fast-case about clj-fast

[bsless]

Thinking about this and #9 , what do you think will be the effects of missed branch predictions on this code? Can we engineer a test case where branch prediction isn't possible to test its effects?

[joinr]

I'm no expert on that, let alone knowing how engineer good degenerate cases. I guess you would have random input over the cases....no idea beyond that.

[bsless]

I ran the following interesting test:

(defn foo [k]
  (fast-case k :a 0 :b 1 :c 2 :d 3 :e 4 :f 5 :g 6 :h 7 :i 9 :j 10 :k 11 :l 12 :m 13 :n 14 :o 15 :p 16 :q 17 :r 18 :s 19 :none))

(defn bar [k]
  (case k :a 0 :b 1 :c 2 :d 3 :e 4 :f 5 :g 6 :h 7 :i 9 :j 10 :k 11 :l
             12 :m 13 :n 14 :o 15 :p 16 :q 17 :r 18 :s 19 :none))

(def ks [:a :b :c :d :e :f :g :h :i :j :k :l :m :n :o :p :q :r :s :none])

(def lots (repeatedly 10000 #(rand-nth ks)))

(do
  (c/bench (doseq [k lots] 0))
  (c/bench (doseq [k lots] (foo k)))
  (c/bench (doseq [k lots] (bar k))))

Overhead: 571.822473 µs
Fast case: 654.098578 µs, minus overhead 85
Core case: 740.666878 µs, minus overhead 170
=> fast case is 2x faster, amortized .

[joinr]

Nice. I tend to use case quit a bit, as does some clojure internals like defrecord, so this seems generally useful.

[bsless]

I wander how critical it actually is. While identity case checks are 2x faster we're still talking about 4 vs 8 ns in the worst case scenario, not to mention I did not measure these influences in a wider context of a running application. Perhaps cache misses will degrade other parts' characteristics, so I'm still wary.
It would be interesting to get a profile of a typical application and how many times it cases on keywords.

[joinr]

If you're using keyword access for records, it will impact your performance. Ended up being a enough in the IFCPC optimization case (where I first ran into this, and defrecord's weaker performance compared to arraymaps..) that I found out about this. If you're not using idiomatic paths that leverage case on hot paths, it may not matter then. That can probably be applied to most of these optimizations though (e.g. context and profiling matter).

[bsless]

Is there a need to introduce the binding to ge and type hint it as an Object? I fail to see the ratilnale behind it. Am I missing anything?

[joinr]

That came from the original source in clojure.core/case. Unsure.