X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Frts%2FGC.c;h=a13cd33afadd346ed79e20ac72235ea4e62887eb;hb=a1b4e3b88a6987deed7bb7f1bd870b30eef1b475;hp=bc8546a115fae765d8524b5a6f651702b6b917c1;hpb=c5cd2343c5a86c8cb5349823a9699b30a269f3e8;p=ghc-hetmet.git

diff --git a/ghc/rts/GC.c b/ghc/rts/GC.c
index bc8546a..a13cd33 100644
--- a/ghc/rts/GC.c
+++ b/ghc/rts/GC.c
@@ -104,6 +104,10 @@ static rtsBool major_gc;
  */
 static nat evac_gen;
 
+/* Whether to do eager promotion or not.
+ */
+static rtsBool eager_promotion;
+
 /* Weak pointers
  */
 StgWeak *old_weak_ptr_list; // also pending finaliser list
@@ -140,6 +144,14 @@ static lnat g0s0_pcnt_kept = 30; // percentage of g0s0 live at last minor GC
 static lnat thunk_selector_depth = 0;
 #define MAX_THUNK_SELECTOR_DEPTH 8
 
+/* Mut-list stats */
+#ifdef DEBUG
+static nat 
+    mutlist_MUTVARS,
+    mutlist_MUTARRS,
+    mutlist_OTHERS;
+#endif
+
 /* -----------------------------------------------------------------------------
    Static function declarations
    -------------------------------------------------------------------------- */
@@ -332,7 +344,7 @@ GarbageCollect ( void (*get_roots)(evac_fn), rtsBool force_major_gc )
 {
   bdescr *bd;
   step *stp;
-  lnat live, allocated, collected = 0, copied = 0, scavd_copied = 0;
+  lnat live, allocated, copied = 0, scavd_copied = 0;
   lnat oldgen_saved_blocks = 0;
   nat g, s, i;
 
@@ -363,6 +375,12 @@ GarbageCollect ( void (*get_roots)(evac_fn), rtsBool force_major_gc )
   memInventory();
 #endif
 
+#ifdef DEBUG
+  mutlist_MUTVARS = 0;
+  mutlist_MUTARRS = 0;
+  mutlist_OTHERS = 0;
+#endif
+
   // Init stats and print par specific (timing) info 
   PAR_TICKY_PAR_START();
 
@@ -571,6 +589,8 @@ GarbageCollect ( void (*get_roots)(evac_fn), rtsBool force_major_gc )
       mark_stack_bdescr = NULL;
   }
 
+  eager_promotion = rtsTrue; // for now
+
   /* -----------------------------------------------------------------------
    * follow all the roots that we know about:
    *   - mutable lists from each generation > N
@@ -792,9 +812,13 @@ GarbageCollect ( void (*get_roots)(evac_fn), rtsBool force_major_gc )
     // Count the mutable list as bytes "copied" for the purposes of
     // stats.  Every mutable list is copied during every GC.
     if (g > 0) {
+	nat mut_list_size = 0;
 	for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
-	    copied += (bd->free - bd->start) * sizeof(StgWord);
+	    mut_list_size += bd->free - bd->start;
 	}
+	copied +=  mut_list_size;
+
+	IF_DEBUG(gc, debugBelch("mut_list_size: %ld (%d vars, %d arrays, %d others)\n", mut_list_size * sizeof(W_), mutlist_MUTVARS, mutlist_MUTARRS, mutlist_OTHERS));
     }
 
     for (s = 0; s < generations[g].n_steps; s++) {
@@ -813,18 +837,6 @@ GarbageCollect ( void (*get_roots)(evac_fn), rtsBool force_major_gc )
       // for generations we collected... 
       if (g <= N) {
 
-	  // rough calculation of garbage collected, for stats output
-	  if (stp->is_compacted) {
-	      collected += (oldgen_saved_blocks - stp->n_old_blocks) * BLOCK_SIZE_W;
-	  } else {
-	      if (g == 0 && s == 0) {
-		  collected += countNurseryBlocks() * BLOCK_SIZE_W;
-		  collected += alloc_blocks;
-	      } else {
-		  collected += stp->n_old_blocks * BLOCK_SIZE_W;
-	      }
-	  }
-
 	/* free old memory and shift to-space into from-space for all
 	 * the collected steps (except the allocation area).  These
 	 * freed blocks will probaby be quickly recycled.
@@ -1142,7 +1154,9 @@ GarbageCollect ( void (*get_roots)(evac_fn), rtsBool force_major_gc )
   ACQUIRE_SM_LOCK;
   
   // send exceptions to any threads which were about to die 
+  RELEASE_SM_LOCK;
   resurrectThreads(resurrected_threads);
+  ACQUIRE_SM_LOCK;
 
   // Update the stable pointer hash table.
   updateStablePtrTable(major_gc);
@@ -1175,7 +1189,7 @@ GarbageCollect ( void (*get_roots)(evac_fn), rtsBool force_major_gc )
 #endif
 
   // ok, GC over: tell the stats department what happened. 
-  stat_endGC(allocated, collected, live, copied, scavd_copied, N);
+  stat_endGC(allocated, live, copied, scavd_copied, N);
 
 #if defined(RTS_USER_SIGNALS)
   // unblock signals again
@@ -1561,11 +1575,11 @@ copy(StgClosure *src, nat size, step *stp)
    * by evacuate()).
    */
   if (stp->gen_no < evac_gen) {
-#ifdef NO_EAGER_PROMOTION    
-    failed_to_evac = rtsTrue;
-#else
-    stp = &generations[evac_gen].steps[0];
-#endif
+      if (eager_promotion) {
+	  stp = &generations[evac_gen].steps[0];
+      } else {
+	  failed_to_evac = rtsTrue;
+      }
   }
 
   /* chain a new block onto the to-space for the destination step if
@@ -1611,11 +1625,11 @@ copy_noscav(StgClosure *src, nat size, step *stp)
    * by evacuate()).
    */
   if (stp->gen_no < evac_gen) {
-#ifdef NO_EAGER_PROMOTION    
-    failed_to_evac = rtsTrue;
-#else
-    stp = &generations[evac_gen].steps[0];
-#endif
+      if (eager_promotion) {
+	  stp = &generations[evac_gen].steps[0];
+      } else {
+	  failed_to_evac = rtsTrue;
+      }
   }
 
   /* chain a new block onto the to-space for the destination step if
@@ -1658,11 +1672,11 @@ copyPart(StgClosure *src, nat size_to_reserve, nat size_to_copy, step *stp)
 
   TICK_GC_WORDS_COPIED(size_to_copy);
   if (stp->gen_no < evac_gen) {
-#ifdef NO_EAGER_PROMOTION    
-    failed_to_evac = rtsTrue;
-#else
-    stp = &generations[evac_gen].steps[0];
-#endif
+      if (eager_promotion) {
+	  stp = &generations[evac_gen].steps[0];
+      } else {
+	  failed_to_evac = rtsTrue;
+      }
   }
 
   if (stp->hp + size_to_reserve >= stp->hpLim) {
@@ -1684,7 +1698,7 @@ copyPart(StgClosure *src, nat size_to_reserve, nat size_to_copy, step *stp)
   SET_EVACUAEE_FOR_LDV(src, size_to_reserve);
   // fill the slop
   if (size_to_reserve - size_to_copy_org > 0)
-    FILL_SLOP(stp->hp - 1, (int)(size_to_reserve - size_to_copy_org)); 
+    LDV_FILL_SLOP(stp->hp - 1, (int)(size_to_reserve - size_to_copy_org)); 
 #endif
   return (StgClosure *)dest;
 }
@@ -1739,11 +1753,11 @@ evacuate_large(StgPtr p)
    */
   stp = bd->step->to;
   if (stp->gen_no < evac_gen) {
-#ifdef NO_EAGER_PROMOTION    
-    failed_to_evac = rtsTrue;
-#else
-    stp = &generations[evac_gen].steps[0];
-#endif
+      if (eager_promotion) {
+	  stp = &generations[evac_gen].steps[0];
+      } else {
+	  failed_to_evac = rtsTrue;
+      }
   }
 
   bd->step = stp;
@@ -1929,7 +1943,8 @@ loop:
   
   switch (info->type) {
 
-  case MUT_VAR:
+  case MUT_VAR_CLEAN:
+  case MUT_VAR_DIRTY:
   case MVAR:
       return copy(q,sizeW_fromITBL(info),stp);
 
@@ -2003,11 +2018,15 @@ loop:
   case THUNK_SELECTOR:
     {
 	StgClosure *p;
+	const StgInfoTable *info_ptr;
 
 	if (thunk_selector_depth > MAX_THUNK_SELECTOR_DEPTH) {
 	    return copy(q,THUNK_SELECTOR_sizeW(),stp);
 	}
 
+	// stashed away for LDV profiling, see below
+	info_ptr = q->header.info;
+
 	p = eval_thunk_selector(info->layout.selector_offset,
 				(StgSelector *)q);
 
@@ -2020,6 +2039,13 @@ loop:
 	    val = evacuate(p);
 	    thunk_selector_depth--;
 
+#ifdef PROFILING
+	    // For the purposes of LDV profiling, we have destroyed
+	    // the original selector thunk.
+	    SET_INFO(q, info_ptr);
+	    LDV_RECORD_DEAD_FILL_SLOP_DYNAMIC(q);
+#endif
+
 	    // Update the THUNK_SELECTOR with an indirection to the
 	    // EVACUATED closure now at p.  Why do this rather than
 	    // upd_evacuee(q,p)?  Because we have an invariant that an
@@ -2029,12 +2055,10 @@ loop:
 	    SET_INFO(q, &stg_IND_info);
 	    ((StgInd *)q)->indirectee = p;
 
-#ifdef PROFILING
-	    // We store the size of the just evacuated object in the
-	    // LDV word so that the profiler can guess the position of
-	    // the next object later.
-	    SET_EVACUAEE_FOR_LDV(q, THUNK_SELECTOR_sizeW());
-#endif
+	    // For the purposes of LDV profiling, we have created an
+	    // indirection.
+	    LDV_RECORD_CREATE(q);
+
 	    return val;
 	}
     }
@@ -2099,7 +2123,8 @@ loop:
       // just copy the block 
       return copy_noscav(q,arr_words_sizeW((StgArrWords *)q),stp);
 
-  case MUT_ARR_PTRS:
+  case MUT_ARR_PTRS_CLEAN:
+  case MUT_ARR_PTRS_DIRTY:
   case MUT_ARR_PTRS_FROZEN:
   case MUT_ARR_PTRS_FROZEN0:
       // just copy the block 
@@ -2150,7 +2175,7 @@ loop:
     }
 
   case BLOCKED_FETCH:
-    ASSERT(sizeofW(StgBlockedFetch) >= MIN_NONUPD_SIZE);
+    ASSERT(sizeofW(StgBlockedFetch) >= MIN_PAYLOD_SIZE);
     to = copy(q,sizeofW(StgBlockedFetch),stp);
     IF_DEBUG(gc,
 	     debugBelch("@@ evacuate: %p (%s) to %p (%s)",
@@ -2161,7 +2186,7 @@ loop:
   case REMOTE_REF:
 # endif
   case FETCH_ME:
-    ASSERT(sizeofW(StgBlockedFetch) >= MIN_UPD_SIZE);
+    ASSERT(sizeofW(StgBlockedFetch) >= MIN_PAYLOAD_SIZE);
     to = copy(q,sizeofW(StgFetchMe),stp);
     IF_DEBUG(gc,
 	     debugBelch("@@ evacuate: %p (%s) to %p (%s)",
@@ -2169,7 +2194,7 @@ loop:
     return to;
 
   case FETCH_ME_BQ:
-    ASSERT(sizeofW(StgBlockedFetch) >= MIN_UPD_SIZE);
+    ASSERT(sizeofW(StgBlockedFetch) >= MIN_PAYLOAD_SIZE);
     to = copy(q,sizeofW(StgFetchMeBlockingQueue),stp);
     IF_DEBUG(gc,
 	     debugBelch("@@ evacuate: %p (%s) to %p (%s)",
@@ -2881,13 +2906,22 @@ scavenge(step *stp)
 	p += sizeofW(StgInd);
 	break;
 
-    case MUT_VAR:
-	evac_gen = 0;
+    case MUT_VAR_CLEAN:
+    case MUT_VAR_DIRTY: {
+	rtsBool saved_eager_promotion = eager_promotion;
+
+	eager_promotion = rtsFalse;
 	((StgMutVar *)p)->var = evacuate(((StgMutVar *)p)->var);
-	evac_gen = saved_evac_gen;
-	failed_to_evac = rtsTrue; // mutable anyhow
+	eager_promotion = saved_eager_promotion;
+
+	if (failed_to_evac) {
+	    ((StgClosure *)q)->header.info = &stg_MUT_VAR_DIRTY_info;
+	} else {
+	    ((StgClosure *)q)->header.info = &stg_MUT_VAR_CLEAN_info;
+	}
 	p += sizeofW(StgMutVar);
 	break;
+    }
 
     case CAF_BLACKHOLE:
     case SE_CAF_BLACKHOLE:
@@ -2928,18 +2962,32 @@ scavenge(step *stp)
 	p += arr_words_sizeW((StgArrWords *)p);
 	break;
 
-    case MUT_ARR_PTRS:
+    case MUT_ARR_PTRS_CLEAN:
+    case MUT_ARR_PTRS_DIRTY:
 	// follow everything 
     {
 	StgPtr next;
-
-	evac_gen = 0;		// repeatedly mutable 
+	rtsBool saved_eager;
+
+	// We don't eagerly promote objects pointed to by a mutable
+	// array, but if we find the array only points to objects in
+	// the same or an older generation, we mark it "clean" and
+	// avoid traversing it during minor GCs.
+	saved_eager = eager_promotion;
+	eager_promotion = rtsFalse;
 	next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
 	for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) {
 	    *p = (StgWord)(StgPtr)evacuate((StgClosure *)*p);
 	}
-	evac_gen = saved_evac_gen;
-	failed_to_evac = rtsTrue; // mutable anyhow.
+	eager_promotion = saved_eager;
+
+	if (failed_to_evac) {
+	    ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_DIRTY_info;
+	} else {
+	    ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_CLEAN_info;
+	}
+
+	failed_to_evac = rtsTrue; // always put it on the mutable list.
 	break;
     }
 
@@ -2967,10 +3015,19 @@ scavenge(step *stp)
     case TSO:
     { 
 	StgTSO *tso = (StgTSO *)p;
-	evac_gen = 0;
+	rtsBool saved_eager = eager_promotion;
+
+	eager_promotion = rtsFalse;
 	scavengeTSO(tso);
-	evac_gen = saved_evac_gen;
-	failed_to_evac = rtsTrue; // mutable anyhow.
+	eager_promotion = saved_eager;
+
+	if (failed_to_evac) {
+	    tso->flags |= TSO_DIRTY;
+	} else {
+	    tso->flags &= ~TSO_DIRTY;
+	}
+
+	failed_to_evac = rtsTrue; // always on the mutable list
 	p += tso_sizeW(tso);
 	break;
     }
@@ -3051,9 +3108,6 @@ scavenge(step *stp)
 	evac_gen = 0;
 	tvar->current_value = evacuate((StgClosure*)tvar->current_value);
 	tvar->first_wait_queue_entry = (StgTVarWaitQueue *)evacuate((StgClosure*)tvar->first_wait_queue_entry);
-#if defined(SMP)
-	tvar->last_update_by = (StgTRecHeader *)evacuate((StgClosure*)tvar->last_update_by);
-#endif
 	evac_gen = saved_evac_gen;
 	failed_to_evac = rtsTrue; // mutable
 	p += sizeofW(StgTVar);
@@ -3253,12 +3307,21 @@ linear_scan:
 		evacuate(((StgInd *)p)->indirectee);
 	    break;
 
-	case MUT_VAR:
-	    evac_gen = 0;
+	case MUT_VAR_CLEAN:
+	case MUT_VAR_DIRTY: {
+	    rtsBool saved_eager_promotion = eager_promotion;
+	    
+	    eager_promotion = rtsFalse;
 	    ((StgMutVar *)p)->var = evacuate(((StgMutVar *)p)->var);
-	    evac_gen = saved_evac_gen;
-	    failed_to_evac = rtsTrue;
+	    eager_promotion = saved_eager_promotion;
+	    
+	    if (failed_to_evac) {
+		((StgClosure *)q)->header.info = &stg_MUT_VAR_DIRTY_info;
+	    } else {
+		((StgClosure *)q)->header.info = &stg_MUT_VAR_CLEAN_info;
+	    }
 	    break;
+	}
 
 	case CAF_BLACKHOLE:
 	case SE_CAF_BLACKHOLE:
@@ -3292,17 +3355,31 @@ linear_scan:
 	    scavenge_AP((StgAP *)p);
 	    break;
       
-	case MUT_ARR_PTRS:
+	case MUT_ARR_PTRS_CLEAN:
+	case MUT_ARR_PTRS_DIRTY:
 	    // follow everything 
 	{
 	    StgPtr next;
-	    
-	    evac_gen = 0;		// repeatedly mutable 
+	    rtsBool saved_eager;
+
+	    // We don't eagerly promote objects pointed to by a mutable
+	    // array, but if we find the array only points to objects in
+	    // the same or an older generation, we mark it "clean" and
+	    // avoid traversing it during minor GCs.
+	    saved_eager = eager_promotion;
+	    eager_promotion = rtsFalse;
 	    next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
 	    for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) {
 		*p = (StgWord)(StgPtr)evacuate((StgClosure *)*p);
 	    }
-	    evac_gen = saved_evac_gen;
+	    eager_promotion = saved_eager;
+
+	    if (failed_to_evac) {
+		((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_DIRTY_info;
+	    } else {
+		((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_CLEAN_info;
+	    }
+
 	    failed_to_evac = rtsTrue; // mutable anyhow.
 	    break;
 	}
@@ -3331,10 +3408,19 @@ linear_scan:
 	case TSO:
 	{ 
 	    StgTSO *tso = (StgTSO *)p;
-	    evac_gen = 0;
+	    rtsBool saved_eager = eager_promotion;
+
+	    eager_promotion = rtsFalse;
 	    scavengeTSO(tso);
-	    evac_gen = saved_evac_gen;
-	    failed_to_evac = rtsTrue;
+	    eager_promotion = saved_eager;
+	    
+	    if (failed_to_evac) {
+		tso->flags |= TSO_DIRTY;
+	    } else {
+		tso->flags &= ~TSO_DIRTY;
+	    }
+	    
+	    failed_to_evac = rtsTrue; // always on the mutable list
 	    break;
 	}
 
@@ -3408,9 +3494,6 @@ linear_scan:
 	    evac_gen = 0;
 	    tvar->current_value = evacuate((StgClosure*)tvar->current_value);
 	    tvar->first_wait_queue_entry = (StgTVarWaitQueue *)evacuate((StgClosure*)tvar->first_wait_queue_entry);
-#if defined(SMP)
-            tvar->last_update_by = (StgTRecHeader *)evacuate((StgClosure*)tvar->last_update_by);
-#endif
 	    evac_gen = saved_evac_gen;
 	    failed_to_evac = rtsTrue; // mutable
 	    break;
@@ -3483,12 +3566,12 @@ linear_scan:
 
 	    // already scavenged?
 	    if (is_marked(oldgen_scan+1,oldgen_scan_bd)) {
-		oldgen_scan += sizeofW(StgHeader) + MIN_NONUPD_SIZE;
+		oldgen_scan += sizeofW(StgHeader) + MIN_PAYLOAD_SIZE;
 		goto loop;
 	    }
 	    push_mark_stack(oldgen_scan);
 	    // ToDo: bump the linear scan by the actual size of the object
-	    oldgen_scan += sizeofW(StgHeader) + MIN_NONUPD_SIZE;
+	    oldgen_scan += sizeofW(StgHeader) + MIN_PAYLOAD_SIZE;
 	    goto linear_scan;
 	}
 
@@ -3572,12 +3655,22 @@ scavenge_one(StgPtr p)
 	break;
     }
     
-    case MUT_VAR:
-	evac_gen = 0;
+    case MUT_VAR_CLEAN:
+    case MUT_VAR_DIRTY: {
+	StgPtr q = p;
+	rtsBool saved_eager_promotion = eager_promotion;
+
+	eager_promotion = rtsFalse;
 	((StgMutVar *)p)->var = evacuate(((StgMutVar *)p)->var);
-	evac_gen = saved_evac_gen;
-	failed_to_evac = rtsTrue; // mutable anyhow
+	eager_promotion = saved_eager_promotion;
+
+	if (failed_to_evac) {
+	    ((StgClosure *)q)->header.info = &stg_MUT_VAR_DIRTY_info;
+	} else {
+	    ((StgClosure *)q)->header.info = &stg_MUT_VAR_CLEAN_info;
+	}
 	break;
+    }
 
     case CAF_BLACKHOLE:
     case SE_CAF_BLACKHOLE:
@@ -3614,17 +3707,31 @@ scavenge_one(StgPtr p)
 	// nothing to follow 
 	break;
 
-    case MUT_ARR_PTRS:
+    case MUT_ARR_PTRS_CLEAN:
+    case MUT_ARR_PTRS_DIRTY:
     {
-	// follow everything 
-	StgPtr next;
-      
-	evac_gen = 0;		// repeatedly mutable 
+	StgPtr next, q;
+	rtsBool saved_eager;
+
+	// We don't eagerly promote objects pointed to by a mutable
+	// array, but if we find the array only points to objects in
+	// the same or an older generation, we mark it "clean" and
+	// avoid traversing it during minor GCs.
+	saved_eager = eager_promotion;
+	eager_promotion = rtsFalse;
+	q = p;
 	next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
 	for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) {
 	    *p = (StgWord)(StgPtr)evacuate((StgClosure *)*p);
 	}
-	evac_gen = saved_evac_gen;
+	eager_promotion = saved_eager;
+
+	if (failed_to_evac) {
+	    ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_DIRTY_info;
+	} else {
+	    ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_CLEAN_info;
+	}
+
 	failed_to_evac = rtsTrue;
 	break;
     }
@@ -3653,11 +3760,19 @@ scavenge_one(StgPtr p)
     case TSO:
     {
 	StgTSO *tso = (StgTSO *)p;
-      
-	evac_gen = 0;		// repeatedly mutable 
+	rtsBool saved_eager = eager_promotion;
+
+	eager_promotion = rtsFalse;
 	scavengeTSO(tso);
-	evac_gen = saved_evac_gen;
-	failed_to_evac = rtsTrue;
+	eager_promotion = saved_eager;
+
+	if (failed_to_evac) {
+	    tso->flags |= TSO_DIRTY;
+	} else {
+	    tso->flags &= ~TSO_DIRTY;
+	}
+
+	failed_to_evac = rtsTrue; // always on the mutable list
 	break;
     }
   
@@ -3732,9 +3847,6 @@ scavenge_one(StgPtr p)
 	evac_gen = 0;
 	tvar->current_value = evacuate((StgClosure*)tvar->current_value);
 	tvar->first_wait_queue_entry = (StgTVarWaitQueue *)evacuate((StgClosure*)tvar->first_wait_queue_entry);
-#if defined(SMP)
-	tvar->last_update_by = (StgTRecHeader *)evacuate((StgClosure*)tvar->last_update_by);
-#endif
 	evac_gen = saved_evac_gen;
 	failed_to_evac = rtsTrue; // mutable
 	break;
@@ -3843,10 +3955,55 @@ scavenge_mutable_list(generation *gen)
 	for (q = bd->start; q < bd->free; q++) {
 	    p = (StgPtr)*q;
 	    ASSERT(LOOKS_LIKE_CLOSURE_PTR(p));
+
+#ifdef DEBUG	    
+	    switch (get_itbl((StgClosure *)p)->type) {
+	    case MUT_VAR_CLEAN:
+		barf("MUT_VAR_CLEAN on mutable list");
+	    case MUT_VAR_DIRTY:
+		mutlist_MUTVARS++; break;
+	    case MUT_ARR_PTRS_CLEAN:
+	    case MUT_ARR_PTRS_DIRTY:
+	    case MUT_ARR_PTRS_FROZEN:
+	    case MUT_ARR_PTRS_FROZEN0:
+		mutlist_MUTARRS++; break;
+	    default:
+		mutlist_OTHERS++; break;
+	    }
+#endif
+
+	    // Check whether this object is "clean", that is it
+	    // definitely doesn't point into a young generation.
+	    // Clean objects don't need to be scavenged.  Some clean
+	    // objects (MUT_VAR_CLEAN) are not kept on the mutable
+	    // list at all; others, such as MUT_ARR_PTRS_CLEAN and
+	    // TSO, are always on the mutable list.
+	    //
+	    switch (get_itbl((StgClosure *)p)->type) {
+	    case MUT_ARR_PTRS_CLEAN:
+		recordMutableGen((StgClosure *)p,gen);
+		continue;
+	    case TSO: {
+		StgTSO *tso = (StgTSO *)p;
+		if ((tso->flags & TSO_DIRTY) == 0) {
+		    // A clean TSO: we don't have to traverse its
+		    // stack.  However, we *do* follow the link field:
+		    // we don't want to have to mark a TSO dirty just
+		    // because we put it on a different queue.
+		    if (tso->why_blocked != BlockedOnBlackHole) {
+			tso->link = (StgTSO *)evacuate((StgClosure *)tso->link);
+		    }
+		    recordMutableGen((StgClosure *)p,gen);
+		    continue;
+		}
+	    }
+	    default:
+		;
+	    }
+
 	    if (scavenge_one(p)) {
-		/* didn't manage to promote everything, so put the
-		 * object back on the list.
-		 */
+		// didn't manage to promote everything, so put the
+		// object back on the list.
 		recordMutableGen((StgClosure *)p,gen);
 	    }
 	}
@@ -4008,6 +4165,32 @@ scavenge_stack(StgPtr p, StgPtr stack_end)
     switch (info->i.type) {
 	
     case UPDATE_FRAME:
+	// In SMP, we can get update frames that point to indirections
+	// when two threads evaluate the same thunk.  We do attempt to
+	// discover this situation in threadPaused(), but it's
+	// possible that the following sequence occurs:
+	//
+	//        A             B
+	//                  enter T
+	//     enter T
+	//     blackhole T
+	//                  update T
+	//     GC
+	//
+	// Now T is an indirection, and the update frame is already
+	// marked on A's stack, so we won't traverse it again in
+	// threadPaused().  We could traverse the whole stack again
+	// before GC, but that seems like overkill.
+	//
+	// Scavenging this update frame as normal would be disastrous;
+	// the updatee would end up pointing to the value.  So we turn
+	// the indirection into an IND_PERM, so that evacuate will
+	// copy the indirection into the old generation instead of
+	// discarding it.
+	if (get_itbl(((StgUpdateFrame *)p)->updatee)->type == IND) {
+	    ((StgUpdateFrame *)p)->updatee->header.info = 
+		(StgInfoTable *)&stg_IND_PERM_info;
+	}
 	((StgUpdateFrame *)p)->updatee 
 	    = evacuate(((StgUpdateFrame *)p)->updatee);
 	p += sizeofW(StgUpdateFrame);
@@ -4439,7 +4622,7 @@ threadPaused(Capability *cap, StgTSO *tso)
 	    bh = ((StgUpdateFrame *)frame)->updatee;
 
 	    if (closure_IND(bh) || bh->header.info == &stg_BLACKHOLE_info) {
-		IF_DEBUG(squeeze, debugBelch("suspending duplicate work: %d words of stack\n", (StgPtr)frame - tso->sp));
+		IF_DEBUG(squeeze, debugBelch("suspending duplicate work: %ld words of stack\n", (StgPtr)frame - tso->sp));
 
 		// If this closure is already an indirection, then
 		// suspend the computation up to this point:
@@ -4507,8 +4690,8 @@ end:
     // Should we squeeze or not?  Arbitrary heuristic: we squeeze if
     // the number of words we have to shift down is less than the
     // number of stack words we squeeze away by doing so.
-    if (1 /*RtsFlags.GcFlags.squeezeUpdFrames == rtsTrue &&
-	    weight < words_to_squeeze*/) {
+    if (RtsFlags.GcFlags.squeezeUpdFrames == rtsTrue &&
+	weight < words_to_squeeze) {
 	stackSqueeze(tso, (StgPtr)frame);
     }
 }