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))
70 # define NEEDS_DPROF_TIMES
73 # define Times(ptr) (dprof_times(aTHX_ ptr))
74 # define NEEDS_DPROF_TIMES
80 # define HZ ((I32)CLK_TCK)
85 # ifdef OS2 /* times() has significant overhead */
86 # define Times(ptr) (dprof_times(aTHX_ ptr))
87 # define NEEDS_DPROF_TIMES
88 # define INCL_DOSPROFILE
89 # define INCL_DOSERRORS
91 # define toLongLong(arg) (*(long long*)&(arg))
92 # define DPROF_HZ g_dprof_ticks
94 # define Times(ptr) (times(ptr))
100 XS(XS_Devel__DProf_END); /* used by prof_mark() */
102 /* Everything is built on times(2). See its manpage for a description
107 clock_t tms_utime; /* cpu time spent in user space */
108 clock_t tms_stime; /* cpu time spent in system */
109 clock_t realtime; /* elapsed real time, in ticks */
115 typedef union prof_any PROFANY;
119 const char* out_file_name; /* output file (defaults to tmon.out) */
120 PerlIO* fp; /* pointer to tmon.out file */
121 Off_t TIMES_LOCATION; /* Where in the file to store the time totals */
122 int SAVE_STACK; /* How much data to buffer until end of run */
123 int prof_pid; /* pid of profiled process */
124 struct tms prof_start;
126 clock_t rprof_start; /* elapsed real time ticks */
137 HV* cv_hash; /* cache of CV to identifier mappings */
138 SV* key_hash; /* key for cv_hash */
147 #ifdef PERL_IMPLICIT_CONTEXT
148 PerlInterpreter *my_perl;
152 prof_state_t g_prof_state;
154 #define g_dprof_ticks g_prof_state.dprof_ticks
155 #define g_out_file_name g_prof_state.out_file_name
156 #define g_fp g_prof_state.fp
157 #define g_TIMES_LOCATION g_prof_state.TIMES_LOCATION
158 #define g_SAVE_STACK g_prof_state.SAVE_STACK
159 #define g_prof_pid g_prof_state.prof_pid
160 #define g_prof_start g_prof_state.prof_start
161 #define g_prof_end g_prof_state.prof_end
162 #define g_rprof_start g_prof_state.rprof_start
163 #define g_rprof_end g_prof_state.rprof_end
164 #define g_wprof_u g_prof_state.wprof_u
165 #define g_wprof_s g_prof_state.wprof_s
166 #define g_wprof_r g_prof_state.wprof_r
167 #define g_otms_utime g_prof_state.otms_utime
168 #define g_otms_stime g_prof_state.otms_stime
169 #define g_orealtime g_prof_state.orealtime
170 #define g_profstack g_prof_state.profstack
171 #define g_profstack_max g_prof_state.profstack_max
172 #define g_profstack_ix g_prof_state.profstack_ix
173 #define g_cv_hash g_prof_state.cv_hash
174 #define g_key_hash g_prof_state.key_hash
175 #define g_total g_prof_state.total
176 #define g_lastid g_prof_state.lastid
177 #define g_default_perldb g_prof_state.default_perldb
178 #define g_depth g_prof_state.depth
179 #ifdef PERL_IMPLICIT_CONTEXT
180 # define g_THX g_prof_state.my_perl
183 # define g_frequ g_prof_state.frequ
184 # define g_start_cnt g_prof_state.start_cnt
187 #ifdef NEEDS_DPROF_TIMES
189 dprof_times(pTHX_ struct tms *t)
196 if (CheckOSError(DosTmrQueryFreq(&g_frequ)))
197 croak("DosTmrQueryFreq: %s", SvPV_nolen(perl_get_sv("!",TRUE)));
199 g_frequ = g_frequ/DPROF_HZ; /* count per tick */
200 if (CheckOSError(DosTmrQueryTime(&cnt)))
201 croak("DosTmrQueryTime: %s",
202 SvPV_nolen_const(perl_get_sv("!",TRUE)));
203 g_start_cnt = toLongLong(cnt);
206 if (CheckOSError(DosTmrQueryTime(&cnt)))
207 croak("DosTmrQueryTime: %s", SvPV_nolen(perl_get_sv("!",TRUE)));
209 return (t->tms_utime = (toLongLong(cnt) - g_start_cnt)/g_frequ);
213 /* Get wall time and convert to 10 ms intervals to
214 * produce the return value dprof expects */
215 # if defined(__DECC) && defined (__ALPHA)
218 _ckvmssts(sys$gettim(&vmstime));
220 retval = vmstime & 0x7fffffff;
222 /* (Older hw or ccs don't have an atomic 64-bit type, so we
223 * juggle 32-bit ints (and a float) to produce a time_t result
224 * with minimal loss of information.) */
225 long int vmstime[2],remainder,divisor = 100000;
226 _ckvmssts(sys$gettim((unsigned long int *)vmstime));
227 vmstime[1] &= 0x7fff; /* prevent overflow in EDIV */
228 _ckvmssts(lib$ediv(&divisor,vmstime,(long int *)&retval,&remainder));
230 /* Fill in the struct tms using the CRTL routine . . .*/
231 times((tbuffer_t *)t);
232 return (clock_t) retval;
233 # else /* !VMS && !OS2 */
235 # include <sys/resource.h>
238 /* Measure offset from start time to avoid overflow */
239 static struct timeval tv0 = { 0, 0 };
242 if (gettimeofday(&tv0, NULL) < 0)
243 croak("gettimeofday: %s", SvPV_nolen_const(perl_get_sv("!",TRUE)));
245 if (getrusage(0, &ru) < 0)
246 croak("getrusage: %s", SvPV_nolen_const(perl_get_sv("!",TRUE)));
248 if (gettimeofday(&tv, NULL) < 0)
249 croak("gettimeofday: %s", SvPV_nolen_const(perl_get_sv("!",TRUE)));
251 t->tms_stime = DPROF_HZ * ru.ru_stime.tv_sec + ru.ru_stime.tv_usec;
252 t->tms_utime = DPROF_HZ * ru.ru_utime.tv_sec + ru.ru_utime.tv_usec;
254 if (tv.tv_usec < tv0.tv_usec)
255 tv.tv_sec--, tv.tv_usec += DPROF_HZ;
257 return DPROF_HZ * (tv.tv_sec - tv0.tv_sec) + tv.tv_usec - tv0.tv_usec;
258 # else /* !VMS && !OS2 && !BSD! */
267 prof_dumpa(pTHX_ opcode ptype, U32 id)
269 if (ptype == OP_LEAVESUB) {
270 PerlIO_printf(g_fp,"- %"UVxf"\n", (UV)id);
272 else if(ptype == OP_ENTERSUB) {
273 PerlIO_printf(g_fp,"+ %"UVxf"\n", (UV)id);
275 else if(ptype == OP_GOTO) {
276 PerlIO_printf(g_fp,"* %"UVxf"\n", (UV)id);
278 else if(ptype == OP_DIE) {
279 PerlIO_printf(g_fp,"/ %"UVxf"\n", (UV)id);
282 PerlIO_printf(g_fp,"Profiler unknown prof code %d\n", ptype);
287 prof_dumps(pTHX_ U32 id, const char *pname, const char *gname)
289 PerlIO_printf(g_fp,"& %"UVxf" %s %s\n", (UV)id, pname, gname);
293 prof_dumpt(pTHX_ long tms_utime, long tms_stime, long realtime)
295 PerlIO_printf(g_fp,"@ %ld %ld %ld\n", tms_utime, tms_stime, realtime);
299 prof_dump_until(pTHX_ long ix)
305 const clock_t realtime1 = Times(&t1);
308 const opcode ptype = g_profstack[base++].ptype;
309 if (ptype == OP_TIME) {
310 const long tms_utime = g_profstack[base++].tms_utime;
311 const long tms_stime = g_profstack[base++].tms_stime;
312 const long realtime = g_profstack[base++].realtime;
314 prof_dumpt(aTHX_ tms_utime, tms_stime, realtime);
316 else if (ptype == OP_GV) {
317 const U32 id = g_profstack[base++].id;
318 const char * const pname = g_profstack[base++].name;
319 const char * const gname = g_profstack[base++].name;
321 prof_dumps(aTHX_ id, pname, gname);
324 const U32 id = g_profstack[base++].id;
325 prof_dumpa(aTHX_ ptype, id);
329 realtime2 = Times(&t2);
330 if (realtime2 != realtime1 || t1.tms_utime != t2.tms_utime
331 || t1.tms_stime != t2.tms_stime) {
332 g_wprof_r += realtime2 - realtime1;
333 g_wprof_u += t2.tms_utime - t1.tms_utime;
334 g_wprof_s += t2.tms_stime - t1.tms_stime;
336 PerlIO_printf(g_fp,"+ & Devel::DProf::write\n");
337 PerlIO_printf(g_fp,"@ %"IVdf" %"IVdf" %"IVdf"\n",
338 /* The (IV) casts are one possibility:
339 * the Painfully Correct Way would be to
341 (IV)(t2.tms_utime - t1.tms_utime),
342 (IV)(t2.tms_stime - t1.tms_stime),
343 (IV)(realtime2 - realtime1));
344 PerlIO_printf(g_fp,"- & Devel::DProf::write\n");
345 g_otms_utime = t2.tms_utime;
346 g_otms_stime = t2.tms_stime;
347 g_orealtime = realtime2;
353 set_cv_key(pTHX_ CV *cv, const char *pname, const char *gname)
355 SvGROW(g_key_hash, sizeof(CV**) + strlen(pname) + strlen(gname) + 3);
356 sv_setpvn(g_key_hash, (char*)&cv, sizeof(CV**));
357 sv_catpv(g_key_hash, pname);
358 sv_catpv(g_key_hash, "::");
359 sv_catpv(g_key_hash, gname);
363 prof_mark(pTHX_ opcode ptype)
366 clock_t realtime, rdelta, udelta, sdelta;
368 SV * const Sub = GvSV(PL_DBsub); /* name of current sub */
371 if (g_profstack_ix + 10 > g_profstack_max) {
372 g_profstack_max = g_profstack_max * 3 / 2;
373 Renew(g_profstack, g_profstack_max, PROFANY);
377 realtime = Times(&t);
378 rdelta = realtime - g_orealtime;
379 udelta = t.tms_utime - g_otms_utime;
380 sdelta = t.tms_stime - g_otms_stime;
381 if (rdelta || udelta || sdelta) {
383 ASSERT(g_profstack_ix + 4 <= g_profstack_max);
384 g_profstack[g_profstack_ix++].ptype = OP_TIME;
385 g_profstack[g_profstack_ix++].tms_utime = udelta;
386 g_profstack[g_profstack_ix++].tms_stime = sdelta;
387 g_profstack[g_profstack_ix++].realtime = rdelta;
389 else { /* Write it to disk now so's not to eat up core */
390 if (g_prof_pid == (int)getpid()) {
391 prof_dumpt(aTHX_ udelta, sdelta, rdelta);
395 g_orealtime = realtime;
396 g_otms_stime = t.tms_stime;
397 g_otms_utime = t.tms_utime;
404 CV * const cv = db_get_cv(aTHX_ Sub);
405 GV * const gv = CvGV(cv);
406 if (isGV_with_GP(gv)) {
407 pname = GvSTASH(gv) ? HvNAME_get(GvSTASH(gv)) : NULL;
408 pname = pname ? pname : (char *) "(null)";
411 gname = pname = (char *) "(null)";
414 set_cv_key(aTHX_ cv, pname, gname);
415 svp = hv_fetch(g_cv_hash, SvPVX_const(g_key_hash), SvCUR(g_key_hash), TRUE);
417 sv_setiv(*svp, id = ++g_lastid);
418 if (CvXSUB(cv) == XS_Devel__DProf_END)
420 if (g_SAVE_STACK) { /* Store it for later recording -JH */
421 ASSERT(g_profstack_ix + 4 <= g_profstack_max);
422 g_profstack[g_profstack_ix++].ptype = OP_GV;
423 g_profstack[g_profstack_ix++].id = id;
424 g_profstack[g_profstack_ix++].name = pname;
425 g_profstack[g_profstack_ix++].name = gname;
427 else { /* Write it to disk now so's not to eat up core */
428 /* Only record the parent's info */
429 if (g_prof_pid == (int)getpid()) {
430 prof_dumps(aTHX_ id, pname, gname);
434 PL_perldb = 0; /* Do not debug the kid. */
443 if (g_SAVE_STACK) { /* Store it for later recording -JH */
444 ASSERT(g_profstack_ix + 2 <= g_profstack_max);
445 g_profstack[g_profstack_ix++].ptype = ptype;
446 g_profstack[g_profstack_ix++].id = id;
448 /* Only record the parent's info */
449 if (g_SAVE_STACK < g_profstack_ix) {
450 if (g_prof_pid == (int)getpid())
451 prof_dump_until(aTHX_ g_profstack_ix);
453 PL_perldb = 0; /* Do not debug the kid. */
457 else { /* Write it to disk now so's not to eat up core */
459 /* Only record the parent's info */
460 if (g_prof_pid == (int)getpid()) {
461 prof_dumpa(aTHX_ ptype, id);
465 PL_perldb = 0; /* Do not debug the kid. */
470 # define defstash PL_defstash
473 /* Counts overhead of prof_mark and extra XS call. */
475 test_time(pTHX_ clock_t *r, clock_t *u, clock_t *s)
477 CV * const cv = perl_get_cv("Devel::DProf::NONESUCH_noxs", FALSE);
478 HV * const oldstash = PL_curstash;
480 const U32 ototal = g_total;
481 const U32 ostack = g_SAVE_STACK;
482 const U32 operldb = PL_perldb;
485 clock_t realtime1 = Times(&t1);
486 clock_t realtime2 = 0;
488 g_SAVE_STACK = 1000000;
492 /* Disable debugging of perl_call_sv on second pass: */
493 PL_curstash = (k == 0 ? PL_defstash : PL_debstash);
494 PL_perldb = g_default_perldb;
497 g_profstack_ix = 0; /* Do not let the stack grow */
499 /* prof_mark(aTHX_ OP_ENTERSUB); */
501 PUSHMARK(PL_stack_sp);
502 perl_call_sv((SV*)cv, G_SCALAR);
504 /* prof_mark(aTHX_ OP_LEAVESUB); */
507 PL_curstash = oldstash;
508 if (k == 0) { /* Put time with debugging */
509 realtime2 = Times(&t2);
510 *r = realtime2 - realtime1;
511 *u = t2.tms_utime - t1.tms_utime;
512 *s = t2.tms_stime - t1.tms_stime;
514 else { /* Subtract time without debug */
515 realtime1 = Times(&t1);
516 *r -= realtime1 - realtime2;
517 *u -= t1.tms_utime - t2.tms_utime;
518 *s -= t1.tms_stime - t2.tms_stime;
523 g_SAVE_STACK = ostack;
528 prof_recordheader(pTHX)
532 /* g_fp is opened in the BOOT section */
533 PerlIO_printf(g_fp, "#fOrTyTwO\n");
534 PerlIO_printf(g_fp, "$hz=%"IVdf";\n", (IV)DPROF_HZ);
535 PerlIO_printf(g_fp, "$XS_VERSION='DProf %s';\n", XS_VERSION);
536 PerlIO_printf(g_fp, "# All values are given in HZ\n");
537 test_time(aTHX_ &r, &u, &s);
539 "$over_utime=%"IVdf"; $over_stime=%"IVdf"; $over_rtime=%"IVdf";\n",
540 /* The (IV) casts are one possibility:
541 * the Painfully Correct Way would be to
543 (IV)u, (IV)s, (IV)r);
544 PerlIO_printf(g_fp, "$over_tests=10000;\n");
546 g_TIMES_LOCATION = PerlIO_tell(g_fp);
548 /* Pad with whitespace. */
549 /* This should be enough even for very large numbers. */
550 PerlIO_printf(g_fp, "%*s\n", 240 , "");
552 PerlIO_printf(g_fp, "\n");
553 PerlIO_printf(g_fp, "PART2\n");
561 /* g_fp is opened in the BOOT section */
563 /* Now that we know the runtimes, fill them in at the recorded
567 prof_dump_until(aTHX_ g_profstack_ix);
569 PerlIO_seek(g_fp, g_TIMES_LOCATION, SEEK_SET);
570 /* Write into reserved 240 bytes: */
572 "$rrun_utime=%"IVdf"; $rrun_stime=%"IVdf"; $rrun_rtime=%"IVdf";",
573 /* The (IV) casts are one possibility:
574 * the Painfully Correct Way would be to
576 (IV)(g_prof_end.tms_utime-g_prof_start.tms_utime-g_wprof_u),
577 (IV)(g_prof_end.tms_stime-g_prof_start.tms_stime-g_wprof_s),
578 (IV)(g_rprof_end-g_rprof_start-g_wprof_r));
579 PerlIO_printf(g_fp, "\n$total_marks=%"IVdf, (IV)g_total);
587 check_depth(pTHX_ void *foo)
589 const U32 need_depth = PTR2UV(foo);
590 if (need_depth != g_depth) {
591 if (need_depth > g_depth) {
592 warn("garbled call depth when profiling");
595 IV marks = g_depth - need_depth;
597 /* warn("Check_depth: got %d, expected %d\n", g_depth, need_depth); */
599 prof_mark(aTHX_ OP_DIE);
601 g_depth = need_depth;
614 SV * const Sub = GvSV(PL_DBsub); /* name of current sub */
616 #ifdef PERL_IMPLICIT_CONTEXT
617 /* profile only the interpreter that loaded us */
620 perl_call_sv((SV*)db_get_cv(aTHX_ Sub), GIMME_V | G_NODEBUG);
625 HV * const oldstash = PL_curstash;
626 const I32 old_scopestack_ix = PL_scopestack_ix;
627 const I32 old_cxstack_ix = cxstack_ix;
631 SAVEDESTRUCTOR_X(check_depth, INT2PTR(void*,g_depth));
634 prof_mark(aTHX_ OP_ENTERSUB);
636 perl_call_sv((SV*)db_get_cv(aTHX_ Sub), GIMME_V | G_NODEBUG);
637 PL_curstash = oldstash;
639 /* Make sure we are on the same context and scope as before the call
640 * to the sub. If the called sub was exited via a goto, next or
641 * last then this will try to croak(), however perl may still crash
642 * with a segfault. */
643 if (PL_scopestack_ix != old_scopestack_ix || cxstack_ix != old_cxstack_ix)
644 croak("panic: Devel::DProf inconsistent subroutine return");
646 prof_mark(aTHX_ OP_LEAVESUB);
655 #ifdef PERL_IMPLICIT_CONTEXT
659 prof_mark(aTHX_ OP_GOTO);
664 #endif /* for_real */
668 MODULE = Devel::DProf PACKAGE = DB
675 HV * const oldstash = PL_curstash;
676 SV * const Sub = GvSV(PL_DBsub); /* name of current sub */
677 /* SP -= items; added by xsubpp */
680 sv_setiv(PL_DBsingle, 0); /* disable DB single-stepping */
682 prof_mark(aTHX_ OP_ENTERSUB);
685 PL_curstash = PL_debstash; /* To disable debugging of perl_call_sv */
686 perl_call_sv(Sub, GIMME_V);
687 PL_curstash = oldstash;
689 prof_mark(aTHX_ OP_LEAVESUB);
691 /* PUTBACK; added by xsubpp */
696 MODULE = Devel::DProf PACKAGE = Devel::DProf
703 /* maybe the process forked--we want only
704 * the parent's profile.
707 #ifdef PERL_IMPLICIT_CONTEXT
710 g_prof_pid == (int)getpid())
712 g_rprof_end = Times(&g_prof_end);
713 DBG_TIMER_NOTIFY("Profiler timer is off.\n");
724 g_TIMES_LOCATION = 42;
725 g_SAVE_STACK = 1<<14;
726 g_profstack_max = 128;
727 #ifdef PERL_IMPLICIT_CONTEXT
731 /* Before we go anywhere make sure we were invoked
732 * properly, else we'll dump core.
735 croak("DProf: run perl with -d to use DProf.\n");
737 /* When we hook up the XS DB::sub we'll be redefining
738 * the DB::sub from the PM file. Turn off warnings
742 const bool warn_tmp = PL_dowarn;
744 newXS("DB::sub", XS_DB_sub, file);
745 newXS("DB::goto", XS_DB_goto, file);
746 PL_dowarn = warn_tmp;
749 sv_setiv(PL_DBsingle, 0); /* disable DB single-stepping */
752 const char *buffer = getenv("PERL_DPROF_BUFFER");
755 g_SAVE_STACK = atoi(buffer);
758 buffer = getenv("PERL_DPROF_TICKS");
761 g_dprof_ticks = atoi(buffer); /* Used under OS/2 only */
767 buffer = getenv("PERL_DPROF_OUT_FILE_NAME");
768 g_out_file_name = savepv(buffer ? buffer : "tmon.out");
771 if ((g_fp = PerlIO_open(g_out_file_name, "w")) == NULL)
772 croak("DProf: unable to write '%s', errno = %d\n",
773 g_out_file_name, errno);
775 g_default_perldb = PERLDBf_NONAME | PERLDBf_SUB | PERLDBf_GOTO;
777 g_key_hash = newSV(256);
778 g_prof_pid = (int)getpid();
780 Newx(g_profstack, g_profstack_max, PROFANY);
781 prof_recordheader(aTHX);
782 DBG_TIMER_NOTIFY("Profiler timer is on.\n");
783 g_orealtime = g_rprof_start = Times(&g_prof_start);
784 g_otms_utime = g_prof_start.tms_utime;
785 g_otms_stime = g_prof_start.tms_stime;
786 PL_perldb = g_default_perldb;