[Webkit-unassigned] [Bug 92475] Use 8 bit atomic strings where possible

[bugzilla-daemon at webkit.org]

https://bugs.webkit.org/show_bug.cgi?id=92475





--- Comment #6 from Dan Carney <dcarney at google.com>  2012-08-03 00:08:22 PST ---
(In reply to comment #3)
> This is an interesting idea but I am missing how did we get here in the first place.
> 
> I wonder if you are fixing this too late in the stack and we could improve previous layer to already use 8 bits characters. From your experiments, from where were those AtomicString initialized and why are they using 16bits string?

In some ways, this is definitely the wrong place to execute the 8 bit check, but strings flow in from a bunch of places where no conversion to 8 bit strings has yet been done - most of the strings are css properties and dom element attributes.  The reason that I want to inject 8 bit strings into the atomic table here is that these strings are often used in string composition downstream, and this means that all compositions are 16 bit as well, since there is no conversion from 16 bit to 8 bit anywhere, only the opposite.  The memory savings I was hoping to achieve was more from composed strings.

I can tell you now that on a typical page, almost 100% of the time, the 8 bit branch is taken, but I'll test the performance a bit and repost, particularly checking how often these strings get up converted to 16 bit and how often they are used in compositions.  If it's often, pre-attaching the original buffer to the 8 bit string might be a good option that would actually lead to memory savings on compositions.

-- 
Configure bugmail: https://bugs.webkit.org/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are the assignee for the bug.