[project @ 2002-12-05 23:49:43 by mthomas]
[ghc-hetmet.git] / ghc / rts / ProfHeap.c
index 2bd9e59..7df27e1 100644 (file)
@@ -1,5 +1,5 @@
 /* -----------------------------------------------------------------------------
- * $Id: ProfHeap.c,v 1.29 2001/11/28 15:01:02 simonmar Exp $
+ * $Id: ProfHeap.c,v 1.39 2002/11/01 11:05:46 simonmar Exp $
  *
  * (c) The GHC Team, 1998-2000
  *
 #include "RetainerProfile.h"
 #include "LdvProfile.h"
 #include "Arena.h"
-
-#ifdef DEBUG_HEAP_PROF
 #include "Printer.h"
-static void fprint_data(FILE *fp);
-#endif
+
+#include <string.h>
 
 /* -----------------------------------------------------------------------------
  * era stores the current time period.  It is the same as the
@@ -41,7 +39,7 @@ static void fprint_data(FILE *fp);
  * 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
@@ -97,8 +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 );
@@ -185,20 +183,27 @@ closureIdentity( StgClosure *p )
 
 #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");
@@ -338,7 +343,9 @@ nextEra( void )
     initEra( &censuses[era] );
 }
 
-/* -------------------------------------------------------------------------- */
+/* -----------------------------------------------------------------------------
+ * DEBUG heap profiling, by info table
+ * -------------------------------------------------------------------------- */
 
 #ifdef DEBUG_HEAP_PROF
 FILE *hp_file;
@@ -358,6 +365,9 @@ void endProfiling( void )
 }
 #endif /* DEBUG_HEAP_PROF */
 
+/* --------------------------------------------------------------------------
+ * Initialize the heap profilier
+ * ----------------------------------------------------------------------- */
 nat
 initHeapProfiling(void)
 {
@@ -365,6 +375,13 @@ initHeapProfiling(void)
         return 0;
     }
 
+#ifdef PROFILING
+    if (doingLDVProfiling() && doingRetainerProfiling()) {
+       prog_belch("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
@@ -441,17 +458,9 @@ 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 (RtsFlags.ProfFlags.bioSelector != NULL) {
+    if (doingLDVProfiling()) {
        nat t;
+       LdvCensusKillAll();
        aggregateCensusInfo();
        for (t = 1; t < era; t++) {
            dumpCensus( &censuses[t] );
@@ -482,6 +491,8 @@ fprint_ccs(FILE *fp, CostCentreStack *ccs, nat max_length)
        return;
     }
 
+    fprintf(fp, "(%d)", ccs->ccsID);
+
     // 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) {
@@ -489,18 +500,28 @@ 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")) {
+#ifdef HAVE_SNPRINTF
            written = snprintf(buf+next_offset, 
                               (int)max_length-3-(int)next_offset,
                               "%s.CAF", ccs->cc->module);
+#else
+           written = sprintf(buf+next_offset, 
+                              "%s.CAF", ccs->cc->module);
+#endif
        } else {
            if (ccs->prevStack != NULL && ccs->prevStack != CCS_MAIN) {
                template = "%s/";
            } else {
                template = "%s";
            }
+#ifdef HAVE_SNPRINTF
            written = snprintf(buf+next_offset, 
                               (int)max_length-3-(int)next_offset,
                               template, ccs->cc->label);
+#else
+           written = sprintf(buf+next_offset, 
+                              template, ccs->cc->label);
+#endif
        }
 
        if (next_offset+written >= max_length-4) {
@@ -512,9 +533,10 @@ 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);
@@ -537,39 +559,42 @@ str_matches_selector( char* str, char* sel )
    }
 }
 
-// Figure out whether a closure should be counted in this census, by
-// testing against all the specified constraints.
+/* -----------------------------------------------------------------------------
+ * Figure out whether a closure should be counted in this census, by
+ * testing against all the specified constraints.
+ * -------------------------------------------------------------------------- */
 rtsBool
 closureSatisfiesConstraints( StgClosure* p )
 {
+#ifdef DEBUG_HEAP_PROF
+    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,
+              b = strMatchesSelector( rs->element[i]->cc->label,
                                         RtsFlags.ProfFlags.retainerSelector );
               if (b) return rtsTrue;
           }
@@ -577,8 +602,8 @@ closureSatisfiesConstraints( StgClosure* p )
        return rtsFalse;
    }
    return rtsTrue;
-}
 #endif /* PROFILING */
+}
 
 /* -----------------------------------------------------------------------------
  * Aggregate the heap census info for biographical profiling
@@ -606,23 +631,8 @@ 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 );
-       }
-       
-       for (t = 1; t < era; t++) { // note: start at 1, not 0
-           fprintf(hp_file, "MARK %f\n", censuses[t].time);
-           fprintf(hp_file, "BEGIN_SAMPLE %f\n", censuses[t].time);
-           fprintf(hp_file, "VOID\t%u\n", censuses[t].void_total * sizeof(W_));
-           fprintf(hp_file, "LAG\t%u\n", 
-                   (censuses[t].not_used - censuses[t].void_total)
-                   * sizeof(W_));
-           fprintf(hp_file, "USE\t%u\n", 
-                   (censuses[t].used - censuses[t].drag_total) * sizeof(W_));
-           fprintf(hp_file, "INHERENT_USE\t%u\n", 
-                   censuses[t].prim * sizeof(W_));
-           fprintf(hp_file, "DRAG\t%u\n", censuses[t].drag_total * sizeof(W_));
-           fprintf(hp_file, "END_SAMPLE %f\n", censuses[t].time);
+           ASSERT( censuses[t].void_total <= censuses[t].not_used );
+           ASSERT( censuses[t].drag_total <= censuses[t].used );
        }
        
        return;
@@ -699,18 +709,34 @@ dumpCensus( Census *census )
 
     fprintf(hp_file, "BEGIN_SAMPLE %0.2f\n", 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);
+       return;
+    }
+#endif
+
     for (ctr = census->ctrs; ctr != NULL; ctr = ctr->next) {
 
 #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
@@ -725,10 +751,10 @@ dumpCensus( Census *census )
 #ifdef DEBUG_HEAP_PROF
        switch (RtsFlags.ProfFlags.doHeapProfile) {
        case HEAP_BY_INFOPTR:
-           fprint_data(hp_file);
+           fprintf(hp_file, "%s", lookupGHCName(ctr->identity));
            break;
        case HEAP_BY_CLOSURE_TYPE:
-           fprint_closure_types(hp_file);
+           fprintf(hp_file, "%s", (char *)ctr->identity);
            break;
        }
 #endif
@@ -736,7 +762,7 @@ 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:
@@ -955,10 +981,18 @@ heapCensus( void )
   }
 #endif
 
+#ifdef PROFILING
   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 {
@@ -998,7 +1032,9 @@ heapCensus( void )
   // we're into the next time period now
   nextEra();
 
+#ifdef PROFILING
   stat_endHeapCensus();
+#endif
 }    
 
 #endif /* PROFILING || DEBUG_HEAP_PROF */