MRI desn't allow defining singleton methods for Floats and Bignums

Patch with test cases. This patch highlights a separate issue:

class << 10
end

should fail with "TypeError: no virtual class for Fixnum" (according to MRI) but instead succeeds. I will raise a separate patch.

Michael, the patch seems mangled and when I apply it testing loops eternally.

Patch applied (commit 2662). Thanks Marcin, Michael

The patch applied does not allow for singleton_method_added being redefined in Numeric to let singleton methods through. Note Marcin's recent example, which should run successfully:

class Numeric
      def singleton_method_added name
          puts "really ?"
      end
end

class << 1.0
     def some_method
     end
end

The behavior appears to be:

• singleton_method_added is called before the method is actually added
• Numeric provides a default impl that throws the TypeError
• redefining singleton_method_added in Numeric can eliminate the TypeError and allow singleton methods through

This behavior could also apply for method_added; I will attach a patch that addresses both.

HOWEVER...this behavior (specifically the order-of-operations regarding the _added methods) must first be confirmed with more test cases or by showing examples of C Ruby source that prove it without a doubt.

numeric_singleton_added.patch modifies the addition of singleton methods to invoke singleton_method_added first, and provides an implementation for Numeric that throws the TypeError referenced by this bug. This resolves Marcin's additional test case.

singleton_method_added is called after the method is actually added and it doesn't change the fact that the above example will run successfully. I'd like to emphasise that recent patches put the "can't define singleton method" TypeError at NodeTypes.DEFNNODE in EvaluationState.evalInternal() which is incorrect.

It makes my mind crazy but it seams that the code:

ruby_frame = ruby_frame->prev; /* pop frame for "singleton_method_added" */
/* Numerics should be values; singleton_methods should not be added to them */

in MRI Numeric#singleton_method_added standard implementation removes the method just after the method has been added

Recording findings for posterity...

The actual behavior for singleton methods on Numerics appears to be:

singleton_method_added is called before the method is actually added
Numeric provides a default impl that throws the TypeError
Redefining singleton_method_added in Numeric can eliminate the TypeError but regardless, the method will still be added to the singleton class

It initially appeared that exception-throwing singleton_method_added methods would cause removal of the associated method. This was not the case - what happened was because any two numeric literals are not the same object then adding singleton methods in the manner below only applies to each instance's singleton class.

E.g

class << 1.0 #Float instance number 1
def hello #will cause a type error to be raised but hello will still be defined
end
end

1.0.hello #Float instance number 2 will fail with NoMethodError

If you rewrite this like follows the true behaviour becomes apparent:

i = 1.0 #One and only Float instance
class << i
def hello #will cause a type error to be raised but hello will still be defined
end
end

i.hello #will succeed

Fixnums and Symbols are still special cases. Singleton methods cannot be used with them and MRI's handles this in the part of the interpreter handling NODE_DEFS .

I'll add this to Rubyspec BTW

The patch removes singletonMethodsAllowed() from IRubyObject and further descendants (RubyObject, RubySymbol, RubyFixnum) - it's not needed (and it doesn't allow to change this behaviour) since proper default Numeric#singleton_method_added has been added with the Numerics compatibility fix
(error message in singleton_method_added and initialize_copy has been slightly changed as well):

EvaluationState is also changed to reflect MRI at NODE_DEFS:

if (FIXNUM_P(recv) || SYMBOL_P(recv)) { rb_raise(rb_eTypeError, "can't define singleton method \"%s\" for %s", rb_id2name(node->nd_mid), rb_obj_classname(recv)); }

in JRuby (at DEFSNODE):

if (receiver.getMetaClass() == runtime.getFixnum() || receiver.getMetaClass() == runtime.getClass("Symbol")) { throw runtime.newTypeError("can't define singleton method \"" + iVisited.getName() + "\" for " + receiver.getType()); }

YARVMachine has been also updated in this manner

now this will succeed:

class Numeric
def singleton_method_added a # turn off the default behaviour
end
end

a = 1.0

class << a

def hello
p "hello"
end

end

a.hello

So is this patch valid? complete? ready? Where do we stand, gents?

JIRA-399.patch is still valid (although it can have some conflicts to resolve now)

the guards cannot be hardcoded in EvaluationState and singletonMethodsAllowed should go away

Fixed in commit 3105. Thanks Marcin