Re: Best Ref-counting algorithms?

Christoffer Lernv <lerno@dragonascendant.com> writes:


> If one would like to implement a fast GC using refcounting, should I
> implement something like in those papers, or are there better
> algorithms out there?


If you want to implement a fast GC, you should not use refcounting. The
main problem with refcounting is that an assignment to a pointer
variable (which would otherwise just be a register-to-register transfer)
requires two memory reads, two memory writes and a conditional jump, so
it stresses the memory system far more than most other GC algorithms.


Refcounts should, IMO, only be used in the following circumstances:


  - Operations on heap pointers are very rare, so even a slow GC doesn't
      matter in the overall picture.


  - You need hard real-time guarantees.


  - The heap is shared between processes/processors with no central
      control.


So refcounts are fine for such things as shared libraries (you don't
very often add or delete references to these), but for a functional or
OO language where you have many small heap structures with short
lifetimes, it will give massive overhead.


Torben

Return to the comp.compilers page.
Search the comp.compilers archives again.