test_limiter_node causes segmentation fault in release mode

[Flamefire]

The test fails with a SIGSEGV in release mode:

44: /tmp/s3248973-EasyBuild/tbb/2021.3.0/GCCcore-10.3.0/oneTBB-2021.3.0/test/tbb/test_limiter_node.cpp:522:
44: TEST CASE:  Message is released if successor does not accept
44: 
44: /tmp/s3248973-EasyBuild/tbb/2021.3.0/GCCcore-10.3.0/oneTBB-2021.3.0/test/tbb/test_limiter_node.cpp:522: FATAL ERROR: test case CRASHED: SIGSEGV - Segmentation violation signal
44: 
44: ===============================================================================
44: [doctest] test cases:      4 |      3 passed |      1 failed |      2 skipped
44: [doctest] assertions: 217181 | 217181 passed |      0 failed |
44: [doctest] Status: FAILURE!
 44/131 Test  #44: test_limiter_node ........................***Exception: SegFault  2.96 sec

It passes in Debug and RelWithDebInfo. Enabling the workaround at https://github.com/oneapi-src/oneTBB/blob/55f9b178643d2925e9d81151efd08ac34dc69625/test/tbb/test_limiter_node.cpp#L358-L365 for the used GCC 10 makes it pass.

According to GDB the crash happens in spawn_in_graph_arena, but due to inlining I can't tell more.

Compiling with UBSAN shows:

include/tbb/../oneapi/tbb/flow_graph.h:2019:37: runtime error: execution reached an unreachable program point

So I guess there is an actual logic error somewhere

[phprus]

Related issue: #342

[Flamefire]

Thanks, and indeed -fno-ipa-cp-clone resolves the issue with TBB as well as with UBSAN and valgrind. So I'd guess it is a GCC bug.

[omalyshe]

@Flamefire Can you close the issue as resolved?

[Flamefire]

Not sure. Although it is a GCC bug, the issue persists, so I'd rather leave this open until either the bug is permanently workaround-ed or the GCC bug is fixed.
I see the workaround has been updated to include GCC 10 and 11, but if not resolved it will still happen for GCC 12: 7567de9

Also note that the above commit erroneously copied the comment //putting the successor back (cc @Iliamish )

[anton-potapov]

@Flamefire as the #647 is landed - could you check if the problem is fixed ?

[isaevil]

Seems like the problem is solved. If any other questions are left, feel free to reopen this issue.