(def sort-by-commonest(l)
    (sort (compare > counts.l) l))

---

I had a longer solution because I wasn't using compare or counts, so thanks for your hints.

Turns out I've been using a hand-rolled implementation of counts called freq. And, you know, I like my version better:

  (def freq(l)
    (ret ans (table)
      (each x l
        (++ (ans x 0)))))

  (def counts (seq (o c (table)))
    (if (no seq)
        c
        (do (++ (c (car seq) 0))
            (counts (cdr seq) c))))

---

One possible solution involves passing in just the keys of counts and repeating them the right number of times. But that feels ugly. In the real world you don't have duplicate objects, you often have objects with equal features but lots of other different aspects.

  (def freq(l (o f idfn))
    (ret ans (table)
      (each x l
        (++ (ans f.x 0)))))

  (def sort-by-commonest(l (o f idfn))
    (let h (freq l f)
      (sort (compare > h:f) l)))

  arc> (sort-by-commonest '((fox 1) (cat 1) (dog 1) (cat 2) (dog 3) (dog 2) (fish 1)) car)
  ((dog 1) (dog 3) (dog 2) (cat 1) (cat 2) (fox 1) (fish 1))

  arc> (sort-by-commonest '(1 3 3 4 4 5 4))
  (4 4 4 3 3 1 5) ; => not 5 1 because it's a stable sort

Song stuck in my head: REM's "What's the frequency, Kenneth?"

  (def sort-by-commonest (xs)
    ((afn (rx rs)
       (if (empty rx) rs
           (let (x n)(commonest rx)
             (self (rem x rx)(join rs (n-of n x))))))
        xs nil))

  arc> (sort-by-commonest '(1 3 3 4 4 5 4))
  (4 4 4 3 3 5 1)
  

  (def sort-by-commonest(l (o f idfn))
    (let h (counts:map f l)
      (sort (compare > h:f) l)))

For testing I am only using the basic case (i.e no accounting for applying functions to items as they are sorted). These following examples are only a preliminary run and intended only to start looking at comparable approaches. One thing I'm noticing right off the bat, is that Method 1 & 2 has code that does already account for applying functions, therefore could be potentially sped up..? Are there other things I should think about here? If the goal were to create the fastest version then are there other ideas that anybody else is willing to pitch in? Or if the goal were to create the most succinct code are there other ideas here too? And, in general, is there anything I am not giving consideration for?

Thanks.

  ; random data - ugly code / quickly hacked
  ; --------------------------------------------------------------------
  (def jumble (seq)
    (let tb (table)
      (each i seq 
        (withs (k (rand 50) v (tb k))
          (fill-table tb (list k (cons i v)))))
     (flat (vals tb))))

  (= data 
    (jumble 
      (let xs nil
	(repeat 100
          (push (with (n (rand 100) x (rand 100))
	    (n-of n x)) xs))
         (flat xs))))
	
	
  ; Method #1 -> succinct, performs very well + stable sort
  ; --------------------------------------------------------------------

  (def freq (l (o f idfn))
    (ret ans (table)
      (each x l
        (++ (ans f.x 0)))))

  (def sort-by-commonest1 (l (o f idfn))
    (let h (freq l f)
      (sort (compare > h:f) l)))

  arc> (time10 (do (sort-by-commonest1 data) nil))
  time: 2399 msec.


  ; Method #2 -> very succinct, performs well + stable sort
  ; --------------------------------------------------------------------

  (def sort-by-commonest2 (l (o f idfn))
    (let h (counts:map f l)
      (sort (compare > h:f) l)))

  arc> (time10 (do (sort-by-commonest2 data) nil))
  time: 2436 msec.


  ; Method #3 -> not so succinct , terrible/terrible performance + unstable sort
  ; --------------------------------------------------------------------
  
  (def sort-by-commonest3 (xs)
    ((afn (rx rs)
       (if (empty rx) rs
           (let (x n)(commonest rx)
             (self (rem x rx)(join rs (n-of n x))))))
        xs nil))

  arc> (time10 (do (sort-by-commonest3 data) nil))
  time: 8052 msec.


  ; Method #4 -> not so succinct, very fast + stable sort 
  ; --------------------------------------------------------------------

  (def sort-by-commonest4 (seq)
    (flat 
      (sort (fn (x y)(> (len x) (len y))) 
        ((afn (xs ys)
           (if (empty xs) ys
               (withs (x  (car xs) 
                       f  (testify [isnt _ x]) 
                       r  (reclist [if (f:car _) _] xs)
                       c1 (len xs)
                       c2 (len r))
                (self r (cons (firstn (- c1 c2) xs) ys)))))
	  (sort > seq) nil))))
	
  arc> (time10 (do (sort-by-commonest4 data) nil))	
  time: 1257 msec.

  ; And Method #5 -> some juice squeezing
  ; --------------------------------------------------------------------

  (def sort-by-commonest5 (seq)
    (mappend (fn (x) x) 	  
      (sort (fn (x y)(> (len x) (len y))) 
	 ((afn (xs cxs ys)
	    (if (empty xs) ys
                (withs (x  (car xs) 
                        f  (testify [isnt _ x]) 
                        r  (reclist [if (f:car _) _] xs)
                        cr (len r))
                (self r cr (cons (firstn (- cxs cr) xs) ys)))))
	     (sort > seq) (len seq) nil))))

  arc> (time10 (do (sort-by-commonest5 data) nil))	
  time: 1035 msec.

Also, I wrote an equivalent function in Clojure (again - for fun):

  (defn sortby-commonest0 [seq]
    (flatten 
      (sort #(> (count %1)(count %2))
  	     (afn [xs ys]
	       (if (empty? xs) ys
                   (let [x  (first xs)  
                         r  (for [i xs :while (= i x)] i)]
                   (self (drop (count r) xs) (cons r ys))))
	         (sort seq) nil))))

  => (time10 (sortby-commonest0 data))
  time: 340 msec.

  (def sort-by-commonest5 (seq)
    (mappend [do _]	  
      (sort (fn (x y)(> len.x len.y)) 
	((afn (xs cxs ys)
	      (if empty.xs ys
                  (withs (x  car.xs
                          f  (testify [isnt _ x]) 
                          r  (reclist [if (f:car _) _] xs)
                          cr len.r)
              (self r cr (cons (firstn (- cxs cr) xs) ys)))))
	   (sort > seq) len.seq nil))))

 arc> (time10 (do (sort-by-commonest5 data) nil))
 time: 1073 msec.


  For sortby-commonest0, freshly loaded, here are the first 10 runs:

  (3081 2102 1803 1585 1564 1529 1596 1813 1571 1580)

  As you can see with the JVM it warms up and running the same code gets faster...
  For example the second run (and my 340 must have been down the road?):

  (581 553 539 540 543 550 541 542 544 536)

  Then I tried a comparable succinct approach in Clojure 
  (no counts, so I used frequencies):

  (defn sort-by-commonest01 [xs]
    (let [h (frequencies xs)]
      (sort (comparitor > #(h %)) xs)))

  first 10:  (4568 2750 2495 2125 2399 2131 2123 2133 2056 1087)
  second 10: (1064 1074 1070 1069 1070 1101 1058 1062 1090 1066)

a) Where you say "mappend [do _]", you can just say "apply + nil". Actually, I've been using "apply join", but it turns out that performs much worse:

  arc> (time10:mappend idfn (n-of 10000 nil))
  time: 468 msec.
  nil
  arc> (time10:apply join (n-of 10000 nil))
  time: 50454 msec.
  nil
  arc> (time10:apply + nil (n-of 10000 nil))
  time: 437 msec.
  nil

c) Not that it'll really save anything when it comes to the overall computational complexity, but this version of list length comparison will only traverse the lists up to the end of the shorter one:

  (def longer-list (a b)
    (if a
      (if b
        (longer-list cdr.a cdr.b)
        t)
      nil))

e) There's no need to call 'testify.

After incorporating a bunch of this feedback--I didn't pit '< against '> --the code might look like this:

  (def sort-by-commonest5r1 (seq)
    (apply + nil
      (sort longer-list
        ( (afn (xs cxs ys)
            (if xs
              (withs (x car.xs
                      f [isnt _ x]
                      r (reclist [if (f:car _) _] xs)
                      cr len.r)
                (self r cr (cons (firstn (- cxs cr) xs) ys)))
              ys))
          (sort > seq) len.seq nil))))

  (def sort-by-commonest5r2 (seq)
    (apply join
      (sort longer-list
        ( (afn (rest bins)
            (iflet (sample . rest) rest
              ( (rfn self2 (bin rest)
                  (if rest
                    (let first car.rest
                      (if (is sample first)
                        (self2 (cons first bin) cdr.rest)
                        (self rest (cons bin bins))))
                    (cons bin bins)))
                list.sample rest)
              bins))
          (sort > seq) nil))))

  arc> (time:repeat 100 sort-by-commonest5r1.data)
  time: 18390 msec.
  nil
  arc> (time:repeat 100 sort-by-commonest5r2.data)
  time: 20437 msec.
  nil
  arc> (time:repeat 100 sort-by-commonest5.data)
  time: 19594 msec.
  nil

1. I made a big mistake on my time calculator for Clojure... it was using nanoTime with rounding errors. I had to re-write it. This is now correct with no rounding:

  (defmacro millitime*
    ([expr]
       `(let [start# (System/currentTimeMillis) 
              ret# ~expr
              stop# (System/currentTimeMillis)]
          (- stop# start#))) 
    ([n expr]
      `(map (fn[~'x](~'millitime* ~expr))(range 0 ~n))))

  2. Clojure runs:

  (def data (read-string (slurp (str rootDir* "data.seq"))))

  (defn sortby-commonest-Clojure-Long [seq]
    (flatten 
      (sort #(> (count %1)(count %2))
  	(afn [xs ys]
	  (if (empty? xs) ys
              (let [x  (first xs)  
                    r  (for [i xs :while (= i x)] i)]
              (self (drop (count r) xs) (cons r ys))))
       (sort seq) nil))))
	
  => (millitime* 10  (sortby-commonest-Clojure-Long data))
  (27 26 23 17 17 16 16 16 17 18)
  
  run2: (6 17 16 17 16 17 16 17 6 7)
  run3: (9 6 6 7 6 6 6 7 6 6)	

  ; restarted & reloaded data

  (defn sort-by-commonest-Clojure-succinct [xs]
    (let [h (frequencies xs)]
      (sort (comparitor > #(h %)) xs)))	
	
  => (millitime* 10 (sort-by-commonest-Clojure-succinct data))	
  (42 28 26 22 22 24 21 21 21 20)

  run2: (10 10 10 10 11 10 10 10 11 10)	
  run3: (10 11 10 10 10 12 11 10 10 10)

  3. Arc Runs:

 ; new time code for arc 
  (mac millitime (expr)
    (w/uniq (t1 t2)
    `(let ,t1 (msec)
       (do ,expr
            (let ,t2 (msec)
                (- ,t2 ,t1))))))

  (mac millitime10 (expr)
    `(accum xs
       (each i (range 1 10)
         (xs (millitime ,expr)))
      xs))


  (def sortby-commonest-Arc-Long (seq)
    (mappend [do _]	  
      (sort (fn (x y)(> len.x len.y)) 
	((afn (xs cxs ys)
	   (if empty.xs ys
               (withs (x  car.xs
                       f  (testify [isnt _ x]) 
                       r  (reclist [if (f:car _) _] xs)
                       cr len.r)
                (self r cr (cons (firstn (- cxs cr) xs) ys)))))
	   (sort > seq) len.seq nil))))

   arc> (millitime10 (sortby-commonest-Arc-Long data))
   (110 103 104 104 105 105 105 105 104 104)

   run2: (105 103 105 105 105 105 105 104 106 103)
   run3: (106 106 106 105 104 101 107 105 105 104)

   ; restarted & reloaded data
   
   (def sort-by-commonest-Arc-succinct (l (o f idfn))
    (let h (counts:map f l)
      (sort (compare > h:f) l)))

   arc> (millitime10 (sort-by-commonest-Arc-succinct data))
   (241 238 238 240 237 237 234 236 235 234)
  
   run2: (238 238 274 283 239 239 239 239 237 240)
   run3: (240 239 238 238 239 239 242 243 239 237)

  Here's rocketnia's new version:
  ; restarted & reloaded data

  arc>(millitime10 (sort-by-commonest5r1 data))
  (110 102 105 104 104 104 101 106 105 104)

  run2: (105 102 104 104 103 104 102 105 105 103)
  run3: (103 106 105 104 102 104 104 105 102 106)
  

I'm not surprised that the readable way is less performant; I'm reassured that the price for readability is quite low. I was totally unsurprised that both require similar optimizations given that both are fairly mature toolchains. Compilers by and large get their performance the same way.

The one thing that can cause the optimal optimizations to shift is different machines with different capacities and hierarchies of caches and RAM. But this is a compute-bound program so even that's not a concern.

I'm not so sure I can agree with the price for readability being quite low. One could call it premature optimization, but a ~50% gain is pretty significant in my mind. Had it been 20% or less I would probably go forward and spend less of my time attempting alternate approaches, but at ~50% I think playing around and learning the boundaries and their benefits can yield positive results for me.

(No criticism intended; no apologies necessary.)

It's 50% if you do just that. In a larger app it's a difference of 1ms.

Now you could argue that everything in an arc program will be slower so I should consider it to be 50%. Lately I've been mulling the tradeoff of creating optimized versions vs just switching languages. When I built keyword extraction for readwarp.com I was shocked how easy it was to add a C library using the FFI. Why bother with a less readable, hard-to-maintain arc version for 2x or 5x if you can get 50-100x using an idiomatic C version?

The whole point of a HLL is to trade off performance for readability; I'd rather back off that tradeoff in a key function or two.

---

Shameless plug section

Wart currently has an utterly simple FFI: create a new C file, write functions that take in Cells and return Cells, run:

  $ wart

But I want to do more than just an FFI. I have this hazy idea of being able to specify optimization hints for my code without actually needing to change it or affect its correctness. They might go into a separate file like the tests currently do.

I think that's generalizing too much. Using that one function with our nominal data set may only cost some ms, but what if you're dealing with hundreds of millions records? It may then, even though only representing .001% of your code base, account for 90% of your operating time - which is when someone normally kicks in with the "premature optimization" argument, which I can't really argue against, other than to say optimizing code is a skill that is generally done well by those who take it into account to begin with.

"Now you could argue that everything in an arc program will be slower so I should consider it to be 50%"

I wouldn't think this to be the case. I'm sure there's a tonne of juice squeezing one can do, but as a general statement, having played around with a lot of arc's code, I would guess most optimizing would yield less than 10%, but these 50+%, while they are few and far between are still worth the effort (to me).

The key, in all this, is to understand these languages well enough to make good judgment calls on where best to invest ones time.

Then just write it in C :)

Let me try to rephrase my argument. Sometimes you care about every last cycle, most of the time you care about making it a little bit faster and then you're done. Sometimes your program needs large scale changes to global data structures that change the work done, sometimes it needs core inner loops to just go faster. Sometimes your program is still evolving, and sometimes you know what you want and don't expect changes.

just a little faster + work smarter => rearchitect within arc.

just a little faster + faster inner loops => rewrite them in C

every last cycle + rigid requirements => gradually rewrite the whole thing in C and then do micro-optimizations on the C sources. You could do arc optimizations here, but this is my bias.

every last cycle + still evolving => ask lkml :)

If you're doing optimizations for fun, more power to you. It's pretty clear I'm more enamored with the readability problem than the performance problem.

I don't want to sound more certain than I am about all this. Lately I spend all my time in the 'just a little faster' quadrant, and I like it that way.

"The key, in all this, is to understand these languages well enough to make good judgment calls on where best to invest ones time."

Let's put put it another way. In my situation, if I have to look a using C then it's already game over[1]. However I can, with my limited amount of time (lunches, evenings and weekends) become proficient enough in Clojure (with a continual thanks to Arc).

What I am suggesting is that knowing the language well enough to easily identify these 50%+ hitters is a matter of finding low hanging fruit and at the same time becoming a better Arc/Clojure programmer. It does not mean I want to change my direction in the heavily-optimized/low-level to high-level/low-maintenance/readable code continuum.

[1] I'm not a programmer[2], I am a business analyst that works with software that simulates oil & gas production from reservoirs that sit several hundred kilometers underground. Simulation scenario's are run for 20 year production periods across hundreds of interdependent wells. The current software tools run the simulations across many cores, they take 8 days to run to completion and they cost about a half a million dollars per seat (+18% per year maintenance fees). This cost can be attributed to the R&D that occurred 8-10 years ago (i.e they required a team(s?) of P.Engs writing software in C to maximize the performance). Eight years ago you couldn't do (pmap[3] (sortby-commonest or whatever.... )) so easily. Nowadays I have the opportunity to create a 70% solution all by my lonesome, costing only a portion of my time. Hence why understanding the language well enough to find the low hanging fruit and not having to use C, is probably bigger deal to me.

[2] Well, maybe I shouldn't say this... rather I should say it's not my day job. I have no formal education in the field. I'm pretty much self taught.

[3] pmap is the same as map only it distributes the load across multiple processors in parallel.

  (sort-by-c­ommonest '(dog­ dog cat fox fox dog fish)­)
  (dog dog dog fox fox cat fish)