As discussed with mmarczyk in irc, the attached patch implements Clojure's hash-set function and modifies the compiler to create sets using a new fromArray static method. To test these, I created three new benchmarks. Here's before and after benchmark results:

$ time ./script/benchmark | grep '#{'
[], #{}, 100000 runs, 77 msecs
[], #{1 2 3}, 100000 runs, 582 msecs
[coll #{1 2 3}], (conj coll 4), 100000 runs, 104 msecs
[], #{}, 100000 runs, 740 msecs
[], #{1 2 3}, 100000 runs, 1809 msecs
[coll #{1 2 3}], (conj coll 4), 100000 runs, 221 msecs
[], #{}, 100000 runs, 281 msecs
[], #{1 2 3}, 100000 runs, 1310 msecs
[coll #{1 2 3}], (conj coll 4), 100000 runs, 288 msecs
./script/benchmark 109.41s user 3.82s system 127% cpu 1:28.96 total

$ time ./script/benchmark | grep '#{'
[], #{}, 100000 runs, 1 msecs
[], #{1 2 3}, 100000 runs, 333 msecs
[coll #{1 2 3}], (conj coll 4), 100000 runs, 100 msecs
[], #{}, 100000 runs, 1 msecs
[], #{1 2 3}, 100000 runs, 1670 msecs
[coll #{1 2 3}], (conj coll 4), 100000 runs, 325 msecs
[], #{}, 100000 runs, 1 msecs
[], #{1 2 3}, 100000 runs, 689 msecs
[coll #{1 2 3}], (conj coll 4), 100000 runs, 321 msecs
./script/benchmark 103.94s user 4.89s system 108% cpu 1:39.88 total
grep '#{' 0.00s user 0.00s system 0% cpu 1:39.85 total