On Wed, Aug 20, 2008 at 01:22:33PM -0400, Will Coleda wrote: > On Wed, Aug 20, 2008 at 1:15 PM, chromatic <[EMAIL PROTECTED]> wrote:
> > The overall concepts of the incremental GC are solid, but a couple of > > details > > of the implementation need polishing. It's difficult to debug these types > > of > > problems, and even more difficult to estimate them. (In particular, > > interleaving GC headers and GC'd elements leads to some troublesome offset > > manipulation.) > You can always try to identify the chunks ex post facto and start the > merge back a chunk at a time; not as easy as identifying the bits > ahead of time, but doable. > If it's *close* (and mostly passing tests) we can always throw it back > into trunk immediately after a monthly release and give ourselves 4 > weeks to clean it up. Right now TGE fails to build, because the Integer PMCs stored in the interpreter class type registry get collected inappropriately. That's on GNU/Linux on x86, which is as forgiving as a platform gets. There are likely platform-specific bugs on PPC and Sparc and 64-bit platforms, not to mention with compilers other than GCC. Tracking down bugs and crashes to a specific checkin will be difficult. If we can figure out the class registry bug and get tests to pass reliably, we might be able to get more platform testing and consider a merge back. As it is now, it's riskier than I like. I don't want to block Rakudo, at least for more than overnight. -- c
