, nm /usr/local/lib*/libgtest.a | grep -c __cxx11 returns 0 on the broken install, ≥1 on a healthy one.

Root Cause

tests/ut/cpp/CMakeLists.txt used to force -D_GLIBCXX_USE_CXX11_ABI=0 on every test target and on every runtime OBJECT lib it links. The matching gtest had to be old-ABI too — fine when FetchContent built it under the same flag, but a system gtest packaged with the default (cxx11) ABI link-failed against tests with the symmetric form of the error.

That _GLIBCXX_USE_CXX11_ABI=0 was historical baggage:

  • src/ runtime build does not set the flag. Production libhost_runtime.so is already cxx11 — nm shows __cxx11::basic_string mangling on exported C++ symbols (e.g. DeviceRunner::init_pmu).
  • CANN libs (libascendcl, libruntime, libhcom) export only extern "C" symbols, so the ABI flag is irrelevant at the .so boundary.

The flag was removed in PR #889 and everything (including the tests that dlopen the runtime .so) is now cxx11 end-to-end. The flip side is that an old-ABI system gtest, which used to be the "fast" path, now fails the link. apt / yum / brew packages have all shipped cxx11 ABI for years; the only way to land in this state is a hand-compiled gtest left over in /usr/local.

Fixes

Pick one:

  1. Rebuild gtest without the ABI flag. If you cloned googletest and configured it with cmake -D_GLIBCXX_USE_CXX11_ABI=0 (or with an old simpler checkout that did so transitively), redo the configure without that flag and make install.

  2. Drop the manually-installed copy, let find_library miss, and FetchContent will rebuild a matching one in the worktree:

sudo rm /usr/local/lib*/libgtest*.a /usr/local/lib*/libgtest_main*.a
  1. One-off override, no global change:
cmake -B tests/ut/cpp/build -S tests/ut/cpp \
      -DGTEST_LIB=GTEST_LIB-NOTFOUND \
      -DGTEST_MAIN_LIB=GTEST_MAIN_LIB-NOTFOUND

Repeat once; the cache is fresh, FetchContent takes over.

History

  • PR #882 added a try_compile ABI probe as a defensive check while the _GLIBCXX_USE_CXX11_ABI=0 flag was still in place: it caught a symmetric mismatch the other way (test=old-ABI vs. a5 self-hosted runner's system gtest=new-ABI). That probe is gone now that the flag is gone — the whole thing reduces to "use what find_library finds; fall back to FetchContent if it finds nothing."
  • PR #889 dropped the 13 -D_GLIBCXX_USE_CXX11_ABI=0 sites and the probe; this doc records the trail.