1 ;;;; This software is part of the SBCL system. See the README file for
4 ;;;; This software is derived from the CMU CL system, which was
5 ;;;; written at Carnegie Mellon University and released into the
6 ;;;; public domain. The software is in the public domain and is
7 ;;;; provided with absolutely no warranty. See the COPYING and CREDITS
8 ;;;; files for more information.
10 (in-package "SB-PROFILE") ; (SB-, not SB!, since we're built in warm load.)
15 ;;;; Thread safe, and reasonably fast: in common case increment is just an
16 ;;;; ATOMIC-INCF, in overflow case grab a lock and increment overflow counter.
18 (defconstant +most-positive-word+ (1- (expt 2 sb-vm:n-word-bits)))
20 (declaim (inline make-counter))
21 (defstruct (counter (:constructor make-counter) (:copier nil))
22 (word 0 :type sb-vm:word)
23 (overflow 0 :type unsigned-byte))
25 (defun incf-counter (counter delta)
26 ;; When running multi-threaded we can easily get negative numbers for the
27 ;; cons-counter. Don't count them at all.
29 (labels ((%incf-overflow (&optional (n 1))
30 ;; Overflow-counter can run into bignums... so we need to loop
31 ;; around CAS till the increment succeeds.
32 (loop for old = (counter-overflow counter)
33 until (eq old (compare-and-swap (counter-overflow counter)
36 ;; Increment the word-sized counter. If it overflows, record the
38 (let ((prev (atomic-incf (counter-word counter) d)))
39 (when (< (logand +most-positive-word+ (+ prev d)) prev)
41 ;; DELTA can potentially be a bignum -- cut it down to word-size.
42 (unless (typep delta 'sb-vm:word)
43 (multiple-value-bind (n r) (truncate delta (1+ +most-positive-word+))
46 ;; ATOMIC-INCF can at most handle SIGNED-WORD: if DELTA doesn't fit that,
48 (if (typep delta 'sb-vm:signed-word)
50 ;; ...and if delta is still too big, split it into four parts: they
51 ;; are guaranteed to fit into a signed word.
52 (multiple-value-bind (n r) (truncate delta 2)
58 (defun counter-count (counter)
59 (+ (counter-word counter)
60 (* (counter-overflow counter) (1+ +most-positive-word+))))
62 ;;;; High resolution timer
64 ;;; FIXME: High resolution this is not. Build a microsecond-accuracy version
65 ;;; on top of unix-getrusage, maybe.
67 (defconstant +ticks-per-second+ internal-time-units-per-second)
69 (declaim (inline get-internal-ticks))
70 (defun get-internal-ticks ()
71 (get-internal-run-time))
73 ;;;; global data structures
75 ;;; We associate a PROFILE-INFO structure with each profiled function
76 ;;; name. This holds the functions that we call to manipulate the
77 ;;; closure which implements the encapsulation.
78 (defvar *profiled-fun-name->info*
80 ;; EQL testing isn't good enough for generalized function names
84 (defstruct (profile-info (:copier nil))
85 (name (missing-arg) :read-only t)
86 (encapsulated-fun (missing-arg) :type function :read-only t)
87 (encapsulation-fun (missing-arg) :type function :read-only t)
88 (read-stats-fun (missing-arg) :type function :read-only t)
89 (clear-stats-fun (missing-arg) :type function :read-only t))
91 ;;; These variables are used to subtract out the time and consing for
92 ;;; recursive and other dynamically nested profiled calls. The total
93 ;;; resource consumed for each nested call is added into the
94 ;;; appropriate variable. When the outer function returns, these
95 ;;; amounts are subtracted from the total.
96 (declaim (counter *enclosed-ticks* *enclosed-consing*))
97 (defvar *enclosed-ticks*)
98 (defvar *enclosed-consing*)
100 ;;; This variable is also used to subtract out time for nested
101 ;;; profiled calls. The time inside the profile wrapper call --
102 ;;; between its two calls to GET-INTERNAL-TICKS -- is accounted
103 ;;; for by the *ENCLOSED-TIME* variable. However, there's also extra
104 ;;; overhead involved, before we get to the first call to
105 ;;; GET-INTERNAL-TICKS, and after we get to the second call. By
106 ;;; keeping track of the count of enclosed profiled calls, we can try
107 ;;; to compensate for that.
108 (declaim (counter *enclosed-profiles*))
109 (defvar *enclosed-profiles*)
111 (declaim (counter *enclosed-gc-run-time*))
112 (defvar *enclosed-gc-run-time*)
114 ;;; the encapsulated function we're currently computing profiling data
115 ;;; for, recorded so that we can detect the problem of
116 ;;; PROFILE-computing machinery calling a function which has itself
118 (defvar *computing-profiling-data-for*)
120 ;;; the components of profiling overhead
121 (defstruct (overhead (:copier nil))
122 ;; the number of ticks a bare function call takes. This is
123 ;; factored into the other overheads, but not used for itself.
124 (call (missing-arg) :type single-float :read-only t)
125 ;; the number of ticks that will be charged to a profiled
126 ;; function due to the profiling code
127 (internal (missing-arg) :type single-float :read-only t)
128 ;; the number of ticks of overhead for profiling that a single
129 ;; profiled call adds to the total runtime for the program
130 (total (missing-arg) :type single-float :read-only t))
132 (declaim (type overhead *overhead*))
133 (makunbound '*overhead*) ; in case we reload this file when tweaking
135 ;;;; profile encapsulations
137 ;;; Return a collection of closures over the same lexical context,
138 ;;; (VALUES ENCAPSULATION-FUN READ-STATS-FUN CLEAR-STATS-FUN).
140 ;;; ENCAPSULATION-FUN is a plug-in replacement for ENCAPSULATED-FUN,
141 ;;; which updates statistics whenever it's called.
143 ;;; READ-STATS-FUN returns the statistics:
144 ;;; (VALUES COUNT TIME CONSING PROFILE).
145 ;;; COUNT is the count of calls to ENCAPSULATION-FUN. TICKS is
146 ;;; the total number of ticks spent in ENCAPSULATED-FUN.
147 ;;; CONSING is the total consing of ENCAPSULATION-FUN. PROFILE is the
148 ;;; number of calls to the profiled function, stored for the purposes
149 ;;; of trying to estimate that part of profiling overhead which occurs
150 ;;; outside the interval between the profile wrapper function's timer
153 ;;; CLEAR-STATS-FUN clears the statistics.
155 ;;; (The reason for implementing this as coupled closures, with the
156 ;;; counts built into the lexical environment, is that we hope this
157 ;;; will minimize profiling overhead.)
158 (defun profile-encapsulation-lambdas (encapsulated-fun)
159 (declare (type function encapsulated-fun))
160 (let* ((count (make-counter))
161 (ticks (make-counter))
162 (consing (make-counter))
163 (profiles (make-counter))
164 (gc-run-time (make-counter)))
165 (declare (counter count ticks consing profiles gc-run-time))
168 (lambda (&more arg-context arg-count)
169 (declare (optimize speed safety))
170 ;; Make sure that we're not recursing infinitely.
171 (when (boundp '*computing-profiling-data-for*)
172 (unprofile-all) ; to avoid further recursion
173 (error "~@<When computing profiling data for ~S, the profiled ~
174 function ~S was called. To get out of this infinite recursion, all ~
175 functions have been unprofiled. (Since the profiling system evidently ~
176 uses ~S in its computations, it looks as though it's a bad idea to ~
178 *computing-profiling-data-for* encapsulated-fun
180 (incf-counter count 1)
183 (inner-enclosed-profiles 0)
185 (declare (truly-dynamic-extent dticks dconsing inner-enclosed-profiles))
187 (let* ((start-ticks (get-internal-ticks))
188 (start-gc-run-time *gc-run-time*)
189 (*enclosed-ticks* (make-counter))
190 (*enclosed-consing* (make-counter))
191 (*enclosed-profiles* (make-counter))
192 (nbf0 *n-bytes-freed-or-purified*)
193 (dynamic-usage-0 (sb-kernel:dynamic-usage))
194 (*enclosed-gc-run-time* (make-counter)))
195 (declare (dynamic-extent *enclosed-ticks* *enclosed-consing* *enclosed-profiles* *enclosed-gc-run-time*))
197 (multiple-value-call encapsulated-fun
198 (sb-c:%more-arg-values arg-context
201 (let ((*computing-profiling-data-for* encapsulated-fun)
202 (dynamic-usage-1 (sb-kernel:dynamic-usage)))
203 (setf dticks (- (get-internal-ticks) start-ticks)
204 dconsing (if (eql *n-bytes-freed-or-purified* nbf0)
205 ;; common special case where we can avoid
207 (- dynamic-usage-1 dynamic-usage-0)
209 (- (get-bytes-consed) nbf0 dynamic-usage-0))
210 inner-enclosed-profiles (counter-count *enclosed-profiles*)
211 dgc-run-time (- *gc-run-time* start-gc-run-time))
212 (incf-counter ticks (- dticks (counter-count *enclosed-ticks*)))
213 (incf-counter gc-run-time (- dgc-run-time (counter-count *enclosed-gc-run-time*)))
214 (incf-counter consing (- dconsing (counter-count *enclosed-consing*)))
215 (incf-counter profiles inner-enclosed-profiles))))
216 (when (boundp '*enclosed-ticks*)
217 (incf-counter *enclosed-ticks* dticks)
218 (incf-counter *enclosed-consing* dconsing)
219 (incf-counter *enclosed-profiles* (1+ inner-enclosed-profiles))
220 (incf-counter *enclosed-gc-run-time* dgc-run-time)))))
223 (values (counter-count count)
224 (counter-count ticks)
225 (counter-count consing)
226 (counter-count profiles)
227 (counter-count gc-run-time)))
230 (setf count (make-counter)
232 consing (make-counter)
233 profiles (make-counter)
234 gc-run-time (make-counter))))))
238 ;;; A symbol or (SETF FOO) list names a function, a string names all
239 ;;; the functions named by symbols in the named package.
240 (defun mapc-on-named-funs (function names)
243 (symbol (funcall function name))
245 (legal-fun-name-or-type-error name)
246 ;; Then we map onto it.
247 (funcall function name))
248 (string (let ((package (find-undeleted-package-or-lose name)))
249 (do-symbols (symbol package)
250 (when (eq (symbol-package symbol) package)
251 (when (and (fboundp symbol)
252 (not (macro-function symbol))
253 (not (special-operator-p symbol)))
254 (funcall function symbol))
255 (let ((setf-name `(setf ,symbol)))
256 (when (fboundp setf-name)
257 (funcall function setf-name)))))))))
260 ;;; Profile the named function, which should exist and not be profiled
262 (defun profile-1-unprofiled-fun (name)
263 (let ((encapsulated-fun (fdefinition name)))
264 (multiple-value-bind (encapsulation-fun read-stats-fun clear-stats-fun)
265 (profile-encapsulation-lambdas encapsulated-fun)
266 (without-package-locks
267 (setf (fdefinition name)
269 (setf (gethash name *profiled-fun-name->info*)
270 (make-profile-info :name name
271 :encapsulated-fun encapsulated-fun
272 :encapsulation-fun encapsulation-fun
273 :read-stats-fun read-stats-fun
274 :clear-stats-fun clear-stats-fun))
277 ;;; Profile the named function. If already profiled, unprofile first.
278 (defun profile-1-fun (name)
279 (cond ((fboundp name)
280 (when (gethash name *profiled-fun-name->info*)
281 (warn "~S is already profiled, so unprofiling it first." name)
282 (unprofile-1-fun name))
283 (profile-1-unprofiled-fun name))
285 (warn "ignoring undefined function ~S" name)))
288 ;;; Unprofile the named function, if it is profiled.
289 (defun unprofile-1-fun (name)
290 (let ((pinfo (gethash name *profiled-fun-name->info*)))
292 (remhash name *profiled-fun-name->info*)
293 (if (eq (fdefinition name) (profile-info-encapsulation-fun pinfo))
294 (without-package-locks
295 (setf (fdefinition name) (profile-info-encapsulated-fun pinfo)))
296 (warn "preserving current definition of redefined function ~S"
299 (warn "~S is not a profiled function." name))))
302 (defmacro profile (&rest names)
306 If no names are supplied, return the list of profiled functions.
308 If names are supplied, wrap profiling code around the named functions.
309 As in TRACE, the names are not evaluated. A symbol names a function.
310 A string names all the functions named by symbols in the named
311 package. If a function is already profiled, then unprofile and
312 reprofile (useful to notice function redefinition.) If a name is
313 undefined, then we give a warning and ignore it. See also
314 UNPROFILE, REPORT and RESET."
316 `(loop for k being each hash-key in *profiled-fun-name->info*
318 `(mapc-on-named-funs #'profile-1-fun ',names)))
320 (defmacro unprofile (&rest names)
322 "Unwrap any profiling code around the named functions, or if no names
323 are given, unprofile all profiled functions. A symbol names
324 a function. A string names all the functions named by symbols in the
325 named package. NAMES defaults to the list of names of all currently
328 `(mapc-on-named-funs #'unprofile-1-fun ',names)
331 (defun unprofile-all ()
332 (dohash ((name profile-info) *profiled-fun-name->info*
334 (declare (ignore profile-info))
335 (unprofile-1-fun name)))
338 "Reset the counters for all profiled functions."
339 (dohash ((name profile-info) *profiled-fun-name->info* :locked t)
340 (declare (ignore name))
341 (funcall (profile-info-clear-stats-fun profile-info))))
343 ;;;; reporting results
345 (defstruct (time-info (:copier nil))
352 ;;; Return our best guess for the run time in a function, subtracting
353 ;;; out factors for profiling overhead. We subtract out the internal
354 ;;; overhead for each call to this function, since the internal
355 ;;; overhead is the part of the profiling overhead for a function that
356 ;;; is charged to that function.
358 ;;; We also subtract out a factor for each call to a profiled function
359 ;;; within this profiled function. This factor is the total profiling
360 ;;; overhead *minus the internal overhead*. We don't subtract out the
361 ;;; internal overhead, since it was already subtracted when the nested
362 ;;; profiled functions subtracted their running time from the time for
363 ;;; the enclosing function.
364 (defun compensate-time (calls ticks profile)
365 (let ((raw-compensated
366 (- (/ (float ticks) (float +ticks-per-second+))
367 (* (overhead-internal *overhead*) (float calls))
368 (* (- (overhead-total *overhead*)
369 (overhead-internal *overhead*))
371 (max raw-compensated 0.0)))
373 (defun report (&key limit (print-no-call-list t))
374 "Report results from profiling. The results are approximately
375 adjusted for profiling overhead. The compensation may be rather
376 inaccurate when bignums are involved in runtime calculation, as in a
377 very-long-running Lisp process.
379 If LIMIT is set to an integer, only the top LIMIT results are
380 reported. If PRINT-NO-CALL-LIST is T (the default) then a list of
381 uncalled profiled functions are listed."
382 (unless (boundp '*overhead*)
385 (let ((time-info-list ())
386 (no-call-name-list ()))
387 (dohash ((name pinfo) *profiled-fun-name->info* :locked t)
388 (unless (eq (fdefinition name)
389 (profile-info-encapsulation-fun pinfo))
390 (warn "Function ~S has been redefined, so times may be inaccurate.~@
391 PROFILE it again to record calls to the new definition."
393 (multiple-value-bind (calls ticks consing profile gc-run-time)
394 (funcall (profile-info-read-stats-fun pinfo))
396 (push name no-call-name-list)
397 (push (make-time-info :name name
399 :seconds (compensate-time calls
403 :gc-run-time gc-run-time)
409 :key #'time-info-seconds)))
411 (if (and limit (> (length times) limit))
412 (subseq times 0 limit)
415 (when (and print-no-call-list no-call-name-list)
416 (format *trace-output*
417 "~%These functions were not called:~%~{~<~%~:; ~S~>~}~%"
418 (sort no-call-name-list #'string<
420 (symbol-name (fun-name-block-name name))))))
425 (defun print-profile-table (time-info-list)
426 (let ((total-seconds 0.0)
429 (total-gc-run-time 0)
430 (seconds-width (length "seconds"))
431 (consed-width (length "consed"))
432 (calls-width (length "calls"))
434 (gc-run-time-width (length "gc"))
436 (dolist (time-info time-info-list)
437 (incf total-seconds (time-info-seconds time-info))
438 (incf total-consed (time-info-consing time-info))
439 (incf total-calls (time-info-calls time-info))
440 (incf total-gc-run-time (time-info-gc-run-time time-info)))
441 (setf seconds-width (max (length (format nil "~10,3F" total-seconds))
443 calls-width (max (length (format nil "~:D" total-calls))
445 consed-width (max (length (format nil "~:D" total-consed))
447 gc-run-time-width (max (length (format nil "~10,3F" (/ total-gc-run-time internal-time-units-per-second)))
451 (dotimes (i (+ seconds-width consed-width calls-width
452 sec/call-width name-width
454 (write-char #\- *trace-output*))
455 (terpri *trace-output*)))
456 (format *trace-output* "~&~@{ ~v:@<~A~>~^|~}~%"
457 seconds-width "seconds"
458 (1+ gc-run-time-width) "gc"
459 (1+ consed-width) "consed"
460 (1+ calls-width) "calls"
461 (1+ sec/call-width) "sec/call"
462 (1+ name-width) "name")
466 (dolist (time-info time-info-list)
467 (format *trace-output* "~v,3F | ~v,3F | ~v:D | ~v:D | ~10,6F | ~S~%"
468 seconds-width (time-info-seconds time-info)
469 gc-run-time-width (/ (time-info-gc-run-time time-info) internal-time-units-per-second)
470 consed-width (time-info-consing time-info)
471 calls-width (time-info-calls time-info)
472 (/ (time-info-seconds time-info)
473 (float (time-info-calls time-info)))
474 (time-info-name time-info)))
478 (format *trace-output* "~v,3F | ~v,3F | ~v:D | ~v:D | | Total~%"
479 seconds-width total-seconds
480 gc-run-time-width (/ total-gc-run-time internal-time-units-per-second)
481 consed-width total-consed
482 calls-width total-calls)
484 (format *trace-output*
485 "~%estimated total profiling overhead: ~4,2F seconds~%"
486 (* (overhead-total *overhead*) (float total-calls)))
487 (format *trace-output*
488 "~&overhead estimation parameters:~% ~Ss/call, ~Ss total profiling, ~Ss internal profiling~%"
489 (overhead-call *overhead*)
490 (overhead-total *overhead*)
491 (overhead-internal *overhead*)))))
494 ;;;; overhead estimation
496 ;;; We average the timing overhead over this many iterations.
498 ;;; (This is a variable, not a constant, so that it can be set in
499 ;;; .sbclrc if desired. Right now, that's an unsupported extension
500 ;;; that I (WHN) use for my own experimentation, but it might
501 ;;; become supported someday. Comments?)
502 (declaim (type unsigned-byte *timer-overhead-iterations*))
503 (defparameter *timer-overhead-iterations*
506 ;;; a dummy function that we profile to find profiling overhead
507 (declaim (notinline compute-overhead-aux))
508 (defun compute-overhead-aux (x)
509 (declare (ignore x)))
511 ;;; Return a newly computed OVERHEAD object.
512 (defun compute-overhead ()
513 (format *debug-io* "~&measuring PROFILE overhead..")
515 (let ((start (get-internal-ticks))
516 (fun (symbol-function 'compute-overhead-aux)))
517 (declare (type function fun))
518 (dotimes (i *timer-overhead-iterations*)
520 (/ (float (- (get-internal-ticks) start))
521 (float +ticks-per-second+)
522 (float *timer-overhead-iterations*)))))
523 (let (;; Measure unprofiled calls to estimate call overhead.
524 (call-overhead (frob))
527 ;; Measure profiled calls to estimate profiling overhead.
530 (profile compute-overhead-aux)
532 (- (frob) call-overhead)))
533 (let* ((pinfo (gethash 'compute-overhead-aux
534 *profiled-fun-name->info*))
535 (read-stats-fun (profile-info-read-stats-fun pinfo))
536 (time (nth-value 1 (funcall read-stats-fun))))
537 (setf internal-overhead
539 (float +ticks-per-second+)
540 (float *timer-overhead-iterations*))))
541 (unprofile compute-overhead-aux))
543 (make-overhead :call call-overhead
544 :total total-overhead
545 :internal internal-overhead)
546 (format *debug-io* "done~%")))))
548 ;;; It would be bad to compute *OVERHEAD*, save it into a .core file,
549 ;;; then load the old *OVERHEAD* value from the .core file into a
550 ;;; different machine running at a different speed. We avoid this by
551 ;;; erasing *CALL-OVERHEAD* whenever we save a .core file.
552 (defun profile-deinit ()
553 (without-package-locks
554 (makunbound '*overhead*)))