Previously, walls were all numbered at 21000+. This is bad because
our blockmap is an array, so this caused our blockmap to take up
80 MiB of RAM in each worker process.
This commit changes the wall numbering, and exploits some bitmasking
to have the wall check run in constant time no matter how many walls
there are. This is done with a simple mask and xor to check the prefix.
RAM usage for the blockmap thus drops to like 44 MiB.
This is in preparation for adding more walls for 1.16.
Instead of trusting that Minecraft will only request up to index
num_palette_entries - 1, we'll assume Mojang occasionally produces
a weird file where it tries to request past that, up to the maximum
palette index it can based on the bits per value. To facilitate this,
this commit extends the translation palette array to this maximum size.
Such invalid palette picks will simply be replaced by ID 0 (aka air),
so it should cause minimal weirdness in the output.
Works around #1797.
Minecraft 1.16.1 loves writing these in singleplayer, and they
cause a whole bunch of corruption warnings if we don't handle them
in a special way like this.
Minecraft occasionally generates chunks which are not yet lit.
In the past, I'd have said to just make them not render, but these
can sometimes be large areas of the world.
Instead, render them with full bright SkyLight. This looks less bad
than whatever is stored in the SkyLight property in these cases.
Closes#1787, probably. Only one person bothered providing a sample file.
Closes#1790.
If a chunk fails reading due to an underlying operating system
IO error, we can invoke the retry logic too. No clue what could
cause this to happen beyond hardware faults.
Also lower the sleep time to 0.25s down from 0.5s because half a
second of sleeping for each chunk this happens on seems incredibly
overkill.
Making the nbt code raise a CorruptChunkError in this case also means
that failing to read a chunk due to an IOError is not a fatal error
anymore, it'll simply skip it and move on.
