Commit | Line | Data |
---|---|---|
47f3ce52 AW |
1 | ;;;; (statprof) -- a statistical profiler for Guile |
2 | ;;;; -*-scheme-*- | |
3 | ;;;; | |
998f8494 | 4 | ;;;; Copyright (C) 2009, 2010, 2011, 2013, 2014 Free Software Foundation, Inc. |
47f3ce52 AW |
5 | ;;;; Copyright (C) 2004, 2009 Andy Wingo <wingo at pobox dot com> |
6 | ;;;; Copyright (C) 2001 Rob Browning <rlb at defaultvalue dot org> | |
7 | ;;;; | |
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. | |
12 | ;;;; | |
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. | |
17 | ;;;; | |
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 | |
21 | ;;;; | |
22 | \f | |
23 | ||
24 | ;;; Commentary: | |
998f8494 | 25 | ;;; |
62fd93e2 | 26 | ;;; @code{(statprof)} is a statistical profiler for Guile. |
998f8494 AW |
27 | ;;; |
28 | ;;; A simple use of statprof would look like this: | |
29 | ;;; | |
30 | ;;; @example | |
188e2ae3 AW |
31 | ;;; (statprof (lambda () (do-something)) |
32 | ;;; #:hz 100 | |
33 | ;;; #:count-calls? #t) | |
998f8494 AW |
34 | ;;; @end example |
35 | ;;; | |
188e2ae3 AW |
36 | ;;; This would run the thunk with statistical profiling, finally |
37 | ;;; displaying a gprof flat-style table of statistics which could | |
38 | ;;; something like this: | |
998f8494 AW |
39 | ;;; |
40 | ;;; @example | |
41 | ;;; % cumulative self self total | |
42 | ;;; time seconds seconds calls ms/call ms/call name | |
43 | ;;; 35.29 0.23 0.23 2002 0.11 0.11 - | |
44 | ;;; 23.53 0.15 0.15 2001 0.08 0.08 positive? | |
45 | ;;; 23.53 0.15 0.15 2000 0.08 0.08 + | |
46 | ;;; 11.76 0.23 0.08 2000 0.04 0.11 do-nothing | |
47 | ;;; 5.88 0.64 0.04 2001 0.02 0.32 loop | |
48 | ;;; 0.00 0.15 0.00 1 0.00 150.59 do-something | |
49 | ;;; ... | |
50 | ;;; @end example | |
51 | ;;; | |
52 | ;;; All of the numerical data with the exception of the calls column is | |
53 | ;;; statistically approximate. In the following column descriptions, and | |
54 | ;;; in all of statprof, "time" refers to execution time (both user and | |
55 | ;;; system), not wall clock time. | |
56 | ;;; | |
57 | ;;; @table @asis | |
58 | ;;; @item % time | |
59 | ;;; The percent of the time spent inside the procedure itself | |
60 | ;;; (not counting children). | |
61 | ;;; @item cumulative seconds | |
62 | ;;; The total number of seconds spent in the procedure, including | |
63 | ;;; children. | |
64 | ;;; @item self seconds | |
65 | ;;; The total number of seconds spent in the procedure itself (not counting | |
66 | ;;; children). | |
67 | ;;; @item calls | |
68 | ;;; The total number of times the procedure was called. | |
69 | ;;; @item self ms/call | |
70 | ;;; The average time taken by the procedure itself on each call, in ms. | |
71 | ;;; @item total ms/call | |
72 | ;;; The average time taken by each call to the procedure, including time | |
73 | ;;; spent in child functions. | |
74 | ;;; @item name | |
75 | ;;; The name of the procedure. | |
76 | ;;; @end table | |
77 | ;;; | |
78 | ;;; The profiler uses @code{eq?} and the procedure object itself to | |
79 | ;;; identify the procedures, so it won't confuse different procedures with | |
80 | ;;; the same name. They will show up as two different rows in the output. | |
81 | ;;; | |
82 | ;;; Right now the profiler is quite simplistic. I cannot provide | |
83 | ;;; call-graphs or other higher level information. What you see in the | |
84 | ;;; table is pretty much all there is. Patches are welcome :-) | |
85 | ;;; | |
86 | ;;; @section Implementation notes | |
87 | ;;; | |
88 | ;;; The profiler works by setting the unix profiling signal | |
89 | ;;; @code{ITIMER_PROF} to go off after the interval you define in the call | |
90 | ;;; to @code{statprof-reset}. When the signal fires, a sampling routine is | |
91 | ;;; run which looks at the current procedure that's executing, and then | |
92 | ;;; crawls up the stack, and for each procedure encountered, increments | |
93 | ;;; that procedure's sample count. Note that if a procedure is encountered | |
94 | ;;; multiple times on a given stack, it is only counted once. After the | |
95 | ;;; sampling is complete, the profiler resets profiling timer to fire | |
96 | ;;; again after the appropriate interval. | |
97 | ;;; | |
98 | ;;; Meanwhile, the profiler keeps track, via @code{get-internal-run-time}, | |
99 | ;;; how much CPU time (system and user -- which is also what | |
100 | ;;; @code{ITIMER_PROF} tracks), has elapsed while code has been executing | |
101 | ;;; within a statprof-start/stop block. | |
102 | ;;; | |
103 | ;;; The profiler also tries to avoid counting or timing its own code as | |
104 | ;;; much as possible. | |
105 | ;;; | |
47f3ce52 AW |
106 | ;;; Code: |
107 | ||
47f3ce52 AW |
108 | (define-module (statprof) |
109 | #:use-module (srfi srfi-1) | |
62fd93e2 | 110 | #:use-module (srfi srfi-9) |
e4a8775d | 111 | #:use-module (srfi srfi-9 gnu) |
47f3ce52 | 112 | #:autoload (ice-9 format) (format) |
e1138ba1 AW |
113 | #:use-module (system vm vm) |
114 | #:use-module (system vm frame) | |
3f9f4a2d | 115 | #:use-module (system vm debug) |
e1138ba1 | 116 | #:use-module (system vm program) |
47f3ce52 AW |
117 | #:export (statprof-active? |
118 | statprof-start | |
119 | statprof-stop | |
120 | statprof-reset | |
121 | ||
122 | statprof-accumulated-time | |
123 | statprof-sample-count | |
124 | statprof-fold-call-data | |
125 | statprof-proc-call-data | |
126 | statprof-call-data-name | |
127 | statprof-call-data-calls | |
128 | statprof-call-data-cum-samples | |
129 | statprof-call-data-self-samples | |
130 | statprof-call-data->stats | |
131 | ||
132 | statprof-stats-proc-name | |
133 | statprof-stats-%-time-in-proc | |
134 | statprof-stats-cum-secs-in-proc | |
135 | statprof-stats-self-secs-in-proc | |
136 | statprof-stats-calls | |
137 | statprof-stats-self-secs-per-call | |
138 | statprof-stats-cum-secs-per-call | |
139 | ||
140 | statprof-display | |
91db6c4f AW |
141 | statprof-display-anomalies |
142 | statprof-display-anomolies ; Deprecated spelling. | |
47f3ce52 AW |
143 | |
144 | statprof-fetch-stacks | |
145 | statprof-fetch-call-tree | |
146 | ||
e1138ba1 | 147 | statprof |
2d239a78 AW |
148 | with-statprof |
149 | ||
150 | gcprof)) | |
47f3ce52 AW |
151 | |
152 | ||
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. | |
156 | ;; | |
157 | ;; Globally the profiler tracks the total time elapsed and the number | |
158 | ;; of times the sampler was fired. | |
159 | ;; | |
160 | ;; Right now, this profiler is not per-thread and is not thread safe. | |
161 | ||
62fd93e2 AW |
162 | (define-record-type <state> |
163 | (make-state accumulated-time last-start-time sample-count | |
19bf8caf | 164 | sampling-period remaining-prof-time profile-level |
cd073eb4 | 165 | call-counts gc-time-taken inside-profiler? |
3f9f4a2d | 166 | prev-sigprof-handler buffer buffer-pos) |
62fd93e2 AW |
167 | state? |
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!) | |
19bf8caf AW |
174 | ;; Microseconds. |
175 | (sampling-period sampling-period set-sampling-period!) | |
62fd93e2 AW |
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!) | |
3f9f4a2d AW |
180 | ;; Hash table mapping ip -> call count, or #f if not counting calls. |
181 | (call-counts call-counts set-call-counts!) | |
62fd93e2 AW |
182 | ;; GC time between statprof-start and statprof-stop. |
183 | (gc-time-taken gc-time-taken set-gc-time-taken!) | |
56bfce7c | 184 | ;; True if we are inside the profiler. |
3072d762 AW |
185 | (inside-profiler? inside-profiler? set-inside-profiler?!) |
186 | ;; True if we are inside the profiler. | |
3f9f4a2d AW |
187 | (prev-sigprof-handler prev-sigprof-handler set-prev-sigprof-handler!) |
188 | ;; Stack samples. | |
189 | (buffer buffer set-buffer!) | |
190 | (buffer-pos buffer-pos set-buffer-pos!)) | |
62fd93e2 AW |
191 | |
192 | (define profiler-state (make-parameter #f)) | |
193 | ||
3f9f4a2d AW |
194 | (define (fresh-buffer) |
195 | (make-vector 1024 #f)) | |
196 | ||
197 | (define (expand-buffer buf) | |
198 | (let* ((size (vector-length buf)) | |
199 | (new (make-vector (* size 2) #f))) | |
200 | (vector-move-left! buf 0 (vector-length buf) new 0) | |
201 | new)) | |
202 | ||
4eb1fb9b | 203 | (define* (fresh-profiler-state #:key (count-calls? #f) |
cd073eb4 AW |
204 | (sampling-period 10000)) |
205 | (make-state 0 #f 0 | |
206 | sampling-period 0 0 | |
207 | (and count-calls? (make-hash-table)) 0 #f | |
208 | #f (fresh-buffer) 0)) | |
4eb1fb9b | 209 | |
62fd93e2 AW |
210 | (define (ensure-profiler-state) |
211 | (or (profiler-state) | |
4eb1fb9b | 212 | (let ((state (fresh-profiler-state))) |
62fd93e2 AW |
213 | (profiler-state state) |
214 | state))) | |
47f3ce52 | 215 | |
45a7de82 AW |
216 | (define (existing-profiler-state) |
217 | (or (profiler-state) | |
218 | (error "expected there to be a profiler state"))) | |
219 | ||
62fd93e2 AW |
220 | (define (accumulate-time state stop-time) |
221 | (set-accumulated-time! state | |
222 | (+ (accumulated-time state) | |
62fd93e2 | 223 | (- stop-time (last-start-time state))))) |
47f3ce52 | 224 | |
47f3ce52 AW |
225 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
226 | ;; SIGPROF handler | |
227 | ||
e4a8775d | 228 | (define (sample-stack-procs state stack) |
3f9f4a2d AW |
229 | (set-sample-count! state (+ (sample-count state) 1)) |
230 | ||
231 | (let lp ((frame (stack-ref stack 0)) | |
232 | (buffer (buffer state)) | |
233 | (pos (buffer-pos state))) | |
234 | (define (write-sample sample) | |
235 | (vector-set! buffer pos sample)) | |
236 | (define (continue pos) | |
237 | (lp (frame-previous frame) buffer pos)) | |
238 | (define (write-sample-and-continue sample) | |
239 | (write-sample sample) | |
240 | (continue (1+ pos))) | |
241 | (cond | |
242 | ((= pos (vector-length buffer)) | |
243 | (lp frame (expand-buffer buffer) pos)) | |
244 | ((not frame) | |
245 | (write-sample #f) | |
246 | (set-buffer! state buffer) | |
247 | (set-buffer-pos! state (1+ pos))) | |
248 | (else | |
249 | (let ((proc (frame-procedure frame))) | |
250 | (cond | |
251 | ((primitive? proc) | |
252 | (write-sample-and-continue (procedure-name proc))) | |
253 | ((program? proc) | |
254 | (write-sample-and-continue (frame-instruction-pointer frame))) | |
255 | (proc (write-sample-and-continue proc)) | |
256 | ;; If proc is false, that would confuse our stack walker. | |
257 | ;; Ignore it. | |
258 | (else (continue pos)))))))) | |
47f3ce52 | 259 | |
19bf8caf | 260 | (define (reset-sigprof-timer usecs) |
e68ed839 AW |
261 | ;; Guile's setitimer binding is terrible. |
262 | (let ((prev (setitimer ITIMER_PROF 0 0 0 usecs))) | |
263 | (+ (* (caadr prev) #e1e6) (cdadr prev)))) | |
19bf8caf | 264 | |
47f3ce52 | 265 | (define (profile-signal-handler sig) |
45a7de82 | 266 | (define state (existing-profiler-state)) |
62fd93e2 | 267 | |
56bfce7c | 268 | (set-inside-profiler?! state #t) |
47f3ce52 AW |
269 | |
270 | ;; FIXME: with-statprof should be able to set an outer frame for the | |
271 | ;; stack cut | |
cad444e3 AW |
272 | (when (positive? (profile-level state)) |
273 | (let* ((stop-time (get-internal-run-time)) | |
274 | ;; cut down to the signal handler. note that this will only | |
275 | ;; work if statprof.scm is compiled; otherwise we get | |
276 | ;; `eval' on the stack instead, because if it's not | |
277 | ;; compiled, profile-signal-handler is a thunk that | |
278 | ;; tail-calls eval. perhaps we should always compile the | |
279 | ;; signal handler instead... | |
280 | (stack (or (make-stack #t profile-signal-handler) | |
546efe25 AW |
281 | (pk 'what! (make-stack #t))))) |
282 | ||
283 | (sample-stack-procs state stack) | |
284 | (accumulate-time state stop-time) | |
285 | (set-last-start-time! state (get-internal-run-time)) | |
286 | ||
19bf8caf | 287 | (reset-sigprof-timer (sampling-period state)))) |
e1138ba1 | 288 | |
56bfce7c | 289 | (set-inside-profiler?! state #f)) |
47f3ce52 AW |
290 | |
291 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; | |
292 | ;; Count total calls. | |
293 | ||
e1138ba1 | 294 | (define (count-call frame) |
3f9f4a2d AW |
295 | (let ((state (existing-profiler-state))) |
296 | (unless (inside-profiler? state) | |
297 | (accumulate-time state (get-internal-run-time)) | |
62fd93e2 | 298 | |
3f9f4a2d AW |
299 | (let* ((key (let ((proc (frame-procedure frame))) |
300 | (cond | |
301 | ((primitive? proc) (procedure-name proc)) | |
302 | ((program? proc) (program-code proc)) | |
303 | (else proc)))) | |
304 | (handle (hashv-create-handle! (call-counts state) key 0))) | |
305 | (set-cdr! handle (1+ (cdr handle)))) | |
47f3ce52 | 306 | |
3f9f4a2d | 307 | (set-last-start-time! state (get-internal-run-time))))) |
47f3ce52 AW |
308 | |
309 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; | |
310 | ||
311 | (define (statprof-active?) | |
312 | "Returns @code{#t} if @code{statprof-start} has been called more times | |
313 | than @code{statprof-stop}, @code{#f} otherwise." | |
45a7de82 AW |
314 | (define state (profiler-state)) |
315 | (and state (positive? (profile-level state)))) | |
47f3ce52 AW |
316 | |
317 | ;; Do not call this from statprof internal functions -- user only. | |
13a977dd | 318 | (define* (statprof-start #:optional (state (ensure-profiler-state))) |
47f3ce52 AW |
319 | "Start the profiler.@code{}" |
320 | ;; After some head-scratching, I don't *think* I need to mask/unmask | |
321 | ;; signals here, but if I'm wrong, please let me know. | |
62fd93e2 | 322 | (set-profile-level! state (+ (profile-level state) 1)) |
cad444e3 | 323 | (when (= (profile-level state) 1) |
19bf8caf AW |
324 | (let ((rpt (remaining-prof-time state))) |
325 | (set-remaining-prof-time! state 0) | |
cad444e3 AW |
326 | ;; FIXME: Use per-thread run time. |
327 | (set-last-start-time! state (get-internal-run-time)) | |
3476a369 | 328 | (set-gc-time-taken! state (assq-ref (gc-stats) 'gc-time-taken)) |
3072d762 AW |
329 | (let ((prev (sigaction SIGPROF profile-signal-handler))) |
330 | (set-prev-sigprof-handler! state (car prev))) | |
19bf8caf | 331 | (reset-sigprof-timer (if (zero? rpt) (sampling-period state) rpt)) |
3f9f4a2d | 332 | (when (call-counts state) |
cad444e3 AW |
333 | (add-hook! (vm-apply-hook) count-call)) |
334 | (set-vm-trace-level! (1+ (vm-trace-level))) | |
335 | #t))) | |
47f3ce52 AW |
336 | |
337 | ;; Do not call this from statprof internal functions -- user only. | |
13a977dd | 338 | (define* (statprof-stop #:optional (state (ensure-profiler-state))) |
47f3ce52 AW |
339 | "Stop the profiler.@code{}" |
340 | ;; After some head-scratching, I don't *think* I need to mask/unmask | |
341 | ;; signals here, but if I'm wrong, please let me know. | |
62fd93e2 | 342 | (set-profile-level! state (- (profile-level state) 1)) |
cad444e3 AW |
343 | (when (zero? (profile-level state)) |
344 | (set-gc-time-taken! state | |
3476a369 | 345 | (- (assq-ref (gc-stats) 'gc-time-taken) |
cad444e3 AW |
346 | (gc-time-taken state))) |
347 | (set-vm-trace-level! (1- (vm-trace-level))) | |
3f9f4a2d | 348 | (when (call-counts state) |
cad444e3 AW |
349 | (remove-hook! (vm-apply-hook) count-call)) |
350 | ;; I believe that we need to do this before getting the time | |
351 | ;; (unless we want to make things even more complicated). | |
19bf8caf | 352 | (set-remaining-prof-time! state (reset-sigprof-timer 0)) |
cad444e3 | 353 | (accumulate-time state (get-internal-run-time)) |
3072d762 AW |
354 | (sigaction SIGPROF (prev-sigprof-handler state)) |
355 | (set-prev-sigprof-handler! state #f) | |
cad444e3 | 356 | (set-last-start-time! state #f))) |
47f3ce52 | 357 | |
e640b440 AW |
358 | (define* (statprof-reset sample-seconds sample-microseconds count-calls? |
359 | #:optional full-stacks?) | |
47f3ce52 AW |
360 | "Reset the statprof sampler interval to @var{sample-seconds} and |
361 | @var{sample-microseconds}. If @var{count-calls?} is true, arrange to | |
362 | instrument procedure calls as well as collecting statistical profiling | |
cd073eb4 AW |
363 | data. (The optional @var{full-stacks?} argument is deprecated; statprof |
364 | always collects full stacks.)" | |
4d0c358b | 365 | (when (statprof-active?) |
4eb1fb9b | 366 | (error "Can't reset profiler while profiler is running.")) |
3072d762 AW |
367 | (profiler-state |
368 | (fresh-profiler-state #:count-calls? count-calls? | |
369 | #:sampling-period (+ (* sample-seconds #e1e6) | |
cd073eb4 | 370 | sample-microseconds))) |
13a977dd | 371 | (values)) |
47f3ce52 | 372 | |
3f9f4a2d | 373 | (define-record-type call-data |
e3997e70 AW |
374 | (make-call-data name source printable |
375 | call-count cum-sample-count self-sample-count) | |
3f9f4a2d AW |
376 | call-data? |
377 | (name call-data-name) | |
e3997e70 | 378 | (source call-data-source) |
3f9f4a2d AW |
379 | (printable call-data-printable) |
380 | (call-count call-data-call-count set-call-data-call-count!) | |
381 | (cum-sample-count call-data-cum-sample-count set-call-data-cum-sample-count!) | |
382 | (self-sample-count call-data-self-sample-count set-call-data-self-sample-count!)) | |
383 | ||
384 | (define (source->string source) | |
385 | (format #f "~a:~a:~a" | |
386 | (or (source-file source) "<current input>") | |
387 | (source-line-for-user source) | |
388 | (source-column source))) | |
389 | ||
390 | (define (program-debug-info-printable pdi) | |
391 | (let* ((addr (program-debug-info-addr pdi)) | |
392 | (name (or (and=> (program-debug-info-name pdi) symbol->string) | |
393 | (string-append "#x" (number->string addr 16)))) | |
394 | (loc (and=> (find-source-for-addr addr) source->string))) | |
395 | (if loc | |
396 | (string-append name " at " loc) | |
397 | name))) | |
398 | ||
399 | (define (addr->pdi addr cache) | |
400 | (cond | |
401 | ((hashv-get-handle cache addr) => cdr) | |
402 | (else | |
403 | (let ((data (find-program-debug-info addr))) | |
404 | (hashv-set! cache addr data) | |
405 | data)))) | |
406 | ||
407 | (define (addr->printable addr pdi) | |
408 | (if pdi | |
409 | (program-debug-info-printable pdi) | |
410 | (string-append "#x" (number->string addr 16)))) | |
411 | ||
412 | (define (inc-call-data-cum-sample-count! cd) | |
413 | (set-call-data-cum-sample-count! cd (1+ (call-data-cum-sample-count cd)))) | |
414 | (define (inc-call-data-self-sample-count! cd) | |
415 | (set-call-data-self-sample-count! cd (1+ (call-data-self-sample-count cd)))) | |
416 | ||
417 | (define (stack-samples->procedure-data state) | |
418 | (let ((table (make-hash-table)) | |
419 | (addr-cache (make-hash-table)) | |
420 | (call-counts (call-counts state)) | |
421 | (buffer (buffer state)) | |
422 | (len (buffer-pos state))) | |
423 | (define (addr->call-data addr) | |
424 | (let* ((pdi (addr->pdi addr addr-cache)) | |
425 | (entry (if pdi (program-debug-info-addr pdi) addr))) | |
426 | (or (hashv-ref table entry) | |
427 | (let ((data (make-call-data (and=> pdi program-debug-info-name) | |
e3997e70 | 428 | (find-source-for-addr entry) |
3f9f4a2d AW |
429 | (addr->printable entry pdi) |
430 | (and call-counts | |
431 | (hashv-ref call-counts entry)) | |
432 | 0 | |
433 | 0))) | |
434 | (hashv-set! table entry data) | |
435 | data)))) | |
436 | ||
437 | (define (callee->call-data callee) | |
438 | (cond | |
439 | ((number? callee) (addr->call-data callee)) | |
440 | ((hashv-ref table callee)) | |
441 | (else | |
442 | (let ((data (make-call-data | |
443 | (cond ((procedure? callee) (procedure-name callee)) | |
444 | ;; a primitive | |
445 | ((symbol? callee) callee) | |
446 | (else #f)) | |
e3997e70 | 447 | #f |
3f9f4a2d AW |
448 | (with-output-to-string (lambda () (write callee))) |
449 | (and call-counts (hashv-ref call-counts callee)) | |
450 | 0 | |
451 | 0))) | |
452 | (hashv-set! table callee data) | |
453 | data)))) | |
454 | ||
455 | (when call-counts | |
456 | (hash-for-each (lambda (callee count) | |
457 | (callee->call-data callee)) | |
458 | call-counts)) | |
459 | ||
460 | (let visit-stacks ((pos 0)) | |
461 | (cond | |
462 | ((< pos len) | |
463 | ;; FIXME: if we are counting all procedure calls, and | |
464 | ;; count-call is on the stack, we need to not count the part | |
465 | ;; of the stack that is within count-call. | |
466 | (inc-call-data-self-sample-count! | |
467 | (callee->call-data (vector-ref buffer pos))) | |
468 | (let visit-stack ((pos pos)) | |
cd073eb4 AW |
469 | (cond |
470 | ((vector-ref buffer pos) | |
471 | => (lambda (callee) | |
472 | (inc-call-data-cum-sample-count! (callee->call-data callee)) | |
473 | (visit-stack (1+ pos)))) | |
474 | (else | |
475 | (visit-stacks (1+ pos)))))) | |
3f9f4a2d AW |
476 | (else table))))) |
477 | ||
cd073eb4 AW |
478 | (define (stack-samples->callee-lists state) |
479 | (let ((buffer (buffer state)) | |
480 | (len (buffer-pos state))) | |
481 | (let visit-stacks ((pos 0) (out '())) | |
482 | (cond | |
483 | ((< pos len) | |
484 | ;; FIXME: if we are counting all procedure calls, and | |
485 | ;; count-call is on the stack, we need to not count the part | |
486 | ;; of the stack that is within count-call. | |
487 | (let visit-stack ((pos pos) (stack '())) | |
488 | (cond | |
489 | ((vector-ref buffer pos) | |
490 | => (lambda (callee) | |
491 | (visit-stack (1+ pos) (cons callee stack)))) | |
492 | (else | |
493 | (visit-stacks (1+ pos) (cons (reverse stack) out)))))) | |
494 | (else (reverse out)))))) | |
495 | ||
47f3ce52 AW |
496 | (define (statprof-fold-call-data proc init) |
497 | "Fold @var{proc} over the call-data accumulated by statprof. Cannot be | |
498 | called while statprof is active. @var{proc} should take two arguments, | |
499 | @code{(@var{call-data} @var{prior-result})}. | |
500 | ||
501 | Note that a given proc-name may appear multiple times, but if it does, | |
502 | it represents different functions with the same name." | |
4d0c358b AW |
503 | (when (statprof-active?) |
504 | (error "Can't call statprof-fold-call-data while profiler is running.")) | |
47f3ce52 AW |
505 | (hash-fold |
506 | (lambda (key value prior-result) | |
507 | (proc value prior-result)) | |
508 | init | |
3f9f4a2d | 509 | (stack-samples->procedure-data (existing-profiler-state)))) |
47f3ce52 AW |
510 | |
511 | (define (statprof-proc-call-data proc) | |
512 | "Returns the call-data associated with @var{proc}, or @code{#f} if | |
513 | none is available." | |
4d0c358b AW |
514 | (when (statprof-active?) |
515 | (error "Can't call statprof-proc-call-data while profiler is running.")) | |
3f9f4a2d AW |
516 | (hashv-ref (stack-samples->procedure-data (existing-profiler-state)) |
517 | (cond | |
518 | ((primitive? proc) (procedure-name proc)) | |
519 | ((program? proc) (program-code proc)) | |
520 | (else (program-code proc))))) | |
47f3ce52 AW |
521 | |
522 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; | |
523 | ;; Stats | |
524 | ||
e3997e70 AW |
525 | (define-record-type stats |
526 | (make-stats proc-name %-time-in-proc cum-secs-in-proc self-secs-in-proc | |
527 | calls self-secs-per-call cum-secs-per-call) | |
528 | stats? | |
529 | (proc-name statprof-stats-proc-name) | |
530 | (%-time-in-proc statprof-stats-%-time-in-proc) | |
531 | (cum-secs-in-proc statprof-stats-cum-secs-in-proc) | |
532 | (self-secs-in-proc statprof-stats-self-secs-in-proc) | |
533 | (calls statprof-stats-calls) | |
534 | (self-secs-per-call statprof-stats-self-secs-per-call) | |
535 | (cum-secs-per-call statprof-stats-cum-secs-per-call)) | |
536 | ||
47f3ce52 AW |
537 | (define (statprof-call-data->stats call-data) |
538 | "Returns an object of type @code{statprof-stats}." | |
45a7de82 | 539 | (define state (existing-profiler-state)) |
62fd93e2 | 540 | |
c165c50d | 541 | (let* ((proc-name (call-data-printable call-data)) |
47f3ce52 AW |
542 | (self-samples (call-data-self-sample-count call-data)) |
543 | (cum-samples (call-data-cum-sample-count call-data)) | |
544 | (all-samples (statprof-sample-count)) | |
545 | (secs-per-sample (/ (statprof-accumulated-time) | |
546 | (statprof-sample-count))) | |
3f9f4a2d AW |
547 | (num-calls (and (call-counts state) |
548 | (statprof-call-data-calls call-data)))) | |
47f3ce52 | 549 | |
e3997e70 AW |
550 | (make-stats proc-name |
551 | (* (/ self-samples all-samples) 100.0) | |
552 | (* cum-samples secs-per-sample 1.0) | |
553 | (* self-samples secs-per-sample 1.0) | |
554 | num-calls | |
555 | (and num-calls ;; maybe we only sampled in children | |
556 | (if (zero? self-samples) 0.0 | |
557 | (/ (* self-samples secs-per-sample) 1.0 num-calls))) | |
558 | (and num-calls ;; cum-samples must be positive | |
559 | (/ (* cum-samples secs-per-sample) | |
560 | 1.0 | |
561 | ;; num-calls might be 0 if we entered statprof during the | |
562 | ;; dynamic extent of the call | |
563 | (max num-calls 1)))))) | |
47f3ce52 AW |
564 | |
565 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; | |
566 | ||
567 | (define (stats-sorter x y) | |
568 | (let ((diff (- (statprof-stats-self-secs-in-proc x) | |
569 | (statprof-stats-self-secs-in-proc y)))) | |
570 | (positive? | |
571 | (if (= diff 0) | |
572 | (- (statprof-stats-cum-secs-in-proc x) | |
573 | (statprof-stats-cum-secs-in-proc y)) | |
574 | diff)))) | |
575 | ||
91db6c4f AW |
576 | (define* (statprof-display #:optional (port (current-output-port)) |
577 | (state (existing-profiler-state))) | |
47f3ce52 AW |
578 | "Displays a gprof-like summary of the statistics collected. Unless an |
579 | optional @var{port} argument is passed, uses the current output port." | |
47f3ce52 AW |
580 | (cond |
581 | ((zero? (statprof-sample-count)) | |
582 | (format port "No samples recorded.\n")) | |
583 | (else | |
584 | (let* ((stats-list (statprof-fold-call-data | |
585 | (lambda (data prior-value) | |
586 | (cons (statprof-call-data->stats data) | |
587 | prior-value)) | |
588 | '())) | |
589 | (sorted-stats (sort stats-list stats-sorter))) | |
590 | ||
591 | (define (display-stats-line stats) | |
3f9f4a2d AW |
592 | (format port "~6,2f ~9,2f ~9,2f" |
593 | (statprof-stats-%-time-in-proc stats) | |
594 | (statprof-stats-cum-secs-in-proc stats) | |
595 | (statprof-stats-self-secs-in-proc stats)) | |
596 | (if (call-counts state) | |
597 | (if (statprof-stats-calls stats) | |
598 | (format port " ~7d ~8,2f ~8,2f " | |
599 | (statprof-stats-calls stats) | |
600 | (* 1000 (statprof-stats-self-secs-per-call stats)) | |
601 | (* 1000 (statprof-stats-cum-secs-per-call stats))) | |
602 | (format port " ")) | |
603 | (display " " port)) | |
47f3ce52 AW |
604 | (display (statprof-stats-proc-name stats) port) |
605 | (newline port)) | |
606 | ||
3f9f4a2d | 607 | (if (call-counts state) |
47f3ce52 AW |
608 | (begin |
609 | (format port "~5a ~10a ~7a ~8a ~8a ~8a ~8@a\n" | |
610 | "% " "cumulative" "self" "" "self" "total" "") | |
611 | (format port "~5a ~9a ~8a ~8a ~8a ~8a ~8@a\n" | |
612 | "time" "seconds" "seconds" "calls" "ms/call" "ms/call" "name")) | |
613 | (begin | |
614 | (format port "~5a ~10a ~7a ~8@a\n" | |
615 | "%" "cumulative" "self" "") | |
616 | (format port "~5a ~10a ~7a ~8@a\n" | |
617 | "time" "seconds" "seconds" "name"))) | |
618 | ||
619 | (for-each display-stats-line sorted-stats) | |
620 | ||
621 | (display "---\n" port) | |
622 | (simple-format #t "Sample count: ~A\n" (statprof-sample-count)) | |
623 | (simple-format #t "Total time: ~A seconds (~A seconds in GC)\n" | |
624 | (statprof-accumulated-time) | |
3476a369 AW |
625 | (/ (gc-time-taken state) |
626 | 1.0 internal-time-units-per-second)))))) | |
47f3ce52 | 627 | |
91db6c4f AW |
628 | (define* (statprof-display-anomalies #:optional (state |
629 | (existing-profiler-state))) | |
630 | "A sanity check that attempts to detect anomalies in statprof's | |
47f3ce52 AW |
631 | statistics.@code{}" |
632 | (statprof-fold-call-data | |
633 | (lambda (data prior-value) | |
3f9f4a2d | 634 | (when (and (call-counts state) |
cad444e3 AW |
635 | (zero? (call-data-call-count data)) |
636 | (positive? (call-data-cum-sample-count data))) | |
637 | (simple-format #t | |
638 | "==[~A ~A ~A]\n" | |
639 | (call-data-name data) | |
640 | (call-data-call-count data) | |
641 | (call-data-cum-sample-count data)))) | |
47f3ce52 AW |
642 | #f) |
643 | (simple-format #t "Total time: ~A\n" (statprof-accumulated-time)) | |
644 | (simple-format #t "Sample count: ~A\n" (statprof-sample-count))) | |
645 | ||
91db6c4f AW |
646 | (define (statprof-display-anomolies) |
647 | (issue-deprecation-warning "statprof-display-anomolies is a misspelling. " | |
648 | "Use statprof-display-anomalies instead.") | |
649 | (statprof-display-anomalies)) | |
650 | ||
651 | (define* (statprof-accumulated-time #:optional (state | |
652 | (existing-profiler-state))) | |
47f3ce52 | 653 | "Returns the time accumulated during the last statprof run.@code{}" |
91db6c4f | 654 | (/ (accumulated-time state) 1.0 internal-time-units-per-second)) |
47f3ce52 | 655 | |
91db6c4f | 656 | (define* (statprof-sample-count #:optional (state (existing-profiler-state))) |
47f3ce52 | 657 | "Returns the number of samples taken during the last statprof run.@code{}" |
91db6c4f | 658 | (sample-count state)) |
47f3ce52 AW |
659 | |
660 | (define statprof-call-data-name call-data-name) | |
661 | (define statprof-call-data-calls call-data-call-count) | |
662 | (define statprof-call-data-cum-samples call-data-cum-sample-count) | |
663 | (define statprof-call-data-self-samples call-data-self-sample-count) | |
664 | ||
91db6c4f | 665 | (define* (statprof-fetch-stacks #:optional (state (existing-profiler-state))) |
47f3ce52 | 666 | "Returns a list of stacks, as they were captured since the last call |
cd073eb4 AW |
667 | to @code{statprof-reset}." |
668 | (stack-samples->callee-lists state)) | |
47f3ce52 AW |
669 | |
670 | (define procedure=? | |
663212bb AW |
671 | (lambda (a b) |
672 | (cond | |
673 | ((eq? a b)) | |
0bd1e9c6 | 674 | ((and (program? a) (program? b)) |
d1100525 | 675 | (eq? (program-code a) (program-code b))) |
663212bb AW |
676 | (else |
677 | #f)))) | |
47f3ce52 AW |
678 | |
679 | ;; tree ::= (car n . tree*) | |
680 | ||
681 | (define (lists->trees lists equal?) | |
682 | (let lp ((in lists) (n-terminal 0) (tails '())) | |
683 | (cond | |
684 | ((null? in) | |
685 | (let ((trees (map (lambda (tail) | |
686 | (cons (car tail) | |
687 | (lists->trees (cdr tail) equal?))) | |
688 | tails))) | |
689 | (cons (apply + n-terminal (map cadr trees)) | |
690 | (sort trees | |
691 | (lambda (a b) (> (cadr a) (cadr b))))))) | |
692 | ((null? (car in)) | |
693 | (lp (cdr in) (1+ n-terminal) tails)) | |
694 | ((find (lambda (x) (equal? (car x) (caar in))) | |
695 | tails) | |
696 | => (lambda (tail) | |
697 | (lp (cdr in) | |
698 | n-terminal | |
699 | (assq-set! tails | |
700 | (car tail) | |
701 | (cons (cdar in) (cdr tail)))))) | |
702 | (else | |
703 | (lp (cdr in) | |
704 | n-terminal | |
705 | (acons (caar in) (list (cdar in)) tails)))))) | |
706 | ||
91db6c4f | 707 | (define* (statprof-fetch-call-tree #:optional (state (existing-profiler-state))) |
47f3ce52 AW |
708 | "Return a call tree for the previous statprof run. |
709 | ||
710 | The return value is a list of nodes, each of which is of the type: | |
711 | @code | |
712 | node ::= (@var{proc} @var{count} . @var{nodes}) | |
713 | @end code" | |
cd073eb4 AW |
714 | (define (callee->printable callee) |
715 | (cond | |
716 | ((number? callee) | |
717 | (addr->printable callee (find-program-debug-info callee))) | |
718 | (else | |
719 | (with-output-to-string (lambda () (write callee)))))) | |
720 | (define (memoizev/1 proc table) | |
721 | (lambda (x) | |
722 | (cond | |
723 | ((hashv-get-handle table x) => cdr) | |
724 | (else | |
725 | (let ((res (proc x))) | |
726 | (hashv-set! table x res) | |
727 | res))))) | |
728 | (let ((callee->printable (memoizev/1 callee->printable (make-hash-table)))) | |
729 | (cons #t (lists->trees (map (lambda (callee-list) | |
730 | (map callee->printable callee-list)) | |
731 | (stack-samples->callee-lists state)) | |
732 | equal?)))) | |
47f3ce52 | 733 | |
e1138ba1 | 734 | (define* (statprof thunk #:key (loop 1) (hz 100) (count-calls? #f) |
cd073eb4 | 735 | (port (current-output-port)) full-stacks?) |
e1138ba1 AW |
736 | "Profiles the execution of @var{thunk}. |
737 | ||
738 | The stack will be sampled @var{hz} times per second, and the thunk itself will | |
739 | be called @var{loop} times. | |
740 | ||
741 | If @var{count-calls?} is true, all procedure calls will be recorded. This | |
cd073eb4 | 742 | operation is somewhat expensive." |
e1138ba1 | 743 | |
13a977dd AW |
744 | (let ((state (fresh-profiler-state #:count-calls? count-calls? |
745 | #:sampling-period | |
cd073eb4 | 746 | (inexact->exact (round (/ 1e6 hz)))))) |
fd5dfcce AW |
747 | (parameterize ((profiler-state state)) |
748 | (dynamic-wind | |
749 | (lambda () | |
13a977dd | 750 | (statprof-start state)) |
fd5dfcce AW |
751 | (lambda () |
752 | (let lp ((i loop)) | |
753 | (unless (zero? i) | |
754 | (thunk) | |
755 | (lp (1- i))))) | |
756 | (lambda () | |
13a977dd AW |
757 | (statprof-stop state) |
758 | (statprof-display port state)))))) | |
e1138ba1 | 759 | |
47f3ce52 AW |
760 | (define-macro (with-statprof . args) |
761 | "Profiles the expressions in its body. | |
762 | ||
763 | Keyword arguments: | |
764 | ||
765 | @table @code | |
766 | @item #:loop | |
767 | Execute the body @var{loop} number of times, or @code{#f} for no looping | |
768 | ||
769 | default: @code{#f} | |
770 | @item #:hz | |
771 | Sampling rate | |
772 | ||
773 | default: @code{20} | |
774 | @item #:count-calls? | |
775 | Whether to instrument each function call (expensive) | |
776 | ||
47f3ce52 AW |
777 | default: @code{#f} |
778 | @end table" | |
779 | (define (kw-arg-ref kw args def) | |
780 | (cond | |
781 | ((null? args) (error "Invalid macro body")) | |
782 | ((keyword? (car args)) | |
783 | (if (eq? (car args) kw) | |
784 | (cadr args) | |
785 | (kw-arg-ref kw (cddr args) def))) | |
786 | ((eq? kw #f def) ;; asking for the body | |
787 | args) | |
788 | (else def))) ;; kw not found | |
e1138ba1 AW |
789 | `((@ (statprof) statprof) |
790 | (lambda () ,@(kw-arg-ref #f args #f)) | |
791 | #:loop ,(kw-arg-ref #:loop args 1) | |
792 | #:hz ,(kw-arg-ref #:hz args 100) | |
793 | #:count-calls? ,(kw-arg-ref #:count-calls? args #f) | |
794 | #:full-stacks? ,(kw-arg-ref #:full-stacks? args #f))) | |
795 | ||
cd073eb4 | 796 | (define* (gcprof thunk #:key (loop 1) full-stacks?) |
2d239a78 AW |
797 | "Do an allocation profile of the execution of @var{thunk}. |
798 | ||
799 | The stack will be sampled soon after every garbage collection, yielding | |
800 | an approximate idea of what is causing allocation in your program. | |
801 | ||
802 | Since GC does not occur very frequently, you may need to use the | |
803 | @var{loop} parameter, to cause @var{thunk} to be called @var{loop} | |
cd073eb4 | 804 | times." |
2d239a78 | 805 | |
cd073eb4 | 806 | (let ((state (fresh-profiler-state))) |
fd5dfcce | 807 | (parameterize ((profiler-state state)) |
fd5dfcce | 808 | (define (gc-callback) |
a7ede58d | 809 | (unless (inside-profiler? state) |
fd5dfcce AW |
810 | (set-inside-profiler?! state #t) |
811 | ||
812 | ;; FIXME: should be able to set an outer frame for the stack cut | |
813 | (let ((stop-time (get-internal-run-time)) | |
814 | ;; Cut down to gc-callback, and then one before (the | |
815 | ;; after-gc async). See the note in profile-signal-handler | |
816 | ;; also. | |
817 | (stack (or (make-stack #t gc-callback 0 1) | |
818 | (pk 'what! (make-stack #t))))) | |
819 | (sample-stack-procs state stack) | |
820 | (accumulate-time state stop-time) | |
821 | (set-last-start-time! state (get-internal-run-time))) | |
cd073eb4 | 822 | |
a7ede58d | 823 | (set-inside-profiler?! state #f))) |
fd5dfcce AW |
824 | |
825 | (dynamic-wind | |
826 | (lambda () | |
a7ede58d AW |
827 | (set-profile-level! state 1) |
828 | (set-last-start-time! state (get-internal-run-time)) | |
829 | (set-gc-time-taken! state (assq-ref (gc-stats) 'gc-time-taken)) | |
830 | (add-hook! after-gc-hook gc-callback)) | |
fd5dfcce AW |
831 | (lambda () |
832 | (let lp ((i loop)) | |
833 | (unless (zero? i) | |
834 | (thunk) | |
835 | (lp (1- i))))) | |
836 | (lambda () | |
a7ede58d AW |
837 | (remove-hook! after-gc-hook gc-callback) |
838 | (set-gc-time-taken! state | |
839 | (- (assq-ref (gc-stats) 'gc-time-taken) | |
840 | (gc-time-taken state))) | |
841 | (accumulate-time state (get-internal-run-time)) | |
842 | (set-profile-level! state 0) | |
fd5dfcce | 843 | (statprof-display)))))) |