remove debugging code accidentally left in
[ghc.git] / rts / ProfHeap.c
1 /* ----------------------------------------------------------------------------
2 *
3 * (c) The GHC Team, 1998-2003
4 *
5 * Support for heap profiling
6 *
7 * --------------------------------------------------------------------------*/
8
9 #include "PosixSource.h"
10 #include "Rts.h"
11 #include "RtsUtils.h"
12 #include "RtsFlags.h"
13 #include "Profiling.h"
14 #include "ProfHeap.h"
15 #include "Stats.h"
16 #include "Hash.h"
17 #include "RetainerProfile.h"
18 #include "LdvProfile.h"
19 #include "Arena.h"
20 #include "Printer.h"
21
22 #include <string.h>
23 #include <stdlib.h>
24 #include <math.h>
25
26 /* -----------------------------------------------------------------------------
27 * era stores the current time period. It is the same as the
28 * number of censuses that have been performed.
29 *
30 * RESTRICTION:
31 * era must be no longer than LDV_SHIFT (15 or 30) bits.
32 * Invariants:
33 * era is initialized to 1 in initHeapProfiling().
34 *
35 * max_era is initialized to 2^LDV_SHIFT in initHeapProfiling().
36 * When era reaches max_era, the profiling stops because a closure can
37 * store only up to (max_era - 1) as its creation or last use time.
38 * -------------------------------------------------------------------------- */
39 unsigned int era;
40 static nat max_era;
41
42 /* -----------------------------------------------------------------------------
43 * Counters
44 *
45 * For most heap profiles each closure identity gets a simple count
46 * of live words in the heap at each census. However, if we're
47 * selecting by biography, then we have to keep the various
48 * lag/drag/void counters for each identity.
49 * -------------------------------------------------------------------------- */
50 typedef struct _counter {
51 void *identity;
52 union {
53 nat resid;
54 struct {
55 int prim; // total size of 'inherently used' closures
56 int not_used; // total size of 'never used' closures
57 int used; // total size of 'used at least once' closures
58 int void_total; // current total size of 'destroyed without being used' closures
59 int drag_total; // current total size of 'used at least once and waiting to die'
60 } ldv;
61 } c;
62 struct _counter *next;
63 } counter;
64
65 STATIC_INLINE void
66 initLDVCtr( counter *ctr )
67 {
68 ctr->c.ldv.prim = 0;
69 ctr->c.ldv.not_used = 0;
70 ctr->c.ldv.used = 0;
71 ctr->c.ldv.void_total = 0;
72 ctr->c.ldv.drag_total = 0;
73 }
74
75 typedef struct {
76 double time; // the time in MUT time when the census is made
77 HashTable * hash;
78 counter * ctrs;
79 Arena * arena;
80
81 // for LDV profiling, when just displaying by LDV
82 int prim;
83 int not_used;
84 int used;
85 int void_total;
86 int drag_total;
87 } Census;
88
89 static Census *censuses = NULL;
90 static nat n_censuses = 0;
91
92 #ifdef PROFILING
93 static void aggregateCensusInfo( void );
94 #endif
95
96 static void dumpCensus( Census *census );
97
98 /* ----------------------------------------------------------------------------
99 Closure Type Profiling;
100 ------------------------------------------------------------------------- */
101
102 #ifndef PROFILING
103 static char *type_names[] = {
104 "INVALID_OBJECT",
105 "CONSTR",
106 "CONSTR_1_0",
107 "CONSTR_0_1",
108 "CONSTR_2_0",
109 "CONSTR_1_1",
110 "CONSTR_0_2",
111 "CONSTR_STATIC",
112 "CONSTR_NOCAF_STATIC",
113 "FUN",
114 "FUN_1_0",
115 "FUN_0_1",
116 "FUN_2_0",
117 "FUN_1_1",
118 "FUN_0_2",
119 "FUN_STATIC",
120 "THUNK",
121 "THUNK_1_0",
122 "THUNK_0_1",
123 "THUNK_2_0",
124 "THUNK_1_1",
125 "THUNK_0_2",
126 "THUNK_STATIC",
127 "THUNK_SELECTOR",
128 "BCO",
129 "AP",
130 "PAP",
131 "AP_STACK",
132 "IND",
133 "IND_OLDGEN",
134 "IND_PERM",
135 "IND_OLDGEN_PERM",
136 "IND_STATIC",
137 "RET_BCO",
138 "RET_SMALL",
139 "RET_BIG",
140 "RET_DYN",
141 "RET_FUN",
142 "UPDATE_FRAME",
143 "CATCH_FRAME",
144 "STOP_FRAME",
145 "CAF_BLACKHOLE",
146 "BLACKHOLE",
147 "SE_BLACKHOLE",
148 "SE_CAF_BLACKHOLE",
149 "MVAR",
150 "ARR_WORDS",
151 "MUT_ARR_PTRS_CLEAN",
152 "MUT_ARR_PTRS_DIRTY",
153 "MUT_ARR_PTRS_FROZEN0",
154 "MUT_ARR_PTRS_FROZEN",
155 "MUT_VAR_CLEAN",
156 "MUT_VAR_DIRTY",
157 "WEAK",
158 "STABLE_NAME",
159 "TSO",
160 "BLOCKED_FETCH",
161 "FETCH_ME",
162 "FETCH_ME_BQ",
163 "RBH",
164 "EVACUATED",
165 "REMOTE_REF",
166 "TVAR_WATCH_QUEUE",
167 "INVARIANT_CHECK_QUEUE",
168 "ATOMIC_INVARIANT",
169 "TVAR",
170 "TREC_CHUNK",
171 "TREC_HEADER",
172 "ATOMICALLY_FRAME",
173 "CATCH_RETRY_FRAME",
174 "CATCH_STM_FRAME",
175 "N_CLOSURE_TYPES"
176 };
177 #endif
178
179 /* ----------------------------------------------------------------------------
180 * Find the "closure identity", which is a unique pointer reresenting
181 * the band to which this closure's heap space is attributed in the
182 * heap profile.
183 * ------------------------------------------------------------------------- */
184 STATIC_INLINE void *
185 closureIdentity( StgClosure *p )
186 {
187 switch (RtsFlags.ProfFlags.doHeapProfile) {
188
189 #ifdef PROFILING
190 case HEAP_BY_CCS:
191 return p->header.prof.ccs;
192 case HEAP_BY_MOD:
193 return p->header.prof.ccs->cc->module;
194 case HEAP_BY_DESCR:
195 return GET_PROF_DESC(get_itbl(p));
196 case HEAP_BY_TYPE:
197 return GET_PROF_TYPE(get_itbl(p));
198 case HEAP_BY_RETAINER:
199 // AFAIK, the only closures in the heap which might not have a
200 // valid retainer set are DEAD_WEAK closures.
201 if (isRetainerSetFieldValid(p))
202 return retainerSetOf(p);
203 else
204 return NULL;
205
206 #else
207 case HEAP_BY_CLOSURE_TYPE:
208 {
209 StgInfoTable *info;
210 info = get_itbl(p);
211 switch (info->type) {
212 case CONSTR:
213 case CONSTR_1_0:
214 case CONSTR_0_1:
215 case CONSTR_2_0:
216 case CONSTR_1_1:
217 case CONSTR_0_2:
218 case CONSTR_STATIC:
219 case CONSTR_NOCAF_STATIC:
220 return GET_CON_DESC(itbl_to_con_itbl(info));
221 default:
222 return type_names[info->type];
223 }
224 }
225
226 #endif
227 default:
228 barf("closureIdentity");
229 }
230 }
231
232 /* --------------------------------------------------------------------------
233 * Profiling type predicates
234 * ----------------------------------------------------------------------- */
235 #ifdef PROFILING
236 STATIC_INLINE rtsBool
237 doingLDVProfiling( void )
238 {
239 return (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
240 || RtsFlags.ProfFlags.bioSelector != NULL);
241 }
242
243 STATIC_INLINE rtsBool
244 doingRetainerProfiling( void )
245 {
246 return (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER
247 || RtsFlags.ProfFlags.retainerSelector != NULL);
248 }
249 #endif /* PROFILING */
250
251 // Precesses a closure 'c' being destroyed whose size is 'size'.
252 // Make sure that LDV_recordDead() is not invoked on 'inherently used' closures
253 // such as TSO; they should not be involved in computing dragNew or voidNew.
254 //
255 // Even though era is checked in both LdvCensusForDead() and
256 // LdvCensusKillAll(), we still need to make sure that era is > 0 because
257 // LDV_recordDead() may be called from elsewhere in the runtime system. E.g.,
258 // when a thunk is replaced by an indirection object.
259
260 #ifdef PROFILING
261 void
262 LDV_recordDead( StgClosure *c, nat size )
263 {
264 void *id;
265 nat t;
266 counter *ctr;
267
268 if (era > 0 && closureSatisfiesConstraints(c)) {
269 size -= sizeofW(StgProfHeader);
270 ASSERT(LDVW(c) != 0);
271 if ((LDVW((c)) & LDV_STATE_MASK) == LDV_STATE_CREATE) {
272 t = (LDVW((c)) & LDV_CREATE_MASK) >> LDV_SHIFT;
273 if (t < era) {
274 if (RtsFlags.ProfFlags.bioSelector == NULL) {
275 censuses[t].void_total += (int)size;
276 censuses[era].void_total -= (int)size;
277 ASSERT(censuses[t].void_total < censuses[t].not_used);
278 } else {
279 id = closureIdentity(c);
280 ctr = lookupHashTable(censuses[t].hash, (StgWord)id);
281 ASSERT( ctr != NULL );
282 ctr->c.ldv.void_total += (int)size;
283 ctr = lookupHashTable(censuses[era].hash, (StgWord)id);
284 if (ctr == NULL) {
285 ctr = arenaAlloc(censuses[era].arena, sizeof(counter));
286 initLDVCtr(ctr);
287 insertHashTable(censuses[era].hash, (StgWord)id, ctr);
288 ctr->identity = id;
289 ctr->next = censuses[era].ctrs;
290 censuses[era].ctrs = ctr;
291 }
292 ctr->c.ldv.void_total -= (int)size;
293 }
294 }
295 } else {
296 t = LDVW((c)) & LDV_LAST_MASK;
297 if (t + 1 < era) {
298 if (RtsFlags.ProfFlags.bioSelector == NULL) {
299 censuses[t+1].drag_total += size;
300 censuses[era].drag_total -= size;
301 } else {
302 void *id;
303 id = closureIdentity(c);
304 ctr = lookupHashTable(censuses[t+1].hash, (StgWord)id);
305 ASSERT( ctr != NULL );
306 ctr->c.ldv.drag_total += (int)size;
307 ctr = lookupHashTable(censuses[era].hash, (StgWord)id);
308 if (ctr == NULL) {
309 ctr = arenaAlloc(censuses[era].arena, sizeof(counter));
310 initLDVCtr(ctr);
311 insertHashTable(censuses[era].hash, (StgWord)id, ctr);
312 ctr->identity = id;
313 ctr->next = censuses[era].ctrs;
314 censuses[era].ctrs = ctr;
315 }
316 ctr->c.ldv.drag_total -= (int)size;
317 }
318 }
319 }
320 }
321 }
322 #endif
323
324 /* --------------------------------------------------------------------------
325 * Initialize censuses[era];
326 * ----------------------------------------------------------------------- */
327
328 STATIC_INLINE void
329 initEra(Census *census)
330 {
331 census->hash = allocHashTable();
332 census->ctrs = NULL;
333 census->arena = newArena();
334
335 census->not_used = 0;
336 census->used = 0;
337 census->prim = 0;
338 census->void_total = 0;
339 census->drag_total = 0;
340 }
341
342 STATIC_INLINE void
343 freeEra(Census *census)
344 {
345 if (RtsFlags.ProfFlags.bioSelector != NULL)
346 // when bioSelector==NULL, these are freed in heapCensus()
347 {
348 arenaFree(census->arena);
349 freeHashTable(census->hash, NULL);
350 }
351 }
352
353 /* --------------------------------------------------------------------------
354 * Increases era by 1 and initialize census[era].
355 * Reallocates gi[] and increases its size if needed.
356 * ----------------------------------------------------------------------- */
357
358 static void
359 nextEra( void )
360 {
361 #ifdef PROFILING
362 if (doingLDVProfiling()) {
363 era++;
364
365 if (era == max_era) {
366 errorBelch("maximum number of censuses reached; use +RTS -i to reduce");
367 stg_exit(EXIT_FAILURE);
368 }
369
370 if (era == n_censuses) {
371 n_censuses *= 2;
372 censuses = stgReallocBytes(censuses, sizeof(Census) * n_censuses,
373 "nextEra");
374 }
375 }
376 #endif /* PROFILING */
377
378 initEra( &censuses[era] );
379 }
380
381 /* ----------------------------------------------------------------------------
382 * Heap profiling by info table
383 * ------------------------------------------------------------------------- */
384
385 #if !defined(PROFILING)
386 FILE *hp_file;
387 static char *hp_filename;
388
389 void initProfiling1 (void)
390 {
391 }
392
393 void freeProfiling1 (void)
394 {
395 }
396
397 void initProfiling2 (void)
398 {
399 if (RtsFlags.ProfFlags.doHeapProfile) {
400 /* Initialise the log file name */
401 hp_filename = stgMallocBytes(strlen(prog_name) + 6, "hpFileName");
402 sprintf(hp_filename, "%s.hp", prog_name);
403
404 /* open the log file */
405 if ((hp_file = fopen(hp_filename, "w")) == NULL) {
406 debugBelch("Can't open profiling report file %s\n",
407 hp_filename);
408 RtsFlags.ProfFlags.doHeapProfile = 0;
409 return;
410 }
411 }
412
413 initHeapProfiling();
414 }
415
416 void endProfiling( void )
417 {
418 endHeapProfiling();
419 }
420 #endif /* !PROFILING */
421
422 static void
423 printSample(rtsBool beginSample, StgDouble sampleValue)
424 {
425 StgDouble fractionalPart, integralPart;
426 fractionalPart = modf(sampleValue, &integralPart);
427 fprintf(hp_file, "%s %" FMT_Word64 ".%02" FMT_Word64 "\n",
428 (beginSample ? "BEGIN_SAMPLE" : "END_SAMPLE"),
429 (StgWord64)integralPart, (StgWord64)(fractionalPart * 100));
430 }
431
432 /* --------------------------------------------------------------------------
433 * Initialize the heap profilier
434 * ----------------------------------------------------------------------- */
435 nat
436 initHeapProfiling(void)
437 {
438 if (! RtsFlags.ProfFlags.doHeapProfile) {
439 return 0;
440 }
441
442 #ifdef PROFILING
443 if (doingLDVProfiling() && doingRetainerProfiling()) {
444 errorBelch("cannot mix -hb and -hr");
445 stg_exit(EXIT_FAILURE);
446 }
447 #endif
448
449 // we only count eras if we're doing LDV profiling. Otherwise era
450 // is fixed at zero.
451 #ifdef PROFILING
452 if (doingLDVProfiling()) {
453 era = 1;
454 } else
455 #endif
456 {
457 era = 0;
458 }
459
460 { // max_era = 2^LDV_SHIFT
461 nat p;
462 max_era = 1;
463 for (p = 0; p < LDV_SHIFT; p++)
464 max_era *= 2;
465 }
466
467 n_censuses = 32;
468 censuses = stgMallocBytes(sizeof(Census) * n_censuses, "initHeapProfiling");
469
470 initEra( &censuses[era] );
471
472 /* initProfilingLogFile(); */
473 fprintf(hp_file, "JOB \"%s", prog_name);
474
475 #ifdef PROFILING
476 {
477 int count;
478 for(count = 1; count < prog_argc; count++)
479 fprintf(hp_file, " %s", prog_argv[count]);
480 fprintf(hp_file, " +RTS");
481 for(count = 0; count < rts_argc; count++)
482 fprintf(hp_file, " %s", rts_argv[count]);
483 }
484 #endif /* PROFILING */
485
486 fprintf(hp_file, "\"\n" );
487
488 fprintf(hp_file, "DATE \"%s\"\n", time_str());
489
490 fprintf(hp_file, "SAMPLE_UNIT \"seconds\"\n");
491 fprintf(hp_file, "VALUE_UNIT \"bytes\"\n");
492
493 printSample(rtsTrue, 0);
494 printSample(rtsFalse, 0);
495
496 #ifdef PROFILING
497 if (doingRetainerProfiling()) {
498 initRetainerProfiling();
499 }
500 #endif
501
502 return 0;
503 }
504
505 void
506 endHeapProfiling(void)
507 {
508 StgDouble seconds;
509
510 if (! RtsFlags.ProfFlags.doHeapProfile) {
511 return;
512 }
513
514 #ifdef PROFILING
515 if (doingRetainerProfiling()) {
516 endRetainerProfiling();
517 }
518 #endif
519
520 #ifdef PROFILING
521 if (doingLDVProfiling()) {
522 nat t;
523 LdvCensusKillAll();
524 aggregateCensusInfo();
525 for (t = 1; t < era; t++) {
526 dumpCensus( &censuses[t] );
527 }
528 }
529 #endif
530
531 #ifdef PROFILING
532 if (doingLDVProfiling()) {
533 nat t;
534 for (t = 1; t <= era; t++) {
535 freeEra( &censuses[t] );
536 }
537 } else {
538 freeEra( &censuses[0] );
539 }
540 #else
541 freeEra( &censuses[0] );
542 #endif
543
544 stgFree(censuses);
545
546 seconds = mut_user_time();
547 printSample(rtsTrue, seconds);
548 printSample(rtsFalse, seconds);
549 fclose(hp_file);
550 }
551
552
553
554 #ifdef PROFILING
555 static size_t
556 buf_append(char *p, const char *q, char *end)
557 {
558 int m;
559
560 for (m = 0; p < end; p++, q++, m++) {
561 *p = *q;
562 if (*q == '\0') { break; }
563 }
564 return m;
565 }
566
567 static void
568 fprint_ccs(FILE *fp, CostCentreStack *ccs, nat max_length)
569 {
570 char buf[max_length+1], *p, *buf_end;
571
572 // MAIN on its own gets printed as "MAIN", otherwise we ignore MAIN.
573 if (ccs == CCS_MAIN) {
574 fprintf(fp, "MAIN");
575 return;
576 }
577
578 fprintf(fp, "(%ld)", ccs->ccsID);
579
580 p = buf;
581 buf_end = buf + max_length + 1;
582
583 // keep printing components of the stack until we run out of space
584 // in the buffer. If we run out of space, end with "...".
585 for (; ccs != NULL && ccs != CCS_MAIN; ccs = ccs->prevStack) {
586
587 // CAF cost centres print as M.CAF, but we leave the module
588 // name out of all the others to save space.
589 if (!strcmp(ccs->cc->label,"CAF")) {
590 p += buf_append(p, ccs->cc->module, buf_end);
591 p += buf_append(p, ".CAF", buf_end);
592 } else {
593 p += buf_append(p, ccs->cc->label, buf_end);
594 if (ccs->prevStack != NULL && ccs->prevStack != CCS_MAIN) {
595 p += buf_append(p, "/", buf_end);
596 }
597 }
598
599 if (p >= buf_end) {
600 sprintf(buf+max_length-4, "...");
601 break;
602 }
603 }
604 fprintf(fp, "%s", buf);
605 }
606 #endif /* PROFILING */
607
608 rtsBool
609 strMatchesSelector( char* str, char* sel )
610 {
611 char* p;
612 // debugBelch("str_matches_selector %s %s\n", str, sel);
613 while (1) {
614 // Compare str against wherever we've got to in sel.
615 p = str;
616 while (*p != '\0' && *sel != ',' && *sel != '\0' && *p == *sel) {
617 p++; sel++;
618 }
619 // Match if all of str used and have reached the end of a sel fragment.
620 if (*p == '\0' && (*sel == ',' || *sel == '\0'))
621 return rtsTrue;
622
623 // No match. Advance sel to the start of the next elem.
624 while (*sel != ',' && *sel != '\0') sel++;
625 if (*sel == ',') sel++;
626
627 /* Run out of sel ?? */
628 if (*sel == '\0') return rtsFalse;
629 }
630 }
631
632 /* -----------------------------------------------------------------------------
633 * Figure out whether a closure should be counted in this census, by
634 * testing against all the specified constraints.
635 * -------------------------------------------------------------------------- */
636 rtsBool
637 closureSatisfiesConstraints( StgClosure* p )
638 {
639 #if !defined(PROFILING)
640 (void)p; /* keep gcc -Wall happy */
641 return rtsTrue;
642 #else
643 rtsBool b;
644
645 // The CCS has a selected field to indicate whether this closure is
646 // deselected by not being mentioned in the module, CC, or CCS
647 // selectors.
648 if (!p->header.prof.ccs->selected) {
649 return rtsFalse;
650 }
651
652 if (RtsFlags.ProfFlags.descrSelector) {
653 b = strMatchesSelector( (GET_PROF_DESC(get_itbl((StgClosure *)p))),
654 RtsFlags.ProfFlags.descrSelector );
655 if (!b) return rtsFalse;
656 }
657 if (RtsFlags.ProfFlags.typeSelector) {
658 b = strMatchesSelector( (GET_PROF_TYPE(get_itbl((StgClosure *)p))),
659 RtsFlags.ProfFlags.typeSelector );
660 if (!b) return rtsFalse;
661 }
662 if (RtsFlags.ProfFlags.retainerSelector) {
663 RetainerSet *rs;
664 nat i;
665 // We must check that the retainer set is valid here. One
666 // reason it might not be valid is if this closure is a
667 // a newly deceased weak pointer (i.e. a DEAD_WEAK), since
668 // these aren't reached by the retainer profiler's traversal.
669 if (isRetainerSetFieldValid((StgClosure *)p)) {
670 rs = retainerSetOf((StgClosure *)p);
671 if (rs != NULL) {
672 for (i = 0; i < rs->num; i++) {
673 b = strMatchesSelector( rs->element[i]->cc->label,
674 RtsFlags.ProfFlags.retainerSelector );
675 if (b) return rtsTrue;
676 }
677 }
678 }
679 return rtsFalse;
680 }
681 return rtsTrue;
682 #endif /* PROFILING */
683 }
684
685 /* -----------------------------------------------------------------------------
686 * Aggregate the heap census info for biographical profiling
687 * -------------------------------------------------------------------------- */
688 #ifdef PROFILING
689 static void
690 aggregateCensusInfo( void )
691 {
692 HashTable *acc;
693 nat t;
694 counter *c, *d, *ctrs;
695 Arena *arena;
696
697 if (!doingLDVProfiling()) return;
698
699 // Aggregate the LDV counters when displaying by biography.
700 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV) {
701 int void_total, drag_total;
702
703 // Now we compute void_total and drag_total for each census
704 // After the program has finished, the void_total field of
705 // each census contains the count of words that were *created*
706 // in this era and were eventually void. Conversely, if a
707 // void closure was destroyed in this era, it will be
708 // represented by a negative count of words in void_total.
709 //
710 // To get the count of live words that are void at each
711 // census, just propagate the void_total count forwards:
712
713 void_total = 0;
714 drag_total = 0;
715 for (t = 1; t < era; t++) { // note: start at 1, not 0
716 void_total += censuses[t].void_total;
717 drag_total += censuses[t].drag_total;
718 censuses[t].void_total = void_total;
719 censuses[t].drag_total = drag_total;
720
721 ASSERT( censuses[t].void_total <= censuses[t].not_used );
722 // should be true because: void_total is the count of
723 // live words that are void at this census, which *must*
724 // be less than the number of live words that have not
725 // been used yet.
726
727 ASSERT( censuses[t].drag_total <= censuses[t].used );
728 // similar reasoning as above.
729 }
730
731 return;
732 }
733
734 // otherwise... we're doing a heap profile that is restricted to
735 // some combination of lag, drag, void or use. We've kept all the
736 // census info for all censuses so far, but we still need to
737 // aggregate the counters forwards.
738
739 arena = newArena();
740 acc = allocHashTable();
741 ctrs = NULL;
742
743 for (t = 1; t < era; t++) {
744
745 // first look through all the counters we're aggregating
746 for (c = ctrs; c != NULL; c = c->next) {
747 // if one of the totals is non-zero, then this closure
748 // type must be present in the heap at this census time...
749 d = lookupHashTable(censuses[t].hash, (StgWord)c->identity);
750
751 if (d == NULL) {
752 // if this closure identity isn't present in the
753 // census for this time period, then our running
754 // totals *must* be zero.
755 ASSERT(c->c.ldv.void_total == 0 && c->c.ldv.drag_total == 0);
756
757 // debugCCS(c->identity);
758 // debugBelch(" census=%d void_total=%d drag_total=%d\n",
759 // t, c->c.ldv.void_total, c->c.ldv.drag_total);
760 } else {
761 d->c.ldv.void_total += c->c.ldv.void_total;
762 d->c.ldv.drag_total += c->c.ldv.drag_total;
763 c->c.ldv.void_total = d->c.ldv.void_total;
764 c->c.ldv.drag_total = d->c.ldv.drag_total;
765
766 ASSERT( c->c.ldv.void_total >= 0 );
767 ASSERT( c->c.ldv.drag_total >= 0 );
768 }
769 }
770
771 // now look through the counters in this census to find new ones
772 for (c = censuses[t].ctrs; c != NULL; c = c->next) {
773 d = lookupHashTable(acc, (StgWord)c->identity);
774 if (d == NULL) {
775 d = arenaAlloc( arena, sizeof(counter) );
776 initLDVCtr(d);
777 insertHashTable( acc, (StgWord)c->identity, d );
778 d->identity = c->identity;
779 d->next = ctrs;
780 ctrs = d;
781 d->c.ldv.void_total = c->c.ldv.void_total;
782 d->c.ldv.drag_total = c->c.ldv.drag_total;
783 }
784 ASSERT( c->c.ldv.void_total >= 0 );
785 ASSERT( c->c.ldv.drag_total >= 0 );
786 }
787 }
788
789 freeHashTable(acc, NULL);
790 arenaFree(arena);
791 }
792 #endif
793
794 /* -----------------------------------------------------------------------------
795 * Print out the results of a heap census.
796 * -------------------------------------------------------------------------- */
797 static void
798 dumpCensus( Census *census )
799 {
800 counter *ctr;
801 int count;
802
803 printSample(rtsTrue, census->time);
804
805 #ifdef PROFILING
806 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV) {
807 fprintf(hp_file, "VOID\t%lu\n", (unsigned long)(census->void_total) * sizeof(W_));
808 fprintf(hp_file, "LAG\t%lu\n",
809 (unsigned long)(census->not_used - census->void_total) * sizeof(W_));
810 fprintf(hp_file, "USE\t%lu\n",
811 (unsigned long)(census->used - census->drag_total) * sizeof(W_));
812 fprintf(hp_file, "INHERENT_USE\t%lu\n",
813 (unsigned long)(census->prim) * sizeof(W_));
814 fprintf(hp_file, "DRAG\t%lu\n",
815 (unsigned long)(census->drag_total) * sizeof(W_));
816 printSample(rtsFalse, census->time);
817 return;
818 }
819 #endif
820
821 for (ctr = census->ctrs; ctr != NULL; ctr = ctr->next) {
822
823 #ifdef PROFILING
824 if (RtsFlags.ProfFlags.bioSelector != NULL) {
825 count = 0;
826 if (strMatchesSelector("lag", RtsFlags.ProfFlags.bioSelector))
827 count += ctr->c.ldv.not_used - ctr->c.ldv.void_total;
828 if (strMatchesSelector("drag", RtsFlags.ProfFlags.bioSelector))
829 count += ctr->c.ldv.drag_total;
830 if (strMatchesSelector("void", RtsFlags.ProfFlags.bioSelector))
831 count += ctr->c.ldv.void_total;
832 if (strMatchesSelector("use", RtsFlags.ProfFlags.bioSelector))
833 count += ctr->c.ldv.used - ctr->c.ldv.drag_total;
834 } else
835 #endif
836 {
837 count = ctr->c.resid;
838 }
839
840 ASSERT( count >= 0 );
841
842 if (count == 0) continue;
843
844 #if !defined(PROFILING)
845 switch (RtsFlags.ProfFlags.doHeapProfile) {
846 case HEAP_BY_CLOSURE_TYPE:
847 fprintf(hp_file, "%s", (char *)ctr->identity);
848 break;
849 }
850 #endif
851
852 #ifdef PROFILING
853 switch (RtsFlags.ProfFlags.doHeapProfile) {
854 case HEAP_BY_CCS:
855 fprint_ccs(hp_file, (CostCentreStack *)ctr->identity, RtsFlags.ProfFlags.ccsLength);
856 break;
857 case HEAP_BY_MOD:
858 case HEAP_BY_DESCR:
859 case HEAP_BY_TYPE:
860 fprintf(hp_file, "%s", (char *)ctr->identity);
861 break;
862 case HEAP_BY_RETAINER:
863 {
864 RetainerSet *rs = (RetainerSet *)ctr->identity;
865
866 // it might be the distinguished retainer set rs_MANY:
867 if (rs == &rs_MANY) {
868 fprintf(hp_file, "MANY");
869 break;
870 }
871
872 // Mark this retainer set by negating its id, because it
873 // has appeared in at least one census. We print the
874 // values of all such retainer sets into the log file at
875 // the end. A retainer set may exist but not feature in
876 // any censuses if it arose as the intermediate retainer
877 // set for some closure during retainer set calculation.
878 if (rs->id > 0)
879 rs->id = -(rs->id);
880
881 // report in the unit of bytes: * sizeof(StgWord)
882 printRetainerSetShort(hp_file, rs);
883 break;
884 }
885 default:
886 barf("dumpCensus; doHeapProfile");
887 }
888 #endif
889
890 fprintf(hp_file, "\t%lu\n", (unsigned long)count * sizeof(W_));
891 }
892
893 printSample(rtsFalse, census->time);
894 }
895
896 /* -----------------------------------------------------------------------------
897 * Code to perform a heap census.
898 * -------------------------------------------------------------------------- */
899 static void
900 heapCensusChain( Census *census, bdescr *bd )
901 {
902 StgPtr p;
903 StgInfoTable *info;
904 void *identity;
905 nat size;
906 counter *ctr;
907 nat real_size;
908 rtsBool prim;
909
910 for (; bd != NULL; bd = bd->link) {
911
912 // HACK: ignore pinned blocks, because they contain gaps.
913 // It's not clear exactly what we'd like to do here, since we
914 // can't tell which objects in the block are actually alive.
915 // Perhaps the whole block should be counted as SYSTEM memory.
916 if (bd->flags & BF_PINNED) {
917 continue;
918 }
919
920 p = bd->start;
921 while (p < bd->free) {
922 info = get_itbl((StgClosure *)p);
923 prim = rtsFalse;
924
925 switch (info->type) {
926
927 case THUNK:
928 size = thunk_sizeW_fromITBL(info);
929 break;
930
931 case THUNK_1_1:
932 case THUNK_0_2:
933 case THUNK_2_0:
934 size = sizeofW(StgThunkHeader) + 2;
935 break;
936
937 case THUNK_1_0:
938 case THUNK_0_1:
939 case THUNK_SELECTOR:
940 size = sizeofW(StgThunkHeader) + 1;
941 break;
942
943 case CONSTR:
944 case FUN:
945 case IND_PERM:
946 case IND_OLDGEN:
947 case IND_OLDGEN_PERM:
948 case CAF_BLACKHOLE:
949 case SE_CAF_BLACKHOLE:
950 case SE_BLACKHOLE:
951 case BLACKHOLE:
952 case FUN_1_0:
953 case FUN_0_1:
954 case FUN_1_1:
955 case FUN_0_2:
956 case FUN_2_0:
957 case CONSTR_1_0:
958 case CONSTR_0_1:
959 case CONSTR_1_1:
960 case CONSTR_0_2:
961 case CONSTR_2_0:
962 size = sizeW_fromITBL(info);
963 break;
964
965 case IND:
966 // Special case/Delicate Hack: INDs don't normally
967 // appear, since we're doing this heap census right
968 // after GC. However, GarbageCollect() also does
969 // resurrectThreads(), which can update some
970 // blackholes when it calls raiseAsync() on the
971 // resurrected threads. So we know that any IND will
972 // be the size of a BLACKHOLE.
973 size = BLACKHOLE_sizeW();
974 break;
975
976 case BCO:
977 prim = rtsTrue;
978 size = bco_sizeW((StgBCO *)p);
979 break;
980
981 case MVAR:
982 case WEAK:
983 case STABLE_NAME:
984 case MUT_VAR_CLEAN:
985 case MUT_VAR_DIRTY:
986 prim = rtsTrue;
987 size = sizeW_fromITBL(info);
988 break;
989
990 case AP:
991 size = ap_sizeW((StgAP *)p);
992 break;
993
994 case PAP:
995 size = pap_sizeW((StgPAP *)p);
996 break;
997
998 case AP_STACK:
999 size = ap_stack_sizeW((StgAP_STACK *)p);
1000 break;
1001
1002 case ARR_WORDS:
1003 prim = rtsTrue;
1004 size = arr_words_sizeW(stgCast(StgArrWords*,p));
1005 break;
1006
1007 case MUT_ARR_PTRS_CLEAN:
1008 case MUT_ARR_PTRS_DIRTY:
1009 case MUT_ARR_PTRS_FROZEN:
1010 case MUT_ARR_PTRS_FROZEN0:
1011 prim = rtsTrue;
1012 size = mut_arr_ptrs_sizeW((StgMutArrPtrs *)p);
1013 break;
1014
1015 case TSO:
1016 prim = rtsTrue;
1017 #ifdef PROFILING
1018 if (RtsFlags.ProfFlags.includeTSOs) {
1019 size = tso_sizeW((StgTSO *)p);
1020 break;
1021 } else {
1022 // Skip this TSO and move on to the next object
1023 p += tso_sizeW((StgTSO *)p);
1024 continue;
1025 }
1026 #else
1027 size = tso_sizeW((StgTSO *)p);
1028 break;
1029 #endif
1030
1031 case TREC_HEADER:
1032 prim = rtsTrue;
1033 size = sizeofW(StgTRecHeader);
1034 break;
1035
1036 case TVAR_WATCH_QUEUE:
1037 prim = rtsTrue;
1038 size = sizeofW(StgTVarWatchQueue);
1039 break;
1040
1041 case INVARIANT_CHECK_QUEUE:
1042 prim = rtsTrue;
1043 size = sizeofW(StgInvariantCheckQueue);
1044 break;
1045
1046 case ATOMIC_INVARIANT:
1047 prim = rtsTrue;
1048 size = sizeofW(StgAtomicInvariant);
1049 break;
1050
1051 case TVAR:
1052 prim = rtsTrue;
1053 size = sizeofW(StgTVar);
1054 break;
1055
1056 case TREC_CHUNK:
1057 prim = rtsTrue;
1058 size = sizeofW(StgTRecChunk);
1059 break;
1060
1061 default:
1062 barf("heapCensus, unknown object: %d", info->type);
1063 }
1064
1065 identity = NULL;
1066
1067 #ifdef PROFILING
1068 // subtract the profiling overhead
1069 real_size = size - sizeofW(StgProfHeader);
1070 #else
1071 real_size = size;
1072 #endif
1073
1074 if (closureSatisfiesConstraints((StgClosure*)p)) {
1075 #ifdef PROFILING
1076 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV) {
1077 if (prim)
1078 census->prim += real_size;
1079 else if ((LDVW(p) & LDV_STATE_MASK) == LDV_STATE_CREATE)
1080 census->not_used += real_size;
1081 else
1082 census->used += real_size;
1083 } else
1084 #endif
1085 {
1086 identity = closureIdentity((StgClosure *)p);
1087
1088 if (identity != NULL) {
1089 ctr = lookupHashTable( census->hash, (StgWord)identity );
1090 if (ctr != NULL) {
1091 #ifdef PROFILING
1092 if (RtsFlags.ProfFlags.bioSelector != NULL) {
1093 if (prim)
1094 ctr->c.ldv.prim += real_size;
1095 else if ((LDVW(p) & LDV_STATE_MASK) == LDV_STATE_CREATE)
1096 ctr->c.ldv.not_used += real_size;
1097 else
1098 ctr->c.ldv.used += real_size;
1099 } else
1100 #endif
1101 {
1102 ctr->c.resid += real_size;
1103 }
1104 } else {
1105 ctr = arenaAlloc( census->arena, sizeof(counter) );
1106 initLDVCtr(ctr);
1107 insertHashTable( census->hash, (StgWord)identity, ctr );
1108 ctr->identity = identity;
1109 ctr->next = census->ctrs;
1110 census->ctrs = ctr;
1111
1112 #ifdef PROFILING
1113 if (RtsFlags.ProfFlags.bioSelector != NULL) {
1114 if (prim)
1115 ctr->c.ldv.prim = real_size;
1116 else if ((LDVW(p) & LDV_STATE_MASK) == LDV_STATE_CREATE)
1117 ctr->c.ldv.not_used = real_size;
1118 else
1119 ctr->c.ldv.used = real_size;
1120 } else
1121 #endif
1122 {
1123 ctr->c.resid = real_size;
1124 }
1125 }
1126 }
1127 }
1128 }
1129
1130 p += size;
1131 }
1132 }
1133 }
1134
1135 void
1136 heapCensus( void )
1137 {
1138 nat g, s;
1139 Census *census;
1140
1141 census = &censuses[era];
1142 census->time = mut_user_time();
1143
1144 // calculate retainer sets if necessary
1145 #ifdef PROFILING
1146 if (doingRetainerProfiling()) {
1147 retainerProfile();
1148 }
1149 #endif
1150
1151 #ifdef PROFILING
1152 stat_startHeapCensus();
1153 #endif
1154
1155 // Traverse the heap, collecting the census info
1156
1157 // First the small_alloc_list: we have to fix the free pointer at
1158 // the end by calling tidyAllocatedLists() first.
1159 tidyAllocateLists();
1160 heapCensusChain( census, small_alloc_list );
1161
1162 // Now traverse the heap in each generation/step.
1163 if (RtsFlags.GcFlags.generations == 1) {
1164 heapCensusChain( census, g0s0->blocks );
1165 } else {
1166 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1167 for (s = 0; s < generations[g].n_steps; s++) {
1168 heapCensusChain( census, generations[g].steps[s].blocks );
1169 // Are we interested in large objects? might be
1170 // confusing to include the stack in a heap profile.
1171 heapCensusChain( census, generations[g].steps[s].large_objects );
1172 }
1173 }
1174 }
1175
1176 // dump out the census info
1177 #ifdef PROFILING
1178 // We can't generate any info for LDV profiling until
1179 // the end of the run...
1180 if (!doingLDVProfiling())
1181 dumpCensus( census );
1182 #else
1183 dumpCensus( census );
1184 #endif
1185
1186
1187 // free our storage, unless we're keeping all the census info for
1188 // future restriction by biography.
1189 #ifdef PROFILING
1190 if (RtsFlags.ProfFlags.bioSelector == NULL)
1191 {
1192 freeHashTable( census->hash, NULL/* don't free the elements */ );
1193 arenaFree( census->arena );
1194 census->hash = NULL;
1195 census->arena = NULL;
1196 }
1197 #endif
1198
1199 // we're into the next time period now
1200 nextEra();
1201
1202 #ifdef PROFILING
1203 stat_endHeapCensus();
1204 #endif
1205 }
1206