[Webkit-unassigned] [Bug 242294] REGRESSION (iOS 16): RELEASE_ASSERT in slow_path_wasm_loop_osr

[bugzilla-daemon at webkit.org]

https://bugs.webkit.org/show_bug.cgi?id=242294

Justin Michaud <justin_michaud at apple.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |justin_michaud at apple.com

--- Comment #9 from Justin Michaud <justin_michaud at apple.com> ---
Hi Kurt,

This issue was being tracked in our internal bug tracker. This also affects Miniclip.com on iOS.

Thank you for your investigation, it matches what I have seen investigating that bug.

Air and B3 have two different approaches to loop osr. In BBQ Air, every loop is included as a potential entry point, so we only need to compile once. For OMG B3, we wish to compile with only one entry point at a time, so that we do not pessimize loops.

1) Your regression test does not reproduce for me, but I have prepared a test case that does. 

2) Your patch seems to be very similar to the fix that I prepared for the Miniclip bug, so I think we should take it. Thank you for your excellent writeup. I will prepare a PR on GitHub. I am happy to go ahead and land it if you want to get back to work, but you should feel free to make your own PR and we can land that to make sure you are publicly credited. 

3) The final solution here is to remove forceB3 completely. We have the WASM LLint now, so I do not think that this mode is useful. There are some possible performance implications of doing that, so we should land this first.

Thanks again, I will post the PR shortly.

-- 
You are receiving this mail because:
You are the assignee for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.webkit.org/pipermail/webkit-unassigned/attachments/20220705/be7aa602/attachment-0001.htm>