On Sunday 24 February 2008 07:33:30 Leopold Toetsch wrote: > Am Sonntag, 24. Februar 2008 10:55 schrieb chromatic:
> > PMCs that *do* need a special mark() are troublesome; they may contain > > pointers to non-constant PObjs that *do* need live marking, lest they get > > swept away during the second half of GC. If these constant PObjs don't > > get marked, there's a problem. > > Constant PMCs that point to non-constant items are IMHO just a bug. The > problem is that when their live bit is set, it isn't cleared, because - as > you have mentioned too - the constant PMC pool isn't swept. Therefore they > don't propagate their liveness. Ergo: all children of constant PMCs have to > be constructed as constants too. That conclusion is impractical; we don't know when we're creating a PMC that it will eventually have a constant parent. However, we *could* check for constantness in pobject_lives and call mark() when appropriate. That's a cheap and easy check. If that doesn't work, we may not be calling pobject_lives() on all const PMCs, which is also a problem. (I suspect we may not be, but I'm willing to be surprised.) > > Allison suggested "Hey, why not register all constant PMCs when you grab > > a new constant PMC header?" and that's an easy patch to write. > > option is to walk through the constant PMC pool at the start of the mark > > phase to look for PMCs with the mark flag set and call those directly. > > (That could simplify some other operations; this could be the root set > > itself, but that's a mere detail.) > All these workarounds render the advantage of constant items towards zero. > You get that workaround much cheaper by ignoring the constant flag during > object construction and never use constant pools. That might wise, from the standpoint of simplicity. > > Another concern to me is that I can't convince myself that the "HEY I AM > > A CONSTANT" flag gets stored in PBC when we freeze such a PMC or gets > > restored when we thaw that PMC. > While not propagating the constness during freeze/thaw is undesirable, it > wouldn't be a problem, if the PMCs have proper mark vtables. ... if they're all reachable appropriately. I'll try the idea you made me think of. -- c
