OK. We are introducing method specialization based on call sites and we ARE introducing strength reduction based on rage analysis, so I guess this is +1 for me.
On Apr 24, 2013, at 12:26 AM, Hannes Wallnoefer <[email protected]> wrote: > The microbenchmark goes from about 250 millis to about 380 for most numbers. > But I guess we'll have to bite the bullet. Also V8 takes around 380 > milliseconds when it actually converts from double (much faster when it's > able to optimize to ints) so I guess it's just the time it takes. > > I also did a few octane runs and couldn't detect any regression there. > > Hannes > > Am 2013-04-23 20:51, schrieb Marcus Lagergren: >> How much is a bit and what did you measure? >> >> On Apr 23, 2013, at 8:33 PM, Hannes Wallnoefer >> <[email protected]> wrote: >> >>> Please review webrev for JDK-8012334: ToUint32, ToInt32, and ToUint16 don't >>> conform to spec: >>> >>> http://cr.openjdk.java.net/~hannesw/8012334/ >>> >>> The code path for small numbers is now a little bit slower but not much. >>> The code path for large numbers that don't fit in 32 bits is quite a bit >>> slower, but it should be correct now. The patch also contains a >>> microbenchmark for ToInt32 (signed right shift) and ToUint32 (unsigned >>> right shift). >>> >>> Hannes >
