We came to the conclusion today that the world would be a little bit better if anyone who suggested using reference counting instead of a real garbage collector could receive a small non-lethal electric shock.
You can implement a real GC in only 600 lines of code or so. It’s not that hard people.
You can in a sane language, but if you program in a sucky language like C++ you can’t.
Reference counting is certainly slower, and has worse locality of reference than “real” GC. However, it does have the virtue of releasing objects as soon as possible (when combined with “real” GC to collect cycles). This is a good fit for interpreted languages (like C python) where the slowness is tiny compared with the interpreter overhead.
Except when it doesn’t release objects because of circular references. This was indeed our problem yesterday — combining closures and reference counting [in Perl] makes it virtually impossible to avoid cycles.
There was a second part to this electic shock theory: people suggesting interpreters instead of writing a good compiler. With LLVM (or even without) there is really no good excuse any more to be writing slow interpreters.
Seconded to this. Spend an two day a while ago, bisecting the code to track down a database leak handle. A recursive closure ftw.