Also tried to sort them by renderer use but that turned out to be without any performance effect, even though the struct is a bit better aligned now and several bytes shorter.
I have to wonder why it had to use such a complicated implementation that provided no advantages whatsoever.
The new code is just 1/5th of the old one's size and much closer to Hexen's original implementation which also was a simple array but with no means to resize the queue.
This involves passing the level explicitly to many functions. What was done here may seem a bit excessive but at least it covers everything.
Most importantly, the global ActiveThinker pointer has been moved into FLevelLocals and is now getting tracked properly by the level without using dangerous assumptions about how the game organizes its data.
- upload only the dynamic AABB subtree using glBufferSubData
- change internal raytracing stack limit from 16 to 32
- update shadowmap AABB tree after R_SetupFrame for proper frame interpolation
Unlike the other classes, the places where variables from this class were accessed were quite scattered so there isn't much scriptified code. Instead, most of these places are now using the script variable access methods.
This was the last remaining subclass of AActor, meaning that class Actor can now be opened for user-side extensions.
This was the only code using the ViewBob member variable.
This also moves the range check for this variable to its application, because a badly behaved mod can just as easily change it at run time instead of just setting an absurdly large value in the class definition.
Since these can be changed on the placed light actor they have to be read from there, so this is now a pointer in FDynamicLight, just like the other properties that can be user-changed.
Also did some cleanup on the interface so that external code doesn't need to dereference the lightflags pointer but can use utility functions for all flags.
backtrace functions are not present in all libc implementations. Cmake has
module to add external libraries into build if needed so use it to fix build on
Unix systems without backtrace in libc.
The current behaviour offsets to the front of the actor rather than the side, due to an oversight in the code, which oddly is not present in the A_FireBullets equivalent.
ReadString allocates a buffer, so saving it in a local variable and then forgetting it will not free the buffer afterward.
(This should probably be refactored to use some safer methods to read the string than this old-school method...)
This should be less of a drag on the playsim than having each light a separate actor. A quick check with ZDCMP2 showed that the light processing time was reduced to 1/3rd from 0.5 ms to 0.17 ms per tic.
It's also one native actor class less.
The loop never checked if the item was still valid and would continue to try to use it, even after it was removed from the inventory and destroyed.
As native code this just failed silently, but with the VM it needs to be explicitly checked.
