d7392f1f3b
# Description This PR adds an internal representation language to Nushell, offering an alternative evaluator based on simple instructions, stream-containing registers, and indexed control flow. The number of registers required is determined statically at compile-time, and the fixed size required is allocated upon entering the block. Each instruction is associated with a span, which makes going backwards from IR instructions to source code very easy. Motivations for IR: 1. **Performance.** By simplifying the evaluation path and making it more cache-friendly and branch predictor-friendly, code that does a lot of computation in Nushell itself can be sped up a decent bit. Because the IR is fairly easy to reason about, we can also implement optimization passes in the future to eliminate and simplify code. 2. **Correctness.** The instructions mostly have very simple and easily-specified behavior, so hopefully engine changes are a little bit easier to reason about, and they can be specified in a more formal way at some point. I have made an effort to document each of the instructions in the docs for the enum itself in a reasonably specific way. Some of the errors that would have happened during evaluation before are now moved to the compilation step instead, because they don't make sense to check during evaluation. 3. **As an intermediate target.** This is a good step for us to bring the [`new-nu-parser`](https://github.com/nushell/new-nu-parser) in at some point, as code generated from new AST can be directly compared to code generated from old AST. If the IR code is functionally equivalent, it will behave the exact same way. 4. **Debugging.** With a little bit more work, we can probably give control over advancing the virtual machine that `IrBlock`s run on to some sort of external driver, making things like breakpoints and single stepping possible. Tools like `view ir` and [`explore ir`](https://github.com/devyn/nu_plugin_explore_ir) make it easier than before to see what exactly is going on with your Nushell code. The goal is to eventually replace the AST evaluator entirely, once we're sure it's working just as well. You can help dogfood this by running Nushell with `$env.NU_USE_IR` set to some value. The environment variable is checked when Nushell starts, so config runs with IR, or it can also be set on a line at the REPL to change it dynamically. It is also checked when running `do` in case within a script you want to just run a specific piece of code with or without IR. # Example ```nushell view ir { |data| mut sum = 0 for n in $data { $sum += $n } $sum } ``` ```gas # 3 registers, 19 instructions, 0 bytes of data 0: load-literal %0, int(0) 1: store-variable var 904, %0 # let 2: drain %0 3: drop %0 4: load-variable %1, var 903 5: iterate %0, %1, end 15 # for, label(1), from(14:) 6: store-variable var 905, %0 7: load-variable %0, var 904 8: load-variable %2, var 905 9: binary-op %0, Math(Plus), %2 10: span %0 11: store-variable var 904, %0 12: load-literal %0, nothing 13: drain %0 14: jump 5 15: drop %0 # label(0), from(5:) 16: drain %0 17: load-variable %0, var 904 18: return %0 ``` # Benchmarks All benchmarks run on a base model Mac Mini M1. ## Iterative Fibonacci sequence This is about as best case as possible, making use of the much faster control flow. Most code will not experience a speed improvement nearly this large. ```nushell def fib [n: int] { mut a = 0 mut b = 1 for _ in 2..=$n { let c = $a + $b $a = $b $b = $c } $b } use std bench bench { 0..50 | each { |n| fib $n } } ``` IR disabled: ``` ╭───────┬─────────────────╮ │ mean │ 1ms 924µs 665ns │ │ min │ 1ms 700µs 83ns │ │ max │ 3ms 450µs 125ns │ │ std │ 395µs 759ns │ │ times │ [list 50 items] │ ╰───────┴─────────────────╯ ``` IR enabled: ``` ╭───────┬─────────────────╮ │ mean │ 452µs 820ns │ │ min │ 427µs 417ns │ │ max │ 540µs 167ns │ │ std │ 17µs 158ns │ │ times │ [list 50 items] │ ╰───────┴─────────────────╯ ```  ## [gradient_benchmark_no_check.nu](https://github.com/nushell/nu_scripts/blob/main/benchmarks/gradient_benchmark_no_check.nu) IR disabled: ``` ╭───┬──────────────────╮ │ 0 │ 27ms 929µs 958ns │ │ 1 │ 21ms 153µs 459ns │ │ 2 │ 18ms 639µs 666ns │ │ 3 │ 19ms 554µs 583ns │ │ 4 │ 13ms 383µs 375ns │ │ 5 │ 11ms 328µs 208ns │ │ 6 │ 5ms 659µs 542ns │ ╰───┴──────────────────╯ ``` IR enabled: ``` ╭───┬──────────────────╮ │ 0 │ 22ms 662µs │ │ 1 │ 17ms 221µs 792ns │ │ 2 │ 14ms 786µs 708ns │ │ 3 │ 13ms 876µs 834ns │ │ 4 │ 13ms 52µs 875ns │ │ 5 │ 11ms 269µs 666ns │ │ 6 │ 6ms 942µs 500ns │ ╰───┴──────────────────╯ ``` ## [random-bytes.nu](https://github.com/nushell/nu_scripts/blob/main/benchmarks/random-bytes.nu) I got pretty random results out of this benchmark so I decided not to include it. Not clear why. # User-Facing Changes - IR compilation errors may appear even if the user isn't evaluating with IR. - IR evaluation can be enabled by setting the `NU_USE_IR` environment variable to any value. - New command `view ir` pretty-prints the IR for a block, and `view ir --json` can be piped into an external tool like [`explore ir`](https://github.com/devyn/nu_plugin_explore_ir). # Tests + Formatting All tests are passing with `NU_USE_IR=1`, and I've added some more eval tests to compare the results for some very core operations. I will probably want to add some more so we don't have to always check `NU_USE_IR=1 toolkit test --workspace` on a regular basis. # After Submitting - [ ] release notes - [ ] further documentation of instructions? - [ ] post-release: publish `nu_plugin_explore_ir` |
||
|---|---|---|
| .. | ||
| fuzz | ||
| src | ||
| tests | ||
| Cargo.toml | ||
| LICENSE | ||
| README.md | ||
nu-parser, the Nushell parser
Nushell's parser is a type-directed parser, meaning that the parser will use type information available during parse time to configure the parser. This allows it to handle a broader range of techniques to handle the arguments of a command.
Nushell's base language is whitespace-separated tokens with the command (Nushell's term for a function) name in the head position:
head1 arg1 arg2 | head2
Lexing
The first job of the parser is to a lexical analysis to find where the tokens start and end in the input. This turns the above into:
<item: "head1">, <item: "arg1">, <item: "arg2">, <pipe>, <item: "head2">
At this point, the parser has little to no understanding of the shape of the command or how to parse its arguments.
Lite parsing
As Nushell is a language of pipelines, pipes form a key role in both separating commands from each other as well as denoting the flow of information between commands. The lite parse phase, as the name suggests, helps to group the lexed tokens into units.
The above tokens are converted the following during the lite parse phase:
Pipeline:
Command #1:
<item: "head1">, <item: "arg1">, <item: "arg2">
Command #2:
<item: "head2">
Parsing
The real magic begins to happen when the parse moves on to the parsing stage. At this point, it traverses the lite parse tree and for each command makes a decision:
- If the command looks like an internal/external command literal: e.g.
fooor/usr/bin/ls, it parses it as an internal or external command - Otherwise, it parses the command as part of a mathematical expression
Types/shapes
Each command has a shape assigned to each of the arguments it reads in. These shapes help define how the parser will handle the parse.
For example, if the command is written as:
where $x > 10
When the parsing happens, the parser will look up the where command and find its Signature. The Signature states what flags are allowed and what positional arguments are allowed (both required and optional). Each argument comes with a Shape that defines how to parse values to get that position.
In the above example, if the Signature of where said that it took three String values, the result would be:
CallInfo:
Name: `where`
Args:
Expression($x), a String
Expression(>), a String
Expression(10), a String
Or, the Signature could state that it takes in three positional arguments: a Variable, an Operator, and a Number, which would give:
CallInfo:
Name: `where`
Args:
Expression($x), a Variable
Expression(>), an Operator
Expression(10), a Number
Note that in this case, each would be checked at compile time to confirm that the expression has the shape requested. For example, "foo" would fail to parse as a Number.
Finally, some Shapes can consume more than one token. In the above, if the where command stated it took in a single required argument, and that the Shape of this argument was a MathExpression, then the parser would treat the remaining tokens as part of the math expression.
CallInfo:
Name: `where`
Args:
MathExpression:
Op: >
LHS: Expression($x)
RHS: Expression(10)
When the command runs, it will now be able to evaluate the whole math expression as a single step rather than doing any additional parsing to understand the relationship between the parameters.
Making space
As some Shapes can consume multiple tokens, it's important that the parser allow for multiple Shapes to coexist as peacefully as possible.
The simplest way it does this is to ensure there is at least one token for each required parameter. If the Signature of the command says that it takes a MathExpression and a Number as two required arguments, then the parser will stop the math parser one token short. This allows the second Shape to consume the final token.
Another way that the parser makes space is to look for Keyword shapes in the Signature. A Keyword is a word that's special to this command. For example in the if command, else is a keyword. When it is found in the arguments, the parser will use it as a signpost for where to make space for each Shape. The tokens leading up to the else will then feed into the parts of the Signature before the else, and the tokens following are consumed by the else and the Shapes that follow.