i think that might actually agree even more with the optimization we were talking about where GC runs in a fork-exec'd process