On 3/12/13 12:56 AM, Glenn Maynard wrote:
(I suppose a simpler optimization is simply copy-on-access: make a copy of
the backing store if the .data property of ImageData is accessed.
This may actually be a harder optimization to make in practice.
For example, Gecko+SpiderMonkey has the .data getter on ImageData
objects annotated as being pure and returning a constant value. This
means that if you have code like this:
img.data[0] = 1;
img.data[1] = 2;
CSE can get rid of the redundant .data gets. Similarly, .data gets can
be loop-hoisted in many cases.
Based on some spot measurements, at first glance WebKit+V8 has somewhat
similar behavior, however they get there (e.g. Gecko used to have a
behavior somewhat like this by simply making ImageData be plain-vanilla
JS objects with an own "data" property instead of WebIDL objects, and
then the JIT can optimize gets using normal alias analysis techniques
for slot gets).
But if .data can have side-effects the ability to do these sort of
optimizations goes out the window and you get a much slower .data
getter. So you get a faster putImageData but the tradeoff is slower
imagedata manipulation unless the script author performs the CSE and
LICM optimizations by hand (which some do and some don't).
So there's no really good way to make this optimization without
degrading performance of things people commonly do....
-Boris
