I found [1] to be a pretty good summary of the key concepts that support the claim. The article, unfortunately, doesn't come right out and say it. But, it talks about how object shapes (a common JS object model) can be used as guards for inline caches (IC) to generate efficient machine code for property access. If you add a new property, you change the shape, which causes the cache guard to fail. That leads to deoptimization since the assumptions made about the call site when the code was compiled no longer hold. That call site is now either polymorphic or megamorphic, depending on the type of IC used and its level of polymorphism.
Comments
Any reference for this?
I found [1] to be a pretty good summary of the key concepts that support the claim. The article, unfortunately, doesn't come right out and say it. But, it talks about how object shapes (a common JS object model) can be used as guards for inline caches (IC) to generate efficient machine code for property access. If you add a new property, you change the shape, which causes the cache guard to fail. That leads to deoptimization since the assumptions made about the call site when the code was compiled no longer hold. That call site is now either polymorphic or megamorphic, depending on the type of IC used and its level of polymorphism.
[1] -- https://mathiasbynens.be/notes/shapes-ics
https://v8.dev/blog/fast-properties
https://v8.dev/docs/hidden-classes