[Openmp-commits] [openmp] [OpenMP] Address __kmp_dist_for_static_init issue (PR #129902)

[via Openmp-commits]

jprotze wrote:

Sure, you are right. I just thought about the nesting from the perspective of lomp data structures, but not whether the code actually makes sense.

With the code from your gist, I can reproduce the issue after I set OMP_NUM_THREADS>1 (while 1 is the default on my test system :)
Since the assertion also triggers with `./a.out T F` in which case team->t.t_parent->t.t_nproc = nthreads-var (OMP_NUM_THREADS) and nteams = 2, the solution proposed by @hansangbae doesn't work.

When spawning the teams, the runtime already forks  nteams*nthreads-var threads, even if they are not used. These are organized in this additional base_team structure. In case of __kmp_fork_call, a special code path is taken `if (__kmp_is_fork_in_teams(...))` is true, but in case of __kmpc_serialized_parallel is used in __kmpc_fork_call_if, a new serialized team is nested into the preallocated team. 

For now, I think, your patch is the best solution to fix the broken assertion. 

As a long-term solution, I think that `__kmpc_serialized_parallel` should eventually also call `__kmp_fork_in_teams` to keep the nesting of teams structures clean and provide better results with OMPD/OMPT.

https://github.com/llvm/llvm-project/pull/129902