A small GC optimisation
[ghc-hetmet.git] / rts / sm / Scav.c
index 9ebd4c5..9ac152a 100644 (file)
 #include "GCThread.h"
 #include "GCUtils.h"
 #include "Compact.h"
+#include "MarkStack.h"
 #include "Evac.h"
 #include "Scav.h"
 #include "Apply.h"
 #include "Trace.h"
 #include "Sanity.h"
 #include "Capability.h"
+#include "LdvProfile.h"
 
 static void scavenge_stack (StgPtr p, StgPtr stack_end);
 
@@ -34,8 +36,8 @@ static void scavenge_large_bitmap (StgPtr p,
 
 #if defined(THREADED_RTS) && !defined(PARALLEL_GC)
 # define evacuate(a) evacuate1(a)
-# define recordMutableGen_GC(a,b) recordMutableGen(a,b)
 # define scavenge_loop(a) scavenge_loop1(a)
+# define scavenge_block(a) scavenge_block1(a)
 # define scavenge_mutable_list(bd,g) scavenge_mutable_list1(bd,g)
 # define scavenge_capability_mut_lists(cap) scavenge_capability_mut_Lists1(cap)
 #endif
@@ -44,63 +46,125 @@ static void scavenge_large_bitmap (StgPtr p,
    Scavenge a TSO.
    -------------------------------------------------------------------------- */
 
-STATIC_INLINE void
-scavenge_TSO_link (StgTSO *tso)
-{
-    // We don't always chase the link field: TSOs on the blackhole
-    // queue are not automatically alive, so the link field is a
-    // "weak" pointer in that case.
-    if (tso->why_blocked != BlockedOnBlackHole) {
-        evacuate((StgClosure **)&tso->_link);
-    }
-}
-
 static void
 scavengeTSO (StgTSO *tso)
 {
     rtsBool saved_eager;
 
-    if (tso->what_next == ThreadRelocated) {
-        // the only way this can happen is if the old TSO was on the
-        // mutable list.  We might have other links to this defunct
-        // TSO, so we must update its link field.
-        evacuate((StgClosure**)&tso->_link);
-        return;
-    }
-
     debugTrace(DEBUG_gc,"scavenging thread %d",(int)tso->id);
 
+    // update the pointer from the Task.
+    if (tso->bound != NULL) {
+        tso->bound->tso = tso;
+    }
+
     saved_eager = gct->eager_promotion;
     gct->eager_promotion = rtsFalse;
 
+    evacuate((StgClosure **)&tso->blocked_exceptions);
+    evacuate((StgClosure **)&tso->bq);
+    
+    // scavange current transaction record
+    evacuate((StgClosure **)&tso->trec);
+
+    evacuate((StgClosure **)&tso->stackobj);
+
+    evacuate((StgClosure **)&tso->_link);
     if (   tso->why_blocked == BlockedOnMVar
        || tso->why_blocked == BlockedOnBlackHole
-       || tso->why_blocked == BlockedOnException
+       || tso->why_blocked == BlockedOnMsgThrowTo
+        || tso->why_blocked == NotBlocked
        ) {
        evacuate(&tso->block_info.closure);
     }
-    evacuate((StgClosure **)&tso->blocked_exceptions);
-    
-    // scavange current transaction record
-    evacuate((StgClosure **)&tso->trec);
-    
-    // scavenge this thread's stack 
-    scavenge_stack(tso->sp, &(tso->stack[tso->stack_size]));
+#ifdef THREADED_RTS
+    // in the THREADED_RTS, block_info.closure must always point to a
+    // valid closure, because we assume this in throwTo().  In the
+    // non-threaded RTS it might be a FD (for
+    // BlockedOnRead/BlockedOnWrite) or a time value (BlockedOnDelay)
+    else {
+        tso->block_info.closure = (StgClosure *)END_TSO_QUEUE;
+    }
+#endif
 
-    if (gct->failed_to_evac) {
-        tso->flags |= TSO_DIRTY;
-        scavenge_TSO_link(tso);
-    } else {
-        tso->flags &= ~TSO_DIRTY;
-        scavenge_TSO_link(tso);
+    tso->dirty = gct->failed_to_evac;
+
+    gct->eager_promotion = saved_eager;
+}
+
+/* -----------------------------------------------------------------------------
+   Mutable arrays of pointers
+   -------------------------------------------------------------------------- */
+
+static StgPtr scavenge_mut_arr_ptrs (StgMutArrPtrs *a)
+{
+    lnat m;
+    rtsBool any_failed;
+    StgPtr p, q;
+
+    any_failed = rtsFalse;
+    p = (StgPtr)&a->payload[0];
+    for (m = 0; (int)m < (int)mutArrPtrsCards(a->ptrs) - 1; m++)
+    {
+        q = p + (1 << MUT_ARR_PTRS_CARD_BITS);
+        for (; p < q; p++) {
+            evacuate((StgClosure**)p);
+        }
         if (gct->failed_to_evac) {
-            tso->flags |= TSO_LINK_DIRTY;
+            any_failed = rtsTrue;
+            *mutArrPtrsCard(a,m) = 1;
+            gct->failed_to_evac = rtsFalse;
         } else {
-            tso->flags &= ~TSO_LINK_DIRTY;
+            *mutArrPtrsCard(a,m) = 0;
         }
     }
 
-    gct->eager_promotion = saved_eager;
+    q = (StgPtr)&a->payload[a->ptrs];
+    if (p < q) {
+        for (; p < q; p++) {
+            evacuate((StgClosure**)p);
+        }
+        if (gct->failed_to_evac) {
+            any_failed = rtsTrue;
+            *mutArrPtrsCard(a,m) = 1;
+            gct->failed_to_evac = rtsFalse;
+        } else {
+            *mutArrPtrsCard(a,m) = 0;
+        }
+    }
+
+    gct->failed_to_evac = any_failed;
+    return (StgPtr)a + mut_arr_ptrs_sizeW(a);
+}
+    
+// scavenge only the marked areas of a MUT_ARR_PTRS
+static StgPtr scavenge_mut_arr_ptrs_marked (StgMutArrPtrs *a)
+{
+    lnat m;
+    StgPtr p, q;
+    rtsBool any_failed;
+
+    any_failed = rtsFalse;
+    for (m = 0; m < mutArrPtrsCards(a->ptrs); m++)
+    {
+        if (*mutArrPtrsCard(a,m) != 0) {
+            p = (StgPtr)&a->payload[m << MUT_ARR_PTRS_CARD_BITS];
+            q = stg_min(p + (1 << MUT_ARR_PTRS_CARD_BITS),
+                        (StgPtr)&a->payload[a->ptrs]);
+            for (; p < q; p++) {
+                evacuate((StgClosure**)p);
+            }
+            if (gct->failed_to_evac) {
+                any_failed = rtsTrue;
+                gct->failed_to_evac = rtsFalse;
+            } else {
+                *mutArrPtrsCard(a,m) = 0;
+            }
+        }
+    }
+
+    gct->failed_to_evac = any_failed;
+    return (StgPtr)a + mut_arr_ptrs_sizeW(a);
 }
 
 /* -----------------------------------------------------------------------------
@@ -249,7 +313,7 @@ scavenge_srt (StgClosure **srt, nat srt_bitmap)
 
   while (bitmap != 0) {
       if ((bitmap & 1) != 0) {
-#if defined(__PIC__) && defined(mingw32_TARGET_OS)
+#if defined(__PIC__) && defined(mingw32_HOST_OS)
          // Special-case to handle references to closures hiding out in DLLs, since
          // double indirections required to get at those. The code generator knows
          // which is which when generating the SRT, so it stores the (indirect)
@@ -259,7 +323,7 @@ scavenge_srt (StgClosure **srt, nat srt_bitmap)
          // If the SRT entry hasn't got bit 0 set, the SRT entry points to a
          // closure that's fixed at link-time, and no extra magic is required.
          if ( (unsigned long)(*srt) & 0x1 ) {
-             evacuate(stgCast(StgClosure**,(stgCast(unsigned long, *srt) & ~0x1)));
+             evacuate( (StgClosure**) ((unsigned long) (*srt) & ~0x1));
          } else {
              evacuate(p);
          }
@@ -298,11 +362,11 @@ scavenge_fun_srt(const StgInfoTable *info)
 /* -----------------------------------------------------------------------------
    Scavenge a block from the given scan pointer up to bd->free.
 
-   evac_step is set by the caller to be either zero (for a step in a
+   evac_gen_no is set by the caller to be either zero (for a step in a
    generation < N) or G where G is the generation of the step being
    scavenged.  
 
-   We sometimes temporarily change evac_step back to zero if we're
+   We sometimes temporarily change evac_gen_no back to zero if we're
    scavenging a mutable object where eager promotion isn't such a good
    idea.  
    -------------------------------------------------------------------------- */
@@ -312,20 +376,18 @@ scavenge_block (bdescr *bd)
 {
   StgPtr p, q;
   StgInfoTable *info;
-  step *saved_evac_step;
   rtsBool saved_eager_promotion;
-  step_workspace *ws;
+  gen_workspace *ws;
 
-  debugTrace(DEBUG_gc, "scavenging block %p (gen %d, step %d) @ %p",
-            bd->start, bd->gen_no, bd->step->no, bd->u.scan);
+  debugTrace(DEBUG_gc, "scavenging block %p (gen %d) @ %p",
+            bd->start, bd->gen_no, bd->u.scan);
 
   gct->scan_bd = bd;
-  gct->evac_step = bd->step;
-  saved_evac_step = gct->evac_step;
+  gct->evac_gen_no = bd->gen_no;
   saved_eager_promotion = gct->eager_promotion;
   gct->failed_to_evac = rtsFalse;
 
-  ws = &gct->steps[bd->step->abs_no];
+  ws = &gct->gens[bd->gen->no];
 
   p = bd->u.scan;
   
@@ -450,7 +512,7 @@ scavenge_block (bdescr *bd)
     gen_obj:
     case CONSTR:
     case WEAK:
-    case STABLE_NAME:
+    case PRIM:
     {
        StgPtr end;
 
@@ -472,23 +534,7 @@ scavenge_block (bdescr *bd)
     }
 
     case IND_PERM:
-      if (bd->gen_no != 0) {
-#ifdef PROFILING
-        // @LDV profiling
-        // No need to call LDV_recordDead_FILL_SLOP_DYNAMIC() because an 
-        // IND_OLDGEN_PERM closure is larger than an IND_PERM closure.
-        LDV_recordDead((StgClosure *)p, sizeofW(StgInd));
-#endif        
-        // 
-        // Todo: maybe use SET_HDR() and remove LDV_RECORD_CREATE()?
-        //
-       SET_INFO(((StgClosure *)p), &stg_IND_OLDGEN_PERM_info);
-
-        // We pretend that p has just been created.
-        LDV_RECORD_CREATE((StgClosure *)p);
-      }
-       // fall through 
-    case IND_OLDGEN_PERM:
+    case BLACKHOLE:
        evacuate(&((StgInd *)p)->indirectee);
        p += sizeofW(StgInd);
        break;
@@ -507,10 +553,25 @@ scavenge_block (bdescr *bd)
        p += sizeofW(StgMutVar);
        break;
 
-    case CAF_BLACKHOLE:
-    case BLACKHOLE:
-       p += BLACKHOLE_sizeW();
-       break;
+    case BLOCKING_QUEUE:
+    {
+        StgBlockingQueue *bq = (StgBlockingQueue *)p;
+        
+       gct->eager_promotion = rtsFalse;
+        evacuate(&bq->bh);
+        evacuate((StgClosure**)&bq->owner);
+        evacuate((StgClosure**)&bq->queue);
+        evacuate((StgClosure**)&bq->link);
+       gct->eager_promotion = saved_eager_promotion;
+
+       if (gct->failed_to_evac) {
+           bq->header.info = &stg_BLOCKING_QUEUE_DIRTY_info;
+       } else {
+           bq->header.info = &stg_BLOCKING_QUEUE_CLEAN_info;
+       }
+        p += sizeofW(StgBlockingQueue);
+        break;
+    }
 
     case THUNK_SELECTOR:
     { 
@@ -546,20 +607,14 @@ scavenge_block (bdescr *bd)
 
     case MUT_ARR_PTRS_CLEAN:
     case MUT_ARR_PTRS_DIRTY:
-       // follow everything 
     {
-       StgPtr next;
+        // 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.
+        gct->eager_promotion = rtsFalse;
 
-       // 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.
-       gct->eager_promotion = rtsFalse;
-       next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
-       for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) {
-           evacuate((StgClosure **)p);
-       }
-       gct->eager_promotion = saved_eager_promotion;
+        p = scavenge_mut_arr_ptrs((StgMutArrPtrs*)p);
 
        if (gct->failed_to_evac) {
            ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_DIRTY_info;
@@ -567,6 +622,7 @@ scavenge_block (bdescr *bd)
            ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_CLEAN_info;
        }
 
+       gct->eager_promotion = saved_eager_promotion;
        gct->failed_to_evac = rtsTrue; // always put it on the mutable list.
        break;
     }
@@ -575,17 +631,12 @@ scavenge_block (bdescr *bd)
     case MUT_ARR_PTRS_FROZEN0:
        // follow everything 
     {
-       StgPtr next;
-
-       next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
-       for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) {
-           evacuate((StgClosure **)p);
-       }
+        p = scavenge_mut_arr_ptrs((StgMutArrPtrs*)p);
 
        // If we're going to put this object on the mutable list, then
        // set its info ptr to MUT_ARR_PTRS_FROZEN0 to indicate that.
        if (gct->failed_to_evac) {
-           ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_FROZEN0_info;
+            ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_FROZEN0_info;
        } else {
            ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_FROZEN_info;
        }
@@ -594,48 +645,40 @@ scavenge_block (bdescr *bd)
 
     case TSO:
     { 
-       StgTSO *tso = (StgTSO *)p;
-        scavengeTSO(tso);
-       p += tso_sizeW(tso);
+        scavengeTSO((StgTSO *)p);
+        p += sizeofW(StgTSO);
        break;
     }
 
-    case TVAR_WATCH_QUEUE:
-      {
-       StgTVarWatchQueue *wq = ((StgTVarWatchQueue *) p);
-       gct->evac_step = 0;
-       evacuate((StgClosure **)&wq->closure);
-       evacuate((StgClosure **)&wq->next_queue_entry);
-       evacuate((StgClosure **)&wq->prev_queue_entry);
-       gct->evac_step = saved_evac_step;
-       gct->failed_to_evac = rtsTrue; // mutable
-       p += sizeofW(StgTVarWatchQueue);
-       break;
-      }
+    case STACK:
+    {
+        StgStack *stack = (StgStack*)p;
 
-    case TVAR:
-      {
-       StgTVar *tvar = ((StgTVar *) p);
-       gct->evac_step = 0;
-       evacuate((StgClosure **)&tvar->current_value);
-       evacuate((StgClosure **)&tvar->first_watch_queue_entry);
-       gct->evac_step = saved_evac_step;
-       gct->failed_to_evac = rtsTrue; // mutable
-       p += sizeofW(StgTVar);
-       break;
-      }
+        gct->eager_promotion = rtsFalse;
+
+        scavenge_stack(stack->sp, stack->stack + stack->stack_size);
+        stack->dirty = gct->failed_to_evac;
+        p += stack_sizeW(stack);
 
-    case TREC_HEADER:
+        gct->eager_promotion = saved_eager_promotion;
+        break;
+    }
+
+    case MUT_PRIM:
       {
-        StgTRecHeader *trec = ((StgTRecHeader *) p);
-        gct->evac_step = 0;
-       evacuate((StgClosure **)&trec->enclosing_trec);
-       evacuate((StgClosure **)&trec->current_chunk);
-       evacuate((StgClosure **)&trec->invariants_to_check);
-       gct->evac_step = saved_evac_step;
+       StgPtr end;
+
+       gct->eager_promotion = rtsFalse;
+
+       end = (P_)((StgClosure *)p)->payload + info->layout.payload.ptrs;
+       for (p = (P_)((StgClosure *)p)->payload; p < end; p++) {
+           evacuate((StgClosure **)p);
+       }
+       p += info->layout.payload.nptrs;
+
+       gct->eager_promotion = saved_eager_promotion;
        gct->failed_to_evac = rtsTrue; // mutable
-       p += sizeofW(StgTRecHeader);
-        break;
+       break;
       }
 
     case TREC_CHUNK:
@@ -643,44 +686,19 @@ scavenge_block (bdescr *bd)
        StgWord i;
        StgTRecChunk *tc = ((StgTRecChunk *) p);
        TRecEntry *e = &(tc -> entries[0]);
-       gct->evac_step = 0;
+       gct->eager_promotion = rtsFalse;
        evacuate((StgClosure **)&tc->prev_chunk);
        for (i = 0; i < tc -> next_entry_idx; i ++, e++ ) {
          evacuate((StgClosure **)&e->tvar);
          evacuate((StgClosure **)&e->expected_value);
          evacuate((StgClosure **)&e->new_value);
        }
-       gct->evac_step = saved_evac_step;
+       gct->eager_promotion = saved_eager_promotion;
        gct->failed_to_evac = rtsTrue; // mutable
        p += sizeofW(StgTRecChunk);
        break;
       }
 
-    case ATOMIC_INVARIANT:
-      {
-        StgAtomicInvariant *invariant = ((StgAtomicInvariant *) p);
-        gct->evac_step = 0;
-       evacuate(&invariant->code);
-       evacuate((StgClosure **)&invariant->last_execution);
-       gct->evac_step = saved_evac_step;
-       gct->failed_to_evac = rtsTrue; // mutable
-       p += sizeofW(StgAtomicInvariant);
-        break;
-      }
-
-    case INVARIANT_CHECK_QUEUE:
-      {
-        StgInvariantCheckQueue *queue = ((StgInvariantCheckQueue *) p);
-        gct->evac_step = 0;
-       evacuate((StgClosure **)&queue->invariant);
-       evacuate((StgClosure **)&queue->my_execution);
-       evacuate((StgClosure **)&queue->next_queue_entry);
-       gct->evac_step = saved_evac_step;
-       gct->failed_to_evac = rtsTrue; // mutable
-       p += sizeofW(StgInvariantCheckQueue);
-        break;
-      }
-
     default:
        barf("scavenge: unimplemented/strange closure type %d @ %p", 
             info->type, p);
@@ -734,14 +752,12 @@ scavenge_mark_stack(void)
 {
     StgPtr p, q;
     StgInfoTable *info;
-    step *saved_evac_step;
+    rtsBool saved_eager_promotion;
 
-    gct->evac_step = &oldest_gen->steps[0];
-    saved_evac_step = gct->evac_step;
+    gct->evac_gen_no = oldest_gen->no;
+    saved_eager_promotion = gct->eager_promotion;
 
-linear_scan:
-    while (!mark_stack_empty()) {
-       p = pop_mark_stack();
+    while ((p = pop_mark_stack())) {
 
        ASSERT(LOOKS_LIKE_CLOSURE_PTR(p));
        info = get_itbl((StgClosure *)p);
@@ -752,8 +768,6 @@ linear_scan:
         case MVAR_CLEAN:
         case MVAR_DIRTY:
         { 
-            rtsBool saved_eager_promotion = gct->eager_promotion;
-            
             StgMVar *mvar = ((StgMVar *)p);
             gct->eager_promotion = rtsFalse;
             evacuate((StgClosure **)&mvar->head);
@@ -836,7 +850,7 @@ linear_scan:
        gen_obj:
        case CONSTR:
        case WEAK:
-       case STABLE_NAME:
+       case PRIM:
        {
            StgPtr end;
            
@@ -861,15 +875,13 @@ linear_scan:
            // no "old" generation.
            break;
 
-       case IND_OLDGEN:
-       case IND_OLDGEN_PERM:
+       case IND:
+        case BLACKHOLE:
            evacuate(&((StgInd *)p)->indirectee);
            break;
 
        case MUT_VAR_CLEAN:
        case MUT_VAR_DIRTY: {
-           rtsBool saved_eager_promotion = gct->eager_promotion;
-           
            gct->eager_promotion = rtsFalse;
            evacuate(&((StgMutVar *)p)->var);
            gct->eager_promotion = saved_eager_promotion;
@@ -882,8 +894,25 @@ linear_scan:
            break;
        }
 
-       case CAF_BLACKHOLE:
-       case BLACKHOLE:
+        case BLOCKING_QUEUE:
+        {
+            StgBlockingQueue *bq = (StgBlockingQueue *)p;
+            
+            gct->eager_promotion = rtsFalse;
+            evacuate(&bq->bh);
+            evacuate((StgClosure**)&bq->owner);
+            evacuate((StgClosure**)&bq->queue);
+            evacuate((StgClosure**)&bq->link);
+            gct->eager_promotion = saved_eager_promotion;
+            
+            if (gct->failed_to_evac) {
+                bq->header.info = &stg_BLOCKING_QUEUE_DIRTY_info;
+            } else {
+                bq->header.info = &stg_BLOCKING_QUEUE_CLEAN_info;
+            }
+            break;
+        }
+
        case ARR_WORDS:
            break;
 
@@ -916,27 +945,21 @@ linear_scan:
        case MUT_ARR_PTRS_DIRTY:
            // follow everything 
        {
-           StgPtr next;
-           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 = gct->eager_promotion;
            gct->eager_promotion = rtsFalse;
-           next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
-           for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) {
-               evacuate((StgClosure **)p);
-           }
-           gct->eager_promotion = saved_eager;
 
-           if (gct->failed_to_evac) {
-               ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_DIRTY_info;
-           } else {
-               ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_CLEAN_info;
-           }
+            scavenge_mut_arr_ptrs((StgMutArrPtrs *)p);
 
+            if (gct->failed_to_evac) {
+                ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_DIRTY_info;
+            } else {
+                ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_CLEAN_info;
+            }
+
+           gct->eager_promotion = saved_eager_promotion;
            gct->failed_to_evac = rtsTrue; // mutable anyhow.
            break;
        }
@@ -945,12 +968,9 @@ linear_scan:
        case MUT_ARR_PTRS_FROZEN0:
            // follow everything 
        {
-           StgPtr next, q = p;
+           StgPtr q = p;
            
-           next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
-           for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) {
-               evacuate((StgClosure **)p);
-           }
+            scavenge_mut_arr_ptrs((StgMutArrPtrs *)p);
 
            // If we're going to put this object on the mutable list, then
            // set its info ptr to MUT_ARR_PTRS_FROZEN0 to indicate that.
@@ -968,81 +988,52 @@ linear_scan:
            break;
        }
 
-       case TVAR_WATCH_QUEUE:
-         {
-           StgTVarWatchQueue *wq = ((StgTVarWatchQueue *) p);
-           gct->evac_step = 0;
-            evacuate((StgClosure **)&wq->closure);
-           evacuate((StgClosure **)&wq->next_queue_entry);
-           evacuate((StgClosure **)&wq->prev_queue_entry);
-           gct->evac_step = saved_evac_step;
-           gct->failed_to_evac = rtsTrue; // mutable
-           break;
-         }
-         
-       case TVAR:
-         {
-           StgTVar *tvar = ((StgTVar *) p);
-           gct->evac_step = 0;
-           evacuate((StgClosure **)&tvar->current_value);
-           evacuate((StgClosure **)&tvar->first_watch_queue_entry);
-           gct->evac_step = saved_evac_step;
-           gct->failed_to_evac = rtsTrue; // mutable
-           break;
-         }
-         
+        case STACK:
+        {
+            StgStack *stack = (StgStack*)p;
+
+            gct->eager_promotion = rtsFalse;
+
+            scavenge_stack(stack->sp, stack->stack + stack->stack_size);
+            stack->dirty = gct->failed_to_evac;
+
+            gct->eager_promotion = saved_eager_promotion;
+            break;
+        }
+
+        case MUT_PRIM:
+        {
+            StgPtr end;
+            
+            gct->eager_promotion = rtsFalse;
+            
+            end = (P_)((StgClosure *)p)->payload + info->layout.payload.ptrs;
+            for (p = (P_)((StgClosure *)p)->payload; p < end; p++) {
+                evacuate((StgClosure **)p);
+            }
+            
+            gct->eager_promotion = saved_eager_promotion;
+            gct->failed_to_evac = rtsTrue; // mutable
+            break;
+        }
+
        case TREC_CHUNK:
          {
            StgWord i;
            StgTRecChunk *tc = ((StgTRecChunk *) p);
            TRecEntry *e = &(tc -> entries[0]);
-           gct->evac_step = 0;
+           gct->eager_promotion = rtsFalse;
            evacuate((StgClosure **)&tc->prev_chunk);
            for (i = 0; i < tc -> next_entry_idx; i ++, e++ ) {
              evacuate((StgClosure **)&e->tvar);
              evacuate((StgClosure **)&e->expected_value);
              evacuate((StgClosure **)&e->new_value);
            }
-           gct->evac_step = saved_evac_step;
-           gct->failed_to_evac = rtsTrue; // mutable
-           break;
-         }
-
-       case TREC_HEADER:
-         {
-           StgTRecHeader *trec = ((StgTRecHeader *) p);
-           gct->evac_step = 0;
-           evacuate((StgClosure **)&trec->enclosing_trec);
-           evacuate((StgClosure **)&trec->current_chunk);
-           evacuate((StgClosure **)&trec->invariants_to_check);
-           gct->evac_step = saved_evac_step;
+           gct->eager_promotion = saved_eager_promotion;
            gct->failed_to_evac = rtsTrue; // mutable
            break;
          }
 
-        case ATOMIC_INVARIANT:
-          {
-            StgAtomicInvariant *invariant = ((StgAtomicInvariant *) p);
-            gct->evac_step = 0;
-           evacuate(&invariant->code);
-           evacuate((StgClosure **)&invariant->last_execution);
-           gct->evac_step = saved_evac_step;
-           gct->failed_to_evac = rtsTrue; // mutable
-            break;
-          }
-
-        case INVARIANT_CHECK_QUEUE:
-          {
-            StgInvariantCheckQueue *queue = ((StgInvariantCheckQueue *) p);
-            gct->evac_step = 0;
-           evacuate((StgClosure **)&queue->invariant);
-           evacuate((StgClosure **)&queue->my_execution);
-            evacuate((StgClosure **)&queue->next_queue_entry);
-           gct->evac_step = saved_evac_step;
-           gct->failed_to_evac = rtsTrue; // mutable
-            break;
-          }
-
        default:
            barf("scavenge_mark_stack: unimplemented/strange closure type %d @ %p", 
                 info->type, p);
@@ -1050,53 +1041,11 @@ linear_scan:
 
        if (gct->failed_to_evac) {
            gct->failed_to_evac = rtsFalse;
-           if (gct->evac_step) {
-               recordMutableGen_GC((StgClosure *)q, gct->evac_step->gen_no);
-           }
-       }
-       
-       // mark the next bit to indicate "scavenged"
-       mark(q+1, Bdescr(q));
-
-    } // while (!mark_stack_empty())
-
-    // start a new linear scan if the mark stack overflowed at some point
-    if (mark_stack_overflowed && oldgen_scan_bd == NULL) {
-       debugTrace(DEBUG_gc, "scavenge_mark_stack: starting linear scan");
-       mark_stack_overflowed = rtsFalse;
-       oldgen_scan_bd = oldest_gen->steps[0].old_blocks;
-       oldgen_scan = oldgen_scan_bd->start;
-    }
-
-    if (oldgen_scan_bd) {
-       // push a new thing on the mark stack
-    loop:
-       // find a closure that is marked but not scavenged, and start
-       // from there.
-       while (oldgen_scan < oldgen_scan_bd->free 
-              && !is_marked(oldgen_scan,oldgen_scan_bd)) {
-           oldgen_scan++;
-       }
-
-       if (oldgen_scan < oldgen_scan_bd->free) {
-
-           // already scavenged?
-           if (is_marked(oldgen_scan+1,oldgen_scan_bd)) {
-               oldgen_scan += sizeofW(StgHeader) + MIN_PAYLOAD_SIZE;
-               goto loop;
+            if (gct->evac_gen_no) {
+                recordMutableGen_GC((StgClosure *)q, gct->evac_gen_no);
            }
-           push_mark_stack(oldgen_scan);
-           // ToDo: bump the linear scan by the actual size of the object
-           oldgen_scan += sizeofW(StgHeader) + MIN_PAYLOAD_SIZE;
-           goto linear_scan;
        }
-
-       oldgen_scan_bd = oldgen_scan_bd->link;
-       if (oldgen_scan_bd != NULL) {
-           oldgen_scan = oldgen_scan_bd->start;
-           goto loop;
-       }
-    }
+    } // while (p = pop_mark_stack())
 }
 
 /* -----------------------------------------------------------------------------
@@ -1111,9 +1060,11 @@ static rtsBool
 scavenge_one(StgPtr p)
 {
     const StgInfoTable *info;
-    step *saved_evac_step = gct->evac_step;
     rtsBool no_luck;
+    rtsBool saved_eager_promotion;
     
+    saved_eager_promotion = gct->eager_promotion;
+
     ASSERT(LOOKS_LIKE_CLOSURE_PTR(p));
     info = get_itbl((StgClosure *)p);
     
@@ -1122,8 +1073,6 @@ scavenge_one(StgPtr p)
     case MVAR_CLEAN:
     case MVAR_DIRTY:
     { 
-       rtsBool saved_eager_promotion = gct->eager_promotion;
-
        StgMVar *mvar = ((StgMVar *)p);
        gct->eager_promotion = rtsFalse;
        evacuate((StgClosure **)&mvar->head);
@@ -1168,6 +1117,7 @@ scavenge_one(StgPtr p)
     case CONSTR_0_2:
     case CONSTR_2_0:
     case WEAK:
+    case PRIM:
     case IND_PERM:
     {
        StgPtr q, end;
@@ -1182,7 +1132,6 @@ scavenge_one(StgPtr p)
     case MUT_VAR_CLEAN:
     case MUT_VAR_DIRTY: {
        StgPtr q = p;
-       rtsBool saved_eager_promotion = gct->eager_promotion;
 
        gct->eager_promotion = rtsFalse;
        evacuate(&((StgMutVar *)p)->var);
@@ -1196,10 +1145,25 @@ scavenge_one(StgPtr p)
        break;
     }
 
-    case CAF_BLACKHOLE:
-    case BLACKHOLE:
-       break;
-       
+    case BLOCKING_QUEUE:
+    {
+        StgBlockingQueue *bq = (StgBlockingQueue *)p;
+        
+        gct->eager_promotion = rtsFalse;
+        evacuate(&bq->bh);
+        evacuate((StgClosure**)&bq->owner);
+        evacuate((StgClosure**)&bq->queue);
+        evacuate((StgClosure**)&bq->link);
+        gct->eager_promotion = saved_eager_promotion;
+        
+        if (gct->failed_to_evac) {
+            bq->header.info = &stg_BLOCKING_QUEUE_DIRTY_info;
+        } else {
+            bq->header.info = &stg_BLOCKING_QUEUE_CLEAN_info;
+        }
+        break;
+    }
+
     case THUNK_SELECTOR:
     { 
        StgSelector *s = (StgSelector *)p;
@@ -1232,28 +1196,21 @@ scavenge_one(StgPtr p)
     case MUT_ARR_PTRS_CLEAN:
     case MUT_ARR_PTRS_DIRTY:
     {
-       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 = gct->eager_promotion;
        gct->eager_promotion = rtsFalse;
-       q = p;
-       next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
-       for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) {
-           evacuate((StgClosure **)p);
-       }
-       gct->eager_promotion = saved_eager;
+
+        scavenge_mut_arr_ptrs((StgMutArrPtrs *)p);
 
        if (gct->failed_to_evac) {
-           ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_DIRTY_info;
+           ((StgClosure *)p)->header.info = &stg_MUT_ARR_PTRS_DIRTY_info;
        } else {
-           ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_CLEAN_info;
+           ((StgClosure *)p)->header.info = &stg_MUT_ARR_PTRS_CLEAN_info;
        }
 
+       gct->eager_promotion = saved_eager_promotion;
        gct->failed_to_evac = rtsTrue;
        break;
     }
@@ -1262,19 +1219,14 @@ scavenge_one(StgPtr p)
     case MUT_ARR_PTRS_FROZEN0:
     {
        // follow everything 
-       StgPtr next, q=p;
-      
-       next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
-       for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) {
-           evacuate((StgClosure **)p);
-       }
-
+        scavenge_mut_arr_ptrs((StgMutArrPtrs *)p);
+        
        // If we're going to put this object on the mutable list, then
        // set its info ptr to MUT_ARR_PTRS_FROZEN0 to indicate that.
        if (gct->failed_to_evac) {
-           ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_FROZEN0_info;
+           ((StgClosure *)p)->header.info = &stg_MUT_ARR_PTRS_FROZEN0_info;
        } else {
-           ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_FROZEN_info;
+           ((StgClosure *)p)->header.info = &stg_MUT_ARR_PTRS_FROZEN_info;
        }
        break;
     }
@@ -1285,83 +1237,58 @@ scavenge_one(StgPtr p)
        break;
     }
   
-    case TVAR_WATCH_QUEUE:
-      {
-       StgTVarWatchQueue *wq = ((StgTVarWatchQueue *) p);
-       gct->evac_step = 0;
-        evacuate((StgClosure **)&wq->closure);
-        evacuate((StgClosure **)&wq->next_queue_entry);
-        evacuate((StgClosure **)&wq->prev_queue_entry);
-       gct->evac_step = saved_evac_step;
-       gct->failed_to_evac = rtsTrue; // mutable
-       break;
-      }
+    case STACK:
+    {
+        StgStack *stack = (StgStack*)p;
 
-    case TVAR:
-      {
-       StgTVar *tvar = ((StgTVar *) p);
-       gct->evac_step = 0;
-       evacuate((StgClosure **)&tvar->current_value);
-        evacuate((StgClosure **)&tvar->first_watch_queue_entry);
-       gct->evac_step = saved_evac_step;
+        gct->eager_promotion = rtsFalse;
+
+        scavenge_stack(stack->sp, stack->stack + stack->stack_size);
+        stack->dirty = gct->failed_to_evac;
+
+        gct->eager_promotion = saved_eager_promotion;
+        break;
+    }
+
+    case MUT_PRIM:
+    {
+       StgPtr end;
+        
+       gct->eager_promotion = rtsFalse;
+        
+       end = (P_)((StgClosure *)p)->payload + info->layout.payload.ptrs;
+       for (p = (P_)((StgClosure *)p)->payload; p < end; p++) {
+           evacuate((StgClosure **)p);
+       }
+
+       gct->eager_promotion = saved_eager_promotion;
        gct->failed_to_evac = rtsTrue; // mutable
        break;
-      }
 
-    case TREC_HEADER:
-      {
-        StgTRecHeader *trec = ((StgTRecHeader *) p);
-        gct->evac_step = 0;
-       evacuate((StgClosure **)&trec->enclosing_trec);
-       evacuate((StgClosure **)&trec->current_chunk);
-        evacuate((StgClosure **)&trec->invariants_to_check);
-       gct->evac_step = saved_evac_step;
-       gct->failed_to_evac = rtsTrue; // mutable
-        break;
-      }
+    }
 
     case TREC_CHUNK:
       {
        StgWord i;
        StgTRecChunk *tc = ((StgTRecChunk *) p);
        TRecEntry *e = &(tc -> entries[0]);
-       gct->evac_step = 0;
+       gct->eager_promotion = rtsFalse;
        evacuate((StgClosure **)&tc->prev_chunk);
        for (i = 0; i < tc -> next_entry_idx; i ++, e++ ) {
          evacuate((StgClosure **)&e->tvar);
          evacuate((StgClosure **)&e->expected_value);
          evacuate((StgClosure **)&e->new_value);
        }
-       gct->evac_step = saved_evac_step;
+       gct->eager_promotion = saved_eager_promotion;
        gct->failed_to_evac = rtsTrue; // mutable
        break;
       }
 
-    case ATOMIC_INVARIANT:
-    {
-      StgAtomicInvariant *invariant = ((StgAtomicInvariant *) p);
-      gct->evac_step = 0;
-      evacuate(&invariant->code);
-      evacuate((StgClosure **)&invariant->last_execution);
-      gct->evac_step = saved_evac_step;
-      gct->failed_to_evac = rtsTrue; // mutable
-      break;
-    }
-
-    case INVARIANT_CHECK_QUEUE:
-    {
-      StgInvariantCheckQueue *queue = ((StgInvariantCheckQueue *) p);
-      gct->evac_step = 0;
-      evacuate((StgClosure **)&queue->invariant);
-      evacuate((StgClosure **)&queue->my_execution);
-      evacuate((StgClosure **)&queue->next_queue_entry);
-      gct->evac_step = saved_evac_step;
-      gct->failed_to_evac = rtsTrue; // mutable
-      break;
-    }
-
-    case IND_OLDGEN:
-    case IND_OLDGEN_PERM:
+    case IND:
+        // IND can happen, for example, when the interpreter allocates
+        // a gigantic AP closure (more than one block), which ends up
+        // on the large-object list and then gets updated.  See #3424.
+    case BLACKHOLE:
     case IND_STATIC:
        evacuate(&((StgInd *)p)->indirectee);
 
@@ -1371,21 +1298,21 @@ scavenge_one(StgPtr p)
        * promoted 
        */
       { 
-       StgPtr start = gen->steps[0].scan;
-       bdescr *start_bd = gen->steps[0].scan_bd;
+       StgPtr start = gen->scan;
+       bdescr *start_bd = gen->scan_bd;
        nat size = 0;
-       scavenge(&gen->steps[0]);
-       if (start_bd != gen->steps[0].scan_bd) {
+       scavenge(&gen);
+       if (start_bd != gen->scan_bd) {
          size += (P_)BLOCK_ROUND_UP(start) - start;
          start_bd = start_bd->link;
-         while (start_bd != gen->steps[0].scan_bd) {
+         while (start_bd != gen->scan_bd) {
            size += BLOCK_SIZE_W;
            start_bd = start_bd->link;
          }
-         size += gen->steps[0].scan -
-           (P_)BLOCK_ROUND_DOWN(gen->steps[0].scan);
+         size += gen->scan -
+           (P_)BLOCK_ROUND_DOWN(gen->scan);
        } else {
-         size = gen->steps[0].scan - start;
+         size = gen->scan - start;
        }
        debugBelch("evac IND_OLDGEN: %ld bytes", size * sizeof(W_));
       }
@@ -1413,8 +1340,10 @@ void
 scavenge_mutable_list(bdescr *bd, generation *gen)
 {
     StgPtr p, q;
+    nat gen_no;
 
-    gct->evac_step = &gen->steps[0];
+    gen_no = gen->no;
+    gct->evac_gen_no = gen_no;
     for (; bd != NULL; bd = bd->link) {
        for (q = bd->start; q < bd->free; q++) {
            p = (StgPtr)*q;
@@ -1423,7 +1352,7 @@ scavenge_mutable_list(bdescr *bd, generation *gen)
 #ifdef DEBUG       
            switch (get_itbl((StgClosure *)p)->type) {
            case MUT_VAR_CLEAN:
-               barf("MUT_VAR_CLEAN on mutable list");
+                // can happen due to concurrent writeMutVars
            case MUT_VAR_DIRTY:
                mutlist_MUTVARS++; break;
            case MUT_ARR_PTRS_CLEAN:
@@ -1444,38 +1373,40 @@ scavenge_mutable_list(bdescr *bd, generation *gen)
            // 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
+           // list at all; others, such as TSO
            // are always on the mutable list.
            //
            switch (get_itbl((StgClosure *)p)->type) {
            case MUT_ARR_PTRS_CLEAN:
-               recordMutableGen_GC((StgClosure *)p,gen->no);
+                recordMutableGen_GC((StgClosure *)p,gen_no);
                continue;
-           case TSO: {
-               StgTSO *tso = (StgTSO *)p;
-               if ((tso->flags & TSO_DIRTY) == 0) {
-                    // Must be on the mutable list because its link
-                    // field is dirty.
-                    ASSERT(tso->flags & TSO_LINK_DIRTY);
-
-                    scavenge_TSO_link(tso);
-                    if (gct->failed_to_evac) {
-                        recordMutableGen_GC((StgClosure *)p,gen->no);
-                        gct->failed_to_evac = rtsFalse;
-                    } else {
-                        tso->flags &= ~TSO_LINK_DIRTY;
-                    }
-                   continue;
-               }
-           }
-           default:
+           case MUT_ARR_PTRS_DIRTY:
+            {
+                rtsBool saved_eager_promotion;
+                saved_eager_promotion = gct->eager_promotion;
+                gct->eager_promotion = rtsFalse;
+
+                scavenge_mut_arr_ptrs_marked((StgMutArrPtrs *)p);
+
+                if (gct->failed_to_evac) {
+                    ((StgClosure *)p)->header.info = &stg_MUT_ARR_PTRS_DIRTY_info;
+                } else {
+                    ((StgClosure *)p)->header.info = &stg_MUT_ARR_PTRS_CLEAN_info;
+                }
+
+                gct->eager_promotion = saved_eager_promotion;
+                gct->failed_to_evac = rtsFalse;
+                recordMutableGen_GC((StgClosure *)p,gen_no);
+               continue;
+            }
+            default:
                ;
            }
 
            if (scavenge_one(p)) {
                // didn't manage to promote everything, so put the
                // object back on the list.
-               recordMutableGen_GC((StgClosure *)p,gen->no);
+                recordMutableGen_GC((StgClosure *)p,gen_no);
            }
        }
     }
@@ -1517,7 +1448,7 @@ scavenge_static(void)
 
   /* Always evacuate straight to the oldest generation for static
    * objects */
-  gct->evac_step = &oldest_gen->steps[0];
+  gct->evac_gen_no = oldest_gen->no;
 
   /* keep going until we've scavenged all the objects on the linked
      list... */
@@ -1602,23 +1533,21 @@ scavenge_static(void)
 static void
 scavenge_large_bitmap( StgPtr p, StgLargeBitmap *large_bitmap, nat size )
 {
-    nat i, b;
+    nat i, j, b;
     StgWord bitmap;
     
     b = 0;
-    bitmap = large_bitmap->bitmap[b];
-    for (i = 0; i < size; ) {
-       if ((bitmap & 1) == 0) {
-           evacuate((StgClosure **)p);
-       }
-       i++;
-       p++;
-       if (i % BITS_IN(W_) == 0) {
-           b++;
-           bitmap = large_bitmap->bitmap[b];
-       } else {
+
+    for (i = 0; i < size; b++) {
+        bitmap = large_bitmap->bitmap[b];
+        j = stg_min(size-i, BITS_IN(W_));
+        i += j;
+        for (; j > 0; j--, p++) {
+            if ((bitmap & 1) == 0) {
+                evacuate((StgClosure **)p);
+            }
            bitmap = bitmap >> 1;
-       }
+        }            
     }
 }
 
@@ -1678,10 +1607,12 @@ scavenge_stack(StgPtr p, StgPtr stack_end)
        // 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.
+       // the updatee would end up pointing to the value.  So we
+       // check whether the value after evacuation is a BLACKHOLE,
+       // and if not, we change the update frame to an stg_enter
+       // frame that simply returns the value.  Hence, blackholing is
+        // compulsory (otherwise we would have to check for thunks
+        // too).
         //
         // Note [upd-black-hole]
         // One slight hiccup is that the THUNK_SELECTOR machinery can
@@ -1692,22 +1623,17 @@ scavenge_stack(StgPtr p, StgPtr stack_end)
         // the updatee is never a THUNK_SELECTOR and we're ok.
         // NB. this is a new invariant: blackholing is not optional.
     {
-        nat type;
-        const StgInfoTable *i;
-        StgClosure *updatee;
-
-        updatee = ((StgUpdateFrame *)p)->updatee;
-        i = updatee->header.info;
-        if (!IS_FORWARDING_PTR(i)) {
-            type = get_itbl(updatee)->type;
-            if (type == IND) {
-                updatee->header.info = &stg_IND_PERM_info;
-            } else if (type == IND_OLDGEN) {
-                updatee->header.info = &stg_IND_OLDGEN_PERM_info;
-            }            
+        StgUpdateFrame *frame = (StgUpdateFrame *)p;
+        StgClosure *v;
+
+        evacuate(&frame->updatee);
+        v = frame->updatee;
+        if (GET_CLOSURE_TAG(v) != 0 ||
+            (get_itbl(v)->type != BLACKHOLE)) {
+            // blackholing is compulsory, see above.
+            frame->header.info = (const StgInfoTable*)&stg_enter_checkbh_info;
         }
-        evacuate(&((StgUpdateFrame *)p)->updatee);
-        ASSERT(GET_CLOSURE_TAG(((StgUpdateFrame *)p)->updatee) == 0);
+        ASSERT(v->header.info != &stg_TSO_info);
         p += sizeofW(StgUpdateFrame);
         continue;
     }
@@ -1716,6 +1642,7 @@ scavenge_stack(StgPtr p, StgPtr stack_end)
     case CATCH_STM_FRAME:
     case CATCH_RETRY_FRAME:
     case ATOMICALLY_FRAME:
+    case UNDERFLOW_FRAME:
     case STOP_FRAME:
     case CATCH_FRAME:
     case RET_SMALL:
@@ -1804,19 +1731,19 @@ scavenge_stack(StgPtr p, StgPtr stack_end)
 /*-----------------------------------------------------------------------------
   scavenge the large object list.
 
-  evac_step set by caller; similar games played with evac_step as with
+  evac_gen set by caller; similar games played with evac_gen as with
   scavenge() - see comment at the top of scavenge().  Most large
-  objects are (repeatedly) mutable, so most of the time evac_step will
+  objects are (repeatedly) mutable, so most of the time evac_gen will
   be zero.
   --------------------------------------------------------------------------- */
 
 static void
-scavenge_large (step_workspace *ws)
+scavenge_large (gen_workspace *ws)
 {
     bdescr *bd;
     StgPtr p;
 
-    gct->evac_step = ws->step;
+    gct->evac_gen_no = ws->gen->no;
 
     bd = ws->todo_large_objects;
     
@@ -1828,15 +1755,15 @@ scavenge_large (step_workspace *ws)
        // the front when evacuating.
        ws->todo_large_objects = bd->link;
        
-       ACQUIRE_SPIN_LOCK(&ws->step->sync_large_objects);
-       dbl_link_onto(bd, &ws->step->scavenged_large_objects);
-       ws->step->n_scavenged_large_blocks += bd->blocks;
-       RELEASE_SPIN_LOCK(&ws->step->sync_large_objects);
+       ACQUIRE_SPIN_LOCK(&ws->gen->sync_large_objects);
+       dbl_link_onto(bd, &ws->gen->scavenged_large_objects);
+       ws->gen->n_scavenged_large_blocks += bd->blocks;
+       RELEASE_SPIN_LOCK(&ws->gen->sync_large_objects);
        
        p = bd->start;
        if (scavenge_one(p)) {
-           if (ws->step->gen_no > 0) {
-               recordMutableGen_GC((StgClosure *)p, ws->step->gen_no);
+           if (ws->gen->no > 0) {
+               recordMutableGen_GC((StgClosure *)p, ws->gen->no);
            }
        }
 
@@ -1848,7 +1775,7 @@ scavenge_large (step_workspace *ws)
 /* ----------------------------------------------------------------------------
    Look for work to do.
 
-   We look for the oldest step that has either a todo block that can
+   We look for the oldest gen that has either a todo block that can
    be scanned, or a block of work on the global queue that we can
    scan.
 
@@ -1867,8 +1794,8 @@ scavenge_large (step_workspace *ws)
 static rtsBool
 scavenge_find_work (void)
 {
-    int s;
-    step_workspace *ws;
+    int g;
+    gen_workspace *ws;
     rtsBool did_something, did_anything;
     bdescr *bd;
 
@@ -1878,11 +1805,8 @@ scavenge_find_work (void)
 
 loop:
     did_something = rtsFalse;
-    for (s = total_steps-1; s >= 0; s--) {
-        if (s == 0 && RtsFlags.GcFlags.generations > 1) { 
-            continue; 
-        }
-        ws = &gct->steps[s];
+    for (g = RtsFlags.GcFlags.generations-1; g >= 0; g--) {
+        ws = &gct->gens[g];
         
         gct->scan_bd = NULL;
 
@@ -1917,11 +1841,8 @@ loop:
 #if defined(THREADED_RTS)
     if (work_stealing) {
         // look for work to steal
-        for (s = total_steps-1; s >= 0; s--) {
-            if (s == 0 && RtsFlags.GcFlags.generations > 1) { 
-                continue; 
-            }
-            if ((bd = steal_todo_block(s)) != NULL) {
+        for (g = RtsFlags.GcFlags.generations-1; g >= 0; g--) {
+            if ((bd = steal_todo_block(g)) != NULL) {
                 scavenge_block(bd);
                 did_something = rtsTrue;
                 break;
@@ -1959,8 +1880,7 @@ loop:
     }
     
     // scavenge objects in compacted generation
-    if (mark_stack_overflowed || oldgen_scan_bd != NULL ||
-       (mark_stack_bdescr != NULL && !mark_stack_empty())) {
+    if (mark_stack_bd != NULL && !mark_stack_empty()) {
        scavenge_mark_stack();
        work_to_do = rtsTrue;
     }