On May 3, 2011, at 5:06 PM, Rémi Forax wrote: > On 05/03/2011 03:59 PM, Christian Thalinger wrote: >> On May 2, 2011, at 9:55 PM, Charles Oliver Nutter wrote: >>> On Thu, Apr 28, 2011 at 11:18 AM, Rémi Forax<fo...@univ-mlv.fr> wrote: >>>> Do you specialize the overflow check depending on the callsite ? >>>> for fib(n - 1), you just have to check if n is different from >>>> Integer.MIN_INT, >>>> for fib(n - 2), if n is<= to Integer.MIN_INT - 1 >>>> and for + use the double xor tricks. >>> Here's a related question. Is there a way we can structure the double >>> xor trick to get Hotspot to intrinsify it as a jc instruction or >>> similar? >>> >>> As I posted earlier, we use the "double xor" trick, which I would >>> expect JVM could recognize as a carry check and do the right thing. Or >>> perhaps there's an intrinsic somewhere in JVM/JDK we could call? >> I look a little into this and as far as I know there is nothing you could >> call. I also tried to hack something into the backend and I think that >> could work. I'm just not sure how many users would trigger that pattern. >> >> -- Christian > > The pattern occurs enough often to be recognized as a pattern. > > Here is the version used by JSR 310 reference implementation: > http://threeten.svn.sourceforge.net/viewvc/threeten/trunk/threeten/src/main/java/javax/time/MathUtils.java?revision=1417&view=markup > (look for safeAdd) > > It uses a && instead of a & in the middle, > I don't know which one is the better.
That's interesting, thank you. Intrinsifying a Java method would definitely be easier than doing something in the backend but I will look at it again. -- Christian _______________________________________________ mlvm-dev mailing list mlvm-dev@openjdk.java.net http://mail.openjdk.java.net/mailman/listinfo/mlvm-dev
