WebAssembly 2.0 First Working Draft
w3.org
w3.org
Exciting to see a major release, though admittedly it seems a little arbitrary if there aren't any breaking changes (and very problematic if there are?).
like say you wanted to make a J2534 DLL http://www.drewtech.com/support/passthru.html
How would you export these DLL functions or even call functions from other DLLs (dlopen/dlsym, etc.)
dlopen/dlsym support is still in the works (for example, see [0]). You can't call function pointers directly from WASM, naturally, but it is possible to wrap dlsym so you can use it from a WASM program. There just isn't a standardized/pre-made way to do it yet, so you'd have to write the glue code yourself.
[edit] The Emscripten WASM toolchain seems to support dlopen/dlsym, see: https://github.com/WebAssembly/tool-conventions/blob/main/Dy...
I don’t know if there’s a better intro article. I believe this is the current iteration of the proposal: https://github.com/WebAssembly/interface-types/blob/main/pro...
Especially the last comment by Alex here seems interesting to me – on one hand the observable result of the execution changes for the same program. On another, who is going to bother writing such programs? Should WebAssembly specification be allowed to assume nobody does and make changes such as these?
Since the original release 1.0 of the WebAssembly specification, a number of proposals for extensions have been integrated. The following sections provide an overview of what has changed.
IIRC tables are used to communicate things like function pointers with the host executor, maybe like a vtable but more general.
Multiple values: Generalized the result type of blocks and functions to allow for multiple values; in addition, introduced the ability to have block parameters
https://github.com/WebAssembly/spec/blob/main/proposals/mult...
Reference types: Added and as new value types and respective instructions
Table instructions: Added instructions to directly access and modify tables
Multiple tables: Added the ability to use multiple tables per module
Bulk memory and table instructions: Added instructions that modify ranges of memory or table entries
https://github.com/WebAssembly/spec/blob/main/proposals/refe...
https://github.com/WebAssembly/spec/blob/main/proposals/bulk...
Vector instructions: Added vector type and instructions that manipulate multiple numeric values in parallel (also known as SIMD, single instruction multiple data)
https://github.com/WebAssembly/spec/tree/main/proposals/simd...
Sign extension instructions: Added new numeric instructions for performing sign extension within integer representations.
https://github.com/WebAssembly/spec/blob/main/proposals/sign...
Non-trapping float-to-int conversions: Added new conversion instructions that avoid trapping when converting a floating-point number to an integer
https://github.com/WebAssembly/spec/blob/main/proposals/nont...
externtype ::= func functype | table tabletype | mem memtype | global globaltype
"During this phase:
"One or more implementations proceed on prototyping the feature to the point that a comprehensive set of tests can be added.
"A test suite is added. These tests need not pass the reference interpreter at this point, but should pass on some implementation."
https://github.com/WebAssembly/meetings/blob/main/process/ph...
Here's the proposal:
https://github.com/WebAssembly/gc
And you can see it's being actively worked on:
Right now wasm is really designed for C/C++/rust
Aren't all of them single-threaded, though? That is, they only work in a VM with a single thread which the GC shares? Since WASM is supposedly finally bringing real multi threading to the web, how would that work? It seems like you'd need a new GC for WASM rather then just repurposing the JS GC.
The problem with bringing your own GC isn't just that, though. It's that using linear memory for everything, you're forced to put external references in a table, which roots them. Tables aren't weak...and even if they were, it's possible to set up a cycle between the linear-memory GC references and external references so that leaks happen. This was a problem in other contexts, for example, in V8 embedded into Chromium, and the ultimate result is that you need cooperative garbage collection across two heaps, with queues. While V8 and Chromium trust each other, both not to screw up, and to be fast, it's hard to see how to make that cooperative garbage collection contract work with untrusted Wasm code.
Anyways, that's why there needs to be support in WebAssembly for GC, because there needs to be a safe way to walk the stack (aside from the obvious JS interop considerations).
It'll work as promised when I can import a C# class in my JS code, instantiate it and then pass one of its methods to a Python module and invoke it from there.
Great times are ahead of us.
Outside of the browser, WebAssembly is still a shadow of CLR capabilities, including sandboxing.
As for the "secure" in WebAssembly, I advise a good read about the security section, specially memory corruption
It was also comically slow and worked only in IE.
I'm well versed in security of Wasm. Much better than anything else available today or in the past, obviously still not perfect.
One of the best features of Wasm is entirely social - seems like everyone has agreed on it, finally. That's enough for me even if it was a 1:1 copy of JVM or CLR.
Lets sell old stuff as something new, never done before, rewriting history.
More recent chapter, application servers with WebAssembly, what a great idea!
And people have seen what was in the past and created a modern, well-composed solution that is accepted by all major players. Excellent if you ask me. Yeah nobody has invented a new wheel here - but that's not necessary, actually it might be counter-productive to the goals of Wasm. Wasm wants to take stable, well-known ideas, improve upon the warts of previous tech like CLR and JVM and put it on 100 billion devices.
Unfortunately, the current GC design for WebAssembly doesn't support interior pointers which is going to be difficult to work around (but I don't think it's impossible). Interior pointers are normally required in Go.
Opens up the target to an additional subset of devs.
I found one of Daniel Lemire's [0] from 2018 which seems to suggest it's basically not much faster. 2018 is long enough away from now though that this post may not be relevant. I don't know.
It's odd to me how WebAssembly seemed to be started as a way to get performance improvements by working with cached instructions (and maybe also explicit memory management and integer types) but I don't hear much about performance these days.
Maybe I'm not paying attention.
Instead I hear more about it as a universal platform or a way you can run C/C++ code in the browser (like tree-sitter and CPython and Ruby, etc.). Or about people using it in edge compute or as the engine for smart contracts. In those ways you could just as well have used the JVM though.
[0] https://lemire.me/blog/2018/10/23/is-webassembly-faster-than...
What difference does it make for you as a user of a language that it compiles to webassembly than to JavaScript or any other VM?
From a user it should be transparent except for performance and FFIs.
Sure you could have implemented unsigned integer semantics on top of bool-arrays in javascript years ago but if the result is too slow you have gained because the compiled c++ code does not run fast enough to be usable.
I wish they were somehow better supported by JSON.parse though.
https://developer.mozilla.org/en-US/docs/Web/JavaScript/Refe...
I usually find it causes more problems than it solves, but I'm hopeful that someday WebAssembly can make compile-for-Web transparent enough that non-JS languages really feel like a great experience in the browser.
There are others however, such as Yew written in Rust, or Blazor in C#.
You are right though they are going to be the first big WASM / Canvas and WebGPU framework as opposed to the generations of JS / DOM. It will be interesting to see where they land with it, I’m quietly hopeful especially once you get AOM sorted in the browser too.
I think the real test is going to be when they decide to migrate over the multiple billion dollars of revenue project that Google Ads is from JS based Dart to WASM based Dart.
Work going on here but not much to look at now just source and no docs https://github.com/dart-lang/sdk/tree/main/pkg/dart2wasm
With SIMD and threads, Wasm can absolutely crush JS performance.
https://www.infoq.com/articles/webassembly-simd-multithreadi...
For example, if you want to implement an editable string, you have to do so using an array of smaller strings and hope the optimizer can figure it out, or slice() and re-join a single “immutable” string. Either way, your javascript code will be slower than the C/rust equivalent because the language is less expressive. Javascript has the same problem with b-trees, skip lists, ropes, gap buffers, AABB trees, and so on. You can implement all of this stuff in javascript - you just end up with much more memory indirection than you want. And for trees with different internal nodes and leaves, your code will give the optimizer indigestion.
In my experience, basically any of these data structures will run about 10-200x slower than their native counterparts. (If both are well optimized). To me this is the big performance advantage wasm has - that we can make high performance data structures.
I’m working on a CRDT in Rust. The fastest JS implementation I know of takes about 1 second to replay an editing trace from a user. In rust I can replay the same trace in 0.01 seconds (100x faster), by using a skip list. In wasm I can do it in about 0.03 seconds.
WebAssembly is less than 10%-20% slower than native code in our benchmarks and tests for algorithms like fast wavelet transform and bit packing.
WA allows us to aggressively optimize ahead of time when compiling, and be less sensitive to JS engine performance pitfalls around JIT optimization and garbage collection.
- The way deleted data is encoded in yjs is awkward. Versions store a list of which content has been deleted. So versions are bigger than they need to be, and you can't replay the editing history without stored snapshots. I think the knowledge of which elements have been deleted should just be saved into the binary format alongside the inserts.
- For javascript objects, yjs pretends each value is actually a list (with everything after the first element ignored) so it can reuse the logic for list editing. This is overcomplicated compared to Shelf or something like it.
- Yjs's internal structure is a linked list with cached offsets. This works ok in practice because inserts usually appear close to each other, but it means yjs has O(n^2) performance to process n edits rather than O(n log n). Diamond types (via wasm) is about 30x faster as a result - though as you say, yjs is still plenty fast enough in practice for most applications. And this gap will narrow once Yrs is working well.
- Yjs can have an interleaving problem when concurrent items are prepended at the same location in a list.
- The current (new) diamond types file format stores the origin-position + parents information instead of storing left-origin / right-origin fields. This lets DT have a smaller file size while storing more information than yjs stores. And I think this format is easier for applications to encode and work with, and it will make pruning easier when we get to it. We should also be able to load & save faster than yjs - but I'm not benchmarking that yet.
But all of this stuff is pretty minor. Yjs works well in practice, and its quite mature. Performance will only get better with the yrs rust implementation. For diamond types, I had the benefit of being able to copy some fantastic implementation ideas from Yjs. If diamond types is better than yjs, its because I'm standing on its shoulders.
I'm building a text editor right now. I'm just using a giant array with one element for each line. I anticipated that it would get slow, and that potentially I can push this into WebAssembly.
The problem is that you're going to need to get the data back out of WASM eventually, back into JavaScript and then back to the DOM.
I focused on other optimizations and strangely my text editor is able to keep 9 million loC file ( over 500mb ) in memory. The trick is to not render everything to the DOM, just the stuff people are looking at. That's the other bottleneck. V8 is surprisingly smart. Random access for a small section of that 9 million element array is still fast. You'd expect this array to be non linear, too. Moreover, even localized shifts are fast. I'm puzzled. For comparison, VSCode uses a more esoteric data structure called a PieceTree but for me becomes unresponsive at 500k LoC. To be fair, they have other sorts of overhead and processing layers to deal with, namely building an AST with potentially thousands of layer of hierarchical / code folding. Speaking of which, VSCode is built entirely in TypeScript and has a reputation of being a fast and snappy editor.
At any rate, I don't buy the argument that because Rust is fast and WebAssembly is comparably fast, then it makes sense to push your data structure, especially a text buffer, down into WebAssembly. The bottleneck is serializing data back and forth between WebAssembly and JavaScript. For text editors, I'm not sure this overhead is worth it, especially when V8 / Chrome is doing some crazy JS optimization behind the scenes. The appealing bit with WebAssembly is potentially that objects will have a smaller memory footprint, therefore we can push further than this 9 million LoC limit where the Chrome tab just runs out of memory, rather than having algorithmic slowdown.
Could you paint everything in a canvas element to avoid the DOM? That is being my approach in a small project, although it admitedly can complicate things.
For example:
- Ctrl+Left / right for moving a word at a time. This is Alt+Left/right on a mac I think. There's another keyboard shortcut to go to the start / end of a line, or the start / end of the input element. These shortcuts are OS-specific and can be overridden in the user's OS level settings - which you don't have access to from javascript. There's a mountain of shortcuts - like Ctrl+A, Ctrl+X/C/V, etc. They all vary per OS. You have to implement them correctly on every operating system.
- International character input methods. Eg, how would you type Japanese / Korean / Chinese characters into your text element?
- Undo/redo support
- Text selection via the mouse, the keyboard and on mobile by touch-dragging to select.
- Accessibility - like voiceover and dictation.
No matter how dedicated you are, I guarantee your custom canvas element will never correctly re-implement all the features already available in a simple HTML text input element.
I see it as an option for a heavy app in the browser. This is AFAIK similar to what Figma does (or flutter). Also, I am not referring specifically to text editors, although I could imagine something like vim working with this approach in the browser without trying to replicate an HTML text input element.
To store a buffer in a plain text editor, with no syntax highlighting? I'm not going to fight you on that. I'm glad pure javascript is fast enough for your needs.
But doing collaborative text editing with a CRDT adds some extra requirements that your text editor's buffer might not have. For example:
- Hydrating the document state on load. To cut down on file size, diamond types files currently just store the entire editing history. When you open them up, we replay the whole editing history into a buffer. And using Wasm + jumprope[1], this is <1ms even for very large documents. The same would not be true in javascript, and I might need to use bigger files.
- Merging concurrent changes. (Like, merging a long lived branch). This currently involves getting fancy with a B-tree. This is fast enough in rust. It would be orders of magnitude slower in javascript.
- Undo support. When you hit undo, you need to ignore any more recent typing from other users and just undo the local user's last change. This is subtle.
- When edits happen, we need to convert between your (line, JS column) tuple and whatever the CRDT uses internally. Diamond types specifies edit positions by counting unicode characters from the start of the document. So we need to convert between (line,col) and unicode offsets whenever local or remote changes happen to broadcast or apply the changes.
Rust + Wasm let me make all of these operations essentially instant. I can do that because I can use an appropriate data structure & algorithm for each operation.
And, sure, Javascript's Array is pretty fast. But if you ever need something more complex than Array, your performance will suffer massively. Personally I feel much better writing this code in rust + wasm.
I'd integrate diamond types with your text editor by just passing edits back and forth across the module boundary. I agree that it probably doesn't make sense to share a text buffer between javascript and wasm.
In this case, the bottleneck at 9 million LoC is not CPU cycles but memory usage. That's where I am considering pushing down into WebAssembly, but whatever function-inlining optimization V8 is doing, it is probably still faster than the overhead of JS<->WebAssembly function call. I have no doubt WebAssembly will edge out again.
> Merging concurrent changes. (Like, merging a long lived branch). This currently involves getting fancy with a B-tree. This is fast enough in rust. It would be orders of magnitude slower in javascript.
I guess my point is why do you need balanced trees? Is this a CRDT specific thing? Can you implement CRDT with just an array of lines / gap buffer?
> Undo support. When you hit undo, you need to ignore any more recent typing from other users and just undo the local user's last change. This is subtle.
From my understanding, the whole point of CDRTs or any command-based design is express changes to the state as the delta. In which case, you only have to stack / remember the set of commands and not have to store the state on every change. I'm not sure if this overlaps with the data structure choice, other than implementation details.
> And, sure, Javascript's Array is pretty fast. But if you ever need something more complex than Array, your performance will suffer massively. Personally I feel much better writing this code in rust + wasm.
I was just looking into TypedArrays. From my understanding, the JS runtime will use a HashMap or DoublyLinkedList if the array elements are homogenous. In the case where the JS runtimes know it is homogenous, they will fallback to those TypedArray / fixed length buffers that are "continuous" in memory. So on the JS side, Arrays can be as fast as TypedArrays thanks to the compiler. Consequently, in some benchmarks they will effectively be the same.
The question is whether the speed of a native array is faster than the speed of a TypedArray such that it pays for the WASM<->JS overhead. And I guess this depends on the application and the access patterns of that interop.
How often does this come up in practice? I can't think of many files I've opened which were 9 million lines long. And you say "LoC" (lines of code). Are you doing syntax highlighting on 9 million lines of source code in javascript? Thats impressive!
> I guess my point is why do you need balanced trees? Is this a CRDT specific thing? Can you implement CRDT with just an array of lines / gap buffer?
Of course! Its just going to be slower. I made a simple reference implementation of Yjs, Automerge and Sync9's list types in javascript here[1]. This code is not optimized, and it takes 30 seconds to process an editing trace that diamond types (in native rust) takes 0.01 seconds to process. We could speed that up - yjs does the same thing in 1 second. But I don't think javascript will ever run as fast as optimized rust code.
The b-tree in diamond types is used for merging. If you're merging 2 branches, we need to map insert locations from the incoming branch into positions in the target (merged) branch. As items are inserted, the mapping changes dynamically. The benchmark I've been using for this is how long it takes to replay (and re-merge) all the changes in the most edited file in the nodejs git repository. That file has just shy of 1M single character insert / delete operations. If you're curious, the causal graph of changes looks like this[2].
Currently it takes 250ms to re-merge the entire causal graph. This is much slower than I'd like, but we can cache the merged positions in about 4kb on disk or something so we only need to do it once. I also want to replace the b-tree with a skip list. I think that'll make the code faster and smaller.
A gap buffer in javascript might work ok... if you're keen, I'd love to see that benchmark. The code to port is here: [3]
> Undo support -> In which case, you only have to stack / remember the set of commands and not have to store the state on every change. I'm not sure if this overlaps with the data structure choice, other than implementation details.
Yeah, I basically never store a snapshot of the state. Not on every change. Not really at all. Everything involves sending around patches. But you can't just roll back the changes when you undo.
Eg: I type "aaa" at position 0 (the start of the document). You type "bbb" at the start of the document. The document is now "bbbaaa". I hit undo. What should happen? Surely, we delete the "aaa" - now at position 3.
Translating from position 0 to position 3 is essentially the same algorithm we need to run in order to merge.
> I was just looking into TypedArrays.
I tried optimizing a physics library a few years ago by putting everything in typedarrays and it was weirdly slower than using raw javascript arrays. I have no idea why - but maybe thats fixed now.
TypedArrays are useful, but they're no panacea. You could probably write a custom b-tree on top of a typedarray in javascript if you really want to - assuming your data also fits into typedarrays. But at that point you may as well just use wasm. It'll be way faster and more ergonomic.
[1] https://github.com/josephg/reference-crdts
[2] https://home.seph.codes/public/node_graph.svg
[3] https://github.com/josephg/diamond-types/tree/master/src/lis...
TL;DR: It depends. https://zaplib.com/docs/blog_ts++.html
Personally https://github.com/WebAssembly/component-model/pull/23 feels like it marks the start of webassembly's first full era. We need interoperation, webassembly needs to really be embeddable, & only then can we really understand & explore what webassembly means to the world.
(Rust got an early start with wasm-bindgen, doing all the lifting on their side, but it's not reasonable & only semi-interoperable having each language rebuild that bindgen system & abingor itself... hence canonical abi & component-model.)
Edit: tux3 has a good comment linking the changelog, noting that indeed numberous specs have been integrated. https://news.ycombinator.com/item?id=31087610
Really excited for multiple table per module support! The idea of a capabilities based sort of system, where a module is handed a couple working sets of data as SharedArrayBuffers to work g communicate with, is quite compelling.
I agree we that interface types/components will be a game changer.
And also that progress on all the important proposals (GC, interface types, linking, stack control, threads) has been painfully slow.
https://github.com/WebAssembly/proposals
You can see that "interface types" is at "Phase 1 - Feature Proposal (CG)", so still early days.
And there's a link to the proposal here, for people like me who don't know about it: https://github.com/WebAssembly/interface-types
I'm asking that because despite all the hype, wasm binaries are still 32-bit (at least in browsers), while 32-bit OS are being deprecated. 64-bit pointers are really good for some applications, with no performance degradation at all.
Does Wasm 2.0 support branch instructions needed for irreducible control flows? (without the relooper hack)
PNaCL also had a lot of limitations though I don’t remember what they actually were.
[1] https://www.w3.org/TR/wasm-core-2/syntax/instructions.html#m...
I fear they will run out of gas before the interesting features there are completed...
But it's not yet ready for mass adoption, yeah.
The only advantage of 64 bit pointers is that you can have a bigger address space than 4GB.
Although an abstract machine could have metadata with each pointer that isn’t part of the official bit size, I suppose.
That probably won't be the case with wasm64 unfortunately. A hidden advantage of wasm using a 32 bit address space while primarily targeting 64 bit platforms is that runtimes can get free implicit bounds checking by carving out a 4GB block of virtual memory for each wasm instance, so out-of-bounds accesses are trapped by the hardware.
That trick won't work for wasm64, so runtimes will have to go back to inserting explicit bounds checks at every memory access, and it'll be a tradeoff where in exchange for being able to access more memory everything will run a bit slower.
Would be cool to see it for more efficient implementation of functional languages.
For example, if you wanted to get involved in Kubernetes you'd go to KubeCon, and governance of the OSS project is pretty understandable?
What is the equivalent in WebAssembly land?
To attend the meetings, first join the W3C WebAssembly Community Group here: https://www.w3.org/groups/cg/webassembly, then email the CG chairs at webassembly-cg-chair@chromium.org to ask for an invite.
From there you'll get a sense of who folks are so you can pair names with faces when contributing to the various proposal discussions on the many proposal repos listed here: https://github.com/webassembly/proposals.
To get a sense of how things are run and decided, read the process documents here: https://github.com/WebAssembly/meetings/tree/main/process. The TL;DR is that the community group and its subgroups decide everything by consensus via votes during the meetings.
- new: vector types - 128-bit width FP and integer types and values for SIMD, with related instructions (and validation thereof) plus memory instructions for wide loads, and support in the runtime for wide values (consts, etc.), and in the text and binary formats
- value types (i32, i64, f32, f64) are now called number types
- new: reference types - basically function pointers and pointers to objects in the host, values of which can be stored in tables, with related instructions for accessing them and checking for null
- new: table instructions - get, set, size, grow, fill, copy, etc.
- element and data segments - not new but expanded definition and given more runtime support (added to the store, can be referenced by address like other runtime objects), plus new indices for referencing segments and their modes (basically, how they are initialized during instantiation)
- limited form of subtyping allowed when importing external values
- result types can be a sequence of values (i.e., multiple return values from functions)
- new instructions to avoid trapping on floating-point to integer conversions ("saturation")
"Generalized the result type of blocks and functions to allow for multiple values; in addition, introduced the ability to have block parameters:
- Function types allow more than one result
- Block types can be arbitrary function types"
[1] https://www.w3.org/TR/wasm-core-2/appendix/changes.html#mult...
The current solution that works only for Chrome and allows to use DWARF is an improvement but it has lot's of limitations: One of them is the limit of 4GiB (or 2GiB - don't remember) of DWARF file. Secondly - the entire browser tends to crash with that plugin on larger projects.
It's supported in Wasmtime, WAVM, and WABT and wasm2c, and not much else so far (https://webassembly.org/roadmap/).
Is it because Rust naturally leant itself to machine code, therefore it was easier to build a compiler?
Is it because the Rust community wants to create more "killer features" for the language and have it stand out?
Are there benefits to using Rust versus TypeScript as the host language?
https://news.ycombinator.com/item?id=26067118
There are probably specific advantages to Rust, but it’s also just nice to work in a language you like. I can compile Go to WebAssembly so I’m not going to write TypeScript. Your preferences may differ.
Edit: Actually it's quite viable as it is and I'm not sure how I did not think of it sooner. Just run everything in WASM and make requests to the database like WASM -> JavaScript -> IPC and latency is not a problem here.
No GC API ?
As far as I understand, something like e.g. a high-performance JVM would be hard to do in WASM due to its machine's Harvard-like architecture.
1. Figma uses Wasm extensively in the browser. 2. Unity ships their game engine to the browser using WebAssembly 3. Amazon Prime uses WebAssembly as a part of their OTT stack 4. A lot of Blockchain companies user WebAssembly to process smart contracts. 5. There are new start-ups lile Hotg.ai that use WebAssembly to deploy ML solutions to connected or semi-connected devices. (Hotg.ai uses Wasmer) 6. Shopify uses AssemblyScript and WebAssembly for their plugin system.
Python makes breaking changes whenever they want in minor version changes.
For now I've got webassembly disabled, and I don't see that changing