ruby--ruby/doc/yjit/yjit.md

YJIT - Yet Another Ruby JIT

DISCLAIMER: Please note that this project is experimental. It is very much a work in progress, it may cause your software to crash, and current performance results will vary widely, especially on larger applications.

YJIT is a lightweight, minimalistic Ruby JIT built inside CRuby. It lazily compiles code using a Basic Block Versioning (BBV) architecture. The target use case is that of servers running Ruby on Rails, an area where MJIT has not yet managed to deliver speedups. To simplify development, we currently support only macOS and Linux on x86-64, but an ARM64 backend is part of future plans. This project is open source and falls under the same license as CRuby.

If you wish to learn more about the approach taken, here are some conference talks and publications:

• RubyKaigi 2021 talk: YJIT: Building a New JIT Compiler Inside CRuby
• Blog post: YJIT: Building a New JIT Compiler Inside CRuby
• MoreVMs 2021 talk: YJIT: Building a New JIT Compiler Inside CRuby
• ECOOP 2016 talk: Interprocedural Type Specialization of JavaScript Programs Without Type Analysis
• ECOOP 2016 paper: Interprocedural Type Specialization of JavaScript Programs Without Type Analysis
• ECOOP 2015 talk: Simple and Effective Type Check Removal through Lazy Basic Block Versioning
• ECOOP 2015 paper: Simple and Effective Type Check Removal through Lazy Basic Block Versioning

To cite this repository in your publications, please use this bibtex snippet:

@misc{yjit_ruby_jit,
  author = {Chevalier-Boisvert, Maxime and Wu, Alan and Patterson, Aaron},
  title = {YJIT - Yet Another Ruby JIT},
  year = {2021},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/Shopify/yjit}},
}

Current Limitations

YJIT is a work in progress and as such may not yet be mature enough for mission-critical software. Below is a list of known limitations, all of which we plan to eventually address:

• No garbage collection for generated code.
• Currently supports only macOS and Linux.
• Currently supports only x86-64 CPUs.

Because there is no GC for generated code yet, your software could run out of executable memory if it is large enough. You can change how much executable memory is allocated using YJIT's command-line options.

Installation

Requirements

You will need to install:

• A C compiler such as GCC or Clang
• GNU Make and Autoconf
• The Rust compiler rustc and Cargo (if you want to build in dev/debug mode)

To install the Rust build toolchain, we suggest following the recommended installation method. Rust also provides first class support for many source code editors.

Building YJIT

Start by cloning the ruby/ruby repository:

git clone https://github.com/ruby/ruby yjit
cd yjit

The YJIT ruby binary can be built with either GCC or Clang. It can be built either in dev (debug) mode or in release mode. For maximum performance, compile YJIT in release mode with GCC. More detailed build instructions are provided in the Ruby README.

# Configure in release mode for maximum performance, build and install
./autogen.sh
./configure --enable-yjit --prefix=$HOME/.rubies/ruby-yjit --disable-install-doc --disable--install-rdoc
make -j install

or

# Configure in dev (debug) mode for development, build and install
./autogen.sh
./configure --enable-yjit=dev --prefix=$HOME/.rubies/ruby-yjit --disable-install-doc --disable--install-rdoc
make -j install

On macOS, you may need to specify where to find some libraries:

# Install dependencies
brew install openssl readline libyaml

# Configure in dev (debug) mode for development, build and install
./autogen.sh
./configure --enable-yjit=dev --prefix=$HOME/.rubies/ruby-yjit --disable-install-doc --disable--install-rdoc --with-opt-dir="$(brew --prefix openssl):$(brew --prefix readline):$(brew --prefix libyaml)"
make -j install

Typically configure will choose the default C compiler. To specify the C compiler, use

# Choosing a specific c compiler
export CC=/path/to/my/chosen/c/compiler

before running ./configure.

You can test that YJIT works correctly by running:

# Quick tests found in /bootstraptest
make btest

# Complete set of tests
make -j test-all

Usage

Examples

Once YJIT is built, you can either use ./miniruby from within your build directory, or switch to the YJIT version of ruby by using the chruby tool:

chruby ruby-yjit
ruby myscript.rb

You can dump statistics about compilation and execution by running YJIT with the --yjit-stats command-line option:

./miniruby --yjit-stats myscript.rb

The machine code generated for a given method can be printed by adding puts RubyVM::YJIT.disasm(method(:method_name)) to a Ruby script. Note that no code will be generated if the method is not compiled.

Command-Line Options

YJIT supports all command-line options supported by upstream CRuby, but also adds a few YJIT-specific options:

• --yjit: enable YJIT (disabled by default)
• --yjit-call-threshold=N: number of calls after which YJIT begins to compile a function (default 2)
• --yjit-exec-mem-size=N: size of the executable memory block to allocate, in MiB (default 256 MiB)
• --yjit-stats: produce statistics after the execution of a program (must compile with cppflags=-DRUBY_DEBUG to use this)
• --yjit-max-versions=N: maximum number of versions to generate per basic block (default 4)
• --yjit-greedy-versioning: greedy versioning mode (disabled by default, may increase code size)

Benchmarking

We have collected a set of benchmarks and implemented a simple benchmarking harness in the yjit-bench repository. This benchmarking harness is designed to disable CPU frequency scaling, set process affinity and disable address space randomization so that the variance between benchmarking runs will be as small as possible. Please kindly note that we are at an early stage in this project.

Performance Tips

This section contains tips on writing Ruby code that will run as fast as possible on YJIT. Some of this advice is based on current limitations of YJIT, while other advice is broadly applicable. It probably won't be practical to apply these tips everywhere in your codebase, but you can profile your code using a tool such as stackprof and refactor the specific methods that make up the largest fractions of the execution time.

• Use exceptions for error recovery only, not as part of normal control-flow
• Avoid redefining basic integer operations (i.e. +, -, <, >, etc.)
• Avoid redefining the meaning of nil, equality, etc.
• Avoid allocating objects in the hot parts of your code
• Use while loops if you can, instead of integer.times
• Minimize layers of indirection
  • Avoid classes that wrap objects if you can
  • Avoid methods that just call another method, trivial one liner methods
• CRuby method calls are costly. Favor larger methods over smaller methods.
• Try to write code so that the same variables always have the same type

You can also compile YJIT in debug mode and use the --yjit-stats command-line option to see which bytecodes cause YJIT to exit, and refactor your code to avoid using these instructions in the hottest methods of your code.

Memory Statistics

YJIT, including in production configuration, keeps track of the size of generated code. If you check YJIT.runtime_stats you can see them:

$ RUBYOPT="--yjit" irb
irb(main):001:0> RubyVM::YJIT.runtime_stats
=> {:inline_code_size=>331945, :outlined_code_size=>272980}

These are the size in bytes of generated inlined code and generated outlined code. If the combined sizes for generated code are very close to the total YJIT exec-mem-size (see above), YJIT will stop generating code once the limit is reached. Try to make sure you have enough exec-mem-size for the program you're running. By default YJIT will allocate 268,435,456 bytes (256 MiB) of space for generated inlined and outlined code.

Other Statistics

If you compile Ruby with RUBY_DEBUG and/or YJIT_STATS defined and run with "--yjit --yjit-stats", YJIT will track and return performance statistics in RubyVM::YJIT.runtime_stats.

$ RUBYOPT="--yjit --yjit-stats" irb
irb(main):001:0> YJIT.runtime_stats
=>
{:inline_code_size=>340745,
 :outlined_code_size=>297664,
 :all_stats=>true,
 :exec_instruction=>1547816,
 :send_callsite_not_simple=>7267,
 :send_kw_splat=>7,
 :send_ivar_set_method=>72,
...

Some of the counters include:

:exec_instruction - how many Ruby bytecode instructions have been executed :binding_allocations - number of bindings allocated :binding_set - number of variables set via a binding :vm_insns_count - number of instructions executed by the Ruby interpreter :compiled_iseq_count - number of bytecode sequences compiled

Counters starting with "exit_" show reasons for YJIT code taking a side exit (return to the interpreter.) See yjit_hacking.md for more details.

Performance counter names are not guaranteed to remain the same between Ruby versions. If you're curious what one does, it's usually best to search the source code for it — but it may change in a later Ruby version.

Contributing

We welcome open source contributors. You should feel free to open new issues to report bugs or just to ask questions. Suggestions on how to make this readme file more helpful for new contributors are most welcome.

Bug fixes and bug reports are very valuable to us. If you find a bug in YJIT, it's very possible be that nobody has reported it before, or that we don't have a good reproduction for it, so please open an issue and provide as much information as you can about your configuration and a description of how you encountered the problem. List the commands you used to run YJIT so that we can easily reproduce the issue on our end and investigate it. If you are able to produce a small program reproducing the error to help us track it down, that is very much appreciated as well.

If you would like to contribute a large patch to YJIT, we suggest opening an issue or a discussion on this repository so that we can have an active discussion. A common problem is that sometimes people submit large pull requests to open source projects without prior communication, and we have to reject them because the work they implemented does not fit within the design of the project. We want to save you time and frustration, so please reach out and we can have a productive discussion as to how you can contribute things we will want to merge into YJIT.

Source Code Organization

The YJIT source code is divided between:

• yjit.c: code YJIT uses to interface with the rest of CRuby
• yjit.h: C definitions YJIT exposes to the rest of the CRuby
• yjit.rb: YJIT Ruby module that is exposed to Ruby
• yjit/src/asm/*: in-memory assembler we use to generate machine code
• yjit/src/codegen.rs: logic for translating Ruby bytecode to machine code
• yjit/src/core.rb: basic block versioning logic, core structure of YJIT
• yjit/src/stats.rs: gathering of run-time statistics
• yjit/src/options.rs: handling of command-line options
• yjit/bindgen/src/main.rs: C bindings exposed to the Rust codebase through bindgen
• yjit/src/cruby.rs: C bindings manually exposed to the Rust codebase
• misc/test_yjit_asm.sh: script to compile and run the in-memory assembler tests
• misc/yjit_asm_tests.c: tests for the in-memory assembler

The core of CRuby's interpreter logic is found in:

• insns.def: defines Ruby's bytecode instructions (gets compiled into vm.inc)
• vm_insnshelper.c: logic used by Ruby's bytecode instructions
• vm_exec.c: Ruby interpreter loop

Generating C bindings with bindgen

In order to expose C functions to the Rust codebase, you will need to generate C bindings:

CC=clang ./configure --enable-yjit=dev
make -j yjit-bindgen

This uses the bindgen tools to generate/update yjit/src/cruby_bindings.inc.rs based on the bindings listed in yjit/bindgen/src/main.rs. Avoid manually editing this file as it could be automatically regenerated at a later time. If you need to manually add C bindings, add them to yjit/cruby.rs instead.

Coding & Debugging Protips

There are 3 test suites:

• make btest (see /bootstraptest)
• make test-all
• make test-spec
• make check runs all of the above

The tests can be run in parallel like this:

make -j test-all RUN_OPTS="--yjit-call-threshold=1"

Or single-threaded like this, to more easily identify which specific test is failing:

make test-all TESTOPTS=--verbose RUN_OPTS="--yjit-call-threshold=1"

To debug a single test in test-all:

make test-all TESTS='test/-ext-/marshal/test_usrmarshal.rb' RUNRUBYOPT=--debugger=lldb RUN_OPTS="--yjit-call-threshold=1"

You can also run one specific test in btest:

make btest BTESTS=bootstraptest/test_ractor.rb RUN_OPTS="--yjit-call-threshold=1"

There are shortcuts to run/debug your own test/repro in test.rb:

make run  # runs ./miniruby test.rb
make lldb # launches ./miniruby test.rb in lldb

You can use the Intel syntax for disassembly in LLDB, keeping it consistent with YJIT's disassembly:

echo "settings set target.x86-disassembly-flavor intel" >> ~/.lldbinit

Running YJIT on M1

It is possible to run YJIT on an Apple M1 via Rosetta. You can find basic instructions below, but there are a few caveats listed further down.

First, install Rosetta:

$ softwareupdate --install-rosetta

Now any command can be run with Rosetta via the arch command line tool.

Then you can start your shell in an x86 environment:

$ arch -x86_64 zsh

You can double check your current architecture via the arch command:

$ arch -x86_64 zsh
$ arch
i386

You may need to set the default target for rustc to x86-64, e.g.

$ rustup default stable-x86_64-apple-darwin

While in your i386 shell, install Cargo and Homebrew, then hack away!

M1 Caveats

1. You must install a version of Homebrew for each architecture
2. Cargo will install in $HOME/.cargo by default, and I don't know a good way to change architectures after install
3. dev won't work if you have i386 Homebrew installed on an M1

If you use Fish shell you can read this link for information on making the dev environment easier.