Reference counting is usually very expensive, because even reading a variable updates the reference count, and ending a scope involves testing the reference count of every variable defined inside the scope and conditionally deallocating the referent.
Without reference counting, here's the end of a hairy function scope that deallocates 15 local variables and restores two callee-saved registers:
11a6: 48 83 c4 78 add $0x78,%rsp
11aa: 5b pop %rbx
11ab: 41 5e pop %r14
11ad: c3 retq
Now imagine looping over 15 local variables to decrement each reference count, test it, and do a conditional jump based on the result; if that's open-coded, your epilogue is humongous, and if it's in a reference-count-decrementing subroutine, it's going to have mispredicted branches all over the place, costing you maybe 15 cycles each, a total of about 100 cycles for this function. We're talking about adding an order of magnitude of cost to subroutine call, or more. (I think this function doesn't really need 15 local variables; that's the compiler's fault.)
This gets worse with multithreading, because writing to the same reference count on different cores would even in the best case require one core stealing the cache line from another in order to modify it, which may stall the core; but often even an atomic reference count increment is more expensive even than that because it involves a memory barrier.
Reference counting can become reasonably cheap if it's done at large granularity (filesystem files or COM objects, not conses); if you can elide almost all of the reference-count updates, as in Rust; or if your language runtime is just so dog-slow at everything that the extra cost of reference counting isn't that important, like CPython.
30 years ago or more, before generational GC had gone mainstream, reference counting was a more reasonable choice, because GC was going to be very slow in any case, and ref counting at least used less memory—especially important on machines without cache or virtual memory.
(Purely immutable (applicative) languages like Haskell and Miranda are usually lazy, since that's the payoff for completely abjuring side effects. But lazy evaluation is implemented at the machine level by mutation.)