The fact that you can fuse multiple passes into one alteration (or reconstruction) of the tree is what's a removal of overhead.
[0] https://mechanical-sympathy.blogspot.com/2012/04/invoke-inte...
>This is exactly what the JVM does. See [0].
The jump table from Martin's blog is the virtual call one.
On top of that JVM has an inline budget and just stops at a point. So the switch stuff ain't happening for megamorph.
[0]:https://blog.h2o.ai/2010/04/inline-caches-and-call-site-opti...
It's a trade off and the article shows no proof the of an improved performance by reducing the 'overhead'. For example object allocation in JVM is really cheap (pointer bump in the TLAB + check&perfectly predicted jump), reusing objects in a cycling list/buffer is similar, etc. In the end the post is about performance and I pointed that the multiple mentions of 'zero overhead' are incorrect.
When I talk about how the OO community has anti-intellectual toxicity can try to buzzword bingo architecture away, Visitor is my go to example. Because that's sure what it seems like to me.
You will 100% end up with pipeline stalls with a sufficiently large, random input anyways. This complaint about vtables is not substantial in this case, and I welcome you to prove they do.
So that seems like quite a stretch.