This commit contains two improvements:
- Support for a Range syntax (and a corresponding Range value)
- Work towards a signature syntax
Implementing the Range syntax resulted in cleaning up how operators in
the core syntax works. There are now two kinds of infix operators
- tight operators (`.` and `..`)
- loose operators
Tight operators may not be interspersed (`$it.left..$it.right` is a
syntax error). Loose operators require whitespace on both sides of the
operator, and can be arbitrarily interspersed. Precedence is left to
right in the core syntax.
Note that delimited syntax (like `( ... )` or `[ ... ]`) is a single
token node in the core syntax. A single token node can be parsed from
beginning to end in a context-free manner.
The rule for `.` is `<token node>.<member>`. The rule for `..` is
`<token node>..<token node>`.
Loose operators all have the same syntactic rule: `<token
node><space><loose op><space><token node>`.
The second aspect of this pull request is the beginning of support for a
signature syntax. Before implementing signatures, a necessary
prerequisite is for the core syntax to support multi-line programs.
That work establishes a few things:
- `;` and newlines are handled in the core grammar, and both count as
"separators"
- line comments begin with `#` and continue until the end of the line
In this commit, multi-token productions in the core grammar can use
separators interchangably with spaces. However, I think we will
ultimately want a different rule preventing separators from occurring
before an infix operator, so that the end of a line is always
unambiguous. This would avoid gratuitous differences between modules and
repl usage.
We already effectively have this rule, because otherwise `x<newline> |
y` would be a single pipeline, but of course that wouldn't work.
Previously, external words accidentally used
ExpansionRule::new().allow_external_command(), when it should have been
ExpansionRule::new().allow_external_word().
External words are the broadest category in the parser, and are the
appropriate category for external arguments. This was just a mistake.
This commit extracts five new crates:
- nu-source, which contains the core source-code handling logic in Nu,
including Text, Span, and also the pretty.rs-based debug logic
- nu-parser, which is the parser and expander logic
- nu-protocol, which is the bulk of the types and basic conveniences
used by plugins
- nu-errors, which contains ShellError, ParseError and error handling
conveniences
- nu-textview, which is the textview plugin extracted into a crate
One of the major consequences of this refactor is that it's no longer
possible to `impl X for Spanned<Y>` outside of the `nu-source` crate, so
a lot of types became more concrete (Value became a concrete type
instead of Spanned<Value>, for example).
This also turned a number of inherent methods in the main nu crate into
plain functions (impl Value {} became a bunch of functions in the
`value` namespace in `crate::data::value`).
This commit extracts Tag, Span, Text, as well as source-related debug
facilities into a new crate called nu_source.
This change is much bigger than one might have expected because the
previous code relied heavily on implementing inherent methods on
`Tagged<T>` and `Spanned<T>`, which is no longer possible.
As a result, this change creates more concrete types instead of using
`Tagged<T>`. One notable example: Tagged<Value> became Value, and Value
became UntaggedValue.
This change clarifies the intent of the code in many places, but it does
make it a big change.
The nu_plugin_match binary wasn't built anymore
after the regex dependency was made non-optional in
https://github.com/nushell/nushell/pull/889, causing
the removal of the regex feature, which nu_plugin_match
depended on.
The original purpose of this PR was to modernize the external parser to
use the new Shape system.
This commit does include some of that change, but a more important
aspect of this change is an improvement to the expansion trace.
Previous commit 6a7c00ea adding trace infrastructure to the syntax coloring
feature. This commit adds tracing to the expander.
The bulk of that work, in addition to the tree builder logic, was an
overhaul of the formatter traits to make them more general purpose, and
more structured.
Some highlights:
- `ToDebug` was split into two traits (`ToDebug` and `DebugFormat`)
because implementations needed to become objects, but a convenience
method on `ToDebug` didn't qualify
- `DebugFormat`'s `fmt_debug` method now takes a `DebugFormatter` rather
than a standard formatter, and `DebugFormatter` has a new (but still
limited) facility for structured formatting.
- Implementations of `ExpandSyntax` need to produce output that
implements `DebugFormat`.
Unlike the highlighter changes, these changes are fairly focused in the
trace output, so these changes aren't behind a flag.
This commit should finish the `coloring_in_tokens` feature, which moves
the shape accumulator into the token stream. This allows rollbacks of
the token stream to also roll back any shapes that were added.
This commit also adds a much nicer syntax highlighter trace, which shows
all of the paths the highlighter took to arrive at a particular coloring
output. This change is fairly substantial, but really improves the
understandability of the flow. I intend to update the normal parser with
a similar tracing view.
In general, this change also fleshes out the concept of "atomic" token
stream operations.
A good next step would be to try to make the parser more
error-correcting, using the coloring infrastructure. A follow-up step
would involve merging the parser and highlighter shapes themselves.
* Moves off of draining between filters. Instead, the sink will pull on the stream, and will drain element-wise. This moves the whole stream to being lazy.
* Adds ctrl-c support and connects it into some of the key points where we pull on the stream. If a ctrl-c is detect, we immediately halt pulling on the stream and return to the prompt.
* Moves away from having a SourceMap where anchor locations are stored. Now AnchorLocation is kept directly in the Tag.
* To make this possible, split tag and span. Span is largely used in the parser and is copyable. Tag is now no longer copyable.
This commit adds the ability to work on features behind a feature flag
that won't be included in normal builds of nu.
These features are not exposed as Cargo features, as they reflect
incomplete features that are not yet stable.
To create a feature, add it to `features.toml`:
```toml
[hintsv1]
description = "Adding hints based on error states in the highlighter"
enabled = false
```
Each feature in `features.toml` becomes a feature flag accessible to `cfg`:
```rs
println!("hintsv1 is enabled");
```
By default, features are enabled based on the value of the `enabled` field.
You can also enable a feature from the command line via the
`NUSHELL_ENABLE_FLAGS` environment variable:
```sh
$ NUSHELL_ENABLE_FLAGS=hintsv1 cargo run
```
You can enable all flags via `NUSHELL_ENABLE_ALL_FLAGS`.
This commit also updates the CI setup to run the build with all flags off and
with all flags on. It also extracts the linting test into its own
parallelizable test, which means it doesn't need to run together with every
other test anymore.
When working on a feature, you should also add tests behind the same flag. A
commit is mergable if all tests pass with and without the flag, allowing
incomplete commits to land on master as long as the incomplete code builds and
passes tests.
This commit replaces the previous naive coloring system with a coloring
system that is more aligned with the parser.
The main benefit of this change is that it allows us to use parsing
rules to decide how to color tokens.
For example, consider the following syntax:
```
$ ps | where cpu > 10
```
Ideally, we could color `cpu` like a column name and not a string,
because `cpu > 10` is a shorthand block syntax that expands to
`{ $it.cpu > 10 }`.
The way that we know that it's a shorthand block is that the `where`
command declares that its first parameter is a `SyntaxShape::Block`,
which allows the shorthand block form.
In order to accomplish this, we need to color the tokens in a way that
corresponds to their expanded semantics, which means that high-fidelity
coloring requires expansion.
This commit adds a `ColorSyntax` trait that corresponds to the
`ExpandExpression` trait. The semantics are fairly similar, with a few
differences.
First `ExpandExpression` consumes N tokens and returns a single
`hir::Expression`. `ColorSyntax` consumes N tokens and writes M
`FlatShape` tokens to the output.
Concretely, for syntax like `[1 2 3]`
- `ExpandExpression` takes a single token node and produces a single
`hir::Expression`
- `ColorSyntax` takes the same token node and emits 7 `FlatShape`s
(open delimiter, int, whitespace, int, whitespace, int, close
delimiter)
Second, `ColorSyntax` is more willing to plow through failures than
`ExpandExpression`.
In particular, consider syntax like
```
$ ps | where cpu >
```
In this case
- `ExpandExpression` will see that the `where` command is expecting a
block, see that it's not a literal block and try to parse it as a
shorthand block. It will successfully find a member followed by an
infix operator, but not a following expression. That means that the
entire pipeline part fails to parse and is a syntax error.
- `ColorSyntax` will also try to parse it as a shorthand block and
ultimately fail, but it will fall back to "backoff coloring mode",
which parsing any unidentified tokens in an unfallible, simple way. In
this case, `cpu` will color as a string and `>` will color as an
operator.
Finally, it's very important that coloring a pipeline infallibly colors
the entire string, doesn't fail, and doesn't get stuck in an infinite
loop.
In order to accomplish this, this PR separates `ColorSyntax`, which is
infallible from `FallibleColorSyntax`, which might fail. This allows the
type system to let us know if our coloring rules bottom out at at an
infallible rule.
It's not perfect: it's still possible for the coloring process to get
stuck or consume tokens non-atomically. I intend to reduce the
opportunity for those problems in a future commit. In the meantime, the
current system catches a number of mistakes (like trying to use a
fallible coloring rule in a loop without thinking about the possibility
that it will never terminate).
The main thrust of this (very large) commit is an overhaul of the
expansion system.
The parsing pipeline is:
- Lightly parse the source file for atoms, basic delimiters and pipeline
structure into a token tree
- Expand the token tree into a HIR (high-level intermediate
representation) based upon the baseline syntax rules for expressions
and the syntactic shape of commands.
Somewhat non-traditionally, nu doesn't have an AST at all. It goes
directly from the token tree, which doesn't represent many important
distinctions (like the difference between `hello` and `5KB`) directly
into a high-level representation that doesn't have a direct
correspondence to the source code.
At a high level, nu commands work like macros, in the sense that the
syntactic shape of the invocation of a command depends on the
definition of a command.
However, commands do not have the ability to perform unrestricted
expansions of the token tree. Instead, they describe their arguments in
terms of syntactic shapes, and the expander expands the token tree into
HIR based upon that definition.
For example, the `where` command says that it takes a block as its first
required argument, and the description of the block syntactic shape
expands the syntax `cpu > 10` into HIR that represents
`{ $it.cpu > 10 }`.
This commit overhauls that system so that the syntactic shapes are
described in terms of a few new traits (`ExpandSyntax` and
`ExpandExpression` are the primary ones) that are more composable than
the previous system.
The first big win of this new system is the addition of the `ColumnPath`
shape, which looks like `cpu."max ghz"` or `package.version`.
Previously, while a variable path could look like `$it.cpu."max ghz"`,
the tail of a variable path could not be easily reused in other
contexts. Now, that tail is its own syntactic shape, and it can be used
as part of a command's signature.
This cleans up commands like `inc`, `add` and `edit` as well as
shorthand blocks, which can now look like `| where cpu."max ghz" > 10`
Kind of touches on #356 by integrating the Starship prompt directly into the shell.
Not finished yet and has surfaced a potential bug in rustyline anyway. It depends on https://github.com/starship/starship/pull/509 being merged so the Starship prompt can be used as a library.
I could have tackled #356 completely and implemented a full custom prompt feature but I felt this was a simpler approach given that Starship is both written in Rust so shelling out isn't necessary and it already has a bunch of useful features built in.
However, I would understand if it would be preferable to just scrap integrating Starship directly and instead implement a custom prompt system which would facilitate simply shelling out to Starship.
This commit migrates Value's numeric types to BigInt and BigDecimal. The
basic idea is that overflow errors aren't great in a shell environment,
and not really necessary.
The main immediate consequence is that new errors can occur when
serializing Nu values to other formats. You can see this in changes to
the various serialization formats (JSON, TOML, etc.). There's a new
`CoerceInto` trait that uses the `ToPrimitive` trait from `num_traits`
to attempt to coerce a `BigNum` or `BigDecimal` into a target type, and
produces a `RangeError` (kind of `ShellError`) if the coercion fails.
Another possible future consequence is that certain performance-critical
numeric operations might be too slow. If that happens, we can introduce
specialized numeric types to help improve the performance of those
situations, based on the real-world experience.
This commit is more substantial than it looks: there was basically no
real support for decimals before, and that impacted values all the way
through.
I also made Size contain a decimal instead of an integer (`1.6kb` is a
reasonable thing to type), which impacted a bunch of code.
The biggest impact of this commit is that it creates many more possible
ways for valid nu types to fail to serialize as toml, json, etc. which
typically can't support the full range of Decimal (or Bigint, which I
also think we should support). This commit makes to-toml fallible, and a
similar effort is necessary for the rest of the serializations.
We also need to figure out how to clearly communicate to users what has
happened, but failing to serialize to toml seems clearly superior to me
than weird errors in basic math operations.
This ended up being a bit of a yak shave. The basic idea in this commit is to
expand `~` in paths, but only in paths.
The way this is accomplished is by doing the expansion inside of the code that
parses literal syntax for `SyntaxType::Path`.
As a quick refresher: every command is entitled to expand its arguments in a
custom way. While this could in theory be used for general-purpose macros,
today the expansion facility is limited to syntactic hints.
For example, the syntax `where cpu > 0` expands under the hood to
`where { $it.cpu > 0 }`. This happens because the first argument to `where`
is defined as a `SyntaxType::Block`, and the parser coerces binary expressions
whose left-hand-side looks like a member into a block when the command is
expecting one.
This is mildly more magical than what most programming languages would do,
but we believe that it makes sense to allow commands to fine-tune the syntax
because of the domain nushell is in (command-line shells).
The syntactic expansions supported by this facility are relatively limited.
For example, we don't allow `$it` to become a bare word, simply because the
command asks for a string in the relevant position. That would quickly
become more confusing than it's worth.
This PR adds a new `SyntaxType` rule: `SyntaxType::Path`. When a command
declares a parameter as a `SyntaxType::Path`, string literals and bare
words passed as an argument to that parameter are processed using the
path expansion rules. Right now, that only means that `~` is expanded into
the home directory, but additional rules are possible in the future.
By restricting this expansion to a syntactic expansion when passed as an
argument to a command expecting a path, we avoid making `~` a generally
reserved character. This will also allow us to give good tab completion
for paths with `~` characters in them when a command is expecting a path.
In order to accomplish the above, this commit changes the parsing functions
to take a `Context` instead of just a `CommandRegistry`. From the perspective
of macro expansion, you can think of the `CommandRegistry` as a dictionary
of in-scope macros, and the `Context` as the compile-time state used in
expansion. This could gain additional functionality over time as we find
more uses for the expansion system.
I'm gonna use a VecDeque now instead of trying to get async streams
working to make progress, but the intent is that we should be able to
use async streams in and out to interleave the work better.
