[OpenMP 4.5] ORDERED SIMD construct in loop SIMD doesn't work as required by the specification

[llvmbot]

Bugzilla Link	51701
Version	10.0
OS	Linux
Reporter	LLVM Bugzilla Contributor

Extended Description

For the following code:

#include <stdio.h>

void work(int j) {
  int i;
  #pragma omp parallel
  #pragma omp for simd ordered
  for (i = 1; i <= 10; i++) {
    printf("%d\n", i);
    #pragma omp ordered simd threads
    {
      printf("%d\n", i * j + 100000);
    }
  }
}

int main() {
  int k = 13;
  work(k);
}

GCC (Ubuntu 9.3.0-11ubuntu0~18.04.1) will generate an executable which prints something like:

9
7
4
5
2
3
1
100013
100026
100039
10
6
8
100052
100065
100078
100091
100104
100117
100130

And this output is in line with what is required by the specification (OpenMP API v4.5): the block demarcated by an ORDERED SIMD construct in a "simd, or loop SIMD region [..] will be exectued in the order of the loop iterations".

But the executable generated by clang will print something like:

1
100013
4
100052
5
100065
7
100091
6
100078
2
100026
3
100039
8
100104
10
100130
9
100117

For another piece of code (using ORDERED SIMD in a SIMD region, not in a loop SIMD region):

#include <stdio.h>

void work(int j) {
  int i;
  #pragma omp simd
  for (i = 1; i <= 10; i++) {
    printf("%d\n", i);
    #pragma omp ordered simd
    {
      printf("%d\n", i * j + 100000);
    }
  }
}

int main() {
  int k = 13;
  work(k);
}

Both sides will output:

1
100013
2
100026
3
100039
4
100052
5
100065
6
100078
7
100091
8
100104
9
100117
10

which is correct according to the specification.