gmon-sol2.c revision 132718
189051Sjake/*-
289051Sjake * Copyright (c) 1991 The Regents of the University of California.
389051Sjake * All rights reserved.
489051Sjake *
589051Sjake * Redistribution and use in source and binary forms, with or without
689051Sjake * modification, are permitted provided that the following conditions
789051Sjake * are met:
889051Sjake * 1. Redistributions of source code must retain the above copyright
989051Sjake *    notice, this list of conditions and the following disclaimer.
1089051Sjake * 2. Redistributions in binary form must reproduce the above copyright
1189051Sjake *    notice, this list of conditions and the following disclaimer in the
1289051Sjake *    documentation and/or other materials provided with the distribution.
1389051Sjake * 3. [rescinded 22 July 1999]
1489051Sjake * 4. Neither the name of the University nor the names of its contributors
1589051Sjake *    may be used to endorse or promote products derived from this software
1689051Sjake *    without specific prior written permission.
1789051Sjake *
1889051Sjake * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
1989051Sjake * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
2089051Sjake * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
2189051Sjake * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
2289051Sjake * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
2389051Sjake * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
2489051Sjake * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
2589051Sjake * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
2689051Sjake * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
2789051Sjake * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
2889051Sjake * SUCH DAMAGE.
2989051Sjake */
3089051Sjake
3189051Sjake/* Mangled into a form that works on SPARC Solaris 2 by Mark Eichin
3289051Sjake * for Cygnus Support, July 1992.
3389051Sjake */
3489051Sjake
3589051Sjake#include "tconfig.h"
3689051Sjake#include "tsystem.h"
3789051Sjake#include <fcntl.h> /* for creat() */
3889051Sjake#include "coretypes.h"
3989051Sjake#include "tm.h"
4089051Sjake
4189051Sjake#if 0
4289051Sjake#include "sparc/gmon.h"
4389051Sjake#else
4489051Sjakestruct phdr {
4589051Sjake  char *lpc;
4689051Sjake  char *hpc;
4789051Sjake  int ncnt;
4889051Sjake};
4989051Sjake#define HISTFRACTION 2
5089051Sjake#define HISTCOUNTER unsigned short
5189051Sjake#define HASHFRACTION 1
5289051Sjake#define ARCDENSITY 2
5389051Sjake#define MINARCS 50
5489051Sjakestruct tostruct {
5589051Sjake  char *selfpc;
5689051Sjake  long count;
5789051Sjake  unsigned short link;
5889051Sjake};
5989051Sjakestruct rawarc {
6089051Sjake    unsigned long       raw_frompc;
6189051Sjake    unsigned long       raw_selfpc;
6289051Sjake    long                raw_count;
6389051Sjake};
6489051Sjake#define ROUNDDOWN(x,y)  (((x)/(y))*(y))
6589051Sjake#define ROUNDUP(x,y)    ((((x)+(y)-1)/(y))*(y))
6689051Sjake
6789051Sjake#endif
6889051Sjake
6989051Sjake/* extern mcount() asm ("mcount"); */
7089051Sjake/*extern*/ char *minbrk /* asm ("minbrk") */;
7189051Sjake
7289051Sjake    /*
7389051Sjake     *	froms is actually a bunch of unsigned shorts indexing tos
7489051Sjake     */
7589051Sjakestatic int		profiling = 3;
7689051Sjakestatic unsigned short	*froms;
7789051Sjakestatic struct tostruct	*tos = 0;
7889051Sjakestatic long		tolimit = 0;
7989051Sjakestatic char		*s_lowpc = 0;
8091617Sjakestatic char		*s_highpc = 0;
8189051Sjakestatic unsigned long	s_textsize = 0;
8289051Sjake
8389051Sjakestatic int	ssiz;
8489051Sjakestatic char	*sbuf;
8589051Sjakestatic int	s_scale;
8689051Sjake    /* see profil(2) where this is describe (incorrectly) */
8789051Sjake#define		SCALE_1_TO_1	0x10000L
8889051Sjake
8989051Sjake#define	MSG "No space for profiling buffer(s)\n"
9089051Sjake
9189051Sjakestatic void moncontrol (int);
9289051Sjakeextern void monstartup (char *, char *);
9389051Sjakeextern void _mcleanup (void);
9491617Sjake
9591617Sjakevoid monstartup(char *lowpc, char *highpc)
9689051Sjake{
9789051Sjake    int			monsize;
9889051Sjake    char		*buffer;
9989051Sjake    register int	o;
10091617Sjake
10191617Sjake	/*
10291617Sjake	 *	round lowpc and highpc to multiples of the density we're using
10391617Sjake	 *	so the rest of the scaling (here and in gprof) stays in ints.
10491617Sjake	 */
10591617Sjake    lowpc = (char *)
10691617Sjake	    ROUNDDOWN((unsigned long)lowpc, HISTFRACTION*sizeof(HISTCOUNTER));
10791617Sjake    s_lowpc = lowpc;
10891617Sjake    highpc = (char *)
10991617Sjake	    ROUNDUP((unsigned long)highpc, HISTFRACTION*sizeof(HISTCOUNTER));
11091617Sjake    s_highpc = highpc;
11191617Sjake    s_textsize = highpc - lowpc;
11291617Sjake    monsize = (s_textsize / HISTFRACTION) + sizeof(struct phdr);
11391617Sjake    buffer = sbrk( monsize );
11491617Sjake    if ( buffer == (char *) -1 ) {
11591617Sjake	write( 2 , MSG , sizeof(MSG) );
11691617Sjake	return;
11791617Sjake    }
11891617Sjake    froms = (unsigned short *) sbrk( s_textsize / HASHFRACTION );
11991617Sjake    if ( froms == (unsigned short *) -1 ) {
12091617Sjake	write( 2 , MSG , sizeof(MSG) );
12191617Sjake	froms = 0;
12291617Sjake	return;
12391617Sjake    }
12489051Sjake    tolimit = s_textsize * ARCDENSITY / 100;
12589051Sjake    if ( tolimit < MINARCS ) {
12689051Sjake	tolimit = MINARCS;
12789051Sjake    } else if ( tolimit > 65534 ) {
12889051Sjake	tolimit = 65534;
12989051Sjake    }
13089051Sjake    tos = (struct tostruct *) sbrk( tolimit * sizeof( struct tostruct ) );
13189051Sjake    if ( tos == (struct tostruct *) -1 ) {
13289051Sjake	write( 2 , MSG , sizeof(MSG) );
13389051Sjake	froms = 0;
13489051Sjake	tos = 0;
13589051Sjake	return;
13689051Sjake    }
13789051Sjake    minbrk = sbrk(0);
13889051Sjake    tos[0].link = 0;
13989051Sjake    sbuf = buffer;
14089051Sjake    ssiz = monsize;
14189051Sjake    ( (struct phdr *) buffer ) -> lpc = lowpc;
14289051Sjake    ( (struct phdr *) buffer ) -> hpc = highpc;
14389051Sjake    ( (struct phdr *) buffer ) -> ncnt = ssiz;
14489051Sjake    monsize -= sizeof(struct phdr);
14589051Sjake    if ( monsize <= 0 )
14689051Sjake	return;
14789051Sjake    o = highpc - lowpc;
14889051Sjake    if( monsize < o )
14991617Sjake#ifndef hp300
15091617Sjake	s_scale = ( (float) monsize / o ) * SCALE_1_TO_1;
15191617Sjake#else /* avoid floating point */
15291617Sjake    {
15391617Sjake	int quot = o / monsize;
15491617Sjake
15591617Sjake	if (quot >= 0x10000)
15691617Sjake		s_scale = 1;
15791617Sjake	else if (quot >= 0x100)
15891617Sjake		s_scale = 0x10000 / quot;
15991617Sjake	else if (o >= 0x800000)
16091617Sjake		s_scale = 0x1000000 / (o / (monsize >> 8));
16191617Sjake	else
16291617Sjake		s_scale = 0x1000000 / ((o << 8) / monsize);
16391617Sjake    }
16491617Sjake#endif
16591617Sjake    else
16691617Sjake	s_scale = SCALE_1_TO_1;
16791617Sjake    moncontrol(1);
16891617Sjake}
16991617Sjake
17091617Sjakevoid
17191617Sjake_mcleanup(void)
17291617Sjake{
17391617Sjake    int			fd;
17489051Sjake    int			fromindex;
17589051Sjake    int			endfrom;
17689051Sjake    char		*frompc;
17789051Sjake    int			toindex;
17889051Sjake    struct rawarc	rawarc;
17989051Sjake    char		*profdir;
18089051Sjake    const char		*proffile;
18189051Sjake    char		*progname;
18289051Sjake    char		 buf[PATH_MAX];
18389051Sjake    extern char	       **___Argv;
18489051Sjake
18589051Sjake    moncontrol(0);
18691617Sjake
18791617Sjake    if ((profdir = getenv("PROFDIR")) != NULL) {
18889051Sjake	/* If PROFDIR contains a null value, no profiling output is produced */
18991617Sjake	if (*profdir == '\0') {
19089051Sjake	    return;
19189051Sjake	}
19289051Sjake
19389051Sjake	progname=strrchr(___Argv[0], '/');
19489051Sjake	if (progname == NULL)
19589051Sjake	    progname=___Argv[0];
19689051Sjake	else
19789051Sjake	    progname++;
19889051Sjake
19989051Sjake	sprintf(buf, "%s/%ld.%s", profdir, (long) getpid(), progname);
20089051Sjake	proffile = buf;
20189051Sjake    } else {
20289051Sjake	proffile = "gmon.out";
20389051Sjake    }
20489051Sjake
20589051Sjake    fd = creat( proffile, 0666 );
20689051Sjake    if ( fd < 0 ) {
20789051Sjake	perror( proffile );
20891617Sjake	return;
20991617Sjake    }
21091617Sjake#   ifdef DEBUG
21189051Sjake	fprintf( stderr , "[mcleanup] sbuf 0x%x ssiz %d\n" , sbuf , ssiz );
21291617Sjake#   endif /* DEBUG */
21391617Sjake    write( fd , sbuf , ssiz );
21491617Sjake    endfrom = s_textsize / (HASHFRACTION * sizeof(*froms));
21591617Sjake    for ( fromindex = 0 ; fromindex < endfrom ; fromindex++ ) {
21691617Sjake	if ( froms[fromindex] == 0 ) {
21791617Sjake	    continue;
21891617Sjake	}
21991617Sjake	frompc = s_lowpc + (fromindex * HASHFRACTION * sizeof(*froms));
22091617Sjake	for (toindex=froms[fromindex]; toindex!=0; toindex=tos[toindex].link) {
22191617Sjake#	    ifdef DEBUG
22291617Sjake		fprintf( stderr ,
22391617Sjake			"[mcleanup] frompc 0x%x selfpc 0x%x count %d\n" ,
22489051Sjake			frompc , tos[toindex].selfpc , tos[toindex].count );
22591617Sjake#	    endif /* DEBUG */
22691617Sjake	    rawarc.raw_frompc = (unsigned long) frompc;
22791617Sjake	    rawarc.raw_selfpc = (unsigned long) tos[toindex].selfpc;
22891617Sjake	    rawarc.raw_count = tos[toindex].count;
22989051Sjake	    write( fd , &rawarc , sizeof rawarc );
23091617Sjake	}
23189051Sjake    }
23289051Sjake    close( fd );
23389051Sjake}
23489051Sjake
23589051Sjake/*
23689051Sjake * The SPARC stack frame is only held together by the frame pointers
23789051Sjake * in the register windows. According to the SVR4 SPARC ABI
23889051Sjake * Supplement, Low Level System Information/Operating System
23989051Sjake * Interface/Software Trap Types, a type 3 trap will flush all of the
24089051Sjake * register windows to the stack, which will make it possible to walk
24189051Sjake * the frames and find the return addresses.
24289051Sjake * 	However, it seems awfully expensive to incur a trap (system
24389051Sjake * call) for every function call. It turns out that "call" simply puts
24491617Sjake * the return address in %o7 expecting the "save" in the procedure to
24589051Sjake * shift it into %i7; this means that before the "save" occurs, %o7
24691617Sjake * contains the address of the call to mcount, and %i7 still contains
24791617Sjake * the caller above that. The asm mcount here simply saves those
24891617Sjake * registers in argument registers and branches to internal_mcount,
24991617Sjake * simulating a call with arguments.
25091617Sjake * 	Kludges:
25191617Sjake * 	1) the branch to internal_mcount is hard coded; it should be
25291617Sjake * possible to tell asm to use the assembler-name of a symbol.
25391617Sjake * 	2) in theory, the function calling mcount could have saved %i7
25489051Sjake * somewhere and reused the register; in practice, I *think* this will
25591617Sjake * break longjmp (and maybe the debugger) but I'm not certain. (I take
25691617Sjake * some comfort in the knowledge that it will break the native mcount
25789051Sjake * as well.)
25891617Sjake * 	3) if builtin_return_address worked, this could be portable.
25991617Sjake * However, it would really have to be optimized for arguments of 0
26091617Sjake * and 1 and do something like what we have here in order to avoid the
26191617Sjake * trap per function call performance hit.
26291617Sjake * 	4) the atexit and monsetup calls prevent this from simply
26389051Sjake * being a leaf routine that doesn't do a "save" (and would thus have
26491617Sjake * access to %o7 and %i7 directly) but the call to write() at the end
26591617Sjake * would have also prevented this.
26691617Sjake *
26791617Sjake * -- [eichin:19920702.1107EST]
26891617Sjake */
26991617Sjake
27091617Sjakestatic void internal_mcount (char *, unsigned short *)
27191617Sjake     __attribute__ ((__unused__));
27291617Sjake
27391617Sjake/* i7 == last ret, -> frompcindex */
27491617Sjake/* o7 == current ret, -> selfpc */
27591617Sjake/* Solaris 2 libraries use _mcount.  */
27691617Sjakeasm(".global _mcount; _mcount: mov %i7,%o1; mov %o7,%o0;b,a internal_mcount");
27791617Sjake/* This is for compatibility with old versions of gcc which used mcount.  */
27891617Sjakeasm(".global mcount; mcount: mov %i7,%o1; mov %o7,%o0;b,a internal_mcount");
27991617Sjake
28091617Sjakestatic void internal_mcount(char *selfpc, unsigned short *frompcindex)
28191617Sjake{
28291617Sjake	register struct tostruct	*top;
28391617Sjake	register struct tostruct	*prevtop;
28491617Sjake	register long			toindex;
28591617Sjake	static char already_setup;
28691617Sjake
28791617Sjake	/*
28891617Sjake	 *	find the return address for mcount,
28989051Sjake	 *	and the return address for mcount's caller.
29091617Sjake	 */
29191617Sjake
29291617Sjake	if(!already_setup) {
29391617Sjake          extern char etext[];
29489051Sjake	  extern char _start[];
29591617Sjake	  extern char _init[];
29691617Sjake	  already_setup = 1;
29791617Sjake	  monstartup(_start < _init ? _start : _init, etext);
29889051Sjake#ifdef USE_ONEXIT
29989051Sjake	  on_exit(_mcleanup, 0);
30089051Sjake#else
30189051Sjake	  atexit(_mcleanup);
30289051Sjake#endif
30391617Sjake	}
30491617Sjake	/*
30591617Sjake	 *	check that we are profiling
30689051Sjake	 *	and that we aren't recursively invoked.
30791066Sphk	 */
30889051Sjake	if (profiling) {
30989051Sjake		goto out;
31089051Sjake	}
31189051Sjake	profiling++;
31289051Sjake	/*
31389051Sjake	 *	check that frompcindex is a reasonable pc value.
31489051Sjake	 *	for example:	signal catchers get called from the stack,
31589051Sjake	 *			not from text space.  too bad.
31689051Sjake	 */
31789051Sjake	frompcindex = (unsigned short *)((long)frompcindex - (long)s_lowpc);
31889051Sjake	if ((unsigned long)frompcindex > s_textsize) {
31989051Sjake		goto done;
32089051Sjake	}
32189051Sjake	frompcindex =
32289051Sjake	    &froms[((long)frompcindex) / (HASHFRACTION * sizeof(*froms))];
32389051Sjake	toindex = *frompcindex;
32489051Sjake	if (toindex == 0) {
32589051Sjake		/*
32689051Sjake		 *	first time traversing this arc
32789051Sjake		 */
32889051Sjake		toindex = ++tos[0].link;
32989051Sjake		if (toindex >= tolimit) {
33089051Sjake			goto overflow;
33189051Sjake		}
33289051Sjake		*frompcindex = toindex;
33389051Sjake		top = &tos[toindex];
33489051Sjake		top->selfpc = selfpc;
33589051Sjake		top->count = 1;
33689051Sjake		top->link = 0;
33789051Sjake		goto done;
33889051Sjake	}
33989051Sjake	top = &tos[toindex];
34089051Sjake	if (top->selfpc == selfpc) {
34189051Sjake		/*
34289051Sjake		 *	arc at front of chain; usual case.
34389051Sjake		 */
34489051Sjake		top->count++;
34589051Sjake		goto done;
34689051Sjake	}
34789051Sjake	/*
34889051Sjake	 *	have to go looking down chain for it.
34989051Sjake	 *	top points to what we are looking at,
35089051Sjake	 *	prevtop points to previous top.
35189051Sjake	 *	we know it is not at the head of the chain.
35289051Sjake	 */
35389051Sjake	for (; /* goto done */; ) {
35489051Sjake		if (top->link == 0) {
35589051Sjake			/*
35689051Sjake			 *	top is end of the chain and none of the chain
35789051Sjake			 *	had top->selfpc == selfpc.
35889051Sjake			 *	so we allocate a new tostruct
35989051Sjake			 *	and link it to the head of the chain.
36089051Sjake			 */
36189051Sjake			toindex = ++tos[0].link;
36289051Sjake			if (toindex >= tolimit) {
36389051Sjake				goto overflow;
36489051Sjake			}
36589051Sjake			top = &tos[toindex];
36689051Sjake			top->selfpc = selfpc;
36789051Sjake			top->count = 1;
36889051Sjake			top->link = *frompcindex;
36989051Sjake			*frompcindex = toindex;
37089051Sjake			goto done;
37189051Sjake		}
37289051Sjake		/*
37389051Sjake		 *	otherwise, check the next arc on the chain.
37489051Sjake		 */
37589051Sjake		prevtop = top;
37689051Sjake		top = &tos[top->link];
37789051Sjake		if (top->selfpc == selfpc) {
37889051Sjake			/*
37989051Sjake			 *	there it is.
38089051Sjake			 *	increment its count
38189051Sjake			 *	move it to the head of the chain.
38289051Sjake			 */
383			top->count++;
384			toindex = prevtop->link;
385			prevtop->link = top->link;
386			top->link = *frompcindex;
387			*frompcindex = toindex;
388			goto done;
389		}
390
391	}
392done:
393	profiling--;
394	/* and fall through */
395out:
396	return;		/* normal return restores saved registers */
397
398overflow:
399	profiling++; /* halt further profiling */
400#   define	TOLIMIT	"mcount: tos overflow\n"
401	write(2, TOLIMIT, sizeof(TOLIMIT));
402	goto out;
403}
404
405/*
406 * Control profiling
407 *	profiling is what mcount checks to see if
408 *	all the data structures are ready.
409 */
410static void moncontrol(int mode)
411{
412    if (mode) {
413	/* start */
414	profil((unsigned short *)(sbuf + sizeof(struct phdr)),
415	       ssiz - sizeof(struct phdr),
416	       (long)s_lowpc, s_scale);
417	profiling = 0;
418    } else {
419	/* stop */
420	profil((unsigned short *)0, 0, 0, 0);
421	profiling = 3;
422    }
423}
424