[project @ 2004-08-13 13:04:50 by simonmar]
[ghc-hetmet.git] / ghc / rts / ProfHeap.c
index f9487e9..49b3813 100644 (file)
@@ -1,7 +1,7 @@
 /* -----------------------------------------------------------------------------
- * $Id: ProfHeap.c,v 1.35 2001/12/12 15:59:33 simonmar Exp $
+ * $Id: ProfHeap.c,v 1.53 2004/08/13 13:10:25 simonmar Exp $
  *
- * (c) The GHC Team, 1998-2000
+ * (c) The GHC Team, 1998-2003
  *
  * Support for heap profiling
  *
 #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.
@@ -69,7 +73,7 @@ typedef struct _counter {
     struct _counter *next;
 } counter;
 
-static inline void
+STATIC_INLINE void
 initLDVCtr( counter *ctr )
 {
     ctr->c.ldv.prim = 0;
@@ -93,8 +97,8 @@ 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 +129,8 @@ static char *type_names[] = {
     , "THUNK_SELECTOR"
 
     , "BCO"
-    , "AP_UPD"
+    , "AP_STACK"
+    , "AP"
 
     , "PAP"
 
@@ -144,7 +149,6 @@ static char *type_names[] = {
     , "UPDATE_FRAME"
     , "CATCH_FRAME"
     , "STOP_FRAME"
-    , "SEQ_FRAME"
 
     , "BLACKHOLE"
     , "BLACKHOLE_BQ"
@@ -174,7 +178,7 @@ 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) {
@@ -212,14 +216,14 @@ 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
@@ -246,6 +250,7 @@ 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) {
@@ -301,7 +306,7 @@ LDV_recordDead( StgClosure *c, nat size )
 /* --------------------------------------------------------------------------
  * Initialize censuses[era];
  * ----------------------------------------------------------------------- */
-static inline void
+STATIC_INLINE void
 initEra(Census *census)
 {
     census->hash  = allocHashTable();
@@ -327,7 +332,8 @@ nextEra( void )
        era++;
 
        if (era == max_era) {
-           barf("maximum number of censuses reached; use +RTS -i to reduce");
+           prog_belch("maximum number of censuses reached; use +RTS -i to reduce");
+           stg_exit(EXIT_FAILURE);
        }
        
        if (era == n_censuses) {
@@ -337,7 +343,7 @@ nextEra( void )
        }
     }
 #endif // PROFILING
-       
+
     initEra( &censuses[era] );
 }
 
@@ -347,6 +353,7 @@ nextEra( void )
 
 #ifdef DEBUG_HEAP_PROF
 FILE *hp_file;
+static char *hp_filename;
 
 void initProfiling1( void )
 {
@@ -354,6 +361,20 @@ 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) {
+      fprintf(stderr, "Can't open profiling report file %s\n", 
+             hp_filename);
+      RtsFlags.ProfFlags.doHeapProfile = 0;
+      return;
+    }
+  }
+  
   initHeapProfiling();
 }
 
@@ -363,6 +384,16 @@ 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
  * ----------------------------------------------------------------------- */
@@ -403,17 +434,17 @@ 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 +455,11 @@ 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 +498,32 @@ 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) {
@@ -491,6 +533,9 @@ fprint_ccs(FILE *fp, CostCentreStack *ccs, nat max_length)
 
     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) {
@@ -498,21 +543,16 @@ 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 {
@@ -555,6 +595,7 @@ rtsBool
 closureSatisfiesConstraints( StgClosure* p )
 {
 #ifdef DEBUG_HEAP_PROF
+    (void)p;   /* keep gcc -Wall happy */
     return rtsTrue;
 #else
    rtsBool b;
@@ -579,12 +620,18 @@ 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;
@@ -695,7 +742,7 @@ 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) {
@@ -708,7 +755,7 @@ 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
@@ -788,7 +835,7 @@ 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);
 }
 
 /* -----------------------------------------------------------------------------
@@ -806,6 +853,15 @@ 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);
@@ -817,6 +873,7 @@ 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:
@@ -842,6 +899,10 @@ heapCensusChain( Census *census, bdescr *bd )
                break;
                
            case BCO:
+               prim = rtsTrue;
+               size = bco_sizeW((StgBCO *)p);
+               break;
+
            case MVAR:
            case WEAK:
            case FOREIGN:
@@ -858,10 +919,14 @@ 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;
@@ -876,9 +941,20 @@ 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");
            }
@@ -973,8 +1049,14 @@ 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 {
@@ -983,7 +1065,7 @@ 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 );
          }
       }
   }