Skip to content

limitations.md

Latest commit

 

History

History
440 lines (314 loc) · 10.5 KB

limitations.md

File metadata and controls

440 lines (314 loc) · 10.5 KB

Limitations and Differences

The philosophy of mruby is to be a lightweight implementation of the Ruby ISO standard. These two objectives are partially contradicting. Ruby is an expressive language with complex implementation details which are difficult to implement in a lightweight manner. To cope with this, limitations to the "Ruby Compatibility" are defined.

This document is collecting these limitations.

Integrity

This document does not contain a complete list of limitations. Please help to improve it by submitting your findings.

Kernel.raise in rescue clause

Kernel.raise without arguments does not raise the current exception within a rescue clause.

begin
  1 / 0
rescue
  raise
end

CRuby

ZeroDivisionError is raised.

mruby

RuntimeError is raised instead of ZeroDivisionError. To re-raise the exception, you have to do:

begin
  1 / 0
rescue => e
  raise e
end

Fiber execution can't cross C function boundary

mruby's Fiber is implemented similarly to Lua's co-routine. This results in the consequence that you can't switch context within C functions. Only exception is mrb_fiber_yield at return.

Array does not support instance variables

To reduce memory consumption Array does not support instance variables.

class Liste < Array
  def initialize(str = nil)
    @field = str
  end
end

p Liste.new "foobar"

CRuby

[]

mruby

ArgumentError is raised.

defined?

The defined? keyword is considered too complex to be fully implemented. It is recommended to use const_defined? and other reflection methods instead.

defined?(Foo)

CRuby

nil

mruby

NameError is raised.

alias on global variables

Aliasing a global variable works in CRuby but is not part of the ISO standard.

alias $a $__a__

CRuby

nil

mruby

Syntax error

Operator modification

Operators on some of the primitive classes cannot be overridden, as they are optimized in the VM.

class String
  def +
  end
end

'a' + 'b'

CRuby

ArgumentError is raised. The re-defined + operator does not accept any arguments.

mruby

'ab' Behavior of the operator wasn't changed.

Kernel#binding is not supported without mruby-binding gem

Kernel#binding method requires the mruby-binding gem (included in the metaprog gembox). Without this gem, binding is not available.

nil? redefinition in conditional expressions

Redefinition of nil? is ignored in conditional expressions.

a = "a"
def a.nil?
  true
end
puts(a.nil? ? "truthy" : "falsy")

Ruby outputs truthy. mruby outputs falsy.

Argument Destructuring

def m(a,(b,c),d); p [a,b,c,d]; end
m(1,[2,3],4)  # => [1,2,3,4]

Destructured arguments (b and c in above example) cannot be accessed from the default expression of optional arguments and keyword arguments, since actual assignment is done after the evaluation of those default expressions. Thus:

def f(a,(b,c),d=b)
  p [a,b,c,d]
end
f(1,[2,3])

CRuby gives [1,2,3,nil]. mruby raises NoMethodError for b.

Keyword argument expansion has similar restrictions. The following example, gives [1, 1] for CRuby, mruby raises NoMethodError for b.

def g(a: 1, b: a)
  p [a,b]
end
g(a:1)

No Double Dispatch in Module Loading

To make implementation simpler, mruby does not use double dispatching in module loading (include/prepend/extend). Those method internally called corresponding actual load methods (append_features/prepend_features/extend_object). But they are rarely overloaded, consumes more memory, and make loading little bit slower. As a Ruby implementation for the smaller device, we decided mruby simpler.

module M
  def self.append_features(mod)
     p :append
  end
end

class C
  include M
end

CRuby

Prints :append.

mruby

Nothing printed (since include does not call append_features internally).

No #hash call for small hashes

For performance reasons, mruby avoids calling the #hash method on keys when a hash table is small. This means that custom #hash methods on key objects may not be executed.

Pattern Matching

Pattern matching is only partially supported in mruby. Currently, only the rightward assignment operator (=>) with simple variable binding is implemented.

expr => var  # Supported: assigns expr to var

CRuby

Full pattern matching with case/in syntax and various pattern types:

case [1, 2, 3]
in [a, b, c]
  puts "#{a}, #{b}, #{c}"  # => "1, 2, 3"
end

case {name: "Alice", age: 30}
in {name:, age:}
  puts "#{name} is #{age}"  # => "Alice is 30"
end

mruby

Only rightward assignment with simple variable binding:

[1, 2, 3] => x
puts x  # => [1, 2, 3]

The following are not supported:

  • case/in syntax
  • Array patterns: in [a, b, c]
  • Hash patterns: in {name:, age:}
  • Guard clauses: in pattern if condition
  • Pin operator: in ^variable
  • Find patterns: in [*, x, *]
  • Alternative patterns: in pattern1 | pattern2
  • Boolean pattern check: value in pattern

Note: mruby does provide Array#deconstruct and Hash#deconstruct_keys methods for future pattern matching compatibility.

No Refinements

Module refinements (refine, using) are not supported in mruby.

No Encoding Class

mruby does not have an Encoding class. Strings are treated as byte sequences by default. UTF-8 aware string operations can be enabled with the MRB_UTF8_STRING compile flag.

Integer Precision Varies by Boxing Mode

Integer size depends on the value boxing configuration:

Configuration Integer range
Word boxing, 64-bit (default) roughly +/- 2^62
Word boxing, 32-bit (default) roughly +/- 2^30
NaN boxing (64-bit only) -2^31 to 2^31-1

Code relying on 64-bit integer precision may behave differently across configurations. The mruby-bigint gem provides arbitrary-precision integers when included.

No ObjectSpace.each_object by Default

ObjectSpace is only available via the mruby-objectspace gem (included in the stdlib gembox). Even with the gem, ObjectSpace.each_object has limited functionality compared to CRuby.

No Implicit Type Conversion (to_int, to_str, to_ary, ...)

mruby does not perform implicit type conversion through methods like to_int, to_str, to_ary, or to_hash. CRuby uses these to let user-defined classes duck-type as built-in types — for example Array#[] calls to_int on its argument, String#+ calls to_str, and multiple assignment calls to_ary on its right-hand side. mruby's built-in operations require the actual built-in type and do not consult these conversion methods.

class MyInt;  def to_int; 42; end;     end
class MyStr;  def to_str; "x"; end;    end
class MyAry;  def to_ary; [1,2,3]; end; end

CRuby

[1,2,3][MyInt.new]   # => nil   (to_int called -> ary[42])
"a" + MyStr.new      # => "ax"  (to_str called)
a, b, c = MyAry.new  # => a=1, b=2, c=3   (to_ary called)

mruby

[1,2,3][MyInt.new]   # TypeError
"a" + MyStr.new      # TypeError
a, b, c = MyAry.new  # a=<MyAry obj>, b=nil, c=nil   (treated as single value)

Identity versions of to_int, to_str, to_sym, and to_hash remain defined on the corresponding built-in types so that respond_to?(:to_str)-style checks work for built-in instances. Float#to_int and Array#to_ary are intentionally not defined.

Explicit conversion methods (to_i, to_s, to_a) work as in CRuby and are called by features such as string interpolation and the splat operator (*obj).

This is a deliberate trade-off: implicit conversion forces every coercion site to go through method dispatch and can silently mask type-mismatch bugs.

Nested def in Singleton-Method Context

def written inside a singleton method (def self.foo) is placed on a different class in mruby than in CRuby. CRuby registers the inner method as an instance method of the lexical enclosing class. mruby registers it as a method of the enclosing receiver's singleton class, which makes it visible as a class method of the enclosing class.

class SomeClass
  def self.class_method
    def nested; 'nested!'; end
  end
end
SomeClass.class_method

CRuby

SomeClass.nested        # NoMethodError
SomeClass.new.nested    # => "nested!"   (instance method)

mruby

SomeClass.nested        # => "nested!"   (class method)
SomeClass.new.nested    # NoMethodError

Writing nested def like this is unusual; this difference rarely surfaces in practical code.

Proc#dup / Proc#clone is Always Orphan

A dup or clone of a block given to a method is always treated as an orphan block in mruby — calling it raises LocalJumpError if the block contains break or return. CRuby is finer-grained: the copy inherits the orphan status of its original, so the copy only becomes orphan once the original yielding method returns.

def m(&b)
  b.dup
end

x = m { break 1 }
x.call

CRuby

LocalJumpError  # raised only after m returns; if called inside m,
                # the dup is still a live block

mruby

LocalJumpError  # always raised — the dup is orphan from the moment
                # it is created

mruby's stricter rule keeps RProc from needing a back-pointer to the original block (which would also enlarge the GC mark set).

Class#initialize Can Be Re-Invoked

CRuby raises TypeError: already initialized class when initialize is invoked on a class that has already been set up. mruby's Class#initialize has no such guard — invoking it on an existing class through __send__, send, or UnboundMethod#bind_call silently succeeds. The superclass argument is ignored in this case, so the call cannot rewrite the class hierarchy; only the block (if any) is evaluated with the class as receiver.

Klass = Class.new
Klass.__send__(:initialize) {}

CRuby

TypeError: already initialized class

mruby

The block is evaluated in the context of Klass; no error is raised.
The superclass is not changed even when one is passed as an argument.

Module#initialize is re-callable in both implementations, so this divergence is Class-specific. Adding the CRuby check would require an additional flag bit on every RClass; mruby leaves the bit unspent because no destructive side effects are possible through this path.