1 ;;;; (statprof) -- a statistical profiler for Guile
4 ;;;; Copyright (C) 2009, 2010, 2011, 2013, 2014 Free Software Foundation, Inc.
5 ;;;; Copyright (C) 2004, 2009 Andy Wingo <wingo at pobox dot com>
6 ;;;; Copyright (C) 2001 Rob Browning <rlb at defaultvalue dot org>
8 ;;;; This library is free software; you can redistribute it and/or
9 ;;;; modify it under the terms of the GNU Lesser General Public
10 ;;;; License as published by the Free Software Foundation; either
11 ;;;; version 3 of the License, or (at your option) any later version.
13 ;;;; This library is distributed in the hope that it will be useful,
14 ;;;; but WITHOUT ANY WARRANTY; without even the implied warranty of
15 ;;;; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 ;;;; Lesser General Public License for more details.
18 ;;;; You should have received a copy of the GNU Lesser General Public
19 ;;;; License along with this library; if not, write to the Free Software
20 ;;;; Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
26 ;;; @code{(statprof)} is a statistical profiler for Guile.
28 ;;; A simple use of statprof would look like this:
31 ;;; (statprof-reset 0 50000 #t)
35 ;;; (statprof-display)
38 ;;; This would reset statprof, clearing all accumulated statistics, then
39 ;;; start profiling, run some code, stop profiling, and finally display a
40 ;;; gprof flat-style table of statistics which will look something like
44 ;;; % cumulative self self total
45 ;;; time seconds seconds calls ms/call ms/call name
46 ;;; 35.29 0.23 0.23 2002 0.11 0.11 -
47 ;;; 23.53 0.15 0.15 2001 0.08 0.08 positive?
48 ;;; 23.53 0.15 0.15 2000 0.08 0.08 +
49 ;;; 11.76 0.23 0.08 2000 0.04 0.11 do-nothing
50 ;;; 5.88 0.64 0.04 2001 0.02 0.32 loop
51 ;;; 0.00 0.15 0.00 1 0.00 150.59 do-something
55 ;;; All of the numerical data with the exception of the calls column is
56 ;;; statistically approximate. In the following column descriptions, and
57 ;;; in all of statprof, "time" refers to execution time (both user and
58 ;;; system), not wall clock time.
62 ;;; The percent of the time spent inside the procedure itself
63 ;;; (not counting children).
64 ;;; @item cumulative seconds
65 ;;; The total number of seconds spent in the procedure, including
67 ;;; @item self seconds
68 ;;; The total number of seconds spent in the procedure itself (not counting
71 ;;; The total number of times the procedure was called.
72 ;;; @item self ms/call
73 ;;; The average time taken by the procedure itself on each call, in ms.
74 ;;; @item total ms/call
75 ;;; The average time taken by each call to the procedure, including time
76 ;;; spent in child functions.
78 ;;; The name of the procedure.
81 ;;; The profiler uses @code{eq?} and the procedure object itself to
82 ;;; identify the procedures, so it won't confuse different procedures with
83 ;;; the same name. They will show up as two different rows in the output.
85 ;;; Right now the profiler is quite simplistic. I cannot provide
86 ;;; call-graphs or other higher level information. What you see in the
87 ;;; table is pretty much all there is. Patches are welcome :-)
89 ;;; @section Implementation notes
91 ;;; The profiler works by setting the unix profiling signal
92 ;;; @code{ITIMER_PROF} to go off after the interval you define in the call
93 ;;; to @code{statprof-reset}. When the signal fires, a sampling routine is
94 ;;; run which looks at the current procedure that's executing, and then
95 ;;; crawls up the stack, and for each procedure encountered, increments
96 ;;; that procedure's sample count. Note that if a procedure is encountered
97 ;;; multiple times on a given stack, it is only counted once. After the
98 ;;; sampling is complete, the profiler resets profiling timer to fire
99 ;;; again after the appropriate interval.
101 ;;; Meanwhile, the profiler keeps track, via @code{get-internal-run-time},
102 ;;; how much CPU time (system and user -- which is also what
103 ;;; @code{ITIMER_PROF} tracks), has elapsed while code has been executing
104 ;;; within a statprof-start/stop block.
106 ;;; The profiler also tries to avoid counting or timing its own code as
107 ;;; much as possible.
111 (define-module (statprof)
112 #:use-module (srfi srfi-1)
113 #:use-module (srfi srfi-9)
114 #:autoload (ice-9 format) (format)
115 #:use-module (system vm vm)
116 #:use-module (system vm frame)
117 #:use-module (system vm program)
118 #:export (statprof-active?
123 statprof-accumulated-time
124 statprof-sample-count
125 statprof-fold-call-data
126 statprof-proc-call-data
127 statprof-call-data-name
128 statprof-call-data-calls
129 statprof-call-data-cum-samples
130 statprof-call-data-self-samples
131 statprof-call-data->stats
133 statprof-stats-proc-name
134 statprof-stats-%-time-in-proc
135 statprof-stats-cum-secs-in-proc
136 statprof-stats-self-secs-in-proc
138 statprof-stats-self-secs-per-call
139 statprof-stats-cum-secs-per-call
142 statprof-display-anomolies
144 statprof-fetch-stacks
145 statprof-fetch-call-tree
153 ;; This profiler tracks two numbers for every function called while
154 ;; it's active. It tracks the total number of calls, and the number
155 ;; of times the function was active when the sampler fired.
157 ;; Globally the profiler tracks the total time elapsed and the number
158 ;; of times the sampler was fired.
160 ;; Right now, this profiler is not per-thread and is not thread safe.
162 (define-record-type <state>
163 (make-state accumulated-time last-start-time sample-count
164 sampling-frequency remaining-prof-time profile-level
165 count-calls? gc-time-taken record-full-stacks?
166 stacks procedure-data inside-profiler?)
168 ;; Total time so far.
169 (accumulated-time accumulated-time set-accumulated-time!)
170 ;; Start-time when timer is active.
171 (last-start-time last-start-time set-last-start-time!)
172 ;; Total count of sampler calls.
173 (sample-count sample-count set-sample-count!)
174 ;; (seconds . microseconds)
175 (sampling-frequency sampling-frequency set-sampling-frequency!)
176 ;; Time remaining when prof suspended.
177 (remaining-prof-time remaining-prof-time set-remaining-prof-time!)
178 ;; For user start/stop nesting.
179 (profile-level profile-level set-profile-level!)
180 ;; Whether to catch apply-frame.
181 (count-calls? count-calls? set-count-calls?!)
182 ;; GC time between statprof-start and statprof-stop.
183 (gc-time-taken gc-time-taken set-gc-time-taken!)
184 ;; If #t, stash away the stacks for future analysis.
185 (record-full-stacks? record-full-stacks? set-record-full-stacks?!)
186 ;; If record-full-stacks?, the stashed full stacks.
187 (stacks stacks set-stacks!)
188 ;; A hash where the key is the function object itself and the value is
189 ;; the data. The data will be a vector like this:
190 ;; #(name call-count cum-sample-count self-sample-count)
191 (procedure-data procedure-data set-procedure-data!)
192 ;; True if we are inside the profiler.
193 (inside-profiler? inside-profiler? set-inside-profiler?!))
195 (define profiler-state (make-parameter #f))
197 (define* (fresh-profiler-state #:key (count-calls? #f)
198 (sampling-frequency '(0 . 10000))
200 (make-state 0.0 #f 0 sampling-frequency #f 0 count-calls? 0.0 #f '()
201 (make-hash-table) #f))
203 (define (ensure-profiler-state)
205 (let ((state (fresh-profiler-state)))
206 (profiler-state state)
209 (define (existing-profiler-state)
211 (error "expected there to be a profiler state")))
213 ;; If you change the call-data data structure, you need to also change
214 ;; sample-uncount-frame.
215 (define (make-call-data proc call-count cum-sample-count self-sample-count)
216 (vector proc call-count cum-sample-count self-sample-count))
217 (define (call-data-proc cd) (vector-ref cd 0))
218 (define (call-data-name cd) (procedure-name (call-data-proc cd)))
219 (define (call-data-printable cd)
220 (or (call-data-name cd)
221 (with-output-to-string (lambda () (write (call-data-proc cd))))))
222 (define (call-data-call-count cd) (vector-ref cd 1))
223 (define (call-data-cum-sample-count cd) (vector-ref cd 2))
224 (define (call-data-self-sample-count cd) (vector-ref cd 3))
226 (define (inc-call-data-call-count! cd)
227 (vector-set! cd 1 (1+ (vector-ref cd 1))))
228 (define (inc-call-data-cum-sample-count! cd)
229 (vector-set! cd 2 (1+ (vector-ref cd 2))))
230 (define (inc-call-data-self-sample-count! cd)
231 (vector-set! cd 3 (1+ (vector-ref cd 3))))
233 (define (accumulate-time state stop-time)
234 (set-accumulated-time! state
235 (+ (accumulated-time state)
236 (- stop-time (last-start-time state)))))
238 (define (get-call-data proc)
239 (define state (ensure-profiler-state))
241 ((program? proc) (program-code proc))
243 (or (hashv-ref (procedure-data state) k)
244 (let ((call-data (make-call-data proc 0 0 0)))
245 (hashv-set! (procedure-data state) k call-data)
248 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
251 (define (sample-stack-procs stack)
252 (let ((stacklen (stack-length stack))
254 (state (existing-profiler-state)))
256 (if (record-full-stacks? state)
257 (set-stacks! state (cons stack (stacks state))))
259 (set-sample-count! state (+ (sample-count state) 1))
260 ;; Now accumulate stats for the whole stack.
261 (let loop ((frame (stack-ref stack 0))
262 (procs-seen (make-hash-table 13))
267 (lambda (proc val accum)
268 (inc-call-data-cum-sample-count!
269 (get-call-data proc)))
272 (and=> (and=> self get-call-data)
273 inc-call-data-self-sample-count!))
274 ((frame-procedure frame)
277 ((eq? proc count-call)
278 ;; We're not supposed to be sampling count-call and
279 ;; its sub-functions, so loop again with a clean
281 (set! hit-count-call? #t)
282 (loop (frame-previous frame) (make-hash-table 13) #f))
284 (hashq-set! procs-seen proc #t)
285 (loop (frame-previous frame)
289 (loop (frame-previous frame) procs-seen self))))
292 (define (profile-signal-handler sig)
293 (define state (existing-profiler-state))
295 (set-inside-profiler?! state #t)
297 ;; FIXME: with-statprof should be able to set an outer frame for the
299 (if (positive? (profile-level state))
300 (let* ((stop-time (get-internal-run-time))
301 ;; cut down to the signal handler. note that this will only
302 ;; work if statprof.scm is compiled; otherwise we get
303 ;; `eval' on the stack instead, because if it's not
304 ;; compiled, profile-signal-handler is a thunk that
305 ;; tail-calls eval. perhaps we should always compile the
306 ;; signal handler instead...
307 (stack (or (make-stack #t profile-signal-handler)
308 (pk 'what! (make-stack #t))))
309 (inside-apply-trap? (sample-stack-procs stack)))
311 (if (not inside-apply-trap?)
313 ;; disabling here is just a little more efficient, but
314 ;; not necessary given inside-profiler?. We can't just
315 ;; disable unconditionally at the top of this function
316 ;; and eliminate inside-profiler? because it seems to
317 ;; confuse guile wrt re-enabling the trap when
318 ;; count-call finishes.
319 (if (count-calls? state)
320 (set-vm-trace-level! (1- (vm-trace-level))))
321 (accumulate-time state stop-time)))
323 (setitimer ITIMER_PROF
325 (car (sampling-frequency state))
326 (cdr (sampling-frequency state)))
328 (if (not inside-apply-trap?)
330 (set-last-start-time! state (get-internal-run-time))
331 (if (count-calls? state)
332 (set-vm-trace-level! (1+ (vm-trace-level))))))))
334 (set-inside-profiler?! state #f))
336 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
337 ;; Count total calls.
339 (define (count-call frame)
340 (define state (existing-profiler-state))
342 (if (not (inside-profiler? state))
344 (accumulate-time state (get-internal-run-time))
346 (and=> (frame-procedure frame)
348 (inc-call-data-call-count!
349 (get-call-data proc))))
351 (set-last-start-time! state (get-internal-run-time)))))
353 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
355 (define (statprof-active?)
356 "Returns @code{#t} if @code{statprof-start} has been called more times
357 than @code{statprof-stop}, @code{#f} otherwise."
358 (define state (profiler-state))
359 (and state (positive? (profile-level state))))
361 ;; Do not call this from statprof internal functions -- user only.
362 (define (statprof-start)
363 "Start the profiler.@code{}"
364 ;; After some head-scratching, I don't *think* I need to mask/unmask
365 ;; signals here, but if I'm wrong, please let me know.
366 (define state (ensure-profiler-state))
367 (set-profile-level! state (+ (profile-level state) 1))
368 (if (= (profile-level state) 1)
369 (let* ((rpt (remaining-prof-time state))
371 (or (positive? (car rpt))
372 (positive? (cdr rpt))))))
373 (set-remaining-prof-time! state #f)
374 (set-last-start-time! state (get-internal-run-time))
375 (set-gc-time-taken! state
376 (cdr (assq 'gc-time-taken (gc-stats))))
378 (setitimer ITIMER_PROF 0 0 (car rpt) (cdr rpt))
379 (setitimer ITIMER_PROF
381 (car (sampling-frequency state))
382 (cdr (sampling-frequency state))))
383 (if (count-calls? state)
384 (add-hook! (vm-apply-hook) count-call))
385 (set-vm-trace-level! (1+ (vm-trace-level)))
388 ;; Do not call this from statprof internal functions -- user only.
389 (define (statprof-stop)
390 "Stop the profiler.@code{}"
391 ;; After some head-scratching, I don't *think* I need to mask/unmask
392 ;; signals here, but if I'm wrong, please let me know.
393 (define state (ensure-profiler-state))
394 (set-profile-level! state (- (profile-level state) 1))
395 (if (zero? (profile-level state))
397 (set-gc-time-taken! state
398 (- (cdr (assq 'gc-time-taken (gc-stats)))
399 (gc-time-taken state)))
400 (set-vm-trace-level! (1- (vm-trace-level)))
401 (if (count-calls? state)
402 (remove-hook! (vm-apply-hook) count-call))
403 ;; I believe that we need to do this before getting the time
404 ;; (unless we want to make things even more complicated).
405 (set-remaining-prof-time! state (setitimer ITIMER_PROF 0 0 0 0))
406 (accumulate-time state (get-internal-run-time))
407 (set-last-start-time! state #f))))
409 (define* (statprof-reset sample-seconds sample-microseconds count-calls?
410 #:optional full-stacks?)
411 "Reset the statprof sampler interval to @var{sample-seconds} and
412 @var{sample-microseconds}. If @var{count-calls?} is true, arrange to
413 instrument procedure calls as well as collecting statistical profiling
414 data. If @var{full-stacks?} is true, collect all sampled stacks into a
415 list for later analysis.
417 Enables traps and debugging as necessary."
418 (when (statprof-active?)
419 (error "Can't reset profiler while profiler is running."))
420 (let ((state (fresh-profiler-state #:count-calls? count-calls?
422 (cons sample-seconds sample-microseconds)
423 #:full-stacks? full-stacks?)))
424 (profiler-state state)
425 (sigaction SIGPROF profile-signal-handler)
428 (define (statprof-fold-call-data proc init)
429 "Fold @var{proc} over the call-data accumulated by statprof. Cannot be
430 called while statprof is active. @var{proc} should take two arguments,
431 @code{(@var{call-data} @var{prior-result})}.
433 Note that a given proc-name may appear multiple times, but if it does,
434 it represents different functions with the same name."
435 (when (statprof-active?)
436 (error "Can't call statprof-fold-call-data while profiler is running."))
438 (lambda (key value prior-result)
439 (proc value prior-result))
441 (procedure-data (existing-profiler-state))))
443 (define (statprof-proc-call-data proc)
444 "Returns the call-data associated with @var{proc}, or @code{#f} if
446 (when (statprof-active?)
447 (error "Can't call statprof-proc-call-data while profiler is running."))
448 (get-call-data proc))
450 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
453 (define (statprof-call-data->stats call-data)
454 "Returns an object of type @code{statprof-stats}."
455 ;; returns (vector proc-name
457 ;; cum-seconds-in-proc
458 ;; self-seconds-in-proc
460 ;; self-secs-per-call
461 ;; total-secs-per-call)
463 (define state (existing-profiler-state))
465 (let* ((proc-name (call-data-printable call-data))
466 (self-samples (call-data-self-sample-count call-data))
467 (cum-samples (call-data-cum-sample-count call-data))
468 (all-samples (statprof-sample-count))
469 (secs-per-sample (/ (statprof-accumulated-time)
470 (statprof-sample-count)))
471 (num-calls (and (count-calls? state) (statprof-call-data-calls call-data))))
474 (* (/ self-samples all-samples) 100.0)
475 (* cum-samples secs-per-sample 1.0)
476 (* self-samples secs-per-sample 1.0)
478 (and num-calls ;; maybe we only sampled in children
479 (if (zero? self-samples) 0.0
480 (/ (* self-samples secs-per-sample) 1.0 num-calls)))
481 (and num-calls ;; cum-samples must be positive
482 (/ (* cum-samples secs-per-sample)
484 ;; num-calls might be 0 if we entered statprof during the
485 ;; dynamic extent of the call
486 (max num-calls 1))))))
488 (define (statprof-stats-proc-name stats) (vector-ref stats 0))
489 (define (statprof-stats-%-time-in-proc stats) (vector-ref stats 1))
490 (define (statprof-stats-cum-secs-in-proc stats) (vector-ref stats 2))
491 (define (statprof-stats-self-secs-in-proc stats) (vector-ref stats 3))
492 (define (statprof-stats-calls stats) (vector-ref stats 4))
493 (define (statprof-stats-self-secs-per-call stats) (vector-ref stats 5))
494 (define (statprof-stats-cum-secs-per-call stats) (vector-ref stats 6))
496 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
498 (define (stats-sorter x y)
499 (let ((diff (- (statprof-stats-self-secs-in-proc x)
500 (statprof-stats-self-secs-in-proc y))))
503 (- (statprof-stats-cum-secs-in-proc x)
504 (statprof-stats-cum-secs-in-proc y))
507 (define* (statprof-display #:optional (port (current-output-port)))
508 "Displays a gprof-like summary of the statistics collected. Unless an
509 optional @var{port} argument is passed, uses the current output port."
510 (define state (existing-profiler-state))
513 ((zero? (statprof-sample-count))
514 (format port "No samples recorded.\n"))
516 (let* ((stats-list (statprof-fold-call-data
517 (lambda (data prior-value)
518 (cons (statprof-call-data->stats data)
521 (sorted-stats (sort stats-list stats-sorter)))
523 (define (display-stats-line stats)
524 (if (count-calls? state)
525 (format port "~6,2f ~9,2f ~9,2f ~7d ~8,2f ~8,2f "
526 (statprof-stats-%-time-in-proc stats)
527 (statprof-stats-cum-secs-in-proc stats)
528 (statprof-stats-self-secs-in-proc stats)
529 (statprof-stats-calls stats)
530 (* 1000 (statprof-stats-self-secs-per-call stats))
531 (* 1000 (statprof-stats-cum-secs-per-call stats)))
532 (format port "~6,2f ~9,2f ~9,2f "
533 (statprof-stats-%-time-in-proc stats)
534 (statprof-stats-cum-secs-in-proc stats)
535 (statprof-stats-self-secs-in-proc stats)))
536 (display (statprof-stats-proc-name stats) port)
539 (if (count-calls? state)
541 (format port "~5a ~10a ~7a ~8a ~8a ~8a ~8@a\n"
542 "% " "cumulative" "self" "" "self" "total" "")
543 (format port "~5a ~9a ~8a ~8a ~8a ~8a ~8@a\n"
544 "time" "seconds" "seconds" "calls" "ms/call" "ms/call" "name"))
546 (format port "~5a ~10a ~7a ~8@a\n"
547 "%" "cumulative" "self" "")
548 (format port "~5a ~10a ~7a ~8@a\n"
549 "time" "seconds" "seconds" "name")))
551 (for-each display-stats-line sorted-stats)
553 (display "---\n" port)
554 (simple-format #t "Sample count: ~A\n" (statprof-sample-count))
555 (simple-format #t "Total time: ~A seconds (~A seconds in GC)\n"
556 (statprof-accumulated-time)
557 (/ (gc-time-taken state) 1.0 internal-time-units-per-second))))))
559 (define (statprof-display-anomolies)
560 "A sanity check that attempts to detect anomolies in statprof's
562 (define state (existing-profiler-state))
564 (statprof-fold-call-data
565 (lambda (data prior-value)
566 (if (and (count-calls? state)
567 (zero? (call-data-call-count data))
568 (positive? (call-data-cum-sample-count data)))
571 (call-data-name data)
572 (call-data-call-count data)
573 (call-data-cum-sample-count data))))
575 (simple-format #t "Total time: ~A\n" (statprof-accumulated-time))
576 (simple-format #t "Sample count: ~A\n" (statprof-sample-count)))
578 (define (statprof-accumulated-time)
579 "Returns the time accumulated during the last statprof run.@code{}"
580 (when (statprof-active?)
581 (error "Can't get accumulated time while profiler is running."))
582 (/ (accumulated-time (existing-profiler-state)) internal-time-units-per-second))
584 (define (statprof-sample-count)
585 "Returns the number of samples taken during the last statprof run.@code{}"
586 (when (statprof-active?)
587 (error "Can't get sample count while profiler is running."))
588 (sample-count (existing-profiler-state)))
590 (define statprof-call-data-name call-data-name)
591 (define statprof-call-data-calls call-data-call-count)
592 (define statprof-call-data-cum-samples call-data-cum-sample-count)
593 (define statprof-call-data-self-samples call-data-self-sample-count)
595 (define (statprof-fetch-stacks)
596 "Returns a list of stacks, as they were captured since the last call
597 to @code{statprof-reset}.
599 Note that stacks are only collected if the @var{full-stacks?} argument
600 to @code{statprof-reset} is true."
601 (define state (existing-profiler-state))
608 ((and (program? a) (program? b))
609 (eq? (program-code a) (program-code b)))
613 ;; tree ::= (car n . tree*)
615 (define (lists->trees lists equal?)
616 (let lp ((in lists) (n-terminal 0) (tails '()))
619 (let ((trees (map (lambda (tail)
621 (lists->trees (cdr tail) equal?)))
623 (cons (apply + n-terminal (map cadr trees))
625 (lambda (a b) (> (cadr a) (cadr b)))))))
627 (lp (cdr in) (1+ n-terminal) tails))
628 ((find (lambda (x) (equal? (car x) (caar in)))
635 (cons (cdar in) (cdr tail))))))
639 (acons (caar in) (list (cdar in)) tails))))))
641 (define (stack->procedures stack)
643 (unfold-right (lambda (x) (not x))
646 (stack-ref stack 0))))
648 (define (statprof-fetch-call-tree)
649 "Return a call tree for the previous statprof run.
651 The return value is a list of nodes, each of which is of the type:
653 node ::= (@var{proc} @var{count} . @var{nodes})
655 (define state (existing-profiler-state))
656 (cons #t (lists->trees (map stack->procedures (stacks state)) procedure=?)))
658 (define* (statprof thunk #:key (loop 1) (hz 100) (count-calls? #f)
660 "Profiles the execution of @var{thunk}.
662 The stack will be sampled @var{hz} times per second, and the thunk itself will
663 be called @var{loop} times.
665 If @var{count-calls?} is true, all procedure calls will be recorded. This
666 operation is somewhat expensive.
668 If @var{full-stacks?} is true, at each sample, statprof will store away the
669 whole call tree, for later analysis. Use @code{statprof-fetch-stacks} or
670 @code{statprof-fetch-call-tree} to retrieve the last-stored stacks."
672 (define state (ensure-profiler-state))
676 (statprof-reset (inexact->exact (floor (/ 1 hz)))
677 (inexact->exact (* 1e6 (- (/ 1 hz)
691 (set-procedure-data! state #f))))
693 (define-macro (with-statprof . args)
694 "Profiles the expressions in its body.
700 Execute the body @var{loop} number of times, or @code{#f} for no looping
708 Whether to instrument each function call (expensive)
712 Whether to collect away all sampled stacks into a list
716 (define (kw-arg-ref kw args def)
718 ((null? args) (error "Invalid macro body"))
719 ((keyword? (car args))
720 (if (eq? (car args) kw)
722 (kw-arg-ref kw (cddr args) def)))
723 ((eq? kw #f def) ;; asking for the body
725 (else def))) ;; kw not found
726 `((@ (statprof) statprof)
727 (lambda () ,@(kw-arg-ref #f args #f))
728 #:loop ,(kw-arg-ref #:loop args 1)
729 #:hz ,(kw-arg-ref #:hz args 100)
730 #:count-calls? ,(kw-arg-ref #:count-calls? args #f)
731 #:full-stacks? ,(kw-arg-ref #:full-stacks? args #f)))
733 (define* (gcprof thunk #:key (loop 1) (full-stacks? #f))
734 "Do an allocation profile of the execution of @var{thunk}.
736 The stack will be sampled soon after every garbage collection, yielding
737 an approximate idea of what is causing allocation in your program.
739 Since GC does not occur very frequently, you may need to use the
740 @var{loop} parameter, to cause @var{thunk} to be called @var{loop}
743 If @var{full-stacks?} is true, at each sample, statprof will store away the
744 whole call tree, for later analysis. Use @code{statprof-fetch-stacks} or
745 @code{statprof-fetch-call-tree} to retrieve the last-stored stacks."
747 (define state (ensure-profiler-state))
750 (if (positive? (profile-level state))
751 (error "Can't reset profiler while profiler is running."))
752 (set-accumulated-time! state 0)
753 (set-last-start-time! state #f)
754 (set-sample-count! state 0)
755 (set-count-calls?! state #f)
756 (set-procedure-data! state (make-hash-table 131))
757 (set-record-full-stacks?! state full-stacks?)
758 (set-stacks! state '()))
760 (define (gc-callback)
762 ((inside-profiler? state))
764 (set-inside-profiler?! state #t)
766 ;; FIXME: should be able to set an outer frame for the stack cut
767 (let ((stop-time (get-internal-run-time))
768 ;; Cut down to gc-callback, and then one before (the
769 ;; after-gc async). See the note in profile-signal-handler
771 (stack (or (make-stack #t gc-callback 0 1)
772 (pk 'what! (make-stack #t)))))
773 (sample-stack-procs stack)
774 (accumulate-time state stop-time)
775 (set-last-start-time! state (get-internal-run-time)))
777 (set-inside-profiler?! state #f))))
780 (set-profile-level! state (+ (profile-level state) 1))
781 (if (= (profile-level state) 1)
783 (set-remaining-prof-time! state #f)
784 (set-last-start-time! state (get-internal-run-time))
785 (set-gc-time-taken! state (cdr (assq 'gc-time-taken (gc-stats))))
786 (add-hook! after-gc-hook gc-callback)
787 (set-vm-trace-level! (1+ (vm-trace-level)))
791 (set-profile-level! state (- (profile-level state) 1))
792 (if (zero? (profile-level state))
794 (set-gc-time-taken! state
795 (- (cdr (assq 'gc-time-taken (gc-stats)))
796 (gc-time-taken state)))
797 (remove-hook! after-gc-hook gc-callback)
798 (accumulate-time state (get-internal-run-time))
799 (set-last-start-time! state #f))))
814 (set-procedure-data! state #f))))