An experiment in YARV bytecode injection.
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.

This repository contains code used while experimenting with YARV bytecode precompilation under the MRI ruby interpreter. For larger rails projects, a considerable portion of load time during development is spent preloading gems and other dependencies that almost never change unless bundler's Gemfile changes. Even though the MRI interpreter can read and write YARV instruction sequences, it does not expose all interfaces publicly.

This is a work in [un]progress. This code is known to NOT WORK with rbenv, but the reason is currently unknown. It should work with rvm, but assumes you do not prefix your commands with bundle ... (for now). It is not currently being maintained due to reasons outlined in the "Practice" section of this README.


In theory, it should be possible to create a library that wraps a series of require statements in an application (or bundler itself) and builds one large precompiled blob of bytecode that can later be injected on subsequent invocations. When that same block of code is reached, instead of executing the contents of that block, a blob of bytecode is injected into the YARV VM and the $LOADED_FEATURES list is updated. The original source files would never be parsed and the MRI interpreter would have to re-navigate a possibly gigantic dependency graph.

#!/usr/bin/env ruby

require 'injectus'

Injectus.capture do
  require 'bundler'

# application code that may or may not change...


In practice, there are multiple tools that try to inject themselves into the code loading process, including both bundler and rails. To complicate thing further, ruby has a concept of autoload (lazy evaluation of dependencies) and rails 3 also implements their own variation of autoload via ActiveSupport::Dependencies.


With the MRI interpreter (at least the 1.9.x series), require statements defined in ruby code can be reliably captured, including the order in which source files are interpreted. Due to legacy issues, ruby code loaded through autoload does not call back out to the ruby's require, contrary to expectation and has not been fixed in 1.9.x or early 2.0.x releases:

Because we can not reliably get a list of all files required in their original order, it is not possible to capture and reload instructions as one giant blob; instead it requires us to manage code loading one source file at a time, and to keep track of whether any other source files have been loaded between invocations of require.

One positive is that autoloading in both ruby and rails appear to be on their way out due to issues with thread-safety, see:

Bytecode Injection

MRI provides an interface for accessing instruction sequences via RubyVM::InstructionSequence. It does not provide a public interface for loading those sequences back into the VM from within ruby code. This can be worked around by exposing rb_iseq_load ourselves using ruby's DL module (

Unfortunately, unmarshalling bytecode into ruby objects before passing it back to the YARV VM causes a major performance hit due to garbage collection and boxing of data structures (some speculation here). The end result is only a marginal performance increase that is further eroded in newer versions of MRI (1.9.3) that make additional improvements in code loading (

Future Viability

This could be a viable project if it was possible to:

  • reliably track the load order of all source files compiled into instruction sequences
  • load a bytecode sequence directly into the VM without first unmarshalling into ruby objects


Primary Resources:

Other links that may be useful: