mcount.c revision 37629
1241823Smarcel/*-
2241823Smarcel * Copyright (c) 1983, 1992, 1993
3241823Smarcel *	The Regents of the University of California.  All rights reserved.
4241823Smarcel *
5241823Smarcel * Redistribution and use in source and binary forms, with or without
6241823Smarcel * modification, are permitted provided that the following conditions
7241823Smarcel * are met:
8241823Smarcel * 1. Redistributions of source code must retain the above copyright
9241823Smarcel *    notice, this list of conditions and the following disclaimer.
10241823Smarcel * 2. Redistributions in binary form must reproduce the above copyright
11241823Smarcel *    notice, this list of conditions and the following disclaimer in the
12241823Smarcel *    documentation and/or other materials provided with the distribution.
13241823Smarcel * 3. All advertising materials mentioning features or use of this software
14241823Smarcel *    must display the following acknowledgement:
15241823Smarcel *	This product includes software developed by the University of
16241823Smarcel *	California, Berkeley and its contributors.
17241823Smarcel * 4. Neither the name of the University nor the names of its contributors
18241823Smarcel *    may be used to endorse or promote products derived from this software
19241823Smarcel *    without specific prior written permission.
20241823Smarcel *
21241823Smarcel * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22241823Smarcel * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23241823Smarcel * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24241823Smarcel * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25241823Smarcel * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26241823Smarcel * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27241823Smarcel * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28267181Sjmmv * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29241823Smarcel * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30241823Smarcel * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * SUCH DAMAGE.
32 */
33
34#if !defined(lint) && !defined(KERNEL) && defined(LIBC_SCCS)
35#if 0
36static char sccsid[] = "@(#)mcount.c	8.1 (Berkeley) 6/4/93";
37#endif
38static const char rcsid[] =
39	"$Id: mcount.c,v 1.12 1997/02/22 09:39:55 peter Exp $";
40#endif
41
42#include <sys/param.h>
43#include <sys/gmon.h>
44#ifdef KERNEL
45#ifndef GUPROF
46#include <sys/systm.h>
47#endif
48#include <vm/vm.h>
49#include <vm/vm_param.h>
50#include <vm/pmap.h>
51void	bintr __P((void));
52void	btrap __P((void));
53void	eintr __P((void));
54void	user __P((void));
55#endif
56
57/*
58 * mcount is called on entry to each function compiled with the profiling
59 * switch set.  _mcount(), which is declared in a machine-dependent way
60 * with _MCOUNT_DECL, does the actual work and is either inlined into a
61 * C routine or called by an assembly stub.  In any case, this magic is
62 * taken care of by the MCOUNT definition in <machine/profile.h>.
63 *
64 * _mcount updates data structures that represent traversals of the
65 * program's call graph edges.  frompc and selfpc are the return
66 * address and function address that represents the given call graph edge.
67 *
68 * Note: the original BSD code used the same variable (frompcindex) for
69 * both frompcindex and frompc.  Any reasonable, modern compiler will
70 * perform this optimization.
71 */
72_MCOUNT_DECL(frompc, selfpc)	/* _mcount; may be static, inline, etc */
73	register uintfptr_t frompc, selfpc;
74{
75#ifdef GUPROF
76	int delta;
77#endif
78	register fptrdiff_t frompci;
79	register u_short *frompcindex;
80	register struct tostruct *top, *prevtop;
81	register struct gmonparam *p;
82	register long toindex;
83#ifdef KERNEL
84	MCOUNT_DECL(s)
85#endif
86
87	p = &_gmonparam;
88#ifndef GUPROF			/* XXX */
89	/*
90	 * check that we are profiling
91	 * and that we aren't recursively invoked.
92	 */
93	if (p->state != GMON_PROF_ON)
94		return;
95#endif
96#ifdef KERNEL
97	MCOUNT_ENTER(s);
98#else
99	p->state = GMON_PROF_BUSY;
100#endif
101	frompci = frompc - p->lowpc;
102
103#ifdef KERNEL
104	/*
105	 * When we are called from an exception handler, frompci may be
106	 * for a user address.  Convert such frompci's to the index of
107	 * user() to merge all user counts.
108	 */
109	if (frompci >= p->textsize) {
110		if (frompci + p->lowpc
111		    >= (uintfptr_t)(VM_MAXUSER_ADDRESS + UPAGES * PAGE_SIZE))
112			goto done;
113		frompci = (uintfptr_t)user - p->lowpc;
114		if (frompci >= p->textsize)
115		    goto done;
116	}
117#endif /* KERNEL */
118
119#ifdef GUPROF
120	if (p->state == GMON_PROF_HIRES) {
121		/*
122		 * Count the time since cputime() was previously called
123		 * against `frompc'.  Compensate for overheads.
124		 *
125		 * cputime() sets its prev_count variable to the count when
126		 * it is called.  This in effect starts a counter for
127		 * the next period of execution (normally from now until
128		 * the next call to mcount() or mexitcount()).  We set
129		 * cputime_bias to compensate for our own overhead.
130		 *
131		 * We use the usual sampling counters since they can be
132		 * located efficiently.  4-byte counters are usually
133		 * necessary.  gprof will add up the scattered counts
134		 * just like it does for statistical profiling.  All
135		 * counts are signed so that underflow in the subtractions
136		 * doesn't matter much (negative counts are normally
137		 * compensated for by larger counts elsewhere).  Underflow
138		 * shouldn't occur, but may be caused by slightly wrong
139		 * calibrations or from not clearing cputime_bias.
140		 */
141		delta = cputime() - cputime_bias - p->mcount_pre_overhead;
142		cputime_bias = p->mcount_post_overhead;
143		KCOUNT(p, frompci) += delta;
144		*p->cputime_count += p->cputime_overhead;
145		*p->mcount_count += p->mcount_overhead;
146	}
147#endif /* GUPROF */
148
149#ifdef KERNEL
150	/*
151	 * When we are called from an exception handler, frompc is faked
152	 * to be for where the exception occurred.  We've just solidified
153	 * the count for there.  Now convert frompci to the index of btrap()
154	 * for trap handlers and bintr() for interrupt handlers to make
155	 * exceptions appear in the call graph as calls from btrap() and
156	 * bintr() instead of calls from all over.
157	 */
158	if ((uintfptr_t)selfpc >= (uintfptr_t)btrap
159	    && (uintfptr_t)selfpc < (uintfptr_t)eintr) {
160		if ((uintfptr_t)selfpc >= (uintfptr_t)bintr)
161			frompci = (uintfptr_t)bintr - p->lowpc;
162		else
163			frompci = (uintfptr_t)btrap - p->lowpc;
164	}
165#endif /* KERNEL */
166
167	/*
168	 * check that frompc is a reasonable pc value.
169	 * for example:	signal catchers get called from the stack,
170	 *		not from text space.  too bad.
171	 */
172	if (frompci >= p->textsize)
173		goto done;
174
175	frompcindex = &p->froms[frompci / (p->hashfraction * sizeof(*p->froms))];
176	toindex = *frompcindex;
177	if (toindex == 0) {
178		/*
179		 *	first time traversing this arc
180		 */
181		toindex = ++p->tos[0].link;
182		if (toindex >= p->tolimit)
183			/* halt further profiling */
184			goto overflow;
185
186		*frompcindex = toindex;
187		top = &p->tos[toindex];
188		top->selfpc = selfpc;
189		top->count = 1;
190		top->link = 0;
191		goto done;
192	}
193	top = &p->tos[toindex];
194	if (top->selfpc == selfpc) {
195		/*
196		 * arc at front of chain; usual case.
197		 */
198		top->count++;
199		goto done;
200	}
201	/*
202	 * have to go looking down chain for it.
203	 * top points to what we are looking at,
204	 * prevtop points to previous top.
205	 * we know it is not at the head of the chain.
206	 */
207	for (; /* goto done */; ) {
208		if (top->link == 0) {
209			/*
210			 * top is end of the chain and none of the chain
211			 * had top->selfpc == selfpc.
212			 * so we allocate a new tostruct
213			 * and link it to the head of the chain.
214			 */
215			toindex = ++p->tos[0].link;
216			if (toindex >= p->tolimit)
217				goto overflow;
218
219			top = &p->tos[toindex];
220			top->selfpc = selfpc;
221			top->count = 1;
222			top->link = *frompcindex;
223			*frompcindex = toindex;
224			goto done;
225		}
226		/*
227		 * otherwise, check the next arc on the chain.
228		 */
229		prevtop = top;
230		top = &p->tos[top->link];
231		if (top->selfpc == selfpc) {
232			/*
233			 * there it is.
234			 * increment its count
235			 * move it to the head of the chain.
236			 */
237			top->count++;
238			toindex = prevtop->link;
239			prevtop->link = top->link;
240			top->link = *frompcindex;
241			*frompcindex = toindex;
242			goto done;
243		}
244
245	}
246done:
247#ifdef KERNEL
248	MCOUNT_EXIT(s);
249#else
250	p->state = GMON_PROF_ON;
251#endif
252	return;
253overflow:
254	p->state = GMON_PROF_ERROR;
255#ifdef KERNEL
256	MCOUNT_EXIT(s);
257#endif
258	return;
259}
260
261/*
262 * Actual definition of mcount function.  Defined in <machine/profile.h>,
263 * which is included by <sys/gmon.h>.
264 */
265MCOUNT
266
267#ifdef GUPROF
268void
269mexitcount(selfpc)
270	uintfptr_t selfpc;
271{
272	struct gmonparam *p;
273	uintfptr_t selfpcdiff;
274
275	p = &_gmonparam;
276	selfpcdiff = selfpc - (uintfptr_t)p->lowpc;
277	if (selfpcdiff < p->textsize) {
278		int delta;
279
280		/*
281		 * Count the time since cputime() was previously called
282		 * against `selfpc'.  Compensate for overheads.
283		 */
284		delta = cputime() - cputime_bias - p->mexitcount_pre_overhead;
285		cputime_bias = p->mexitcount_post_overhead;
286		KCOUNT(p, selfpcdiff) += delta;
287		*p->cputime_count += p->cputime_overhead;
288		*p->mexitcount_count += p->mexitcount_overhead;
289	}
290}
291
292void
293empty_loop()
294{
295	int i;
296
297	for (i = 0; i < CALIB_SCALE; i++)
298		;
299}
300
301void
302nullfunc()
303{
304}
305
306void
307nullfunc_loop()
308{
309	int i;
310
311	for (i = 0; i < CALIB_SCALE; i++)
312		nullfunc();
313}
314#endif /* GUPROF */
315