Lots of extra copies of objspace, potentially refactor later to use a st_table <objspace*, mark_queue_node_t*> and remove that pointer from mark_queue_node_t to save space?

We need to think about refactoring this - a tiny malloc on every node overshadows any possible gains from parallelism

I agree--I think the best approach is probably to use a circular buffer to implement
fixed size queues (similar to what the current implementation does for the mark stack.)

On Nov 8, 2012, at 12:25 PM, Kevin Sweeney notifications@github.com wrote:

In gc.c:

@@ -1650,6 +1670,45 @@ struct mark_tbl_arg {
}

void
+gc_mark_defer(rb_objspace_t *objspace, VALUE ptr, int lev) {

• mark_queue_node_t* node =
• (mark_queue_node_t*) malloc(sizeof(mark_queue_node_t));
  We need to think about refactoring this - a tiny malloc on every node overshadows any possible gains from parallelism

—
Reply to this email directly or view it on GitHub.