X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=rts%2FStats.c;h=2e1561313515557b6c930658e760f2b5819b6746;hb=4cc1ed3b25dd2c565341665b3b4175dfdbc4328f;hp=461dabddf80a68587d445b0941dcbe5cd7d3f4f9;hpb=a75a8790410ce3ffb439bfd0b7c3999e7df72eb1;p=ghc-hetmet.git

diff --git a/rts/Stats.c b/rts/Stats.c
index 461dabd..2e15613 100644
--- a/rts/Stats.c
+++ b/rts/Stats.c
@@ -17,8 +17,6 @@
 #include "Profiling.h"
 #include "GetTime.h"
 #include "GC.h"
-#include "GCUtils.h"
-#include "Evac.h"
 
 #if USE_PAPI
 #include "Papi.h"
@@ -68,6 +66,7 @@ static Ticks HCe_start_time, HCe_tot_time = 0;   // heap census prof elap time
 static lnat MaxResidency = 0;     // in words; for stats only
 static lnat AvgResidency = 0;
 static lnat ResidencySamples = 0; // for stats only
+static lnat MaxSlop = 0;
 
 static lnat GC_start_faults = 0, GC_end_faults = 0;
 
@@ -132,8 +131,56 @@ mut_user_time_during_heap_census( void )
 }
 #endif /* PROFILING */
 
+// initStats0() has no dependencies, it can be called right at the beginning
 void
-initStats(void)
+initStats0(void)
+{
+    ElapsedTimeStart = 0;
+
+    InitUserTime     = 0;
+    InitElapsedTime  = 0;
+    InitElapsedStamp = 0;
+
+    MutUserTime      = 0;
+    MutElapsedTime   = 0;
+    MutElapsedStamp  = 0;
+
+    ExitUserTime     = 0;
+    ExitElapsedTime  = 0;
+
+    GC_tot_alloc     = 0;
+    GC_tot_copied    = 0;
+    GC_par_max_copied = 0;
+    GC_par_avg_copied = 0;
+    GC_start_time = 0;
+    GC_tot_time  = 0;
+    GCe_start_time = 0;
+    GCe_tot_time = 0;
+
+#ifdef PROFILING
+    RP_start_time  = 0;
+    RP_tot_time  = 0;
+    RPe_start_time = 0;
+    RPe_tot_time = 0;
+
+    HC_start_time = 0;
+    HC_tot_time = 0;
+    HCe_start_time = 0;
+    HCe_tot_time = 0;
+#endif
+
+    MaxResidency = 0;
+    AvgResidency = 0;
+    ResidencySamples = 0;
+    MaxSlop = 0;
+
+    GC_start_faults = 0;
+    GC_end_faults = 0;
+}    
+
+// initStats1() can be called after setupRtsFlags()
+void
+initStats1 (void)
 {
     nat i;
   
@@ -153,7 +200,7 @@ initStats(void)
 	GC_coll_times[i] = 0;
 	GC_coll_etimes[i] = 0;
     }
-}    
+}
 
 /* -----------------------------------------------------------------------------
    Initialisation time...
@@ -298,7 +345,7 @@ stat_startGC(void)
 
 void
 stat_endGC (lnat alloc, lnat live, lnat copied, lnat gen,
-            lnat max_copied, lnat avg_copied)
+            lnat max_copied, lnat avg_copied, lnat slop)
 {
     if (RtsFlags.GcFlags.giveStats != NO_GC_STATS) {
 	Ticks time, etime, gc_time, gc_etime;
@@ -353,6 +400,8 @@ stat_endGC (lnat alloc, lnat live, lnat copied, lnat gen,
 	    ResidencySamples++;
 	    AvgResidency += live;
 	}
+
+        if (slop > MaxSlop) MaxSlop = slop;
     }
 
     if (rub_bell) {
@@ -501,6 +550,7 @@ StgInt TOTAL_CALLS=1;
   statsPrintf("  (SLOW_CALLS_" #arity ") %% of (TOTAL_CALLS) : %.1f%%\n", \
 	      SLOW_CALLS_##arity * 100.0/TOTAL_CALLS)
 
+extern lnat hw_alloc_blocks;
 
 void
 stat_exit(int alloc)
@@ -545,8 +595,13 @@ stat_exit(int alloc)
 		statsPrintf("%16s bytes maximum residency (%ld sample(s))\n",
 			temp, ResidencySamples);
 	    }
-	    statsPrintf("%16ld MB total memory in use\n\n", 
-		    mblocks_allocated * MBLOCK_SIZE / (1024 * 1024));
+
+	    ullong_format_string(MaxSlop*sizeof(W_), temp, rtsTrue/*commas*/);
+	    statsPrintf("%16s bytes maximum slop\n", temp);
+
+	    statsPrintf("%16ld MB total memory in use (%ld MB lost due to fragmentation)\n\n", 
+                        mblocks_allocated * MBLOCK_SIZE_W / (1024 * 1024 / sizeof(W_)),
+                        (mblocks_allocated * MBLOCK_SIZE_W - hw_alloc_blocks * BLOCK_SIZE_W) / (1024 * 1024 / sizeof(W_)));
 
 	    /* Print garbage collections in each gen */
 	    for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
@@ -638,6 +693,21 @@ stat_exit(int alloc)
 #if USE_PAPI
             papi_stats_report();
 #endif
+#if defined(THREADED_RTS) && defined(PROF_SPIN)
+            {
+                nat g, s;
+                
+                statsPrintf("recordMutableGen_sync: %"FMT_Word64"\n", recordMutableGen_sync.spin);
+                statsPrintf("gc_alloc_block_sync: %"FMT_Word64"\n", gc_alloc_block_sync.spin);
+                statsPrintf("whitehole_spin: %"FMT_Word64"\n", whitehole_spin);
+                for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
+                    for (s = 0; s < generations[g].n_steps; s++) {
+                        statsPrintf("gen[%d].steps[%d].sync_todo: %"FMT_Word64"\n", g, s, generations[g].steps[s].sync_todo.spin);
+                        statsPrintf("gen[%d].steps[%d].sync_large_objects: %"FMT_Word64"\n", g, s, generations[g].steps[s].sync_large_objects.spin);
+                    }
+                }
+            }
+#endif
 	}
 
 	if (RtsFlags.GcFlags.giveStats == ONELINE_GC_STATS) {
@@ -655,22 +725,6 @@ stat_exit(int alloc)
 		    TICK_TO_DBL(GC_tot_time), TICK_TO_DBL(GCe_tot_time));
 	}
 
-#if defined(THREADED_RTS) && defined(PROF_SPIN)
-        {
-            nat g, s;
-            
-            statsPrintf("recordMutableGen_sync: %"FMT_Word64"\n", recordMutableGen_sync.spin);
-            statsPrintf("gc_alloc_block_sync: %"FMT_Word64"\n", gc_alloc_block_sync.spin);
-            statsPrintf("whitehole_spin: %"FMT_Word64"\n", whitehole_spin);
-            for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
-                for (s = 0; s < generations[g].n_steps; s++) {
-                    statsPrintf("gen[%d].steps[%d].sync_todo: %"FMT_Word64"\n", g, s, generations[g].steps[s].sync_todo.spin);
-                    statsPrintf("gen[%d].steps[%d].sync_large_objects: %"FMT_Word64"\n", g, s, generations[g].steps[s].sync_large_objects.spin);
-                }
-            }
-        }
-#endif
-
 	statsFlush();
 	statsClose();
     }