Why does malloc take a byte count, just to divide it by sizeof(u32)? Why not embrace the fact that natural "word" in the JavaScript VM is a boxed value that stores an int, float, or string, and define sizeof(u8) === sizeof(u16) === sizeof(u32) === 1? Unless you really want each cell in your memory arena to literally emulate a byte and always store a value between 0 and 255?
In a similar vein, why not use the native JavaScript object type to back the structs? This will cooperate nicely with the inline caching in typical JavaScript JITs. The result will be native assembler that indexes directly into the structs in the way you'd hope, since your static typing system forces every object reference to be monomorphic.
Does this language actually eliminate GC pauses? Isn't the GC still going to run, except now it has to walk every single cell in your memory arena every time, defeating the very heuristics that GCs use to reduce pauses?
There's definitely something alluring about the static typing, but I'm less sold on the manual memory management. Maybe there is a way that the language could help you generate less garbage, without actually requiring you to manually control the lifetime of each object?
From your linked-list example, it's eminently clear that it needs a C++-style template system and a port of the STL :) If you wrote a C-to-*JS translator, you could hypothetically use a pre-existing cfront (C++-to-C translator) and STL implementation to get that working.
> Does this language actually eliminate GC pauses? Isn't the GC still going to run, except now it has to walk every single cell in your memory arena every time, defeating the very heuristics that GCs use to reduce pauses?
Yes, this eliminates GC pauses.
The JS GC will only run if allocations occur, allocations in the sense of actual JS objects. If you do malloc which amounts to manually handling indexes in a single array, no JS objects are created and destroyed.
Also, JS GC's would not walk the memory arena. The memory arena is just an array of numbers, it does not contain native JS references which is what the JS GC traces.
I meant the synthetic memory arena, the JS array that malloc() and free() manage. That certainly could contain JS references and needs to be examined by the GC.
I don't think you can actually avoid GC by avoiding JS object allocations. Any form of string manipulation, any form of IO (event handling, DOM manipulation, XHR), any use of timers, etc, is going to allocate objects, right? Also, the JS runtime could generate garbage internally. For example, since JS has closures, it wouldn't be totally unreasonable for the runtime to manage the lifetime of activation records (JS stack frames) using the GC.
The memory arena is a typed array, which cannot contain object references and so does not need to be scanned by the GC.
If you keep JavaScript object allocations to a minimum, then GC pauses will be short and infrequent. This could actually be useful to write a game render loop or any other code that can't tolerate pauses.
Managing JS stack frames using the GC would be very bad for performance. You almost never need to do that unless your language has something like Scheme's call/cc. You may need to promote closed-over variables to the heap, but you don't need to migrate the entire stack frame there.
As modeless said, the memory arena is a typed array. It contains numbers, not JS references. The GC does not look at the contents of typed arrays at all.
String creation does create JS objects, yes, as does creating closures. It is hard to avoid garbage entirely, that's true.
In practice, most safe languages provide back doors and foreign interfaces to unsafe code. Look at Obj.magic in Ocaml, or unsafePerformIO in haskell, or ffi/unsafe in Racket.
But on the web, you obviously can't tolerate true unsafe back doors. This project gives you something that is "unsafe" in that you can read or write to parts of the allocated "heap" buffer that you aren't supposed to, but it's not possible for that to violate JavaScript's safety. And the compiler is generating extremely optimizable code, so it JITs very, very well.
On top of that, the fact that it's all implemented on top of a simple API (typed arrays) means that you could compile in a safe mode that does extra checks, or even trap all memory traffic and build valgrind-like diagnostic tools. It's starting out as a virtualized abstract machine, which means it's easy to hook into and provide great development tools.
I probably wouldn't reach for this kind of dialect for most purposes. But for very performance-sensitive kernels, it's got promise. But it's an experiment, so we'll see!
There would always be the option of running the code out into a raw VM host using NaCL or Xax style container (I know Mozilla dont see these approaches as necessarily compatible with the open web). But to take an approach like boot to gecko to the next level and being compatible with low power usage, I dream of running near native code - inline SSE/Neon with a fallback to generic code.
Sure! Right now there's just Michael's GitHub repo (this is literally only a few weeks old) but I've created an IRC channel -- join us at #lljs on irc.mozilla.org. You may have better luck getting attention during the work week, though. ;-)
Why not extend the typed array API to include SIMD operations on multiple array elements, we can compile to use those and special case them in the JIT.
I'm also thinking how the vm/language could self extend in the field. How the SIMD extensions (or any feature) could be introduced a-posteriori or independent from browser release cycle.
Cool. Does anyone else write their code with the * after the type, instead of before the variable? Like int* x; as opposed to int * x; to me int* is the type of x, so keeping it together makes sense. Like a function returning an int pointer would be - int* myfunction(); or type casting something (int* ). Putting the * before the variable and not as part of the type just seems unintuitive. Anyone else agree/disagree?
edit: another point - putting * before the variable makes me think of a dereferencing operation, and that's part of why putting it before the variable name in declaration is unintuitive, and very confusing to people learning C.
This style rests on the premise that C declaration syntax is:
<type> <space> <name>
It isn't, and using the style is misleading people to think that it is.
You can't write:
int[10] a;
You write:
int a[10];
Basically any type that involves more than a base-type and a pointer is going to fail to work with that style, and then you will lose uniformity as well.
There's actually a pretty good reason for the way C syntax is: The declaration and use syntax are virtually identical, and the precedence rules are identical. It's nice to only have to learn that once.
The fact the * before the variable looks like dereferencing is intentional, you're supposed to read this:
int *x;
as: "Declare the dereference of x to be an int".
And this:
int (*(*x)(void))[10];
As "The dereference of x is a function of void, the dereference of which is an array of 10 ints".
The declaration and use syntax of pointers is not 'virtually identical.' int* a; could be used to set int* b = a; or int c = * a; or int* d = &a;
Also for arrays it's unfortunate the syntax is like it is, because int a[10];, is of course a pointer and would be a bit more intuitive in the form of int[] a = new int[10];
Also you say * before the variable 'looks like dereferencing is intentional.' but is it really? what's your source?
defines an array, not a pointer. typeof(a) is int[10], and not (int ptr). The thing is, whenever you take the value of an array in C, it is degraded to a pointer to the array's first element, and this confuses people to no end that arrays are pointers.
For example, if we declare:
int a[10];
int *b;
Then sizeof(a) and sizeof(b) are very different. typeof(&a) is pointer-to-int[10], whereas typeof(&b) is pointer-to-pointer-to-int.
void f(int (*x)[10]) { ...
}
If you call:
f(&a); // will type-check
f(&b); // will not
So the syntax for declaring arrays is fine, and when you declare an array, no pointer is being declared at all.
This is C syntax for type declarations:
<base type> <type declaration here>
The "<type declaration here>" part uses virtually the same syntax as use syntax.
For example, if x is a pointer to a function that returns a pointer to an array of floats, you could reach a float via this syntax:
(*(*x)())[1]
You could declare x using this syntax:
float (*(*x)())[10];
The main differences are:
* Array size instead of index
* There's no requirement to dereference a function pointer to call it, but there is a requirement to use the dereference syntax to declare it.
* Array values can be used as pointers to their first elements (due to the automatic degradation), but must be declared as arrays.
Yes, it is unintuitive. However, there's a reason people do it the unintuitive way, and that's due to the C spec.
float* foo, bar;
This code declares bar as a float, not a pointer! To make things less misleading, most C programmers have adopted the convention of putting the star next to the variable name instead of the type:
I do what you're doing. I do this not because it accurately supposes what C's syntax is (it doesn't; however that doesn't matter because I follow the 1 pointer per line convention), but because it was most common in codebases I've worked on and the notation commonly shown in school.
The emphasis with 'int *x' is that the variable is no longer really a variable anymore. It is a pointer to your variable. It doesn't really have anything to do with the type.
Yes it is still a variable, a variable of type int* , not int. Which can be returned by a function that returns type int* or type casted int* . Variables of type int are assigned numbers, while variables of type int* are assigned memory addresses.
second edit: there's a link on reddit to the c++ coding standard for the joint strike fighter and it says pointers should be written like int* x; for exactly the reasons above. pretty cool - http://www2.research.att.com/~bs/JSF-AV-rules.pdf see section 4.9.6
What I mean to say is, if T is a type, then a pointer to T is certainly different from T, but a "pointer to T" isn't a new type. The C type system includes the single type "pointer."
* C/Javascript/Lisp on an augmented LuaJit (structs, efficient array slices, machine word types).
* dual syntax interchangable S-EXPR and C{}.
* lang(machine) { instr over machine } blocks. Where machine is either a syntactic transformation or a way of hosting binary code in another language and interchanging objects.
* raw continuations and stack management.
* explicit and implicit memory management.
* language profiles to allow assert() and compilation failures based on feature set used to allow different restricted forms of the language to be used according to the position in the software stack the code is targeting (embedded, client, server, enterprise boilerplate). eg. profile(no-gc, no-lang, must-types)
* full ability to specify code to execute and participate at: edit/view, compile (pre-process, parse, code-gen etc), link, load and runtime (debugging, re-jit, trace analysis).
Well, the * was supposed to be a pointer. JS is dereferencing JS, so JS would be more appropriate since it implies static typing and pointers, but it sounds awkward when you say it, so we went with *JS for now.
Just a temporary name. We're tossing around name ideas. Michael's a true hacker. He writes code first, asks questions later. I'm more useless, so I like to think about names. ;-)
Since the IRC channel is called 'lljs', then 'low level javascript' is a fairly functional - if not totally imaginative - name.
On a more bizarre level, use the same logic as coffeescript (java - coffee), and get something like 'C-matra' (Sumatra being another Indonesian isle, like Java).
Not really into high speed javascript, but wouldn't you be able to achieve the same kind of memory reuse by simply using factory methods patterns that would reuse object instance from a pool instead of reallocating ?
It's the same kind of tricks we use in ios / android programming when we want to avoid memory allocation or garbage collections...
Using the pattern you describe you could only reuse the same kinds of objects. With a general heap you could reuse the memory for arbitrary structures.
Objects in JavaScript are not cheap, they need to carry
around lots of extra information and can be many times
larger than their C style counterparts, moreover property
access is slow.
It's easy to say "X is not cheap, Y is slow".
It's much more valuable to actually prove it with numbers that show where it's fast and where it's slow.
In my experience, object creation and property access is almost never the bottleneck, and GC pauses only are relevant (in v8 at least) if you're leaking a large number of objects, especially where those objects are of mixed generations. (Ie, have a lot of long-lived and short-lived objects that refer to one another and occasionally leak.) That is, if you are getting hit with long GC pauses, then it's worth re-thinking your design and tracking down objects that may be leaking.
Newer JS engines (ie, the ones that implement TypedArrays) already highly optimize object creation and property access. The benefit therefor seems slim and highly niche. Looking at the generated code, it seems like the debugging cost is going to be very high, and it drastically increases the number of lines that the programmer has to write, which will increase bugs.
I could see the value in certain niche situations, but I would really like to see the performance characteristics explored more fully.
Yeah, I probably wouldn't have worded it that way. I think the better way to look at it is, it's hard to know how to predict the performance of JavaScript, given the complex, dynamic, heuristic optimizations performed by modern engines. This project is better thought of as an attempt to build a dialect of JS that can be more easily tuned for performance, because its performance model is simpler than that of JS.
If I were going to try to make that claim more precise, I might do some experiments to demonstrate the high variance of object performance in modern JS engines.
I could see the value in certain niche situations...
That's really the idea. This isn't meant to be an alternative to JS, but rather a tool for performance-sensitive kernels. Interoperability will therefore be key, because it has to be easy to write just a small component of a larger app in this dialect while smoothly integrating into the rest of the app.
...but I would really like to see the performance characteristics explored more fully.
Perhaps. I'm skeptical of our ability to accurately measure general claims like "objects are slow" or "X is easier to program in than Y" or "*JS is easier to performance tune than JS." But I do think it's easier to evaluate those claims, at least informally, after you've built something you can experiment with. Hence this experiment!
It's easy to say "situation X does not occur often".
It's much more valuable to actually prove it with numbers that show how often it is used across the sum of problem domains out there that you might not be aware of.
We are all solving different problems. I'm running in memory stream processing. Objects in Javascript and associate GC overhead are a massive hindrance in my domain.
This! There are no benchmarks given or otherwise proof that this is faster.
With the example given of the linked list, it's claimed that the JS version is slow. However running it myself, I find their JS version actually runs faster in FF and Chrome then the JC one.
Think of it more like C (really more like a hybrid of C and JS) compiling to JS. It runs on the web, but it compiles to a very stylized kind of JS: it allocates a large typed array that represents a C-like heap and does its memory access on that typed array. This is the same basic approach that Emscripten uses for running C/C++ programs.
The purpose of this project is to experiment with a dialect of JS that integrates well with regular JS, so you can write performance-sensitive parts of your application in this dialect, and they can interoperate smoothly with the other parts of your normal JS code.
Cool! For a while I've been pondering a similar idea with the goal of cross-compiling the language to idiomatic C and JavaScript. The JavaScript translation would necessarily be less than perfectly idiomatic, but it would be more idiomatic and have better performance and smaller size than the code generated by Emscripten. The intent was to use it for high-performance applications like games.
Serious question. Do you browse the web with Internet Explorer?
V8 and the new Mozilla engine are optimized to hell already. And they compile to native code. http://kailaspatil.blogspot.com/2011/08/jaegermonkey-archite... And they are efficient.. These engines have proven that you don't need pointers or explicit types to be an efficient Javascript implementation.
The only reason that I can see for these types of features like explicit types and pointers are to support systems programming, where they are required for certain activities (probably less that you would think though). For that, I would really love to see a CoffeeScript with those features available for use optionally. I think to do that you would want to find a way to compile directly to machine code or assembly or LLVM bytecode or something like that.
There is still a lot of performance slippage out there in high level languages (eg look at the need for assembly when people write raw codecs in C - x264 etc). We are rapidly approaching the battery wall on mobile devices and the clock wall is already here. ISA's may be extended to support dynamic languages more efficiently (tags, direct uop translation from user specified instructions) - but right now at a given TDP static execution is greener for many applications than dynamic, and history has shown us cpu vendors are unwilling to go down this path.
Javascript is anything but efficient right now - some code paths are fast but with a high startup cost. Look at the startup problem with Chrome/V8 for instance, or watch a nodejs application hit the GC wall with 800MB of live data.
There are times when static is faster than dynamic and vice-versa. Times when heavy upfront compilation is better than incremental run time analysis and vice-versa. As the clock wall, TDP wall and process communication wall hits we need all the tools available to exploit maximum performance.
I'm proposing to compile as much upfront as possible, more than the default for those engines. But for something like JavaScript, without the types specified or inferred in all cases, you can't necessarily compile everything upfront. Type inference would be much better than requiring manual specification of all types.
I think that it is just much better software engineering to improve the GC and JIT compilation rather than to code all of the memory management and types manually or use pointer tricks. If you are building a codec or critical part of an operating system then you may need assembly or well-defined types for static compilation.
It would be nice to have the memory management and other features available for when you need them but I don't think they should be the default.
Anyway I think it could actually be useful to find ways to remove the separation between assembly-level coding and higher level coding. For example, if I were writing a codec in CoffeeScript, I could would probably write something like interleave.highBytesFromQuads rather than PUNPCKHBW.
I agree totally about removing the separation between assembly level coding and higher level coding.
With current ISAs no matter how much genius is thrown at GC/JIT its always going to yield a layer of overhead. The pipes are a fixed with, the caches a fixed size - the plumbing is static.
Below this floor a thinner abstraction will yield greater performance. The thinner abstraction is useful to implement GC/JIT. Any language that wont let you bust outside of the GC heap is always going to hit a pain point.
Until a language can self host with no significant efficiency loss there will always be another language to wedge under it. Until a platform has a mechanism to expose its raw feature set up to a hosted language we will forever be in a world of software rasterizers, sluggish Java UI, where software developers re-implement functionality that highly tuned hardware pipes already provide.
Better to be able to write using interleave.highBytesFromQuads and better to able to include your implementation of this using a punch thru to PUNPCKHBW where available or an emulation where not. I guess its the old argument of high level interfaces versus low level. Useful high level abstractions appear over time, but without access to the low level we cannot experiment and build them on a rapid cycle.
I'm sure some OS vendors would like to keep the browser crippled - because in the natural end game their OS doesnt need to exist as an expensive product. Its good to see Mozilla pushing the boundaries.
JS JITs are wonders of modern engineering, for sure. But there's a weird paradox in that as they get smarter, it doesn't get easier to write optimized code, it actually gets harder! Nobody really understands how to write optimizable code. Nor, in some sense, should they. But I think there's a real need for subset languages with more predictable performance models than JavaScript, that have compilers tuned for the kinds of optimization done by modern JS engines.
One nice description of this general problem was by Slava Egorov:
As a C programmer I like the concept but debugging is going to be a nightmare. Look at how pointers translate to Javascript. I don't foresee being able to debug anything with code like that in Chrome Developer Tools.
The code is still very young (less than a month old). Debugging in C is done by printing hex values and using valgrind, so debugging here can be done by printing hex values (array indexes in the heap or stack) and creating a valgrind that runs at about the same speed as a non-valgrind session, since valgrind can be implemented in the js engine.
Actually, you could build something like Valgrind quite easily. You can replace the underlying heap with a JS proxy and trap all read/writes, it would be a fun project.
I was skeptical at first, but I've been writing so much Go lately that the struct syntax, typing and pointers would be lovely to be able to utilize in a web app without having to go so extreme as to use Dart.
In a similar vein, why not use the native JavaScript object type to back the structs? This will cooperate nicely with the inline caching in typical JavaScript JITs. The result will be native assembler that indexes directly into the structs in the way you'd hope, since your static typing system forces every object reference to be monomorphic.
Does this language actually eliminate GC pauses? Isn't the GC still going to run, except now it has to walk every single cell in your memory arena every time, defeating the very heuristics that GCs use to reduce pauses?
There's definitely something alluring about the static typing, but I'm less sold on the manual memory management. Maybe there is a way that the language could help you generate less garbage, without actually requiring you to manually control the lifetime of each object?
From your linked-list example, it's eminently clear that it needs a C++-style template system and a port of the STL :) If you wrote a C-to-*JS translator, you could hypothetically use a pre-existing cfront (C++-to-C translator) and STL implementation to get that working.