Generating the sync report takes ~twice as long as
a normal tick, and occurs once every 3 ticks.
These reports record of all of the synced state
(separate to the sync hash, which is still calculated)
in order to generate the syncreport.log of the game
desyncs. This perf overhead is completely unnecessary
when we know that we won't have other syncreports to
compare against (singleplayer, replays).
Disabling report generation in these cases gives
us an easy 40% average tick-time win.
This greatly improves performance by not syncing the world state for every single order processed as this becomes very expensive, at the cost of being unable to directly pinpoint the order that causes a desync. Instead the granularity of detecting desyncs is reduced to the tick level.
- Implement IDisposable interface correctly, with sealed classes where possible for simplicity.
- Add using statement around undisposed local variables.
- Made private methods static where possible (runtime can elide checking the object for null).
- Declared attribute classes as sealed (allows reflection on attributes to complete faster).
- Moved some static cctor's into field initializers (static cctor's are slower than static field initializers).
- Made classes static if they contained only static methods (can't create instances of useless objects).
- Use inferable Exts.Lazy and not new Lazy<T>().
- Added required STAThread attribute to CrashDialog.
- Removed unused parameters in private methods.
- Added Serializable attribute to exceptions.
- Added parameter name in calls to ArgumentNullException.
- Use of as operator instead of is + cast.
- Changed (x as Foo).Bar anti-pattern into ((Foo)x).Bar. Results in sensible cast exceptions on error rather than null dereferences.
- Removed unused method in NullShader.
