Running Lua in a browser via Emscripten(mozakai.blogspot.com)
mozakai.blogspot.com
Running Lua in a browser via Emscripten
http://mozakai.blogspot.com/2013/05/lua-in-javascript-running-vm-in-vm.html
10 comments
Wow! Awesome! Sorry I missed it when you posted this the other day. (https://news.ycombinator.com/item?id=5775859) OpenSCAD in JavaScript is worth a top-post all by itself. I know of OpenJSCad and CoffeeSCAD, but I prefer the simplicity of plain old OpenSCAD. I do all my 3D design with OpenSCAD.
Curiously, even though compilation is indeed near real-time in Firefox, the rendering of the model itself is really jaggy (when rotating). This is not the case with Chrome, where it renders smoothly.
For reference, I'm using 2012 MBA with Firefox 23.0a2 (2013-05-30) (clean profile, just to be sure) and Chrome 27.0.1453.93.
For reference, I'm using 2012 MBA with Firefox 23.0a2 (2013-05-30) (clean profile, just to be sure) and Chrome 27.0.1453.93.
IIRC FF still doesn't use vsync for requestAnimationFrame so that could cause rotating to look less smooth.
Ah... that's interesting. Thanks for the info Kevin, I'll look into that (my poor WebGL coding notwithstanding!)
Yes this is something I noticed too (as in the stuttering in FF compared to Chrome). I implemented an image based CSG algorithm directly in WebGL (the one place where Emscripten couldn't help), so I would imagine that could be vastly improved, as its the first time I've ever done anything in OpenGL of any variety.
To my surprise, lua.vm.js gzipped is about the same size (within 2K) as my Win32 Lua DLL build gzipped. The latter has the Microsoft C runtime statically linked, making the comparison more fair IIUC, since code built with Emscripten has to include a libc implementation.
This finding demonstrates that asm.js isn't significantly less efficient than LLVM bitcode or native code, even though it's text.
This finding demonstrates that asm.js isn't significantly less efficient than LLVM bitcode or native code, even though it's text.
By "efficient" I assume you mean the object code is compact, filesize-wise? It's fairly typical that byte-code or otherwise high-level object code is more compact than machine code. For example, Java .class files will be smaller than their equivalent native-compiled object code.
> It turns out that the entire compiled Lua VM fits in 200K when gzipped. That's too much for some use cases, but certainly acceptable for others (especially with proper caching).
Tell that to someone on a mobile device with poor connectivity.
> In particular, remember that the Lua VM is often significantly faster than other dynamic languages like Python and Ruby. These languages are useful in many cases even if they are not super-fast.
Something people often overlook is that performance is highly dependent on where the code runs. Ruby is fine for server-side programs because you can always go fast by throwing more hardware at it.
Languages that run on end-user machines don't have that luxury. This is why the playing field for client-side programming languages is much more constrained and why C++ is still hugely popular there. It's also why so much work has gone into optimizing JavaScript.
Being half as fast as JS could be tolerable for some apps, but that means your app will likely be slow and stuttery in some cases. When it is, there's little you can do about it. Is Lua that much of an improvement over JS to justify that?
> There are however some tricky issues, for example we can't do cross-VM cycle collection - if a Lua object and a JavaScript object are both not referred to by anything, but do refer to each other, then to be able to free them we would need to be able to traverse the entire heap on both sides, and basically do our own garbage collection in place of the browser's - for normal JavaScript objects, not just a new type of objects like Lua ones.
This is a huge deal. It basically means if you use this, your app is very likely to leak memory unless you are very careful. The difficulty of being appropriately careful is exactly why we moved to GC languages in the first place.
WebKit (now Blink on the Google side) actually has a similar problem already: WebKit manages memory for the DOM separately from V8's garbage collector. This adds a bunch of complexity to the browser to deal with those cycles and has, I think, a significant performance cost.
It's enough of an issue that the Chrome team is starting a new project ("oilpan") to provide a unified GC shared by both V8 and the DOM.
Don't get me wrong, I think this is a very cool hack. But I don't think it says much about the viability for using something like this for real apps, at least not yet.
Tell that to someone on a mobile device with poor connectivity.
> In particular, remember that the Lua VM is often significantly faster than other dynamic languages like Python and Ruby. These languages are useful in many cases even if they are not super-fast.
Something people often overlook is that performance is highly dependent on where the code runs. Ruby is fine for server-side programs because you can always go fast by throwing more hardware at it.
Languages that run on end-user machines don't have that luxury. This is why the playing field for client-side programming languages is much more constrained and why C++ is still hugely popular there. It's also why so much work has gone into optimizing JavaScript.
Being half as fast as JS could be tolerable for some apps, but that means your app will likely be slow and stuttery in some cases. When it is, there's little you can do about it. Is Lua that much of an improvement over JS to justify that?
> There are however some tricky issues, for example we can't do cross-VM cycle collection - if a Lua object and a JavaScript object are both not referred to by anything, but do refer to each other, then to be able to free them we would need to be able to traverse the entire heap on both sides, and basically do our own garbage collection in place of the browser's - for normal JavaScript objects, not just a new type of objects like Lua ones.
This is a huge deal. It basically means if you use this, your app is very likely to leak memory unless you are very careful. The difficulty of being appropriately careful is exactly why we moved to GC languages in the first place.
WebKit (now Blink on the Google side) actually has a similar problem already: WebKit manages memory for the DOM separately from V8's garbage collector. This adds a bunch of complexity to the browser to deal with those cycles and has, I think, a significant performance cost.
It's enough of an issue that the Chrome team is starting a new project ("oilpan") to provide a unified GC shared by both V8 and the DOM.
Don't get me wrong, I think this is a very cool hack. But I don't think it says much about the viability for using something like this for real apps, at least not yet.
> Is Lua that much of an improvement over JS to justify that?
This depends on the project, of course. For existing software that's already written in Lua and now needs to run inside a browser, running the Lua VM in JS may be the best way to go. However, I've decided that for new code, Lua doesn't have enough advantages over JS to justify the problems with running a VM in a VM.
Here are the things that I consider significant advantages of Lua over JS, and my thoughts on each.
1. Coroutines: JS is getting a form of these via generators, though I don't know how soon that feature will become ubiquitous. In the meantime, a compiler can turn code that uses coroutines into continuation-passing style.
2. Weak references: It's unfortunate that JS doesn't have these. But as the OP pointed out, running a VM within a VM introduces other memory management problems, since the guest VM has its own GC. To avoid those issues, I can live without weak references in the language.
3. Metamethods: The most well-known metamethods are for overriding table operations, so one can implement properties, proxies, or other dynamic behavior. I've sometimes found these useful on previous Lua projects, but for new code, I can do without them in order to avoid the problems of running a VM in a VM.
4. Non-string keys for tables/objects: There are ways around this. For example, instead of using another object as a key, one can assign a unique ID to the object, then use that as the key.
This depends on the project, of course. For existing software that's already written in Lua and now needs to run inside a browser, running the Lua VM in JS may be the best way to go. However, I've decided that for new code, Lua doesn't have enough advantages over JS to justify the problems with running a VM in a VM.
Here are the things that I consider significant advantages of Lua over JS, and my thoughts on each.
1. Coroutines: JS is getting a form of these via generators, though I don't know how soon that feature will become ubiquitous. In the meantime, a compiler can turn code that uses coroutines into continuation-passing style.
2. Weak references: It's unfortunate that JS doesn't have these. But as the OP pointed out, running a VM within a VM introduces other memory management problems, since the guest VM has its own GC. To avoid those issues, I can live without weak references in the language.
3. Metamethods: The most well-known metamethods are for overriding table operations, so one can implement properties, proxies, or other dynamic behavior. I've sometimes found these useful on previous Lua projects, but for new code, I can do without them in order to avoid the problems of running a VM in a VM.
4. Non-string keys for tables/objects: There are ways around this. For example, instead of using another object as a key, one can assign a unique ID to the object, then use that as the key.
0. Lua is much simpler, with a much smaller WAT-quotient, without loss of power.
(For what I'm doing that's outweighed by the convenience of direct browser support, but to me that's the big attraction.)
(For what I'm doing that's outweighed by the convenience of direct browser support, but to me that's the big attraction.)
ES6 will introduce Maps and WeakMaps, at least, which addresses two and four; similarly, proxies will enable three.
WeakMaps don't quite give you everything that weak references do in other languages. Sometimes weak refs can be used to check if the object referred to has gone away (the weak ref becomes null in that case), that is not possible with WeakMaps. You also cannot create a list of weak references and iterate over them with a WeakMap.
There has been some discussion about adding more powerful weak refs to JavaScript, but I'm not sure where that discussion stands.
There has been some discussion about adding more powerful weak refs to JavaScript, but I'm not sure where that discussion stands.
I wrote a parenthetical about weak references. I wonder why I deleted it. In short: weakrefs are unlikely to appear any time soon in ES, as they open up all kinds of fun around cross-origin objects (being able to detect their approximate life-time, and that's a fairly major side-channel attack) in some lovely edge-cases.
If you have any more details about this, or even if you don't, please post to es-discuss about it. Currently there is broad consensus on TC39 to add weak references in JS in the future.
I was imprecise; Lua has weak tables, not general weak references. So I'd be happy with weak maps in JS.
It means you can port your game to the web without having to change its addons written in Lua to something else.
If the choice is lua.js vs porting code to Dart you know 'good enough' is going to win.
If the choice is lua.js vs porting code to Dart you know 'good enough' is going to win.
Lua is famously easy to embed. Some implementations are fast as hell. It's comparatively a very rational and minimal language. It doesn't have to immediately interact with Javascript objects in any interesting way.
Please just embed the damn thing in a browser already. It's been way too long without any competiton to Javascript being allowed.
Please just embed the damn thing in a browser already. It's been way too long without any competiton to Javascript being allowed.
A lot of people want their language embedded in the browser suh as Lua, Dart, Ruby, etc.
But the way things are standing now it's not gonna happen. JavaScript, for the better and worse, is what we are stuck with. No browser vendor will be able to push though another language, and no standards committee would suggest a second scripting language to the web that runs alongside JavaScript.
If you absolutely cannot stand programming in JavaScript, which is the only native option for the web, you have to make do with the "language written on top of another language" solutions that are so popular these days, along with it's somewhat costy overhead.
But the way things are standing now it's not gonna happen. JavaScript, for the better and worse, is what we are stuck with. No browser vendor will be able to push though another language, and no standards committee would suggest a second scripting language to the web that runs alongside JavaScript.
If you absolutely cannot stand programming in JavaScript, which is the only native option for the web, you have to make do with the "language written on top of another language" solutions that are so popular these days, along with it's somewhat costy overhead.
The only real option is presumably the Dart route: Work really hard to make it compile to fast javascript and integrate well, and do a native VM as a separate option that can at some point be slotted in as an "optimisation" if the language gets enough traction. I don't think that alternative looks that bad. I'd rather have that, than go back to plugin hell or have browsers bloated with support for tons of languages.
Direct link for those who want to try it live: http://kripken.github.io/lua.vm.js/repl.html
Well, Chromium and Firefox are open source. Give it a go, if you know C/C++ it shouldn't be that difficult.
If this is the real deal, then why not ship common language VMs (in asm.js) with the browser? Then you don't need to worry about 200k at all and all code can run in the browser? I mean, if google is shipping NaCL and Dart VMs, why not? Updates could ship with browser updates, or you can still load your own rolled version if you like.
Hmm, you could avoid some downloads by basically populating the browser cache by default with some common libraries, like jQuery, and maybe a Lua VM if that catches on, yeah. But this would be totally different from shipping a nonstandard VM, since there is no violation of the standards process if you are basically just optimizing away first downloads of common JS libraries. It does have the cost of increasing the initial browser download though, which is perhaps why this isn't already done (also things like jQuery have many versions which makes things harder).
Note however that Google has not actually shipped NaCl or Dart. (NaCl technically ships in Chrome, but is disabled on the web by default, it is only usable in the Chrome Store.) So currently no browser is shipping a new nonstandard VM - which is good, because if a major browser did that it could fragment the web.
Note however that Google has not actually shipped NaCl or Dart. (NaCl technically ships in Chrome, but is disabled on the web by default, it is only usable in the Chrome Store.) So currently no browser is shipping a new nonstandard VM - which is good, because if a major browser did that it could fragment the web.
FYI, current mobile browser caches are very weak and unreliable. For instance, iOS clears all cached data just by exiting Safari. Android is also in a similar situation (unpredictable and can't explain).
I guess it is caused by the storage limitations. We can't rely on the mobile browser cache, at least for now.
I guess it is caused by the storage limitations. We can't rely on the mobile browser cache, at least for now.
I really feel like the right approach for this kind of thing is something like emscripten, but not emscripten. It just doesn't make sense to duplicate the garbage collection (never mind that interactions between different GCs can be strange and unpredictable) and run a whole VM for it.
What we need is a similarly narrow target JS subset for dynamic languages where the outer GC is actually functional and lookup/inline cache kinds of optimizations can be performed on correctly generated code.
What we need is a similarly narrow target JS subset for dynamic languages where the outer GC is actually functional and lookup/inline cache kinds of optimizations can be performed on correctly generated code.
I like the idea, and it could work in many cases, but it's hard to make it truly universal. For example languages like Lua, Java and C# have finalizers, and JavaScript GCs do not, so you can't reuse the outer GC. Other possible issues are weak refs, etc.
So even if we duplicate by compiling another GC, we are enabling new types of GCing, it isn't an exact clone.
So even if we duplicate by compiling another GC, we are enabling new types of GCing, it isn't an exact clone.
While Java does have finalizers, there is no guarantee that they will run. In other words, a valid implementation could just skip running them at all.
That's a good point, and a shame as well. Hopefully they both get added to the language at some point and we can move towards something like this.
> It just doesn't make sense to duplicate the garbage collection
This is reinventing the wheel so that you have the ability to ensure the wheels you are using are actually round.
While there are potential performance benefits in using a provided-system, something like Garbage Collection is a pillar upon which most of the program rests. I wouldn't trust the behaviour of all of the existing implementations to seamlessly work the same way reliably. If you discover a flaw in the system, being able to fix it can allow you to continue working. That is much better than filing a bug-report and waiting for the development cycle to churn out a released fix.
This is reinventing the wheel so that you have the ability to ensure the wheels you are using are actually round.
While there are potential performance benefits in using a provided-system, something like Garbage Collection is a pillar upon which most of the program rests. I wouldn't trust the behaviour of all of the existing implementations to seamlessly work the same way reliably. If you discover a flaw in the system, being able to fix it can allow you to continue working. That is much better than filing a bug-report and waiting for the development cycle to churn out a released fix.
If we're going to get into metaphor then I'd like to point out that running a garbage collector on top of another garbage collector is a bit like driving a car on a treadmill mounted on a truck. ;)
It would get really interesting if someone modified LuaJIT to emit asm.js code.
In this case, the Lua code could be even faster than the equivalent JS code on some applications where LuaJIT generates better code than V8/IonMonkey (even considering the 2x slowdown of asm.js wrt native).
The relevant yo dawg joke would be "I heard you like JITs so we put LuaJIT on your OdinMonkey so you can JIT while you JIT"
In this case, the Lua code could be even faster than the equivalent JS code on some applications where LuaJIT generates better code than V8/IonMonkey (even considering the 2x slowdown of asm.js wrt native).
The relevant yo dawg joke would be "I heard you like JITs so we put LuaJIT on your OdinMonkey so you can JIT while you JIT"
In the spirit of a vm in a vm, has anyone successfully compiled emacs in emscripten yet? I ran a quick test yesterday and it didn't go cleanly.
The benchmark[1] includes scimark, which measures FLOPS. I get 0.62 MFLOPS. In scimark running in the JVM[2], I get 353.4 MFLOPS (units explained here[3]). So in conclusion, the Oracle JVM on x86-32 is 570 x faster than lua.vm.js on V8 on same platform.
[1]: http://kripken.github.io/lua.vm.js/lua.vm.js.html
[2]: http://math.nist.gov/scimark2/run.html
[3]: http://math.nist.gov/scimark2/faq.html
[1]: http://kripken.github.io/lua.vm.js/lua.vm.js.html
[2]: http://math.nist.gov/scimark2/run.html
[3]: http://math.nist.gov/scimark2/faq.html
I assume the JS code generated here is using heavily the asm.js subset, so it can be run faster on asm.js-capable browsers.
Therefore V8 is not the ideal platform for benchmarking this implementation. You could try with a nightly Firefox.
EDIT: Indeed, 8 MFLOPS on FF Aurora, 1.2 MFLOPS in Chrome Canary.
Therefore V8 is not the ideal platform for benchmarking this implementation. You could try with a nightly Firefox.
EDIT: Indeed, 8 MFLOPS on FF Aurora, 1.2 MFLOPS in Chrome Canary.
What do you get in the JVM?
Did Lua's use of setjmp and longjmp for error handling pose any particular challenge? I always thought that was an odd aspect of Lua's implementation.
setjmp/longjmp works in emscripten, it does prevent some optimizations in functions where it is called though. But luckily it doesn't seem like it is called in anything performance-sensitive in the benchmarks I've tested.
Not sure if this is related, but you often have to declare local variables volatile when using setjmp/longjmp because they do not preserve registers. It's one case where unoptimized C code will run fine, then -O will push variables into registers and break it.
I wonder how difficult it would be to convert Lua to C++ with exceptions.
I wonder how difficult it would be to convert Lua to C++ with exceptions.
This isn't an issue for Lua, because so much effort has gone into making it extremely portable. Regardless, a port to C++ would be possible, though I'd (without thinking about it much) guess it would take a performance hit due to the increased code size from to the EH tables causing icache misses.
setjmp() isn't particularly rare; libpng also uses it for error reporting, for example.
With asm.js in Firefox nightly it's nearly native speed.