Optimization of Body's hidden class

[bchevalier]

I noticed that the number of properties on the bodies were not always the same. It varied between 45 and 48. Which is bad for performance as the JS compiler and optimizer cannot take advantage of consistent hidden classes when compiling JS code.
That is to say, having different number of properties per body meant that the optimizer could not optimally optimize some methods (apparently the most unoptimized methods were solverVelocity and Body.update).

Changes

• Made sure all 48 properties were initialized at body creation time.
• Bonus change: simplified the Body.set method as it was never called with a string as a 2nd argument.

Results

The gain is especially high on test using mixed shapes where more than 2 hidden structures had to be supported. The following benchmark was done on the Mixed Shape test:
Before

After

On a fairly similar profiling duration we can notice that the consumption of the engine is about 25% lower after the optimization!

Gains are usually not that high and seem to indicate a performance boost of 10% in most cases.

Note: Taking a memory snapshot both before and after seemed to indicate that the memory footprint increase was unnoticeable:

(Snapshot 4 is before and Snapshot 5 after)

[photonstorm]

Just tested this across several light and heavy tests. On average I'm seeing Engine.update go from 63.2% down to 51.8% with no significant memory difference.

Am seeing regular gc flushes though, causing a traditional peaked mountain effect. It did this before making this change too though. Resolver.solveVelocity remains the main CPU hog across all tests.

[bchevalier]

I think it's possible to reduce object creation (by creating pool of reusable objects) although I haven't experienced it as being an important issue yet. What browser are you testing on?

[liabru]

Looks great other than the change to the signature of Body.set. I know that internally the string option isn't used but it's intentionally there as a shorthand for users.

Can you revert that part so that feature remains? Thanks!

[bchevalier]

Done!