As part of investigating #7106 we ran into a new issue where a call to one of the registry functions in registry.rb produced the following trace:

>bin\jruby -e "require 'resolv'"
ArgumentError: cannot convert parameter to native pointer
         OpenKey at C:/Users/User/jruby/lib/ruby/stdlib/win32/registry.rb:289
            open at C:/Users/User/jruby/lib/ruby/stdlib/win32/registry.rb:431
            open at C:/Users/User/jruby/lib/ruby/stdlib/win32/registry.rb:542
   get_hosts_dir at C:/Users/User/jruby/lib/ruby/stdlib/win32/resolv.rb:142
  get_hosts_path at C:/Users/User/jruby/lib/ruby/stdlib/win32/resolv.rb:15
   <class:Hosts> at C:/Users/User/jruby/lib/ruby/stdlib/resolv.rb:176
  <class:Resolv> at C:/Users/User/jruby/lib/ruby/stdlib/resolv.rb:172
          <main> at C:/Users/User/jruby/lib/ruby/stdlib/resolv.rb:38
         require at org/jruby/RubyKernel.java:1018
         require at C:/Users/User/jruby/lib/ruby/stdlib/rubygems/core_ext/kernel_require.rb:85
          <main> at -e:1

The issue here appears to stem from the fact that our Fiddle implementation is just a thin shim over FFI, and while the latter requires more exact parameter types (e.g. Pointer instances for :pointer args), the former appears to either be looser about these requirements or better at making conversions.

This was not observed in the past due to registry.rb using our FFI-based win32api. When it was upgraded to use Fiddle as part of 3.0/3.1 updates in JRuby 9.4, the breakage was exposed.

This is not the first problem of its type in Fiddle, and the best fix would be to reimplement fiddle, either as a standalone library atop our Java FFI layer JNR, or as a better wrapper around Ruby FFI.

I will file a separate issue for the registry.rb issues, which keep networking from functioning properly on Windows.