"A sufficiently advanced compiler will optimize all the overhead away" is the standard line for all high-level languages with automatic memory management. It works for tracing GC too, you just need "sufficiently advanced" escape analysis.
Unfortunately it means there are always performance cliffs: you make some change to your code that looks innocuous and suddenly your performance collapses (e.g. because a function is now too big to be automatically inlined, so certain optimizations don't work anymore) --- and you don't understand why.
I agree that could happen, but a parser would have to be coded pretty strangely to disable these basic ARC optimizations. I don't imagine a parser needing to hold and release references beyond the call stack.
In this vein, both Swift and Rust performance will succumb to any number of their IL-level (and LLVM's IR-level) optimizations being disabled by some seemingly orthogonal change to the source code.
Either way, I hope Apple (and others) can squash this entire category of bugs soon!
Unfortunately it means there are always performance cliffs: you make some change to your code that looks innocuous and suddenly your performance collapses (e.g. because a function is now too big to be automatically inlined, so certain optimizations don't work anymore) --- and you don't understand why.