getThreadNum() always returns 0 (Bug #3529)

Hi,

I know that it depends on the underlying parallelism library chosen.
In my case, I am using tbb, on OsX.

The following code :

#include <opencv2/core/core.hpp>
#if CV_MAJOR_VERSION >= 3
#include <opencv2/core/utility.hpp>
#endif
#
#include <iostream>
#include <unistd.h>

using namespace std;
using namespace cv;

struct ThreadIdGetter
{
    ThreadIdGetter() {
        AutoLock lock(mutex);
        int i, count = used.size();
        for (i=0; i<count; ++i) if (!used[i]) { used[i] = true; break; }
        if (i==count) used.push_back(true);
        id = i;
    }
    ~ThreadIdGetter() {
        AutoLock lock(mutex);
        used[id] = false;
    }
    int operator()() const {
        return id;
    }
    int id;
    static vector<bool> used;
    static Mutex mutex;
};
vector<bool> ThreadIdGetter::used;
Mutex ThreadIdGetter::mutex;

struct Doer : public ParallelLoopBody {
    void operator()(const Range& range) const {
        ThreadIdGetter id;
        for (int i=range.start; i<range.end; ++i) {
            mutex.lock();
            cout << "Thread id : " << getThreadNum() << " " << id() << endl;
            mutex.unlock();
            usleep(rand()%5000+5000);
        }
    }
    mutable Mutex mutex;
};

int main(int argc, char *argv[]) {
    parallel_for_(Range(0,10), Doer(), 10);
    return 0;
}

yields, e.g, the following output :

Thread id : 0 0
Thread id : 0 1
Thread id : 0 2
Thread id : 0 3
Thread id : 0 2
Thread id : 0 1
Thread id : 0 0
Thread id : 0 3
Thread id : 0 1
Thread id : 0 2

The first number is returned by getThreadNum() and is always 0, while the second is returned from my code.
I used a simple structure with a critical section to get a unique id for each concurrent thread.
I guess this is not the best solution, but at least, it gives a different id for each thread that is executing in parallel.

Fixing this implies having a look at each implementation of getThreadNum() ? or maybe just tbb's is wrong ?