1 #+TITLE: Generalizers: New Metaobjects for Generalized Dispatch
2 #+AUTHOR: Christophe Rhodes, Jan Moringen, David Lichteblau
5 #+LaTeX_HEADER: \usepackage[margin=1in]{geometry}
8 1. This paper introduces a new metaobject, the generalizer, which
9 complements the existing specializer metaobject.
10 2. With the help of examples, we show that this metaobject allows for
11 the efficient implementation of complex non-class-based dispatch
12 within the framework of existing metaobject protocols
13 3. We present the generalizer protocol, implemented within the SBCL
14 implementation of Common Lisp
15 4. In combination with previous work, this produces a fully-functional
16 extension of the existing mechanism for method selection and
17 effective method computation, including support for standard and
18 user-defined method combination independent from method selection.
22 The revisions to the original Common Lisp language \cite{CLtL1}
23 included the detailed specification of an object system, known as
24 the Common Lisp Object System (CLOS), which was eventually
25 standardized as part of the ANSI Common Lisp standard \cite{CLtS}.
26 The object system as presented to the standardization committee was
27 formed of three parts, the first two of which covered XXX [what?]
28 and were incorporated into the final standard, and the third,
29 covering a Metaobject Protocol (MOP) for CLOS, was not.
31 Nevertheless, the CLOS MOP has proven to be a robust design, and
32 while many implementations have derived their implementations of
33 CLOS from either the Closette illustrative implementation in
34 \cite{AMOP}, or the Portable Common Loops implementation of CLOS
35 from Xerox Parc, there have been from-scratch reimplementations of
36 CLOS (in at least CLISP; check for others -- ABCL? Lisp500?!)
37 incorporating the majority of the Metaobject Protocol as described.
39 Although it has stood the test of time, the MOP is neither without
40 issues (e.g. M-M-L considered harmful; slot-definition initargs
41 issue) nor a complete framework for the metaprogrammer to implement
42 all conceivable variations of object-oriented behaviour; indeed,
43 while metaprogramming offers some possibilities for customization of
44 the object system behaviour, those possibilities cannot extend
45 arbitrarily in all directions. There is still an expectation that
46 functionality is implemented with methods on generic functions,
47 acting on objects with slots. [XXX find Paepke picture here? Not
48 Paepke; AMOP?]. XXX include typical examples of MOP: object
49 persistence; maybe ref. Kizcales "MOPs: why we want them and what
50 else they can do"? (Fig. 2 in that is good) ORMs; sparse slots.
52 + introspection, e.g. documentation generation
53 + programmatic construction of classes and generic functions
54 e.g. for IDL compilers, model transformations
56 One area of functionality where there is scope for customization by
57 the metaprogrammer is in the mechanics and semantics of method
58 applicability and dispatch. While in principle AMOP allows
59 customization of dispatch in various different ways (the
60 metaprogrammer can define methods on protocol functions such as
61 =compute-applicable-methods=,
62 =compute-applicable-methods-using-classes=), for example, in
63 practice implementation support for this was weak until relatively
64 recently (ref. closer, also check how ContextL and filtered dispatch
66 jmoringe: filtered dispatch uses a custom method combination, i
69 Another potential mechanism for customizing dispatch is implicit in
70 the class structure defined by AMOP: standard specializer objects
71 (instances of =class= and =eql-specializer=) are generalized
72 instances of the =specializer= protocol class, and in principle
73 there are no restrictions on the metaprogrammer constructing
74 additional subclasses. Previous work [Newton/Rhodes] has explored
75 the potential for customizing generic function dispatch using
76 extended specializers, but as of that work the metaprogrammer must
77 override the entirety of the generic function invocation protocol
78 (from =compute-discriminating-function= on down), leading to toy
79 implementations and duplicated effort.
81 This paper introduces a protocol for efficient and controlled
82 handling of arbitrary subclasses of =specializer=. In particular,
83 it introduces the =generalizer= protocol class, which generalizes
84 (ahem) the return value of =class-of=, and allows the metaprogrammer
85 to hook into cacheing schemes to avoid needless recomputation of
86 effective methods for sufficiently similar generic function
87 arguments (See Figure\nbsp\ref{fig:dispatch}).
89 #+CAPTION: Dispatch Comparison
91 #+ATTR_LATEX: width=0.9\linewidth float
92 [[file:figures/dispatch-comparison.pdf]]
94 The remaining sections in this paper can be read in any order. We
95 give some motivating examples in section XX, including
96 reimplementations of examples from previous work, as well as
97 examples which are poorly supported by previous protocols. We
98 describe the protocol itself in section YY, describing each protocol
99 function in detail and, where applicable, relating it to existing
100 protocol functions within the CLOS MOP. We survey related work in
101 more detail in section ZZ, touching on work on customized dispatch
102 schemes in other environments. Finally, we draw our conclusions
103 from this work, and indicate directions for further development, in
104 section WW; reading that section before the others indicates
105 substantial trust in the authors' work.
107 In this section, we present a number of examples of dispatch
108 implemented using our protocol, which we describe in section YY.
109 For reasons of space, the metaprogram code examples in this section
110 do not include some of the necessary support code to run; complete
111 implementations of each of these cases are included in an appendix /
112 in the accompanying repository snapshot / at this location.
114 A note on terminology: we will attempt to distinguish between the
115 user of an individual case of generalized dispatch (the
116 “programmer”), the implementor of a particular case of generalized
117 dispatch (the “metaprogrammer”), and the authors as the designers
118 and implementors of our generalized dispatch protocol (the
119 “metametaprogammer”, or more likely ”we”).
121 - [ ] =cons-specializer= (can be done using filtered dispatch)
122 - [ ] factorial (like filtered dispatch)
123 - [ ] HTTP Accept header
125 - [ ] prototype/multimethod
127 We start by presenting our original use case, performing
128 dispatching on the first element of lists. Semantically, we allow
129 the programmer to specialize any argument of methods with a new
130 kind of specializer, =cons-specializer=, which is applicable if and
131 only if the corresponding object is a =cons= whose =car= is =eql=
132 to the symbol associated with the =cons-specializer=; these
133 specializers are more specific than the =cons= class, but less
134 specific than an =eql-specializer= on any given =cons=.
136 One motivation for the use of this specializer is in an extensible
137 code walker: a new special form can be handled simply by writing an
138 additional method on the walking generic function, seamlessly
139 interoperating with all existing methods.
141 The programmer code using these specializers is unchanged from
142 \cite{Newton.Rhodes.2008}; the benefits of the protocol described
143 here are centered on performance: in an application such as walking
144 source code, we would expect to encounter special forms
145 (distinguished by particular atoms in the =car= position) multiple
146 times, and hence to dispatch to the same effective method
149 (defclass cons-specializer (specializer)
150 ((%car :reader %car :initarg :car)))
151 (defclass cons-generalizer (generalizer)
152 ((%car :reader %car :initarg :car)))
153 (defmethod generalizer-of-using-class ((gf cons-generic-function) arg)
155 ((cons symbol) (make-instance 'cons-generalizer :car (car arg)))
156 (t (call-next-method))))
157 (defmethod generalizer-equal-hash-key ((gf cons-generic-function)
158 (g cons-generalizer))
160 (defmethod specializer-accepts-generalizer-p ((gf cons-generic-function)
162 (g cons-generalizer))
163 (if (eql (%car s) (%car g))
166 (defmethod specializer-accepts-p ((s cons-specializer) o)
167 (and (consp o) (eql (car o) (%car s))))
169 #| less interesting methods elided: jmoringe: (un)parsing, specializer<?, more? |#
171 #| XXX insert motivating example from Newton/Rhodes here |#
174 TODO Timings: generalizer-enabled vs not vs case (car x) for CL special
175 operators. Tweak so that the overhead of =case= is actually important.
177 Note that in this example there is no strict need for
178 =cons-specializer= and =cons-generalizer= to be distinct classes –
179 just as in the normal protocol involving
180 =compute-applicable-methods= and
181 =compute-applicable-methods-using-classes=, the specializer object
182 for mediating dispatch contains the same information as the object
183 representing the equivalence class of objects to which that
184 specializer is applicable: here it is the =car= of the =cons=
185 object; in the standard dispatch it is the =class= of the object.
186 This feature also characterizes those use cases where the
187 metaprogrammer could straightforwardly use filtered dispatch
188 \cite{Costanza.etal:2008} to implement their dispatch semantics.
189 We will see in section XX.x an example of a case where filtered
190 dispatch is incapable of efficiently implementing the dispatch, but
191 first we present our implementation of the motivating case from
192 \cite{Costanza.etal:2008}.
194 Our second example of the implementation and use of generalized
195 specializers is a reimplementation of one of the examples in
196 \cite{Costanza.etal:2008}: specifically, the factorial function.
197 Here, we will perform dispatch based on the =signum= of the
198 argument, and again, at most one method with a =signum= specializer
199 will be appliable to any given argument, which makes the structure
200 of the specializer implementation very similar to the =cons=
201 specializers in the previous section.
203 We have chosen to compare signum values using \texttt{=}, which
204 means that a method with specializer =(signum 1)= will be
205 applicable to positive floating-point arguments (see the first
206 method on =specializer-accepts-generalizer-p= and the method on
207 =specializer=accepts-p= below). This leads to one subtle
208 difference in behaviour compared to that of the =cons=
209 specializers: in the case of =signum= specializers, the /next/
210 method after any =signum= specializer can be different, depending
211 on the class of the argument. This aspect of the dispatch is
212 handled by the second method on =specializer-accepts-generalizer-p=
215 (defclass signum-specializer (specializer)
216 ((%signum :reader %signum :initarg :signum)))
217 (defclass signum-generalizer (generalizer)
218 ((%signum :reader %signum :initarg :signum)))
219 (defmethod generalizer-of-using-class ((gf signum-generic-function) arg)
221 (real (make-instance 'signum-generalizer :signum (signum arg)))
222 (t (call-next-method))))
223 (defmethod generalizer-equal-hash-key ((gf signum-generic-function)
224 (g signum-specializer))
225 (%signum g)) ; this will create multiple entries for the same emf, but that's OK
226 (defmethod specializer-accepts-generalizer-p ((gf signum-generic-function)
227 (s signum-specializer)
228 (g signum-generalizer))
229 (if (= (%signum s) (%signum g)) ; or EQL?
233 ;; this method is perhaps interesting enough to talk about?
234 (defmethod specializer-accepts-generalizer-p ((gf signum-generic-function) (specializer sb-mop:specializer) (thing signum-specializer))
235 (specializer-accepts-generalizer-p gf specializer (class-of (%signum thing))))
238 (defmethod specializer-accepts-p ((s signum-specializer) o)
239 (and (realp o) (= (%signum s) (signum o))))
241 #| again elide more boring methods |#
244 Given these definitions, and some more straightforward ones elided
245 for reasons of space, we can implement the factorial function as
250 (:generic-function-class signum-generic-function))
251 (defmethod fact ((n (signum 0))) 1)
252 (defmethod fact ((n (signum 1))) (* n (fact (1- n))))
255 We do not need to include a method on =(signum -1)=, as the
256 standard =no-applicable-method= protocol will automatically apply to
257 negative real or non-real arguments.
259 Benchmarketing: we chose to benchmark 20! because that is the
260 largest factorial whose answer fits in SBCL's 63-bit fixnums, so as
261 to attempt to measure the maximum effect of dispatch (unobscured by
262 allocation / gc issues)
265 (progn (gc :full t) (time (dotimes (i 10000) (%fact 20))))
268 | fact (signum-gf) | %fact (fun) | %%fact (gf / 1meth) | fact (signum-gf / 1arg hash-special-case) |
269 |------------------+-------------+---------------------+-------------------------------------------|
270 | 0.284 | 0.004 | 0.016 | 0.032 |
271 | 0.076 | 0.008 | 0.012 | 0.024 |
272 | 0.072 | 0.004 | 0.012 | 0.116 |
273 | 0.264 | 0.004 | 0.008 | 0.120 |
274 | 0.292 | 0.008 | 0.012 | 0.120 |
275 | 0.264 | 0.004 | 0.016 | 0.084 |
276 | 0.276 | 0.008 | 0.012 | 0.092 |
277 | 0.264 | 0.008 | 0.012 | 0.036 |
278 | 0.276 | 0.008 | 0.004 | 0.104 |
279 | 0.272 | 0.004 | 0.012 | 0.020 |
281 We could allow the metaprogrammer to improve on the one-argument
282 performance by constructing a specialized cache: for =signum=
283 arguments of =rational= arguments, the logical cache structure is
284 to index a three-element vector with =(1+ signum)=. The current
285 protocol does not provide a way of eliding the two generic function
286 calls for the generic cache; we discuss possible approaches in
288 ** HTTP Accept header
289 In this section, we implement a non-trivial form of dispatch. The
290 application in question is a web server, and specifically to allow
291 the programmer to support RFC 2616 \cite{rfc2616} content
292 negotiation, of particular interest to publishers and consumers of
295 The basic mechanism in content negotiation is as follows: the web
296 client sends an HTTP request with an =Accept:= header, which is a
297 string describing the media types it is willing to receive as a
298 response to the request, along with numerical preferences. The web
299 server compares these stated client preferences with the resources
300 it has available to satisfy this request, and sends the best
301 matching resource in its response.
303 In the case where there are static files on the filesystem, and the
304 web server must merely select between them, there is not much more
305 to say. However, it is not unusual for a web service to be backed
306 by some other form of data, and responses computed and sent on the
307 fly, and in these circumstances the web server must compute which
308 of its known output formats it can use to satisfy the request
309 before actually generating the best matching response.
311 The =accept-specializer= below implements the dispatch. It depends
312 on a lazily-computed =tree= to represent the information in the
313 accept header, and a function =q= to compute the (defaulted)
314 preference level for a given content-type and =tree=; then, method
315 selection and ordering involves finding the =q= for each
316 =accept-specializer='s content type given the =tree=, and sorting
317 them according to the preference level.
320 (defclass accept-specializer (extended-specializer)
321 ((media-type :initarg :media-type :reader media-type)))
322 (defclass accept-generalizer ()
323 ((header :initarg :header :reader header)
325 (next :initarg :next :reader next)))
326 (defmethod generalizer-equal-hash-key
327 ((gf accept-generic-function) (g accept-generalizer))
328 `(accept-generalizer ,(header g)))
329 (defmethod specializer-accepts-generalizer-p ((gf accept-generic-function) (s acc
330 ept-specializer) (generalizer accept-generalizer))
331 (values (q (media-type s) (tree generalizer)) t))
332 (defmethod specializer-accepts-generalizer-p ((gf accept-generic-function) (s sb-
333 mop:specializer) (generalizer accept-generalizer))
334 (specializer-accepts-generalizer-p gf s (next generalizer)))
336 (defmethod specializer< ((gf accept-generic-function) (s1 accept-specializer) (s2
337 accept-specializer) generalizer)
339 ((string= (media-type s1) (media-type s2)) '=)
340 (t (let ((q1 (q (media-type s1) (tree generalizer)))
341 (q2 (q (media-type s2) (tree generalizer))))
347 ;; here are the only methods that actually know about TBNL
348 (defmethod generalizer-of-using-class ((gf accept-generic-function) (arg tbnl:request))
349 (make-instance 'accept-generalizer
350 :header (tbnl:header-in :accept arg)
351 :next (class-of arg)))
352 (defmethod specializer-accepts-p ((specializer accept-specializer) (obj tbnl:requ
354 (q (media-type specializer) (parse-accept-string (tbnl:header-in :accept obj)))
358 This dispatch can't be done with filtered dispatch, except by
359 generating anonymous classes with all the right mime-types as
360 direct superclasses in dispatch order,so the filter does
362 (ensure-class nil :direct-superclasses '(text/html image/webp ...))
364 and dispatch is defined by using those classes. And that's even
365 more awkward than it sounds, because that means that in principle
366 all the mime types in the universe need an existence as classes, to
367 cater for arbitrary mime types in accept headers. And handling
368 wildcards is pretty much impossible, too. See that in
369 =specializer<= which involves a nontrivial ordering of methods
370 (whereas in two above previous cases only a single extended
371 specializer could be applicable to any given argument)
373 Also of interest: note that we can have these
374 specializer/generalizers handle arbitrary objects: we can define
375 methods on =string=, completely independently from the methods on
376 =tbnl:request= methods are independent of each other; this
377 generalizes to dealing with multiple web server libraries.
380 ;; we can define methods on STRING too, for debugging/simulation purposes
381 (defmethod generalizer-of-using-class ((gf accept-generic-function) (s string))
382 (make-instance 'accept-generalizer
385 (defmethod specializer-accepts-p ((s accept-specializer) (string string))
386 (q (media-type s) (parse-accept-string string)))
389 jmoringe: the name =accept-specializer=, while sensible, may
390 confusing in this context because "accept" occurs as part of the
391 protocol with a different semantic.
392 ** Pattern / xpattern / regex / optima
393 Here's the /really/ interesting bit, but on the other hand we're
394 probably going to run out of space, and the full description of
395 these is going to take us into =make-method-lambda= territory.
396 A second paper? Future work?
399 - [ ] =generalizer-of-using-class= (NB class of gf not class of object)
400 - [ ] =compute-applicable-methods-using-generalizers=
401 - [ ] =generalizer-equal-hash-key=
402 - [ ] =specializer-accepts-generalizer-p=
403 - [ ] =specializer-accepts-p=
405 jmoringe: If I remember correctly, closette has
406 =method-more-specific-p= should we aim for parity with that and
407 use =specializer-more-specific-p=? The downside would be that
408 =-p= indicates a Boolean return value which is not the case here.
410 Description and specification left for reasons of space (we'll see?)
411 - [ ] =same-specializer-p=
412 - [ ] =parse/unparse-specializer-using-class=
413 - [ ] =make-method-specializers-form=
414 - [ ] jmoringe: In an email, I suggested
415 =make-specializer-form-using-class=:
418 Could we change =make-method-specializers-form='s default
419 behaviour to call a new generic function
421 make-specializer-form-using-class gf method name env
423 with builtin methods on =sb-mop:specializer=, =symbol=, =cons= (for
424 eql-specializers)? This would make it unnecessary to repeat
425 boilerplate along the lines of
427 (flet ((make-parse-form (name)
428 (if <name-is-interesting>
429 <handle-interesting-specializer>
430 <repeat-handling-of-standard-specializers>)))
431 `(list ,@(mapcar #'make-parse-form specializer-names)))
433 for each generic function class.
435 - [ ] =make-method-lambda= revision (use environment arg?)
437 jmoringe: would only be relevant for pattern dispatch, right? I
438 think, we didn't finish the discussion regarding special
439 variables vs. environment vs. new protocol function
442 - [ ] filtered dispatch -- the point is that our work continues to
443 be useful in cases where there are unbounded numbers of
444 equivalence classes but each given invokation involves a small
446 - [ ] ContextL / context-oriented programming -- dispatch occurs on
447 hidden layer argument being an instance of an anonymous class with
448 suitably arranged superclasses -- OK because set of layers is
449 bounded and under programmer control
450 - [ ] http://soft.vub.ac.be/Publications/2010/vub-tr-soft-10-04.pdf
451 - [ ] http://soft.vub.ac.be/lambic/files/lambic-ilc09.pdf
452 - [ ] http://soft.vub.ac.be/Publications/2011/vub-soft-phd-11-03.pdf
453 - [ ] Prototypes with Multiple Dispatch
454 http://sauerbraten.org/lee/ecoop.pdf -- extension of Self-style
455 object system to handle multiple equally-privileged "receivers".
456 A good test case for our protocol; handled adequately with
457 generalizer being the tuple of (roles,delegations), with some
458 thought needed for method redefinitions but otherwise working
463 We thank Lee Salzman, Pascal Costanza, Mikel Evins for their