Welcome to Language Agnostic, the blog of Inaimathi! It's now built on top of the Clojure framework http-kit. I'm reasonably confident it's no longer the least performant blog on the internet.

Enjoy the various programming-themed writings availble on offer. The latest post is available below and the archive link is directly above this text.


Subverting Common Lisp Types And Emacs Interaction For Clj

Sun May 24, 2020

Ok, so profiling the extra-low-hanging-fruit in terms of generic function performance revealed that it didn't do much in our situation. My next idea was to subvert the Common Lisp type system to give our set and map primitives some hints about what kind of equality operations to use.

I'm once again starting this piece before having written the code it'll be explaining, so this is less a thoughtful tour and more a stream-of-consciousness account of the writing.

Defining Types

Assuming the thing you're defining fits into the pre-existing Common Lisp types, you're fine. As soon as you want to do something like define polymorphic key/value structures you are, near as I can tell, on your fucking own bucko.

So I guess I'm rolling my own here?

Ok, the good news is that I'm in just enough of a hacky mood that I don't give a flying fuck how shitty this is going to be. That... might come back to bite me later, but we'll burn that bridge and salt it as we pass.

Here's how I have to define type map.

(deftype map (&optional keys vals)
  (let ((sym (intern (format nil "MAP-TYPE-~a-~a" keys vals) :clj)))

    (unless (fboundp sym)
      (setf (fdefinition sym) (kv-types keys vals)))

    `(and (satisfies map?) (satisfies ,sym))))

This feels batshit insane. In order to properly define a polymorphic key/value type, I have to manually intern predicates that deal with the specific types in question at declaration time. The problem is that satisfies specifically only accepts a symbol that must refer to a function of one argument that's meant to return a boolean. If it could take lambda terms, I could do something like

(defun kv-type (k-type v-type)
  (lambda (thing)
    (and (map? thing)
	 (every (lambda (pair)
		  (and (typep (car pair) k-type)
		       (typep (cdr pair) v-type)))
		(values thing)))))
...
(satisfies (kv-type 'keyword 'integer))

This is, unfortunately, off the table. Oh well. The complete definitions for both map and set types is

(defun map? (thing)
  (typep thing 'cl-hamt:hash-dict))

(defun map-type? (type)
  (and type
       (listp type)
       (eq (car type) 'map)))

(defun kv-types (k-type v-type)
  (lambda (map)
    (cl-hamt:dict-reduce
     (lambda (memo k v)
       (and memo (typep k k-type) (typep v v-type)))
     map t)))

(deftype map (&optional keys vals)
  (let ((sym (intern (format nil "MAP-TYPE-~a-~a" keys vals) :clj)))

    (unless (fboundp sym)
      (setf (fdefinition sym) (kv-types keys vals)))

    `(and (satisfies map?) (satisfies ,sym))))

(defun set? (thing)
  (typep thing 'cl-hamt:hash-set))

(defun set-type? (type)
  (and type
       (listp type)
       (eq (car type) 'set)))

(defun seq-types (v-type)
  (lambda (set)
    (cl-hamt:set-reduce
     (lambda (memo elem)
       (and memo (typep elem v-type)))
     set t)))

(deftype set (&optional vals)
  (let ((sym (intern (format nil "SET-TYPE-~a" vals) :clj)))

    (unless (fboundp sym)
      (setf (fdefinition sym) (seq-types vals)))

    `(and (satisfies set?) (satisfies ,sym))))

Once I've got that, I can declare things. Like,

CLJ> (let ((a {:a 1 :b 2}))
  (declare (type (map keyword t) a))
  a)
{:A 1 :B 2}
CLJ>

Checking for equalities

There's some more work to do. The whole point of this exercise is Once I've got a type declared, I need to do the work I actually care about. Which is: figure out which of the built-in structural equality operations is the most efficient I can use while also being as correct as possible.

(defun fullest-equality (equalities)
  (find-if
   (lambda (e) (member e equalities :test #'eq))
   '(clj:== cl:equalp cl:equal cl:eql cl:eq cl:string= cl:=)))

(defun equality-function (name) (fdefinition name))

(defun equality-of (type)
  (cond
    ((member type '(integer number float ratio rational bignum bit complex long-float short-float signed-byte unsigned-byte single-float double-float fixnum))
     'cl:=)
    ((member type '(string simple-string))
     'cl:string=)
    ((member type '(atom symbol keyword package readtable null stream random-state))
     'cl:eq)
    ((member type '(standard-char character pathname))
     'cl:eql)
    ((member type '(cons list))
     'cl:equal)
    ((and (listp type) (eq 'or (first type)))
     (fullest-equality (mapcar #'equality-of (rest type))))
    ((member type '(hash-table sequence array bit-vector simple-array simple-bit-vector simple-vector vector))
     'cl:equalp)
    ((and (listp type) (member (car type) '(array simple-array simple-bit-vector simple-vector vector)))
     'cl:equalp)
    ((member type '(compiled-function function))
     nil)
    (t 'clj:==)))

It's a fairly naive binding table, completely inextensible for the moment, that maps a type to the name of an equality operation that will accurately compare them. Hopefully, I mean. As long as I didn't fuck something up.

CLJ> (equality-of '(map keyword t))
==
CLJ> (equality-of 'keyword)
EQ
CLJ> (equality-of 'list)
EQUAL
CLJ> (equality-of 'hash-table)
EQUALP
CLJ> (equality-of 'string)
STRING=
CLJ>

Seems legit.

Putting it all together

The next step is, we want to use this equality selection procedure to make our map and set constructors pick a better one than == if it can.

(defparameter *type* nil)
...
(defun alist->map (alist &key equality)
  (let ((equality (or equality
		      (if (map-type? *type*)
			  (equality-function (equality-of (second *type*)))
			  #'==))))
    (loop with dict = (cl-hamt:empty-dict :test equality)
       for (k . v) in alist do (setf dict (cl-hamt:dict-insert dict k v))
       finally (return dict))))

(defun list->map (lst &key equality)
  (assert (evenp (length lst)) nil "Map literal must have an even number of elements")
  (let ((equality (or equality
		      (if (map-type? *type*)
			  (equality-function (equality-of (second *type*)))
			  #'==))))
    (loop with dict = (cl-hamt:empty-dict :test equality)
       for (k v) on lst by #'cddr
       do (setf dict (cl-hamt:dict-insert dict k v))
       finally (return dict))))
...
(defun list->set (lst)
  (let ((equality (if (set-type? *type*)
		      (equality-function (equality-of (second *type*)))
		      #'==)))
    (reduce
     (lambda (set elem)
       (cl-hamt:set-insert set elem))
     lst :initial-value (cl-hamt:empty-set :test equality))))

So, we've got a *type* special var that we can use to declare the type of the map/set we're defining, and if it's set, we use it to pick an appropriate equality. Otherwise, we just go with #'==, because that's as general as it gets.

CLJ> (list->set (list 1 2 3 4))
#{3 2 1 4}
CLJ> (cl-hamt::hamt-test (list->set (list 1 2 3 4)))
#<STANDARD-GENERIC-FUNCTION CLJ:== (8)>
CLJ> (let ((*type* '(set integer))) (list->set (list 1 2 3 4)))
#{3 2 1 4}
CLJ> (let ((*type* '(set integer))) (cl-hamt::hamt-test (list->set (list 1 2 3 4))))
#<FUNCTION =>
CLJ> (list->map (list :a 1 :b 2 :c 3))
{:A 1 :C 3 :B 2}
CLJ> (cl-hamt::hamt-test (list->map (list :a 1 :b 2 :c 3)))
#<STANDARD-GENERIC-FUNCTION CLJ:== (8)>
CLJ> (let ((*type* '(map keyword t))) (cl-hamt::hamt-test (list->map (list :a 1 :b 2 :c 3))))
#<FUNCTION EQ>
CLJ>
Nice.

It doesn't fit all of our use cases though.

CLJ> {:a 1 :b 2 :c 3}
{:A 1 :C 3 :B 2}
CLJ> (let ((*type* '(map keyword t))) {:a 1 :b 2 :c 3})
{:A 1 :C 3 :B 2}
CLJ> (let ((*type* '(map keyword t))) (cl-hamt::hamt-test {:a 1 :b 2 :c 3}))
#<STANDARD-GENERIC-FUNCTION CLJ:== (8)>
CLJ>

The problem is that, because we have reader syntax for our maps and sets, this decision kicks in too late to deal with them. We unfortunately also need a reader macro to handle type declarations.

Reader Macro for Type Declaration

The naive solution here is

...
(defun type-literal-reader (stream sub-char numarg)
  (declare (ignore sub-char numarg))
  (let* ((*type* (read stream))
	 (form (read stream))
	 (val (eval form)))
    (assert (typep val *type*) nil "Type checking failure ~s ~s" *type* form)
    val))

...
  (:dispatch-macro-char #\# #\# #'type-literal-reader))

I don't really want to define this as using :: because of the headache-inducing implications of doing (make-dispatch-macro-character #\:). I'm trying to avoid those for the moment. Same story with #:, because uninterned symbols are common and I don't want to stomp them here. So, I had to pick something else, and arbitrarily accepted ## even though #t or #T would have been equally reasonable choices.

This technically works.

CLJ> {:a 1 :b 2 :c 3}
{:A 1 :C 3 :B 2}
CLJ> ## (map keyword t) {:a 1 :b 2 :c 3}
{:A 1 :C 3 :B 2}
CLJ> (cl-hamt::hamt-test ## (map keyword t) {:a 1 :b 2 :c 3})
#<FUNCTION EQ>
CLJ>

But I want to avoid calling eval as part of it. The more macro-like version would look something more like

(defun type-literal-reader (stream sub-char numarg)
  (declare (ignore sub-char numarg))
  (let* ((*type* (read stream))
	 (form (read stream))
	 (res (gensym)))
    (if *type*
	`(let ((,res ,form))
	   (check-type ,res ,*type*)
	   ,res)
	res)))

It still works...

CLJ> ## (map keyword t) {:a 1 :b 2}
{:A 1 :B 2}
CLJ> (cl-hamt::hamt-test ## (map keyword t) {:a 1 :b 2})
#<FUNCTION EQ>
CLJ>

... but has the added bonuses of not calling eval and also making use of check-type, which we couldn't do if we wanted to do that check inline at read time.

I don't really like the syntax1, but that's good enough for now2.

Performance implications

(defun untyped-benchmark (&key (times 10000))
  (loop repeat times
     do (let* ((m {:a 1 :b "two" :c :three :d 44})
	       (inserted (insert m (cons :test-key :test-value))))
	  (list (len m)
		(lookup inserted :test-key)
		(len inserted)))))

(defun typed-benchmark (&key (times 10000))
  (loop repeat times
     do (let* ((m ## (map keyword t) {:a 1 :b "two" :c :three :d 44})
	       (inserted (insert m (cons :test-key :test-value))))
	  (list (len m)
		(lookup inserted :test-key)
		(len inserted)))))

With the above defined in benchmark.lisp, running the benchmarks and reporting them with M-x slime-profile-report slime-profile-reset gives us...

CLJ> (untyped-benchmark :times 1000000)
NIL
measuring PROFILE overhead..done
  seconds  |     gc     |     consed    |   calls   |  sec/call  |  name
---------------------------------------------------------------
     1.230 |      0.068 | 1,076,698,880 | 1,000,000 |   0.000001 | CLJ::INSERT
     0.931 |      0.000 |        32,768 | 2,000,000 |   0.000000 | CLJ::LEN
     0.617 |      0.000 |     1,679,216 | 1,000,000 |   0.000001 | CLJ::LOOKUP
     0.000 |      0.018 |    59,768,832 |         1 |   0.000000 | CLJ::UNTYPED-BENCHMARK
     0.000 |      0.000 |             0 | 1,000,000 |   0.000000 | CLJ:==
---------------------------------------------------------------
     2.778 |      0.086 | 1,138,179,696 | 5,000,001 |            | Total

estimated total profiling overhead: 9.09 seconds
overhead estimation parameters:
  1.8e-8s/call, 1.8179999e-6s total profiling, 8.8e-7s internal profiling

These functions were not called:
 CLJ:ALIST->MAP CLJ::EQUALITY-FUNCTION CLJ::EQUALITY-OF
 CLJ::FULLEST-EQUALITY CLJ::KV-TYPES CLJ::LIST->MAP CLJ:LIST->SET
 CLJ::MAP-LITERAL-READER CLJ::MAP-TYPE-KEYWORD-T CLJ::MAP-TYPE?
 CLJ::MAP? CLJ::SEQ-TYPES CLJ::SET-LITERAL-READER CLJ::SET-TYPE?
 CLJ::SET? CLJ::TYPE-LITERAL-READER CLJ::TYPED-BENCHMARK

CLJ> (typed-benchmark :times 1000000)
NIL
  seconds  |     gc     |     consed    |   calls   |  sec/call  |  name
---------------------------------------------------------------
     1.195 |      0.040 | 1,076,307,616 | 1,000,000 |   0.000001 | CLJ::INSERT
     0.768 |      0.000 |             0 | 2,000,000 |   0.000000 | CLJ::LEN
     0.605 |      0.000 |             0 | 1,000,000 |   0.000001 | CLJ::LOOKUP
     0.000 |      0.004 |    59,703,296 |         1 |   0.000000 | CLJ::TYPED-BENCHMARK
---------------------------------------------------------------
     2.568 |      0.044 | 1,136,010,912 | 4,000,001 |            | Total

estimated total profiling overhead: 7.27 seconds
overhead estimation parameters:
  1.8e-8s/call, 1.8179999e-6s total profiling, 8.8e-7s internal profiling

These functions were not called:
 CLJ:== CLJ:ALIST->MAP CLJ::EQUALITY-FUNCTION CLJ::EQUALITY-OF
 CLJ::FULLEST-EQUALITY CLJ::KV-TYPES CLJ::LIST->MAP CLJ:LIST->SET
 CLJ::MAP-LITERAL-READER CLJ::MAP-TYPE-KEYWORD-T CLJ::MAP-TYPE?
 CLJ::MAP? CLJ::SEQ-TYPES CLJ::SET-LITERAL-READER CLJ::SET-TYPE?
 CLJ::SET? CLJ::TYPE-LITERAL-READER CLJ::UNTYPED-BENCHMARK
CLJ>

A pretty goddamn tiny difference. I'm not sure this approach is worth much more effort, but I'll plug away for a bit longer to see how elegant I can make it. In the meantime,

Emacs Interaction Improvements

So, the sad thing about all of this is that I've been lying to you. Whenever I show you those nice readouts from the SLIME repl that says something like {:a 1 :b 2}, that's a result of me correcting it. Because, by default, when I type {, what I get is {$. Which I then have to manually backspace and correct. Using this shiny new syntax in Common Lisp mode is also less than ideal, because the default paredit doesn't provide s-exp support for curly braces. It's not as simple as adding

"{" 'paredit-open-curly
"}" 'paredit-close-curly

to a mode-map somewhere, because that does pair them, but doesn't help with navigation.

After messing around with modifying existing syntax-tables, redefining matching-paren, and poking around in paredit internals, the solution I settled on was just adding a mode-hook to a bunch of lisp modes and slime-repl modes that activates the clojure-mode-syntax-table.

You can do this in your .emacs file by doing something like

(defun use-clojure-syntax-table () (set-syntax-table (set-syntax-table clojure-mode-syntax-table)))
(add-hook 'common-lisp-mode-hook 'use-clojure-syntax-table)
(add-hook 'slime-mode-hook 'use-clojure-syntax-table)
(add-hook 'slime-repl-mode-hook 'use-clojure-syntax-table)

I added it to my .emacs by doing

(hooks (common-lisp lisp emacs-lisp scheme lisp-interaction slime clojure slime-repl)
       (lambda ()
	 (setq autopair-dont-activate t)
	 (autopair-mode -1))
       'enable-paredit-mode
       (lambda () (set-syntax-table (set-syntax-table clojure-mode-syntax-table))))

Which is both more thorough and more extensive, but requires me to define some conveniences first.

The next time I write about CLJ, the SLIME repl snippets will not be a lie.

  1. Ideally, the type annotation would be declared like (:: type form), :: type form, or possibly type :: form. However, infix operators are more complicated to deal with, and : already has various meanings in Common Lisp that would make using it as a read-macro-char more complicated than I'd like. Specifically, as hinted at above, doing (make-dispatch-macro-character #\:) instantly complicates the parsing of keywords, uninterned-symbols and any qualified name you end up typing. I'll read up on it a bit and see if there's a way to fall through to the default behavior somehow, but in the absence of that option, this is absolutely more hassle than it's worth.
  2. Possible future improvements include inferring the type of a map literal based on its initial values, and storing the type annotation somehow so that it can be checked against by insert later. I'm not sure any of this is worth the time, and once we pick an appropriate interface, it'll be easy to change internals later.


Creative Commons License

all articles at langnostic are licensed under a Creative Commons Attribution-ShareAlike 3.0 Unported License

Reprint, rehost and distribute freely (even for profit), but attribute the work and allow your readers the same freedoms. Here's a license widget you can use.

The menu background image is Jewel Wash, taken from Dan Zen's flickr stream and released under a CC-BY license