Feature #16163


Reduce the output of `RubyVM::InstructionSequence#to_binary`

Added by NagayamaRyoga (Nagayama Ryoga) about 2 years ago. Updated about 2 years ago.

Target version:



The output of RubyVM::InstructionSequence#to_binary is extremely large.
We have reduced the output of #to_binary by more than 70%.

The execution speed of RubyVM::InstructionSequence.load_from_binary is about 7% slower, but when reading a binary from a file, it may be faster than the master.

Since Bootsnap gem uses #to_binary, this proposal reduces the compilation cache size of Rails projects to about 1/4.


#to_binary and .load_from_binary are used by Bootsnap gem
that is installed by default in Rails projects since Rails 5.2.
Improving #to_binary output also reduces the compilation cache generated by it.



  1. Prevented unnecessary structure fields from being output.

    i.e. MJIT information in struct rb_iseq_constant_body.

  2. Output integer value in variable length format such as UTF-8.

    * Small uint serialization
    * 0x00000000_00000000 - 0x00000000_0000007f: 1byte | XXXX XXX1 |
    * 0x00000000_00000080 - 0x00000000_00003fff: 2byte | XXXX XX10 | XXXX XXXX |
    * 0x00000000_00004000 - 0x00000000_001fffff: 3byte | XXXX X100 | XXXX XXXX | XXXX XXXX |
    * 0x00000000_00020000 - 0x00000000_0fffffff: 4byte | XXXX 1000 | XXXX XXXX | XXXX XXXX | XXXX XXXX |
    * ...
    * 0x00010000_00000000 - 0x00ffffff_ffffffff: 8byte | 1000 0000 | XXXX XXXX | XXXX XXXX | XXXX XXXX | XXXX XXXX | XXXX XXXX | XXXX XXXX | XXXX XXXX |
    * 0x01000000_00000000 - 0xffffffff_ffffffff: 9byte | 0000 0000 | XXXX XXXX | XXXX XXXX | XXXX XXXX | XXXX XXXX | XXXX XXXX | XXXX XXXX | XXXX XXXX | XXXX XXXX |
  3. We integrated ID output mechanism and object serialization.



OS: Ubuntu 16.04 LTS
CPU: Intel(R) Core(TM) i7-6700 CPU @ 3.40GHz
Memory: 32GB

Simple benchmark

First, We combined the files in the benchmark/ and generated a huge .rb file with 5400 lines.
And We measured the output size of #to_binary and the time taken to load it.

The benchmark code:

  • master (ruby 2.7.0dev (2019-08-17T11:20:04Z master 2a65498ca2) [x86_64-linux])

    size: 1963764B
                                  user    system      total      real
    load_from_binary               4.276000  0.000000   4.276000 (  4.277652) + load_from_binary   5.060000  0.536000   5.596000 (  5.593620)
  • This proposal

    size: 463776B
                                  user    system      total      real
    load_from_binary               4.576000  0.004000   4.580000 (  4.580691) + load_from_binary   4.856000  0.080000   4.936000 (  4.934168)

The output size of #to_binary is about 24% (4 times smaller!) of the output of master's.

.load_from_binary is about 7% slower.
However, loading the binary from a file and decoding it ( + load_from_binary), it is about 12% faster than master.

A Rails project with Bootsnap

Next, We measured the startup time of the simple Rails project generated with $ rails new.
Bootsnap caches the compilation results at the first boot and uses them to load the application from the next time.


  • master

    • Cache (tmp/): 32MB
    • The first boot: Average 1.700s (N=10)
    • Boot from cache: Average 0.588s (N=10)
  • proposal

    • Cache (tmp/): 9.4MB
    • The first boot: Average 1.684s (N=10)
    • Boot from cache: Average 0.592s (N=10)

The cache size is now about 30%.
There was no impact on project startup time.


Passed make test-all with RUBY_ISEQ_DUMP_DEBUG='to_binary'.

$ make test-all -j8 RUBY_ISEQ_DUMP_DEBUG=to_binary
../../ruby-dev/revision.h unchanged
Run options: "--ruby=./miniruby -I../../ruby-dev/lib -I. -I.ext/common  ../../ruby-dev/tool/runruby.rb --extout=.ext  -- --disable-gems" --excludes-dir=../../ruby-dev/test/excludes --name=!/memory_leak/

# Running tests:

Finished tests in 46.252333s, 452.6258 tests/s, 57576.1656 assertions/s.
20935 tests, 2663032 assertions, 0 failures, 0 errors, 92 skips

ruby -v: ruby 2.7.0dev (2019-09-05T09:20:11Z alt-bytecode/load_.. 8aa0a1cc4c) [x86_64-linux]


The output size of RubyVM::InstructionSequence#to_binary is about 1/4 of the master.
The impact on speed is negligible.
Passed all tests.


Also available in: Atom PDF