X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Frts%2FProfHeap.c;h=932c0694233b4e37247476fcc69a5a4e6f6928c8;hb=3160a6546bf6e57d0d5ef7bc084fcbcb10297ff6;hp=b57cd68ce137f17b3c41fc902a02520698ee0f2a;hpb=20977132895cce0550b9957c7c3b35d1a212f8b4;p=ghc-hetmet.git diff --git a/ghc/rts/ProfHeap.c b/ghc/rts/ProfHeap.c index b57cd68..932c069 100644 --- a/ghc/rts/ProfHeap.c +++ b/ghc/rts/ProfHeap.c @@ -1,7 +1,6 @@ /* ----------------------------------------------------------------------------- - * $Id: ProfHeap.c,v 1.31 2001/11/28 17:45:13 simonmar Exp $ * - * (c) The GHC Team, 1998-2000 + * (c) The GHC Team, 1998-2003 * * Support for heap profiling * @@ -30,6 +29,10 @@ #include "Arena.h" #include "Printer.h" +#include +#include +#include + /* ----------------------------------------------------------------------------- * era stores the current time period. It is the same as the * number of censuses that have been performed. @@ -37,13 +40,13 @@ * RESTRICTION: * era must be no longer than LDV_SHIFT (15 or 30) bits. * Invariants: - * era is initialized to 0 in initHeapProfiling(). + * era is initialized to 1 in initHeapProfiling(). * * max_era is initialized to 2^LDV_SHIFT in initHeapProfiling(). * When era reaches max_era, the profiling stops because a closure can * store only up to (max_era - 1) as its creation or last use time. * -------------------------------------------------------------------------- */ -nat era; +unsigned int era; static nat max_era; /* ----------------------------------------------------------------------------- @@ -69,7 +72,7 @@ typedef struct _counter { struct _counter *next; } counter; -static inline void +STATIC_INLINE void initLDVCtr( counter *ctr ) { ctr->c.ldv.prim = 0; @@ -93,8 +96,8 @@ typedef struct { int drag_total; } Census; -Census *censuses = NULL; -nat n_censuses = 0; +static Census *censuses = NULL; +static nat n_censuses = 0; #ifdef PROFILING static void aggregateCensusInfo( void ); @@ -125,7 +128,8 @@ static char *type_names[] = { , "THUNK_SELECTOR" , "BCO" - , "AP_UPD" + , "AP_STACK" + , "AP" , "PAP" @@ -144,7 +148,6 @@ static char *type_names[] = { , "UPDATE_FRAME" , "CATCH_FRAME" , "STOP_FRAME" - , "SEQ_FRAME" , "BLACKHOLE" , "BLACKHOLE_BQ" @@ -174,27 +177,34 @@ static char *type_names[] = { * the band to which this closure's heap space is attributed in the * heap profile. * ------------------------------------------------------------------------- */ -static inline void * +STATIC_INLINE void * closureIdentity( StgClosure *p ) { switch (RtsFlags.ProfFlags.doHeapProfile) { #ifdef PROFILING case HEAP_BY_CCS: - return ((StgClosure *)p)->header.prof.ccs; + return p->header.prof.ccs; case HEAP_BY_MOD: - return ((StgClosure *)p)->header.prof.ccs->cc->module; + return p->header.prof.ccs->cc->module; case HEAP_BY_DESCR: - return (get_itbl((StgClosure *)p))->prof.closure_desc; + return get_itbl(p)->prof.closure_desc; case HEAP_BY_TYPE: - return (get_itbl((StgClosure *)p))->prof.closure_type; + return get_itbl(p)->prof.closure_type; case HEAP_BY_RETAINER: - return retainerSetOf((StgClosure *)p); + // AFAIK, the only closures in the heap which might not have a + // valid retainer set are DEAD_WEAK closures. + if (isRetainerSetFieldValid(p)) + return retainerSetOf(p); + else + return NULL; + #else // DEBUG case HEAP_BY_INFOPTR: return (void *)((StgClosure *)p)->header.info; case HEAP_BY_CLOSURE_TYPE: return type_names[get_itbl(p)->type]; + #endif default: barf("closureIdentity"); @@ -205,14 +215,14 @@ closureIdentity( StgClosure *p ) * Profiling type predicates * ----------------------------------------------------------------------- */ #ifdef PROFILING -static inline rtsBool +STATIC_INLINE rtsBool doingLDVProfiling( void ) { return (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV || RtsFlags.ProfFlags.bioSelector != NULL); } -static inline rtsBool +STATIC_INLINE rtsBool doingRetainerProfiling( void ) { return (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER @@ -239,6 +249,7 @@ LDV_recordDead( StgClosure *c, nat size ) if (era > 0 && closureSatisfiesConstraints(c)) { size -= sizeofW(StgProfHeader); + ASSERT(LDVW(c) != 0); if ((LDVW((c)) & LDV_STATE_MASK) == LDV_STATE_CREATE) { t = (LDVW((c)) & LDV_CREATE_MASK) >> LDV_SHIFT; if (t < era) { @@ -294,7 +305,7 @@ LDV_recordDead( StgClosure *c, nat size ) /* -------------------------------------------------------------------------- * Initialize censuses[era]; * ----------------------------------------------------------------------- */ -static inline void +STATIC_INLINE void initEra(Census *census) { census->hash = allocHashTable(); @@ -320,7 +331,8 @@ nextEra( void ) era++; if (era == max_era) { - barf("maximum number of censuses reached; use +RTS -i to reduce"); + errorBelch("maximum number of censuses reached; use +RTS -i to reduce"); + stg_exit(EXIT_FAILURE); } if (era == n_censuses) { @@ -330,7 +342,7 @@ nextEra( void ) } } #endif // PROFILING - + initEra( &censuses[era] ); } @@ -340,6 +352,7 @@ nextEra( void ) #ifdef DEBUG_HEAP_PROF FILE *hp_file; +static char *hp_filename; void initProfiling1( void ) { @@ -347,6 +360,20 @@ void initProfiling1( void ) void initProfiling2( void ) { + if (RtsFlags.ProfFlags.doHeapProfile) { + /* Initialise the log file name */ + hp_filename = stgMallocBytes(strlen(prog_name) + 6, "hpFileName"); + sprintf(hp_filename, "%s.hp", prog_name); + + /* open the log file */ + if ((hp_file = fopen(hp_filename, "w")) == NULL) { + debugBelch("Can't open profiling report file %s\n", + hp_filename); + RtsFlags.ProfFlags.doHeapProfile = 0; + return; + } + } + initHeapProfiling(); } @@ -356,6 +383,16 @@ void endProfiling( void ) } #endif /* DEBUG_HEAP_PROF */ +static void +printSample(rtsBool beginSample, StgDouble sampleValue) +{ + StgDouble fractionalPart, integralPart; + fractionalPart = modf(sampleValue, &integralPart); + fprintf(hp_file, "%s %d.%02d\n", + (beginSample ? "BEGIN_SAMPLE" : "END_SAMPLE"), + (int)integralPart, (int)(fractionalPart * 100 + 0.5)); +} + /* -------------------------------------------------------------------------- * Initialize the heap profilier * ----------------------------------------------------------------------- */ @@ -366,6 +403,13 @@ initHeapProfiling(void) return 0; } +#ifdef PROFILING + if (doingLDVProfiling() && doingRetainerProfiling()) { + errorBelch("cannot mix -hb and -hr"); + stg_exit(1); + } +#endif + // we only count eras if we're doing LDV profiling. Otherwise era // is fixed at zero. #ifdef PROFILING @@ -389,17 +433,17 @@ initHeapProfiling(void) initEra( &censuses[era] ); - fprintf(hp_file, "JOB \"%s", prog_argv[0]); + /* initProfilingLogFile(); */ + fprintf(hp_file, "JOB \"%s", prog_name); #ifdef PROFILING { int count; for(count = 1; count < prog_argc; count++) fprintf(hp_file, " %s", prog_argv[count]); - fprintf(hp_file, " +RTS "); + fprintf(hp_file, " +RTS"); for(count = 0; count < rts_argc; count++) - fprintf(hp_file, "%s ", rts_argv[count]); - fprintf(hp_file, "\n"); + fprintf(hp_file, " %s", rts_argv[count]); } #endif /* PROFILING */ @@ -410,11 +454,11 @@ initHeapProfiling(void) fprintf(hp_file, "SAMPLE_UNIT \"seconds\"\n"); fprintf(hp_file, "VALUE_UNIT \"bytes\"\n"); - fprintf(hp_file, "BEGIN_SAMPLE 0.00\n"); - fprintf(hp_file, "END_SAMPLE 0.00\n"); + printSample(rtsTrue, 0); + printSample(rtsFalse, 0); #ifdef DEBUG_HEAP_PROF - DEBUG_LoadSymbols(prog_argv[0]); + DEBUG_LoadSymbols(prog_name); #endif #ifdef PROFILING @@ -442,17 +486,9 @@ endHeapProfiling(void) #endif #ifdef PROFILING - // Note: - // We do not need to perform a major garbage collection because all the - // closures created since the last census will not affect the profiling - // statistics anyhow. - if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV) - LdvCensusKillAll(); -#endif - -#ifdef PROFILING if (doingLDVProfiling()) { nat t; + LdvCensusKillAll(); aggregateCensusInfo(); for (t = 1; t < era; t++) { dumpCensus( &censuses[t] ); @@ -461,21 +497,32 @@ endHeapProfiling(void) #endif seconds = mut_user_time(); - fprintf(hp_file, "BEGIN_SAMPLE %0.2f\n", seconds); - fprintf(hp_file, "END_SAMPLE %0.2f\n", seconds); + printSample(rtsTrue, seconds); + printSample(rtsFalse, seconds); fclose(hp_file); } #ifdef PROFILING +static size_t +buf_append(char *p, const char *q, char *end) +{ + int m; + + for (m = 0; p < end; p++, q++, m++) { + *p = *q; + if (*q == '\0') { break; } + } + return m; +} + static void fprint_ccs(FILE *fp, CostCentreStack *ccs, nat max_length) { - char buf[max_length+1]; + char buf[max_length+1], *p, *buf_end; nat next_offset = 0; nat written; - char *template; // MAIN on its own gets printed as "MAIN", otherwise we ignore MAIN. if (ccs == CCS_MAIN) { @@ -483,6 +530,11 @@ fprint_ccs(FILE *fp, CostCentreStack *ccs, nat max_length) return; } + fprintf(fp, "(%d)", ccs->ccsID); + + p = buf; + buf_end = buf + max_length + 1; + // keep printing components of the stack until we run out of space // in the buffer. If we run out of space, end with "...". for (; ccs != NULL && ccs != CCS_MAIN; ccs = ccs->prevStack) { @@ -490,21 +542,16 @@ fprint_ccs(FILE *fp, CostCentreStack *ccs, nat max_length) // CAF cost centres print as M.CAF, but we leave the module // name out of all the others to save space. if (!strcmp(ccs->cc->label,"CAF")) { - written = snprintf(buf+next_offset, - (int)max_length-3-(int)next_offset, - "%s.CAF", ccs->cc->module); + p += buf_append(p, ccs->cc->module, buf_end); + p += buf_append(p, ".CAF", buf_end); } else { if (ccs->prevStack != NULL && ccs->prevStack != CCS_MAIN) { - template = "%s/"; - } else { - template = "%s"; + p += buf_append(p, "/", buf_end); } - written = snprintf(buf+next_offset, - (int)max_length-3-(int)next_offset, - template, ccs->cc->label); + p += buf_append(p, ccs->cc->label, buf_end); } - - if (next_offset+written >= max_length-4) { + + if (p >= buf_end) { sprintf(buf+max_length-4, "..."); break; } else { @@ -513,12 +560,13 @@ fprint_ccs(FILE *fp, CostCentreStack *ccs, nat max_length) } fprintf(fp, "%s", buf); } +#endif // PROFILING -static rtsBool -str_matches_selector( char* str, char* sel ) +rtsBool +strMatchesSelector( char* str, char* sel ) { char* p; - // fprintf(stderr, "str_matches_selector %s %s\n", str, sel); + // debugBelch("str_matches_selector %s %s\n", str, sel); while (1) { // Compare str against wherever we've got to in sel. p = str; @@ -537,7 +585,6 @@ str_matches_selector( char* str, char* sel ) if (*sel == '\0') return rtsFalse; } } -#endif // PROFILING /* ----------------------------------------------------------------------------- * Figure out whether a closure should be counted in this census, by @@ -547,38 +594,43 @@ rtsBool closureSatisfiesConstraints( StgClosure* p ) { #ifdef DEBUG_HEAP_PROF + (void)p; /* keep gcc -Wall happy */ return rtsTrue; #else rtsBool b; - if (RtsFlags.ProfFlags.modSelector) { - b = str_matches_selector( ((StgClosure *)p)->header.prof.ccs->cc->module, - RtsFlags.ProfFlags.modSelector ); - if (!b) return rtsFalse; + + // The CCS has a selected field to indicate whether this closure is + // deselected by not being mentioned in the module, CC, or CCS + // selectors. + if (!p->header.prof.ccs->selected) { + return rtsFalse; } + if (RtsFlags.ProfFlags.descrSelector) { - b = str_matches_selector( (get_itbl((StgClosure *)p))->prof.closure_desc, + b = strMatchesSelector( (get_itbl((StgClosure *)p))->prof.closure_desc, RtsFlags.ProfFlags.descrSelector ); if (!b) return rtsFalse; } if (RtsFlags.ProfFlags.typeSelector) { - b = str_matches_selector( (get_itbl((StgClosure *)p))->prof.closure_type, + b = strMatchesSelector( (get_itbl((StgClosure *)p))->prof.closure_type, RtsFlags.ProfFlags.typeSelector ); if (!b) return rtsFalse; } - if (RtsFlags.ProfFlags.ccSelector) { - b = str_matches_selector( ((StgClosure *)p)->header.prof.ccs->cc->label, - RtsFlags.ProfFlags.ccSelector ); - if (!b) return rtsFalse; - } if (RtsFlags.ProfFlags.retainerSelector) { RetainerSet *rs; nat i; - rs = retainerSetOf((StgClosure *)p); - if (rs != NULL) { - for (i = 0; i < rs->num; i++) { - b = str_matches_selector( rs->element[i]->cc->label, - RtsFlags.ProfFlags.retainerSelector ); - if (b) return rtsTrue; + // We must check that the retainer set is valid here. One + // reason it might not be valid is if this closure is a + // a newly deceased weak pointer (i.e. a DEAD_WEAK), since + // these aren't reached by the retainer profiler's traversal. + if (isRetainerSetFieldValid((StgClosure *)p)) { + rs = retainerSetOf((StgClosure *)p); + if (rs != NULL) { + for (i = 0; i < rs->num; i++) { + b = strMatchesSelector( rs->element[i]->cc->label, + RtsFlags.ProfFlags.retainerSelector ); + if (b) return rtsTrue; + } } } return rtsFalse; @@ -613,8 +665,8 @@ aggregateCensusInfo( void ) drag_total += censuses[t].drag_total; censuses[t].void_total = void_total; censuses[t].drag_total = drag_total; - ASSERT( censuses[t].void_total < censuses[t].not_used ); - ASSERT( censuses[t].drag_total < censuses[t].used ); + ASSERT( censuses[t].void_total <= censuses[t].not_used ); + ASSERT( censuses[t].drag_total <= censuses[t].used ); } return; @@ -643,8 +695,8 @@ aggregateCensusInfo( void ) // totals *must* be zero. ASSERT(c->c.ldv.void_total == 0 && c->c.ldv.drag_total == 0); - // fprintCCS(stderr,c->identity); - // fprintf(stderr," census=%d void_total=%d drag_total=%d\n", + // debugCCS(c->identity); + // debugBelch(" census=%d void_total=%d drag_total=%d\n", // t, c->c.ldv.void_total, c->c.ldv.drag_total); } else { d->c.ldv.void_total += c->c.ldv.void_total; @@ -689,7 +741,7 @@ dumpCensus( Census *census ) counter *ctr; int count; - fprintf(hp_file, "BEGIN_SAMPLE %0.2f\n", census->time); + printSample(rtsTrue, census->time); #ifdef PROFILING if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV) { @@ -702,7 +754,7 @@ dumpCensus( Census *census ) census->prim * sizeof(W_)); fprintf(hp_file, "DRAG\t%u\n", census->drag_total * sizeof(W_)); - fprintf(hp_file, "END_SAMPLE %0.2f\n", census->time); + printSample(rtsFalse, census->time); return; } #endif @@ -712,13 +764,13 @@ dumpCensus( Census *census ) #ifdef PROFILING if (RtsFlags.ProfFlags.bioSelector != NULL) { count = 0; - if (str_matches_selector("lag", RtsFlags.ProfFlags.bioSelector)) + if (strMatchesSelector("lag", RtsFlags.ProfFlags.bioSelector)) count += ctr->c.ldv.not_used - ctr->c.ldv.void_total; - if (str_matches_selector("drag", RtsFlags.ProfFlags.bioSelector)) + if (strMatchesSelector("drag", RtsFlags.ProfFlags.bioSelector)) count += ctr->c.ldv.drag_total; - if (str_matches_selector("void", RtsFlags.ProfFlags.bioSelector)) + if (strMatchesSelector("void", RtsFlags.ProfFlags.bioSelector)) count += ctr->c.ldv.void_total; - if (str_matches_selector("use", RtsFlags.ProfFlags.bioSelector)) + if (strMatchesSelector("use", RtsFlags.ProfFlags.bioSelector)) count += ctr->c.ldv.used - ctr->c.ldv.drag_total; } else #endif @@ -744,7 +796,7 @@ dumpCensus( Census *census ) #ifdef PROFILING switch (RtsFlags.ProfFlags.doHeapProfile) { case HEAP_BY_CCS: - fprint_ccs(hp_file, (CostCentreStack *)ctr->identity, 30); + fprint_ccs(hp_file, (CostCentreStack *)ctr->identity, 25); break; case HEAP_BY_MOD: case HEAP_BY_DESCR: @@ -782,7 +834,7 @@ dumpCensus( Census *census ) fprintf(hp_file, "\t%d\n", count * sizeof(W_)); } - fprintf(hp_file, "END_SAMPLE %0.2f\n", census->time); + printSample(rtsFalse, census->time); } /* ----------------------------------------------------------------------------- @@ -800,6 +852,15 @@ heapCensusChain( Census *census, bdescr *bd ) rtsBool prim; for (; bd != NULL; bd = bd->link) { + + // HACK: ignore pinned blocks, because they contain gaps. + // It's not clear exactly what we'd like to do here, since we + // can't tell which objects in the block are actually alive. + // Perhaps the whole block should be counted as SYSTEM memory. + if (bd->flags & BF_PINNED) { + continue; + } + p = bd->start; while (p < bd->free) { info = get_itbl((StgClosure *)p); @@ -811,6 +872,7 @@ heapCensusChain( Census *census, bdescr *bd ) case FUN: case THUNK: case IND_PERM: + case IND_OLDGEN: case IND_OLDGEN_PERM: case CAF_BLACKHOLE: case SE_CAF_BLACKHOLE: @@ -836,6 +898,10 @@ heapCensusChain( Census *census, bdescr *bd ) break; case BCO: + prim = rtsTrue; + size = bco_sizeW((StgBCO *)p); + break; + case MVAR: case WEAK: case FOREIGN: @@ -852,10 +918,14 @@ heapCensusChain( Census *census, bdescr *bd ) size = sizeofW(StgHeader) + MIN_UPD_SIZE; break; + case AP: case PAP: - case AP_UPD: size = pap_sizeW((StgPAP *)p); break; + + case AP_STACK: + size = ap_stack_sizeW((StgAP_STACK *)p); + break; case ARR_WORDS: prim = rtsTrue; @@ -870,9 +940,20 @@ heapCensusChain( Census *census, bdescr *bd ) case TSO: prim = rtsTrue; +#ifdef DEBUG_HEAP_PROF size = tso_sizeW((StgTSO *)p); break; - +#else + if (RtsFlags.ProfFlags.includeTSOs) { + size = tso_sizeW((StgTSO *)p); + break; + } else { + // Skip this TSO and move on to the next object + p += tso_sizeW((StgTSO *)p); + continue; + } +#endif + default: barf("heapCensus"); } @@ -967,8 +1048,14 @@ heapCensus( void ) stat_startHeapCensus(); #endif - // traverse the heap, collecting the census info + // Traverse the heap, collecting the census info + + // First the small_alloc_list: we have to fix the free pointer at + // the end by calling tidyAllocatedLists() first. + tidyAllocateLists(); heapCensusChain( census, small_alloc_list ); + + // Now traverse the heap in each generation/step. if (RtsFlags.GcFlags.generations == 1) { heapCensusChain( census, g0s0->to_blocks ); } else { @@ -977,7 +1064,7 @@ heapCensus( void ) heapCensusChain( census, generations[g].steps[s].blocks ); // Are we interested in large objects? might be // confusing to include the stack in a heap profile. - // heapCensusChain( census, generations[g].steps[s].large_objects ); + heapCensusChain( census, generations[g].steps[s].large_objects ); } } }