Comment by mananaysiempre

Rejection sampling, that is. Unexpectedly, but obviously in retrospect, the common (ancient) wisdom of using that in the continuous case instead of figuring out tricky combinations of special functions to get the result from a bounded number of uniform samples... fails miserably on LuaJIT. Of course atan2() is not the fastest thing in the world, but on a compiler that only really understands loops with linear bodies calling it even via FFI beats adding an unpredictable nested loop.