That's okay, I've implemented the instrumentation and produced some stats now. :-)
On average the oldest targets are called significantly more frequently than later ones, with the average chain position remaining below 1.5 (0 being the oldest target) up to our standard maximum gwt chain length. I then significantly increased the maximum length of our chains un to 20, and the average chain position remains below 3 up to a chain length of 15, and only starts going significantly above 3 when we hit a length of 18. Increasing the chain length to 50 (!) showed that even then the average chain position was approximately 16, so there's still a fairly strong bias towards older targets. These figures almost certainly do not cover the bootstrap process as I had to reset the counts as part of call site invalidation, so some bootstrap processes will have wiped out the earliest stats. This suggests that in our case it would be well worth ensuring the oldest target is always first in the chain, and that chain may be worth keeping in place even once we've passed our the limit for inlining further targets. It is worth noting that this is all based on a very basic initial cut of the stats. I'd need to output the raw data and do some number crunching to check variances etc. on the individual call sites. On 21/12/2012 15:37, "Remi Forax" <fo...@univ-mlv.fr> wrote: >On 12/21/2012 04:07 PM, MacGregor, Duncan (GE Energy Management) wrote: >> Did you ever get a result from this? I haven't got round to >> instrumentation for our gwt chains that finely, but I'm curious what >> results you got. > >It's still on my TODO list, Java lambda stuff claim most of my time >devoted to OpenJDK ... > >Rémi > >> >> On 30/11/2012 08:12, "Remi Forax" <fo...@univ-mlv.fr> wrote: >> >>> There is a nice article about the JIT that Facebook uses in his HipHop >>>VM. >>> >>>https://www.facebook.com/notes/facebook-engineering/speeding-up-php-base >>>d- >>> development-with-hiphop-vm/10151170460698920 >>> >>> They have an inline cache with several guards and they found that it >>>was >>> better from the perf point of view to link the guard by putting the >>> older first instead of putting the newer first. I wonder if there is >>>the >>> same effect for other languages, I will test with Java + a dedicated >>> indy logic with week end. >>> >>> Rémi >>> >>> -- >>> You received this message because you are subscribed to the Google >>>Groups >>> "JVM Languages" group. >>> To post to this group, send email to jvm-languages@googlegroups.com. >>> To unsubscribe from this group, send email to >>> jvm-languages+unsubscr...@googlegroups.com. >>> For more options, visit this group at >>> http://groups.google.com/group/jvm-languages?hl=en. >>> > >-- >You received this message because you are subscribed to the Google Groups >"JVM Languages" group. >To post to this group, send email to jvm-languages@googlegroups.com. >To unsubscribe from this group, send email to >jvm-languages+unsubscr...@googlegroups.com. >For more options, visit this group at >http://groups.google.com/group/jvm-languages?hl=en. > -- You received this message because you are subscribed to the Google Groups "JVM Languages" group. To post to this group, send email to jvm-languages@googlegroups.com. To unsubscribe from this group, send email to jvm-languages+unsubscr...@googlegroups.com. For more options, visit this group at http://groups.google.com/group/jvm-languages?hl=en.
