But tie isn't the only kind of magic there is, and all magic (except overload) is triggered by the same code checks. Most of the variables listed in perlvar use some form of magic (admittedly, a few do so uselessly; for example $^O).

I thought untie within a FETCH used to work and got broken at some point along the way and then fixed. This is an area of valid complaint: magic really isn't well represented in perl's regression tests.

Speaking of pervasive performance penalties, perl has an "optimization" of not reference counting items on the stack (including parameters in @_ in most cases). To keep things from getting garbage collected while still on the stack, it has tons of special code to mark things that should be kept even with a refcount of 0, and special code when dealing with arrays to check if the array at hand is @_ without refcounts. I've often wondered what the gain or loss would be if that mechanism were ripped out and regular reference counting were used for the stack, but its really to late now; there'd be too much work and too much danger of introducing bugs.