Ladybird WebAssembly Update #3

It's been a great few weeks since my last Ladybird browser update, where I'm focusing on the WebAssembly implementation, LibWasm. A few months ago, I set a goal: get LibWasm to pass the entire Wasm 2.0 specification test suite by the end of the summer.¹

I'm happy to announce that we've done it! As of PR #851, there are zero failing Wasm spec-tests.

It's been an amazing journey, and I'll go over the specific progress we made in this post!

Improvements

A lot of my effort these past weeks has been directed towards SIMD, or rather, WebAssembly's version of SIMD.

I implemented/fixed the many of the Wasm SIMD instructions, including:

Load and store (PR)
Min/max (PR)
Comparisons ( $\lt$ , $\ge$ , etc.) (PR)
Integer conversions (PR)
Float conversions (PR)
Bit shift right (PR)
Shuffle and swizzle² (PR)
A bunch more miscellaneous ones

iNxM PR Co-Author

I co-authored a large PR that implemented the vast majority of the SIMD integer operations. The original PR was written by Enver Balalic, but the author ran into some annoying issues caught by UBSan, which I decided to try to help fix after a few weeks.

It felt fantastic to check out someone's in-progress work and help out! Somewhat anticlimactically, the undefined behavior was caused by C++'s implicit integer conversions, which I've become more and more annoyed by over time...

gcc Weirdness

Getting the Wasm SIMD testsuite to pass uniformly across macOS, Linux, clang, and gcc was quite a chore. One such issue arose when implementing the swizzle instructions.

Essentially, gcc has access to an intrinsic called __builtin_shuffle. Our goal is to use this intrinsic for our SIMD swizzle instructions. However, __builtin_shuffle and Wasm's swizzle actually have different behavior in edge-cases!

A swizzle in Wasm should look like this:

(i8x16.swizzle
  (v128.const i8x16 10 241 -3 0 1 -15 3 23 24 25 90 244 180 4 77 31)
  (v128.const i8x16 15 1 2 3 4 5 6 7 8 9 10 11 12 13 14 0))
;; This will return a vector with the first (index 0) and last (index 15)
;; elements swapped!
;; (v128.const i8x16 31 241 -3 0 1 -15 3 23 24 25 90 244 180 4 77 10)

When index vector has indices that are out of bounds, (i.e. -10 or 18), Wasm says to put a 0 in the output vector, like so:

(i8x16.swizzle
  (v128.const i8x16 10 241 -3 0 1 -15 3 23 24 25 90 244 180 4 77 31)
  (v128.const i8x16 -10 16 100 -1 87 99 18 -1 -1 -1 -1 -1 -1 -1 -1 15))
;; All zeroes except for the last element:
;; (v128.const i8x16 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 31)

How does __builtin_shuffle work? It wraps those invalid indices modulo 16 and proceeds from there, which is very different behavior!

Since I'm on macOS, clang is an unescapable default,³ and clang doesn't have the __builtin_shuffle intrinsic! I tried to compile Ladybird using gcc, but ran into a ton of problems, so fixing this was a little difficult.

The fix wasn't too difficult itself, but actually getting gcc working was.

Optimizations

I optimized the parser quite a bit, making it fast enough to instantiate spidermonkey.wasm (a 5.4mb file) in 140ms, compared to the initial 450ms.

I'm going to write a dedicated post about this in the future, so stay tuned for that!

Going Forward

It's taken a good amount of work to the point of full spec-compliance, which was made possible by Ali, who maintains LibWasm and gave feedback on a lot of my PRs. But, even though my goal has been completed, I'm nowhere near done hacking on Ladybird!

Right now, complex Wasm websites are nearly unusable because of how slow the bytecode interpreter is. In the next few weeks, I think I'll focus my efforts on that front!⁴

Also, I've recently been branching out a little bit into other areas of the browser. I implemented the :has() pseudo-class in this PR. More of that work is to come!

You can read my original post for more details! ↩︎
Fun fact: both shuffle and swizzle instructions re-order the lanes of a vector based on a set of indices. The difference is that a shuffle's indices are statically defined! ↩︎
The gcc command on macOS is actually just a symlink to clang... ↩︎
My integral calculator website does work, though! ↩︎