WebAssembly
Nick Desaulniers - @LostOracleToday's topics
- What is it?
- Why should I care?
- Hasn't this been done before?
- Does it allow us to run <favorite language> in the browser?
- Why not use LLVM bytecode?
- Where can I learn more?
...but first, a story about Instruction Set Architecture and processor design.
What is WASM?
A bytecode explicitly designed for portability, stability, compressability, security, and minimal nondeterminism.
Why should I care?
Loading large amounts of asm.js code or using a language other than JS in the browser.
Otherwise, WASM might not be of any interest to you, yet.
What's asm.js?
A type stable subset of JavaScript that can be compiled ahead of time by existing JS engines.
Luke Wagner
The key strength of the web is safe execution of remote code.
Hasn't this been done before?
All prior attempts taught us things like:
Sandboxing is a good idea.
We want something that is easily integrated across all browsers.
How is WASM different?
Start Up Performance
30% smaller than gzipped asm.js
Can be parsed 23x faster than large asm.js modules
View Source
Like LLVM IR, you'll be able to convert binary to text.
No need for unary plus and bitwise or operators (asm.js type hints).
Text can be shown in dev tools, source maps can show original source.
No text representation today, still working on necessary opcodes.
Working with DOM APIs
Not replacing them.
Can we run other languages in the browser?
Translation vs runtime
Two approaches:
- Translate one language, statement per statement, to JS.
- Prone to correctness bugs.
- Compile runtime from C/C++ to JS with emscripten.
- Binary bloat, GC coordination, ABI.
GC coordination
We want to dynamically create DOM elements from a managed language and have its GC be in charge of lifetime.
Who cleans up around here?
"Supporting multiple VMs is a big infrastructural challenge ... You'll need to make sure that the different GCs can coordinate ... Otherwise you will either not have an story of object reclamation, or you'll have memory leaks ... Doing this right is hard ... See [my paper and another which found] >5% throughput regression." ~ Filip Pizlo
What's an ABI?
a contract that two binaries compiled possibly from two different compilers at two different points in time can interoperate
ABI
Need to agree upon:
- size of types (ILP32), padding of structs, alignment
- calling convention (sending & retrieving values from function calls)
- binary representation (Mach-O, ELF, PE32+)
- function name mangling
Why not reuse LLVM bytecode?
That's a lot of LLVM IR, let's just use that!
| wasm | llvm ir |
|---|---|
| portable | has architecture specific opcodes |
| stable | breaking changes between versions |
| small encoding | not a design goal |
| fast decoding | not a design goal |
| fast compiling | not a design goal |
| minimal nondeterminism | nondeterminism and undefined behavior |
"... LLVM IR is a poor system for building a Platform ... LLVM isn't actually a virtual machine. It's widely acknoledged that the name 'LLVM' is a historical artifact ... LLVM IR is a compiler IR." ~ Dan Gohman
WASM is more of an ISA than a compiler IR.