On a side note, as D also uses MT19937 as the main RNG, random2 designers may want to address the same accuracy issue which Python did. More on that topic here: http://bugs.python.org/issue9025
Sorry for replying to myself again, but the abovementioned accuracy issue is not about MT19937. It's about how we transform the (uniform enough) bits from MT19937 to an integer in some range [0, LIMIT). Looks like the inaccuracy is already addressed in D, so, sorry again for the bother. Still, the current implementation uses divisions (which is slow on most CPUs), and for a good source of random bits such as MT19937, the implementation which repeatedly gets exactly core.bitop.bsr(LIMIT)+1 bits until the result is in [0, LIMIT) may be faster.
----- Ivan Kazmenko
