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 *cv = db_get_cv(aTHX_ Sub);
42 GV *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)
187 if (CheckOSError(DosTmrQueryFreq(&g_frequ)))
188 croak("DosTmrQueryFreq: %s", SvPV(perl_get_sv("!",TRUE),n_a));
190 g_frequ = g_frequ/DPROF_HZ; /* count per tick */
191 if (CheckOSError(DosTmrQueryTime(&cnt)))
192 croak("DosTmrQueryTime: %s",
193 SvPV(perl_get_sv("!",TRUE), n_a));
194 g_start_cnt = toLongLong(cnt);
197 if (CheckOSError(DosTmrQueryTime(&cnt)))
198 croak("DosTmrQueryTime: %s", SvPV(perl_get_sv("!",TRUE), n_a));
200 return (t->tms_utime = (toLongLong(cnt) - g_start_cnt)/g_frequ);
204 /* Get wall time and convert to 10 ms intervals to
205 * produce the return value dprof expects */
206 # if defined(__DECC) && defined (__ALPHA)
209 _ckvmssts(sys$gettim(&vmstime));
211 retval = vmstime & 0x7fffffff;
213 /* (Older hw or ccs don't have an atomic 64-bit type, so we
214 * juggle 32-bit ints (and a float) to produce a time_t result
215 * with minimal loss of information.) */
216 long int vmstime[2],remainder,divisor = 100000;
217 _ckvmssts(sys$gettim((unsigned long int *)vmstime));
218 vmstime[1] &= 0x7fff; /* prevent overflow in EDIV */
219 _ckvmssts(lib$ediv(&divisor,vmstime,(long int *)&retval,&remainder));
221 /* Fill in the struct tms using the CRTL routine . . .*/
222 times((tbuffer_t *)t);
223 return (clock_t) retval;
224 # else /* !VMS && !OS2 */
231 prof_dumpa(pTHX_ opcode ptype, U32 id)
233 if (ptype == OP_LEAVESUB) {
234 PerlIO_printf(g_fp,"- %"UVxf"\n", (UV)id);
236 else if(ptype == OP_ENTERSUB) {
237 PerlIO_printf(g_fp,"+ %"UVxf"\n", (UV)id);
239 else if(ptype == OP_GOTO) {
240 PerlIO_printf(g_fp,"* %"UVxf"\n", (UV)id);
242 else if(ptype == OP_DIE) {
243 PerlIO_printf(g_fp,"/ %"UVxf"\n", (UV)id);
246 PerlIO_printf(g_fp,"Profiler unknown prof code %d\n", ptype);
251 prof_dumps(pTHX_ U32 id, char *pname, char *gname)
253 PerlIO_printf(g_fp,"& %"UVxf" %s %s\n", (UV)id, pname, gname);
257 prof_dumpt(pTHX_ long tms_utime, long tms_stime, long realtime)
259 PerlIO_printf(g_fp,"@ %ld %ld %ld\n", tms_utime, tms_stime, realtime);
263 prof_dump_until(pTHX_ long ix)
267 clock_t realtime1, realtime2;
269 realtime1 = Times(&t1);
272 opcode ptype = g_profstack[base++].ptype;
273 if (ptype == OP_TIME) {
274 long tms_utime = g_profstack[base++].tms_utime;
275 long tms_stime = g_profstack[base++].tms_stime;
276 long realtime = g_profstack[base++].realtime;
278 prof_dumpt(aTHX_ tms_utime, tms_stime, realtime);
280 else if (ptype == OP_GV) {
281 U32 id = g_profstack[base++].id;
282 char *pname = g_profstack[base++].name;
283 char *gname = g_profstack[base++].name;
285 prof_dumps(aTHX_ id, pname, gname);
288 U32 id = g_profstack[base++].id;
289 prof_dumpa(aTHX_ ptype, id);
293 realtime2 = Times(&t2);
294 if (realtime2 != realtime1 || t1.tms_utime != t2.tms_utime
295 || t1.tms_stime != t2.tms_stime) {
296 g_wprof_r += realtime2 - realtime1;
297 g_wprof_u += t2.tms_utime - t1.tms_utime;
298 g_wprof_s += t2.tms_stime - t1.tms_stime;
300 PerlIO_printf(g_fp,"+ & Devel::DProf::write\n");
301 PerlIO_printf(g_fp,"@ %"IVdf" %"IVdf" %"IVdf"\n",
302 /* The (IV) casts are one possibility:
303 * the Painfully Correct Way would be to
305 (IV)(t2.tms_utime - t1.tms_utime),
306 (IV)(t2.tms_stime - t1.tms_stime),
307 (IV)(realtime2 - realtime1));
308 PerlIO_printf(g_fp,"- & Devel::DProf::write\n");
309 g_otms_utime = t2.tms_utime;
310 g_otms_stime = t2.tms_stime;
311 g_orealtime = realtime2;
317 set_cv_key(pTHX_ CV *cv, char *pname, char *gname)
319 SvGROW(g_key_hash, sizeof(CV**) + strlen(pname) + strlen(gname) + 3);
320 sv_setpvn(g_key_hash, (char*)&cv, sizeof(CV**));
321 sv_catpv(g_key_hash, pname);
322 sv_catpv(g_key_hash, "::");
323 sv_catpv(g_key_hash, gname);
327 prof_mark(pTHX_ opcode ptype)
330 clock_t realtime, rdelta, udelta, sdelta;
332 SV *Sub = GvSV(PL_DBsub); /* name of current sub */
335 if (g_profstack_ix + 10 > g_profstack_max) {
336 g_profstack_max = g_profstack_max * 3 / 2;
337 Renew(g_profstack, g_profstack_max, PROFANY);
341 realtime = Times(&t);
342 rdelta = realtime - g_orealtime;
343 udelta = t.tms_utime - g_otms_utime;
344 sdelta = t.tms_stime - g_otms_stime;
345 if (rdelta || udelta || sdelta) {
347 ASSERT(g_profstack_ix + 4 <= g_profstack_max);
348 g_profstack[g_profstack_ix++].ptype = OP_TIME;
349 g_profstack[g_profstack_ix++].tms_utime = udelta;
350 g_profstack[g_profstack_ix++].tms_stime = sdelta;
351 g_profstack[g_profstack_ix++].realtime = rdelta;
353 else { /* Write it to disk now so's not to eat up core */
354 if (g_prof_pid == (int)getpid()) {
355 prof_dumpt(aTHX_ udelta, sdelta, rdelta);
359 g_orealtime = realtime;
360 g_otms_stime = t.tms_stime;
361 g_otms_utime = t.tms_utime;
370 cv = db_get_cv(aTHX_ Sub);
372 pname = GvSTASH(gv) ? HvNAME_get(GvSTASH(gv)) : 0;
373 pname = pname ? pname : (char *) "(null)";
376 set_cv_key(aTHX_ cv, pname, gname);
377 svp = hv_fetch(g_cv_hash, SvPVX_const(g_key_hash), SvCUR(g_key_hash), TRUE);
379 sv_setiv(*svp, id = ++g_lastid);
380 if (CvXSUB(cv) == XS_Devel__DProf_END)
382 if (g_SAVE_STACK) { /* Store it for later recording -JH */
383 ASSERT(g_profstack_ix + 4 <= g_profstack_max);
384 g_profstack[g_profstack_ix++].ptype = OP_GV;
385 g_profstack[g_profstack_ix++].id = id;
386 g_profstack[g_profstack_ix++].name = pname;
387 g_profstack[g_profstack_ix++].name = gname;
389 else { /* Write it to disk now so's not to eat up core */
390 /* Only record the parent's info */
391 if (g_prof_pid == (int)getpid()) {
392 prof_dumps(aTHX_ id, pname, gname);
396 PL_perldb = 0; /* Do not debug the kid. */
405 if (g_SAVE_STACK) { /* Store it for later recording -JH */
406 ASSERT(g_profstack_ix + 2 <= g_profstack_max);
407 g_profstack[g_profstack_ix++].ptype = ptype;
408 g_profstack[g_profstack_ix++].id = id;
410 /* Only record the parent's info */
411 if (g_SAVE_STACK < g_profstack_ix) {
412 if (g_prof_pid == (int)getpid())
413 prof_dump_until(aTHX_ g_profstack_ix);
415 PL_perldb = 0; /* Do not debug the kid. */
419 else { /* Write it to disk now so's not to eat up core */
421 /* Only record the parent's info */
422 if (g_prof_pid == (int)getpid()) {
423 prof_dumpa(aTHX_ ptype, id);
427 PL_perldb = 0; /* Do not debug the kid. */
432 # define defstash PL_defstash
435 /* Counts overhead of prof_mark and extra XS call. */
437 test_time(pTHX_ clock_t *r, clock_t *u, clock_t *s)
439 CV *cv = perl_get_cv("Devel::DProf::NONESUCH_noxs", FALSE);
441 HV *oldstash = PL_curstash;
443 clock_t realtime1 = 0, realtime2 = 0;
444 U32 ototal = g_total;
445 U32 ostack = g_SAVE_STACK;
446 U32 operldb = PL_perldb;
448 g_SAVE_STACK = 1000000;
449 realtime1 = Times(&t1);
453 /* Disable debugging of perl_call_sv on second pass: */
454 PL_curstash = (k == 0 ? PL_defstash : PL_debstash);
455 PL_perldb = g_default_perldb;
458 g_profstack_ix = 0; /* Do not let the stack grow */
460 /* prof_mark(aTHX_ OP_ENTERSUB); */
462 PUSHMARK(PL_stack_sp);
463 perl_call_sv((SV*)cv, G_SCALAR);
465 /* prof_mark(aTHX_ OP_LEAVESUB); */
468 PL_curstash = oldstash;
469 if (k == 0) { /* Put time with debugging */
470 realtime2 = Times(&t2);
471 *r = realtime2 - realtime1;
472 *u = t2.tms_utime - t1.tms_utime;
473 *s = t2.tms_stime - t1.tms_stime;
475 else { /* Subtract time without debug */
476 realtime1 = Times(&t1);
477 *r -= realtime1 - realtime2;
478 *u -= t1.tms_utime - t2.tms_utime;
479 *s -= t1.tms_stime - t2.tms_stime;
484 g_SAVE_STACK = ostack;
489 prof_recordheader(pTHX)
493 /* g_fp is opened in the BOOT section */
494 PerlIO_printf(g_fp, "#fOrTyTwO\n");
495 PerlIO_printf(g_fp, "$hz=%"IVdf";\n", (IV)DPROF_HZ);
496 PerlIO_printf(g_fp, "$XS_VERSION='DProf %s';\n", XS_VERSION);
497 PerlIO_printf(g_fp, "# All values are given in HZ\n");
498 test_time(aTHX_ &r, &u, &s);
500 "$over_utime=%"IVdf"; $over_stime=%"IVdf"; $over_rtime=%"IVdf";\n",
501 /* The (IV) casts are one possibility:
502 * the Painfully Correct Way would be to
504 (IV)u, (IV)s, (IV)r);
505 PerlIO_printf(g_fp, "$over_tests=10000;\n");
507 g_TIMES_LOCATION = PerlIO_tell(g_fp);
509 /* Pad with whitespace. */
510 /* This should be enough even for very large numbers. */
511 PerlIO_printf(g_fp, "%*s\n", 240 , "");
513 PerlIO_printf(g_fp, "\n");
514 PerlIO_printf(g_fp, "PART2\n");
522 /* g_fp is opened in the BOOT section */
524 /* Now that we know the runtimes, fill them in at the recorded
528 prof_dump_until(aTHX_ g_profstack_ix);
530 PerlIO_seek(g_fp, g_TIMES_LOCATION, SEEK_SET);
531 /* Write into reserved 240 bytes: */
533 "$rrun_utime=%"IVdf"; $rrun_stime=%"IVdf"; $rrun_rtime=%"IVdf";",
534 /* The (IV) casts are one possibility:
535 * the Painfully Correct Way would be to
537 (IV)(g_prof_end.tms_utime-g_prof_start.tms_utime-g_wprof_u),
538 (IV)(g_prof_end.tms_stime-g_prof_start.tms_stime-g_wprof_s),
539 (IV)(g_rprof_end-g_rprof_start-g_wprof_r));
540 PerlIO_printf(g_fp, "\n$total_marks=%"IVdf, (IV)g_total);
548 check_depth(pTHX_ void *foo)
550 U32 need_depth = PTR2UV(foo);
551 if (need_depth != g_depth) {
552 if (need_depth > g_depth) {
553 warn("garbled call depth when profiling");
556 IV marks = g_depth - need_depth;
558 /* warn("Check_depth: got %d, expected %d\n", g_depth, need_depth); */
560 prof_mark(aTHX_ OP_DIE);
562 g_depth = need_depth;
574 SV *Sub = GvSV(PL_DBsub); /* name of current sub */
576 #ifdef PERL_IMPLICIT_CONTEXT
577 /* profile only the interpreter that loaded us */
580 perl_call_sv((SV*)db_get_cv(aTHX_ Sub), GIMME_V | G_NODEBUG);
585 HV *oldstash = PL_curstash;
586 I32 old_scopestack_ix = PL_scopestack_ix;
587 I32 old_cxstack_ix = cxstack_ix;
591 SAVEDESTRUCTOR_X(check_depth, INT2PTR(void*,g_depth));
594 prof_mark(aTHX_ OP_ENTERSUB);
596 perl_call_sv((SV*)db_get_cv(aTHX_ Sub), GIMME_V | G_NODEBUG);
597 PL_curstash = oldstash;
599 /* Make sure we are on the same context and scope as before the call
600 * to the sub. If the called sub was exited via a goto, next or
601 * last then this will try to croak(), however perl may still crash
602 * with a segfault. */
603 if (PL_scopestack_ix != old_scopestack_ix || cxstack_ix != old_cxstack_ix)
604 croak("panic: Devel::DProf inconsistent subroutine return");
606 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 *oldstash = PL_curstash;
635 SV *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 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 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;