#2126
#2137
#2140
#2172
#2212
#2213
#2241
#2267
#2322
#2348
Meta-Bug for performance bugs affecting 0 A.D.: Bug 897962
Overview
We are currently using Spidermonkey 1.8.5. That's a very old version and we need to upgrade to a newer version for various reasons. Unfortunately that's not very easy because the API changes quite often.
Also check the forum thread for the current status.
Reasons for upgrading:
- Switching earlier avoids writing too much code for the older API that has to be changed later
- Only when using a supported version it's possible to fix Spidermonkey security issues (I know that 0 A.D. itself most likely has more security issues at the moment)
- Learning how the new API works with multi-threading is important for our AI multi-threading design.
- Performance: During the work on upgrading we figured out that performance currently isn't much better in the new version compared to v1.8.5. The huge improvements from benchmarks don't translate to most real world applications. Loading random maps is a lot faster, but that's the only thing at the moment. It's still possible that the performance will improve once the most important bugs in Ion Monkey and Baseline are fixed.
- If we find bugs in v1.8.5 we will never get support from Mozilla. If we use the current version or if the same bug also affects the current version, getting help is much more likely.
- Getting new features into the official version of Spidermonkey is also only possible when we use the current version.
- Many Linux distributions will use the new ESR versions of the standalone library. It's better to use that library provided by the system instead of bundling our own version. v1.8.5 won't be supported by Linux distributions forever.
Version
Mozilla plans to release standalone versions of Spidermonkey for each ESR-release.
Currently we don't know which version of Spidermonkey we will use because some performance bugs need to be fixed in the current version of Spidermonkey.
JIT
The Just In Time compiler creates optimized bytecode or even native assembler code for functions or loops that are called often. That's something different than the compiling which translates the scripts to bytecode in the first step.
It's currently not possible to JIT-compile the scripts before they are executed or even store them in JIT-compiled form on the disc.
Using js::NotifyAnimationActivity it's possible to prevent the garbage collection from collecting generated JIT-code (and forcing recompilation). Unfortunately this only works if js::NotifyAnimationActivity is called at least once a second. With big lag-spikes, especially in debug mode, this can't be guaranteed at the moment.
At the moment I'm working around this by modifying Spidermonkey's Runtime.cpp directly. I'll have to find a better solution.
Garbage collection
Garbage collection and memory allocation is a major bottleneck for certain parts of JS code that update a lot of data very often. I made some measurements here:
http://www.wildfiregames.com/forum/index.php?showtopic=16721&#entry262183
Garbage collection could be improved by improving the way we handle compartments, especially related to the AI. That's a bit vague but I'm still trying to figure out more about that.
An interesting article: http://andreasgal.com/2010/10/13/compartments/ .
That article is quite old, but I think the basic concepts explained there are still valid.
Garbage collection can now even be done in a separate thread or we could run it more often (each frame for example) to avoid lag-spikes when it happens. For the AI that problem would probably be solved anyway when we move it to a separate thread.
I asked about the JIT topic and Ion-Monkey in the jsapi IRC channel. Logs here.
Summary: JIT compiled code can't be saved on the harddisk and used later because it contains memory addresses. However, a function should be compiled with Ion-Monkey when it has been run 1000 times, so everything that runs once per frame should be compiled after about 30 seconds with 30-35 fps, which should be fine for the moment.
Garbage collection occasionally throws away the JIT-ed code. To make sure this doesn't happen, calling "js::NotifyAnimationActivity" should be enough. Check Bug 753630 for more background information about that.