[OpenMP] Fix tasking + parallel bug

[jlpeyton]

Summary

From the bug report, the runtime needs to initialize the nproc variables (inside middle init) for each root when the task is encountered, otherwise, a segfault can occur.

Bugzilla: https://bugs.llvm.org/show_bug.cgi?id=36720

[Hahnfeld]

This doesn't seem to work with GOMP interface. I think the initialization should be moved to __kmp_task_alloc (or __kmp_omp_task, is there a reason this has to happen when allocating?)?

[AndreyChurbanov]

In D49996#1180711, @Hahnfeld wrote:

This doesn't seem to work with GOMP interface. I think the initialization should be moved to __kmp_task_alloc (or __kmp_omp_task, is there a reason this has to happen when allocating?)?

Yes, should be fine to move it to __kmp_task_alloc. An alternative would be to change middle initialization to initialize nproc not only for current task, but for all its parents, that looks a bit trickier to write at first glance...

Regarding __kmp_omp_task - this looks worse, because it may be skipped. E.g. Intel compiler can skip it for "task if(0)". So the allocation routine looks preferable.

[Hahnfeld]

In D49996#1180769, @AndreyChurbanov wrote:

Regarding __kmp_omp_task - this looks worse, because it may be skipped. E.g. Intel compiler can skip it for "task if(0)". So the allocation routine looks preferable.

Ah yes, that makes sense.

[jlpeyton]

Moved middle init to __kmp_task_alloc()

[Hahnfeld]

LGTM, fixes the provided test case. Thanks for the fast fix!

openmp/trunk/runtime/src/kmp_tasking.cpp

kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
                              kmp_routine_entry_t task_entry) {
   kmp_task_t *task;
   kmp_taskdata_t *taskdata;
   kmp_info_t *thread = __kmp_threads[gtid];
   kmp_team_t *team = thread->th.th_team;
   kmp_taskdata_t *parent_task = thread->th.th_current_task;
   size_t shareds_offset;
 
+  if (!TCR_4(__kmp_init_middle))
+    __kmp_middle_initialize();
+
   KA_TRACE(10, ("__kmp_task_alloc(enter): T#%d loc=%p, flags=(0x%x) "
                 "sizeof_task=%ld sizeof_shared=%ld entry=%p\n",
                 gtid, loc_ref, *((kmp_int32 *)flags), sizeof_kmp_task_t,
                 sizeof_shareds, task_entry));
 
   if (parent_task->td_flags.final) {
     if (flags->merged_if0) {
     }

openmp/trunk/runtime/test/tasking/bug_36720.c

+// RUN: %libomp-compile-and-run
+
+/*
+Bugzilla: https://bugs.llvm.org/show_bug.cgi?id=36720
+
+Assertion failure at kmp_runtime.cpp(1715): nthreads > 0.
+OMP: Error #13: Assertion failure at kmp_runtime.cpp(1715).
+
+The assertion fails even with OMP_NUM_THREADS=1. If the second task is removed,
+everything runs to completion. If the "omp parallel for" directives are removed
+from inside the tasks, once again everything runs fine.
+*/
+
+#define N 1024
+
+int main() {
+  #pragma omp task
+  {
+    #pragma omp parallel for
+    for (int i = 0; i < N; i++)
+      (void)0;
+  }
+
+  #pragma omp task
+  {
+    #pragma omp parallel for
+    for (int i = 0; i < N; ++i)
+      (void)0;
+  }
+
+  #pragma omp taskwait
+
+  return 0;
+}