Evaluating Ruby in Ruby

Jan 26, 2020

TL;DR #

This article is about instruction sequences and evaluating them using pure Ruby.

The repository is available here .

Is it a Ruby implementation?

No. It’s just a runner of instructions. It is similar to MRI’s virtual machine, but it lacks many features and it’s 100 times slower.

Can I use it in my applications?

Of course, no. Well, if you want.

Does it work at all?

Yes, and it even passes most language specs from RubySpec test suite.

How Ruby evaluates your code. #

Well, I think I should start with explaining basics. There is a plenty of articles about it, so I’ll be short:

  1. First, Ruby parses your code into AST (parse.y)
  2. Then, it compiles it into instruction sequence (compile.c)
  3. And every time when you let’s say call a method it evaluates these instructions. (insn.def, vm_eval.c, vm_insnhelper.c)

Long time ago there was no YARV and Ruby used to evaluate AST.

1+2 is just a small syntax tree with a send node in the root and two children: int(1) and int(2). To evaluate it you need to “traverse” it by walking down recursively. Primitive nodes are getting substituted with values (integers 1 and 2 respectively), send node calls + on pre-reduced children and returns 3.

On the other side, YARV is a stack-based Virtual Machine (VM) that evaluates an assembly-like language that consist of more than 100 predefined instructions that store their input/output values on the VM’s stack.

These instructions have arguments, some of them are primitive values, some are inlined Ruby objects, some are flags.

You can view these instructions by running Ruby with --dump=insns flag:

$ ruby --dump=insns -e 'puts 2 + 3'
== disasm: #<ISeq:<main>@-e:1 (1,0)-(1,10)> (catch: FALSE)
0000 putself                                                          (   1)[Li]
0001 putobject                    2
0003 putobject                    3
0005 opt_plus                     <callinfo!mid:+, argc:1, ARGS_SIMPLE>, <callcache>
0008 opt_send_without_block       <callinfo!mid:puts, argc:1, FCALL|ARGS_SIMPLE>, <callcache>
0011 leave

As you can see, there are 5 instructions:

  1. putself - pushes self at the top of the stack
  2. putobject - pushes given object (numbers 2 an 3)
  3. opt_plus - optimized instruction for + method
  4. opt_send_without_block - optimized instruction for a generic method call without block
  5. leave - an equivalent of return

The structure of the instruction sequence (ISeq) #

Let’s start with the example above.

MRI has an API to compile an arbitrary code into instructions:

> pp RubyVM::InstructionSequence.compile("2+3").to_a
["YARVInstructionSequence/SimpleDataFormat",
 2,
 6,
 1,
 {:arg_size=>0,
  :local_size=>0,
  :stack_max=>2,
  :node_id=>4,
  :code_location=>[1, 0, 1, 3]},
 "<compiled>",
 "<compiled>",
 "<compiled>",
 1,
 :top,
 [],
 {},
 [],
 [1,
  :RUBY_EVENT_LINE,
  [:putobject, 2],
  [:putobject, 3],
  [:opt_plus, {:mid=>:+, :flag=>16, :orig_argc=>1}, false],
  [:leave]]]

Our code gets compiled into an array where:

Each instruction is either:

Only arrays are “real” instructions, numbers and symbols are special debug entries that are used internally by the VM. In our case a number followed by the :RUBY_EVENT_LINE is a mark that MRI uses to know what is the number of the line that is currently being interpreted (for example, all backtrace entries include these numbers)

Building a PoC #

How can we evaluate instructions above? Well, we definitely need a stack:

$stack = []

Then, we need to iterate over instructions (the last item of the iseq) and evaluate them one by one. We could write a giant case-when, but I promise that it won’t fit on 10 screens. Let’s use some meta-programming and dynamic dispatching:

def execute_insn(insn)
  name, *payload = insn
  send(:"execute_#{name}", payload)
end

This way we need to write a method per instruction type, let’s start with putobject:

def execute_putobject(object)
  $stack.push(object)
end

def execute_opt_plus(options, flag)
  rhs = $stack.pop
  lhs = $stack.pop
  $stack.push(lhs + rhs)
end

def execute_leave
  # nothing here so far
end

This code is enough to get 5 in the stack once it’s executed. Here’s the runner:

iseq = RubyVM::InstructionSequence.compile("2+3").to_a
insns = iseq[13].grep(Array) # ignore numbers and symbols for now
insns.each { |insn| execute_insn(insn) }
pp $stack

You should see [5].

All instructions above simply pull some values from the stack, do some computations and push the result back.

Self #

Let’s think about self for a minute:

> pp RubyVM::InstructionSequence.compile("self").to_a[13]
[1, :RUBY_EVENT_LINE, [:putself], [:leave]]

self is stored somewhere internally, and even more, it’s dynamic:

puts self
class X
  puts self
end

The first puts self prints main, while the second one prints X.

Here comes the concept of frames.

Frames #

Frame is an object inside the virtual machine that represents a closure. Or a binding. It’s an isolated “world” with its own set of locals, its own self, its own file/line information, it’s own rescue and ensure handlers etc.

Frame also has a type:

  1. it can be a top frame that wraps all the code in your file. All variables set in the global scope of your file belong to the top frame. Each parsed and evaluated file creates its own top frame.
  2. it can be a method frame. All methods create it. One method frame per one method call.
  3. it can be a block frame. All blocks and lambdas create it. One block frame per one block call.
  4. it can be a class frame, however it does not mean that instantiating a class creates it. The whole class X; ... end does it. Later when you do X.new Ruby does not create any class frames. This frame represents a class definition.
  5. it can be a rescue frame that represents a code inside rescue => e; <...>; end block
  6. it can be an ensure frame (well, I’m sure you get it)
  7. there are also module (for a module body), sclass (for a singleton class) and a very unique eval frame.

And they are stored internally as a stack.

When you invoke caller method you see this stack (or some information based on it). Each entry in the error backtrace is based on the state of this stack when the error is thrown.

OK, let’s write some code to extend our VM:

class FrameClass
  COMMON_FRAME_ATTRIBUTES = %i[
    _self
    nesting
    locals
    file
    line
    name
  ].freeze

  def self.new(*arguments, &block)
    Struct.new(
      *COMMON_FRAME_ATTRIBUTES,
      *arguments,
      keyword_init: true
    ) do
      class_eval(&block)

      def self.new(iseq:, **attributes)
        instance = allocate

        instance.file = iseq.file
        instance.line = iseq.line
        instance.name = iseq.name
        instance.locals = {}

        instance.send(:initialize, **attributes)

        instance
      end
    end
  end
end

I like builders and I don’t like inheritance (the fact that frames share some common attributes does not mean that they should be inherited from an AbstractFrame class; also I don’t like abstract classes, they are dead by definition)

Here’s the first version of the TopFrame:

TopFrame = FrameClass.new do
  def initialize(**)
    self._self = TOPLEVEL_BINDING.eval('self')
    self.nesting = [Object]
  end

  def pretty_name
    "TOP #{file}"
  end
end

Let’s walk through the code:

  1. each frame has a common set of attributes:
  2. _self - what self returns inside the frame
  3. nesting - what Module.nesting returns (used for the relative constant lookup)
  4. locals - a set of local variables
  5. file and line - currently running __FILE__:__LINE__
  6. name - a human-readable name of the frame, we will use it mostly for debugging
  7. FrameClass is a builder that is capable of building a custom Frame class (similar to Struct class)
  8. FrameClass.new takes:
  9. a list of custom frame-class-specific attributes
  10. and a block that is evaluated in a context of the created frame class

So, the TopFrame class is a Struct-like class that:

  1. has a constructor that takes only common attributes (because we haven’t specified any in FrameClass.new)
  2. has a custom behavior in the constructor that sets _self and nesting
  3. has a custom pretty_name instance method

We will create as many classes as we need to cover all kinds of frames (I will return to it later, I promise)

Wrapping the ISeq #

I don’t like working with plain arrays, and as I mentioned above there’s a ton of useful information in the instruction sequence that we get from RubyVM::InstructionSequence.compile("code").to_a.

Let’s create a wrapper that knows the meaning of array items:

class ISeq
  attr_reader :insns

  def initialize(ruby_iseq)
    @ruby_iseq = ruby_iseq
    @insns = @ruby_iseq[13].dup
  end

  def file
    @ruby_iseq[6]
  end

  def line
    @ruby_iseq[8]
  end

  def kind
    @ruby_iseq[9]
  end

  def name
    @ruby_iseq[5]
  end

  def lvar_names
    @ruby_iseq[10]
  end

  def args_info
    @ruby_iseq[11]
  end
end

Instance methods are self-descriptive, but just in case:

  1. file/line return file/line where the iseq has been created
  2. kind returns a Symbol that we will later use to distinguish frames (:top for a TopFrame)
  3. insns returns a list of instructions
  4. name is an internal name of the frame that is used in stacktraces (for class X; end it returns <class:X>)
  5. lvar_names is an array of all local variable names that are used in the frame
  6. args_info is a special Hash with a meta-information about arguments (empty for all frames except methods)

Frame stack #

Frames are organized as a stack internally, every time when we enter a frame Ruby pushes it on a stack. When the frame ends (i.e. when its list of instruction ends or there’s a special [:leave] instruction) it pops it.

class FrameStack
  attr_reader :stack

  include Enumerable

  def initialize
    @stack = []
  end

  def each
    return to_enum(:each) unless block_given?
    @stack.each { |frame| yield frame }
  end

  def push(frame)
    @stack << frame
    frame
  end

  def push_top(**args)
    push TopFrame.new(**args)
  end

  def pop
    @stack.pop
  end

  def top
    @stack.last
  end

  def size
    @stack.size
  end

  def empty?
    @stack.empty?
  end
end

Each entry in the stack is a frame that we entered at some point, so we can quickly build a caller:

class BacktraceEntry < String
  def initialize(frame)
    super("#{frame.file}:#{frame.line}:in `#{frame.name}'")
  end
end

stack = FrameStack.new
code = '2 + 3'
iseq = ISeq.new(RubyVM::InstructionSequence.compile(code, 'test.rb', '/path/to/test.rb', 42).to_a)

stack.push_top(iseq: iseq)
caller = stack.map { |frame| BacktraceEntry.new(frame) }.join("\n")
puts caller
# => "/path/to/test.rb:42: in `<compiled>'"

Writing the evaluator #

Let’s write it in a “script style” (it is simplified for a good reason, real code is much more complicated):

iseq = ISeq.new(RubyVM::InstructionSequence.compile_file('path/to/file.rb').to_a)
frame_stack = FrameStack.new
stack = []

frame_stack.push_top(iseq: iseq)

until frame_stack.empty?
  current_frame = frame_stack.top
  if current_frame.nil?
    break
  end

  if current_frame.insns.empty?
    frame_stack.pop_frame
    next
  end

  current_insn = current_frame.insns.shift
  execute_insn(current_insn)
end

Generally speaking, the code above is the core of the VM. Once it’s executed both frame_stack and stack must be empty. I added a bunch of consistency checks in my implementation, but for the sake of simplicity I’m going to omit them here.

Instructions #

I’ll try to be short here, there are about 100 instructions in Ruby, and some of them look similar.

putself, putobject, putnil, putstring, putiseq #

All of these guys push a simple object at the top of the stack. putnil pushes a known global nil object, others have an argument that is used in stack.push(argument)

optimized instructions like opt_plus #

Ruby has a mode (that is turned on by default) that optimizes some frequently used method calls, like + or .size. It is possible to turn them off by manipulating RubyVM::InstructionSequence.compile_option (if you set :specialized_instruction to false you’ll get a normal method call instead of the specialized instruction).

All of them do one specific thing, here’s an example of the opt_size:

def execute_opt_size(_)
  push(pop.size)
end

Of course we do it this way because we cannot optimize it. MRI does a different thing:

  1. if there’s a string/array/hash on top of the stack
  2. and String#size (or Array#size if it’s an array) is not redefined
  3. then it directly calls a C method rb_str_length (or RARRAY_LEN if it’s an array)
  4. otherwise (if it’s an object of some other type or a method has been redefined) it calls a method through the regular method dispatch mechanism (which is obviously slower)

We could do the same sequence of steps, but we can’t invoke a C method, and so calling a check + .size afterwards is even slower. It’s better for us to fall to the slow branch from the beginning.

You can print all available specialized instructions by running

RubyVM::INSTRUCTION_NAMES.grep(/\Aopt_/)

On Ruby 2.6.4 there are 34 of them.

opt_send_without_block (or send if specialized instructions are disabled) #

This is an instruction that is used to invoke methods. puts 123 looks like this:

[:opt_send_without_block, {:mid=>:puts, :flag=>20, :orig_argc=>1}, false]

It has 2 arguments:

  1. a hash with options
  2. mid - a method ID (method name)
  3. flag - a bitmask with a metadata about the invocation
  4. orig_argc - a number of arguments passed to a method call
  5. a boolean flag that is called CALL_DATA in C. I have no idea what it does

Here’s the rough implementation:

def execute_opt_send_without_block(options, _)
  mid = options[:mid]
  args = options[:orig_argc].times.map { pop }.reverse
  recv = pop
  result = recv.send(mid, *args)
  push(result)
end

So here we

  1. take a method from the options hash (mid)
  2. then we pop N arguments from the stack (args)
  3. then we pop the receiver of the method (recv)
  4. then call recv.send(mid, *args) (in our case it’s self.send(:puts, *[123])
  5. and then we push the result back to the stack

method definition #

I intentionally started with method calls because Ruby defines methods via method calls. Yes.

Ruby has a special singleton object called Frozen Core. When you define a method via def m; end Ruby invokes frozen_core.send("core#define_method", method_iseq):

$ ruby --dump=insns -e 'def m; end'
== disasm: #<ISeq:<main>@-e:1 (1,0)-(1,10)> (catch: FALSE)
0000 putspecialobject             1                                   (   1)[Li]
0002 putobject                    :m
0004 putiseq                      m
0006 opt_send_without_block       <callinfo!mid:core#define_method, argc:2, ARGS_SIMPLE>, <callcache>
0009 leave

The object itself is defined here .

Of course, we don’t have access to the Frozen Core. But we have an instruction that pushes it at the top of the stack. We can create our own FrozenCore = Object.new.freeze and check if recv is equal to this frozen core.

As you may notice there are also putobject :m and putiseq instructions. And argc of the method call is 2. Hmmm.

core#define_method takes two arguments:

  1. a method name that is pushed by a putobject instruction
  2. and an iseq that is … pushed by the putiseq instruction. Yes, instruction is an argument for another instruction.

Here’s the code:

if recv.equal?(FrozenCore) && mid == :'core#define_method'
  method_name, body_iseq = *args
  result = __define_method(method_name: method_name, body_iseq: body_iseq)
end

Here’s how __define_method looks:

def __define_method(method_name:, body_iseq:)
  parent_nesting = current_frame.nesting
  define_on = MethodDefinitionScope.new(current_frame)

  define_on.define_method(method_name) do |*method_args, &block|
    execute(body_iseq, _self: self, method_args: method_args, block: block, parent_nesting: parent_nesting)
  end

  method_name
end

When we enter a method in Ruby it inherits Module.nesting of a frame that defines it. This is why we also copy current_frame.nesting to a method frame.

define_on = MethodDefinitionScope.new(current_frame) is also quite simple:

class MethodDefinitionScope
  def self.new(frame)
    case frame._self
    when Class, Module
      frame._self
    when TOPLEVEL_BINDING.eval('self')
      Object
    else
      frame._self.singleton_class
    end
  end
end

These code constructions are equivalent:

def m; end
# and
Object.define_method(:m) {}

class X; def m; end; end
# and
X.define_method(:m) {}

o = Object.new
o.instance_eval { def m; end }
# and
o.singleton_class.define_method(:m) {}

Then comes this part:

define_on.define_method(method_name) do |*method_args, &block|
  execute(body_iseq, _self: self, method_args: method_args, block: block, parent_nesting: parent_nesting)
end

We define a method that takes any arguments (it breaks Method#parameters, but let’s ignore it) and an optional block and executes the ISeq of the method body in a context of self.

I admit that it’s a very hacky trick, but it allows us to dynamically assign self.

Plus, we pass all other things that can (and in most cases will) be used in a method body:

  1. method_args - what was given to a particular invocation of our method
  2. block - a block given to a method call
  3. parent_nesting - Module.nesting in the outer scope. We have to store in the beginning of the method definition because it may change before the method gets called.

execute(iseq, **options) is a tiny wrapper that pushes the frame into the frame_stack depending on the kind of the given iseq:

def execute(iseq, **payload)
  iseq = ISeq.new(iseq)
  push_frame(iseq, **payload)
  evaluate_last_frame
  pop_frame
end

def push_frame(iseq, **payload)
  case iseq.kind
  when :top
    @frame_stack.push_top(
      iseq: iseq
    )
  when :method
    @frame_stack.push_method(
      iseq: iseq,
      parent_nesting: payload[:parent_nesting],
      _self: payload[:_self],
      arg_values: payload[:method_args],
      block: payload[:block]
    )
  else
    raise NotImplementedError, "Unknown iseq kind #{iseq.kind.inspect}"
  end
end

local variables #

there are 2 most-commonly used instructions to get/set locals:

Both take two arguments:

Here’s an example

> pp RubyVM::InstructionSequence.compile('a = 10; b = 20; a; b').to_a[13]
[[:putobject, 10],
 [:setlocal_WC_0, 4],
 [:putobject, 20],
 [:setlocal_WC_0, 3],
 [:getlocal_WC_0, 3],
 [:leave]]

We push 10 to the stack, then we pop it and assign to a variable with ID = 4 in the current frame (setlocal_WC_0 here is a specialized instruction that is setlocal 0, 4 when the optimization is turned off).

Here’s the code to maintain locals:

require 'set'

# A simple struct that represents a single local variable;
# has a name, an ID and a value (or no value)
Local = Struct.new(:name, :id, :value, keyword_init: true) do
  def get
    value
  end

  def set(value)
    self.value = value
    value
  end
end

# A wrapper around "Set" that holds all locals for some frame;
# Absolutely each frame has its own instance of "Locals"
class Locals
  UNDEFINED = Object.new
  def UNDEFINED.inspect; 'UNDEFINED'; end

  def initialize(initial_names)
    @set = Set.new

    initial_names.reverse_each.with_index(3) do |arg_name, idx|
      # implicit args (like a virtual attribute that holds mlhs value) have numeric names
      arg_name += 1 if arg_name.is_a?(Integer)
      declare(name: arg_name, id: idx).set(Locals::UNDEFINED)
    end
  end

  def declared?(name: nil, id: nil)
    !find_if_declared(name: name, id: id).nil?
  end

  def declare(name: nil, id: nil)
    local = Local.new(name: name, id: id, value: nil)
    @set << local
    local
  end

  def find_if_declared(name: nil, id: nil)
    if name
      @set.detect { |var| var.name == name }
    elsif id
      @set.detect { |var| var.id == id }
    else
      raise NotImplementedError, "At least one of name:/id: is required"
    end
  end

  def find(name: nil, id: nil)
    result = find_if_declared(name: name, id: id)

    if result.nil?
      raise InternalError, "No local name=#{name.inspect}/id=#{id.inspect}"
    end

    result
  end

  def pretty
    @set
      .map { |local| ["#{local.name}(#{local.id})", local.value] }
      .sort_by { |(name, value)| name }
      .to_h
  end
end

So locals inside a frame is just a set. It is possible to declare a local, to check if it’s declared and to get it.

Here’s the implementation of getlocal:

def execute_getlocal(local_var_id, n)
  frame = n.times.inject(current_frame) { |f| f.parent_frame }
  local = frame.locals.find(id: local_var_id)
  value = local.get
  if value.equal?(Locals::UNDEFINED)
    value = nil
  end
  push(value)
end

We jump out n times to get the Nth frame, we find a local, we return nil if it’s undefined and we push the value back to the stack (so the result can be used by a subsequent instruction)

Here’s the implementation of setlocal:

def execute_setlocal(local_var_id, n)
  value = pop
  frame = n.times.inject(current_frame) { |f| f.parent_frame }
  local =
    if (existing_local = frame.locals.find_if_declared(id: local_var_id))
      existing_local
    elsif frame.equal?(current_frame)
      frame.locals.declare(id: local_var_id)
    else
      raise InternalError, 'locals are malformed'
    end

  local.set(value)
end

This one is a bit more complicated:

  1. first, we pop the value from the stack (it was pushed by a previous instruction putobject 10)
  2. then we find Nth frame
  3. then we get a local from this frame
  4. if it’s there we use it; if not - we declare it in the current frame
  5. and then we set the value

method arguments #

Every method has a list of arguments. Yes, sometimes it’s empty, but even in such case we do an arity check. In general arguments initialization is a part of every method call.

This part of Ruby is really complicated, because we have 12 argument types:

First, let’s take a look at the iseq to see what we have:

> pp RubyVM::InstructionSequence.compile('def m(a, b = 42, *c, d); end').to_a[13][4][1]
["YARVInstructionSequence/SimpleDataFormat",
 2,
 6,
 1,
 {:arg_size=>4,
  :local_size=>4,
  :stack_max=>1,
  :node_id=>7,
  :code_location=>[1, 0, 1, 28]},
 "m",
 "<compiled>",
 "<compiled>",
 1,
 :method,
 [:a, :b, :c, :d],
 {:opt=>[:label_0, :label_4],
  :lead_num=>1,
  :post_num=>1,
  :post_start=>3,
  :rest_start=>2},
 [],
 [:label_0,
  1,
  [:putobject, 42],
  [:setlocal_WC_0, 5],
  :label_4,
  [:putnil],
  :RUBY_EVENT_RETURN,
  [:leave]]]

There are two entries that we are interested in:

  1. a list of argument names (well, it’s a list of all variables, but it works for our case)
  2. a hash with arguments information

Let’s prepare and group it first, it’s hard to work with such format:

class CategorizedArguments
  attr_reader :req, :opt, :rest, :post, :kw, :kwrest, :block

  def initialize(arg_names, args_info)
    @req = []
    @opt = []
    @rest = nil
    @post = []

    parse!(arg_names.dup, args_info.dup)
  end

  def parse!(arg_names, args_info)
    (args_info[:lead_num] || 0).times do
      req << take_arg(arg_names)
    end

    opt_info = args_info[:opt].dup || []
    opt_info.shift
    opt_info.each do |label|
      opt << [take_arg(arg_names), label]
    end

    if args_info[:rest_start]
      @rest = take_arg(arg_names)
    end

    (args_info[:post_num] || 0).times do
      post << take_arg(arg_names)
    end
  end

  def take_arg(arg_names)
    arg_name_or_idx = arg_names.shift

    if arg_name_or_idx.is_a?(Integer)
      arg_name_or_idx += 1
    end

    arg_name_or_idx
  end
end

I intentionally skip keyword arguments here, but they are not that much different from other types of arguments. The only noticeable difference is that optional keyword arguments have “inlined” default values if they are simple enough (like plain strings or numbers, but not expressions like 2+2). If you are interested you can go to the repository and check this file.

Then, we should parse arguments and assign them into local variables when we push a method frame (so they are available once we start executing instructions of a method body):

class MethodArguments
  attr_reader :args, :values, :locals

  def initialize(iseq:, values:, locals:)
    @values = values.dup
    @locals = locals
    @iseq = iseq

    @args = CategorizedArguments.new(
      iseq.lvar_names,
      iseq.args_info
    )
  end

  def extract(arity_check: false)
    if arity_check && values.length < args.req.count + args.post.count
      raise ArgumentError, 'wrong number of arguments (too few)'
    end

    # Required positional args
    args.req.each do |name|
      if arity_check && values.empty?
        raise ArgumentError, 'wrong number of arguments (too few)'
      end

      value = values.shift
      locals.find(name: name).set(value)
    end

    # Optional positional args
    args.opt.each do |(name, label)|
      break if values.length <= args.post.count

      value = values.shift
      locals.find(name: name).set(value)

      VM.jump(label)
    end

    # Rest positional argument
    if (name = args.rest)
      value = values.first([values.length - args.post.length, 0].max)
      @values = values.last(args.post.length)

      locals.find(name: name).set(value)
    end

    # Required post positional arguments
    args.post.each do |name|
      if arity_check && values.empty?
        raise ArgumentError, 'Broken arguments, cannot extract required argument'
      end

      value = values.shift
      locals.find(name: name).set(value)
    end

    # Make sure there are no arguments left
    if arity_check && values.any?
      raise ArgumentError, 'wrong number of arguments (too many)'
    end
  end
end

Here values is what we get in a method call in *method_args, locals is equal to MethodFrame#locals that is set to Locals.new by default.

Let’s write MethodFrame class!

MethodFrame = FrameClass.new do
  attr_reader :arg_values

  attr_reader :block

  def initialize(parent_nesting:, _self:, arg_values:, block:)
    self._self = _self
    self.nesting = parent_nesting

    @block = block

    MethodArguments.new(
      iseq: iseq,
      values: arg_values,
      locals: locals,
      block: iseq.args_info[:block_start] ? block : nil
    ).extract(arity_check: true)
  end

  def pretty_name
    "#{_self.class}##{name}"
  end
end

Method frame is just a regular frame that extracts arguments during its initialization.

constants #

A regular constant assignment (like A = 1) is based on a scope (Module.nesting for relative lookup) and two instructions:

  1. setconstant
  2. getconstant

Both have a single argument - a constant name. But how does Ruby distinguish relative and absolute constant lookup? I mean, what’s the difference between A and ::A?

Ruby uses a special instruction to set a “constant scope” that:

Let’s take a look at the non-optimized mode (because we can’t optimize it anyway):

> pp RubyVM::InstructionSequence.compile('A; ::B; Kernel::D').to_a[13]
[[:putnil],
 [:getconstant, :A],
 [:pop],

 [:putobject, Object],
 [:getconstant, :B],
 [:pop],

 [:putnil],
 [:getconstant, :Kernel],
 [:getconstant, :D],

 [:leave]]

A constant performs a relative lookup, so putnil is used.

::B constant performs a global lookup on the Object that is a known object, and so it’s inlined in the putobject instruction.

Kernel::D first searches for Kernel constant locally, then it uses it as a “scope” for a constant D.

Quite easy, right? Not so fast. Ruby uses Module.nesting to perform a bottom-top search. This is why it’s so important to maintain nesting value in frames. Thus, the lookup is performed on current_scope.nesting in reverse order:

def execute_getconstant(name)
  scope = pop

  search_in = scope.nil? ? current_scope.nesting.reverse : [scope]

  search_in.each do |mod|
    if mod.const_defined?(name)
      const = mod.const_get(name)
      push(const)
      return
    end
  end

  raise NameError, "uninitialized constant #{name}"
end

If the scope is given (via push in a previous instruction) we use it. Otherwise we have a relative lookup and so we must use current_scope.nesting.reverse.

setconstant is a bit simpler, because it always defines a constant on a scope set by a previous instruction:

> pp RubyVM::InstructionSequence.compile('A = 10; ::B = 20; Kernel::D = 30').to_a[13]
[[:putobject, 10],
 [:putspecialobject, 3],
 [:setconstant, :A],

 [:putobject, 20],
 [:putobject, Object],
 [:setconstant, :B],

 [:putobject, 30],
 [:dup],
 [:putnil],
 [:getconstant, :Kernel],
 [:setconstant, :D],

 [:leave]]

putspecialobject is an instruction that is (when called with 3) pushes a “current” scope.

def putspecialobject(kind)
  case kind
  when 3
    push(current_frame.nesting.last) # push "current" scope
  else
    raise NotImplementedError, "Unknown special object #{kind}"
  end
end

def setconstant(name)
  scope = pop
  value = pop

  scope.const_set(name, value)
end

Instance/Class variables #

Instance variables are always picked from the self of the current frame (they literally look like a simplified version of local variables that are always stored in self of the current scope):

> pp RubyVM::InstructionSequence.compile('@a = 42; @a').to_a[13]
[[:putobject, 42],
 [:setinstancevariable, :@a, 0],
 [:getinstancevariable, :@a, 0],
 [:leave]]

I guess you know how the code should look like:

def execute_getinstancevariable(name, _)
  value = current_frame._self.instance_variable_get(name)
  push(value)
end

def execute_setinstancevariable(name, _)
  value = pop
  current_frame._self.instance_variable_set(name, value)
end

Class variables are similar, but it is possible to get it in the instance method, so it uses self if our current frame is a ClassFrame or self.class otherwise:

> pp RubyVM::InstructionSequence.compile('@@a = 42; @@a').to_a[13]
[[:putobject, 42],
 [:setclassvariable, :@@a],
 [:getclassvariable, :@@a],
 [:leave]]

def execute_setclassvariable(name)
  value = pop
  klass = current_frame._self
  klass = klass.class unless klass.is_a?(Class)
  klass.class_variable_set(name, value)
end

def execute_getclassvariable(name)
  klass = current_frame._self
  klass = klass.class unless klass.is_a?(Class)
  value = klass.class_variable_get(name)
  push(value)
end

Literals #

But how can we construct arrays, hashes?

> pp RubyVM::InstructionSequence.compile('[ [:foo,a,:bar], [4,5], 42 ]').to_a[13]
[[:putobject, :foo],
 [:putself],
 [:send, {:mid=>:a, :flag=>28, :orig_argc=>0}, false, nil],
 [:putobject, :bar],
 [:newarray, 3],

 [:duparray, [4, 5]],

 [:putobject, 42],
 [:newarray, 3],
 [:leave]]

As you can see the strategy of building an array depends on its dynamicity:

The whole array is also dynamic (because one of its elements is also dynamic). Let’s define them:

def execute_duparray(array)
  push(array.dup)
end

def execute_newarray(size)
  array = size.times.map { pop }.reverse
  push(array)
end

Do hashes support inlining?

> pp RubyVM::InstructionSequence.compile('{ primitive: { foo: :bar }, dynamic: { c: d } }').to_a[13]
[[:putobject, :primitive],
 [:duphash, {:foo=>:bar}],

 [:putobject, :dynamic],
 [:putobject, :c],
 [:putself],
 [:send, {:mid=>:d, :flag=>28, :orig_argc=>0}, false, nil],
 [:newhash, 2],

 [:newhash, 4],
 [:leave]]

Yes! duphash contains an inlined hash that should be pushed to the stack as is. newhash has a numeric argument that represents the number of keys and values on the hash (i.e. keys * 2 or values * 2, there’s no difference). And once again, if at least one element of the hash is dynamic, the whole has is also dynamic and so it uses newhash:

def execute_duphash(hash)
  push(hash.dup)
end

def execute_newhash(size)
  hash = size.times.map { pop }.reverse.each_slice(2).to_h
  push(hash)
end

Why do we need .dup in duphash and duparray? The reason is simple: this instruction can be executed multiple times (if it’s a part of a method or block, for example), and so the same value will be pushed to the stack multiple times. One of the next instructions can modify it but literals have to stay static no matter what. Without using .dup the code like

2.times do
  p [1, 2, 3].pop
end

would print 3 and 2.

Splats #

Splat is one of the most beautiful features of Ruby. Splat is foo, bar = *baz (and also [*foo, *bar]):

> pp RubyVM::InstructionSequence.compile('a, b = *c, 42').to_a[13]
[[:putself],
 [:send, {:mid=>:c, :flag=>28, :orig_argc=>0}, false, nil],
 [:splatarray, true],
 [:putobject, 42],
 [:newarray, 1],
 [:concatarray],
 [:dup],

 [:expandarray, 2, 0],

 [:setlocal_WC_0, 4],
 [:setlocal_WC_0, 3],
 [:leave]]

splatarray pops the object from the stack, converts it to Array by calling to_a (if it’s not an array; otherwise there’s no type casting), and pushes the result back to the stack.

concatarray constructs an array from two top elements and pushes it back. So it changes the stack [a, b] to [ [a,b] ]. If items are arrays it expands and merges them.

expandarray expands it by doing pop and pushing items back to the stack. It takes the number of elements that need to be returned, so if an array is bigger it drops some items, if it’s too small - it pushes as many nils as needed.

def execute_splatarray(_)
  array = pop
  array = array.to_a unless array.is_a?(Array)
  push(array)
end

def execute_concatarray(_)
  last = pop
  first = pop
  push([*first, *last])
end

def execute_expandarray(size, _flag)
  array = pop

  if array.size < size
    array.push(nil) until array.size == size
  elsif array.size > size
    array.pop until array.size == size
  else
    # they are equal
  end

  array.reverse_each { |item| push(item) }
end

In fact expandarray is much, much more complicated, you can go to the repository and check it if you want.

Keyword splats (like { **x, **y }) are really similar to array splats, I’m not going to cover them here.

conditions (if/unless) #

To handle conditions Ruby uses local goto (just like in C). Target of the goto-like instruction is a label:

> pp RubyVM::InstructionSequence.compile("a = b = c = 42; if a; b; else; c; end").to_a[13]
[[:putobject, 42],
 [:dup],
 [:setlocal_WC_0, 3],
 [:dup],
 [:setlocal_WC_0, 4],
 [:setlocal_WC_0, 5],

 [:getlocal_WC_0, 5],
 [:branchunless, :label_20],
 [:jump, :label_16],

 :label_16,
 [:getlocal_WC_0, 4],
 [:jump, :label_22],

 :label_20,
 [:getlocal_WC_0, 3],

 :label_22,
 [:leave]]

Do you see these :label_<NN> symbols? They are used as markers. branchunless takes a single argument: a label that it to jump to if the value on the top of the stack is false or nil. If it’s true-like it does nothing.

def execute_branchunless(label)
  cond = pop
  unless cond
    jump(label)
  end
end

def jump(label)
  insns = current_frame.iseq.insns
  insns.drop_while { |insn| insn != label }
  insns.shift # to drop the label too
end

Here we do pop, check it and call jump if it’s false. jump skips instructions until it sees a given label.

MRI also has branchif and branchnil:

String interpolation/concatenation #

Ruby has a few compile-time optimizations that optimize code like

"a""b"
"#{'a'}#{'b'}"

into a string "ab". However more complicated cases with dynamic interpolation involve a few new instructions:

> pp RubyVM::InstructionSequence.compile('"#{a}#{:sym}"').to_a[13]
[[:putobject, ""],

 [:putself],
 [:send, {:mid=>:a, :flag=>28, :orig_argc=>0}, false, nil],
 [:dup],
 [:checktype, 5],
 [:branchif, :label_18],
 [:dup],
 [:send, {:mid=>:to_s, :flag=>20, :orig_argc=>0}, false, nil],
 [:tostring],
 :label_18,

 [:putobject, :sym],
 [:dup],
 [:checktype, 5],
 [:branchif, :label_31],
 [:dup],
 [:send, {:mid=>:to_s, :flag=>20, :orig_argc=>0}, false, nil],
 [:tostring],
 :label_31,

 [:concatstrings, 3],
 [:leave]]

Parts above are split into sections.

  1. There’s an “invisible” beginning of the string (putobject "")
  2. Then we have an interpolated method call a:
  3. First we call a via send
  4. then we run checktype instruction that checks for an argument 5 that what’s popped is a string. it pushes back a boolean value
  5. then we conditionally invoke to_s if the object is not a string
  6. then we have an interpolated symbol :sym that gets interpolated in the same way
  7. and finally we invoke concatstrings 3 that does pop 3 times, concatenates 3 strings and pushes the result back to the stack

First let’s take a look at the checktype instruction:

CHECK_TYPE = ->(klass, obj) {
  klass === obj
}.curry

RB_OBJ_TYPES = {
  0x00 => ->(obj) { raise NotImplementedError },     # RUBY_T_NONE

  0x01 => CHECK_TYPE[Object],                        # RUBY_T_OBJECT
  0x02 => ->(obj) { raise NotImplementedError },     # RUBY_T_CLASS
  0x03 => ->(obj) { raise NotImplementedError },     # RUBY_T_MODULE
  0x04 => ->(obj) { raise NotImplementedError },     # RUBY_T_FLOAT
  0x05 => CHECK_TYPE[String],                        # RUBY_T_STRING
  0x06 => ->(obj) { raise NotImplementedError },     # RUBY_T_REGEXP
  0x07 => ->(obj) { raise NotImplementedError },     # RUBY_T_ARRAY
  0x08 => ->(obj) { raise NotImplementedError },     # RUBY_T_HASH
  0x09 => ->(obj) { raise NotImplementedError },     # RUBY_T_STRUCT
  0x0a => ->(obj) { raise NotImplementedError },     # RUBY_T_BIGNUM
  0x0b => ->(obj) { raise NotImplementedError },     # RUBY_T_FILE
  0x0c => ->(obj) { raise NotImplementedError },     # RUBY_T_DATA
  0x0d => ->(obj) { raise NotImplementedError },     # RUBY_T_MATCH
  0x0e => ->(obj) { raise NotImplementedError },     # RUBY_T_COMPLEX
  0x0f => ->(obj) { raise NotImplementedError },     # RUBY_T_RATIONAL

  0x11 => ->(obj) { raise NotImplementedError },     # RUBY_T_NIL
  0x12 => ->(obj) { raise NotImplementedError },     # RUBY_T_TRUE
  0x13 => ->(obj) { raise NotImplementedError },     # RUBY_T_FALSE
  0x14 => ->(obj) { raise NotImplementedError },     # RUBY_T_SYMBOL
  0x15 => ->(obj) { raise NotImplementedError },     # RUBY_T_FIXNUM
  0x16 => ->(obj) { raise NotImplementedError },     # RUBY_T_UNDEF

  0x1a => ->(obj) { raise NotImplementedError },     # RUBY_T_IMEMO
  0x1b => ->(obj) { raise NotImplementedError },     # RUBY_T_NODE
  0x1c => ->(obj) { raise NotImplementedError },     # RUBY_T_ICLASS
  0x1d => ->(obj) { raise NotImplementedError },     # RUBY_T_ZOMBIE
  0x1e => ->(obj) { raise NotImplementedError },     # RUBY_T_MOVED

  0x1f => ->(obj) { raise NotImplementedError },     # RUBY_T_MASK
}.freeze

def execute_checktype(type)
  item_to_check = pop
  check = RB_OBJ_TYPES.fetch(type) { raise InternalError, "checktype - unknown type #{type}" }
  result = check.call(item_to_check)
  push(result)
end

I blindly took it from MRI and yes, this instruction supports many types. I implemented only two of them, but the rest looks simple (except imemo and friends). Honestly I have no idea why, but about 95% of specs from the RubySpec (only language group, I did not check the whole test suite) are passing with these missing parts. I have no idea how to trigger MRI to use them. Maybe it uses them internally?

concatstrings looks just like newarray:

def execute_concatstrings(count)
  strings = count.times.map { pop }.reverse
  push(strings.join)
end

Blocks #

Blocks are passed to method calls as a third argument:

> pp RubyVM::InstructionSequence.compile('m { |a| a + 42 }').to_a[13]
[[:putself],
 [:send,
  {:mid=>:m, :flag=>4, :orig_argc=>0},
  false,
  ["YARVInstructionSequence/SimpleDataFormat",
   2,
   6,
   1,
   {:arg_size=>1,
    :local_size=>1,
    :stack_max=>2,
    :node_id=>7,
    :code_location=>[1, 2, 1, 16]},
   "block in <compiled>",
   "<compiled>",
   "<compiled>",
   1,
   :block,
   [:a],
   {:lead_num=>1, :ambiguous_param0=>true},
   [[:redo, nil, :label_1, :label_9, :label_1, 0],
    [:next, nil, :label_1, :label_9, :label_9, 0]],
   [1,
    :RUBY_EVENT_B_CALL,
    [:nop],
    :label_1,
    :RUBY_EVENT_LINE,
    [:getlocal_WC_0, 3],
    [:putobject, 42],
    [:send, {:mid=>:+, :flag=>16, :orig_argc=>1}, false, nil],
    :label_9,
    [:nop],
    :RUBY_EVENT_B_RETURN,
    [:leave]]]]]

Block definitely needs a frame that looks pretty much like a MethodFrame:

BlockFrame = FrameClass.new(:arg_values) do
  def initialize(arg_values:)
    self._self = parent_frame._self
    self.nesting = parent_frame.nesting

    MethodArguments.new(
      iseq: iseq,
      values: arg_values,
      locals: locals,
      block: nil
    ).extract(arity_check: false)
  end

  def pretty_name
    name
  end
end

(For simplicity let’s ignore that blocks can also take blocks; also let’s ignore lambdas, we will return to them later)

The code above looks almost like a method frame. The only difference is the arity_check value that we pass to the MethodArguments class.

But when should we create this frame? And how can we get a proc from it?

VM_CALL_ARGS_BLOCKARG   = (0x01 << 1)

def execute_send(options, flag, block_iseq)
  _self = self
  mid = options[:mid]

  args = []

  block =
    if block_iseq
      proc do |*args|
        execute(block_iseq, self: _self, arg_values: args)
      end
    elsif flag & VM_CALL_ARGS_BLOCKARG
      pop
    else
      nil
    end

  args = options[:orig_argc].times.map { pop }.reverse
  recv = pop

  result = recv.send(mid, *args, &block)

  push(result)
end

It looks like a more generalized version of opt_send_without_block, because opt_send_without_block is a specialized implementation of send.

This instruction also pops a receiver and arguments, but what’s important, it also computes the block.

  1. If block_iseq is given we create a proc that (once called) executes a block instruction (i.e. a block body) with given arguments. This block uses self of the place where it was created. (i.e. self == proc { self }.call always returns true)
  2. If there’s no block_iseq the block can be given via a &block argument. MRI marks method call as VM_CALL_ARGS_BLOCKARG (this flag is just a bitmask)
  3. and then we simply call a method with a generated proc object.

Implicit block like b = proc {}; m(&b) does not need any additional implementation. Method proc here takes a block (handled by the first if branch), it gets stored in a local variable and we pass it to the method as a block argument (elseif branch).

Lambdas #

It’s complicated and I don’t have a complete solution that covers all cases (I guess because MRI does not expose enough APIs to do it. Or I’m just not smart enough).

Arrow lambda (->(){}) is just a method call FrozenCore#lambda, and so we can easily determine that it’s a lambda and not a proc. But what about lambda {}? It can be overwritten.

An incomplete (and somewhat unreliable) solution is to check that our receiver does not override lambda method inherited from a Kernel module:

creating_a_lambda = false

if mid == :lambda
  if recv.equal?(FrozenCore)
    # ->{} syntax
    creating_a_lambda = true
  end

  if recv.class.instance_method(:lambda).owner == Kernel
    if Kernel.instance_method(:lambda) == RubyRb::REAL_KERNEL_LAMBDA
      # an original "lambda" method from a Kernel module
      creating_a_lambda = true
    end
  end
end

Then we can set it on our block frame as an attribute.

# in the branch that creates a proc from the `block_iseq`
proc do |*args|
  execute(block_iseq, self: _self, arg_values: args, is_lambda: creating_a_lambda)
end

BlockFrame = FrameClass.new(:arg_values, :is_lambda) do
  def initialize(arg_values:, is_lambda:)
    self.is_lambda = is_lambda
    self._self = parent_frame._self
    self.nesting = parent_frame.nesting

    MethodArguments.new(
      iseq: iseq,
      values: arg_values,
      locals: locals,
      block: nil
    ).extract(arity_check: is_lambda)
  end
end

Arity check is enabled only if our proc is a lambda.

Calling a block #

If you remember when we define a method we tell it to save given block in a method frame:

define_on.define_method(method_name) do |*method_args, &block|
  execute(body_iseq, _self: self, method_args: method_args, block: block, parent_nesting: parent_nesting)
end

And the frame itself saves it in the attr_reader.

So both explicit and implicit blocks are available in a method body via current_frame.block. It’s possible to invoke it by calling block.call(arguments) (if it’s available as an explicit block argument) or to call yield(arguments) (in such case it does not even have to be declared in a method signature).

pp RubyVM::InstructionSequence.compile('def m; yield; end').to_a[13][4][1][13]
[[:invokeblock, {:mid=>nil, :flag=>16, :orig_argc=>0}],
 [:leave]]

Honestly even before I started working on this article I expected MRI to do something like this. yield is equivalent to <current block>.call(args):

def execute_invokeblock(options)
  args = options[:orig_argc].times.map { pop }.reverse

  frame = current_frame
  frame = frame.parent_frame until frame.can_yield?

  result = frame.block.call(*args)
  push(result)
end

Do you see frame = frame.parent_frame until frame.can_yield?? The reason for this line is that you may have a code like

def m
  [1,2,3].each { |item| yield item }
end

^ yield here belongs to the method m, not to the BlockFrame of the .each method. There can be more nested blocks, so we have to go up until we see something that supports yield. Well, we know that only one frame can do yield: it’s a MethodFrame.

Our frame class factory need to be extended to generate this method by default and return false from it. MethodFrame has to override it and return true. Polymorphism!

Super #

Calling super is very similar to calling yield: it can be replaced with method(__method__).super_method.call(args).

__method__ can be retrieved from current_frame.name, args are processed using options[:orig_argc]:

def execute_invokesuper(options, _, _)
  recv = current_frame._self
  mid = current_frame.name
  super_method = recv.method(mid).super_method
  args = options[:orig_argc].times.map { pop }.reverse

  result = super_method.call(*args)
  push(result)
end

This implementation is incorrect, it can’t handle a sequence of super calls (class A < B < C, each has a method that calls super). I guess it’s possible to implement it by recording the class where the method was defined (i.e. by storing current_frame._self before calling define_method and passing it to the MethodFrame constructor as a defined_in attribute). This way we could do something like this:

def execute_invokesuper(options, _, _)
  recv = current_frame._self
  mid = current_frame.name

  dispatchers = recv.class.ancestors
  current_dispatcher_idx = dispatchers.index(current_frame.defined_in)
  next_dispatcher = dispatchers[current_dispatcher_idx + 1]

  super_method = next_dispatcher.instance_method(mid).bind(recv)
  args = options[:orig_argc].times.map { pop }.reverse

  result = super_method.call(*args)
  push(result)
end

I did not implement it because MSpec does not rely on it and I usually try to avoid sequences of super calls.

Global variables #

Similar to locals and instance variables, there are getglobal/setglobal instructions. They also take a variable name as an argument.

Unfortunately, Ruby has no API to dynamically get/set global variables. But we have eval!

def execute_getglobal((name))
  push eval(name.to_s)
end

def execute_setglobal((name))
  # there's no way to set a gvar by name/value
  # but eval can reference locals
  value = pop
  eval("#{name} = value")
end

defined? keyword #

As you may know this keyword can handle pretty much anything:

> pp RubyVM::InstructionSequence.compile('defined?(42)').to_a[13]
[:putobject, "expression"],
 [:leave]]

> pp RubyVM::InstructionSequence.compile('a = 42; defined?(a)').to_a[13]
[[:putobject, 42],
 [:setlocal_WC_0, 3],
 [:putobject, "local-variable"],
 [:leave]]

In some simple cases it does not do any computations. It’s obvious that 42 is an expression and a is a local variable (and there’s no way to remove it by any code between assignment and defined? check)

More advanced checks use a defined instruction:

> pp RubyVM::InstructionSequence.compile('@a = 42; defined?(@a)').to_a[13]
[[:putobject, 42],
 [:setinstancevariable, :@a, 0],
 [:putnil],
 [:defined, 2, :@a, true],
 [:leave]]

The first argument is a special enum flag that specifies what are we trying to check here:

module DefinedType
  DEFINED_NOT_DEFINED = 0
  DEFINED_NIL = 1
  DEFINED_IVAR = 2
  DEFINED_LVAR = 3
  DEFINED_GVAR = 4
  DEFINED_CVAR = 5
  DEFINED_CONST = 6
  DEFINED_METHOD = 7
  DEFINED_YIELD = 8
  DEFINED_ZSUPER = 9
  DEFINED_SELF = 10
  DEFINED_TRUE = 11
  DEFINED_FALSE = 12
  DEFINED_ASGN = 13
  DEFINED_EXPR = 14
  DEFINED_IVAR2 = 15
  DEFINED_REF = 16
  DEFINED_FUNC = 17
end

I’ll show you the branch that handles instance variables:

def execute_defined(defined_type, obj, needstr)
  # used only in DEFINED_FUNC/DEFINED_METHOD branches
  # but we still have to do `pop` here (even if it's unused)
  context = pop

  verdict =
    case defined_type
    when DefinedType::DEFINED_IVAR
      ivar_name = obj
      if current_frame._self.instance_variable_defined?(ivar_name)
        'instance-variable'
      end
    # ... other branches
    end

  push(verdict)
end

All other branches are similar, they do some check and push a constant string or nil back to the stack.

Range literals #

For static ranges (like (1..2)) Ruby uses inlining and a well-known putobject instruction. But what if it’s dynamic? Like (a..b)

> pp RubyVM::InstructionSequence.compile('a = 3; b = 4; p (a..b); p (a...b)').to_a[13]
[[:putobject, 3],
 [:setlocal_WC_0, 4],
 [:putobject, 4],
 [:setlocal_WC_0, 3],
 [:putself],

 [:getlocal_WC_0, 4],
 [:getlocal_WC_0, 3],
 [:newrange, 0],

 [:send, {:mid=>:p, :flag=>20, :orig_argc=>1}, false, nil],
 [:pop],
 [:putself],

 [:getlocal_WC_0, 4],
 [:getlocal_WC_0, 3],
 [:newrange, 1],

 [:send, {:mid=>:p, :flag=>20, :orig_argc=>1}, false, nil],
 [:leave]]

There’s a special newrange instruction that takes a flag as an argument to specify inclusion of the right side (i.e. to distinguish .. vs ...)

def execute_newrange(flag)
  high = pop
  low = pop
  push(Range.new(low, high, flag == 1))
end

Long jumps #

This is probably the most complicated part. What if you have a method that has a loop inside a loop that does return? You want to stop executing both loops and simply exit the method, right?

def m
  2.times do |i|
    3.times do |j|
      return [i,j]
    end
  end
end

m

Of course you can just find the closest frame that supports return (i.e. a MethodFrame), but you also need to stop execution of two running methods and blocks. In our case it’s even more complicated because we don’t control them (they are written in C).

The only way I was able to find is to throw an exception. An exception destroys all frames (including YARV’s C frames) until it finds someone who can catch and handle it. If there’s no such frame the programs exits with an error.

Let’s create a special exception class called VM::LongJumpError. Each frame class has to know what it can handle (for example, you can do break in a block, but not in a method; return is normally supported only by methods and lambdas, etc):

class LongJumpError  < InternalError
  attr_reader :value

  def initialize(value)
    @value = value
  end

  def do_jump!
    raise InternalError, 'Not implemented'
  end

  def message
    "#{self.class}(#{@value.inspect})"
  end
end

class ReturnError < LongJumpError
  def do_jump!
    frame = current_frame

    if frame.can_return?
      # swallow and consume
      frame.returning = self.value
    else
      pop_frame(reason: "longjmp (return) #{self}")
      raise self
    end
  end
end

Each longjmp exception wraps the value that it “returns” with (or “breaks” with, for break we need a separate class, but I’m going to skip it here. break/next and other friends are really similar to return).

But we need to catch them, right? Without a rescue handler we will have something conceptually similar to segfault:

def execute(iseq, **payload)
  iseq = ISeq.new(iseq)
  push_frame(iseq, **payload)
  # here comes the difference:
  # we wrap executing instructions into a rescue handler
  begin
    evaluate_last_frame
  rescue LongJumpError => e
    e.do_jump!
  end
  pop_frame
end

The only missing thing is the implementation of can_return? method in our frames. All frames except MethodFrame (and BlockFrame it it’s marked as lambda) must return false, MethodFrame must return true.

MRI uses a special instruction called throw that has a single argument that is a throw_type (an enum, for return it’s 1, break is 3, next is 4, there are also retry/redo and a few more). The value that must be attached to the thrown exception comes from the stack (so this instruction does a single pop)

VM_THROW_STATE_MASK = 0xff

RUBY_TAG_NONE = 0x0
RUBY_TAG_RETURN = 0x1
RUBY_TAG_BREAK = 0x2
RUBY_TAG_NEXT = 0x3
RUBY_TAG_RETRY = 0x4
RUBY_TAG_REDO = 0x5
RUBY_TAG_RAISE = 0x6
RUBY_TAG_THROW = 0x7
RUBY_TAG_FATAL = 0x8
RUBY_TAG_MASK = 0xf

def execute_throw(throw_type)
  throw_type = throw_type & VM_THROW_STATE_MASK
  throw_obj = pop

  case throw_type
  when RUBY_TAG_RETURN
    raise VM::ReturnError, throw_obj
  when RUBY_TAG_BREAK
    raise VM::BreakError, throw_obj
  when RUBY_TAG_NEXT
    raise VM::NextError, throw_obj
  # ...
  end
end

longjmp in MRI #

But does it work in the same way in MRI? C does not have exceptions. And at the same time there is a bunch of places where MRI does something like

if (len > ARY_MAX_SIZE) {
    rb_raise(rb_eArgError, "array size too big");
}
// handle validated data

This rb_raise somehow exits a C function. Well, here’s the trick: non-local goto.

There are C calls that perform a goto to any place (that was previously marked of course, similar to jump to labels for a local goto):

setjmp() saves the stack context/environment in env for later use by longjmp

… also known as “context switch”. And it’s relatively expensive.

Even if you don’t raise an exception and only do begin; ...; rescue; end in your code you still have to save the context (to jump to it once you raise an error). MRI does not know at compile time which methods can throw an error (and do you throw them at all), so each rescue produces a setjmp call (and each raise triggers a longjmp and passes closest rescue -> saved env as an argument)

rescue/ensure #

So now we know that raise/rescue works via long jumps under the hood. Let’s implement our own exceptions.

By sticking to MRI exceptions we can unwrap both internal and our stacks at the same time. I’m not going to override raise, it should do what it originally does, but we still need to support our own rescue blocks. Let’s see what MRI gives us:

> pp RubyVM::InstructionSequence.compile('begin; p "x"; rescue A; p "y"; end').to_a
[ # ...snip
 [[:rescue,
   [ # ...snip
    "rescue in <compiled>",
    "<compiled>",
    "<compiled>",
    1,
    :rescue,
    [:"\#$!"],
    {},
    [],
    [1,
     [:getlocal_WC_0, 3],
     [:putnil],
     [:getconstant, :A],
     [:checkmatch, 3],
     [:branchif, :label_11],
     [:jump, :label_19],
     :label_11,
     :RUBY_EVENT_LINE,
     [:putself],
     [:putstring, "y"],
     [:send, {:mid=>:p, :flag=>20, :orig_argc=>1}, false, nil],
     [:leave],
     :label_19,
     [:getlocal_WC_0, 3],
     [:throw, 0]]],
   :label_0,
   :label_7,
   :label_8,
   0],
  [:retry, nil, :label_7, :label_8, :label_0, 0]],
 [:label_0,
  1,
  :RUBY_EVENT_LINE,
  [:putself],
  [:putstring, "x"],
  [:send, {:mid=>:p, :flag=>20, :orig_argc=>1}, false, nil],
  :label_7,
  [:nop],
  :label_8,
  [:leave]]]

An instruction sequence that has some rescue blocks inside includes all information about them (in the element #12, right above an instructions list). Each rescue handler is a frame with its own list of variables and instructions. Its kind is :rescue and is has at least one local variable: $!. It starts with a dollar sign, but it’s a local variable. According to its semantics it has to be a local variable, but unfortunately it can’t look like a local variable (because it’d would potentially conflict with method calls). I mean, that’s how I explain it to myself, I don’t know for sure what was the initial reason to design it this way.

It also has a few labels at the bottom - :label_7, :label_8, :label_0:

A top-level instruction also contains these labels, and the meaning of them is:

Let’s code it!

class ISeq
  attr_reader :rescue_handlers
  attr_reader :ensure_handlers

  def initialize(ruby_iseq)
    @ruby_iseq = ruby_iseq
    reset!
    setup_rescue_handlers!
    setup_ensure_handlers!
  end

  # ... other existing methods on ISeq class

  def setup_rescue_handlers!
    @rescue_handler = @ruby_iseq[12]
      .select { |handler| handler[0] == :rescue }
      .map { |(_, iseq, begin_label, end_label, exit_label)| Handler.new(iseq, begin_label, end_label, exit_label) }
  end

  def setup_ensure_handlers!
    @ensure_handler = @ruby_iseq[12]
      .select { |handler| handler[0] == :ensure }
      .map { |(_, iseq, begin_label, end_label, exit_label)| Handler.new(iseq, begin_label, end_label, exit_label) }
  end

  class Handler
    attr_reader :iseq
    attr_reader :begin_label, :end_label, :exit_label
    def initialize(iseq, begin_label, end_label, exit_label)
      @iseq = iseq
      @begin_label, @end_label, @exit_label = begin_label, end_label, exit_label
    end
  end
end

So now each iseq object has two getters:

Frames must know which handlers are active (but not instruction sequences, because methods can recursively call themselves and so the same iseq will be reused; it’s a per-frame property):

class FrameClass
  def self.new
    Struct.new {
      # Both must be set to `Set.new` in the constructor
      attr_accessor :enabled_rescue_handlers
      attr_accessor :enabled_ensure_handlers
    }
  end
end

So this way all frames have it too. Every time when we see a label in our execution loop we need to check if it matches any begin_label or end_label of our current_frame.iseq.rescue_handlers (or ensure_handlers):

def on_label(label)
  {
    current_iseq.rescue_handlers => current_frame.enabled_rescue_handlers,
    current_iseq.ensure_handlers => current_frame.enabled_ensure_handlers,
  }.each do |all_handlers, enabled_handlers|
    all_handlers
      .select { |handler| handler.begin_label == label }
      .each { |handler| enabled_handlers << handler }

    all_handlers
      .select { |handler| handler.end_label == label }
      .each { |handler| enabled_handlers.delete(handler) }
  end
end

side note: when we do a local jump we should also walk through skipped instructions and enable/disable our handlers; this is important

OK, now the only missing part is the reworked execution loop:

# a generic runner that is used inside the loop
def execute_insn(insn)
  case insn
  when [:leave]
    current_frame.returning = stack.pop
  when Array
    name, *payload = insn

    with_error_handling do
      send(:"execute_#{name}", payload)
    end
  # -- new branch for labels --
  when Regexp.new("label_\\d+")
    on_label(insn)
  else
    # ignore
  end
end

# a wrapper that catches and handles error
def with_error_handling
  yield
rescue VM::InternalError => e
  # internal errors like LongJumpError should be invisible for users
  raise
rescue Exception => e
  handle_error(e)
end

def handle_error(error)
  if (rescue_handler = current_frame.enabled_rescue_handler.first)
    result = execute(rescue_handler.iseq, caught: error, exit_to: rescue_handler.exit_label)
    stack.push(result)
  else
    raise error
  end
end

# This guy also needs customization to support `jump(exit_label)`
def pop_frame
  frame = @frame_stack.pop
  if frame.is_a?(RescueFrame)
    jump(frame.exit_to)
  end
  frame.returning
end

# And here's the rescue frame implementation
RescueFrame = FrameClass.new do
  attr_reader :parent_frame, :caught, :exit_to

  def initialize(parent_frame:, caught:, exit_to:)
    self._self = parent_frame._self
    self.nesting = parent_frame.nesting

    @parent_frame = parent_frame
    @caught = caught
    @exit_to = exit_to

    # $! always has an ID = 3
    locals.declare(id: 3, name: :"\#$!")
    locals.find(id: 3).set(caught)
  end
end

But why do we handle only the first handler? Can there be multiple handlers? The answer is no, because:

throw/catch methods #

As a side (and I personally think a very interesting) note, while I was working on this project I have realized that specs for Kernel#throw are not working for me at all. They were literally completely broken (even after I finished working on a very basic implementation of exceptions):

catch(:x) do
  begin
    throw :x
  rescue Exception => e
    puts e
  end
end

This code does not print anything. However, if you do just throw :x you get an exception:

> throw :x
Traceback (most recent call last):
        2: from (irb):14
        1: from (irb):14:in `throw'
UncaughtThrowError (uncaught throw :x)

Huh, what’s going on? Let’s take a look at the implementation of Kernel#throw:

void
rb_throw_obj(VALUE tag, VALUE value)
{
    rb_execution_context_t *ec = GET_EC();
    struct rb_vm_tag *tt = ec->tag;

    while (tt) {
      if (tt->tag == tag) {
          tt->retval = value;
          break;
      }
      tt = tt->prev;
    }
    if (!tt) {
      VALUE desc[3];
      desc[0] = tag;
      desc[1] = value;
      desc[2] = rb_str_new_cstr("uncaught throw %p");
      rb_exc_raise(rb_class_new_instance(numberof(desc), desc, rb_eUncaughtThrow));
    }

    ec->errinfo = (VALUE)THROW_DATA_NEW(tag, NULL, TAG_THROW);
    EC_JUMP_TAG(ec, TAG_THROW);
}

This code raises an instance of rb_eUncaughtThrow only in one case: if there’s no frame above that may “catch” it (like in the case when we did just throw :x).

However if there’s a frame somewhere above that has the same tag MRI performs a manual longjmp. This is why we can’t catch this exception. There’s simply no exception if there’s a catch(:x) above (but there would be an exception if we would do catch(:y) { throw :x }).

Is it faster? Let’s see

require 'benchmark/ips'

Benchmark.ips do |x|
  x.config(:time => 3)

  x.report('raise') do
    begin
      raise 'x'
    rescue => e
    end
  end

  x.report('throw') do
    catch(:x) do
      throw :x
    end
  end

  x.compare!
end

$ ruby benchmark.rb
Warming up --------------------------------------
               raise   101.832k i/100ms
               throw   256.893k i/100ms
Calculating -------------------------------------
               raise      1.310M (± 4.2%) i/s -      3.971M in   3.037942s
               throw      4.853M (± 3.0%) i/s -     14.643M in   3.020227s

Comparison:
               throw:  4852821.4 i/s
               raise:  1309915.6 i/s - 3.70x  slower

As expected, it’s faster. But what’s more important, let’s see what if there are more frames between raise/rescue and throw/catch. First, let’s create a small method that wraps given code into N frames:

def in_n_frames(depth = 0, n, blk)
  if depth == n
    blk.call
  else
    in_n_frames(depth + 1, n, blk)
  end
end

begin
  in_n_frames(1000, proc { raise 'err' })
rescue => e
  p e.backtrace.length
end

It prints 1003, because there’s also a TopFrame and a BlockFrame (and a small bug that does +1), but that’s absolutely fine for us.

$ ruby benchmark.rb
Warming up --------------------------------------
               raise     1.061k i/100ms
               throw     1.115k i/100ms
Calculating -------------------------------------
               raise     10.628k (± 1.3%) i/s -     32.891k in   3.095347s
               throw     11.183k (± 1.5%) i/s -     34.565k in   3.091514s

Comparison:
               throw:    11183.1 i/s
               raise:    10627.8 i/s - 1.05x  slower

There’s almost no difference! The reason is simple: the only thing that is different is creation of the exception object. throw does not do it.

Utility instructions #

There are also a few interesting instructions that MRI uses to evaluate your complex code:

Final words #

First of all, I’d like to say thank you to everyone who made YARV. I was not able to find a single place where MRI behaves inefficiently (and I spent many hours looking into instructions).

Once again, the code is available here , feel free to create issues/message me in Twitter. And please, don’t use it in the real world.

Page views