This required the following:
- A generation task (taskA) to already be running for any chunk (chunkA)
- A chunk (chunkB) is requested for generation, and the task (taskB) to do the generation
is commenced immediately
- chunkB generation promise is aborted (e.g. due to chunk unload) and
taskB is orphaned
- chunkB is subsequently re-requested, but ends up in the generation
queue because taskB is still running
- taskA completes and drains the generation queue
- chunkB attempts to be populated a second time, but taskB has not yet
been collected, resulting in an assertion failure.
This bug has been appearing intermittently ever since PM 4.0 release.
For most users there is no obvious effect since production servers don't
have assertions enabled; however, it's unclear what kind of weird side
effects this bug may have had.
we don't usually add VanillaItems entries for blocks since they already exist in VanillaBlocks, but air has a special use case specifically as an itemstack, so we make an exception for this case.
this fixes a wide range of blocks with invalid blockstates becoming update! blocks on the client.
The most common occurrence of this was air with nonzero metadata left behind by world editors which set blockIDs but not block metadata. This caused large ghost structures of update! blocks to appear from nowhere.
The performance impact of this is very minimal (20 microseconds per chunk load in timings, compared to average 660 microseconds to load tiles).
if the worker selected previously had a generator registered, but has since been shutdown, the workerStartHook that cleans up generatorRegisteredWorkers won't yet have been called.
This results in the worker being started by the submission of PopulationTask, and the generator doesn't get preemptively registered.
if this happened, the index would stay set in activeChunkPopulationTasks, eventually causing the generation queue to jam up completely and non-forced generation to come to a standstill.
When profiling this, I noticed that we spend a stupidly large amount of time creating useless Position objects in the case of update=true, because Vector3->sides() calls Position->getSide(), which calls Position::fromObject(parent::getSide()). This is stupid because the update logic doesn't require Positions anywhere (as evidenced by this change needing no other alterations.
A rough profile shows that this improves setBlock() performance by about 25% in the update=true case, which is a pretty big margin.
As an added bonus, it gets rid of some unrealized cyclic dependencies in World->changedBlocks.
if the request/cancel/re-request happens all in the time before the queue gets drained, chunk hashes may appear multiple times in the queue. We don't want to process them twice if this happens (although it's mostly harmless anyway).
