> On May 5, 2015, at 11:55 AM, Tab Atkins Jr. <jackalm...@gmail.com> wrote: > > On Tue, May 5, 2015 at 11:20 AM, Ryosuke Niwa <rn...@apple.com> wrote: >>> On May 4, 2015, at 10:20 PM, Anne van Kesteren <ann...@annevk.nl> wrote: >>> >>>> On Tue, May 5, 2015 at 6:58 AM, Elliott Sprehn <espr...@chromium.org> >>>> wrote: >>>> We can solve this >>>> problem by running the distribution code in a separate scripting context >>>> with a restricted (distribution specific) API as is being discussed for >>>> other extension points in the platform. >>> >>> That seems like a lot of added complexity, but yeah, that would be an >>> option I suppose. Dimitri added something like this to the imperative >>> API proposal page a couple of days ago. >>> >>> >>>> One thing to consider here is that we very much consider distribution a >>>> style concept. It's about computing who you inherit style from and where >>>> you >>>> should be in the box tree. It just so happens it's also leveraged in event >>>> dispatch too (like pointer-events). It happens asynchronously from DOM >>>> mutation as needed just like style and reflow though. >>> >>> I don't really see it that way. The render tree is still computed from >>> the composed tree. The composed tree is still a DOM tree, just >>> composed from various other trees. In the "open" case you can access >>> it synchronously through various APIs (e.g. >>> if we keep that for >>> querySelector() selectors and also deepPath). >> >> I agree. I don't see any reason node distribution should be considered as a >> style concept. It's a DOM concept. There is no CSS involved here. > > Yes there is. As Elliot stated in the elided parts of his quoted > response above, most of the places where we update distribution are > for CSS or related concerns: > > # 3 event related > # 3 shadow dom JS api
These two are nothing to do with styles or CSS. >> I have issues with the argument that we should do it lazily. On one hand, >> if node distribution is so expensive that we need to do it lazily, then it's >> unacceptable to make event dispatching so much slower. On the other hand, >> if node distribution is fast, as it should be, then there is no reason we >> need to do it lazily. >> >> The problem is really the redistributions. If we instead had explicit >> insertion points under each shadow host, then we wouldn't really need >> redistributions at all, and node distribution can happen in O(1) per child >> change. > > As repeatedly stated, redistribution appears to be a necessity for > composition to work in all but the most trivial cases. Where? I have not yet to see a use case for which selective redistribution of nodes (i.e. redistributing only a non-empty strict subset of nodes from an insertion point) are required. - R. Niwa
