[gclist] ref-counting performance cost

Daniel Wang danwang@CS.Princeton.EDU
11 Oct 2000 18:12:00 -0400

Andrew Chen <chenandr@cse.msu.edu> writes:
{stuff deleted}
> "correctly implement" implies two possibilities:
> 	provably correct (use of a formal automated proof system)
> 	probably correct (use of someone else's encoded knowledge)
> You seem to be advocating the latter, correct? (The former I flirted 
> with but ultimately have decided to abandon. One of my former 
> advisors was into formal methods - one of the reasons that person is 
> a _former_ advisor.)

I'm advocating both...  

A strong guarantee of partial correctness is also good... i.e. the scheme
doesn't free and live data, but may not actually free any
data..... debugging a space leak in your scheme is a bit more preferable to
debugging the reuse of space that shouldn't be reused... 

In any case, I think (and sure hope) the research on GC/memory management
hasn't devolved into a pure numbers games... The work on region based type
systems is quite cool and will in the long run be yet one more useful memory
managment technique....