Rethinking rendering

(Moderator: jchaffraix, Scribe: dpranke)

How can we get more people hacking on rendering?

• accelerated compositing is tied to a layer (a very old concept)
  • can we make layers better or faster?

• can we talk about security and the impact of specific design decisions in the render tree (inferno, bethdakin?)

• jchaffraix: from study tables, thinks layers are the wrong abstraction
  • as a general rule, the render object knows what its own structure should be
    • because layers are generic, they don't know and have to do more work
    • they do too much
      • repainting
      • clipping
      • compositing
      • stacking
      • scrolling

• eseidel: layers are designed to allow the rendering tree to be lightweight
  • layers exist to handle the more "advanced" features of rendering
  • smfr: layers designed to solve problems like scrolling and clipping that affect a set of render objects in one go

• eseidel: what are the problems we're trying to solve?
  • (??) need render layers to support z-index (for svg 2.0 and z-stacking not in document order)
  • jamesr: we want to composited into svg
  • enne: rendering / layers are hard to test
  • dave barton: mathml is not normal inline or block - math layout is bottom-up from the children of the node, not top-down left-to-right on the page
  • jamesr / jchaffraix: tables are very similar
  • eseidel: if you call setNeedsLayout(False), you should be able to do whatever you want inside your own elements
    • jchaffraix: this means we don't reuse other code
    • eseidel: there are other layout models other than the renderboxmodel
    • eseidel: also there's computePreferredWidths()
    • dave barton: this is kinda what he does in mathml
    • jamesr: maybe mathml shouldn't be deriving from renderboxmodelobject?
    • dave barton: but we do use a lot of it - maybe this is more a question about writing your own render object and should be in the other talk?
  • jchaffraix: maybe you shouldn't inherit from RenderBlock if you aren't actually a "block" in the CSS box model sense

• eseidel: a render layer defines its own coordinate space (and a z-order) - so when you need your own coordinates (e.g., abs positioning), you need a layer

Security Issues w/ the current architecture

• jchaffraix: if the render tree gets confused, the code does bad things with pointers
• jchaffraix: potentially layout should never modify the render tree (during layout)
  • example: don't destroy / reparent things during run-ins
  • inferno: e.g., during flexbox layout, this can leave other objects with dangling pointers to stale objects
• rniwa: how do you handle continuations (generated content, bidirectional text)?
  • during style recalc?

• eseidel: the render tree is really just a big cache, and maybe that's the problem
  • we hope that nothing is stale at that point
  • layout is supposed to make things not be stale

• jchaffraix: are continuations strictly defined by the dom tree?
  • inferno: they shouldn't be (aren't) created during layout (?)
  • eseidel: that is not the invariant today - blocks are created and destroyed during layout
  • jamesr: that only happens during style recalc
  • eseidel: I am misinformed :)

• rniwa: why do counters need to compute the counter during layout (and not style calc)
  • jchaffraix: dunno, need to redesign counters? possible could be done during pre layout

• render trees use raw pointers are over the place
  • could we do something like weak ptrs instead so that collection is well defined (as long as it doesn't impact performance)
  • what happened to the experiment of holding references to dom nodes?
    • jamesr: 5% perf impact (follow up w/ ojan?)
    • rniwa: you could hold on to a lot of nodes until script invocation ends
      • rniwa: needs a lot more investigation

• rniwa: jamesr, are you adding assertions / invariants to the code?
  • jamesr: did some, but there were too many exceptions to be able to turn things on
  • jamesr: whenever an invariant failed, it always lead to a security bug, it seemed
  • jamesr: we need to document what the invariants should be (e.g., anonymous objects shouldn't have a layer)
  • jchaffraix: perhaps this is part of "create a new object"

• jchaffraix: the whole problem of the render tree is corner cases - you always forget one (or 10%)
  • rniwa: e.g., people always forget editing / designMode
• jchaffraix: responsibility is also spread out, e.g. overflow
• smfr: a lot of this has been done for optimization
• jchaffraix: should we be more strict about who is allowed to know about who?
  • smfr: probably, but we're very performance-sensitive
  • eseidel: we might be too sensitive here

Repainting

• jchaffraix: who knows how repainting actually works? (laughter in the room)
  • there are lots of ways that it can get triggered
  • what would a good architecture be for a general way of repainting?
  • smfr: it is definitely confusing and buggy - compositing made things worse - but it's not clear what better ideas are
  • jamesr: it's also very hard to construct good tests - maybe we need better hooks for testing?
    • paint everything, make some changes that hopefully cause invalidation, paint again, and hope that the PNG displays something useful
    • we don't necessarily just take the union of invalid rects
    • maybe you can use layoutTestController.display() - see a list of rectangles that got sent out, rather than the end result
  • jamesr: scrolling also complicates things (don't necessarily get the blit at all)

Splitting up RenderLayer?

• jchaffraix: first, hack off some of the bits to other things
• then split up the core object
• ex: overflow: hidden
  • just needs to hold the clipped rect?
  • doesn't need scroll bars or a full layer
  • (smfr: of course, you can still scroll via js)
  • smfr: we should fix the cost of updateLayerPositions - we don't have a dirty bit
  • jchaffraix: last time he tried to figure out what was dirty, it was hard
• ??: can we do both at the same time - create a lightweight render layer *and* split out / abstract things better

• eseidel: do most render layers have graphics layers?
  • smfr / jamesr: no - only a few can / do

• eseidel: there's at least two roles:
  • i have rare (uncommon) data
  • i have something that needs to be in the compositing stack
  • smfr: maybe it's more something that modifies hit testing (e.g., filters)
  • smfr: maybe the responsibility for doing painting should be a separate class, hit testing should be a separate class, stacking should be a separate class?

• eseidel: maybe (Following on the regions discussion) break things out more things into an optional hasA relationship?
  • smfr: possibly fix parts of SVG so that they can use these things (e.g., foreign objects)
  • jamesr: you need painting, hit testing
  • jchaffraix: maybe z-ordering is optional (?)
  • smfr: certainly some layers need z-ordering
  • smfr: it's not clear that you wouldn't recreate the same algorithm (go layer at a time and then paint every object in the layer)

• eseidel: some things are very specific to the CSS box model (the 7 steps of rendering a box) but maybe some things can be reused more generally (e.g., z-ordering)
  • smfr: things may not be so separable, but maybe some things can be optional

• (??): perhaps the confusion is over mixing functionality and ownership - there's only one tree, but we split the functionality over multiple classes
  • maybe you could have one tree with different attributes instead of different subtrees (?) - this is a different way to look at the problem but hasn't really been investigated
  • smfr: you might end up re-creating trees just so you can keep track of what's going on during traversal
    • (??): maybe they would at least be homogenous trees in that case

• (??): maybe render layers shouldn't be boxes, because they don't need all of the aspects of a box