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