https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128859

            Bug ID: 128859
           Summary: [OpenMP] wrong code: statement before doacross(sink:
                    omp_cur_iteration - 1) runs only in first inner
                    iteration
           Product: gcc
           Version: 17.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: yqtian668 at gmail dot com
  Target Milestone: ---

Created attachment 66218
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=66218&action=edit
patch and test

With `#pragma omp for ordered(N)` where N is larger than the collapse count,
the statement immediately before `#pragma omp ordered doacross(sink:
omp_cur_iteration - 1)` is executed only in the first iteration of the inner
(non-collapsed) loops.

```c
int a[64][64], b;

void
f (int n)
{
  int i, j;
#pragma omp for ordered(2)
  for (i = 0; i < n; i++)
    for (j = 0; j < n; j++)
      {
        b++;
#pragma omp ordered doacross(sink: omp_cur_iteration - 1)
        a[i][j] = b;
#pragma omp ordered doacross(source: omp_cur_iteration)
      }
}

int
main ()
{
  f (5);
  if (b != 25)
    __builtin_abort ();
  return 0;
}
```

```sh
gcc -O2 -fopenmp t.c && OMP_NUM_THREADS=1 ./a.out
```

`b` is 5 instead of 25 (also at -O0).  Tested with trunk
e32f3d31e252d378450484b3a9250631d8efb36b and GCC 13.3.

If the statement before the construct is a debug statement, it is the debug
marker that gets misplaced instead, so -g and -g0 differ.  Writing `b++; ;`
gives b=25 with -g, b=5 with -g0, and `-fcompare-debug` fails.

In `expand_omp_ordered_sink`, for `fd->ordered > fd->collapse`, `gsi3` is the
last statement before the construct, and the block is split before it:

```c
          else
            {
              gsi_prev (&gsi3);
              e4 = split_block (gsi_bb (gsi3), gsi_stmt (gsi3));
            }
```

so that statement moves into the block guarded by `if (.first != 0)`.
Splitting after `gsi3` (dropping the `gsi_prev`) guards only the
inserted statements.  I have a patch with a libgomp runtime test.

Reply via email to