> We've reached the point where it's better for our development velocity to work in the open on the bleeding edge branch of V8. That's also better for our collaborators who are working on ports to more platforms.
I wonder who are the collaborators with access to Google's private V8 repo and what platforms they're porting to. If merging TurboFan to the open repo didn't reveal their partners' proprietary plans now, then why not develop in the open sooner?
I interpreted that differently - that Google's private, secret v8 repo is hidden from their partners as well. Therefore they decided to unveil the work publicly so that their partners can access it, start to work to port turbofan to the other platforms, and so forth.
As to who the partners are, you can see commits from Intel adding x87 support and Imagination doing MIPS, for example.
As to why not develop in the open sooner, good question - this is the second time v8 does this, after CrankShaft (third time if you count the initial unveiling of chrome and v8). Maybe it's just how they work.
I can imagine that they want to have some initial implementation before they let the public see it. They have a fairly good idea of what they want with it and want to make sure it goes in that direction before they let the public comment/commit.
in the context of V8 "x87-platform" means "x87-only for floating point" as opposed to normal ia32 platform where V8 assumes SSE2 is present. It's implemented as a separate platform port, not as a bunch of if's inside the ia32 port.
azakai didn't say it's a platform. But it is a part of the x86 instruction set that you can entirely avoid. So it makes perfect sense for support for it to be added later.
It would be nice to understand the goals of this project in terms of both approach as well as the overall performance improvements and how it would compare with the new LLVM-based approach taken by Safari/Webkit.
As a first guess, I'm not sure what the v8 strategy is here. The new compiler seems to use the "sea of nodes" approach as opposed to SSA form. A comparison of the two is here
The "sea of nodes" approach can give some speedups, but they don't appear huge - 10-20% in that link. Not sure how representative that data is. But it is interesting that modern compilers, like gcc and LLVM, typically use SSA form and not the approach v8 is taking, as further evidence that the "sea of nodes" is not clearly superior.
Perhaps the v8 designers believe there is some special advantage for JS that the new model provides? Otherwise this seems surprising. But hard to guess as to such a thing. If anything, JS has lots of possible surprises everywhere, which makes control flow complex (this can throw, that can cause a deopt or bailout, etc.), and not the best setting for the new approach.
Furthermore, the "sea of nodes" approach tends to take longer to compile, even as it emits somewhat better code. Compilation times are already a big concern in JS engines, more perhaps than any other type of compiler.
Perhaps v8 intends to keep crankshaft, and have turbofan as a third tier (baseline-crankshaft-turbofan)? That would let it only run the slower turbofan when justified. But that seems like a path that is hard to maintain - 2 register allocators, etc., - and turbofan seems like in part a cleanup of the crankshaft codebase (no large code duplications anymore, etc.), not a parallel addition.
Overall the Safari and Firefox strategies make sense to me: Safari pushes the limits by using LLVM as the final compiler backend, and Firefox aside from general improvements has also focused efforts on particular aspects of code or code styles, like float32 and asm.js. Both of those strategies have been proven to be very successful. I don't see, at first glance, what Chrome is planning here. However, the codebase has some intriguing TODOs, so maybe the cool stuff is yet to appear.
The "sea of nodes" approach is just a data structure for representing a program in SSA form. It's orthogonal to anything that has an impact on speed. E.g. GCC uses a tree representation, LLVM uses a CFG, and Hotspot (C2 and Graal) uses a "sea of nodes" representation, but they all represent code in SSA form and that representation is orthogonal to the quality of particular optimizations implemented within the framework.
The speedup reported in that paper is from running constant propagation and dead code elimination at the same time instead of doing them separately, which finds more constants and dead code because the two problems are coupled. The same process can be implemented in a more traditional CFG representation (and generally is--sparse conditional constant propagation).
Too late to edit this, but I should clarify: "data structure" is probably not the right word. To be more precise, "SSA form" is a property of variables in a program IR. It means variables are assigned only once, that defs dominate uses, and value flow is merged at control flow merge points with phi nodes. You can have different program representations that all represent values in SSA form, but differ in how they represent other things. Where the "sea of nodes" representation differs is that it explicitly represents control dependencies. In LLVM, you always have a control flow graph, with basic blocks and edges between them. Control dependencies between instructions are implicit from their placement in particular basic blocks. In a "sea of nodes" IR, there are no basic blocks.[1] Control dependencies are represented explicitly with control inputs to nodes, just as data dependencies are represented explicitly with data inputs.
This makes certain things easier in a "sea of nodes" IR. Normally, during optimization you don't have to worry about maintaining a legal schedule of instructions within and between the basic blocks. You just have to respect the control dependencies. However, in order to get executable code you have to impose a schedule on the nodes, whereas with a more conventional CFG IR, you already have a schedule in the form of the basic blocks and the ordering within them.
[1] See Section 2.2-2.4 of Click's paper: http://paperhub.s3.amazonaws.com/24842c95fb1bc5d7c5da2ec735e.... His IR replaces basic blocks with Region nodes. The only instructions that must be linked to Region nodes are ones that inherently have a control dependency. E.g. an "If" node takes a data input and produces two control outputs, which can be consumed by region nodes. "Phi" nodes must also have control inputs, so they can properly associate different control flow paths with the data values that are merged along those paths.
Here's something that's been bothering me since Chrome dropped the experimental support for `position: sticky` (and CSS Regions before) and I didn't find the right place to ask (nor is a submission about the JS engine an appropriate venue for this question, I know), so I'm going to hijack this thread:
We know some properties are expensive and when you use them a few times (or with certain values) you get sub-60fps scrolling — but why are they expensive? Are they inherently hard to optimize (e.g. different GPUs across mobile devices), or is it that nobody got to optimize them yet?
Because, in general, they can trigger layout/reflow, so they can't run on the compositor. For example, changing things like "top" on an absolutely positioned box can trigger reflow of the contents inside, because of the way CSS works (for example, if "bottom" is set, then the height changes, which can affect the size of things with percentage heights, which can cause floats to be repositioned, etc. etc.)
In traditional browser engines it's even worse because layout runs on the main thread, which is also shared with your JavaScript, so painting ends up transitively blocked waiting for your scripts to finish. That is not the case in Servo (disclaimer: I work on Servo), but making layout run off the main thread is hard for many reasons—some inherent to the problem, some historical in the design of current engines—so all current browsers run JS and layout together.
To elaborate on how this works in Gecko. The rendering pipeline has several optional stages:
requestAnimationFrame (Scripts) -> Style flush -> Reflow flush -> display list construction -> Layer construction (recycling) -> invalidation -> Paint/Rasterization -> Compositing (on it's own thread).
Gecko tries to only run each stage of the pipeline only if they are needed. Fast operations like a CSS transition on an opacity or transform will only activate the Compositing phase. WebGL only canvas drawing will only activate rAF + Compositing. Meanwhile a JS animation on "top" will run all of these phases.
> Because, in general, they can trigger layout/reflow, so they can't run on the compositor.
If you're doing absolute layout, it's easier/better performing to just set top/left to zero, set width and height to their correct values, and then use CSS transforms to position the element where you want it on the page. :)
(Not meant for you, pcwalton, but for others who might not know about a reasonable workaround.)
I wonder who are the collaborators with access to Google's private V8 repo and what platforms they're porting to. If merging TurboFan to the open repo didn't reveal their partners' proprietary plans now, then why not develop in the open sooner?