[OpenMP] host rtl: fixed bug in implementation of distribute construct.

[AndreyChurbanov]

Summary

When the "target teams distribute" construct located inside parallel region,
scheduler mistakenly shared to loop iterations between teams from different target regions.
The patch fixes this problem.

[jlpeyton]

According to the pre-merge results: https://buildkite.com/llvm-project/premerge-checks/builds/76212#f61fe509-f800-4780-9571-cfe70624ba18, the test still depends on libomptarget even though it is run on the host. I think the test may have to go into the libomptarget/test directory. @jdoerfert could you advise on integrating the test?

[AndreyChurbanov]

Removed target triple from test compilation.
This eliminates the dependency on libomptarget.

[jlpeyton]

Ah ok that works for me!

openmp/runtime/src/kmp_sched.cpp

#endif
   // Although there are schedule enumerations above kmp_ord_upper which are not
   // schedules for "distribute", the only ones which are useful are dynamic, so
   // cannot be seen here, since this codepath is only executed for static
   // schedules.
   if (schedtype > kmp_ord_upper) {
     // we are in DISTRIBUTE construct
     schedtype += kmp_sch_static -
                  kmp_distribute_static; // AC: convert to usual schedule type
+    if (th->th.th_team->t.t_serialized > 1) {
+      tid = 0;
+      team = th->th.th_team;
+    } else {
       tid = th->th.th_team->t.t_master_tid;
       team = th->th.th_team->t.t_parent;
+    }
   } else {
     tid = __kmp_tid_from_gtid(global_tid);
     team = th->th.th_team;
   }
 
   /* determine if "for" loop is an active worksharing construct */
   if (team->t.t_serialized) {
     /* serialized parallel, each thread executes whole iteration space */

openmp/runtime/test/teams/teams-distr-on-host.c

+// The test supposes no offload, pure host execution.
+// It checks that the bug in implementation of distribute construct is fixed.
+
+// RUN: %libomp-compile-and-run
+// UNSUPPORTED: icc
+
+#include <stdio.h>
+#include <omp.h>
+
+int main()
+{
+  const int size = 4;
+  int wrong_counts = 0;
+  omp_set_num_threads(2);
+  #pragma omp parallel reduction(+:wrong_counts)
+  {
+    int i;
+    int A[size];
+    int th = omp_get_thread_num();
+    for(i = 0; i < size; i++)
+      A[i] = 0;
+
+    #pragma omp target teams distribute map(tofrom: A[:size]) private(i)
+    for(i = 0; i < size; i++)
+    {
+      A[i] = i;
+      printf("th %d, team %d, i %d\n", th, omp_get_team_num(), i);
+    }
+    #pragma omp critical
+    {
+      printf("tid = %d\n", th);
+      for(i = 0; i < size; i++)
+      {
+        if (A[i] != i) wrong_counts++;
+        printf("  %d", A[i]);
+      }
+      printf("\n");
+    }
+  }
+  if (wrong_counts) {
+    printf("failed\n");
+  } else {
+    printf("passed\n");
+  }
+  return wrong_counts;
+}