[ruby-core:40018] Re: 2.0 feature questionnaire

(10/07/2011 02:19 PM), Evan Phoenix wrote:
>> (10/02/2011 05:34 PM), Jeremy Kemper wrote:
>>>> |3. MVM, with inter-vm message passing.
>>>>
>>>> I have not decided yet, but since MVM requires incompatible changes to
>>>> C API, so we might pend it to Ruby 3.0.
>>>
>>> I understand @shyouhei's branch is backward compatible for extensions
>>> that don't use MVM features.
>>
> 
> I did quite of work into getting MVM working on Rubinius and ended up
> stopping development because I can't see how you can rectify it with how
> most C extensions are written.
> 
> Specifically, most extensions use static VALUEs and IDs variables. To make
> matters worse, extensions commonly store VALUE's for Modules and Classes in
> static variables without calling rb_global_variable() (This is itself a
> latent bug that almost no one hits because how the MRI GC works).
> 
> Any C extension that uses any static VALUE or ID will crash the whole
> process very quickly, meaning that most existing C extensions are not
> compatible with MVM UNLESS they are recoded to remove static VALUEs and IDs.
> 
> I haven't seen @shoyuhei's branch in a while, how is this issue dealt
> with?

No, it isn't.  MVM-aware extensions shall obey the MVM-safe APIs.

But MVM-unaware extensions are OK if and only if the whole ruby
process don't use MVM features at all.  And even if you don't
touch MVM things, binary compatibility is broken.  Recompilation
is mandatory for everything.