[project @ 2003-10-01 10:57:39 by wolfgang]
[ghc-hetmet.git] / ghc / rts / ProfHeap.c
index 38a9aaa..27a1a41 100644 (file)
@@ -1,7 +1,7 @@
 /* -----------------------------------------------------------------------------
- * $Id: ProfHeap.c,v 1.34 2001/12/12 15:01:25 simonmar Exp $
+ * $Id: ProfHeap.c,v 1.48 2003/09/23 15:38:36 simonmar Exp $
  *
- * (c) The GHC Team, 1998-2000
+ * (c) The GHC Team, 1998-2003
  *
  * Support for heap profiling
  *
@@ -30,6 +30,9 @@
 #include "Arena.h"
 #include "Printer.h"
 
+#include <string.h>
+#include <stdlib.h>
+
 /* -----------------------------------------------------------------------------
  * era stores the current time period.  It is the same as the
  * number of censuses that have been performed.
@@ -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"
@@ -327,7 +330,8 @@ 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 +341,7 @@ nextEra( void )
        }
     }
 #endif // PROFILING
-       
+
     initEra( &censuses[era] );
 }
 
@@ -403,7 +407,7 @@ initHeapProfiling(void)
 
     initEra( &censuses[era] );
 
-    fprintf(hp_file, "JOB \"%s", prog_argv[0]);
+    fprintf(hp_file, "JOB \"%s", prog_name);
 
 #ifdef PROFILING
     {
@@ -428,7 +432,7 @@ initHeapProfiling(void)
     fprintf(hp_file, "END_SAMPLE 0.00\n");
 
 #ifdef DEBUG_HEAP_PROF
-    DEBUG_LoadSymbols(prog_argv[0]);
+    DEBUG_LoadSymbols(prog_name);
 #endif
 
 #ifdef PROFILING
@@ -475,13 +479,24 @@ endHeapProfiling(void)
 
 
 #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) {
@@ -489,6 +504,11 @@ 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) {
@@ -496,21 +516,16 @@ 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 {
@@ -577,12 +592,18 @@ 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;
@@ -748,7 +769,7 @@ 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:
@@ -804,6 +825,15 @@ 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);
@@ -815,6 +845,7 @@ 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:
@@ -840,6 +871,10 @@ heapCensusChain( Census *census, bdescr *bd )
                break;
                
            case BCO:
+               prim = rtsTrue;
+               size = bco_sizeW((StgBCO *)p);
+               break;
+
            case MVAR:
            case WEAK:
            case FOREIGN:
@@ -856,10 +891,14 @@ 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;
@@ -874,9 +913,20 @@ 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");
            }
@@ -971,8 +1021,14 @@ 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 {
@@ -981,7 +1037,7 @@ 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 );
          }
       }
   }