On 12/21/2016 6:50 AM, thedeemon wrote:
Have you seen this one? http://www.infognition.com/blog/2014/the_real_problem_with_gc_in_d.html
Although I had called them write gates, write barriers are the same thing. Yes, that's the problem with implementing a generational collector in D.
I once tried to implement write barriers by using the hardware VM system. I'd mark the old generation pages as read-only. When the program would write to those pages, a seg fault happened. This would then run a handler in the GC code which would mark that page as "dirty", then write-enable the page, and restart the program at the point where it seg faulted.
This worked great. The only trouble was that the seg faulting path at runtime was so slow it ruined the speed advantage of not have write barriers. So I had to abandon it.
But that was long ago. Maybe the tradeoff is better these days with modern hardware. But I suspect that if other GC developers are not using this technique, it is still too slow.
