Previously, many concatenated texture field strings were turned to text, then had that string run through get_number, then had it converted back into a string to become a stringarg.
Now, the concatenated string is copied directly to the relevant stringarg, with no intermediary steps.
This fixes an issue where a map with object or state properties would have "context drift" - breaking when the object or state list changed, due to differently ordered freeslots, or new hardcoded objects.
Affected types:
- doomednum 1110/Action 9 (Custom Mace/Chain)
- doomednum 700 and those immediately following (Ambient Sound)
- Action 4 and 414 (dash pad, play sound effect)
- Action 14 (bustable block parameters)
- Action 434 (Award Power)
- doomednum 757/Action 15 (fan particle generator)
- doomednum 1202 (bumpable hazard rock/apple spawner)
- This one has undefined behaviour in the binary map format which was not previously forbidden. This undefined behaviour is EXTREMELY vulnerable to context drift, and so it's simply not worth creating a system to write numerical values into object types - we write an explicit name only for the intended range, and otherwise report the threat of context drift when converting.
In addition, to reduce duplicated zone memory, (sidedef_t).text and (linedef_t).text have been removed from all but the Lua API. In Lua, in binary maps, they point to the host line's stringargs - the line's text and a frontside's text will return stringargs[0], while a backside's text will return stringargs[1]. I've done my best to match the previous API as closely possible, to the point of reporting false nils if the line didn't previously have text available.
However, there are four linedef Actions for which the sidedef text will be different between builds containing and not containing this commit - 331, 332, 333 (the Character-specific linedef executors), and 443 (Call Lua Script), and only if the text is long enough to go over two lines. Given that both halves would be incomplete nonsense in this case, I'm willing to wager this minor point of discrepancy is not a breaking issue.
Pretty much removed everything except for time operations. File operations steps on the toes of the IO library, and the others were useless or invasive. The only one I wasn't sure about was `os.tmpname()`, but I think that's something that could be implemented for IO library instead if necessary.
The main reason I want this is because os.time provides an OK way to initially seed RNG on game first boot. But it would also be cute to let Lua modders be able to re-implement Christmas mode / day&night cycles themselves :p
- Prevent connection timeout during the waiting gamestate from the last commit.
- Keep client connections alive during fades / other internal loops.
- More consistently timeout clients when they reach the end of BACKUPTICS.
- Dedicated servers will not run any game logic if no nodes are sending packets to it, to reduce CPU usage when there is no one interacting with your server.
- Unlike SRB2Kart, the amount of time is configurable with the "dedicatedidletime" console variable. Setting this to 0 will disable this feature.
- CL_SendClientCmd uses exact packet types instead of magic number offsets.
All gamestates besides GS_LEVEL are unsupported by the save game functions. This commit forces players joining during these gamestates into GS_WAITINGPLAYERS, which is a basic gamestate that just maintains the connection until we can enter the start of a new one. Also provides an extremely simple drawer for GS_WAITINGPLAYERS so the joining player knows what's going on.
