See the discussion in #1352 - Netty uses the system one by default,
so no sense creating our own.
Also make sure we through the HTTP error every time, not just on the
first failure. Otherwise we get cryptic connection dropped errors.
Given an input like f(x), which is both a valid statement and
expression, both parsers would accept the whole input. However, this was
treated the same as both parsers rejecting the input, resulting in a
crash when trying to print the error.
We now return immediately when any parser accepts the input.
Fixes#1354
If someone had a recursive table (created with an IIFE), then we'd throw
an error inside reserialize. We now catch this error and silently drop
the value.
I'm not thrilled by this behaviour - there's an argument we should
return false instead - but it's consistent with what we currently do.
Closes#1337.
Previously it was possible to access all methods of the multishell
redirect by calling term.current(). This is definitely not intended, as
it leaks all sorts of internals to the user.
Also bump illuaminate - the new version is about twice as fast on my
machine.
Woops!
- Fix the REPL not printing values, as exception.try didn't return
values. It did originally, and then I tried to simplify it >_>
- Change repl_exprs to run an expression and program parser in
parallel, rather than handling the parallelism on the grammar side -
that has a few shift/reduce conflicts which result in bad parse
errors.
- Bump Cobalt to 0.6.0. We now track both line and column numbers of
each bytecode instruction, allowing us to map an error to a concrete
position.
- `loadfile` (and similar functions) now use the full path, rather than
the file name. Cobalt truncates this to 30 characters (rather than
the previous 60) so this should be less noisy.
- The shell, edit and Lua REPL now display the corresponding source
code alongside an error.
Note this is incredibly limited right now - it won't cope with errors
which cross coroutine boundaries. Supporting this is on the roadmap,
but requires some careful API design.
- Encode the DFA as a virtual machine (identical to lrgrep) rather than
compiling it to a series of Lua functions. While this is a little
slower and uglier, it's much more space efficient, shaving off 16Kb.
- Minimise the DFA properly. This only shaves off a few states, but
every little helps.
- Run the error handling code from a non-reduced parser stack. This was
incredibly nasty to get right (and positions are still not correctly
handled), but it fixes several broken error messages.
- Move modem recipes out of the usage section.
- Add missing argument names to BinaryWriableHandle.write. Illuaminate
really should catch this, but for now I did a grep and couldn't find
any more instances of this.
- Add several (internal) modules for lexing and parsing Lua code. These
allow us to provide (hopefully) higher quality error messages than
Lua's built-in messages.
- `shell.run`, `edit` and `lua` now use this parser when fed invalid
code. This allows us to provide better syntax errors, while not
having any impact on the happy path.
Note this does not affect any other mechanism for loading code
(`load`, `require`, `dofile`).
There's still a lot of work to do here in improving error message
quality, but hopefully this provides a good starting point.
Just ran[^1] over the codebase. Turns out we'd duplicated one of the
changelog entries entirely - I suspect due to a version merge gone
wrong!
[^1]: https://github.com/crate-ci/typos/
I originally put cc.import in a separate directory from the main
modules. This means that programs must extend the package path in order
to import these modules.
However, this ends up being a mixed blessing: while it makes it much
harder for users to accidentally require user code, it also means we
can't expose a public interface which wraps a private module.
Instead, cc.import now lives on the main package path, but lives under
the cc.internal namespace and is not documented anywhere. Hopefully this
should be enough of a clue that one shouldn't use it :p.
I've been meaning to fix this for over 6 years, and just kept
forgetting.
Previously ILuaContext.executeMainThreadTask worked by running
ILuaContext.issueMainThreadTask, pulling task_complete events, and then
returning the results.
While this makes the implementation simple, it means that the task's
results were converted into Lua values (in order to queue the event) and
then back into Java ones (when the event was pulled), before eventually
being converted into Lua once more.
Not only is this inefficient, as roundtripping isn't lossless, you
couldn't return functions or rich objects from main thread functions
(see https://github.com/dan200/ComputerCraft/issues/125).
We now store the return value on the Java side and then return that when
the receiving the task_complete event - the event no longer carries the
result. Note this does not affect methods using issueMainThreadTask!
- Correctly handle FileOperationExceptions for the root mount.
- Remove some checks from MountWrapper: Mount/WritableMount should do
these already!
- Normalise file paths, always using a '/'.
- Separate FileMount into separate FileMount and WritableFileMount
classes. This separates the (relatively simple) read-only code from
the (soon to be even more complex) read/write code.
It also allows you to create read-only mounts which don't bother with
filesystem accounting, which is nice.
- Make openForWrite/openForAppend always return a SeekableFileHandle.
Appendable files still cannot be seeked within, but that check is now
done on the FS side.
- Refactor the various mount tests to live in test contract interfaces,
allowing us to reuse them between mounts.
- Clean up our error handling a little better. (Most) file-specific code
has been moved to FileMount, and ArchiveMount-derived classes now
throw correct path-localised exceptions.
- Flip http.websocket and http.websocketAsync docs (fixes#1244)
- Fix http.request queuing a http_failure event with no URL when
passing a malformed URL
- Fix http.websocketAsync not queuing websocket_failure events on
immediate failure.
In classic squid tradition: 20% code, and 80% test logic.
Closes#962. Alas, whoever reported this has deleted their account, so
they can't even be happy about it :(.
