Re: New GC concept

[Daniel Mendler <mail@daniel-mendler.de>]

Thank you for answering my questions. Did you also consider the approach
of using a non-copying collector, keeping the classical mark and sweep?
Andrea mentioned in his mail that the focus should be on reactivity.
With mark and sweep it is possible to avoid the copying step, which
scales with the live size, such that one can achieve constant pause times.

On the other hand the copying step is probably quick for the expected
Emacs heap sizes. Furthermore with m&s, you have the fragmentation
problem, it is harder to use such a bump-style allocator and it is
harder to separate the generations, which is a requirement for the
hardware write barrier. So I think overall your design is a sound
approach as long as the heap stays at a reasonable size.

Your approach seems to be quite general purpose and does not require
intrusive changes to the Emacs C code; it seems to be relatively
decoupled from the Elisp runtime. Do you think it is realistic to
implement the GC as a "library" behind some abstract interface? Of
course there is some dependence on the object memory layout, but the GC
interface could offer APIs to request objects of different types. It
should then be possible to use different GCs which use a similar
approach (conservative stack scanning, no explicit read/write barriers).

Daniel Mendler