Consider the following example:

```
global a;

probe oneshot {
    a[1] <<< 32;

    a[2] <<< 78;
    a[2] <<< 0;
    a[2] <<< 0;

    a[3] <<< 96;
    a[3] <<< 0;

    foreach ([i] in a @count-) {
        printf("%d: %d\n", i, @sum(a[i]));
    }
}
```

This example generates random output across different runs:

```
$ stap a.stp
1: 32
3: 96
2: 78

$ stap a.stp
2: 78
3: 96
1: 32

$ stap a.stp
3: 96
1: 32
2: 78
```

The MAP operations generated by the stap translator in the kernel source code misses the COUNT_OP flag in the `_stp_pmap_add_ix()` runtim C function call:

```
      { int rc = _stp_pmap_add_ix (global(s___global_a), ((int64_t)1LL), ((int64_t)32LL), 0, 1, 0, 0, 0); if (unlikely(rc)) { c->last_error = "Array overflow, check MAXMAPENTRIES"; goto out; }};
```

Using the implicit `-` sorting direction in the stap script example above instead of `@count-` has the same problem.

Both the kernel and dyninst runtime modes have this issue.

Use of the `@count()` operator inside the `foreach()` loop *body* does make the problem disappear.

The issue lies in the translator frontend where the use of aggregate operators in the foreach sorting specification does not affect the stat_op combinations of the `stat_op` data structure.