1 ;;;; the basics of the PCL wrapper cache mechanism
3 ;;;; This software is part of the SBCL system. See the README file for
6 ;;;; This software is derived from software originally released by Xerox
7 ;;;; Corporation. Copyright and release statements follow. Later modifications
8 ;;;; to the software are in the public domain and are provided with
9 ;;;; absolutely no warranty. See the COPYING and CREDITS files for more
12 ;;;; copyright information from original PCL sources:
14 ;;;; Copyright (c) 1985, 1986, 1987, 1988, 1989, 1990 Xerox Corporation.
15 ;;;; All rights reserved.
17 ;;;; Use and copying of this software and preparation of derivative works based
18 ;;;; upon this software are permitted. Any distribution of this software or
19 ;;;; derivative works must comply with all applicable United States export
22 ;;;; This software is made available AS IS, and Xerox Corporation makes no
23 ;;;; warranty about the software, its performance or its conformity to any
26 ;;;; Note: as of SBCL 1.0.6.3 it is questionable if cache.lisp can
27 ;;;; anymore be considered to be "derived from software originally
28 ;;;; released by Xerox Corporation", as at that time the whole cache
29 ;;;; implementation was essentially redone from scratch.
39 ;;;; emit-cache-lookup
41 ;;;; hash-table-to-cache
43 ;;;; This is a thread and interrupt safe reimplementation loosely
44 ;;;; based on the original PCL cache by Kickzales and Rodrigues,
45 ;;;; as described in "Efficient Method Dispatch in PCL".
47 ;;;; * Writes to cache are made atomic using compare-and-swap on
48 ;;;; wrappers. Wrappers are never moved or deleted after they have
49 ;;;; been written: to clean them out the cache need to be copied.
51 ;;;; * Copying or expanding the cache drops out incomplete and invalid
54 ;;;; * Since the cache is used for memoization only we don't need to
55 ;;;; worry about which of simultaneous replacements (when expanding
56 ;;;; the cache) takes place: the loosing one will have its work
57 ;;;; redone later. This also allows us to drop entries when the
58 ;;;; cache is about to grow insanely huge.
60 ;;;; The cache is essentially a specialized hash-table for layouts, used
61 ;;;; for memoization of effective methods, slot locations, and constant
64 ;;;; Subsequences of the cache vector are called cache lines.
66 ;;;; The cache vector uses the symbol SB-PCL::..EMPTY.. as a sentinel
67 ;;;; value, to allow storing NILs in the vector as well.
69 (defstruct (cache (:constructor %make-cache)
70 (:copier %copy-cache))
71 ;; Number of keys the cache uses.
72 (key-count 1 :type (integer 1 (#.call-arguments-limit)))
73 ;; True if we store values in the cache.
75 ;; Number of vector elements a single cache line uses in the vector.
76 ;; This is always a power of two, so that the vector length can be both
77 ;; an exact multiple of this and a power of two.
78 (line-size 1 :type (integer 1 #.most-positive-fixnum))
79 ;; Cache vector, its length is always both a multiple of line-size
80 ;; and a power of two. This is so that we can calculate
81 ;; (mod index (length vector))
83 (vector #() :type simple-vector)
84 ;; The bitmask used to calculate (mod (* line-size line-hash) (length vector))).
86 ;; Current probe-depth needed in the cache.
88 ;; Maximum allowed probe-depth before the cache needs to expand.
89 (limit 0 :type index))
91 (defun compute-cache-mask (vector-length line-size)
92 ;; Since both vector-length and line-size are powers of two, we
93 ;; can compute a bitmask such that
95 ;; (logand <mask> <combined-layout-hash>)
97 ;; is "morally equal" to
99 ;; (mod (* <line-size> <combined-layout-hash>) <vector-length>)
101 ;; This is it: (1- vector-length) is #b111... of the approriate size
102 ;; to get the MOD, and (- line-size) gives right the number of zero
103 ;; bits at the low end.
104 (logand (1- vector-length) (- line-size)))
106 ;;; The smallest power of two that is equal to or greater then X.
107 (declaim (inline power-of-two-ceiling))
108 (defun power-of-two-ceiling (x)
109 (ash 1 (integer-length (1- x))))
111 (defun cache-statistics (cache)
112 (let* ((vector (cache-vector cache))
113 (size (length vector))
114 (line-size (cache-line-size cache))
115 (total-lines (/ size line-size))
116 (free-lines (loop for i from 0 by line-size below size
117 unless (eq (svref vector i) '..empty..)
119 (values (- total-lines free-lines) total-lines
120 (cache-depth cache) (cache-limit cache))))
122 ;;; Don't allocate insanely huge caches.
123 (defconstant +cache-vector-max-length+ (expt 2 14))
125 ;;; Compute the maximum allowed probe depth as a function of cache size.
126 ;;; Cache size refers to number of cache lines, not the length of the
129 ;;; FIXME: It would be nice to take the generic function optimization
130 ;;; policy into account here (speed vs. space.)
131 (declaim (inline compute-limit))
132 (defun compute-limit (size)
133 (ceiling (sqrt (sqrt size))))
135 ;;; Returns VALUE if it is not ..EMPTY.., otherwise executes ELSE:
136 (defmacro non-empty-or (value else)
137 (with-unique-names (n-value)
138 `(let ((,n-value ,value))
139 (if (eq ,n-value '..empty..)
143 ;;; Fast way to check if a thing found at the position of a cache key is one:
144 ;;; it is always either a wrapper, or the ..EMPTY.. symbol.
145 (declaim (inline cache-key-p))
146 (defun cache-key-p (thing)
147 (not (symbolp thing)))
149 (eval-when (:compile-toplevel :load-toplevel :execute)
150 (sb-kernel:define-structure-slot-compare-and-swap compare-and-swap-cache-depth
154 ;;; Utility macro for atomic updates without locking... doesn't
155 ;;; do much right now, and it would be nice to make this more magical.
156 (defmacro compare-and-swap (place old new)
157 (unless (consp place)
158 (error "Don't know how to compare and swap ~S." place))
161 `(simple-vector-compare-and-swap ,@(cdr place) ,old ,new))
163 `(compare-and-swap-cache-depth ,@(cdr place) ,old ,new))))
165 ;;; Atomically update the current probe depth of a cache.
166 (defun note-cache-depth (cache depth)
167 (loop for old = (cache-depth cache)
168 while (and (< old depth)
169 (not (eq old (compare-and-swap (cache-depth cache)
172 ;;; Compute the starting index of the next cache line in the cache vector.
173 (declaim (inline next-cache-index))
174 (defun next-cache-index (mask index line-size)
175 (logand mask (+ index line-size)))
177 ;;; Returns the hash-value for layout, or executes ELSE if the layout
179 (defmacro hash-layout-or (layout else)
180 (with-unique-names (n-hash)
181 `(let ((,n-hash (layout-clos-hash ,layout)))
186 ;;; Compute cache index for the cache and a list of layouts.
187 (declaim (inline compute-cache-index))
188 (defun compute-cache-index (cache layouts)
189 (let ((index (hash-layout-or (car layouts)
190 (return-from compute-cache-index nil))))
191 (declare (fixnum index))
192 (dolist (layout (cdr layouts))
193 (mixf index (hash-layout-or layout (return-from compute-cache-index nil))))
194 ;; align with cache lines
195 (logand index (cache-mask cache))))
197 ;;; Emit code that does lookup in cache bound to CACHE-VAR using
198 ;;; layouts bound to LAYOUT-VARS. Go to MISS-TAG on event of a miss or
199 ;;; invalid layout. Otherwise, if VALUE-VAR is non-nil, set it to the
200 ;;; value found. (VALUE-VAR is non-nil only when CACHE-VALUE is true.)
202 ;;; In other words, produces inlined code for COMPUTE-CACHE-INDEX when
203 ;;; number of keys and presence of values in the cache is known
205 (defun emit-cache-lookup (cache-var layout-vars miss-tag value-var)
206 (let ((line-size (power-of-two-ceiling (+ (length layout-vars)
207 (if value-var 1 0)))))
208 (with-unique-names (n-index n-vector n-depth n-pointer n-mask
209 MATCH-WRAPPERS EXIT-WITH-HIT)
210 `(let* ((,n-index (hash-layout-or ,(car layout-vars) (go ,miss-tag)))
211 (,n-vector (cache-vector ,cache-var))
212 (,n-mask (cache-mask ,cache-var)))
213 (declare (index ,n-index))
214 ,@(mapcar (lambda (layout-var)
215 `(mixf ,n-index (hash-layout-or ,layout-var (go ,miss-tag))))
217 ;; align with cache lines
218 (setf ,n-index (logand ,n-index ,n-mask))
219 (let ((,n-depth (cache-depth ,cache-var))
220 (,n-pointer ,n-index))
221 (declare (index ,n-depth ,n-pointer))
227 (eq ,layout-var (svref ,n-vector ,n-pointer))
231 `((setf ,value-var (non-empty-or (svref ,n-vector ,n-pointer)
237 (setf ,n-index (next-cache-index ,n-mask ,n-index ,line-size)
242 ;;; Probes CACHE for LAYOUTS.
244 ;;; Returns two values: a boolean indicating a hit or a miss, and a secondary
245 ;;; value that is the value that was stored in the cache if any.
246 (defun probe-cache (cache layouts)
247 (unless (consp layouts)
248 (setf layouts (list layouts)))
249 (let ((vector (cache-vector cache))
250 (key-count (cache-key-count cache))
251 (line-size (cache-line-size cache))
252 (mask (cache-mask cache)))
253 (flet ((probe-line (base)
255 (loop for offset from 0 below key-count
256 for layout in layouts do
257 (unless (eq layout (svref vector (+ base offset)))
260 ;; all layouts match!
261 (let ((value (when (cache-value cache)
262 (non-empty-or (svref vector (+ base key-count))
264 (return-from probe-cache (values t value)))
266 (return-from probe-line (next-cache-index mask base line-size)))))
267 (let ((index (compute-cache-index cache layouts)))
269 (loop repeat (1+ (cache-depth cache)) do
270 (setf index (probe-line index)))))))
273 ;;; Tries to write LAYOUTS and VALUE at the cache line starting at
274 ;;; the index BASE. Returns true on success, and false on failure.
275 (defun try-update-cache-line (cache base layouts value)
276 (declare (index base))
277 (let ((vector (cache-vector cache))
279 ;; If we unwind from here, we will be left with an incomplete
280 ;; cache line, but that is OK: next write using the same layouts
281 ;; will fill it, and reads will treat an incomplete line as a
282 ;; miss -- causing it to be filled.
283 (loop for old = (compare-and-swap (svref vector base) '..empty.. new) do
284 (when (and (cache-key-p old) (not (eq old new)))
285 ;; The place was already taken, and doesn't match our key.
286 (return-from try-update-cache-line nil))
288 ;; All keys match or succesfully saved, save our value --
289 ;; just smash it in. Until the first time it is written
290 ;; there is ..EMPTY.. here, which probes look for, so we
291 ;; don't get bogus hits. This is necessary because we want
292 ;; to be able store arbitrary values here for use with
293 ;; constant-value dispatch functions.
294 (when (cache-value cache)
295 (setf (svref vector (1+ base)) value))
296 (return-from try-update-cache-line t))
297 (setf new (pop layouts))
300 ;;; Tries to write LAYOUTS and VALUE somewhere in the cache. Returns
301 ;;; true on success and false on failure, meaning the cache is too
303 (defun try-update-cache (cache layouts value)
304 (let ((vector (cache-vector cache))
305 (index (or (compute-cache-index cache layouts)
306 ;; At least one of the layouts was invalid: just
307 ;; pretend we updated the cache, and let the next
308 ;; read pick up the mess.
309 (return-from try-update-cache t)))
310 (line-size (cache-line-size cache))
311 (mask (cache-mask cache)))
312 (declare (index index))
313 (loop for depth from 0 upto (cache-limit cache) do
314 (when (try-update-cache-line cache index layouts value)
315 (note-cache-depth cache depth)
316 (return-from try-update-cache t))
317 (setf index (next-cache-index mask index line-size)))))
319 ;;; Constructs a new cache.
320 (defun make-cache (&key (key-count (missing-arg)) (value (missing-arg))
322 (let* ((line-size (power-of-two-ceiling (+ key-count (if value 1 0))))
323 (adjusted-size (power-of-two-ceiling size))
324 (length (* adjusted-size line-size)))
325 (if (<= length +cache-vector-max-length+)
326 (%make-cache :key-count key-count
328 :vector (make-array length :initial-element '..empty..)
330 :mask (compute-cache-mask length line-size)
331 :limit (compute-limit adjusted-size))
332 ;; Make a smaller one, then
333 (make-cache :key-count key-count :value value :size (ceiling size 2)))))
335 ;;;; Copies and expands the cache, dropping any invalidated or
336 ;;;; incomplete lines.
337 (defun copy-and-expand-cache (cache)
338 (let ((copy (%copy-cache cache))
339 (length (length (cache-vector cache))))
340 (when (< length +cache-vector-max-length+)
341 (setf length (* 2 length)))
344 (setf (cache-vector copy) (make-array length :initial-element '..empty..)
346 (cache-mask copy) (compute-cache-mask length (cache-line-size cache))
347 (cache-limit copy) (compute-limit (/ length (cache-line-size cache))))
348 (map-cache (lambda (layouts value)
349 (unless (try-update-cache copy layouts value)
350 ;; If the cache would grow too much we drop the
351 ;; remaining the entries that don't fit. FIXME:
352 ;; It would be better to drop random entries to
353 ;; avoid getting into a rut here (best done by
354 ;; making MAP-CACHE map in a random order?), and
355 ;; possibly to downsize the cache more
356 ;; aggressively (on the assumption that most
357 ;; entries aren't getting used at the moment.)
358 (when (< length +cache-vector-max-length+)
359 (setf length (* 2 length))
364 (defun cache-has-invalid-entries-p (cache)
365 (let ((vector (cache-vector cache))
366 (line-size (cache-line-size cache))
367 (key-count (cache-key-count cache))
368 (mask (cache-mask cache))
371 ;; Check if the line is in use, and check validity of the keys.
372 (let ((key1 (svref vector index)))
373 (when (cache-key-p key1)
374 (if (zerop (layout-clos-hash key1))
375 ;; First key invalid.
376 (return-from cache-has-invalid-entries-p t)
377 ;; Line is in use and the first key is valid: check the rest.
378 (loop for offset from 1 below key-count
379 do (let ((thing (svref vector (+ index offset))))
380 (when (or (not (cache-key-p thing))
381 (zerop (layout-clos-hash thing)))
382 ;; Incomplete line or invalid layout.
383 (return-from cache-has-invalid-entries-p t)))))))
384 ;; Line empty of valid, onwards.
385 (setf index (next-cache-index mask index line-size))
388 (return-from cache-has-invalid-entries-p nil)))))
390 (defun hash-table-to-cache (table &key value key-count)
391 (let ((cache (make-cache :key-count key-count :value value
392 :size (hash-table-count table))))
393 (maphash (lambda (class value)
394 (setq cache (fill-cache cache (class-wrapper class) value)))
398 ;;; Inserts VALUE to CACHE keyd by LAYOUTS. Expands the cache if
399 ;;; necessary, and returns the new cache.
400 (defun fill-cache (cache layouts value)
402 ((%fill-cache (cache layouts value)
403 (cond ((try-update-cache cache layouts value)
405 ((cache-has-invalid-entries-p cache)
406 ;; Don't expand yet: maybe there will be enough space if
407 ;; we just drop the invalid entries.
408 (%fill-cache (copy-cache cache) layouts value))
410 (%fill-cache (copy-and-expand-cache cache) layouts value)))))
412 (%fill-cache cache layouts value)
413 (%fill-cache cache (list layouts) value))))
415 ;;; Calls FUNCTION with all layouts and values in cache.
416 (defun map-cache (function cache)
417 (let* ((vector (cache-vector cache))
418 (key-count (cache-key-count cache))
419 (valuep (cache-value cache))
420 (line-size (cache-line-size cache))
421 (mask (cache-mask cache))
422 (fun (if (functionp function)
424 (fdefinition function)))
430 (loop for offset from 0 below key-count
431 collect (non-empty-or (svref vector (+ offset index))
433 (let ((value (when valuep
434 (non-empty-or (svref vector (+ index key-count))
436 ;; Let the callee worry about invalid layouts
437 (funcall fun layouts value)))
439 (setf index (next-cache-index mask index line-size))
440 (unless (zerop index)
444 ;;; Copying a cache without expanding it is very much like mapping it:
445 ;;; we need to be carefull because there may be updates while we are
446 ;;; copying it, and we don't want to copy incomplete entries or invalid
448 (defun copy-cache (cache)
449 (let* ((vector (cache-vector cache))
450 (copy (make-array (length vector) :initial-element '..empty..))
451 (line-size (cache-line-size cache))
452 (key-count (cache-key-count cache))
453 (valuep (cache-value cache))
454 (mask (cache-mask cache))
455 (size (/ (length vector) line-size))
461 (let ((layouts (loop for offset from 0 below key-count
462 collect (non-empty-or (svref vector (+ index offset))
464 ;; Check validity & compute primary index.
465 (let ((primary (or (compute-cache-index cache layouts)
467 ;; Check & copy value.
469 (setf (svref copy (+ index key-count))
470 (non-empty-or (svref vector (+ index key-count))
473 (loop for offset from 0 below key-count do
474 (setf (svref copy (+ index offset)) (pop layouts)))
475 ;; Update probe depth.
476 (let ((distance (/ (- index primary) line-size)))
477 (setf depth (max depth (if (minusp distance)
478 ;; account for wrap-around
482 (setf index (next-cache-index mask index line-size))
483 (unless (zerop index)
485 (%make-cache :vector copy
487 :key-count (cache-key-count cache)
491 :limit (cache-limit cache))))
493 ;;;; For debugging & collecting statistics.
495 (defun map-all-caches (function)
496 (dolist (p (list-all-packages))
498 (when (eq p (symbol-package s))
499 (dolist (name (list s
503 (slot-boundp-name s)))
505 (let ((fun (fdefinition name)))
506 (when (typep fun 'generic-function)
507 (let ((cache (gf-dfun-cache fun)))
509 (funcall function name cache)))))))))))
511 (defun check-cache-consistency (cache)
512 (let ((table (make-hash-table :test 'equal)))
513 (map-cache (lambda (layouts value)
514 (declare (ignore value))
515 (if (gethash layouts table)
516 (cerror "Check futher."
517 "Multiple appearances of ~S." layouts)
518 (setf (gethash layouts table) t)))