Watching a macro as it expands

;; macros are code transformers
;; in a lisp, that means tree transformers
;; let's look at the -> macro, built into clojure.

(use 'clojure.contrib.repl-utils)
(source ->)

;; (defmacro ->
;;   "Threads the expr through the forms. Inserts x as the
;;   second item in the first form, making a list of it if it is not a
;;   list already. If there are more forms, inserts the first form as the
;;   second item in second form, etc."
;;   ([x] x)
;;   ([x form] (if (seq? form)
;;               (with-meta `(~(first form) ~x ~@(next form)) (meta form))
;;               (list form x)))
;;   ([x form & more] `(-> (-> ~x ~form) ~@more)))



;; What's 20C in Fahrenheit?
;; Multiply by nine, divide by 5, add 32
(-> 20 (* 9) (/ 5) (+ 32))

;; let's look at the tree for that
(use 'clojure.inspector)
(inspect-tree '(-> 20 (* 9) (/ 5) (+ 32)))

;; What does that look like after macroexpansion?
(use 'clojure.walk)
(macroexpand-all '(-> 20 (* 9) (/ 5) (+ 32)))
(inspect-tree (macroexpand-all '(-> 20 (* 9) (/ 5) (+ 32))))


;; How does the threading macro work?
(macroexpand-1 '(-> 20 (* 9) (/ 5) (+ 32)))

;; We can follow the recursion by using macroexpand-1 on each subexpression in turn.

;;slightly tidied up, stages are:

(macroexpand-1 '(-> 20 (* 9) (/ 5) (+ 32)))

(macroexpand-1
'(-> 
  (-> 20 (* 9))
  (/ 5) (+ 32)))

(macroexpand-1
 '(-> 
   (-> 
    (-> 20 (* 9)) 
    (/ 5)) 
   (+ 32)))

(macroexpand-1
 '(+ 
   (-> 
    (-> 20 (* 9)) 
    (/ 5))
   32))

`(+ 
  ~(macroexpand-1 '
    (-> 
     (-> 20 (* 9)) 
     (/ 5)))
  32)

`(+ 
  (/ 
   ~(macroexpand-1 '
     (-> 20 (* 9)))
   5)
  32)

(+ 
 (/ 
  (* 20 9) 
  5) 
 32)

;;This is a bit long-winded, how would we do this automatically?

(defn slow-tree-transformer [fn tree]
  (cond (not (seq? tree)) (fn tree)
        (empty? tree) '()
        :else (let [nxt (fn tree)]
                (if (not (= nxt tree)) nxt
                    (let [nxt (fn (first tree))]
                      (if (= nxt (first tree))
                        (let [nxt (slow-tree-transformer fn (first tree))]
                          (if (= nxt (first tree))
                            (cons (first tree) (slow-tree-transformer fn (rest tree)))
                            (cons nxt (rest tree))))
                        (cons nxt (rest tree))))))))


(def f (fn[x] (slow-tree-transformer macroexpand-1 x)))

(defn iterate-to-stable [fn start]
  (let [next (fn start)]
    (if (= next start) (list start)
        (cons start (iterate-to-stable fn next)))))

(use 'clojure.contrib.pprint)
(println "-----------------------------------")
(doall (map (fn [x] (pprint x) (println)) (iterate-to-stable f '(-> 20 (* 9) (/ 5) (+ 32)))))
(println "-----------------------------------")

Jihad Scanner

There's a very promising looking new lisp book out, Doug Hoyte's "Let over Lambda" http://letoverlambda.com/

My copy arrived this morning, and to keep me honest while reading it, I wondered how much of it could be translated into clojure. Here's the first example I found interesting:
 

;; The jihad scanner from let over lambda. Watch a communications channel for a word.

(defn make-scanner [ watchword ]
  (let [ curr (atom watchword) ]
    (fn [ chunk ]
      (doseq [c chunk]
        (if @curr
          (swap! curr (fn [cv] (cond (= (first cv) c)        (next cv) 
                                     (= (first watchword) c) (next watchword)
                                     :else                    watchword)))))
      (nil? @curr))))


;; I've added a clause so that jjihad still fires the detector. I think the original version doesn't catch jjihad
;; but does catch qjihad, which seems inconsistent.

;; Compared to the Common Lisp version, the requirement to wrap mutable state in an atom has cost us
;; a couple of spurious @s, and required "(swap! curr (fn [cv]" instead of "(setq" without really 
;; buying us anything. 

;; On the other hand, the fact that strings are just another form of seq has simplified the function
;; and generalised it to an arbitrary subsequence detector.


(let [a (make-scanner "jihad")]
  (a "We will start")
  (a "the jih")
  (a "ad tomorrow"))


(let [a (make-scanner "jihad")]
  (a "The jijihad starts tomorrow"))

(let [a (make-scanner "jihad")]
  (a "The jij ihad starts tomorrow"))

(let [a (make-scanner "jihad")]
  (a "Don't mention the j-word"))

(let [a (make-scanner '(97 98 99))]
  (for [x (partition 3 (range 110))] (a x)))

(let [a (make-scanner '[a b c])]
  (for [ x '((a b) (c d)) ] (a x)))

(let [a (make-scanner '[a b c])]
  (for [ x '((a b) [c d]) ] (a x)))

Requiring all possible namespaces

It often irritates me that I can't use clojure's built in documentation unless I've already loaded the library I need.

Sometimes it turns out that the generic-looking function I need already exists, but it can be hard to find it. Sometimes I can remember the function I want, but not its precise name or library namespace.

It's also often the case that the best way to understand a function is to read the source.

The other day it occurred to me that a function to find all the namespaces available on the classpath might help quite a bit, and I started to write one.

And at that point I had the obvious recursive thought.

   

;; Seek and ye shall find. Thank you Stuart Sierra.
(use 'clojure.contrib.find-namespaces)

;; First let's see what our classpath is:
(println "Classpath")
(doseq [p (seq (.getURLs (java.lang.ClassLoader/getSystemClassLoader)))] (println (str p)))


;; In my case, that's the clojure.jar and clojure.contrib.jar files, and also the directory
;; containing the files for swank-clojure, the library which allows it to talk to emacs.
(println "Swank namespaces")
(println (for [s (for [n (all-ns)] (name (ns-name n))) :when (. s contains "swank")] s))


;; Because I'm still a beginner, and not brave enough to modify swank, I'm more interested in 
;; the namespaces provided by clojure and clojure.contrib

;; So far, the following have managed to get themselves loaded:
(println "None-swank namespaces")
(use 'clojure.contrib.pprint)
(pprint (sort (for [s (for [n (all-ns)] (name (ns-name n))) :when (not (. s contains "swank"))] s)))


;; But the following are in fact available
(pprint (find-namespaces-on-classpath))

;; Again, let's ignore the swank-related ones.
(pprint (filter #(not (. (str %) startsWith "swank")) (find-namespaces-on-classpath)))


;;For me there are only namespaces starting with either swank or clojure, because I don't have much
;;of a classpath. But it's easy enough to check that.
(defn find-namespaces-starting-with [strng]
  (filter #(. (str %) startsWith strng) (find-namespaces-on-classpath)))

(map count (map find-namespaces-starting-with '("clojure" "swank" ""))) ; should add up


;; Some of the clojure.contrib libraries fail to load for me, so we have to bullet-proof require
(defn require-may-fail [ns]
  (try
   (print "Attempting to require " ns ": ")
   (require ns)
   (println "success")
   (catch Exception e (println "couldn't require " ns "\nException\n" e "\n\n"))))


;; Now we can load everything
(doall (map require-may-fail (filter #(. (str %) startsWith "clojure") (find-namespaces-on-classpath))))


;; Now we have lots and lots of namespaces available: The number of 'batteries included' is impressive.
(println "None-swank namespaces")
(pprint (sort (for [s (for [n (all-ns)] (name (ns-name n))) :when (not (. s contains "swank"))] s)))


;; And the find-doc, doc, and source commands become really useful
;; Of course we not only have to find source, we actually can, now.
(find-doc "source code")

;; Also note that all this requiring hasn't brought any new functions into play yet.
;; To use something without qualification, you need to refer as well.
(refer 'clojure.contrib.repl-utils)
(doc source)
(source source)
(source get-source)
(:file (meta (resolve 'source)))

Other sources of enjoyment are the find and grep commands, and browsing clojure.contrib.jar and clojure.jar . These work particularly nicely under emacs. A good arrangement of find and grep is (assuming that your clojure sources are under ~/opt):

find ~/opt/ -name "*.clj" -and -type f -print0 | xargs -0 -e grep -nH -e "repl-util"

Using the trace function in clojure.contrib

There's a fine trace macro in clojure.contrib.trace, which really can help when looking at recursive functions.
Here's a very simple example with our dear friend 'naive fibonacci'.

   
user> (defn fib [n]
        (if (< n 2) n
            (+ (fib (- n 1)) (fib (- n 2)))))
#'user/fib
user> (use 'clojure.contrib.trace)
nil
user> (dotrace (fib) (fib 3))

TRACE t1880: (fib 3)
TRACE t1881: |    (fib 2)
TRACE t1882: |    |    (fib 1)
TRACE t1882: |    |    => 1
TRACE t1883: |    |    (fib 0)
TRACE t1883: |    |    => 0
TRACE t1881: |    => 1
TRACE t1884: |    (fib 1)
TRACE t1884: |    => 1
TRACE t1880: => 2
2
user> 

How the Clojure dojo actually worked in practice.

We decided that I'd try out (the first half of) my Clojure dojo on the Chemistry Department.

The audience was the non-lispers in our group, so there were about eight very bright Java guys, one of whom had done some ML before, and one of whom had been through this very exercise with me in a pub some months before as his first contact with LISP since he met McCarthy in the sixties(!).

The dojo format worked very well, and I mainly sat back and let them get on with it.

I didn't say a great deal after posing the problem in the terms outlined at the start. When they completed a subproblem I gave them a new problem and maybe a hint or two. I was surprised how much they managed to infer from limited information.

At first everything went absolutely swimmingly, and the syntax didn't faze them at all. Questions like 'What's the sum of the squares of 3 and 4' were answered without breaking stride. They understood Heron's method immediately. When I suggested that they turned it into a function to improve guesses, they spontaneously put in a target variable instead of hard-coding 10, and started talking about recursion.

They couldn't figure out how to do it, so I suggested trying to write factorial. This they managed to work out for themselves.

There was then much excitement trying to turn that into a solver, but they couldn't do it.

I suggested trying to write good-enough?

They realised the need for an absolute value function, and to my utter amazement, guessed the syntax for calling Math/abs, and then for Math/sqrt!

They declared the problem solved.

So I asked them to generalize the solution to make a general root-finder, which got them back to trying to get Heron's method to work.

This was the last time I said anything helpful. Much more thinking, and they got on the right track. They defined an okguess function which took the tolerance as a parameter, and were well on the way to solving the problem, when the ML guy and the guy who'd done it before took the chairs.

I said they were only allowed a couple of minutes, but they had the thing done almost immediately.

I still wanted to go back and show them iterate and map, so I asked them to make a function that returned a function. Can we make a function that takes a number and returns a function which multiplies things by it?

I hinted that defn was a combination of def and fn, and that was enough

That solved, I asked them to make a function which took a target and gave back the relevant improver.

Then I told them to try (doc iterate), and see if they could figure out what it was for. And then I told them to look up map and take. They were playing happily with infinite sequences when our hour and a half ran out.

Understanding the REPL

A REPL is a subtle thing.

Consider what happens when you type in 23

user>23
23

The program reads in the sequence of characters 2, 3, <return>, and prints the sequence of characters 2, 3, <return>

It looks very much as if nothing has happened at all!

But in fact, a great deal has happened.

A REPL is a read-eval-print loop.

Any interaction with it involves reading in a stream of characters, which are passed to the function read. The output from read is passed to the evaluator, eval The output from eval is passed to the print function, and turned into a sequence of characters, which are then rendered on the screen by a mysterious process known as the operating system.

We can forget about the process which reads the keyboard and interprets physical movements as a stream of characters We can also forget about the process which prints the characters on the screen.

Clojure would work in just the same way if it lived on a computer half-a-mile away, and you sent and recieved characters to it by Morse code.

So let us concentrate only on the three processes which take the character stream coming in and return the character stream going out.

Read, Eval, and Print.

One way to understand what is going on is to write your own read, eval, and print functions.

One way to get started on that project is to borrow the read and print functions from an existing lisp, and then use that lisp to write your own eval function, which is both the most interesting bit of the system, and the easiest to construct.

This is a traditional rite of passage for lispers, and I refer you to chapter 4 of the Structure and Interpretation of Computer Programs if you wish to pursue it.

Now most lisps will give us the power to call their read, eval, and print functions, and very useful they are too.

But Clojure is a lisp written in java, and it also gives us the power to create objects in the underlying java system.

And we can use this power to conduct experiments on the three functions.

Let's look a little at the reader first.

The reader reads characters from any stream which is of type java.io.PushbackReader

We can create one of these streams from our REPL!

Remember that clojure is just a java program. We could have created these streams in a java program, but we will use clojure's REPL to manipulate them. We must remember that whenever anything is typed in, it is the reader and the evaluator that interpret them, and the printer that displays the result, but if we are careful we can pick the process apart.

So here's our input stream with the characters 2,3, in it.

(java.io.PushbackReader. (java.io.StringReader. "23\n"))
#<PushbackReader java.io.PushbackReader@1f9f0f2>

Remember, we typed in (...). The reader interpreted that stream of characters, and passed it to the evaluator.

The evaluator, as a result created a java.lang.String, wrapped that in a java.io.StringReader, wrapped that in a java.io.PushbackReader, and then gave that object to the printer, which decided that an appropriate printed representation would be the sequence of characters #,<,P,u,s,...........2,>,

Let's instead assign it to a value

(def inputstream (java.io.PushbackReader. (java.io.StringReader. "23\n")))
#'user/inputstream

Now the evaluator has done much the same thing, but then rather than handing the PushbackReader to the print function, it has attached it to the var #'user/inputstream, noted that the symbol foo should be associated in the user namespace with the var #'user/inputstream, and then passed the var to the printer, which has decided that #,',u,s,....,m is an appropriate sequence of characters with which to represent the var to the user.

We can check that the symbol->var mapping is now in the user namespace:

(ns-publics 'user)
{inputstream #'user/inputstream}

This is a side-effect of evaluating a def form. In fact the return value of the expression is not particularly interesting. It is the side-effect that we are after.

Since we've only just started our REPL, the user namespace is empty apart from the association that the evaluator has just made.

Now the question in which we are interested is: What does the reader produce when it is given a PushbackReader which yields the characters 2,3, ?

We can ask the REPL

(read inputstream)
23

And it tells us that it is a thing which the printer chooses to represent as 2,3,

PushbackReaders are mutable objects, and this one is now exhausted. It has no more characters to give.

Let us create another similar stream, pass it again to the reader, and this time save the value that the reader gives us.

(def read23 (read (java.io.PushbackReader. (java.io.StringReader. "23\n"))))
#'user/read23

Another var has been created by the def, and the printer has turned it into a stream of characters for us.

But we can examine the object to which read23 refers:

(type read23)
java.lang.Integer

Aha! the reader returned, not the string of characters 2,3,, but an object of type java.lang.Integer I wonder what else we can learn about it?

Remember that it is a java class, and that we can use clojure to call java methods on it.

(.getMethods (class read23))
#

(map str (.getMethods (class read23)))

java.lang.Integer is a class with many methods and we will have an easier time understanding it if we cheat.

(use 'clojure.contrib.repl-utils)
(javadoc read23)

let's try a method. doubleValue returns the value as a string.

(.doubleValue read23)
23.0

Unless the printer is being incredibly perverse, we are dealing with a java.lang.Integer whose value is 23.

But notice that this is a very different thing from a character stream made from 2,3,

Let's forget all about the details of creating streams, and instead use the read-string function, which can be fed strings directly.

What does the reader return when we give it the characters (,+,,2,,2,)?

(def twoplustwo (read-string "(+ 2 2)\n"))
#'user/twoplustwo

(type twoplustwo)
clojure.lang.PersistentList

That's a new one.

(.count twoplustwo)
3

A list with three things in it.

(map type twoplustwo)
(clojure.lang.Symbol java.lang.Integer java.lang.Integer)

A list of one symbol, and two integers.

(javadoc clojure.lang.Symbol)

(.getName (first twoplustwo))
"+"

(.doubleValue (first (rest twoplustwo)))
2.0

So the brackets have gone. The reader's interpretation of the string "(+ 2 2)" is a list of a clojure.lang.Symbol whose name is the (java.lang.)String "+", and two java.lang.Integers whose value as doubles is 2.0.

I reckon we have this reader thing pretty well sorted out. But look at this:

(def line-noise (read-string "'foo"))
#'user/line-noise

(type line-noise)
clojure.lang.Cons

(.first line-noise)
quote

(type (.next line-noise))
clojure.lang.PersistentList

(.count (.next line-noise))
1

(type (.first (.next line-noise)))
clojure.lang.Symbol

(.getName (.first (.next line-noise)))
"foo"

So it appears that when the reader's presented with ',f,o,o It produces a Cons, whose first element is the Symbol whose name is "quote", and whose next element is a list, whose sole element is the Symbol "foo".

The printer represents this little tree as (quote foo).

But if we wrote our own printer we might well call it (cons (symbol "quote") (list (symbol "foo")))

indeed this is a clojure expression:

(cons (symbol "quote") (list (symbol "foo")))

which we can type in at the REPL, which will return the same string as if we typed in (read-string "'foo")

It appears that there are some subtleties to the reader.

xxdiff and hg view with mercurial

Not really clojure, but now I'm using it properly on a project where we're using mercurial for version control

Enabling xxdiff

You can link xxdiff into mercurial, but you have to modify the hgrc file


add to  $HOME/.hgrc

[extensions]
hgext.extdiff =

and then we can say

hg extdiff -p program -o options

e.g.

$ hg extdiff -p xxdiff -o -r

which causes xxdiff -r to run
or

$ hg extdiff -p kdiff3

We can be even more subtle by adding an explicit command for xxdiff -r

[extdiff]

cmd.xxdiff = xxdiff
opts.xxdiff = -r

which allows us to use

$ hg xxdiff

to call xxdiff directly in its recursive mode.

Enabling hg view

While we're messing about with .hgrc, add

[extensions]
hgext.extdiff=
hgk=

which will enable the hg view command.