Basically, they key difference between StonedMC and CarnMC is that when multiple ticks want to run at the same byond tick, we divvy up the tick between the subsystems, rather then allow one subsystem to hog it all.
The key difference between StonedMC and GoonPS is that we allow the subsystems to tell us how to divvy up the tick using flags and priority.
The new SS_ flags allows us to select behaviors that used to be piggybacked as side effects of dynamic wait or default but sometimes unneeded behavior.
Dynamic wait is 100% gone, lower priority and SS_BACKGROUND are better more refined ways of doing this when combined with MC_TICK_CHECK
I have by design never looked at the inners of goonPS, so this is all original code but I know it uses two loops because of comments by goon devs on reddit threads, that design didn't make sense before, but when I can tell a SS how much of a byond tick it is allowed to have, knowing how many need to run this tick is helpful I also know a bit more about how it works from piecing together comments in #vgstation.
Detailed list of changes:
Subsystems now have flags, allowing fine grain control over things like rather or not it processes, inits, rather it's wait is how long between runs (post run timing) or how long between starts, and rather or not late fires should cause the next fire to be earlier.
Mc now has two loops One loop handles queuing shit, one loop handles running shit.
MC now splits up tick allotment rather than first come first serve Subsystems can even request a bigger share using higher priorities. (It will even resume subsystems it paused if other subsystems hadn't used as much as it predicted they might need)
Default fps is now 20 This is related enough to the MC and it's a change that's really long since over due
All code oddities are most likely to be necessities to lower overhead on the mc since it runs every tick
Misc:
+Fixes unreported issue with initializing lighting on a specific zlevel
+Fixes two similar issues with moveElement and moveRange. Where fromIndex or toIndex could be adjusted incorrectly in certain conditions. Potentially causing bad-sorts, or out of bound errors.
+Rewrites listclearnulls(list/L) to no longer iterate through L.len elements for every null in the list (plus 1). i.e. went from L.len*(number_of_nulls+1) list-element reads (best-case), to L.len list-element reads (worst-case)
+New proc/getElementByVar(list/L, varname, value) which finds the first datum in a list, with a variable named varname, which equals value. You can also feed it atoms instead of lists due to the way the in operator functions.
+Fixes an unreported issue with Yota's list2text rewrite. Under certain conditions, the first element would not be converted into a string. Causing type-mismatch runtimes.
+New global map_ready variable. This is not fully implemented yet, but will be used to avoid duplicate calls to initialize() for map objects.
+All turfs now maintain references to all lights currently illuminating them. This will mean higher memory use unfortunately, due to the huge number of turfs. However, it will speed up updateAffectingLights significantly. I've used list husbandry to reduce baseline memory usage, so it shouldn't be any worse than some past atmos modifications memory-wise.
-Removed 'quadratic lighting', can add this back at some point. Sorry.
+modified the way lum() works slightly, to allow turfs to have overridden delta-lumen. i.e. space cannot be illuminated more than its default ambiance. This allowed removal of some iffy special-snowflake lighting areas implemented by somebody else.
+Lighting images in the dmi can now use arbitrary naming schemes. It is reliant on order now. This allows the dmi to be replaced by simply dropping in a new dmi.
-Removed all subtypes of /area/shuttle. Shuttles now create duplicate 'rooms' of /area/shuttle. (More on this later). This will conflict with most maps. Guide on how to fix to follow.
+All verbs/tools relating to world.tick_lag were refactored to use world.fps. However old config text for setting tick_lag will still work (it converts the value to fps for you)
+MC stats improved using smoothing. They now have their own tab so they dont get in the way when you're playing as an admin.
-removed the push_mob_back stuff due to conflicting changes. Sorry Giacom.
_OK, NOW THE ACTUAL INTERESTING STUFF_
Following systems moved over to subsystem datums:
air_master
garbage_manager
lighting_controller
process_mobs (aka Life())
nanomanager
power
sun
pipenets
AFK kick loops
shuttle_controller (aka emergency shuttle/pods), supply_shuttle and other shuttles
voting
bots
radio
diseases
events
jobs
objects
ticker
Subsystems hooks and variables should be commented fairly in-depth. If anything isn't particularly clear, please make an issue.
Many system-specific global variables have been refactored into
All tickers which previously used world.timeofday now use world.time
some subsystems can iterate before round start. this resolves the issue with votes not working pregame
RESULTS:
sorting 10 random lists of length 3 to 303 in increments of 3
(Meh, I forgot to refresh this one, there were only 338 trials rather than 1010, can't be bothered to recode the test)
Profile results (total time)
Proc Name Self CPU Total CPU Real Time Calls
------------------------------------------------- --------- --------- --------- ---------
/proc/sortList 0.672 16.141 16.243 171226 <--TG's current mergesort(recursive, hence the higher number of calls)
/proc/sortTim 0.008 3.278 3.274 338 <--TimSort
/proc/sortMerge 0.011 2.839 2.855 338 <--new mergesort
/proc/sortInsert 0.010 2.124 2.103 338 <--binary insertion
Sorting 10 presorted lists with 3 inversions (3 elements shuffled up), Lists of length 3 to 303 (increments of 3)
Profile results (total time)
Proc Name Self CPU Total CPU Real Time Calls
------------------------------------------------- --------- --------- --------- ---------
/proc/sortList 1.290 23.056 23.254 308050 <--rather cataclysmic
/proc/sortMerge 0.015 4.077 4.068 1010 <--
/proc/sortInsert 2.639 3.472 3.464 1010 <--
/proc/sortTim 0.014 1.567 1.576 1010 <--TimSort is faaar more effective in these cases,
Timsort can exploit runs effectively
sorting 10 presorted lists which have been reversed
Profile results (total time)
Proc Name Self CPU Total CPU Real Time Calls
------------------------------------------------- --------- --------- --------- ---------
/proc/sortList 1.234 23.193 23.295 308050
/proc/sortMerge 0.023 4.681 4.686 1010
/proc/sortInsert 2.875 3.750 3.765 1010
/proc/sortTim 0.020 3.294 3.284 1010 //This can be lower by using a different comparison method
*Corrected: /proc/sortTim 0.017 0.665 0.663 1010 //Using a non-strictly ascending comparison
sorting 10 presorted lists
Profile results (total time)
Proc Name Self CPU Total CPU Real Time Calls
------------------------------------------------- --------- --------- --------- ---------
/proc/sortList 1.199 21.391 21.517 308050
/proc/sortMerge 0.018 3.724 3.729 1010
/proc/sortInsert 2.497 3.302 3.309 1010
/proc/sortTim 0.024 0.586 0.584 1010
Summary, all the new procs are faster than the old ones. TimSort is ever so slightly slower than Insertion and Merging on random lists. But on lists with natural runs (partially sorted data) it is far faster than all others.
The old merge sort was removed and replaced with timSort. Other algorithms are provided as alternatives.
All algorithms use a central datum, so accept many of the same parameters. For instance, setting associative=1 will make them sort associative lists by their associated values, rather than keys.
They also accept a cmp argument. This allows sorting of lists of datums, text, numbers or whatever. The pre-existing helpers in lists.dm were rewritten as examples.
