On 05/01/2011 09:58 PM, Charles Oliver Nutter wrote: > On Sun, May 1, 2011 at 1:01 PM, Mark Roos<mr...@roos.com> wrote: >> With respect to chaining gwt handles. >> >> Any thoughts on at what depth the chain should be converted to some other >> form of dispatch? >> I am currently just letting them build >> >> In smalltalk it seems that around 10-20 inline tests are about the same as a >> hash lookup but I >> would think it could be different after the JIT gets to it. > I think it's going to vary widely by how much code is in the target > methods, how expensive the test is, how "balanced" the chain is (e.g. > most frequently-hit at the front) and so on. Obviously if you have a > megamorphic site that's seeing hundreds of methods you wouldn't want > to chain GWT to that length. > > I'm going to write a benchmark with variable-morphology(?) to see > where the line is in JRuby. > > The other question is whether a mostly mono or bimorphic call site > should keep those at the head of the chain, speeding up those calls, > and just have the IC at the end of the chain as a catch-all. I'm also > trying to think of low-cost ways of tracking which branches are > getting followed in the chain, so they could potentially be rebalanced > periodically. >
Could you also include a test with -XX:+TieredCompilation ? I wonder how profile information are propagated between tiers. > - Charlie cheers, Rémi _______________________________________________ mlvm-dev mailing list mlvm-dev@openjdk.java.net http://mail.openjdk.java.net/mailman/listinfo/mlvm-dev
