1 /* -----------------------------------------------------------------------------
2 * $Id: ProfHeap.c,v 1.43 2003/02/20 15:39:59 simonmar Exp $
4 * (c) The GHC Team, 1998-2000
6 * Support for heap profiling
8 * ---------------------------------------------------------------------------*/
10 #if defined(DEBUG) && !defined(PROFILING)
11 #define DEBUG_HEAP_PROF
13 #undef DEBUG_HEAP_PROF
16 #if defined(PROFILING) || defined(DEBUG_HEAP_PROF)
18 #include "PosixSource.h"
22 #include "Profiling.h"
28 #include "RetainerProfile.h"
29 #include "LdvProfile.h"
36 /* -----------------------------------------------------------------------------
37 * era stores the current time period. It is the same as the
38 * number of censuses that have been performed.
41 * era must be no longer than LDV_SHIFT (15 or 30) bits.
43 * era is initialized to 1 in initHeapProfiling().
45 * max_era is initialized to 2^LDV_SHIFT in initHeapProfiling().
46 * When era reaches max_era, the profiling stops because a closure can
47 * store only up to (max_era - 1) as its creation or last use time.
48 * -------------------------------------------------------------------------- */
52 /* -----------------------------------------------------------------------------
55 * For most heap profiles each closure identity gets a simple count
56 * of live words in the heap at each census. However, if we're
57 * selecting by biography, then we have to keep the various
58 * lag/drag/void counters for each identity.
59 * -------------------------------------------------------------------------- */
60 typedef struct _counter {
65 int prim; // total size of 'inherently used' closures
66 int not_used; // total size of 'never used' closures
67 int used; // total size of 'used at least once' closures
68 int void_total; // current total size of 'destroyed without being used' closures
69 int drag_total; // current total size of 'used at least once and waiting to die'
72 struct _counter *next;
76 initLDVCtr( counter *ctr )
79 ctr->c.ldv.not_used = 0;
81 ctr->c.ldv.void_total = 0;
82 ctr->c.ldv.drag_total = 0;
86 double time; // the time in MUT time when the census is made
91 // for LDV profiling, when just displaying by LDV
99 static Census *censuses = NULL;
100 static nat n_censuses = 0;
103 static void aggregateCensusInfo( void );
106 static void dumpCensus( Census *census );
108 /* -----------------------------------------------------------------------------
109 Closure Type Profiling;
111 PROBABLY TOTALLY OUT OF DATE -- ToDo (SDM)
112 -------------------------------------------------------------------------- */
114 #ifdef DEBUG_HEAP_PROF
115 static char *type_names[] = {
121 , "CONSTR_NOCAF_STATIC"
159 , "MUT_ARR_PTRS_FROZEN"
173 #endif /* DEBUG_HEAP_PROF */
175 /* -----------------------------------------------------------------------------
176 * Find the "closure identity", which is a unique pointer reresenting
177 * the band to which this closure's heap space is attributed in the
179 * ------------------------------------------------------------------------- */
181 closureIdentity( StgClosure *p )
183 switch (RtsFlags.ProfFlags.doHeapProfile) {
187 return p->header.prof.ccs;
189 return p->header.prof.ccs->cc->module;
191 return get_itbl(p)->prof.closure_desc;
193 return get_itbl(p)->prof.closure_type;
194 case HEAP_BY_RETAINER:
195 // AFAIK, the only closures in the heap which might not have a
196 // valid retainer set are DEAD_WEAK closures.
197 if (isRetainerSetFieldValid(p))
198 return retainerSetOf(p);
203 case HEAP_BY_INFOPTR:
204 return (void *)((StgClosure *)p)->header.info;
205 case HEAP_BY_CLOSURE_TYPE:
206 return type_names[get_itbl(p)->type];
210 barf("closureIdentity");
214 /* --------------------------------------------------------------------------
215 * Profiling type predicates
216 * ----------------------------------------------------------------------- */
218 static inline rtsBool
219 doingLDVProfiling( void )
221 return (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
222 || RtsFlags.ProfFlags.bioSelector != NULL);
225 static inline rtsBool
226 doingRetainerProfiling( void )
228 return (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER
229 || RtsFlags.ProfFlags.retainerSelector != NULL);
233 // Precesses a closure 'c' being destroyed whose size is 'size'.
234 // Make sure that LDV_recordDead() is not invoked on 'inherently used' closures
235 // such as TSO; they should not be involved in computing dragNew or voidNew.
237 // Even though era is checked in both LdvCensusForDead() and
238 // LdvCensusKillAll(), we still need to make sure that era is > 0 because
239 // LDV_recordDead() may be called from elsewhere in the runtime system. E.g.,
240 // when a thunk is replaced by an indirection object.
244 LDV_recordDead( StgClosure *c, nat size )
250 if (era > 0 && closureSatisfiesConstraints(c)) {
251 size -= sizeofW(StgProfHeader);
252 if ((LDVW((c)) & LDV_STATE_MASK) == LDV_STATE_CREATE) {
253 t = (LDVW((c)) & LDV_CREATE_MASK) >> LDV_SHIFT;
255 if (RtsFlags.ProfFlags.bioSelector == NULL) {
256 censuses[t].void_total += (int)size;
257 censuses[era].void_total -= (int)size;
259 id = closureIdentity(c);
260 ctr = lookupHashTable(censuses[t].hash, (StgWord)id);
261 ASSERT( ctr != NULL );
262 ctr->c.ldv.void_total += (int)size;
263 ctr = lookupHashTable(censuses[era].hash, (StgWord)id);
265 ctr = arenaAlloc(censuses[era].arena, sizeof(counter));
267 insertHashTable(censuses[era].hash, (StgWord)id, ctr);
269 ctr->next = censuses[era].ctrs;
270 censuses[era].ctrs = ctr;
272 ctr->c.ldv.void_total -= (int)size;
276 t = LDVW((c)) & LDV_LAST_MASK;
278 if (RtsFlags.ProfFlags.bioSelector == NULL) {
279 censuses[t+1].drag_total += size;
280 censuses[era].drag_total -= size;
283 id = closureIdentity(c);
284 ctr = lookupHashTable(censuses[t+1].hash, (StgWord)id);
285 ASSERT( ctr != NULL );
286 ctr->c.ldv.drag_total += (int)size;
287 ctr = lookupHashTable(censuses[era].hash, (StgWord)id);
289 ctr = arenaAlloc(censuses[era].arena, sizeof(counter));
291 insertHashTable(censuses[era].hash, (StgWord)id, ctr);
293 ctr->next = censuses[era].ctrs;
294 censuses[era].ctrs = ctr;
296 ctr->c.ldv.drag_total -= (int)size;
304 /* --------------------------------------------------------------------------
305 * Initialize censuses[era];
306 * ----------------------------------------------------------------------- */
308 initEra(Census *census)
310 census->hash = allocHashTable();
312 census->arena = newArena();
314 census->not_used = 0;
317 census->void_total = 0;
318 census->drag_total = 0;
321 /* --------------------------------------------------------------------------
322 * Increases era by 1 and initialize census[era].
323 * Reallocates gi[] and increases its size if needed.
324 * ----------------------------------------------------------------------- */
329 if (doingLDVProfiling()) {
332 if (era == max_era) {
333 prog_belch("maximum number of censuses reached; use +RTS -i to reduce");
334 stg_exit(EXIT_FAILURE);
337 if (era == n_censuses) {
339 censuses = stgReallocBytes(censuses, sizeof(Census) * n_censuses,
345 initEra( &censuses[era] );
348 /* -----------------------------------------------------------------------------
349 * DEBUG heap profiling, by info table
350 * -------------------------------------------------------------------------- */
352 #ifdef DEBUG_HEAP_PROF
355 void initProfiling1( void )
359 void initProfiling2( void )
364 void endProfiling( void )
368 #endif /* DEBUG_HEAP_PROF */
370 /* --------------------------------------------------------------------------
371 * Initialize the heap profilier
372 * ----------------------------------------------------------------------- */
374 initHeapProfiling(void)
376 if (! RtsFlags.ProfFlags.doHeapProfile) {
381 if (doingLDVProfiling() && doingRetainerProfiling()) {
382 prog_belch("cannot mix -hb and -hr");
387 // we only count eras if we're doing LDV profiling. Otherwise era
390 if (doingLDVProfiling()) {
398 { // max_era = 2^LDV_SHIFT
401 for (p = 0; p < LDV_SHIFT; p++)
406 censuses = stgMallocBytes(sizeof(Census) * n_censuses, "initHeapProfiling");
408 initEra( &censuses[era] );
410 fprintf(hp_file, "JOB \"%s", prog_argv[0]);
415 for(count = 1; count < prog_argc; count++)
416 fprintf(hp_file, " %s", prog_argv[count]);
417 fprintf(hp_file, " +RTS ");
418 for(count = 0; count < rts_argc; count++)
419 fprintf(hp_file, "%s ", rts_argv[count]);
420 fprintf(hp_file, "\n");
422 #endif /* PROFILING */
424 fprintf(hp_file, "\"\n" );
426 fprintf(hp_file, "DATE \"%s\"\n", time_str());
428 fprintf(hp_file, "SAMPLE_UNIT \"seconds\"\n");
429 fprintf(hp_file, "VALUE_UNIT \"bytes\"\n");
431 fprintf(hp_file, "BEGIN_SAMPLE 0.00\n");
432 fprintf(hp_file, "END_SAMPLE 0.00\n");
434 #ifdef DEBUG_HEAP_PROF
435 DEBUG_LoadSymbols(prog_argv[0]);
439 if (doingRetainerProfiling()) {
440 initRetainerProfiling();
448 endHeapProfiling(void)
452 if (! RtsFlags.ProfFlags.doHeapProfile) {
457 if (doingRetainerProfiling()) {
458 endRetainerProfiling();
463 if (doingLDVProfiling()) {
466 aggregateCensusInfo();
467 for (t = 1; t < era; t++) {
468 dumpCensus( &censuses[t] );
473 seconds = mut_user_time();
474 fprintf(hp_file, "BEGIN_SAMPLE %0.2f\n", seconds);
475 fprintf(hp_file, "END_SAMPLE %0.2f\n", seconds);
483 fprint_ccs(FILE *fp, CostCentreStack *ccs, nat max_length)
485 char buf[max_length+1];
490 // MAIN on its own gets printed as "MAIN", otherwise we ignore MAIN.
491 if (ccs == CCS_MAIN) {
496 fprintf(fp, "(%d)", ccs->ccsID);
498 // keep printing components of the stack until we run out of space
499 // in the buffer. If we run out of space, end with "...".
500 for (; ccs != NULL && ccs != CCS_MAIN; ccs = ccs->prevStack) {
502 // CAF cost centres print as M.CAF, but we leave the module
503 // name out of all the others to save space.
504 if (!strcmp(ccs->cc->label,"CAF")) {
506 written = snprintf(buf+next_offset,
507 (int)max_length-3-(int)next_offset,
508 "%s.CAF", ccs->cc->module);
510 written = sprintf(buf+next_offset,
511 "%s.CAF", ccs->cc->module);
514 if (ccs->prevStack != NULL && ccs->prevStack != CCS_MAIN) {
520 written = snprintf(buf+next_offset,
521 (int)max_length-3-(int)next_offset,
522 template, ccs->cc->label);
524 written = sprintf(buf+next_offset,
525 template, ccs->cc->label);
529 if (next_offset+written >= max_length-4) {
530 sprintf(buf+max_length-4, "...");
533 next_offset += written;
536 fprintf(fp, "%s", buf);
541 strMatchesSelector( char* str, char* sel )
544 // fprintf(stderr, "str_matches_selector %s %s\n", str, sel);
546 // Compare str against wherever we've got to in sel.
548 while (*p != '\0' && *sel != ',' && *sel != '\0' && *p == *sel) {
551 // Match if all of str used and have reached the end of a sel fragment.
552 if (*p == '\0' && (*sel == ',' || *sel == '\0'))
555 // No match. Advance sel to the start of the next elem.
556 while (*sel != ',' && *sel != '\0') sel++;
557 if (*sel == ',') sel++;
559 /* Run out of sel ?? */
560 if (*sel == '\0') return rtsFalse;
564 /* -----------------------------------------------------------------------------
565 * Figure out whether a closure should be counted in this census, by
566 * testing against all the specified constraints.
567 * -------------------------------------------------------------------------- */
569 closureSatisfiesConstraints( StgClosure* p )
571 #ifdef DEBUG_HEAP_PROF
576 // The CCS has a selected field to indicate whether this closure is
577 // deselected by not being mentioned in the module, CC, or CCS
579 if (!p->header.prof.ccs->selected) {
583 if (RtsFlags.ProfFlags.descrSelector) {
584 b = strMatchesSelector( (get_itbl((StgClosure *)p))->prof.closure_desc,
585 RtsFlags.ProfFlags.descrSelector );
586 if (!b) return rtsFalse;
588 if (RtsFlags.ProfFlags.typeSelector) {
589 b = strMatchesSelector( (get_itbl((StgClosure *)p))->prof.closure_type,
590 RtsFlags.ProfFlags.typeSelector );
591 if (!b) return rtsFalse;
593 if (RtsFlags.ProfFlags.retainerSelector) {
596 // We must check that the retainer set is valid here. One
597 // reason it might not be valid is if this closure is a
598 // a newly deceased weak pointer (i.e. a DEAD_WEAK), since
599 // these aren't reached by the retainer profiler's traversal.
600 if (isRetainerSetFieldValid((StgClosure *)p)) {
601 rs = retainerSetOf((StgClosure *)p);
603 for (i = 0; i < rs->num; i++) {
604 b = strMatchesSelector( rs->element[i]->cc->label,
605 RtsFlags.ProfFlags.retainerSelector );
606 if (b) return rtsTrue;
613 #endif /* PROFILING */
616 /* -----------------------------------------------------------------------------
617 * Aggregate the heap census info for biographical profiling
618 * -------------------------------------------------------------------------- */
621 aggregateCensusInfo( void )
625 counter *c, *d, *ctrs;
628 if (!doingLDVProfiling()) return;
630 // Aggregate the LDV counters when displaying by biography.
631 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV) {
632 int void_total, drag_total;
634 // Now we compute void_total and drag_total for each census
637 for (t = 1; t < era; t++) { // note: start at 1, not 0
638 void_total += censuses[t].void_total;
639 drag_total += censuses[t].drag_total;
640 censuses[t].void_total = void_total;
641 censuses[t].drag_total = drag_total;
642 ASSERT( censuses[t].void_total <= censuses[t].not_used );
643 ASSERT( censuses[t].drag_total <= censuses[t].used );
649 // otherwise... we're doing a heap profile that is restricted to
650 // some combination of lag, drag, void or use. We've kept all the
651 // census info for all censuses so far, but we still need to
652 // aggregate the counters forwards.
655 acc = allocHashTable();
658 for (t = 1; t < era; t++) {
660 // first look through all the counters we're aggregating
661 for (c = ctrs; c != NULL; c = c->next) {
662 // if one of the totals is non-zero, then this closure
663 // type must be present in the heap at this census time...
664 d = lookupHashTable(censuses[t].hash, (StgWord)c->identity);
667 // if this closure identity isn't present in the
668 // census for this time period, then our running
669 // totals *must* be zero.
670 ASSERT(c->c.ldv.void_total == 0 && c->c.ldv.drag_total == 0);
672 // fprintCCS(stderr,c->identity);
673 // fprintf(stderr," census=%d void_total=%d drag_total=%d\n",
674 // t, c->c.ldv.void_total, c->c.ldv.drag_total);
676 d->c.ldv.void_total += c->c.ldv.void_total;
677 d->c.ldv.drag_total += c->c.ldv.drag_total;
678 c->c.ldv.void_total = d->c.ldv.void_total;
679 c->c.ldv.drag_total = d->c.ldv.drag_total;
681 ASSERT( c->c.ldv.void_total >= 0 );
682 ASSERT( c->c.ldv.drag_total >= 0 );
686 // now look through the counters in this census to find new ones
687 for (c = censuses[t].ctrs; c != NULL; c = c->next) {
688 d = lookupHashTable(acc, (StgWord)c->identity);
690 d = arenaAlloc( arena, sizeof(counter) );
692 insertHashTable( acc, (StgWord)c->identity, d );
693 d->identity = c->identity;
696 d->c.ldv.void_total = c->c.ldv.void_total;
697 d->c.ldv.drag_total = c->c.ldv.drag_total;
699 ASSERT( c->c.ldv.void_total >= 0 );
700 ASSERT( c->c.ldv.drag_total >= 0 );
704 freeHashTable(acc, NULL);
709 /* -----------------------------------------------------------------------------
710 * Print out the results of a heap census.
711 * -------------------------------------------------------------------------- */
713 dumpCensus( Census *census )
718 fprintf(hp_file, "BEGIN_SAMPLE %0.2f\n", census->time);
721 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV) {
722 fprintf(hp_file, "VOID\t%u\n", census->void_total * sizeof(W_));
723 fprintf(hp_file, "LAG\t%u\n",
724 (census->not_used - census->void_total) * sizeof(W_));
725 fprintf(hp_file, "USE\t%u\n",
726 (census->used - census->drag_total) * sizeof(W_));
727 fprintf(hp_file, "INHERENT_USE\t%u\n",
728 census->prim * sizeof(W_));
729 fprintf(hp_file, "DRAG\t%u\n", census->drag_total *
731 fprintf(hp_file, "END_SAMPLE %0.2f\n", census->time);
736 for (ctr = census->ctrs; ctr != NULL; ctr = ctr->next) {
739 if (RtsFlags.ProfFlags.bioSelector != NULL) {
741 if (strMatchesSelector("lag", RtsFlags.ProfFlags.bioSelector))
742 count += ctr->c.ldv.not_used - ctr->c.ldv.void_total;
743 if (strMatchesSelector("drag", RtsFlags.ProfFlags.bioSelector))
744 count += ctr->c.ldv.drag_total;
745 if (strMatchesSelector("void", RtsFlags.ProfFlags.bioSelector))
746 count += ctr->c.ldv.void_total;
747 if (strMatchesSelector("use", RtsFlags.ProfFlags.bioSelector))
748 count += ctr->c.ldv.used - ctr->c.ldv.drag_total;
752 count = ctr->c.resid;
755 ASSERT( count >= 0 );
757 if (count == 0) continue;
759 #ifdef DEBUG_HEAP_PROF
760 switch (RtsFlags.ProfFlags.doHeapProfile) {
761 case HEAP_BY_INFOPTR:
762 fprintf(hp_file, "%s", lookupGHCName(ctr->identity));
764 case HEAP_BY_CLOSURE_TYPE:
765 fprintf(hp_file, "%s", (char *)ctr->identity);
771 switch (RtsFlags.ProfFlags.doHeapProfile) {
773 fprint_ccs(hp_file, (CostCentreStack *)ctr->identity, 25);
778 fprintf(hp_file, "%s", (char *)ctr->identity);
780 case HEAP_BY_RETAINER:
782 RetainerSet *rs = (RetainerSet *)ctr->identity;
784 // it might be the distinguished retainer set rs_MANY:
785 if (rs == &rs_MANY) {
786 fprintf(hp_file, "MANY");
790 // Mark this retainer set by negating its id, because it
791 // has appeared in at least one census. We print the
792 // values of all such retainer sets into the log file at
793 // the end. A retainer set may exist but not feature in
794 // any censuses if it arose as the intermediate retainer
795 // set for some closure during retainer set calculation.
799 // report in the unit of bytes: * sizeof(StgWord)
800 printRetainerSetShort(hp_file, rs);
804 barf("dumpCensus; doHeapProfile");
808 fprintf(hp_file, "\t%d\n", count * sizeof(W_));
811 fprintf(hp_file, "END_SAMPLE %0.2f\n", census->time);
814 /* -----------------------------------------------------------------------------
815 * Code to perform a heap census.
816 * -------------------------------------------------------------------------- */
818 heapCensusChain( Census *census, bdescr *bd )
828 for (; bd != NULL; bd = bd->link) {
830 while (p < bd->free) {
831 info = get_itbl((StgClosure *)p);
834 switch (info->type) {
840 case IND_OLDGEN_PERM:
842 case SE_CAF_BLACKHOLE:
847 case CONSTR_CHARLIKE:
861 size = sizeW_fromITBL(info);
872 size = sizeW_fromITBL(info);
875 case THUNK_1_0: /* ToDo - shouldn't be here */
876 case THUNK_0_1: /* " ditto " */
878 size = sizeofW(StgHeader) + MIN_UPD_SIZE;
883 size = pap_sizeW((StgPAP *)p);
887 size = ap_stack_sizeW((StgAP_STACK *)p);
892 size = arr_words_sizeW(stgCast(StgArrWords*,p));
896 case MUT_ARR_PTRS_FROZEN:
898 size = mut_arr_ptrs_sizeW((StgMutArrPtrs *)p);
903 #ifdef DEBUG_HEAP_PROF
904 size = tso_sizeW((StgTSO *)p);
907 if (RtsFlags.ProfFlags.includeTSOs) {
908 size = tso_sizeW((StgTSO *)p);
911 // Skip this TSO and move on to the next object
912 p += tso_sizeW((StgTSO *)p);
923 #ifdef DEBUG_HEAP_PROF
926 // subtract the profiling overhead
927 real_size = size - sizeofW(StgProfHeader);
930 if (closureSatisfiesConstraints((StgClosure*)p)) {
932 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV) {
934 census->prim += real_size;
935 else if ((LDVW(p) & LDV_STATE_MASK) == LDV_STATE_CREATE)
936 census->not_used += real_size;
938 census->used += real_size;
942 identity = closureIdentity((StgClosure *)p);
944 if (identity != NULL) {
945 ctr = lookupHashTable( census->hash, (StgWord)identity );
948 if (RtsFlags.ProfFlags.bioSelector != NULL) {
950 ctr->c.ldv.prim += real_size;
951 else if ((LDVW(p) & LDV_STATE_MASK) == LDV_STATE_CREATE)
952 ctr->c.ldv.not_used += real_size;
954 ctr->c.ldv.used += real_size;
958 ctr->c.resid += real_size;
961 ctr = arenaAlloc( census->arena, sizeof(counter) );
963 insertHashTable( census->hash, (StgWord)identity, ctr );
964 ctr->identity = identity;
965 ctr->next = census->ctrs;
969 if (RtsFlags.ProfFlags.bioSelector != NULL) {
971 ctr->c.ldv.prim = real_size;
972 else if ((LDVW(p) & LDV_STATE_MASK) == LDV_STATE_CREATE)
973 ctr->c.ldv.not_used = real_size;
975 ctr->c.ldv.used = real_size;
979 ctr->c.resid = real_size;
997 census = &censuses[era];
998 census->time = mut_user_time();
1000 // calculate retainer sets if necessary
1002 if (doingRetainerProfiling()) {
1008 stat_startHeapCensus();
1011 // Traverse the heap, collecting the census info
1013 // First the small_alloc_list: we have to fix the free pointer at
1014 // the end by calling tidyAllocatedLists() first.
1015 tidyAllocateLists();
1016 heapCensusChain( census, small_alloc_list );
1018 // Now traverse the heap in each generation/step.
1019 if (RtsFlags.GcFlags.generations == 1) {
1020 heapCensusChain( census, g0s0->to_blocks );
1022 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1023 for (s = 0; s < generations[g].n_steps; s++) {
1024 heapCensusChain( census, generations[g].steps[s].blocks );
1025 // Are we interested in large objects? might be
1026 // confusing to include the stack in a heap profile.
1027 heapCensusChain( census, generations[g].steps[s].large_objects );
1032 // dump out the census info
1034 // We can't generate any info for LDV profiling until
1035 // the end of the run...
1036 if (!doingLDVProfiling())
1037 dumpCensus( census );
1039 dumpCensus( census );
1043 // free our storage, unless we're keeping all the census info for
1044 // future restriction by biography.
1046 if (RtsFlags.ProfFlags.bioSelector == NULL)
1049 freeHashTable( census->hash, NULL/* don't free the elements */ );
1050 arenaFree( census->arena );
1051 census->hash = NULL;
1052 census->arena = NULL;
1055 // we're into the next time period now
1059 stat_endHeapCensus();
1063 #endif /* PROFILING || DEBUG_HEAP_PROF */