Bug in performance test.

[yiyuanliu]

During performance test, I got some strange results. I noticed that in class PerformanceTest, it create a tmp buffer and copy data into tmp buffer before run sort. But sort function runs directly on the origin buffer instead of tmp buffer. This causes the remaining tests will be performed on ordered data。

// speed.cc class PerformanceTest
template <typename SORT_FUNCTION>
uint64_t run(SORT_FUNCTION sort) {

    uint64_t time = 0;

    int k = iterations;
    while (k--) {
        memcpy(tmp, input.pointer(), input.size());

        uint64_t t1, t2;

#ifdef USE_RDTSC
        RDTSC_START(t1);
#else
        t1 = get_time();
#endif
        // !!! should sort tmp buffer, not origin buffer !!!
        sort(input.pointer(), 0, input.count() - 1);

#ifdef USE_RDTSC
        RDTSC_START(t2);
#else
        t2 = get_time();
#endif

        const uint64_t dt = t2 - t1;

        if (time == 0) {
            time = dt;
        } else if (dt < time) {
            time = dt;
        }
    }

    return time;
}

[bbqz007]

even thought

        if (time == 0) {
            time = dt;
        } else if (dt < time) {
            time = dt;
        }

it use the max dt not the average time.