False positive reports for stack variable access

[jprotze]

Replacing the memoize version of Fibonacci in benchmarks/RacyFibonacci.cc with a bruteforce version, I get dozens of false positive reports for the race-free version:

long comp_fib_numbers(int n) {
  if ( n == 0 || n == 1 ) return n;
  long a=0, b=0;                                              // line 19

  #pragma omp task firstprivate(n) shared(a)                  // line 21
  {
    a = comp_fib_numbers(n - 1);                              // line 23
  }

  #pragma omp task firstprivate(n) shared(b)                  // line 26
  {
    b = comp_fib_numbers(n - 2);                              // line 28
  }

  #pragma omp taskwait // comment for racy version
  long fn = a+b;                                              // line 32
  // #pragma omp taskwait // uncomment for racy version

  return fn;
}

    19 (19)  <--> 23 (23) on 12 memory addresses
    19 (19)  <--> 28 (28) on 11 memory addresses
    19 (19)  <--> 32 (32) on 24 memory addresses
    21 (21)  <--> 21 (21) on 10 memory addresses
    21 (21)  <--> 23 (23) on 13 memory addresses
    21 (21)  <--> 26 (26) on 15 memory addresses
    21 (21)  <--> 28 (28) on 15 memory addresses
    23 (23)  <--> 23 (23) on 9 memory addresses
    23 (23)  <--> 26 (26) on 15 memory addresses
    23 (23)  <--> 32 (32) on 14 memory addresses
    26 (26)  <--> 26 (26) on 7 memory addresses
    26 (26)  <--> 28 (28) on 13 memory addresses
    28 (28)  <--> 28 (28) on 11 memory addresses
    28 (28)  <--> 32 (32) on 14 memory addresses

My suspicion is that reuse of stack addresses causes these reports. The reuse of stack addresses happens, when tasks are executed at the same call stack depth.
Is this a fundamental problem of the analysis approach, or can this be fixed?