Anyway, nice to see that somebody is doing some work in that direction... :)
Marek On Wednesday, April 17, 2013 8:10:37 PM UTC+2, Ghadi Shayban wrote: > > Brandon, > Not really, but I don't want to deter anyone from exploration as its own > goal. There are a lot of interesting ways that Clojure can leverage indy, > this was an experiment. I'm going to try to put it in more high value > places, and let the ideas marinate. This modification does nothing but > actually use the instruction and pass all the tests. Obviously a lot of the > really useful call sites for indy require much rework (and will break ABI > and JVM compatibility, ick). The microarchitecture of how a language > runtime can interact with its host platform is fascinating to explore. > Btw, Attila Szegedi wrote the dynalink [1] library that has some > abstractions over a lot of the machinery inside java.lang.invoke, > definitely worth a look through. > > I think that performance is something exciting for developers, and I think > indy will help with that...eventually. This needs hammock time, Clojure is > definitely not slow today. > > [1] https://github.com/szegedi/dynalink/wiki/User-Guide-0.5 > > Ghadi > @smashthepast > > > On Tuesday, April 16, 2013 10:58:50 PM UTC-4, Brandon Bloom wrote: >> >> Do you have any measurements, results, or conclusions to report? >> >> On Tuesday, April 16, 2013 12:12:15 PM UTC-4, Ghadi Shayban wrote: >>> >>> I've added a very minimal usage of invokedynamic to the compiler. >>> Basically the smallest delta without having to change internals of ObjExpr >>> or break ABI compatibility. This is minimal and raw. There are many many >>> usages of indy that will really help the Clojure runtime, this ain't one of >>> them. No benchmarks here, it's probably slower. >>> >>> In current Clojure mainline, a Fn has reference slots to any vars it >>> needs in its constant pool, like a cache. Anytime a non-dynamic Var's value >>> needs to be accessed, the Var goes from the constant pool to the stack, and >>> getRawRoot() is invoked on it. >>> >>> With this change, an invokedynamic instruction instead creates a >>> ConstantCallSite, which closes over a looked-up Var, and then binds the >>> call site to invoke getRawRoot() on it directly. This is only for >>> non-dynamic Vars. >>> >>> Simple todos: >>> cache the CallSite as a member on the Var itself so that all identical >>> indy lookup instructions have fast bootstrapping. >>> emit a similar call for dynamic vars >>> remove the emission of Vars into the constant pool of a class >>> >>> Lots of really interesting use cases for invokedynamic and all the >>> associated combinators in java.lang.invoke: >>> >>> Better protocol callsite caching >>> CallSite "middleware" for things like CLJ specific instrumentation >>> equality could be a special instruction >>> KeywordCallSite could be its own instruction as well >>> (apply) argument "spreading"/varargs array collection (through the >>> combinators) >>> potentially removing IFn.invoke(*) and using MethodHandle invocation >>> instead (fat chance) >>> >>> You can pull down the changes at github.com/ghadishayban/clojure. >>> mvn clean package, tested on OpenJDK 1.7 >>> >> -- -- You received this message because you are subscribed to the Google Groups "Clojure" group. To post to this group, send email to clojure@googlegroups.com Note that posts from new members are moderated - please be patient with your first post. To unsubscribe from this group, send email to clojure+unsubscr...@googlegroups.com For more options, visit this group at http://groups.google.com/group/clojure?hl=en --- You received this message because you are subscribed to the Google Groups "Clojure" group. To unsubscribe from this group and stop receiving emails from it, send an email to clojure+unsubscr...@googlegroups.com. For more options, visit https://groups.google.com/groups/opt_out.
