Bug 258302 – WebAssembly SIMD results in incorrect float values and non working bitwise operations

• Status: RESOLVED
• Resolution: FIXED
• Component: WebAssembly

[John Hankinson]

I've not yet been able to narrow down the exact instructions that cause the issues, but something is very broken. I've noticed incorrect float values after some arithmetic operations, and my suspicion is that something
with package integer bitwise ops and comparisons is not working as expected.

The same code works on Chrome/Android/PC/Edge and Safari on Mac.
(IE. everything else).

The following site demonstrates the issues:
https://www.terraspace.co.uk

For which the output is identical everywhere except iOS 16.x (since wasm simd supported landed).
Blocks of 4x pixels are not masking and updating as they should be.

[John Hankinson]

https://github.com/microsoft/onnxruntime/issues/15644

https://forum.babylonjs.com/t/webassembly-error-havok-project-in-old-ios-device-iphone-6s/40213

https://github.com/emscripten-core/emscripten/issues/19619

Are reporting, possibly related issues.

[Radar WebKit Bug Importer]

<rdar://problem/111050621>

[Justin Michaud]

I wonder if this is already fixed by https://bugs.webkit.org/show_bug.cgi?id=257842

[John Hankinson]

The description of the fix certainly sounds very similar to the issue encountered. Will this be landing in 16.6 (possibly in the beta)?

[Justin Michaud]

(In reply to John Hankinson from comment #4)
> The description of the fix certainly sounds very similar to the issue
> encountered. Will this be landing in 16.6 (possibly in the beta)?

Apple does not comment on the content of future releases.

I checked with a nightly though, and it seems that the bug still reproduces.

Thanks for the links to other issues, this is super helpful!

[Justin Michaud]

Is there any chance I can get source for a smaller test case that reproduces this, or source for your site itself? It would be super helpful for bisecting the root cause.

On my end, I have confirmed that in an nightly build, turning off our highest tier fixes this bug: 

```
265333@main % __XPC_JSC_useOMGJIT=0 ./run-webkit-archive
```

setting defaultB3OptLevel=0 or disabling wasm call inlining does not fix the bug.

Next I would like to bisect by function index.

[John Hankinson]

I've emailed you the WAT functions involved, the source is quite long, but I think we can narrow it down to a single function as they all perform similar operations. Interestingly, I have been informed, although cannot verify it myself, that this issue is not present on an M1 mac, which I would have expected to generate the same aarch64 output from codegen as the mobile soc?

[John Hankinson]

Let me know if there is anything I can help with re. the WAT functions or testing.

[John Hankinson]

Are there any updates on progress with this yet?

[Justin Michaud]

```__XPC_JSC_validateOptions=1 __XPC_JSC_dumpOMGDisassembly=1 __XPC_JSC_omgAllowlist=/Users/justin_michaud/Downloads/265333@main/allow.txt __XPC_JSC_useWebAssemblyOSR=0 __XPC_JSC_useConcurrentJIT=0 __XPC_JSC_maximumWasmDepthForInlining=0 __XPC_JSC_maximumWasmCalleeSizeForInlining=0 DYLD_FRAMEWORK_PATH=/Volumes/WebKit/ReleaseVersion/OpenSource/WebKitBuild/Release __XPC_DYLD_FAMEWORK_PATH=/Volumes/WebKit/ReleaseVersion/OpenSource/WebKitBuild/Release /Volumes/WebKit/ReleaseVersion/OpenSource/WebKitBuild/Release/Safari.app/Contents/MacOS/Safari -ExtensionsEnabled NO```

with sandbox disabled (see patch) and allow list:
```
12
```

reproduces the issue, producing the attached assembly

[Justin Michaud]

Created attachment 466936 [details]
Bad disassembly

[Justin Michaud]

Created attachment 466937 [details]
Disable sandbox to load allow list

[Justin Michaud]

The culprit:

(call $probe_begin (i32.const 22))
          (call $probe (i64x2.extract_lane 0 (local.get $l25)))
          (call $probe (i64x2.extract_lane 1 (local.get $l25)))
          (call $probe (i64x2.extract_lane 0 (local.get $l63)))
          (call $probe (i64x2.extract_lane 1 (local.get $l63)))
          (call $probe (i64x2.extract_lane 0 (local.get $l42)))
          (call $probe (i64x2.extract_lane 1 (local.get $l42)))
          (call $probe32 (local.get $p0))
          (call $probe (i64x2.extract_lane 0 (v128.load offset=16 (local.get $p0))))
          (call $probe (i64x2.extract_lane 1 (v128.load offset=16 (local.get $p0))))

          (i32.eqz
              (v128.any_true
                (local.get $l42)))
          (local.set $debug32)
          (call $probe32 (local.get $debug32))

          (v128.not
              (local.get $l42)) <---------------------------------------Here is the first difference
          (local.set $debug128)
          (call $probe (i64x2.extract_lane 0 (local.get $debug128)))
          (call $probe (i64x2.extract_lane 1 (local.get $debug128)))
(call $probe_end (i32.const 22))

[Justin Michaud]

Pull request: https://github.com/WebKit/WebKit/pull/18827

[EWS]

Committed 269080@main (f9a3a2147af0): <https://commits.webkit.org/269080@main>

Reviewed commits have been landed. Closing PR #18827 and removing active labels.