Adds a basic bucketing system to move loops.
This should hopefully save a lot of cpu time, and allow for more load while gaining better smoothness.
The idea is very similar to SStimer, but my implementation is much more simple, since I have to worry less about long delays and redundant buckets.
Insertion needs to be cheaper too, since I'm making a system that by design holds a lot of looping things
It comes with some minor tradeoffs, we can't have constant rechecking of loops if a move "fails", not that we really want that anyway
We also lose direct control over the timer var, but I think that's better, don't want people manipulating that directly
Not that it even really worked very well back when we did have it
Removes the sleep from singularity code
Rather then using sleep to store the state of our iteration, we instead queue the iteration in a list.
We then use a custom singulo processing subsystem to call our "digest" proc several times per full eat, with the hope of staying on top of
our queue
This rarely happens because the queue is too large, god why is a sm powered singulo 24x24 tiles.
I've also A: cached our dist checks, and B: Added dist checks to prevent attempting to pull things out of range
This might look a bit worse, but it saves a lot of work
Oh right and I made the singulo unable to eat while it still has tiles to digest. The hope is to prevent
overwork and list explosion.
Hopefully this will prevent singulo server stoppage, though I've seen some other worrying things in testing.
See title. Also refactors caltrops into a component because they use connect_loc_behalf which requires them to hold the state.
This also fixes COMPONENT_DUPE_SELECTIVE from just outright not working.
connect_loc_behalf doesn't make sense as an element because it tries to hold states. There is also no way to maintain current behaviour and not have the states that it needs.
Due to the fact that it tries to hold states, it means the code itself is a lot more buggy because it's a lot harder to successfully manage these states without runtimes or bugs.
On metastation, there is only 2519 connect_loc_behalf components at roundstart. MrStonedOne has told me that datums take up this much space:
image
If we do the (oversimplified) math, there are only ever 5 variables that'll likely be changed on most connect_loc_behalf components at runtime:
connections,
tracked,
signal_atom,
parent,
signal_procs
This means that on metastation at roundstart, we take up this amount: (24 + 16 * 5) * 2519 = 261.97600 kilobytes
This is not really significant and the benefits of moving this to a component greatly outweighs the memory cost.
(Basically the memory cost is outweighed by the maint cost of tracking down issues with the thing. It's too buggy to be viable longterm basically)
* Makes turfs persist signals
* Splits connect_loc up into two elements, one for stuff that wishes to connect on behalf of something, and one for stuff that just wants to connect normally. Connecting on behalf of someone has a significant amount of overhead, so let's do this to keep things clear
* Converts all uses of connect_loc over to the new patterns
* Adds some comments, actually makes turfs persist signals
* There's no need to detach connect loc anymore, since all it does is unregister signals. Unregisters a signal from formorly decal'd turfs, and makes the changeturf signal persistance stuff actually work
* bro fuck documentation
* Changes from a var to a proc, prevents admemems and idiots
* Extra detail on why we do the copy post qdel
Enter(), Entered(), Exit() and Exited() all passed the old loc forward, but everything except a single a case cared about the direction of the movement more than about the specific source.
Since moving multi-tile objects will have multiple sources of movement but a single direction, this change makes it easier to track their movement.
Cleaned up a lot of code around and made proc inputs compatible.
I'll add opacity support for multi-tile objects in a different PR after this is merged, as this has grown large enough and I don't want to compromise the reviewability.
Tested this locally and as expected it didn't impair movement nor produced any runtimes.
