AS3 Bitwise gems - fast integer math • Hacker News Board

mahmud wrote: post #2

Interested parties can also look at how the Tamarin VM implements those bitwise opcodes natively. http://www.mozilla.org/projects/tamarin/ Flash opcodes into Nanojit IR: http://hg.mozilla.org/tamarin-central/raw-file/4eb9f961a087/... Nanojit IR to native x86 translation tables? http://hg.mozilla.org/tamarin-central/raw-file/4eb9f961a087/...

Nice! That's great to hear, since it's always perplexing when native code ISN'T the fastest... Kind of makes you have to ask, "If there is a better way, why isn't that the native way?" But, for Tamarin, that question will be moot.

Re: AS3 Bitwise gems - fast integer math

#4

Post by pmjordan » Sun, May 24, 2009, 4:06 PM UTC

I don't do flash, but I was under the impression that there's some kind of compilation going on and ActionScript isn't just interpreted. How on earth does a modern compiler not optimise

  a / 64

to

  a >> 6

? I could understand it in a dynamically typed language, where a, and therefore the result, could be non-integer, but ActionScript seems to have static type declarations for variables.

Re: AS3 Bitwise gems - fast integer math

#5

Post by _cbsz » Sun, May 24, 2009, 9:00 PM UTC

pmjordan wrote: post #4

I don't do flash, but I was under the impression that there's some kind of compilation going on and ActionScript isn't just interpreted. How on earth does a modern compiler not optimise a / 64 to a >> 6 ? I could understand it in a dynamically typed language, where a, and therefore the result, could be non-integer, but ActionScript seems to have static type declarations for variables.

This isn't even "modern" in compiler terms, it's a simple peephole optimization. (grep through the IR, check for "X / 2^n", replace by "X >> n") To get it right when you have

    int a;
    int b = 64;
    a / b

requires dataflow analysis so you can do constant propagation, copy propagation, and dead code elimination. This is more work to get right, but that's not a very good excuse for not at least doing the peephole optimization.