2 Mar
2010
2 Mar
'10
9:11 a.m.
Both concurrent GC and incremental GC tend to add overheads to the mutator, because they need a read barrier. There was an incremental GC for GHC once [1], taking advantage of the built-in read barrier that we have whereby most closures are "entered"
Was there a specific reason why that GC implementation chose to use a read barrier rather than a write barrier? I would have thought that in general, a write barrier is cheaper to implement. Doesn't ghc update fewer thunks than it enters? Regards, Malcolm