Comment by haberman | Hacker News Reader

Comment by haberman | Hacker News Reader

Parent Full thread

There is a wide gap between fully embracing a technology and spreading misinformation about it. I respect Mozilla's decision not to integrate NaCl, to argue that it's premature to talk about doing so while it's underspecified and coupled to Chromium, and to set criteria that it must meet before it will be on the table for further discussion. I can understand concerns about cost and governance and an unwillingness to jump on what is perceived as a "Google treadmill." None of my comment was about any of that.

What I can't understand is the fundamentalist reaction to the very idea of native code, the ignoring/dismissing of serious work to solve the problem of portability, the liberal use of words like "never" and "non-starter," spread of FUD by invoking inaccurate comparisons like ActiveX (vis a vis its security model) and DLL Hell, and the spreading of misinformation. For example, PNaCl is not, and as far as I can tell never has been, dependent on machine word size. The link you cited doesn't apply because it is arguing against a different approach than what PNaCl actually does.

PNaCl works by defining a little-endian ILP32 machine as the target and fixing all important characteristics of this machine independently of the characteristics of the underlying CPU. This abstract machine's characteristics are defined in such a way that they can efficiently be translated to native code on any modern CPU. This is all covered in the introductory doc: http://src.chromium.org/viewvc/native_client/data/site/pnacl...

> a cross-browser approach targeting JS VMs, which are already there and getting fast enough with typed array memory models to compete with PNaCl? (We aim to demonstrate this.)

This is a far more reasonable and compelling story. By all means talk up your stuff and argue that you can win in the free market of ideas. I'm not arguing that I or anybody else should be able to dictate to developers what technology they use; on the contrary it is the Mozilla argument of "no one gets to the machine except through our VM and our GC" that paternalistically ties developers' hands and limits their options.

You are still being free with accusations of spreading misinformation and other evils. If you want to have a real exchange, cool it! Just try to imagine how a hardball from me casting aspersions on you for suspected bad or unfair (to Mozilla; "fair" to Google) motives might feel.

Thanks for the PNaCl pointer. My comment was based on LLVM bitcode having machine word size dependencies. This was an issue a while ago. I should have checked to see if it remained one.

This correction doesn't alter the general unreadiness of PNaCl for the web, on several fronts. Pepper is one, but PNaCl performance lagging NaCl is another. The Chrome Web Store features games ported via NaCl, for performance -- not PNaCl, which would be significantly slower. On this basis alone, it's premature for you to push PNaCl ahead of Google.

> This is a far more reasonable and compelling story.

Well, gee, thanks a ton! :-|

I've been telling this story clearly since Fluent in May. That you chose not to hear it and instead flung accusations and told sob-stories about big bad Mozilla is your doing, not mine.

Here's a final clue: all browser vendors, definitely including Chrome, make the rule (not an argument) "no one gets to the machine except through our VM(s) and GC(s)" -- outside of a few dying plugins, which are even source-licensed and co-released.

And that brings back my final point: NaCl is for safer plugins, which are OS-specific anyway. The likeliest evolution of SFI or CFI enforcing compilers and runtimes as plugin hosts is via the OS, not the browser. Write a letter to Microsoft and Apple, not to Mozilla!

These are shallow arguments: * pepper is "inspired" by nsapi, clarify your point. * PNaCl performance lagging isn't a solid argument, you know it'll get better, the solution might even be to cut LLVM out save for bitcode. * "nobody does this at the moment" so why does it belong in the OS?

In reverse order:

* Why in the OS? I didn't say "belong", just "likelier". That is because plugins are native code compiled by OS-dependent toolchains, and OS vendors are few (three that matter) and lock up native code these days via SDK licenses, app store rules, and even kernel-level restrictions.

In contrast, there are four or five competitive browsers, only one of which has Pepper and the rest do not -- and will not.

* I do not know how much better PNaCl can get. The shallow argument here is your assertion that "you know it'll get better". The same could be speculated about JS performance at Emscripten-generated code, and that works cross-browser. That's the cross-browser path of least resistance, compared to the practically unpassable Pepper barrier.

* Pepper is "inspired" by lots of APIs, but here the shallow shoe fits your new-HN-user drive-by. NPAPI is a sunk cost all browsers save modern IE have paid out for years. Pepper is new and much bigger. Have you even read all the interfaces?

The bottom line is that whatever PNaCl performance wins may lie in the future -- and I will believe them when Google does as shown by Chrome Web Store games being PNaCl'ed not NaCl'ed -- Pepper is the blocker for any cross-browser adoption in reality.

This ignores principled objections to more native code on the web, as a "social ill". Let's take that up separately, because it could override any technical argument. I'm happy to stop on the Pepper point for now, since Google manifestly is stuck there.

Why differentiate plugins? What makes a VM with JIT not a plugin save the browser vendor shipping it with the browser?

Why wouldn't other browsers have Pepper?

Compilers are as good as what they've been tuned for. In my view PNaCl's shortcoming is startup time because it lacks a JIT and LLVM's back end is too slow for now. Speed up the backend or JIT code and you'll get close to GCC performance while being portable and somewhat language agnostic.

Yes I have seen pepper, and most of the interface relates to the GPU. How is sunk cost better, when a big part of the API can be backed by what canvas relies on?

You would consider adopting PNaCl and pepper in FF if there were games that targeted them? If the code were contributed to Mozilla?

What do you mean by "more native code"? Can't view source?

I appreciate the answers.

>Why differentiate plugins? What makes a VM with JIT not a plugin save the browser vendor shipping it with the browser?

I think you answered your own question with the "save" part.

The vendor shipping it with the browser means it controls it, it has responsibility for it, it secures it, and it allows it. End of story.

>Why wouldn't other browsers have Pepper?

JS is a necessity for a web browser/vendor, and is already present in it. Pepper is not, and there are NO signs it will be. Do you see any movement towards adoption as of now? I see the opposite, the abandonment of even old style plugins.

I'll rephrase: why is any new VM to be relegated to the OS? The presence of incumbent VMs?

JS is an incumbent. Pepper is similar to nsapi, and has nice features which are compatible with HTML5's implementation (as in canvas). Saying it shouldn't be adopted because Nobody is adopting it is circular.

What are old-style plugins? Anything not JS?

I said clearly why Pepper is not being adopted: it is a gigantic pile of API and implementation specified only by the C++ in chromium.org svn. Other browsers cannot port all that OS and WebKit glue code except at very high cost, direct and opportunity -- and even then on a bet that Pepper + NaCl wins, and again on a treadmill far behind Chrome.

Do you actually work on a browser codebase? If so, have you worked on competing browsers' codebases at all? Do you begin to see the problem? It's not quite Active X (open source is a small help), but it's on that slope and uphill only a bit.

> Other browsers cannot port all that OS and WebKit glue code except at very high cost

Why would any other browser need that glue code? The Pepper API is large but fairly straightforward and doesn't change dramatically between revisions. In addition, I don't believe Google has ever said that they wouldn't make the development process around those changes more open (at least making them public before pushing the new implementation out to the world).

> Do you actually work on a browser codebase? If so, have you worked on competing browsers' codebases at all? Do you begin to see the problem? It's not quite Active X (open source is a small help), but it's on that slope and uphill only a bit.

I've only worked on Webkit a small amount (mainly doing security analysis) but I worked with Pepper a good deal and I've worked on Gecko for a decent while now. I really don't see the incompatibility; there are plenty of good arguments against NaCl, but I don't think there's a fundamental problem there. I can definitely understand not wanting to allocate resources to the issue, but not being opposed to the issue in general.

> Why would any other browser need that glue code?

Because other browsers do not use WebKit, or at least chromium WebKit. Are you really asserting that no glue code is required on any other browser?

> The Pepper API is large but fairly straightforward

Where is the spec? You are not in the real world here.

There are plenty of differences between Gecko's audio APIs and Pepper's. If you really work on Gecko, mail me about this. I have reason to doubt your claims here.

If I sound argumentative and fired up, it's because I feel like Mozilla has been casting stones on this issue for years. Imagine how you would feel if Google executives were publicly criticizing Mozilla efforts like Persona, arguing that they would never support them and no one else will either, basing their criticisms on issues that you are actively fixing.

(For what it's worth, Persona looks promising to me personally, and I also like Rust very much, a lot more than Go. I say this to demonstrate that I'm not just a Google partisan and that I admire a lot of what comes from Mozilla).

I am much happier to discuss this dispassionately on a technical basis. I'm much happier if I don't have to argue against what to me are very unfair accusations, like being as proprietary as Silverlight.

> Here's a final clue: all browser vendors, definitely including Chrome, make the rule (not an argument) "no one gets to the machine except through our VM(s) and GC(s)"

I don't understand the argument you are making, (P)NaCl are specifically designed to allow execution of untrusted code without making it run on top of a VM or GC. And (P)NaCl executables are OS-independent. I don't understand what you're getting at here.

>(P)NaCl are specifically designed to allow execution of untrusted code without making it run on top of a VM or GC

And this is the argument he's making: that does not fly by browser vendors. They DON'T want to have code run OUTSIDE their VM/GC.

He said "including Chrome." Chrome supports NaCl. This does not compute.

It's not hard: Chrome sandboxes NaCl'ed code and links it against a runtime, Pepper. That's a "VM" by any other name.

Thanks for the support, but that's not what I meant. NaCl + Pepper is like a VM where the compiler does the heavy lifting so the native code can run safely (Software Fault Isolation, SFI -- wild pointers lead to a safe non-exploitable crash), rather than a JITting or MMU- or hypervisor-based VM doing the heavy lifting at runtime.

It's quite clever, but still enough of a new thing that Chrome also sandboxes NaCl'ed code out of process. Belt and braces are good. No silver bullets.

But a VM is as a VM does. This is part of Google's VM-set and not any other browsers. The rule still applies.

Truly unsafe native code in plugins (e.g., un-NaCl'ed Flash) runs out of process too, and sandboxed to some extent, but it can cause problems that are not contained (and did at the last CanSecWest Pwn2Own contest, IIRC).

The link he cited still does apply. It discusses several different issues. PNaCl's portability only covers a subset of them.

Thanks! I happen to agree with with Dan Gohman (http://comments.gmane.org/gmane.comp.compilers.llvm.devel/43...), but I'm not sure where Chris Lattner ended up on this.

Much is possible in software, so perhaps some day, or under some transformation, LLVM bitcode would be suitable as a stable long-term object file format.

There's still a point here: PNaCl is pushing a stone up a very tall hill. ANDF and other Universal Object formats go back to the 70s if not earlier. It's very difficult to standardize such things, never mind Pepper.

FWIW, the aim for LLVM is to avoid breaking the bitcode format now 3.0 has shipped — not that it's platform independent or anything else yet.

Work with say Khronos group to establish an OpenCPU standard with a source code and possibly intermediate representation.

Socialize amongst CPU vendors, and interest platform makers in the mobile and desktop space.

Watch it absorbed by web standards.

Reply on news.ycombinator.com