Awesome
Loxido
A Rust implementation of the Lox programming language from the amazing book Crafting Interpreters by Bob Nystrom.
This is based on the clox reference implementation in C. It implements a byte-code compiler and VM.
You can read a detailed blog post about this here:
https://ceronman.com/2021/07/22/my-experience-crafting-an-interpreter-with-rust/
This implementation passes the 243 integration tests from clox and jlox.
Two branches
This repository contains two main branches:
-
The master branch is written using purely safe code. As a result it's quite slower than the clox reference implementation.
-
The unsafe branch is based on the master code but it introduces some unsafe Rust code. It has a different GC implementation and some hand-written data structures, including a Stack and HashMap. This version is almost as fast as clox (see benchmarks bellow).
Differences between loxido and clox
The safe implementation follows the design of clox, but trying to be Rust idiomatic. Some of the differences are:
-
The design of the GC is quite different to adapt to Rust safety limitations. A vector is used to store objects allocated and another vector of tombstones is used to keep track of deleted objects. Trait based polymorphism is used to implement different tracing strategies for different objects.
-
The GC doesn't run while compiling code, only when the VM is running. This simplifies dealing with the borrow checker.
-
The GC is not able to precisely keep track of all the allocated bytes; only an approximation. For this reason, the GC behavior might be less predictable.
-
For the most part, data structures from the Rust standard library are used. This includes vectors, hashmaps and strings.
-
The VM uses fixed length instructions, as opposed to the variable length from clox. This allows using enums for instructions, which makes the code nicer to write.
-
This implementation avoids NaN Boxing completely.
Changes in the unsafe branch
The unsafe branch introduces some unsafe code to improve performance. The changes are:
-
Using raw pointers to keep track of the current frame and bytecode chunk on every cycle of the VM.
-
Redesigned GC using raw pointers, very similar to the one used in clox. Although objects are represented as an enum instead of trait objects to avoid dynamic dispatch.
-
Stack is implemented using an array and a raw pointer indicating the top. Similar to clox.
-
The Program Counter (PC) is implemented using a raw pointer.
-
Custom HashMap implementation based on the code from clox.
How to run
To compile and run this project you need Cargo.
To run the REPL, just go for:
cargo run
To run a specific file:
cargo run program.lox
Running the tests
To run the tests just run:
cargo test
Performance
I measured the running times of the different benchmarks available in the lox repository. I also ported the benchmarks to Python and Perl. Here are the compared running times of different implementations of Lox, including mind.
These were run on my Thinkpad X1 laptop with Intel Core i7-8565U CPU
note: The horizontal axis is capped at 15 seconds to improve visibility.
clox | jlox | loxido | loxido unsafe | Python 3.9.2 | Perl 5.32.1 | |
---|---|---|---|---|---|---|
arithmetic | 0.6304s | 3.3200s | 5.4746s | 0.9706s | 1.6815s | 1.0798s |
binary_tree | 2.6283s | 7.0670s | 8.4780s | 3.4534s | 4.1730s | 6.3510s |
equality | 1.3740s | 3.6209s | 11.337s | 1.9433s | 3.4657s | 2.2632s |
fib | 0.9067s | 4.2789s | 5.0318s | 1.2042s | 2.1110s | 4.2134s |
instantiation | 0.5264s | 1.5280s | 7.2826s | 0.9068s | 1.4895s | 3.8319s |
invocation | 0.2916s | 1.3600s | 1.6865s | 0.3566s | 0.7789s | 0.9535s |
method_call | 0.1574s | 1.9440s | 1.1231s | 0.2255s | 0.5696s | 0.9560s |
properties | 0.3716s | 3.9539s | 2.2317s | 0.5132s | 1.4074s | 1.4339s |
trees | 1.9714s | 21.677s | 14.478s | 2.9840s | 6.4791s | 13.019s |
zoo | 0.3132s | 2.8380s | 1.6914s | 0.3836s | 0.9739s | 1.1222s |
Tree walker implementation
I also wrote another implementation of the tree walk interpreter in Clojure called cloxure.
Copyright note
Most of the test and benchmark files are copied from Crafting Interpreters repository. The copyright for these belongs to Bob Nystrom and are copied here because their license allows it (MIT).
The source code for this implementation is copyright of Manuel Cerón. And it's also licensed as MIT (see LICENCE for details.)