1 ;;; Copyright (C) 2003 Gerd Moellmann <gerd.moellmann@t-online.de>
2 ;;; All rights reserved.
4 ;;; Redistribution and use in source and binary forms, with or without
5 ;;; modification, are permitted provided that the following conditions
8 ;;; 1. Redistributions of source code must retain the above copyright
9 ;;; notice, this list of conditions and the following disclaimer.
10 ;;; 2. Redistributions in binary form must reproduce the above copyright
11 ;;; notice, this list of conditions and the following disclaimer in the
12 ;;; documentation and/or other materials provided with the distribution.
13 ;;; 3. The name of the author may not be used to endorse or promote
14 ;;; products derived from this software without specific prior written
17 ;;; THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS
18 ;;; OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
19 ;;; WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 ;;; ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE
21 ;;; LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
22 ;;; CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT
23 ;;; OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
24 ;;; BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
25 ;;; LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26 ;;; (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
27 ;;; USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
30 ;;; Statistical profiler.
34 ;;; This profiler arranges for SIGPROF interrupts to interrupt a
35 ;;; running program at regular intervals. Each time a SIGPROF occurs,
36 ;;; the current program counter and return address is recorded in a
37 ;;; vector, until a configurable maximum number of samples have been
40 ;;; A profiling report is generated from the samples array by
41 ;;; determining the Lisp functions corresponding to the recorded
42 ;;; addresses. Each program counter/return address pair forms one
43 ;;; edge in a call graph.
47 ;;; The code being generated on x86 makes determining callers reliably
48 ;;; something between extremely difficult and impossible. Example:
50 ;;; 10979F00: .entry eval::eval-stack-args(arg-count)
51 ;;; 18: pop dword ptr [ebp-8]
52 ;;; 1B: lea esp, [ebp-32]
57 ;;; 29: mov [ebp-12], edi
58 ;;; 2C: mov dword ptr [ebp-16], #x28F0000B ; nil
59 ;;; ; No-arg-parsing entry point
60 ;;; 33: mov dword ptr [ebp-20], 0
62 ;;; 3C: L0: mov edx, esp
64 ;;; 41: mov eax, [#x10979EF8] ; #<FDEFINITION object for eval::eval-stack-pop>
66 ;;; 49: mov [edx-4], ebp
68 ;;; 4E: call dword ptr [eax+5]
71 ;;; Suppose this function is interrupted by SIGPROF at 4E. At that
72 ;;; point, the frame pointer EBP has been modified so that the
73 ;;; original return address of the caller of eval-stack-args is no
74 ;;; longer where it can be found by x86-call-context, and the new
75 ;;; return address, for the call to eval-stack-pop, is not yet on the
76 ;;; stack. The effect is that x86-call-context returns something
77 ;;; bogus, which leads to wrong edges in the call graph.
79 ;;; One thing that one might try is filtering cases where the program
80 ;;; is interrupted at a call instruction. But since the above example
81 ;;; of an interrupt at a call instruction isn't the only case where
82 ;;; the stack is something x86-call-context can't really cope with,
83 ;;; this is not a general solution.
85 ;;; Random ideas for implementation:
87 ;;; * Space profiler. Sample when new pages are allocated instead of
90 ;;; * Record a configurable number of callers up the stack. That
91 ;;; could give a more complete graph when there are many small
94 ;;; * Print help strings for reports, include hints to the problem
97 ;;; * Make flat report the default since call-graph isn't that
100 (defpackage #:sb-sprof
101 (:use #:cl #:sb-ext #:sb-unix #:sb-alien #:sb-sys)
102 (:export #:*sample-interval* #:*max-samples*
103 #:start-sampling #:stop-sampling #:with-sampling
104 #:with-profiling #:start-profiling #:stop-profiling
107 (in-package #:sb-sprof)
112 (defstruct (vertex (:constructor make-vertex)
113 (:constructor make-scc (scc-vertices edges)))
114 (visited nil :type boolean)
115 (root nil :type (or null vertex))
117 (edges () :type list)
118 (scc-vertices () :type list))
121 (vertex (sb-impl::missing-arg) :type vertex))
124 (vertices () :type list))
126 (declaim (inline scc-p))
127 (defun scc-p (vertex)
128 (not (null (vertex-scc-vertices vertex))))
130 (defmacro do-vertices ((vertex graph) &body body)
131 `(dolist (,vertex (graph-vertices ,graph))
134 (defmacro do-edges ((edge edge-to vertex) &body body)
135 `(dolist (,edge (vertex-edges ,vertex))
136 (let ((,edge-to (edge-vertex ,edge)))
139 (defun self-cycle-p (vertex)
140 (do-edges (e to vertex)
144 (defun map-vertices (fn vertices)
146 (setf (vertex-visited v) nil))
148 (unless (vertex-visited v)
151 ;;; Eeko Nuutila, Eljas Soisalon-Soininen, around 1992. Improves on
152 ;;; Tarjan's original algorithm by not using the stack when processing
153 ;;; trivial components. Trivial components should appear frequently
154 ;;; in a call-graph such as ours, I think. Same complexity O(V+E) as
156 (defun strong-components (vertices)
157 (let ((in-component (make-array (length vertices)
158 :element-type 'boolean
159 :initial-element nil))
163 (labels ((min-root (x y)
164 (let ((rx (vertex-root x))
165 (ry (vertex-root y)))
166 (if (< (vertex-dfn rx) (vertex-dfn ry))
170 (aref in-component (vertex-dfn v)))
171 ((setf in-component) (in v)
172 (setf (aref in-component (vertex-dfn v)) in))
174 (> (vertex-dfn x) (vertex-dfn y)))
176 (setf (vertex-dfn v) (incf dfn)
179 (vertex-visited v) t)
181 (unless (vertex-visited w)
183 (unless (in-component w)
184 (setf (vertex-root v) (min-root v w))))
185 (if (eq v (vertex-root v))
186 (loop while (and stack (vertex-> (car stack) v))
188 collect w into this-component
189 do (setf (in-component w) t)
191 (setf (in-component v) t)
192 (push (cons v this-component) components))
194 (map-vertices #'visit vertices)
197 ;;; Given a dag as a list of vertices, return the list sorted
198 ;;; topologically, children first.
199 (defun topological-sort (dag)
202 (labels ((rec-sort (v)
203 (setf (vertex-visited v) t)
204 (setf (vertex-dfn v) (incf dfn))
205 (dolist (e (vertex-edges v))
206 (unless (vertex-visited (edge-vertex e))
207 (rec-sort (edge-vertex e))))
209 (map-vertices #'rec-sort dag)
212 ;;; Reduce graph G to a dag by coalescing strongly connected components
213 ;;; into vertices. Sort the result topologically.
214 (defun reduce-graph (graph &optional (scc-constructor #'make-scc))
215 (sb-int:collect ((sccs) (trivial))
216 (dolist (c (strong-components (graph-vertices graph)))
217 (if (or (cdr c) (self-cycle-p (car c)))
218 (sb-int:collect ((outgoing))
223 (sccs (funcall scc-constructor c (outgoing))))
226 (dolist (v (trivial))
228 (when (member w (vertex-scc-vertices scc))
229 (setf (edge-vertex e) scc)))))
230 (setf (graph-vertices graph)
231 (topological-sort (nconc (sccs) (trivial))))))
236 ;;; An AA tree is a red-black tree with the extra condition that left
237 ;;; children may not be red. This condition simplifies the red-black
238 ;;; algorithm. It eliminates half of the restructuring cases, and
239 ;;; simplifies the delete algorithm.
241 (defstruct (aa-node (:conc-name aa-))
242 (left nil :type (or null aa-node))
243 (right nil :type (or null aa-node))
244 (level 0 :type integer)
248 (let ((node (make-aa-node)))
249 (setf (aa-left node) node)
250 (setf (aa-right node) node)
254 (root *null-node* :type aa-node))
256 (declaim (inline skew split rotate-with-left-child rotate-with-right-child))
258 (defun rotate-with-left-child (k2)
259 (let ((k1 (aa-left k2)))
260 (setf (aa-left k2) (aa-right k1))
261 (setf (aa-right k1) k2)
264 (defun rotate-with-right-child (k1)
265 (let ((k2 (aa-right k1)))
266 (setf (aa-right k1) (aa-left k2))
267 (setf (aa-left k2) k1)
271 (if (= (aa-level (aa-left aa)) (aa-level aa))
272 (rotate-with-left-child aa)
276 (when (= (aa-level (aa-right (aa-right aa)))
278 (setq aa (rotate-with-right-child aa))
279 (incf (aa-level aa)))
282 (macrolet ((def (name () &body body)
283 (let ((name (sb-int::symbolicate 'aa- name)))
284 `(defun ,name (item tree &key
285 (test-< #'<) (test-= #'=)
286 (node-key #'identity) (item-key #'identity))
287 (let ((.item-key. (funcall item-key item)))
288 (flet ((item-< (node)
289 (funcall test-< .item-key.
290 (funcall node-key (aa-data node))))
292 (funcall test-= .item-key.
293 (funcall node-key (aa-data node)))))
294 (declare (inline item-< item-=))
298 (labels ((insert-into (aa)
299 (cond ((eq aa *null-node*)
300 (setq aa (make-aa-node :data item
302 :right *null-node*)))
304 (return-from insert-into aa))
306 (setf (aa-left aa) (insert-into (aa-left aa))))
308 (setf (aa-right aa) (insert-into (aa-right aa)))))
310 (setf (aa-tree-root tree)
311 (insert-into (aa-tree-root tree)))))
314 (let ((deleted-node *null-node*)
316 (labels ((remove-from (aa)
317 (unless (eq aa *null-node*)
320 (setf (aa-left aa) (remove-from (aa-left aa)))
322 (setq deleted-node aa)
323 (setf (aa-right aa) (remove-from (aa-right aa)))))
324 (cond ((eq aa last-node)
326 ;; If at the bottom of the tree, and item
327 ;; is present, delete it.
328 (when (and (not (eq deleted-node *null-node*))
329 (item-= deleted-node))
330 (setf (aa-data deleted-node) (aa-data aa))
331 (setq deleted-node *null-node*)
332 (setq aa (aa-right aa))))
334 ;; Otherwise not at bottom of tree; rebalance.
335 ((or (< (aa-level (aa-left aa))
337 (< (aa-level (aa-right aa))
340 (when (> (aa-level (aa-right aa)) (aa-level aa))
341 (setf (aa-level (aa-right aa)) (aa-level aa)))
343 (setf (aa-right aa) (skew (aa-right aa)))
344 (setf (aa-right (aa-right aa))
345 (skew (aa-right (aa-right aa))))
347 (setf (aa-right aa) (split (aa-right aa))))))
349 (setf (aa-tree-root tree)
350 (remove-from (aa-tree-root tree))))))
353 (let ((current (aa-tree-root tree)))
354 (setf (aa-data *null-node*) item)
356 (cond ((eq current *null-node*)
357 (return (values nil nil)))
359 (return (values (aa-data current) t)))
361 (setq current (aa-left current)))
363 (setq current (aa-right current))))))))
368 ;;; Sort the subsequence of Vec in the interval [From To] using
369 ;;; comparison function Test. Assume each element to sort consists of
370 ;;; Element-Size array slots, and that the slot Key-Offset contains
372 (defun qsort (vec &key (element-size 1) (key-offset 0)
373 (from 0) (to (- (length vec) element-size)))
374 (declare (type fixnum to from element-size key-offset))
375 (declare (type (simple-array address) vec))
376 (labels ((rotate (i j)
377 (declare (fixnum i j))
378 (loop repeat element-size
379 for i from i and j from j do
380 (rotatef (aref vec i) (aref vec j))))
382 (aref vec (+ i key-offset)))
384 (declare (fixnum to from))
386 (let* ((mid (* element-size
387 (round (+ (/ from element-size)
391 (j (+ to element-size))
393 (declare (fixnum mid i j))
396 (loop do (incf i element-size)
398 ;; QSORT used to take a test
399 ;; parameter which was funcalled
400 ;; here. This caused some consing,
401 ;; which is problematic since
402 ;; QSORT is indirectly called in
403 ;; an after-gc-hook. So just
404 ;; hardcode >, which would've been
405 ;; used for the test anyway.
408 (loop do (decf j element-size)
409 until (or (<= j from)
412 (when (< j i) (return))
415 (rec-sort from (- j element-size))
424 "Type used for addresses, for instance, program counters,
425 code start/end locations etc."
426 '(unsigned-byte #.sb-vm::n-machine-word-bits))
428 (defconstant +unknown-address+ 0
429 "Constant representing an address that cannot be determined.")
431 ;;; A call graph. Vertices are NODE structures, edges are CALL
433 (defstruct (call-graph (:include graph)
434 (:constructor %make-call-graph))
435 ;; the value of *Sample-Interval* at the time the graph was created
436 (sample-interval (sb-impl::missing-arg) :type number)
437 ;; number of samples taken
438 (nsamples (sb-impl::missing-arg) :type sb-impl::index)
439 ;; sample count for samples not in any function
440 (elsewhere-count (sb-impl::missing-arg) :type sb-impl::index)
441 ;; a flat list of NODEs, sorted by sample count
442 (flat-nodes () :type list))
444 ;;; A node in a call graph, representing a function that has been
445 ;;; sampled. The edges of a node are CALL structures that represent
446 ;;; functions called from a given node.
447 (defstruct (node (:include vertex)
448 (:constructor %make-node))
449 ;; A numeric label for the node. The most frequently called function
450 ;; gets label 1. This is just for identification purposes in the
452 (index 0 :type fixnum)
453 ;; start and end address of the function's code
454 (start-pc 0 :type address)
455 (end-pc 0 :type address)
456 ;; the name of the function
458 ;; sample count for this function
459 (count 0 :type fixnum)
460 ;; count including time spent in functions called from this one
461 (accrued-count 0 :type fixnum)
462 ;; list of NODEs for functions calling this one
463 (callers () :type list))
465 ;;; A cycle in a call graph. The functions forming the cycle are
466 ;;; found in the SCC-VERTICES slot of the VERTEX structure.
467 (defstruct (cycle (:include node)))
469 ;;; An edge in a call graph. EDGE-VERTEX is the function being
471 (defstruct (call (:include edge)
472 (:constructor make-call (vertex)))
473 ;; number of times the call was sampled
474 (count 1 :type sb-impl::index))
476 ;;; Info about a function in dynamic-space. This is used to track
477 ;;; address changes of functions during GC.
478 (defstruct (dyninfo (:constructor make-dyninfo (code start end)))
479 ;; component this info is for
480 (code (sb-impl::missing-arg) :type sb-kernel::code-component)
481 ;; current start and end address of the component
482 (start (sb-impl::missing-arg) :type address)
483 (end (sb-impl::missing-arg) :type address)
484 ;; new start address of the component, after GC.
485 (new-start 0 :type address))
487 (defmethod print-object ((call-graph call-graph) stream)
488 (print-unreadable-object (call-graph stream :type t :identity t)
489 (format stream "~d samples" (call-graph-nsamples call-graph))))
491 (defmethod print-object ((node node) stream)
492 (print-unreadable-object (node stream :type t :identity t)
493 (format stream "~s [~d]" (node-name node) (node-index node))))
495 (defmethod print-object ((call call) stream)
496 (print-unreadable-object (call stream :type t :identity t)
497 (format stream "~s [~d]" (node-name (call-vertex call))
498 (node-index (call-vertex call)))))
500 (deftype report-type ()
501 '(member nil :flat :graph))
503 (defvar *sample-interval* 0.01
504 "Default number of seconds between samples.")
505 (declaim (number *sample-interval*))
507 (defvar *max-samples* 50000
508 "Default number of samples taken.")
509 (declaim (type sb-impl::index *max-samples*))
511 (defconstant +sample-size+
515 (defvar *samples* nil)
516 (declaim (type (or null (vector address)) *samples*))
518 (defvar *samples-index* 0)
519 (declaim (type sb-impl::index *samples-index*))
521 (defvar *profiling* nil)
522 (defvar *sampling* nil)
523 (declaim (type boolean *profiling* *sampling*))
525 (defvar *dynamic-space-code-info* ())
526 (declaim (type list *dynamic-space-code-info*))
528 (defvar *show-progress* nil)
530 (defvar *old-sampling* nil)
532 (defun turn-off-sampling ()
533 (setq *old-sampling* *sampling*)
534 (setq *sampling* nil))
536 (defun turn-on-sampling ()
537 (setq *sampling* *old-sampling*))
539 (defun show-progress (format-string &rest args)
540 (when *show-progress*
541 (apply #'format t format-string args)
544 (defun start-sampling ()
545 "Switch on statistical sampling."
548 (defun stop-sampling ()
549 "Switch off statistical sampling."
550 (setq *sampling* nil))
552 (defmacro with-sampling ((&optional (on t)) &body body)
553 "Evaluate body with statistical sampling turned on or off."
554 `(let ((*sampling* ,on))
557 (defun sort-samples (key-offset)
558 "Sort *Samples* using comparison Test. Key must be one of
559 :Pc or :Return-Pc for sorting by pc or return pc."
560 (when (plusp *samples-index*)
563 :to (- *samples-index* +sample-size+)
564 :element-size +sample-size+
565 :key-offset key-offset)))
568 (declare (type address pc))
569 (setf (aref *samples* *samples-index*) pc)
570 (incf *samples-index*))
572 ;;; SIGPROF handler. Record current PC and return address in
575 (defun sigprof-handler (signal code scp)
576 (declare (ignore signal code) (type system-area-pointer scp))
577 (sb-sys:with-interrupts
578 (when (and *sampling*
579 (< *samples-index* (length *samples*)))
580 (sb-sys:without-gcing
581 (locally (declare (optimize (inhibit-warnings 2)))
582 (with-alien ((scp (* os-context-t) :local scp))
583 ;; For some reason completely bogus small values for the
584 ;; frame pointer are returned every now and then, leading
585 ;; to segfaults. Try to avoid these cases.
587 ;; FIXME: Do a more thorough sanity check on ebp, or figure
588 ;; out why this is happening.
589 ;; -- JES, 2005-01-11
590 (when (< (sb-vm::context-register scp #.sb-vm::ebp-offset)
592 (dotimes (i +sample-size+)
594 (return-from sigprof-handler nil))
595 (let* ((pc-ptr (sb-vm:context-pc scp))
596 (fp (sb-vm::context-register scp #.sb-vm::ebp-offset)))
597 (record (sap-int pc-ptr))
598 (let ((fp (int-sap fp))
600 (dotimes (i (1- +sample-size+))
603 (sb-di::x86-call-context fp :depth i))
608 (record 0))))))))))))
610 ;; FIXME: On non-x86 platforms we don't yet walk the call stack deeper
613 (defun sigprof-handler (signal code scp)
614 (declare (ignore signal code))
615 (sb-sys:with-interrupts
616 (when (and *sampling*
617 (< *samples-index* (length *samples*)))
618 (sb-sys:without-gcing
619 (with-alien ((scp (* os-context-t) :local scp))
620 (locally (declare (optimize (inhibit-warnings 2)))
621 (let* ((pc-ptr (sb-vm:context-pc scp))
622 (fp (sb-vm::context-register scp #.sb-vm::cfp-offset))
625 (* sb-vm::lra-save-offset sb-vm::n-word-bytes))))
626 (record (sap-int pc-ptr))
629 ;;; Map function FN over code objects in dynamic-space. FN is called
630 ;;; with two arguments, the object and its size in bytes.
631 (defun map-dynamic-space-code (fn)
632 (flet ((call-if-code (obj obj-type size)
633 (declare (ignore obj-type))
634 (when (sb-kernel:code-component-p obj)
635 (funcall fn obj size))))
636 (sb-vm::map-allocated-objects #'call-if-code :dynamic)))
638 ;;; Return the start address of CODE.
639 (defun code-start (code)
640 (declare (type sb-kernel:code-component code))
641 (sap-int (sb-kernel:code-instructions code)))
643 ;;; Return start and end address of CODE as multiple values.
644 (defun code-bounds (code)
645 (declare (type sb-kernel:code-component code))
646 (let* ((start (code-start code))
647 (end (+ start (sb-kernel:%code-code-size code))))
650 (defun record-dyninfo ()
651 (setf *dynamic-space-code-info* nil)
652 (flet ((record-address (code size)
653 (declare (ignore size))
654 (multiple-value-bind (start end)
656 (push (make-dyninfo code start end)
657 *dynamic-space-code-info*))))
658 (map-dynamic-space-code #'record-address)))
660 (defun adjust-samples (offset)
661 (sort-samples offset)
663 (declare (type sb-impl::index sidx))
664 (dolist (info *dynamic-space-code-info*)
665 (unless (= (dyninfo-new-start info) (dyninfo-start info))
666 (let ((pos (do ((i sidx (+ i +sample-size+)))
667 ((= i *samples-index*) nil)
668 (declare (type sb-impl::index i))
669 (when (<= (dyninfo-start info)
670 (aref *samples* (+ i offset))
675 (loop with delta = (- (dyninfo-new-start info)
676 (dyninfo-start info))
677 for j from sidx below *samples-index* by +sample-size+
678 as pc = (aref *samples* (+ j offset))
679 while (<= (dyninfo-start info) pc (dyninfo-end info)) do
680 (incf (aref *samples* (+ j offset)) delta)
681 (incf sidx +sample-size+))))))))
683 ;;; This runs from *AFTER-GC-HOOKS*. Adjust *SAMPLES* for address
684 ;;; changes of dynamic-space code objects.
685 (defun adjust-samples-for-address-changes ()
686 (sb-sys:without-gcing
688 (setq *dynamic-space-code-info*
689 (sort *dynamic-space-code-info* #'> :key #'dyninfo-start))
690 (dolist (info *dynamic-space-code-info*)
691 (setf (dyninfo-new-start info)
692 (code-start (dyninfo-code info))))
694 (dotimes (i +sample-size+)
696 (dolist (info *dynamic-space-code-info*)
697 (let ((size (- (dyninfo-end info) (dyninfo-start info))))
698 (setf (dyninfo-start info) (dyninfo-new-start info))
699 (setf (dyninfo-end info) (+ (dyninfo-new-start info) size))))
702 (defmacro with-profiling ((&key (sample-interval '*sample-interval*)
703 (max-samples '*max-samples*)
706 (report nil report-p))
708 "Repeatedly evaluate Body with statistical profiling turned on.
709 The following keyword args are recognized:
711 :Sample-Interval <seconds>
712 Take a sample every <seconds> seconds. Default is
716 Repeat evaluating body until <max> samples are taken.
717 Default is *Max-Samples*.
720 If specified, call Report with :Type <type> at the end.
723 It true, call Reset at the beginning."
724 (declare (type report-type report))
725 `(let ((*sample-interval* ,sample-interval)
726 (*max-samples* ,max-samples))
727 ,@(when reset '((reset)))
730 (when (>= *samples-index* (length *samples*))
732 ,@(when show-progress
733 `((format t "~&===> ~d of ~d samples taken.~%"
734 (/ *samples-index* +sample-size+)
736 (let ((.last-index. *samples-index*))
738 (when (= .last-index. *samples-index*)
739 (warn "No sampling progress; possibly a profiler bug.")
742 ,@(when report-p `((report :type ,report)))))
744 (defun start-profiling (&key (max-samples *max-samples*)
745 (sample-interval *sample-interval*)
747 "Start profiling statistically if not already profiling.
748 The following keyword args are recognized:
750 :Sample-Interval <seconds>
751 Take a sample every <seconds> seconds. Default is
755 Maximum number of samples. Default is *Max-Samples*.
758 If true, the default, start sampling right away.
759 If false, Start-Sampling can be used to turn sampling on."
761 (multiple-value-bind (secs usecs)
762 (multiple-value-bind (secs rest)
763 (truncate sample-interval)
764 (values secs (truncate (* rest 1000000))))
765 (setq *samples* (make-array (* max-samples +sample-size+)
766 :element-type 'address))
767 (setq *samples-index* 0)
768 (setq *sampling* sampling)
769 ;; Disabled for now, since this was causing some problems with the
770 ;; sampling getting turned off completely. --JES, 2004-06-19
772 ;; BEFORE-GC-HOOKS have exceedingly bad interactions with
773 ;; threads. -- CSR, 2004-06-21
775 ;; (pushnew 'turn-off-sampling *before-gc-hooks*)
776 (pushnew 'adjust-samples-for-address-changes *after-gc-hooks*)
778 (sb-sys:enable-interrupt sb-unix:sigprof #'sigprof-handler)
779 (unix-setitimer :profile secs usecs secs usecs)
780 (setq *profiling* t)))
783 (defun stop-profiling ()
784 "Stop profiling if profiling."
786 (setq *after-gc-hooks*
787 (delete 'adjust-samples-for-address-changes *after-gc-hooks*))
788 (unix-setitimer :profile 0 0 0 0)
789 ;; Even with the timer shut down we cannot be sure that there is
790 ;; no undelivered sigprof. Besides, leaving the signal handler
791 ;; installed won't hurt.
792 (setq *sampling* nil)
793 (setq *profiling* nil))
797 "Reset the profiler."
799 (setq *sampling* nil)
800 (setq *dynamic-space-code-info* ())
802 (setq *samples-index* 0)
805 ;;; Make a NODE for debug-info INFO.
806 (defun make-node (info)
808 (sb-kernel::code-component
809 (multiple-value-bind (start end)
811 (%make-node :name (or (sb-disassem::find-assembler-routine start)
812 (format nil "~a" info))
813 :start-pc start :end-pc end)))
814 (sb-di::compiled-debug-fun
815 (let* ((name (sb-di::debug-fun-name info))
816 (cdf (sb-di::compiled-debug-fun-compiler-debug-fun info))
817 (start-offset (sb-c::compiled-debug-fun-start-pc cdf))
818 (end-offset (sb-c::compiled-debug-fun-elsewhere-pc cdf))
819 (component (sb-di::compiled-debug-fun-component info))
820 (start-pc (code-start component)))
821 (%make-node :name name
822 :start-pc (+ start-pc start-offset)
823 :end-pc (+ start-pc end-offset))))
825 (%make-node :name (sb-di::debug-fun-name info)))
827 (%make-node :name (coerce info 'string)))))
829 ;;; Return something serving as debug info for address PC. If we can
830 ;;; get something from SB-DI:DEBUG-FUNCTION-FROM-PC, return that.
831 ;;; Otherwise, if we can determine a code component, return that.
832 ;;; Otherwise return nil.
833 (defun debug-info (pc)
834 (declare (type address pc))
835 (let ((ptr (sb-di::component-ptr-from-pc (int-sap pc))))
836 (cond ((sap= ptr (int-sap 0))
837 (let ((name (sap-foreign-symbol (int-sap pc))))
839 (format nil "foreign function ~a" name))))
841 (let* ((code (sb-di::component-from-component-ptr ptr))
842 (code-header-len (* (sb-kernel:get-header-data code)
845 (- (sb-kernel:get-lisp-obj-address code)
846 sb-vm:other-pointer-lowtag)
848 (df (ignore-errors (sb-di::debug-fun-from-pc code
854 ;;; One function can have more than one COMPILED-DEBUG-FUNCTION with
855 ;;; the same name. Reduce the number of calls to Debug-Info by first
856 ;;; looking for a given PC in a red-black tree. If not found in the
857 ;;; tree, get debug info, and look for a node in a hash-table by
858 ;;; function name. If not found in the hash-table, make a new node.
861 (defvar *name->node*)
863 (defmacro with-lookup-tables (() &body body)
864 `(let ((*node-tree* (make-aa-tree))
865 (*name->node* (make-hash-table :test 'equal)))
868 (defun tree-find (item)
869 (flet ((pc/node-= (pc node)
870 (<= (node-start-pc node) pc (node-end-pc node)))
872 (< pc (node-start-pc node))))
873 (aa-find item *node-tree* :test-= #'pc/node-= :test-< #'pc/node-<)))
875 (defun tree-insert (item)
876 (flet ((node/node-= (x y)
877 (<= (node-start-pc y) (node-start-pc x) (node-end-pc y)))
879 (< (node-start-pc x) (node-start-pc y))))
880 (aa-insert item *node-tree* :test-= #'node/node-= :test-< #'node/node-<)))
882 ;;; Find or make a new node for address PC. Value is the NODE found
883 ;;; or made; NIL if not enough information exists to make a NODE for
885 (defun lookup-node (pc)
886 (declare (type address pc))
888 (let ((info (debug-info pc)))
890 (let* ((new (make-node info))
891 (found (gethash (node-name new) *name->node*)))
893 (setf (node-start-pc found)
894 (min (node-start-pc found) (node-start-pc new)))
895 (setf (node-end-pc found)
896 (max (node-end-pc found) (node-end-pc new)))
899 (setf (gethash (node-name new) *name->node*) new)
903 ;;; Return a list of all nodes created by LOOKUP-NODE.
904 (defun collect-nodes ()
905 (loop for node being the hash-values of *name->node*
908 ;;; Value is a CALL-GRAPH for the current contents of *SAMPLES*.
909 (defun make-call-graph-1 (depth)
910 (let ((elsewhere-count 0)
912 (with-lookup-tables ()
913 (loop for i below (1- *samples-index*) ;; by +sample-size+
914 as pc = (aref *samples* i)
915 as return-pc = (aref *samples* (1+ i))
916 as callee = (lookup-node pc)
918 (when (and callee (/= return-pc +unknown-address+))
919 (let ((caller (lookup-node return-pc)))
923 (when (and *show-progress* (plusp i))
924 (cond ((zerop (mod i 1000))
925 (show-progress "~d" i))
927 (show-progress "."))))
928 (when (< (mod i +sample-size+) depth)
929 (when (= (mod i +sample-size+) 0)
930 (setf visited-nodes nil)
932 (incf (node-accrued-count callee))
933 (incf (node-count callee)))
935 (incf elsewhere-count))))
937 (push callee visited-nodes))
939 (unless (member caller visited-nodes)
940 (incf (node-accrued-count caller)))
942 (let ((call (find callee (node-edges caller)
943 :key #'call-vertex)))
944 (pushnew caller (node-callers callee))
946 (unless (member caller visited-nodes)
947 (incf (call-count call)))
948 (push (make-call callee) (node-edges caller))))))))
949 (let ((sorted-nodes (sort (collect-nodes) #'> :key #'node-count)))
950 (loop for node in sorted-nodes and i from 1 do
951 (setf (node-index node) i))
952 (%make-call-graph :nsamples (/ *samples-index* +sample-size+)
953 :sample-interval *sample-interval*
954 :elsewhere-count elsewhere-count
955 :vertices sorted-nodes)))))
957 ;;; Reduce CALL-GRAPH to a dag, creating CYCLE structures for call
959 (defun reduce-call-graph (call-graph)
961 (flet ((make-one-cycle (vertices edges)
962 (let* ((name (format nil "<Cycle ~d>" (incf cycle-no)))
963 (count (loop for v in vertices sum (node-count v))))
964 (make-cycle :name name
967 :scc-vertices vertices
969 (reduce-graph call-graph #'make-one-cycle))))
971 ;;; For all nodes in CALL-GRAPH, compute times including the time
972 ;;; spent in functions called from them. Note that the call-graph
973 ;;; vertices are in reverse topological order, children first, so we
974 ;;; will have computed accrued counts of called functions before they
975 ;;; are used to compute accrued counts for callers.
976 (defun compute-accrued-counts (call-graph)
977 (do-vertices (from call-graph)
978 (setf (node-accrued-count from) (node-count from))
979 (do-edges (call to from)
980 (incf (node-accrued-count from)
981 (round (* (/ (call-count call) (node-count to))
982 (node-accrued-count to)))))))
984 ;;; Return a CALL-GRAPH structure for the current contents of
985 ;;; *SAMPLES*. The result contain a list of nodes sorted by self-time
986 ;;; in the FLAT-NODES slot, and a dag in VERTICES, with call cycles
987 ;;; reduced to CYCLE structures.
988 (defun make-call-graph (depth)
990 (show-progress "~&Computing call graph ")
991 (let ((call-graph (without-gcing (make-call-graph-1 depth))))
992 (setf (call-graph-flat-nodes call-graph)
993 (copy-list (graph-vertices call-graph)))
994 (show-progress "~&Finding cycles")
995 (reduce-call-graph call-graph)
996 (show-progress "~&Propagating counts")
997 #+nil (compute-accrued-counts call-graph)
1003 (defun print-separator (&key (length 72) (char #\-))
1004 (format t "~&~V,,,V<~>~%" length char))
1006 (defun samples-percent (call-graph count)
1008 (* 100.0 (/ count (call-graph-nsamples call-graph)))
1011 (defun print-call-graph-header (call-graph)
1012 (let ((nsamples (call-graph-nsamples call-graph))
1013 (interval (call-graph-sample-interval call-graph))
1014 (ncycles (loop for v in (graph-vertices call-graph)
1016 (format t "~2&Number of samples: ~d~%~
1017 Sample interval: ~f seconds~%~
1018 Total sampling time: ~f seconds~%~
1019 Number of cycles: ~d~2%"
1022 (* nsamples interval)
1025 (defun print-flat (call-graph &key (stream *standard-output*) max
1026 min-percent (print-header t))
1027 (let ((*standard-output* stream)
1028 (*print-pretty* nil)
1031 (min-count (if min-percent
1032 (round (* (/ min-percent 100.0)
1033 (call-graph-nsamples call-graph)))
1036 (print-call-graph-header call-graph))
1037 (format t "~& Self Cumul Total~%")
1038 (format t "~& Nr Count % Count % Count % Function~%")
1040 (let ((elsewhere-count (call-graph-elsewhere-count call-graph))
1042 (dolist (node (call-graph-flat-nodes call-graph))
1043 (when (or (and max (> (incf i) max))
1044 (< (node-count node) min-count))
1046 (let* ((count (node-count node))
1047 (percent (samples-percent call-graph count))
1048 (accrued-count (node-accrued-count node))
1049 (accrued-percent (samples-percent call-graph accrued-count)))
1050 (incf total-count count)
1051 (incf total-percent percent)
1052 (format t "~&~4d ~6d ~5,1f ~6d ~5,1f ~6d ~5,1f ~s~%"
1063 (format t "~& ~6d ~5,1f elsewhere~%"
1065 (samples-percent call-graph elsewhere-count)))))
1067 (defun print-cycles (call-graph)
1068 (when (some #'cycle-p (graph-vertices call-graph))
1069 (format t "~& Cycle~%")
1070 (format t "~& Count % Parts~%")
1071 (do-vertices (node call-graph)
1072 (when (cycle-p node)
1073 (flet ((print-info (indent index count percent name)
1074 (format t "~&~6d ~5,1f ~11@t ~V@t ~s [~d]~%"
1075 count percent indent name index)))
1077 (format t "~&~6d ~5,1f ~a...~%"
1079 (samples-percent call-graph (cycle-count node))
1081 (dolist (v (vertex-scc-vertices node))
1082 (print-info 4 (node-index v) (node-count v)
1083 (samples-percent call-graph (node-count v))
1088 (defun print-graph (call-graph &key (stream *standard-output*)
1090 (let ((*standard-output* stream)
1091 (*print-pretty* nil))
1092 (print-call-graph-header call-graph)
1093 (print-cycles call-graph)
1094 (flet ((find-call (from to)
1095 (find to (node-edges from) :key #'call-vertex))
1096 (print-info (indent index count percent name)
1097 (format t "~&~6d ~5,1f ~11@t ~V@t ~s [~d]~%"
1098 count percent indent name index)))
1099 (format t "~& Callers~%")
1100 (format t "~& Cumul. Function~%")
1101 (format t "~& Count % Count % Callees~%")
1102 (do-vertices (node call-graph)
1105 ;; Print caller information.
1106 (dolist (caller (node-callers node))
1107 (let ((call (find-call caller node)))
1108 (print-info 4 (node-index caller)
1110 (samples-percent call-graph (call-count call))
1111 (node-name caller))))
1112 ;; Print the node itself.
1113 (format t "~&~6d ~5,1f ~6d ~5,1f ~s [~d]~%"
1115 (samples-percent call-graph (node-count node))
1116 (node-accrued-count node)
1117 (samples-percent call-graph (node-accrued-count node))
1121 (do-edges (call called node)
1122 (print-info 4 (node-index called)
1124 (samples-percent call-graph (call-count call))
1125 (node-name called))))
1128 (print-flat call-graph :stream stream :max max
1129 :min-percent min-percent :print-header nil))))
1131 (defun report (&key (type :graph) max min-percent call-graph
1132 (stream *standard-output*) ((:show-progress *show-progress*)))
1133 "Report statistical profiling results. The following keyword
1134 args are recognized:
1137 Specifies the type of report to generate. If :FLAT, show
1138 flat report, if :GRAPH show a call graph and a flat report.
1139 If nil, don't print out a report.
1142 Specify a stream to print the report on. Default is
1146 Don't show more than <max> entries in the flat report.
1148 :Min-Percent <min-percent>
1149 Don't show functions taking less than <min-percent> of the
1150 total time in the flat report.
1152 :Show-Progress <bool>
1153 If true, print progress messages while generating the call graph.
1156 Print a report from <graph> instead of the latest profiling
1159 Value of this function is a Call-Graph object representing the
1160 resulting call-graph."
1161 (let ((graph (or call-graph (make-call-graph (1- +sample-size+)))))
1164 (print-flat graph :stream stream :max max :min-percent min-percent))
1166 (print-graph graph :stream stream :max max :min-percent min-percent))
1170 ;;; Interface to DISASSEMBLE
1172 (defun add-disassembly-profile-note (chunk stream dstate)
1173 (declare (ignore chunk stream))
1174 (unless (zerop *samples-index*)
1176 (+ (sb-disassem::seg-virtual-location
1177 (sb-disassem:dstate-segment dstate))
1178 (sb-disassem::dstate-cur-offs dstate)))
1179 (samples (loop for x from 0 below *samples-index* by +sample-size+
1180 summing (if (= (aref *samples* x) location)
1183 (unless (zerop samples)
1184 (sb-disassem::note (format nil "~A/~A samples"
1185 samples (/ *samples-index* +sample-size+))
1188 (pushnew 'add-disassembly-profile-note sb-disassem::*default-dstate-hooks*)
1192 (defun test-0 (n &optional (depth 0))
1193 (declare (optimize (debug 3)))
1196 (test-0 n (1+ depth))
1197 (test-0 n (1+ depth)))))
1200 (with-profiling (:reset t :max-samples 1000 :report :graph)