[pbsdp]

> What is your goal - speed? Code size? Portability? Security?

That depends on the target. I selected them because they represent different aspects of the state of the art.

> ... and I would argue that modern JS VMs are competitive with them in many respects

But not all, and in many places, not even close. JS VMs make trade-offs that we don't need if we abandon JS, and on top of it all, you have Mozilla insisting against shared-state multi-threading supported and exposed by the VM, which discards enormously important optimization opportunities) -- and that's just the tip of the iceberg.

Get rid of JS, get rid of Mozilla's self-enforced constraints on the browser, and look very seriously at something like NaCL which blends native execution performance (including being able to drop to SIMD instructions) with security sandboxing.

Let language authors target the low-level machine, and target your high-level general purpose bytecode.

Once you've built a runtime in which I could implement a JS runtime that's as performant as yours, we will no longer be operating in a two-tier universe in which browser makers think that JS is good enough for everyone but themselves.

[azakai]

JS VMs make tradeoffs, yes. Regarding your specific examples:

1. Shared-state multithreading. This has been proposed by various people in the past, and is still being debated. There are people for it and against it in various places. It might happen, if there is consensus to standardize it.

2. SIMD: PNaCl (which I mention since you mention NaCl in that context) does not have SIMD. But of course it can add SIMD, just like Mono and Dart have, and there is a proposal for JS as well, hopefully that will get standardized.

So even those limitations are in principle resolvable. They do depend on standardization, of course, but so would any VM you want to run on the web.

> Let language authors target the low-level machine, and target your high-level general purpose bytecode.

asm.js is meant to get performance similar to a low-level machine. It's already very close to native on many benchmarks.

> Once you've built a runtime in which I could implement a JS runtime that's as performant as yours

I don't follow that. You can't make a fast JS runtime in your previous examples (the JVM, PNaCl, .NET, etc.).

[pbsdp]

> I don't follow that. You can't make a fast JS runtime in your previous examples (the JVM, PNaCl, .NET, etc.).

I included ones you could, such a NaCL. And I should probably have also included the native platforms (Apple, MS, Google), because they're the competition, even if their sandboxing or runtime environment isn't what one might call 'state of the art'.

At the end of the day, it's the two-tier universe that bothers me the most. You're content to foist the JS runtime on everyone but yourselves. Once you implement Firefox entirely as an application running in a JS VM, the argument that JS is good enough might carry some weight.

[azakai]

The point is that NaCl is not portable, just like a native platform such as Windows.

Nonportable native platforms are there, you can write apps for them. They are even the dominant platforms on mobile.

But the web's entire purpose for existing is to be portable. So you do need something like JS or PNaCl or the JVM. And all of those, due to being portable and secure, impose limitations, such as limiting the native operations that you perform. That's unavoidable. But again, if you don't like that, develop directly for a native platform.

[pbsdp]

> The point is that NaCl is not portable, just like a native platform such as Windows.

So provide fat binaries. Maximize performance on ARM and i386, with fallback to PNaCL for future platforms that are neither.

> They are even the dominant platforms on mobile.

For a good reason. Adopting their strengths while eschewing their weaknesses (proprietary and single-vendor) would benefit the entire industry greatly.

[azakai]

Fat binaries are limited, and people will simply provide the parts for the platforms they currently care about, without care for future platforms.

That's fine for most things, but web content is something that we do want to always be accessible.

It is hard to adopt the strengths of native execution using all the lowest-level tweaks specific to one platform, because that inherently limit portability by definition (and often also security).

I share your goals, but don't think there is an obvious better compromise than the one we are all already making on the web.

[slacka]

The teams at Mozilla and Google have done an amazing job of improving JavaScript performance. But, I also have to agree with PBS, that Mozilla has a knee-jerk reaction to many of Google’s promising technologies like WebP and PNaCl. Like many HTML5 technologies that are being shoehorned into areas they were not originally designed for, JS was never meant to be a bytecode and likely will never be able to achieve the performance possible with PNaCl.

If W3C and WHATWG were competent and delivered a viable web application platform, we wouldn't be hiring Android and iOS developers now. Mozilla needs to be more open to new technologies like PNaCl that enable the web to be competitive.

[comex]

> I included ones you could, such a NaCL.

You cannot do it in NaCl, as being able to map pages rwx would break the security model.

[pbsdp]

All you need to do is perform NaCL validation of pages before marking them executable; this is what NaCL already does.

You can either do this through a high-level "JIT API" that generates safe machine code from a validated IR (aka PNaCL), or through a fancier version of mprotect() that validates the actual machine code (aka NaCL).

In a wondrous hypothetical future where processors support a NaCL restricted operating mode, you wouldn't even need to validate; just set the processor state to thumb^Wnacl mode, and define a syscall instruction that switches to direct execution (without actually requiring a context switch to the kernel, just flip an execution flag).

This is why NaCL is so damn interesting, and JS/asm.js is not. NaCL has the possibility of turning our design of restricted execution environments on its head, in a very good (for performance) way.

[comex]

You can already do that with NaCl - or even asm.js with eval, I bet it wouldn't be that much slower. But current JITs apparently gain a significant amount of performance with inline caching, which expects code to be able to be rewritten very quickly.

Edit to respond to your edit: although it would be cool to be able to have a "sandboxed mode" that can somehow be switched out of more cheaply than an interrupt, the whole thing seems like a massive hack to me. After all, NaCl does not take advantage of its pseudo-ability to do so: NaCl code runs in its own process and already incurs a context switch whenever it communicates with the browser, so there is no inherent hardware reason NaCl couldn't just run directly under the kernel and have the kernel provide the same level of sandboxing as the NaCl runtime currently does. It's just an issue of getting such an approach to work portably with existing kernels... hardware support might be able to make it easier to get that to work, but it's probably unnecessary.

[azakai]

> In a wondrous hypothetical future where processors support a NaCL restricted operating mode

If we speculate wildly, why not an "asm.js restricted operating mode"? Not saying that's a good idea, but I'm not sure why a NaCl one would be either. Both PNaCl and asm.js should reach pretty much native speed anyhow.

Btw, the L is not capitalized in NaCl.

[pbsdp]

> If we speculate wildly, why not an "asm.js restricted operating mode"?

Because we already had Jazelle, and it sucked, and we learned our lesson. A Jazelle that operates on ASCII-encoded bytecode with traps for unsupported JS constructs? No thanks. :)

> Not saying that's a good idea, but I'm not sure why a NaCl one would be either. Both PNaCl and asm.js should reach pretty much native speed anyhow.

Pretty much native and actually native are very different things.

I've had to sit down and hand-optimize critical paths that would not have been viable otherwise, and would have meant discarding a feature, or introducing a significant impact on usability -- and that's on platforms where similar investments in runtime library optimization were already made by the platform vendor, too.

If we're going to throw away performance, it has to be for a good reason. As someone who doesn't spend all day on platform development (as interesting as it might be), my focus is in providing the best possible user experience.

I'd love to do that on the web, but the web needs to punting throwing away solid technology decisions because of the bad technology decisions made in the early 90s when none of us knew what the hell we were doing.

> Btw, the L is not capitalized in NaCl.

Whoops. Thanks. Now I will look less stupid in front of software engineers AND my chemistry buddies. :)