instance_variable_set is a silent no-op on File/StringIO/Dir/IO::Buffer/Array/Hash/String/Proc (ivarTable returns nil) #672

What

instance_variable_set on most of rbgo's natively-backed objects is a silent no-op. It does not raise; the write is discarded and the next read answers nil.

f = File.open("x", "r")
f.instance_variable_set(:@zz, 7)
p f.instance_variable_get(:@zz)   # MRI: 7      rbgo: nil

Measured across the natively-backed kinds (instance_variable_set(:@zz, 7) then get):

ivarTable (internal/vm/object_model.go) enumerates the kinds that have ivar storage — RObject, RClass, object.Main, SinatraCtx, ActionViewBase, and the boxed method values — and returns nil for everything else. Every caller then has nowhere to write and, apparently, no complaint.

The silence is the dangerous part: attr_accessor on a singleton class of such an object appears to work — the reader and writer are defined and callable — and simply never stores anything.

Measured cost

Two examples in core/file/open_spec.rb, via core/file/fixtures/common.rb's FileSpecs.make_closer:

core/file/open_spec.rb    rbgo = 78    MRI 4.0.5 = 82

def self.make_closer(obj, exc=nil)
  class << obj
    attr_accessor :close_exception
    alias_method :original_close, :close
    def close
      original_close
      ScratchPad << :file_closed
      raise @close_exception if @close_exception   # @close_exception is always nil in rbgo
    end
  end
  obj.close_exception = exc                       # discarded
end

So "propagates StandardErrors produced by close" and "propagates non-StandardErrors produced by close" both see no exception. The File.open ensure machinery is correct — ioYieldAndClose already re-raises a #close exception and filters the IOError "closed stream" case, and it is witnessed working when the override raises unconditionally:

File.open(p, "r") do |f|
  class << f
    alias_method :oc, :close
    def close; oc; raise StandardError, "boom"; end   # this DOES propagate in rbgo
  end
end

The defect is only that the fixture's exception never gets stored.

Given attr_accessor on a singleton is a common spec-fixture idiom, and String/Array/Hash/Proc are on the list too, this very likely costs more than these two examples across the suite.

Suggested shape

Give the natively-backed values an ivar table the way boxed already does — a small embedded struct with a lazily-created map — rather than extending the type switch one kind at a time. Time and the boxed method values already show the pattern.

Meanwhile, a caller that cannot find a table should probably not stay silent.

Where

internal/vm/object_model.go (ivarTable) and internal/object. Both outside the file set of #668, which is why this is filed rather than fixed there.