This is an attempt to enforce better separation between ComputerThread
and ComputerExecutor. Both of these classes are pretty complex in their
own right, and the way the two bleed into each other makes it all the
more confusing!
This effectively splits the ComputerExecutor into two separate classes:
- ComputerScheduler.Executor (with the actual implementation inside
ComputerThread): This holds all the ComputerThread-related logic
which used to be in ComputerExecutor, including:
- before/after work hooks
- is-on-thread tracking
- virtual runtime computation
- ComputerScheduler.Worker: This encapsulates all the computer-related
behaviour. The actual implementation remains in ComputerExecutor.
The boundaries are still a little fuzzy here, and it's all definitely
more coupled then I'd like, but still an improvement!
There are several additional changes at the same time:
- TimeoutState has also been split up, to better define the boundary
between consumers (such as ComputerExecutor and ILuaMachine) and
controllers (ComputerThread).
The getters still live in TimeoutState, but the core logic lives in
ManagedTimeoutState.
- We no longer track cumulative time in the TimeoutState. Instead, we
allow varying the timeout of a computer. When a computer is paused,
we store the remaining time, and restore it when resuming again.
This also allows us give a longer timeout for computer
startup/shutdown, hopefully avoiding some of those class-not-found
issues we've seen.
- We try to make the state machine of how ComputerExecutors live on the
queue a little more explicit. This is very messy/confusing -
something I want to property test in the future.
I'm sure there's more to be done here, especially in ComputerExecutor,
but hopefully this makes future changes a little less intimidating.
Allows registering arbitrary block lookup functions instead of a
platform-specific capability. This is roughly what Fabric did before,
but generalised to also take an invalidation callback.
This callback is a little nasty - it needs to be a NonNullableConsumer
on Forge, but that class isn't available on Fabric. For now, we make the
lookup function (and thus the generic peripheral provider) generic on
some <T extends Runnable> type, then specialise that on the Forge side.
Hopefully we can clean this up when NeoForge reworks capabilities.
This is the second time I've rewritten our class generation in a little
over a month. Oh dear!
Back in d562a051c7 we started using method
handles inside our generated ASM, effectively replacing a direct call
with .invokeExact on a constant method handle.
This goes one step further and removes our ASM entirely, building up a
MethodHandle that checks arguments and then wraps the return value.
Rather than generating a class, we just return a new LuaFunction
instance that invokeExacts the method handle.
This is definitely slower than what we had before, but in the order of
8ns vs 12ns (in the worst case, sometimes they're much more comparable),
so I'm not too worried in practice.
However, generation of the actual method is now a bit faster. I've not
done any proper benchmarking, but it's about 20-30% faster.
This also gives us a bit more flexibility in the future, for instance
uisng bound MethodHandles in generation (e.g. for instance methods on
GenericSources). Not something I'm planning on doing right now, but is
an option.
Or rather, being published to the wrong place. The java-convention
plugin sets the group, but that was applied after the publishing one - I
was hoping it'd read that property lazy, but clearly not!
Wow, some of this is /old/. All the Maps.newHashMap stuff dates back to
Java 6, so must originally be CCTweaks code?!
We're unlikely to drop our Guava dependency (we use too much other
stuff), but we should make the most of the stdlib where possible.
This should be significantly faster than LoadingCache (2.5x in my
benchmarks, but not sure they're representative). This isn't super
important - a lookup only takes 6us - but still worth using!
Previously we had the invariant that if we had a server monitor, we also
had a terminal. When a monitor shrank into a place, we deleted the
monitor, and then recreated it when a peripheral was requested.
As of ab785a0906 this has changed
slightly, and we now just delete the terminal (keeping the ServerMonitor
around). However, we didn't adjust the peripheral code accordingly,
meaning we didn't recreate the /terminal/ when a peripheral was
requested.
The fix for this is very simple - most of the rest of this commit is
some additional code for ensuring monitor invariants hold, so we can
write tests with a little more confidence.
I'm not 100% sold on this approach. It's tricky having a double layer of
nullable state (ServerMonitor, and then the terminal). However, I think
this is reasonable - the ServerMonitor is a reference to the multiblock,
and the Terminal is part of the multiblock's state.
Even after all the refactors, monitor code is still nastier than I'd
like :/.
Fixes#1608
Currently redirects would be returned from the proxy, and then
immediately followed by XMLHTTPRequest. The proxy now follows requests
(when requested), so that should no longer happen.
We should probably switch over to fetch(...) here, to allow setting
follow_redirects to false, but that's a job for another day.
Haha, so many web emulator related commits of late. This'll die down
soon.
We can't use FriendlyByte.readCollection to read to a
pre-allocated/array-backed NonNullList, as that doesn't implement
List.add. Instead, we just need to do a normal loop.
We add a couple of tests to round-trip our recipe specs. Unfortunately
we can't test the recipes themselves as our own registries aren't set
up, so this'll have to do for now.
- Update to Rollup 4.x
- Replace terser and postcss with swc and lightningcss. This is
definitely more code for us to write (maybe I should turn them into
proper plugins we can depend on), but both speedier and fewer
dependencies.
- Drop dependency on glob - we can get away with fs.readdir for what we
needed it for.
Oh, this was a really nasty bug to reproduce. I'm not sure why - it's
very simple - I guess I've only just seen screenshots of it, and never
sat down to try myself. Reminder to actually report your bugs folks!
In this case:
1. Place down three down three monitors and then a computer.
2. Display something on the monitor (monitor left paint a) is my go-to.
3. Break the middle monitor.
We'd expect the left most monitor to be cleared, however it actually
preserves the monitor contents, resizing (and skewing it) to fit on its
new size!
This is because we clear the server monitor, but never sync that over to
the client, so the client monitor retains the old contents. To fix that,
instead of nulling out the server monitor, we null out the underlying
Terminal. This causes the change to be synced, fixing the bug.
Paint implements its menu slightly differently to edit, in that it takes
control of the event loop until the menu is closed. This means that the
term_resize event is ignored, and so the canvas not redrawn when the
menu is open.
Historically we've used copy-cat to provide a web-based emulator for
running example code on our documentation site. However, copy-cat is
often out-of-date with CC:T, which means example snippets fail when you
try to run them!
This commit vendors in copy-cat (or rather an updated version of it)
into CC:T itself, allowing us to ensure the emulator is always in sync
with the mod.
While the ARCHITECTURE.md documentation goes into a little bit more
detail here, the general implementation is as follows
- In project/src/main we implement the core of the emulator. This
includes a basic reimplementation of some of CC's classes to work on
the web (mostly the HTTP API and ComputerThread), and some additional
code to expose the computers to Javascript.
- This is all then compiled to Javascript using [TeaVM][1] (we actually
use a [personal fork of it][2] as there's a couple of changes I've
not upstreamed yet).
- The Javascript side then pulls in the these compiled classes (and
the CC ROM) and hooks them up to [cc-web-term][3] to display the
actual computer.
- As we're no longer pulling in copy-cat, we can simplify our bundling
system a little - we now just compile to ESM modules directly.
[1]: https://github.com/konsoletyper/teavm
[2]: https://github.com/SquidDev/teavm/tree/squid-patches
[3]: https://github.com/squiddev-cc/cc-web-term
This moves MemoryMount to the main core module, and converts it to be a
"proper" WritableMount. It's still naively implemented - definitely
would be good to flesh out our tests in the future - but enough for what
we need it for.
We also do the following:
- Remove the FileEntry.path variable, and instead pass the path around
as a variable.
- Clean up BinaryReadableHandle to use ByteBuffers in a more idiomatic
way.
- Add a couple more tests to our FS tests. These are in a bit of an odd
place, where we want both Lua tests (for emulator compliance) and
Java tests (for testing different implementations) - something to
think about in the future.
- Move the frontend code into src/frontend
- Move our custom element SSR system into src/htmlTransform.
This is mostly in prep for merging in copy-cat's core, as that's a whole
bunch of extra code.
This attempts to reduce some duplication in recipe serialisation (and
deserialisation) by moving the structure of a recipe (group, category,
ingredients, result) into seprate types.
- Add ShapedRecipeSpec and ShapelessRecipeSpec, which store the core
properties of shaped and shapeless recipes. There's a couple of
additional classes here for handling some of the other shared or
complex logic.
- These classes are now used by two new Custom{Shaped,Shapeless}Recipe
classes, which are (mostly) equivalent to Minecraft's
shaped/shapeless recipes, just with support for nbt in results.
- All the other similar recipes now inherit from these base classes,
which allows us to reuse a lot of this serialisation code. Alas, the
total code size has still gone up - maybe there's too much
abstraction here :).
- Mostly unrelated, but fix the skull recipes using the wrong UUID
format.
This allows us to remove our mixin for nbt in recipes (as we just use
our custom recipe now) and simplify serialisation a bit - hopefully
making the switch to codecs a little easier.
- Add AbstractInMemoryMount, which contains all of ArchiveMount's file
tree logic, but not the caching functionality.
- Convert MemoryMount to inherit from AbstractInMemoryMount.
- Add a helper method to add a file to an AbstractInMemoryMount, and
use that within {Resource,Jar}Mount.
There's definitely more work to be done here - it might be nice to split
FileEntry into separate Directory and File interfaces, or at least make
them slightly more immutable, but that's definitely a future job.
- Add a new WebsocketClient interface, which WebsocketHandle uses for
sending messages and closing. This reduces coupling between Websocket
and WebsocketHandle, which is nice, though admitedly only use for
copy-cat :).
- WebsocketHandle now uses Websocket(Client).isClosed(), rather than
tracking the closed state itself - this makes the class mostly a thin
Lua wrapper over the client, which is nice.
- Convert Options into a record.
- Clarify the behaviour of ws.close() and the websocket_closed event.
Our previous test was incorrect as it called WebsocketHandle.close
(rather than WebsocketHandle.doClose), which had slightly different
semantics in whether the event is queued.
We already use preact for the copy-cat integration, so it makes sense to
use it during the static pass too. This allows us to drop a dependency
on react.
- Placing a command computer requires the player to be in creative and
opped.
- Breaking a command computer now requires the player to be opped, as
well as in creative.
As we've now got a dedicated item class for command comptuers, we move
the command-specific IMedia override to that class.
Fixes#1582.
As this is responsible for interrupting computers, we should make sure
its priority is higher than the background threads. It spends most of
its time sleeping, so should be fine.
Rather than having a mess of lambdas, we now move the bulk of the
implemetation to their own methods. The lambdas now just do argument
extraction - it's all stringly typed, so good to keep that with the
argument definition.
This also removes a couple of exception keys (and thus their translation
keys) as we no longer use them.
When the target method is in a different class loader to CC, our
generated method fails, as it cannot find the target class. To get
around that, we create a MethodHandle to the target method, and then
inject that into the generated class (with Java's new dynamic constant
system). We can then invoke the MethodHandle in our generated code,
avoiding any references to the target class/method.
I removed this in aa0d544bba, way back in
late 2021. Looks like it's been updating in the meantime and I hadn't
noticed, so add it back.
I've simplified the code a little bit, to make use of our new capability
helpers, but otherwise it's almost exactly the same :D.
Wow, this is old. It looks like it's a legacy of when this method was on
TileGeneric (and so returned false by default). As all implementations
now return true (turtle tools no longer block redstone), we don't really
need this any more.
- Split buttons.png into individual textures.
- Split corners_xyz.png into the following:
- borders_xyz.png: A nine-sliced texture of the computer borders.
- pocket_bottom_xyz.png: A horizontally 3-sliced texture of the
bottom part of a pocket computer.
- sidebar_xyz.png: A vertically 3-sliced texture of the computer
sidebar.
While not splitting the sliced textures into smaller ones may seem a
little odd, it's consistent with what vanilla does in 1.20.2, and I
think will make editing them easier than juggling 9 textures.
I do want to make this more data-driven in the future, but that will
have to wait until the changes in 1.20.2.
This also adds a tools/update-resources.py program, which performs this
transformation on a given resource pack.
- Add a generic PermissionRegistry interface. This behaves similarly to
our ShaderMod interface, searching all providers until it finds a
compatible one.
We could just make this part of the platform code instead, but this
allows us to support multiple systems on Fabric, where things are
less standardised.
This interface behaves like a registry, rather than a straight
`getPermission(node, player)` method, as Forge requires us to list
our nodes up-front.
- Add Forge (using the built-in system) and Fabric (using
fabric-permissions-api) implementations of the above interface.
- Register permission nodes for our commands, and use those
instead. This does mean that the permissions check for the root
/computercraft command now requires enumerating all child
commands (and so potential does 7 permission lookups), but hopefully
this isn't too bad in practice.
- Remove UserLevel.OWNER - we never used this anywhere, and I can't
imagine we'll want to in the future.
Should be max_websocket_message, not just websocket_message.
Also add some additional validation to address rules, to check no
unrecognised keys are present.
Closes#1566.
