On Tue, Dec 23, 2014 at 1:51 PM, L. David Baron <dba...@dbaron.org> wrote:
> On Tuesday 2014-12-23 13:14 -0800, Eric Rescorla wrote: > > Just to be clear, is your problem the implicit conversion itself > > or the reference count increment/decrement? > > The latter -- the problem is that there's an implicit conversion > that has surprising side-effects. > > (It sounds from this discussion like you've written a bunch of code > that you wrote the way you did to avoid excess reference-counting, > yet there may have been a bunch of excess reference-counting that > you didn't know about.) I could look, but I don't think so, for two reasons: - I try to use UniquePtr<T> when I can, and for obvious reasons, UniquePtr<T> doesn't do implicit conversion. - I generally don't do a lot of passing around of SmartPtr<Base>, just because the type of code I write tends not to have a lot of this kind of inheritance. > On the specific topic of what users expect, what you expect may not > > be what I expect, but I would observe that both C++11 shared_ptr > > and Boost intrusive_ptr (the analog to our RefPtrs> automatically > > convert from SmartPtr<Derived> to SmartPtr<Base> and appear to > > increment the reference count in the same way that RefPtr does; > > this seems to suggest that at least some people do expect it to > > behave this way. > > Ugh. > > Then again, this wouldn't be the first thing in the standard library > that didn't make sense for users who care about performance. This may be a much longer argument, but I'm not convinced that sacrificing what would otherwise be good programming practice (never unboxing your pointers) at the altar of performance is a good idea. I'd note that you only pay this penalty when you pass an object of type RefPtr<Derived> to something that takes RefPtr<Base>, and depending on your code, this may happen infrequently or at least only happen a lot of times at a fairly small number of call sites, which can then be optimized in the usual fashion by changing to a raw pointer call. -Ekr _______________________________________________ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
