Openmp 在并行块后测试值时出现意外的未指定行为

Question

考虑以下内容：

// compile with
// gcc -fopenmp test.c

#include <stdio.h>    // to display on the screen
#include <stdlib.h>   // to access EXIT_SUCCESS
#include <omp.h>      // OpenMP library.

int main() {
  unsigned long long input = 100;
  unsigned long long b = 0, c = 0;
  #pragma omp parallel
  {
    #pragma omp for
    for(unsigned long long a = 0 ;  a < input ; a++) {
      b++;
      c++;
    }
  } // <= There should be a barrier here.
  // Check
  if (b != c || b != input){
      fprintf(stdout, "\"Error\" : %llu and %llu are different from %llu", b, c, input);
      return EXIT_FAILURE;
  }
  else{
    return EXIT_SUCCESS;
  }
}

我目前的理解是，#pragma omp parallel { 块的末尾应该有一个隐式屏障，以便当代码到达 if 语句。

然而，我得到的结果不一致：

gcc -fopenmp test.c
./a.out 
"Error" : 70 and 69 are different from 100
./a.out 
"Error" : 75 and 78 are different from 100
./a.out 
"Error" : 83 and 81 are different from 100
./a.out 
"Error" : 91 and 86 are different from 100
./a.out  <= Nothing, which means it "worked"
./a.out 
"Error" : 88 and 86 are different from 100

我的理解有什么缺陷？