GitLab

In preparation for working on Inko full-time, I'm adding this note. This
should hopefully prevent people from spending a lot of time on changes,
only for me to turn them down. In the long term we may relax the
contribution policy, if we find it won't cost us more effort than it
brings benefits.

Rust 1.50 includes some changes related to UNIX file descriptors,
essentially making -1 a reserved value. This means we need to use a
different value to indicate a file that has already been closed. We
don't want to use Option<File> here as this would require double
checking: once to unwrap the option, and once at the OS level to check
if the file descriptor is valid. If/when we move to single ownership
this won't be a problem, as we can just drop the entire File when an
object is dropped.

In addition, we fix some atomic operation deprecations, due to Rust 1.50
deprecating the use of compare_and_swap() in favour of
compare_exchange().

In practise these methods are only used with booleans. By defining these
methods only on Boolean, objects are free to define their own methods
using these names.

This reverts commit 17a342b6.

Converting an Array to an Array doesn't make much sense, as the input is
already an Array. In addition, users may think `to_array` creates a new
Array when in reality it just returns `self`. Cloning the entire array
and its values would be expensive, and wouldn't bring any benefits.

Error no longer requires a `message` method, instead only requiring the
`ToString` implementation. This simplifies implementing the trait for
custom error types, and means the user doesn't have to ask themselves
whether they should use `to_string` or `message`.

This is much simpler compared to the previous perfect hashing approach,
and also performs better due to not needing as many method calls. In the
future the compiler may optimise this further, but it's good enough for
the time being.

This reduces the binary size by about 200 KB, and is possible now that
we no longer catch Rust panics in the VM. In addition, aborting
immediately ensures we don't end up making the panic worse by running
cleanup routines or other code.

In dev mode we still use panic=unwind, as this may make debugging in
certain cases easier.

The std::test module has been refactored to no longer depend on global
state. Instead, each module must define a `tests` method that takes a
`Tests` type as its argument. This type is used for registering test
groups and tests, without relying on global state (using a process). In
addition, the assertion methods are now located in std::test itself, and
use the naming pattern `assert_X`. They also throw instead of panicking,
requiring the use of `try`.

These changes are made to reduce the places where we panic, to
provide a nicer unit testing interface, and to not rely on global state
(provided through a process).

As part of these changes, support for panic handlers has been removed.
Panics are meant for errors that you can't recover from, but panic
handlers contradict that goal. In addition, they introduce a fair bit of
complexity, and were never truly guaranteed to run. The removal of panic
handlers simplifies the VM, makes it more clear what panics are for,
and allows the VM to be compiled with `panic=abort`.

To handle panics, one should use a system monitoring service of some
sort, such as runit or systemd. Because panics are the result of
software bugs, you can't handle these at runtime. This means the only
sane response is to log it somewhere, and restart the entire program. A
well written program won't panic, so this shouldn't pose a problem in
practise. In addition, external monitoring solutions will always be more
reliable than an in-process monitoring/logging solution, as they aren't
affected by the mayhem caused in the process being monitored.

This fixes #205.

Now that std::env.executable is available, the fixture method
"inko_executable" is unnecessary.

This method allows one to get the path to the current executable.

This adds std::os::Command and various related types. The Command type
is used as a builder type for spawning a child process. The API is
loosely based on Rust's Command API, albeit with some differences here
and there.

The type std::io::Error now is a custom class, wrapping both an IO error
code and message. The error code makes it easier to determine at runtime
what lead to a failure. Storing the message in addition to the error
allows for custom error messages where needed.

This fixes #236

The instructions moved over are instructions that aren't frequently
used, or likely can't be made faster in the future with a JIT (when/if
we add one). As such, these have been moved out of the interpreter loop,
reducing the instruction set to 98 instructions.

Inko's instruction set was nearing almost 190 instructions, many of
which are high-level instructions (e.g. writing to STDOUT). We want to
add the ability to spawn OS commands, which would need (based on our
current estimates) around 5 VM instructions.

Every time an instruction is added, it adds more code to the main
interpreter loop. We also need to extend the compiler to support the new
instructions, and only have limited type-checking applied when these
instructions are used.

To solve these problems, we introduce the concept of external functions.
An external function is a function with a fixed signature, registered in
the VM to a name. The Inko runtime can load these functions using the
ExternalFunctionLoad instruction, then call them using the
ExternalFunctionCall instruction. Loading these functions is done using
the syntax `extern def` like so:

    extern def stdout_write_string(input: String) !! String -> Integer

These functions are then called ...

It's now less extreme/strict about the use of keyword arguments, based
on our experience with Inko's standard library use of them thus far.

Lambdas now allow the use of the keywords `return`, `try`, and `throw`.
This removes the need for the more verbose `local try` and `local throw`
keywords, and allows early returns in lambdas; something that wasn't
possible before.

The CLI wouldn't set the exit status if a program terminated without a
panic or CLI related error (e.g. an invalid flag).

Having to type out "lambda" makes using lambdas annoying. Using "fn"
makes this more pleasant, and keeps the keyword in line with the "do"
keyword used for closures.

The Inko type checker now correctly disallows the use of the constructor
syntax for built-in types, such as ByteArray and Array.

Back when Inko was more dynamic/prototype based, the use of `object`
made more sense. As we're planning on changing the memory layout to
resemble a more traditional class based approach, the use of "object"
instead of "class" can be confusing. In addition, users coming from
other languages may wonder why we're not just using "class" instead of
"object".

This commit changes the "object" keyword to "class", solving these
problems.

When using a closure or lambda in a type signature, leaving out the
return type infers it as Any; not Nil. Inferring the type as Nil can
lead to soundness problems, such as this:

    import std::mirror
    import std::stdio::stdout

    def example(callback: do) -> Nil {
      callback.call
    }

    let result = example { 10 }

    stdout.print(mirror.reflect(result).inspect)

This will print `10` to STDOUT, instead of `Nil`. By inferring
`callback` as `do -> Any` instead of `do -> Nil` we can solve this, as
the above `example` method wouldn't compile (as `Any` isn't compatible
with `Nil`).

When using type parameters inside a block, other types were incorrectly
treated as being compatible with the parameter (assuming its
requirements are met). This can lead to soundness issues, such as the
following:

    def push!(A)(values: Array!(A)) {
      values.push(10)
    }

    push(Array.new('foo'))

Here `push` is unsound because the exact type of `A` isn't known until
runtime, meaning it's not safe to push `10` into the array.

To solve this, we introduce the concept of "rigid type parameters".
These are type parameters that can't be further inferred as/assigned new
types, and other types aren't compatible with these type parameters.

The term "rigid type parameter" originates from Haskell. Rigid type
parameters are also called "skolem type parameter", but we don't use
that term because it's unlikely to be widely known or understood.

When generating the type name of a type parameter, the name of the type
parameter is now always included. This makes it a bit easier to read
type errors, as you're not left wondering if a type is a trait or a type
parameter.

This adds support for looking up type parameter methods in the Inko type
checker.

Early local returns in closures/lambdas are basically impossible to use,
as all control flow goes through closures and thus can't return from
nested closures (but not the surrounding method).

This commit removes all support for `local return`, in favour of just
using implicit returns.

While still experimental, there may be cases where it can prove useful.

The contents of std::ffi::types have been moved into std::ffi. The
various Type constants (VOID, Pointer, etc) have all been removed in
favour of the module methods that returned these types. The
corresponding unit tests have been updated to not depend on constants
containing complex objects (e.g. an FFI library).

In preparation for #212, a
variety of constants containing non-primitive data (e.g. arrays) have
been refactored away. Some cases remain, such as
std::test::DEFAULT_CLIENT. These requires further changes that we'll
make in separate commits.

Similar to Array.filled(), this method is used to create a ByteArray
filled with a given byte.

While semantically ?T and Option!(T) are the same, syntactically they
are not. In addition, desugaring ?T at the parser level means a consumer
of the AST can't determine if ?T or Option!(T) was used.

This changes the self-hosting parser so it produces an OptionType AST
node, instead of desugaring ?T into a Constant node.