This seems very little cost and matches the old behavior more
closely. This will cause some more falling nodes to get added
to falling clusters. With the efficiency of the algorithm, this
really doesn't do much damage.
This took me a while to figure out. We no longer visit all 9 block
around and with the touched node, but instead visit adjacent plus
self. We then walk -non- recursively through all neigbors and if
they cause a nodeupdate, we just keep walking until it ends. On
the way back we prune the tail.
I've tested this with 8000+ sand nodes. Video result is here:
https://youtu.be/liKKgLefhFQ
Took ~ 10 seconds to process and return to normal.
2016-03-27 15:47:01: ERROR[Main]: ServerError: Lua: Runtime error from mod '*builtin*' in callback luaentity_Step(): Node name is not set or is not a string!
2016-03-27 15:47:01: ERROR[Main]: stack traceback:
2016-03-27 15:47:01: ERROR[Main]: [C]: in function 'add_node'
2016-03-27 15:47:01: ERROR[Main]: /usr/share/minetest/builtin/game/falling.lua:96: in function </usr/share/minetest/builtin/game/falling.lua:43>
Commit
65c09a96f4 "Set acceleration only once in falling node"
has made the acceleration being set only once.
But this has introduced a regression.
Fix#3884.
