lag/drag/void: add an extra assertion, and some commentary
[ghc-hetmet.git] / ghc / rts / ProfHeap.c
index 73ca19d..312bee7 100644 (file)
@@ -1,7 +1,6 @@
 /* -----------------------------------------------------------------------------
- * $Id: ProfHeap.c,v 1.33 2001/12/12 14:31:42 simonmar Exp $
  *
- * (c) The GHC Team, 1998-2000
+ * (c) The GHC Team, 1998-2003
  *
  * Support for heap profiling
  *
 #include "ProfHeap.h"
 #include "Stats.h"
 #include "Hash.h"
-#include "StrHash.h"
 #include "RetainerProfile.h"
 #include "LdvProfile.h"
 #include "Arena.h"
 #include "Printer.h"
 
+#include <string.h>
+#include <stdlib.h>
+#include <math.h>
+
 /* -----------------------------------------------------------------------------
  * era stores the current time period.  It is the same as the
  * number of censuses that have been performed.
@@ -43,7 +45,7 @@
  * 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 +71,7 @@ typedef struct _counter {
     struct _counter *next;
 } counter;
 
-static inline void
+STATIC_INLINE void
 initLDVCtr( counter *ctr )
 {
     ctr->c.ldv.prim = 0;
@@ -93,8 +95,8 @@ 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 +127,8 @@ static char *type_names[] = {
     , "THUNK_SELECTOR"
 
     , "BCO"
-    , "AP_UPD"
+    , "AP_STACK"
+    , "AP"
 
     , "PAP"
 
@@ -144,20 +147,19 @@ static char *type_names[] = {
     , "UPDATE_FRAME"
     , "CATCH_FRAME"
     , "STOP_FRAME"
-    , "SEQ_FRAME"
 
     , "BLACKHOLE"
-    , "BLACKHOLE_BQ"
     , "MVAR"
 
     , "ARR_WORDS"
 
-    , "MUT_ARR_PTRS"
+    , "MUT_ARR_PTRS_CLEAN"
+    , "MUT_ARR_PTRS_DIRTY"
     , "MUT_ARR_PTRS_FROZEN"
-    , "MUT_VAR"
+    , "MUT_VAR_CLEAN"
+    , "MUT_VAR_DIRTY"
 
     , "WEAK"
-    , "FOREIGN"
   
     , "TSO"
 
@@ -174,7 +176,7 @@ 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) {
@@ -212,20 +214,20 @@ 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
            || RtsFlags.ProfFlags.retainerSelector != NULL);
 }
-#endif // PROFILING
+#endif /* PROFILING */
 
 // Precesses a closure 'c' being destroyed whose size is 'size'.
 // Make sure that LDV_recordDead() is not invoked on 'inherently used' closures
@@ -246,12 +248,14 @@ 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) {
                if (RtsFlags.ProfFlags.bioSelector == NULL) {
                    censuses[t].void_total   += (int)size;
                    censuses[era].void_total -= (int)size;
+                   ASSERT(censuses[t].void_total < censuses[t].not_used);
                } else {
                    id = closureIdentity(c);
                    ctr = lookupHashTable(censuses[t].hash, (StgWord)id);
@@ -301,7 +305,7 @@ LDV_recordDead( StgClosure *c, nat size )
 /* --------------------------------------------------------------------------
  * Initialize censuses[era];
  * ----------------------------------------------------------------------- */
-static inline void
+STATIC_INLINE void
 initEra(Census *census)
 {
     census->hash  = allocHashTable();
@@ -327,7 +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) {
@@ -336,8 +341,8 @@ nextEra( void )
                                       "nextEra");
        }
     }
-#endif // PROFILING
-       
+#endif /* PROFILING */
+
     initEra( &censuses[era] );
 }
 
@@ -347,6 +352,7 @@ nextEra( void )
 
 #ifdef DEBUG_HEAP_PROF
 FILE *hp_file;
+static char *hp_filename;
 
 void initProfiling1( void )
 {
@@ -354,6 +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();
 }
 
@@ -363,6 +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));
+}
+
 /* --------------------------------------------------------------------------
  * Initialize the heap profilier
  * ----------------------------------------------------------------------- */
@@ -375,8 +405,8 @@ initHeapProfiling(void)
 
 #ifdef PROFILING
     if (doingLDVProfiling() && doingRetainerProfiling()) {
-       prog_belch("cannot mix -hb and -hr");
-       stg_exit(1);
+       errorBelch("cannot mix -hb and -hr");
+       stg_exit(EXIT_FAILURE);
     }
 #endif
 
@@ -403,17 +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 */
 
@@ -424,11 +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
@@ -467,21 +497,30 @@ 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];
-    nat next_offset = 0;
-    nat written;
-    char *template;
+    char buf[max_length+1], *p, *buf_end;
 
     // MAIN on its own gets printed as "MAIN", otherwise we ignore MAIN.
     if (ccs == CCS_MAIN) {
@@ -489,6 +528,11 @@ fprint_ccs(FILE *fp, CostCentreStack *ccs, nat max_length)
        return;
     }
 
+    fprintf(fp, "(%ld)", 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) {
@@ -496,36 +540,29 @@ 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 {
-           next_offset += written;
        }
     }
     fprintf(fp, "%s", buf);
 }
-#endif // PROFILING
+#endif /* PROFILING */
 
 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;
@@ -553,6 +590,7 @@ rtsBool
 closureSatisfiesConstraints( StgClosure* p )
 {
 #ifdef DEBUG_HEAP_PROF
+    (void)p;   /* keep gcc -Wall happy */
     return rtsTrue;
 #else
    rtsBool b;
@@ -577,12 +615,18 @@ closureSatisfiesConstraints( StgClosure* p )
    if (RtsFlags.ProfFlags.retainerSelector) {
        RetainerSet *rs;
        nat i;
-       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;
+       // 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;
@@ -610,6 +654,15 @@ aggregateCensusInfo( void )
        int void_total, drag_total;
 
        // Now we compute void_total and drag_total for each census
+       // After the program has finished, the void_total field of
+       // each census contains the count of words that were *created*
+       // in this era and were eventually void.  Conversely, if a
+       // void closure was destroyed in this era, it will be
+       // represented by a negative count of words in void_total.
+       //
+       // To get the count of live words that are void at each
+       // census, just propagate the void_total count forwards:
+
        void_total = 0;
        drag_total = 0;
        for (t = 1; t < era; t++) { // note: start at 1, not 0
@@ -617,8 +670,15 @@ 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 );
+           // should be true because: void_total is the count of
+           // live words that are void at this census, which *must*
+           // be less than the number of live words that have not
+           // been used yet.
+
+           ASSERT( censuses[t].drag_total <= censuses[t].used );
+           // similar reasoning as above.
        }
        
        return;
@@ -647,8 +707,8 @@ 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;
@@ -693,20 +753,20 @@ 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) {
-       fprintf(hp_file, "VOID\t%u\n", census->void_total * sizeof(W_));
-       fprintf(hp_file, "LAG\t%u\n", 
-               (census->not_used - census->void_total) * sizeof(W_));
-       fprintf(hp_file, "USE\t%u\n", 
-               (census->used - census->drag_total) * sizeof(W_));
-       fprintf(hp_file, "INHERENT_USE\t%u\n", 
-               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);
+      fprintf(hp_file, "VOID\t%lu\n", (unsigned long)(census->void_total) * sizeof(W_));
+       fprintf(hp_file, "LAG\t%lu\n", 
+               (unsigned long)(census->not_used - census->void_total) * sizeof(W_));
+       fprintf(hp_file, "USE\t%lu\n", 
+               (unsigned long)(census->used - census->drag_total) * sizeof(W_));
+       fprintf(hp_file, "INHERENT_USE\t%lu\n", 
+               (unsigned long)(census->prim) * sizeof(W_));
+       fprintf(hp_file, "DRAG\t%lu\n",
+               (unsigned long)(census->drag_total) * sizeof(W_));
+       printSample(rtsFalse, census->time);
        return;
     }
 #endif
@@ -748,7 +808,7 @@ 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:
@@ -783,10 +843,10 @@ dumpCensus( Census *census )
        }
 #endif
 
-       fprintf(hp_file, "\t%d\n", count * sizeof(W_));
+       fprintf(hp_file, "\t%lu\n", (unsigned long)count * sizeof(W_));
     }
 
-    fprintf(hp_file, "END_SAMPLE %0.2f\n", census->time);
+    printSample(rtsFalse, census->time);
 }
 
 /* -----------------------------------------------------------------------------
@@ -804,6 +864,15 @@ 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);
@@ -811,16 +880,31 @@ heapCensusChain( Census *census, bdescr *bd )
            
            switch (info->type) {
 
+           case THUNK:
+               size = thunk_sizeW_fromITBL(info);
+               break;
+
+           case THUNK_1_1:
+           case THUNK_0_2:
+           case THUNK_2_0:
+               size = sizeofW(StgThunkHeader) + 2;
+               break;
+
+           case THUNK_1_0:
+           case THUNK_0_1:
+           case THUNK_SELECTOR:
+               size = sizeofW(StgThunkHeader) + 1;
+               break;
+
            case CONSTR:
            case FUN:
-           case THUNK:
            case IND_PERM:
+           case IND_OLDGEN:
            case IND_OLDGEN_PERM:
            case CAF_BLACKHOLE:
            case SE_CAF_BLACKHOLE:
            case SE_BLACKHOLE:
            case BLACKHOLE:
-           case BLACKHOLE_BQ:
            case CONSTR_INTLIKE:
            case CONSTR_CHARLIKE:
            case FUN_1_0:
@@ -828,9 +912,6 @@ heapCensusChain( Census *census, bdescr *bd )
            case FUN_1_1:
            case FUN_0_2:
            case FUN_2_0:
-           case THUNK_1_1:
-           case THUNK_0_2:
-           case THUNK_2_0:
            case CONSTR_1_0:
            case CONSTR_0_1:
            case CONSTR_1_1:
@@ -838,47 +919,95 @@ heapCensusChain( Census *census, bdescr *bd )
            case CONSTR_2_0:
                size = sizeW_fromITBL(info);
                break;
-               
+
+           case IND:
+               // Special case/Delicate Hack: INDs don't normally
+               // appear, since we're doing this heap census right
+               // after GC.  However, GarbageCollect() also does
+               // resurrectThreads(), which can update some
+               // blackholes when it calls raiseAsync() on the
+               // resurrected threads.  So we know that any IND will
+               // be the size of a BLACKHOLE.
+               size = BLACKHOLE_sizeW();
+               break;
+
            case BCO:
+               prim = rtsTrue;
+               size = bco_sizeW((StgBCO *)p);
+               break;
+
            case MVAR:
            case WEAK:
-           case FOREIGN:
            case STABLE_NAME:
-           case MUT_VAR:
-           case MUT_CONS:
+           case MUT_VAR_CLEAN:
+           case MUT_VAR_DIRTY:
                prim = rtsTrue;
                size = sizeW_fromITBL(info);
                break;
 
-           case THUNK_1_0:             /* ToDo - shouldn't be here */
-           case THUNK_0_1:             /* "  ditto  " */
-           case THUNK_SELECTOR:
-               size = sizeofW(StgHeader) + MIN_UPD_SIZE;
+           case AP:
+               size = ap_sizeW((StgAP *)p);
                break;
 
            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;
                size = arr_words_sizeW(stgCast(StgArrWords*,p));
                break;
                
-           case MUT_ARR_PTRS:
+           case MUT_ARR_PTRS_CLEAN:
+           case MUT_ARR_PTRS_DIRTY:
            case MUT_ARR_PTRS_FROZEN:
+           case MUT_ARR_PTRS_FROZEN0:
                prim = rtsTrue;
                size = mut_arr_ptrs_sizeW((StgMutArrPtrs *)p);
                break;
                
            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
+
+           case TREC_HEADER: 
+               prim = rtsTrue;
+               size = sizeofW(StgTRecHeader);
+               break;
+
+           case TVAR_WAIT_QUEUE:
+               prim = rtsTrue;
+               size = sizeofW(StgTVarWaitQueue);
+               break;
                
+           case TVAR:
+               prim = rtsTrue;
+               size = sizeofW(StgTVar);
+               break;
+               
+           case TREC_CHUNK:
+               prim = rtsTrue;
+               size = sizeofW(StgTRecChunk);
+               break;
+
            default:
-               barf("heapCensus");
+               barf("heapCensus, unknown object: %d", info->type);
            }
            
            identity = NULL;
@@ -971,17 +1100,23 @@ 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 );
+      heapCensusChain( census, g0s0->blocks );
   } else {
       for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
          for (s = 0; s < generations[g].n_steps; s++) {
              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 );
          }
       }
   }