Speaking of object representation, this old-ish MIT paper came up when talking to Patrick the other day: http://cscott.net/Publications/lctes03-talk.pdf
In our mobile WebKit benchmarks, ~20% of the power is in memory, and perf quickly gets blocked on it. So, it's worth considering if the object representation can go down by a magnitude. Aggressive compression is especially attractive for parallel computing because you can trade time for space and win. I recall three especially cool ideas from the paper: 1. Compress 'mostly' constant fields: only allocate for the typical small case, and use a bit flag to indicate an instance where you need to instead do a hash lookup. This can be because a value is generally a default, or has a small range. 2. Automatically do this w/ program analysis or profiling. Perhaps a more important question for Rust is whether Rust can make optimized access safe/transparent, even without autotuning. 3. Hashing the speculation indicators. For example, 1 bit for predictable style attributes, another bit for DOM. - Leo On Jul 24, 2013, at 12:00 PM, [email protected] wrote: > Send dev-servo mailing list submissions to > [email protected] > > To subscribe or unsubscribe via the World Wide Web, visit > https://lists.mozilla.org/listinfo/dev-servo > or, via email, send a message with subject or body 'help' to > [email protected] > > You can reach the person managing the list at > [email protected] > > When replying, please edit your Subject line so it is more specific > than "Re: Contents of dev-servo digest..." > > > Today's Topics: > > 1. Inline DOM storage progress report (Josh Matthews) > 2. Re: Inline DOM storage progress report (Boris Zbarsky) > 3. Re: Inline DOM storage progress report (Patrick Walton) > > > ---------------------------------------------------------------------- > > Message: 1 > Date: Tue, 23 Jul 2013 18:34:10 -0400 > From: Josh Matthews <[email protected]> > To: [email protected] > Subject: [dev-servo] Inline DOM storage progress report > Message-ID: <[email protected]> > Content-Type: text/plain; charset=ISO-8859-1; format=flowed > > I took the shortest route for prototyping right now and added really > basic APIs that did what I needed for our local embedding of > Spidermonkey. Accordingly, there are now branches of servo, rust-mozjs, > and mozjs (all named 'inline') that contain my work to date on storing > DOM objects inline in their JS wrappers' fixed slots. There are some > rooting problems that make it crash once in a while, but otherwise > test_bindings.html runs to completion on a regular basis. > > Thing to note: > * Each Option<AbstractNode<ScriptView>> is 16 bytes, so just the base > Node is already way over the inline storage limit, leading to the > NodeOutOfLine structure in src/script/dom/node.rs. > > See > https://github.com/mozilla/servo/commit/5370fc15acbe5d3407d5a97599f6b2fdb780714b > > for the juicy stuff. > > Cheers, > Josh > > > ------------------------------ > > Message: 2 > Date: Tue, 23 Jul 2013 22:08:36 -0400 > From: Boris Zbarsky <[email protected]> > To: [email protected] > Subject: Re: [dev-servo] Inline DOM storage progress report > Message-ID: <[email protected]> > Content-Type: text/plain; charset=ISO-8859-1; format=flowed > > On 7/23/13 6:34 PM, Josh Matthews wrote: >> Accordingly, there are now branches of servo, rust-mozjs, >> and mozjs (all named 'inline') that contain my work to date on storing >> DOM objects inline in their JS wrappers' fixed slots. There are some >> rooting problems that make it crash once in a while, but otherwise >> test_bindings.html runs to completion on a regular basis. > > How does the resulting rust-side code look? Is is sufficiently like > idiomatic rust? > > Basically, would like to avoid us just pushing the complexity of > implementing a DOM object from CC boilerplate to access-to-members > boilerplate... > >> Thing to note: >> * Each Option<AbstractNode<ScriptView>> is 16 bytes, so just the base >> Node is already way over the inline storage limit, leading to the >> NodeOutOfLine structure in src/script/dom/node.rs. > > Hm. What are we storing inline, then? From the diff it looks like > basic stuff like parent and owner document and prevsibling is out of > line... What's our inline storage limit? > > -Boris > > > ------------------------------ > > Message: 3 > Date: Tue, 23 Jul 2013 19:17:10 -0700 > From: Patrick Walton <[email protected]> > To: [email protected] > Subject: Re: [dev-servo] Inline DOM storage progress report > Message-ID: <[email protected]> > Content-Type: text/plain; charset=ISO-8859-1; format=flowed > > On 7/23/13 7:08 PM, Boris Zbarsky wrote: >> On 7/23/13 6:34 PM, Josh Matthews wrote: >>> Accordingly, there are now branches of servo, rust-mozjs, >>> and mozjs (all named 'inline') that contain my work to date on storing >>> DOM objects inline in their JS wrappers' fixed slots. There are some >>> rooting problems that make it crash once in a while, but otherwise >>> test_bindings.html runs to completion on a regular basis. >> >> How does the resulting rust-side code look? Is is sufficiently like >> idiomatic rust? >> >> Basically, would like to avoid us just pushing the complexity of >> implementing a DOM object from CC boilerplate to access-to-members >> boilerplate... > > Note that the Rust compiler is missing a feature needed for automatic > generation of trace hooks: external #[deriving] implementations. This is > a sub-bug of "allow procedural macros to be defined outside the > compiler" (and should be fairly straightforward once that is > implemented). This feature is scoped for Rust 1.0 but is not yet > implemented. > > Also, I imagine this feature will benefit from the yet-to-be-implemented > compiler support for custom smart pointers (and in fact this sort of > thing in Servo was much of the reason for my pushing this feature). This > should make accessing members of these DOM nodes more idiomatic. My > "Removing Garbage Collection from the Rust Language" blog post describes > the plan very briefly, but Niko and I need to iron out the exact details > still. Work related to this is underway in the Rust compiler (in > particular removing `@fn`). > > Patrick > > > > ------------------------------ > > _______________________________________________ > dev-servo mailing list > [email protected] > https://lists.mozilla.org/listinfo/dev-servo > > > End of dev-servo Digest, Vol 17, Issue 16 > ***************************************** _______________________________________________ dev-servo mailing list [email protected] https://lists.mozilla.org/listinfo/dev-servo

