Hyperthreaded Micro Threads for Aspect Oriented Programming

I recently discovered a series of articles written by Jon "Hannibal" Stokes on arstechnica.com. Usually, I stick to learning weird, exotic languages, but I'm really fascinated by his descriptions of modern processors. I'm especially fascinated by the idea of tuning modern day processors and programming techniques so that they take advantage of one another. For instance, C++'s use of automatic variables results in so many function calls, that if these function calls are not inlined, they can play hell on the pipeline, unless branch prediction works well.

Having read Hannibal's articles on hyperthreading, SMT, and the Itanium-64, IA-64, I became intriqued by the performance aspects of SMT on aspect oriented programming, AOP. AOP is currently confined mostly to the Java world, so it doesn't have a direct effect on the processor, per se, but I wondered if a natively compiled AOP compiler could better take advantage of an IA-64 processor. It would do this via explicitly parallel instruction computing, EPIC, and what I call micro threading.

Specifically, AOP permits a separation of concerns. Hence, a function that looks like this:

void do_f() {
log();
authenticate();
do_stuff();
log();
}

can be torn apart into something like:

void do_f() {
do_stuff();
}

// Blatant pseudo code.
match function do_* {
before { log(); authenticate(); }
after { log(); }
}

I propose that the above AOP could possibly be extended to:

match function do_* {
before { thread_fork log(); inline authenticate(); }
after { thread_fork log(); }
}

Hence, in the above example, for every function do_*, there would be a "micro" thread to take are of logging at the start of the function and another micro thread to take care of logging at the end of the function, but the authentication would be inlined. Thanks to EPIC, the compiler could give a hint to the processor and the OS that it would be best to run these micro threads at the same time as the main function. Paraphrasing the words of Intel, I'm creating code-level parallelism in places where there was none before. AOP was designed to make the programmer's life easier, but with the addition of hyperthreaded micro threads, it can possibly result in substantial performance increases as well.

Well, I hope the above made some sense. I most definitely welcome your comments. Hopefully, someone more intelligent than me can make it happen ;)

Comments

I actually posted this "micro-article" on my Web site about two years ago. Now that I have a blog, I finally have a place for it. For any of you that have seen it, please forgive the duplication.