Google I/O 2010 - A JIT Compiler for Android's Dalvik VM

Google I/O 2010 - A JIT Compiler for Android's Dalvik VM

Android 301

Ben Cheng, Bill Buzbee

In this session we will outline the design of a JIT Compiler suitable for embedded Android devices. Topics will include an architectural overview, the rationale for design decisions and the special support for JIT verification, testing and tuning.

For all I/O 2010 sessions, please go to http://code.google.com/events/io/2010...

• Category

  Science & Technology

[alexgrinkov]

By call overhead I mean the actual overhead to call a function. I don't know how ARM processors or whatever Android runs on work, but on x86, to call a function, you need to establish a stack frame by moving registers around, by pushing required arguments in stack, by saving caller-saved registers used at the point of function call, by performing a jump and then after the call, you destruct the frame, and jump back... And all of this can be done using 1 vectorized instruction.

[SharifAjmal]

I haven't programmed in Java in ages, so I don't remember exactly how this works, but I know there was some sort of pass-by-reference-esque thing you an do... (I think you just wrap the contents of the stuff in an object or something.) So the only overhead is creating the activation record. (Negligible relative to the work the function will do for you.)

I should probably try to write some Java again at some point. :p

[alexgrinkov]

But games require lots of matrix/vector math. There's a lot of overhead to call functions to do, for example, vector multiplication instead of doing it inline in native code.

[alexgrinkov]

"ask questions on Google Wave"

haha

[SharifAjmal]

So if you're running something like a game that depends heavily on library code, you're not gonna get much of a performance boost from your translation cache... But you're also not gonna get much of a slowdown either (relative to purely native code); most of your time will be spent in library code, which is already compiled native. So there's not much translation overhead. (Plus the Dalvik interpreter was already twice as fast as the JVM interpreter; probably even faster now.)

[deathxblade]

So your saying its taking away CPU resources? But WITHOUT JIT you will have the interpreter taking up CPU resources. The only time JIT is bad is when all of your code is already in native code, which maybe you are thinking of since many 3D graphic libraries might be all in native code? Not sure.

[VoLtrex31]

3D intense software such as video games .

your CPU will have to do 2 tasks - 1 is to load data as you play.

2 to process the game.

while when JIT is turned off you will first load 100% of the game data and "Then" your CPU will Only process the game .

as for Quad CPU then the CPU can do it both with out much issues.

hope i made it more clear. i do not claim this information is right- IMO.

[deathxblade]

Your comment doesn't really make sense. What about "high end" programs make it run slower with JIT on?

[Graham Steffaniak]

5:21 "its really a smallish amount of code"