forked from servo/servo
-
Notifications
You must be signed in to change notification settings - Fork 0
Minutes Hackathon 2012 03 27
nikomatsakis edited this page Mar 28, 2012
·
1 revision
Here are some notes from our Hackathon on 2012-03-07.
- What kind of data structure do we use?
- RCU
- clean/dirty, has a rd/wr ptr
- when making dirty, copy over and adjust wr ptr
- keep a linked list of dirty nodes
- when you join layout, walk over and reset reader ptrs
- However:
- Refresh driver, particularly around animations, can async trigger layout
- What do we do when JS is running and the timer goes off?
- is it enough to just set the 'layout is running' flag?
- have to be sure that it's always safe at any time
- Danger points:
- about to join
- JS is about to write:
- it needs to know if it can make a copy
- much simpler:
- make use of the operation callback to stop JS
- operation callback every .5ms to 1ms seems sufficient
- an alternative:
- on every dom mutation grab a spin lock
- try to get smart with compare-and-swap? but that's just a spin lock
- How many pipeline stages for the layout?
- Clear stages:
- Computing styles
- Create Layout
- Layout Layout
- Create paint
- Paint (rasterize / composite)
- When we want to join the layout, what do we need:
- sometimes the DOM wants style information, but not layout
- we probably want to have
- In general, layout is probably two stage:
- computing class information
- computing geometry
- sometimes we just join the class info
- sometimes we tell layout to stop after class info
- Clear stages:
- create paint:
- needs to run in "the layout task"
- display list
- contains display items: things that paint themselves
- once we have the display list, do we still need the DOM?
- bz says ... "mostly done with the DOM"
- do need DOM to some extent during DL construction because you have to sort things by Z order
- One "obvious" exception
- Canvas: it has a bitmap backing store
- we'll probably need a double buffered setup
- today: HW accelerated canvas could cause issues
- WebGL'd canvas
- connection between canvas and paint is the backing store
- Canvas: it has a bitmap backing store
- right now we do not persist the display items
- but we do the layer tree
- these are cases where rasterization is expensive
- we think this can be handled locally by a (theoretical) paint task
- Parallelizing layout
- Floats: contaminate up to the block formatting context (BFC)
- Layout: multi-step process
- first, compute intrinsic widths
- CSS layout:
- width is input, height is output
- for a typical block, you determine its width, then layout kids
to determine height
- computing width:
- look at style, border, margins, etc
- some things do "shrink-wrapping", where width depends on children
- for floats with autowidth-
- find intrinsic / preferred minimum width
- find max of longest word and (min of avail and preferred)
- computing intrinsic, preferred width is completely parallelizable
- does not depend on layout, parent
- same applies to table cells
- so even within these painful things there are emb. par. portions
- what about floats with fixed width---
- still don't know height, but things are somewhat easier
- computing width:
- is the CSS layout truly so complex?
- another question: flexibility
- highly tuned, manually written code is very inflexible
- automatically generated, highly tuned code is fast but of medium flexibility
- simple, hand-written code may be slow but very flexible
- another question: flexibility
- DOM.js
- we should move to an IDL-based autogeneration mechanism like in Gecko
- something we need---
- fast way to validate type of arguments, not necc. easy in JS
- should be able to repurpose the front end (Python program that generates a data structure)
- should be able to tweak and repurpose the code gen to produce code
- to some extent it's properly setup