We're always trying to make JRuby faster, and I think we've done a pretty good job. But a recent contender has shown me some areas we really need to improve.
MacRuby recently started to implement their backend compilation on top of LLVM, and it seems to be paying substantial dividends for performance. On integer math benchmarks, they're several times faster than JRuby. On closure-performance benchmarks, they're at least twice as fast. Most of their performance boost seems to come from a couple specific optimizations: * True fixnums! I don't think we'll ever be able to compete on math performance until we either get fixnums in the JVM or have enough information to eliminate constructing them. My early attempts to do this have involved adding some "long RHS" call paths, which certainly helps. But we're still constructing too many fixnums, and nothing seems to get around that fact. Even Java 7 escape analysis combined with my dyncall inlining work has no effect. * Something like alloca. MacRuby generally allocates runtime structures with alloca, which appears to reduce their overhead substantially. In our case, closures are allocated with normal heap-based objects, and we pay the full cost for doing so. If we could reify local variables into a heap structure lazily, or stand up a scoping structure that only lived on the stack, we might get closer. * Lack of an optimizing compiler framework for JVM. The truth is that we're expecting too much of the JVM. We need a better compiler framework that can do some type inference, constant propagation, etc, and be aware of what bytecode and runtime logic optimizes well. The work Tobias is starting on ssa-compiler will be important...we need to find a way to move it along faster. I have a couple books and papers to read so I can be useful in this regard. I'd like to hear what other languages, especially dynamic languages, are doing to take the next performance steps. I'd also like input on what we can do to continue improving JRuby, and hopefully improve the state of dynlangs on JVM at the same time. - Charlie --~--~---------~--~----~------------~-------~--~----~ 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 -~----------~----~----~----~------~----~------~--~---
