1 #define PERL_NO_GET_CONTEXT
6 /* define DBG_SUB to cause a warning on each subroutine entry. */
9 /* define DBG_TIMER to cause a warning when the timer is turned on and off. */
10 /*#define DBG_TIMER 1 */
13 #define ASSERT(x) assert(x)
19 db_get_cv(pTHX_ SV *sv)
23 if (SvIOK(sv)) { /* if (PERLDB_SUB_NN) { */
24 cv = INT2PTR(CV*,SvIVX(sv));
27 cv = get_cv(SvPVX_const(sv), TRUE);
28 } else if (SvROK(sv)) {
31 croak("DProf: don't know what subroutine to profile");
38 # define DBG_SUB_NOTIFY(A) dprof_dbg_sub_notify(aTHX_ A)
40 dprof_dbg_sub_notify(pTHX_ SV *Sub) {
41 CV * const cv = db_get_cv(aTHX_ Sub);
42 GV * const gv = cv ? CvGV(cv) : NULL;
44 warn("XS DBsub(%s::%s)\n",
45 ((GvSTASH(gv) && HvNAME_get(GvSTASH(gv))) ?
46 HvNAME_get(GvSTASH(gv)) : "(null)"),
49 warn("XS DBsub(unknown) at %x", Sub);
53 # define DBG_SUB_NOTIFY(A) /* nothing */
58 # define DBG_TIMER_NOTIFY(A) warn(A)
60 # define DBG_TIMER_NOTIFY(A) /* nothing */
63 /* HZ == clock ticks per second */
65 # define HZ ((I32)CLK_TCK)
67 # include <starlet.h> /* prototype for sys$gettim() */
68 # include <lib$routines.h>
69 # define Times(ptr) (dprof_times(aTHX_ ptr))
73 # define HZ ((I32)CLK_TCK)
78 # ifdef OS2 /* times() has significant overhead */
79 # define Times(ptr) (dprof_times(aTHX_ ptr))
80 # define INCL_DOSPROFILE
81 # define INCL_DOSERRORS
83 # define toLongLong(arg) (*(long long*)&(arg))
84 # define DPROF_HZ g_dprof_ticks
86 # define Times(ptr) (times(ptr))
91 XS(XS_Devel__DProf_END); /* used by prof_mark() */
93 /* Everything is built on times(2). See its manpage for a description
98 clock_t tms_utime; /* cpu time spent in user space */
99 clock_t tms_stime; /* cpu time spent in system */
100 clock_t realtime; /* elapsed real time, in ticks */
106 typedef union prof_any PROFANY;
110 char* out_file_name; /* output file (defaults to tmon.out) */
111 PerlIO* fp; /* pointer to tmon.out file */
112 Off_t TIMES_LOCATION; /* Where in the file to store the time totals */
113 int SAVE_STACK; /* How much data to buffer until end of run */
114 int prof_pid; /* pid of profiled process */
115 struct tms prof_start;
117 clock_t rprof_start; /* elapsed real time ticks */
128 HV* cv_hash; /* cache of CV to identifier mappings */
129 SV* key_hash; /* key for cv_hash */
138 #ifdef PERL_IMPLICIT_CONTEXT
139 PerlInterpreter *my_perl;
143 prof_state_t g_prof_state;
145 #define g_dprof_ticks g_prof_state.dprof_ticks
146 #define g_out_file_name g_prof_state.out_file_name
147 #define g_fp g_prof_state.fp
148 #define g_TIMES_LOCATION g_prof_state.TIMES_LOCATION
149 #define g_SAVE_STACK g_prof_state.SAVE_STACK
150 #define g_prof_pid g_prof_state.prof_pid
151 #define g_prof_start g_prof_state.prof_start
152 #define g_prof_end g_prof_state.prof_end
153 #define g_rprof_start g_prof_state.rprof_start
154 #define g_rprof_end g_prof_state.rprof_end
155 #define g_wprof_u g_prof_state.wprof_u
156 #define g_wprof_s g_prof_state.wprof_s
157 #define g_wprof_r g_prof_state.wprof_r
158 #define g_otms_utime g_prof_state.otms_utime
159 #define g_otms_stime g_prof_state.otms_stime
160 #define g_orealtime g_prof_state.orealtime
161 #define g_profstack g_prof_state.profstack
162 #define g_profstack_max g_prof_state.profstack_max
163 #define g_profstack_ix g_prof_state.profstack_ix
164 #define g_cv_hash g_prof_state.cv_hash
165 #define g_key_hash g_prof_state.key_hash
166 #define g_total g_prof_state.total
167 #define g_lastid g_prof_state.lastid
168 #define g_default_perldb g_prof_state.default_perldb
169 #define g_depth g_prof_state.depth
170 #ifdef PERL_IMPLICIT_CONTEXT
171 # define g_THX g_prof_state.my_perl
174 # define g_frequ g_prof_state.frequ
175 # define g_start_cnt g_prof_state.start_cnt
179 dprof_times(pTHX_ struct tms *t)
186 if (CheckOSError(DosTmrQueryFreq(&g_frequ)))
187 croak("DosTmrQueryFreq: %s", SvPV(perl_get_sv("!",TRUE),n_a));
189 g_frequ = g_frequ/DPROF_HZ; /* count per tick */
190 if (CheckOSError(DosTmrQueryTime(&cnt)))
191 croak("DosTmrQueryTime: %s",
192 SvPV_nolen_const(perl_get_sv("!",TRUE)));
193 g_start_cnt = toLongLong(cnt);
196 if (CheckOSError(DosTmrQueryTime(&cnt)))
197 croak("DosTmrQueryTime: %s", SvPV(perl_get_sv("!",TRUE), n_a));
199 return (t->tms_utime = (toLongLong(cnt) - g_start_cnt)/g_frequ);
203 /* Get wall time and convert to 10 ms intervals to
204 * produce the return value dprof expects */
205 # if defined(__DECC) && defined (__ALPHA)
208 _ckvmssts(sys$gettim(&vmstime));
210 retval = vmstime & 0x7fffffff;
212 /* (Older hw or ccs don't have an atomic 64-bit type, so we
213 * juggle 32-bit ints (and a float) to produce a time_t result
214 * with minimal loss of information.) */
215 long int vmstime[2],remainder,divisor = 100000;
216 _ckvmssts(sys$gettim((unsigned long int *)vmstime));
217 vmstime[1] &= 0x7fff; /* prevent overflow in EDIV */
218 _ckvmssts(lib$ediv(&divisor,vmstime,(long int *)&retval,&remainder));
220 /* Fill in the struct tms using the CRTL routine . . .*/
221 times((tbuffer_t *)t);
222 return (clock_t) retval;
223 # else /* !VMS && !OS2 */
230 prof_dumpa(pTHX_ opcode ptype, U32 id)
232 if (ptype == OP_LEAVESUB) {
233 PerlIO_printf(g_fp,"- %"UVxf"\n", (UV)id);
235 else if(ptype == OP_ENTERSUB) {
236 PerlIO_printf(g_fp,"+ %"UVxf"\n", (UV)id);
238 else if(ptype == OP_GOTO) {
239 PerlIO_printf(g_fp,"* %"UVxf"\n", (UV)id);
241 else if(ptype == OP_DIE) {
242 PerlIO_printf(g_fp,"/ %"UVxf"\n", (UV)id);
245 PerlIO_printf(g_fp,"Profiler unknown prof code %d\n", ptype);
250 prof_dumps(pTHX_ U32 id, const char *pname, const char *gname)
252 PerlIO_printf(g_fp,"& %"UVxf" %s %s\n", (UV)id, pname, gname);
256 prof_dumpt(pTHX_ long tms_utime, long tms_stime, long realtime)
258 PerlIO_printf(g_fp,"@ %ld %ld %ld\n", tms_utime, tms_stime, realtime);
262 prof_dump_until(pTHX_ long ix)
268 const clock_t realtime1 = Times(&t1);
271 const opcode ptype = g_profstack[base++].ptype;
272 if (ptype == OP_TIME) {
273 const long tms_utime = g_profstack[base++].tms_utime;
274 const long tms_stime = g_profstack[base++].tms_stime;
275 const long realtime = g_profstack[base++].realtime;
277 prof_dumpt(aTHX_ tms_utime, tms_stime, realtime);
279 else if (ptype == OP_GV) {
280 const U32 id = g_profstack[base++].id;
281 const char * const pname = g_profstack[base++].name;
282 const char * const gname = g_profstack[base++].name;
284 prof_dumps(aTHX_ id, pname, gname);
287 const U32 id = g_profstack[base++].id;
288 prof_dumpa(aTHX_ ptype, id);
292 realtime2 = Times(&t2);
293 if (realtime2 != realtime1 || t1.tms_utime != t2.tms_utime
294 || t1.tms_stime != t2.tms_stime) {
295 g_wprof_r += realtime2 - realtime1;
296 g_wprof_u += t2.tms_utime - t1.tms_utime;
297 g_wprof_s += t2.tms_stime - t1.tms_stime;
299 PerlIO_printf(g_fp,"+ & Devel::DProf::write\n");
300 PerlIO_printf(g_fp,"@ %"IVdf" %"IVdf" %"IVdf"\n",
301 /* The (IV) casts are one possibility:
302 * the Painfully Correct Way would be to
304 (IV)(t2.tms_utime - t1.tms_utime),
305 (IV)(t2.tms_stime - t1.tms_stime),
306 (IV)(realtime2 - realtime1));
307 PerlIO_printf(g_fp,"- & Devel::DProf::write\n");
308 g_otms_utime = t2.tms_utime;
309 g_otms_stime = t2.tms_stime;
310 g_orealtime = realtime2;
316 set_cv_key(pTHX_ CV *cv, const char *pname, const char *gname)
318 SvGROW(g_key_hash, sizeof(CV**) + strlen(pname) + strlen(gname) + 3);
319 sv_setpvn(g_key_hash, (char*)&cv, sizeof(CV**));
320 sv_catpv(g_key_hash, pname);
321 sv_catpv(g_key_hash, "::");
322 sv_catpv(g_key_hash, gname);
326 prof_mark(pTHX_ opcode ptype)
329 clock_t realtime, rdelta, udelta, sdelta;
331 SV * const Sub = GvSV(PL_DBsub); /* name of current sub */
334 if (g_profstack_ix + 10 > g_profstack_max) {
335 g_profstack_max = g_profstack_max * 3 / 2;
336 Renew(g_profstack, g_profstack_max, PROFANY);
340 realtime = Times(&t);
341 rdelta = realtime - g_orealtime;
342 udelta = t.tms_utime - g_otms_utime;
343 sdelta = t.tms_stime - g_otms_stime;
344 if (rdelta || udelta || sdelta) {
346 ASSERT(g_profstack_ix + 4 <= g_profstack_max);
347 g_profstack[g_profstack_ix++].ptype = OP_TIME;
348 g_profstack[g_profstack_ix++].tms_utime = udelta;
349 g_profstack[g_profstack_ix++].tms_stime = sdelta;
350 g_profstack[g_profstack_ix++].realtime = rdelta;
352 else { /* Write it to disk now so's not to eat up core */
353 if (g_prof_pid == (int)getpid()) {
354 prof_dumpt(aTHX_ udelta, sdelta, rdelta);
358 g_orealtime = realtime;
359 g_otms_stime = t.tms_stime;
360 g_otms_utime = t.tms_utime;
367 CV * const cv = db_get_cv(aTHX_ Sub);
368 GV * const gv = CvGV(cv);
369 pname = GvSTASH(gv) ? HvNAME_get(GvSTASH(gv)) : 0;
370 pname = pname ? pname : (char *) "(null)";
373 set_cv_key(aTHX_ cv, pname, gname);
374 svp = hv_fetch(g_cv_hash, SvPVX_const(g_key_hash), SvCUR(g_key_hash), TRUE);
376 sv_setiv(*svp, id = ++g_lastid);
377 if (CvXSUB(cv) == XS_Devel__DProf_END)
379 if (g_SAVE_STACK) { /* Store it for later recording -JH */
380 ASSERT(g_profstack_ix + 4 <= g_profstack_max);
381 g_profstack[g_profstack_ix++].ptype = OP_GV;
382 g_profstack[g_profstack_ix++].id = id;
383 g_profstack[g_profstack_ix++].name = pname;
384 g_profstack[g_profstack_ix++].name = gname;
386 else { /* Write it to disk now so's not to eat up core */
387 /* Only record the parent's info */
388 if (g_prof_pid == (int)getpid()) {
389 prof_dumps(aTHX_ id, pname, gname);
393 PL_perldb = 0; /* Do not debug the kid. */
402 if (g_SAVE_STACK) { /* Store it for later recording -JH */
403 ASSERT(g_profstack_ix + 2 <= g_profstack_max);
404 g_profstack[g_profstack_ix++].ptype = ptype;
405 g_profstack[g_profstack_ix++].id = id;
407 /* Only record the parent's info */
408 if (g_SAVE_STACK < g_profstack_ix) {
409 if (g_prof_pid == (int)getpid())
410 prof_dump_until(aTHX_ g_profstack_ix);
412 PL_perldb = 0; /* Do not debug the kid. */
416 else { /* Write it to disk now so's not to eat up core */
418 /* Only record the parent's info */
419 if (g_prof_pid == (int)getpid()) {
420 prof_dumpa(aTHX_ ptype, id);
424 PL_perldb = 0; /* Do not debug the kid. */
429 # define defstash PL_defstash
432 /* Counts overhead of prof_mark and extra XS call. */
434 test_time(pTHX_ clock_t *r, clock_t *u, clock_t *s)
436 CV * const cv = perl_get_cv("Devel::DProf::NONESUCH_noxs", FALSE);
437 HV * const oldstash = PL_curstash;
439 const U32 ototal = g_total;
440 const U32 ostack = g_SAVE_STACK;
441 const U32 operldb = PL_perldb;
444 clock_t realtime1 = Times(&t1);
445 clock_t realtime2 = 0;
447 g_SAVE_STACK = 1000000;
451 /* Disable debugging of perl_call_sv on second pass: */
452 PL_curstash = (k == 0 ? PL_defstash : PL_debstash);
453 PL_perldb = g_default_perldb;
456 g_profstack_ix = 0; /* Do not let the stack grow */
458 /* prof_mark(aTHX_ OP_ENTERSUB); */
460 PUSHMARK(PL_stack_sp);
461 perl_call_sv((SV*)cv, G_SCALAR);
463 /* prof_mark(aTHX_ OP_LEAVESUB); */
466 PL_curstash = oldstash;
467 if (k == 0) { /* Put time with debugging */
468 realtime2 = Times(&t2);
469 *r = realtime2 - realtime1;
470 *u = t2.tms_utime - t1.tms_utime;
471 *s = t2.tms_stime - t1.tms_stime;
473 else { /* Subtract time without debug */
474 realtime1 = Times(&t1);
475 *r -= realtime1 - realtime2;
476 *u -= t1.tms_utime - t2.tms_utime;
477 *s -= t1.tms_stime - t2.tms_stime;
482 g_SAVE_STACK = ostack;
487 prof_recordheader(pTHX)
491 /* g_fp is opened in the BOOT section */
492 PerlIO_printf(g_fp, "#fOrTyTwO\n");
493 PerlIO_printf(g_fp, "$hz=%"IVdf";\n", (IV)DPROF_HZ);
494 PerlIO_printf(g_fp, "$XS_VERSION='DProf %s';\n", XS_VERSION);
495 PerlIO_printf(g_fp, "# All values are given in HZ\n");
496 test_time(aTHX_ &r, &u, &s);
498 "$over_utime=%"IVdf"; $over_stime=%"IVdf"; $over_rtime=%"IVdf";\n",
499 /* The (IV) casts are one possibility:
500 * the Painfully Correct Way would be to
502 (IV)u, (IV)s, (IV)r);
503 PerlIO_printf(g_fp, "$over_tests=10000;\n");
505 g_TIMES_LOCATION = PerlIO_tell(g_fp);
507 /* Pad with whitespace. */
508 /* This should be enough even for very large numbers. */
509 PerlIO_printf(g_fp, "%*s\n", 240 , "");
511 PerlIO_printf(g_fp, "\n");
512 PerlIO_printf(g_fp, "PART2\n");
520 /* g_fp is opened in the BOOT section */
522 /* Now that we know the runtimes, fill them in at the recorded
526 prof_dump_until(aTHX_ g_profstack_ix);
528 PerlIO_seek(g_fp, g_TIMES_LOCATION, SEEK_SET);
529 /* Write into reserved 240 bytes: */
531 "$rrun_utime=%"IVdf"; $rrun_stime=%"IVdf"; $rrun_rtime=%"IVdf";",
532 /* The (IV) casts are one possibility:
533 * the Painfully Correct Way would be to
535 (IV)(g_prof_end.tms_utime-g_prof_start.tms_utime-g_wprof_u),
536 (IV)(g_prof_end.tms_stime-g_prof_start.tms_stime-g_wprof_s),
537 (IV)(g_rprof_end-g_rprof_start-g_wprof_r));
538 PerlIO_printf(g_fp, "\n$total_marks=%"IVdf, (IV)g_total);
546 check_depth(pTHX_ void *foo)
548 const U32 need_depth = PTR2UV(foo);
549 if (need_depth != g_depth) {
550 if (need_depth > g_depth) {
551 warn("garbled call depth when profiling");
554 IV marks = g_depth - need_depth;
556 /* warn("Check_depth: got %d, expected %d\n", g_depth, need_depth); */
558 prof_mark(aTHX_ OP_DIE);
560 g_depth = need_depth;
573 SV * const Sub = GvSV(PL_DBsub); /* name of current sub */
575 #ifdef PERL_IMPLICIT_CONTEXT
576 /* profile only the interpreter that loaded us */
579 perl_call_sv((SV*)db_get_cv(aTHX_ Sub), GIMME_V | G_NODEBUG);
584 HV * const oldstash = PL_curstash;
585 const I32 old_scopestack_ix = PL_scopestack_ix;
586 const I32 old_cxstack_ix = cxstack_ix;
590 SAVEDESTRUCTOR_X(check_depth, INT2PTR(void*,g_depth));
593 prof_mark(aTHX_ OP_ENTERSUB);
595 perl_call_sv((SV*)db_get_cv(aTHX_ Sub), GIMME_V | G_NODEBUG);
596 PL_curstash = oldstash;
598 /* Make sure we are on the same context and scope as before the call
599 * to the sub. If the called sub was exited via a goto, next or
600 * last then this will try to croak(), however perl may still crash
601 * with a segfault. */
602 if (PL_scopestack_ix != old_scopestack_ix || cxstack_ix != old_cxstack_ix)
603 croak("panic: Devel::DProf inconsistent subroutine return");
605 prof_mark(aTHX_ OP_LEAVESUB);
614 #ifdef PERL_IMPLICIT_CONTEXT
618 prof_mark(aTHX_ OP_GOTO);
623 #endif /* for_real */
627 MODULE = Devel::DProf PACKAGE = DB
634 HV * const oldstash = PL_curstash;
635 SV * const Sub = GvSV(PL_DBsub); /* name of current sub */
636 /* SP -= items; added by xsubpp */
639 sv_setiv(PL_DBsingle, 0); /* disable DB single-stepping */
641 prof_mark(aTHX_ OP_ENTERSUB);
644 PL_curstash = PL_debstash; /* To disable debugging of perl_call_sv */
645 perl_call_sv(Sub, GIMME_V);
646 PL_curstash = oldstash;
648 prof_mark(aTHX_ OP_LEAVESUB);
650 /* PUTBACK; added by xsubpp */
655 MODULE = Devel::DProf PACKAGE = Devel::DProf
662 /* maybe the process forked--we want only
663 * the parent's profile.
666 #ifdef PERL_IMPLICIT_CONTEXT
669 g_prof_pid == (int)getpid())
671 g_rprof_end = Times(&g_prof_end);
672 DBG_TIMER_NOTIFY("Profiler timer is off.\n");
683 g_TIMES_LOCATION = 42;
684 g_SAVE_STACK = 1<<14;
685 g_profstack_max = 128;
686 #ifdef PERL_IMPLICIT_CONTEXT
690 /* Before we go anywhere make sure we were invoked
691 * properly, else we'll dump core.
694 croak("DProf: run perl with -d to use DProf.\n");
696 /* When we hook up the XS DB::sub we'll be redefining
697 * the DB::sub from the PM file. Turn off warnings
701 const bool warn_tmp = PL_dowarn;
703 newXS("DB::sub", XS_DB_sub, file);
704 newXS("DB::goto", XS_DB_goto, file);
705 PL_dowarn = warn_tmp;
708 sv_setiv(PL_DBsingle, 0); /* disable DB single-stepping */
711 const char *buffer = getenv("PERL_DPROF_BUFFER");
714 g_SAVE_STACK = atoi(buffer);
717 buffer = getenv("PERL_DPROF_TICKS");
720 g_dprof_ticks = atoi(buffer); /* Used under OS/2 only */
726 buffer = getenv("PERL_DPROF_OUT_FILE_NAME");
727 g_out_file_name = savepv(buffer ? buffer : "tmon.out");
730 if ((g_fp = PerlIO_open(g_out_file_name, "w")) == NULL)
731 croak("DProf: unable to write '%s', errno = %d\n",
732 g_out_file_name, errno);
734 g_default_perldb = PERLDBf_NONAME | PERLDBf_SUB | PERLDBf_GOTO;
736 g_key_hash = newSV(256);
737 g_prof_pid = (int)getpid();
739 Newx(g_profstack, g_profstack_max, PROFANY);
740 prof_recordheader(aTHX);
741 DBG_TIMER_NOTIFY("Profiler timer is on.\n");
742 g_orealtime = g_rprof_start = Times(&g_prof_start);
743 g_otms_utime = g_prof_start.tms_utime;
744 g_otms_stime = g_prof_start.tms_stime;
745 PL_perldb = g_default_perldb;