On 01/15/2013 06:50 PM, Peter Levart wrote: > On 01/15/2013 03:38 PM, Aleksey Shipilev wrote: >> On 01/15/2013 06:33 PM, Peter Levart wrote: >>> One more thing, not related to serialization: >>> >>> If a TLR reference is somehow passed from the thread that obtained it >>> via TLR.current() to some other thread that did never call TLR.current() >>> and this other thread calls methods on such instance (nextInt(), ...), >>> it will start the random sequence from the zero seed, bypassing >>> localInit() call... >>> >>> Is this ok? >> I think this counts as "accidental" sharing, and Javadoc recommends to >> always do TLR.current().* to take the appropriate TLR; hence I think >> this is the adopted failure scenario. The upside for *not* fixing this >> is skipping the initialized checks from each of next*() methods, which >> are needed in corner cases only, thus saving the hot-path performance. > > If javadoc recommends to always do TLR.current() then this method is > also hot-path. So why not move the initialization check from current() > to next(int bits)? Well, the following usage pattern can save some > checks in the original case: > > TLR r = TLR.current(); > > r.next*(); .... r.next*(); ....
I'm not sure I understand your comment, Peter. Your code calls TLR.current() first, and that is a good move, as per Javadoc; it should be probably mentioned that TLR should be *acquired* via current(). This does not prevent acquiring TLR once per scope, or even per thread. Before, the performance bets were off for rogue non-current() TLRs, now we should extend that to the statistical properties as well. -Aleksey.
