No -- I get zero hits for 4 and 7 with this approach.

csl · on Nov 26, 2008

Ok, just to limit my embarrassment, I've implemented an alternative solution -- though not as good as the others.

First convert rand5() to rand2(). The LSB from rand5() has a uniform distribution for the integers 0--3:

    int rand2()
    {
        int n = rand5();
        return n!=4 ? n & 1 : rand2();
    }

Now we simply build a three bit number:

    int rand7()
    {
        int n = rand2();
        n |= rand2() << 1;
        n |= rand2() << 2;
        return n;
    }

This gives the proper distribution as well, but it's not branch free, so really nothing new here.