Issue #14541 has been updated by jeremyevans0 (Jeremy Evans).

Assignee set to matz (Yukihiro Matsumoto)
Status changed from Closed to Assigned

jeremyevans (Jeremy Evans) wrote in #note-17:
> Applied in changeset commit:git|900e83b50115afda3f79712310e4cb95e4508972.
> 
> ----------
> Turn class variable warnings into exceptions

@alanwu pointed out this this commit only raised for class variable overtaken in verbose mode, because previously the warning was only issued in verbose mode.  I've added a pull request to fix this and raise in all cases ( https://github.com/ruby/ruby/pull/3210), but are we OK jumping directly from verbose-mode warning in 2.7 to RuntimeError in 3.0, without issuing a non-verbose warning?

Note that the code that previously issued the warning and now raises the error is when accessing a class variable that is defined in both a superclass and an included/prepended module.

----------------------------------------
Bug #14541: Class variables have broken semantics, let's fix them
https://bugs.ruby-lang.org/issues/14541#change-86083

* Author: Eregon (Benoit Daloze)
* Status: Assigned
* Priority: Normal
* Assignee: matz (Yukihiro Matsumoto)
* ruby -v: ruby 2.6.0dev (2018-01-29 trunk 62091) [x86_64-linux]
* Backport: 2.3: UNKNOWN, 2.4: UNKNOWN, 2.5: UNKNOWN
----------------------------------------
Class variables have the weird semantics of being tied to the class hierarchy and being inherited between classes.
I think this is counter-intuitive, dangerous and basically nobody expects this behavior.

To illustrate that, we can break the tmpdir stdlib by defining a top-level class variable:

    $ ruby -rtmpdir -e '$SAFE=1; @@systmpdir=42; p Dir.mktmpdir {}'
    -e:1: warning: class variable access from toplevel
    Traceback (most recent call last):
    	3: from -e:1:in `<main>'
    	2: from /home/eregon/prefix/ruby-trunk/lib/ruby/2.6.0/tmpdir.rb:86:in `mktmpdir'
    	1: from /home/eregon/prefix/ruby-trunk/lib/ruby/2.6.0/tmpdir.rb:125:in `create'
    /home/eregon/prefix/ruby-trunk/lib/ruby/2.6.0/tmpdir.rb:125:in `join': no implicit conversion of Integer into String (TypeError)

Or even simpler in RubyGems:

    $ ruby -e '@@all=42; p Gem.ruby_version'
    -e:1: warning: class variable access from toplevel
    Traceback (most recent call last):
    	3: from -e:1:in `<main>'
    	2: from /home/eregon/prefix/ruby-trunk/lib/ruby/2.6.0/rubygems.rb:984:in `ruby_version'
    	1: from /home/eregon/prefix/ruby-trunk/lib/ruby/2.6.0/rubygems/version.rb:199:in `new'
    /home/eregon/prefix/ruby-trunk/lib/ruby/2.6.0/rubygems/version.rb:199:in `[]': no implicit conversion of String into Integer (TypeError)

So defining a class variable on Object removes class variables in all classes inheriting from Object.
Maybe @@systmpdir is not so prone to conflict, but how about @@identifier, @@context, @@locales, @@sequence, @@all, etc which are class variables of the standard library?

Moreover, class variables are extremely complex to implement correctly and very difficult to optimize due to the complex semantics.
In fact, none of JRuby, TruffleRuby, Rubinius and MRuby implement the "setting a class var on Object removes class vars in subclasses".
It seems all implementations but MRI print :foo twice here (instead of :foo :toplevel for MRI):

~~~ ruby
class Foo
  @@cvar = :foo
  def self.read
    @@cvar
  end
end

p Foo.read
@@cvar = :toplevel
p Foo.read
~~~


Is there any library actually taking advantage that class variables are inherited between classes? I would guess not or very few.
Therefore, I propose to give class variable intuitive semantics: no inheritance, they behave just like variables of that specific class, much like class-level instance variables (but separate for compatibility).

Another option is to remove them completely, but that's likely too hard for compatibility.

Thoughts?



-- 
https://bugs.ruby-lang.org/

Unsubscribe: <mailto:ruby-core-request / ruby-lang.org?subject=unsubscribe>
<http://lists.ruby-lang.org/cgi-bin/mailman/options/ruby-core>