On Fri, 24 Aug 2012 08:27:09 +0200, Jacob Carlborg <[email protected]> wrote:
On 2012-08-23 16:38, Alex Rønne Petersen wrote:
Yes, but parallelization of the mark phase is fairly trivial, and
something we should probably look into.
The GC will probably always be STW unless we get compiler support for
inserting GC barriers.
Would a thread local GC be possible, and desirable? To my understanding,
which is not much, that means the GC will only run in one thread (or
multiple) and only needs to stop that/those thread(s). That also means
it only need to search for dead objects in the heap/storage area for
that particular thread (and where these point to).
If I understand this correctly this would be perfect for D since
everything is thread local by default.
There's also a global heap for global objects or objects shared between
threads.
It certainly would be possible. But I believe it requires thread-local
heaps, and some way of keeping track of when an object changes owning
thread.
Perhaps a mark phase could run on each thread's heap, and unreferenced
objects be moved to a global list of potentially dead objects. If after
all threads have run a mark, none have claimed the objects, they're
collected.
Then again, I'm hardly a GC architect.
--
Simen
