[project @ 2005-04-05 12:19:54 by simonmar]
[ghc-hetmet.git] / ghc / rts / Schedule.c
index b73f530..6e363a6 100644 (file)
@@ -1,7 +1,6 @@
 /* ---------------------------------------------------------------------------
- * $Id: Schedule.c,v 1.190 2004/02/27 15:56:25 simonmar Exp $
  *
- * (c) The GHC Team, 1998-2003
+ * (c) The GHC Team, 1998-2004
  *
  * Scheduler
  *
@@ -10,7 +9,7 @@
  *
  * WAY  Name     CPP flag  What's it for
  * --------------------------------------
- * mp   GUM      PAR          Parallel execution on a distrib. memory machine
+ * mp   GUM      PARALLEL_HASKELL          Parallel execution on a distrib. memory machine
  * s    SMP      SMP          Parallel execution on a shared memory machine
  * mg   GranSim  GRAN         Simulation of parallel execution
  * md   GUM/GdH  DIST         Distributed execution (based on GUM)
 #include "SchedAPI.h"
 #include "RtsUtils.h"
 #include "RtsFlags.h"
+#include "BlockAlloc.h"
+#include "OSThreads.h"
 #include "Storage.h"
 #include "StgRun.h"
-#include "StgStartup.h"
 #include "Hooks.h"
 #define COMPILING_SCHEDULER
 #include "Schedule.h"
 #include "StgMiscClosures.h"
-#include "Storage.h"
 #include "Interpreter.h"
 #include "Exception.h"
 #include "Printer.h"
 #include "Signals.h"
 #include "Sanity.h"
 #include "Stats.h"
+#include "STM.h"
 #include "Timer.h"
 #include "Prelude.h"
 #include "ThreadLabels.h"
+#include "LdvProfile.h"
+#include "Updates.h"
 #ifdef PROFILING
 #include "Proftimer.h"
 #include "ProfHeap.h"
 #endif
-#if defined(GRAN) || defined(PAR)
+#if defined(GRAN) || defined(PARALLEL_HASKELL)
 # include "GranSimRts.h"
 # include "GranSim.h"
 # include "ParallelRts.h"
@@ -74,7 +76,6 @@
 #endif
 #include "Sparks.h"
 #include "Capability.h"
-#include "OSThreads.h"
 #include  "Task.h"
 
 #ifdef HAVE_SYS_TYPES_H
 #include <errno.h>
 #endif
 
+// Turn off inlining when debugging - it obfuscates things
+#ifdef DEBUG
+# undef  STATIC_INLINE
+# define STATIC_INLINE static
+#endif
+
 #ifdef THREADED_RTS
 #define USED_IN_THREADED_RTS
 #else
  */
 StgMainThread *main_threads = NULL;
 
-/* Thread queues.
- * Locks required: sched_mutex.
- */
 #if defined(GRAN)
 
 StgTSO* ActiveTSO = NULL; /* for assigning system costs; GranSim-Light only */
@@ -134,14 +138,23 @@ StgTSO *ccalling_threadss[MAX_PROC];
 
 #else /* !GRAN */
 
+/* Thread queues.
+ * Locks required: sched_mutex.
+ */
 StgTSO *run_queue_hd = NULL;
 StgTSO *run_queue_tl = NULL;
 StgTSO *blocked_queue_hd = NULL;
 StgTSO *blocked_queue_tl = NULL;
+StgTSO *blackhole_queue = NULL;
 StgTSO *sleeping_queue = NULL;    /* perhaps replace with a hash table? */
 
 #endif
 
+/* The blackhole_queue should be checked for threads to wake up.  See
+ * Schedule.h for more thorough comment.
+ */
+rtsBool blackholes_need_checking = rtsFalse;
+
 /* Linked list of all threads.
  * Used for detecting garbage collected threads.
  */
@@ -153,19 +166,22 @@ StgTSO *all_threads = NULL;
  */
 static StgTSO *suspended_ccalling_threads;
 
-static StgTSO *threadStackOverflow(StgTSO *tso);
-
 /* KH: The following two flags are shared memory locations.  There is no need
        to lock them, since they are only unset at the end of a scheduler
        operation.
 */
 
 /* flag set by signal handler to precipitate a context switch */
-nat context_switch = 0;
+int context_switch = 0;
 
 /* if this flag is set as well, give up execution */
 rtsBool interrupted = rtsFalse;
 
+/* If this flag is set, we are running Haskell code.  Used to detect
+ * uses of 'foreign import unsafe' that should be 'safe'.
+ */
+static rtsBool in_haskell = rtsFalse;
+
 /* Next thread ID to allocate.
  * Locks required: thread_id_mutex
  */
@@ -201,6 +217,10 @@ StgTSO *CurrentTSO;
  */
 StgTSO dummy_tso;
 
+# if defined(SMP)
+static Condition gc_pending_cond = INIT_COND_VAR;
+# endif
+
 static rtsBool ready_to_gc;
 
 /*
@@ -210,13 +230,6 @@ static rtsBool ready_to_gc;
  */
 static rtsBool shutting_down_scheduler = rtsFalse;
 
-void            addToBlockedQueue ( StgTSO *tso );
-
-static void     schedule          ( StgMainThread *mainThread, Capability *initialCapability );
-       void     interruptStgRts   ( void );
-
-static void     detectBlackHoles  ( void );
-
 #if defined(RTS_SUPPORTS_THREADS)
 /* ToDo: carefully document the invariants that go together
  *       with these synchronisation objects.
@@ -226,7 +239,7 @@ Mutex     term_mutex        = INIT_MUTEX_VAR;
 
 #endif /* RTS_SUPPORTS_THREADS */
 
-#if defined(PAR)
+#if defined(PARALLEL_HASKELL)
 StgTSO *LastTSO;
 rtsTime TimeOfLastYield;
 rtsBool emitSchedule = rtsTrue;
@@ -234,6 +247,7 @@ rtsBool emitSchedule = rtsTrue;
 
 #if DEBUG
 static char *whatNext_strs[] = {
+  "(unknown)",
   "ThreadRunGHC",
   "ThreadInterpret",
   "ThreadKilled",
@@ -242,7 +256,66 @@ static char *whatNext_strs[] = {
 };
 #endif
 
-#if defined(PAR)
+/* -----------------------------------------------------------------------------
+ * static function prototypes
+ * -------------------------------------------------------------------------- */
+
+#if defined(RTS_SUPPORTS_THREADS)
+static void taskStart(void);
+#endif
+
+static void schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
+                     Capability *initialCapability );
+
+//
+// These function all encapsulate parts of the scheduler loop, and are
+// abstracted only to make the structure and control flow of the
+// scheduler clearer.
+//
+static void schedulePreLoop(void);
+static void scheduleHandleInterrupt(void);
+static void scheduleStartSignalHandlers(void);
+static void scheduleCheckBlockedThreads(void);
+static void scheduleCheckBlackHoles(void);
+static void scheduleDetectDeadlock(void);
+#if defined(GRAN)
+static StgTSO *scheduleProcessEvent(rtsEvent *event);
+#endif
+#if defined(PARALLEL_HASKELL)
+static StgTSO *scheduleSendPendingMessages(void);
+static void scheduleActivateSpark(void);
+static rtsBool scheduleGetRemoteWork(rtsBool *receivedFinish);
+#endif
+#if defined(PAR) || defined(GRAN)
+static void scheduleGranParReport(void);
+#endif
+static void schedulePostRunThread(void);
+static rtsBool scheduleHandleHeapOverflow( Capability *cap, StgTSO *t );
+static void scheduleHandleStackOverflow( StgTSO *t);
+static rtsBool scheduleHandleYield( StgTSO *t, nat prev_what_next );
+static void scheduleHandleThreadBlocked( StgTSO *t );
+static rtsBool scheduleHandleThreadFinished( StgMainThread *mainThread, 
+                                            Capability *cap, StgTSO *t );
+static void scheduleDoHeapProfile(void);
+static void scheduleDoGC(void);
+
+static void unblockThread(StgTSO *tso);
+static rtsBool checkBlackHoles(void);
+static SchedulerStatus waitThread_(/*out*/StgMainThread* m,
+                                  Capability *initialCapability
+                                  );
+static void scheduleThread_ (StgTSO* tso);
+static void AllRoots(evac_fn evac);
+
+static StgTSO *threadStackOverflow(StgTSO *tso);
+
+static void raiseAsync_(StgTSO *tso, StgClosure *exception, 
+                       rtsBool stop_at_atomically);
+
+static void printThreadBlockage(StgTSO *tso);
+static void printThreadStatus(StgTSO *tso);
+
+#if defined(PARALLEL_HASKELL)
 StgTSO * createSparkThread(rtsSpark spark);
 StgTSO * activateSpark (rtsSpark spark);  
 #endif
@@ -254,11 +327,11 @@ StgTSO * activateSpark (rtsSpark spark);
 #if defined(RTS_SUPPORTS_THREADS)
 static rtsBool startingWorkerThread = rtsFalse;
 
-static void taskStart(void);
 static void
 taskStart(void)
 {
   ACQUIRE_LOCK(&sched_mutex);
+  startingWorkerThread = rtsFalse;
   schedule(NULL,NULL);
   RELEASE_LOCK(&sched_mutex);
 }
@@ -275,12 +348,35 @@ startSchedulerTaskIfNecessary(void)
       // just because the last one hasn't yet reached the
       // "waiting for capability" state
       startingWorkerThread = rtsTrue;
-      startTask(taskStart);
+      if (!startTask(taskStart)) {
+         startingWorkerThread = rtsFalse;
+      }
     }
   }
 }
 #endif
 
+/* -----------------------------------------------------------------------------
+ * Putting a thread on the run queue: different scheduling policies
+ * -------------------------------------------------------------------------- */
+
+STATIC_INLINE void
+addToRunQueue( StgTSO *t )
+{
+#if defined(PARALLEL_HASKELL)
+    if (RtsFlags.ParFlags.doFairScheduling) { 
+       // this does round-robin scheduling; good for concurrency
+       APPEND_TO_RUN_QUEUE(t);
+    } else {
+       // this does unfair scheduling; good for parallelism
+       PUSH_ON_RUN_QUEUE(t);
+    }
+#else
+    // this does round-robin scheduling; good for concurrency
+    APPEND_TO_RUN_QUEUE(t);
+#endif
+}
+    
 /* ---------------------------------------------------------------------------
    Main scheduling loop.
 
@@ -316,18 +412,17 @@ startSchedulerTaskIfNecessary(void)
      This is not the ugliest code you could imagine, but it's bloody close.
 
    ------------------------------------------------------------------------ */
+
 static void
 schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
           Capability *initialCapability )
 {
   StgTSO *t;
-  Capability *cap = initialCapability;
+  Capability *cap;
   StgThreadReturnCode ret;
 #if defined(GRAN)
   rtsEvent *event;
-#elif defined(PAR)
-  StgSparkPool *pool;
-  rtsSpark spark;
+#elif defined(PARALLEL_HASKELL)
   StgTSO *tso;
   GlobalTaskId pe;
   rtsBool receivedFinish = rtsFalse;
@@ -335,62 +430,58 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
   nat tp_size, sp_size; // stats only
 # endif
 #endif
-  rtsBool was_interrupted = rtsFalse;
-  StgTSOWhatNext prev_what_next;
+  nat prev_what_next;
   
   // Pre-condition: sched_mutex is held.
-#if defined(RTS_SUPPORTS_THREADS)
-  //
-  // in the threaded case, the capability is either passed in via the
-  // initialCapability parameter, or initialized inside the scheduler
-  // loop 
-  //
-  IF_DEBUG(scheduler,
-          sched_belch("### NEW SCHEDULER LOOP (main thr: %p, cap: %p)",
-                      mainThread, initialCapability);
-      );
-#else
+  // We might have a capability, passed in as initialCapability.
+  cap = initialCapability;
+
+#if !defined(RTS_SUPPORTS_THREADS)
   // simply initialise it in the non-threaded case
   grabCapability(&cap);
 #endif
 
-#if defined(GRAN)
-  /* set up first event to get things going */
-  /* ToDo: assign costs for system setup and init MainTSO ! */
-  new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
-           ContinueThread, 
-           CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
-
-  IF_DEBUG(gran,
-          fprintf(stderr, "GRAN: Init CurrentTSO (in schedule) = %p\n", CurrentTSO);
-          G_TSO(CurrentTSO, 5));
-
-  if (RtsFlags.GranFlags.Light) {
-    /* Save current time; GranSim Light only */
-    CurrentTSO->gran.clock = CurrentTime[CurrentProc];
-  }      
+  IF_DEBUG(scheduler,
+          sched_belch("### NEW SCHEDULER LOOP (main thr: %p, cap: %p)",
+                      mainThread, initialCapability);
+      );
 
-  event = get_next_event();
+  schedulePreLoop();
 
-  while (event!=(rtsEvent*)NULL) {
-    /* Choose the processor with the next event */
-    CurrentProc = event->proc;
-    CurrentTSO = event->tso;
+  // -----------------------------------------------------------
+  // Scheduler loop starts here:
 
-#elif defined(PAR)
+#if defined(PARALLEL_HASKELL)
+#define TERMINATION_CONDITION        (!receivedFinish)
+#elif defined(GRAN)
+#define TERMINATION_CONDITION        ((event = get_next_event()) != (rtsEvent*)NULL) 
+#else
+#define TERMINATION_CONDITION        rtsTrue
+#endif
 
-  while (!receivedFinish) {    /* set by processMessages */
-                               /* when receiving PP_FINISH message         */ 
+  while (TERMINATION_CONDITION) {
 
-#else // everything except GRAN and PAR
+#if defined(GRAN)
+      /* Choose the processor with the next event */
+      CurrentProc = event->proc;
+      CurrentTSO = event->tso;
+#endif
 
-  while (1) {
+      IF_DEBUG(scheduler, printAllThreads());
 
+#if defined(SMP)
+      // 
+      // Wait until GC has completed, if necessary.
+      //
+      if (ready_to_gc) {
+         if (cap != NULL) {
+             releaseCapability(cap);
+             IF_DEBUG(scheduler,sched_belch("waiting for GC"));
+             waitCondition( &gc_pending_cond, &sched_mutex );
+         }
+      }
 #endif
 
-     IF_DEBUG(scheduler, printAllThreads());
-
 #if defined(RTS_SUPPORTS_THREADS)
       // Yield the capability to higher-priority tasks if necessary.
       //
@@ -408,238 +499,466 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
       // We now have a capability...
 #endif
 
+    // Check whether we have re-entered the RTS from Haskell without
+    // going via suspendThread()/resumeThread (i.e. a 'safe' foreign
+    // call).
+    if (in_haskell) {
+         errorBelch("schedule: re-entered unsafely.\n"
+                    "   Perhaps a 'foreign import unsafe' should be 'safe'?");
+         stg_exit(1);
+    }
+
+    scheduleHandleInterrupt();
+
+#if defined(not_yet) && defined(SMP)
     //
-    // If we're interrupted (the user pressed ^C, or some other
-    // termination condition occurred), kill all the currently running
-    // threads.
+    // Top up the run queue from our spark pool.  We try to make the
+    // number of threads in the run queue equal to the number of
+    // free capabilities.
     //
-    if (interrupted) {
-       IF_DEBUG(scheduler, sched_belch("interrupted"));
-       interrupted = rtsFalse;
-       was_interrupted = rtsTrue;
-#if defined(RTS_SUPPORTS_THREADS)
-       // In the threaded RTS, deadlock detection doesn't work,
-       // so just exit right away.
-       prog_belch("interrupted");
-       releaseCapability(cap);
-       RELEASE_LOCK(&sched_mutex);
-       shutdownHaskellAndExit(EXIT_SUCCESS);
-#else
-       deleteAllThreads();
-#endif
+    {
+       StgClosure *spark;
+       if (EMPTY_RUN_QUEUE()) {
+           spark = findSpark(rtsFalse);
+           if (spark == NULL) {
+               break; /* no more sparks in the pool */
+           } else {
+               createSparkThread(spark);         
+               IF_DEBUG(scheduler,
+                        sched_belch("==^^ turning spark of closure %p into a thread",
+                                    (StgClosure *)spark));
+           }
+       }
     }
+#endif // SMP
 
+    scheduleStartSignalHandlers();
+
+    // Only check the black holes here if we've nothing else to do.
+    // During normal execution, the black hole list only gets checked
+    // at GC time, to avoid repeatedly traversing this possibly long
+    // list each time around the scheduler.
+    if (EMPTY_RUN_QUEUE()) { scheduleCheckBlackHoles(); }
+
+    scheduleCheckBlockedThreads();
+
+    scheduleDetectDeadlock();
+
+    // Normally, the only way we can get here with no threads to
+    // run is if a keyboard interrupt received during 
+    // scheduleCheckBlockedThreads() or scheduleDetectDeadlock().
+    // Additionally, it is not fatal for the
+    // threaded RTS to reach here with no threads to run.
     //
-    // Go through the list of main threads and wake up any
-    // clients whose computations have finished.  ToDo: this
-    // should be done more efficiently without a linear scan
-    // of the main threads list, somehow...
-    //
-#if defined(RTS_SUPPORTS_THREADS)
-    { 
-       StgMainThread *m, **prev;
-       prev = &main_threads;
-       for (m = main_threads; m != NULL; prev = &m->link, m = m->link) {
-         if (m->tso->what_next == ThreadComplete
-             || m->tso->what_next == ThreadKilled)
-         {
-           if (m == mainThread)
-           {
-              if (m->tso->what_next == ThreadComplete)
-              {
-                if (m->ret)
-                {
-                  // NOTE: return val is tso->sp[1] (see StgStartup.hc)
-                  *(m->ret) = (StgClosure *)m->tso->sp[1]; 
-                }
-                m->stat = Success;
-              }
-              else
-              {
-                if (m->ret)
-                {
-                  *(m->ret) = NULL;
-                }
-                if (was_interrupted)
-                {
-                  m->stat = Interrupted;
-                }
-                else
-                {
-                  m->stat = Killed;
-                }
-              }
-              *prev = m->link;
-           
-#ifdef DEBUG
-             removeThreadLabel((StgWord)m->tso->id);
+    // win32: might be here due to awaitEvent() being abandoned
+    // as a result of a console event having been delivered.
+    if ( EMPTY_RUN_QUEUE() ) {
+#if !defined(RTS_SUPPORTS_THREADS) && !defined(mingw32_HOST_OS)
+       ASSERT(interrupted);
 #endif
-              releaseCapability(cap);
-              return;
-            }
-            else
-            {
-                // The current OS thread can not handle the fact that
-                // the Haskell thread "m" has ended.  "m" is bound;
-                // the scheduler loop in it's bound OS thread has to
-                // return, so let's pass our capability directly to
-                // that thread.
-               passCapability(&m->bound_thread_cond);
-               continue;
-            }
-          }
-       }
+       continue; // nothing to do
     }
-    
-#else /* not threaded */
 
-# if defined(PAR)
-    /* in GUM do this only on the Main PE */
-    if (IAmMainThread)
-# endif
-    /* If our main thread has finished or been killed, return.
-     */
-    {
-      StgMainThread *m = main_threads;
-      if (m->tso->what_next == ThreadComplete
-         || m->tso->what_next == ThreadKilled) {
-#ifdef DEBUG
-       removeThreadLabel((StgWord)m->tso->id);
-#endif
-       main_threads = main_threads->link;
-       if (m->tso->what_next == ThreadComplete) {
-           // We finished successfully, fill in the return value
-           // NOTE: return val is tso->sp[1] (see StgStartup.hc)
-           if (m->ret) { *(m->ret) = (StgClosure *)m->tso->sp[1]; };
-           m->stat = Success;
-           return;
-       } else {
-         if (m->ret) { *(m->ret) = NULL; };
-         if (was_interrupted) {
-           m->stat = Interrupted;
-         } else {
-           m->stat = Killed;
-         }
-         return;
-       }
-      }
+#if defined(PARALLEL_HASKELL)
+    scheduleSendPendingMessages();
+    if (EMPTY_RUN_QUEUE() && scheduleActivateSpark()) 
+       continue;
+
+#if defined(SPARKS)
+    ASSERT(next_fish_to_send_at==0);  // i.e. no delayed fishes left!
+#endif
+
+    /* If we still have no work we need to send a FISH to get a spark
+       from another PE */
+    if (EMPTY_RUN_QUEUE()) {
+       if (!scheduleGetRemoteWork(&receivedFinish)) continue;
+       ASSERT(rtsFalse); // should not happen at the moment
+    }
+    // from here: non-empty run queue.
+    //  TODO: merge above case with this, only one call processMessages() !
+    if (PacketsWaiting()) {  /* process incoming messages, if
+                               any pending...  only in else
+                               because getRemoteWork waits for
+                               messages as well */
+       receivedFinish = processMessages();
     }
 #endif
 
+#if defined(GRAN)
+    scheduleProcessEvent(event);
+#endif
 
-#if defined(RTS_USER_SIGNALS)
-    // check for signals each time around the scheduler
-    if (signals_pending()) {
-      RELEASE_LOCK(&sched_mutex); /* ToDo: kill */
-      startSignalHandlers();
-      ACQUIRE_LOCK(&sched_mutex);
-    }
+    // 
+    // Get a thread to run
+    //
+    ASSERT(run_queue_hd != END_TSO_QUEUE);
+    POP_RUN_QUEUE(t);
+
+#if defined(GRAN) || defined(PAR)
+    scheduleGranParReport(); // some kind of debuging output
+#else
+    // Sanity check the thread we're about to run.  This can be
+    // expensive if there is lots of thread switching going on...
+    IF_DEBUG(sanity,checkTSO(t));
 #endif
 
-    /* Check whether any waiting threads need to be woken up.  If the
-     * run queue is empty, and there are no other tasks running, we
-     * can wait indefinitely for something to happen.
-     */
-    if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue) 
 #if defined(RTS_SUPPORTS_THREADS)
-               || EMPTY_RUN_QUEUE()
-#endif
-        )
+    // Check whether we can run this thread in the current task.
+    // If not, we have to pass our capability to the right task.
     {
-      awaitEvent( EMPTY_RUN_QUEUE() );
+      StgMainThread *m = t->main;
+      
+      if(m)
+      {
+       if(m == mainThread)
+       {
+         IF_DEBUG(scheduler,
+           sched_belch("### Running thread %d in bound thread", t->id));
+         // yes, the Haskell thread is bound to the current native thread
+       }
+       else
+       {
+         IF_DEBUG(scheduler,
+           sched_belch("### thread %d bound to another OS thread", t->id));
+         // no, bound to a different Haskell thread: pass to that thread
+         PUSH_ON_RUN_QUEUE(t);
+         passCapability(&m->bound_thread_cond);
+         continue;
+       }
+      }
+      else
+      {
+       if(mainThread != NULL)
+        // The thread we want to run is bound.
+       {
+         IF_DEBUG(scheduler,
+           sched_belch("### this OS thread cannot run thread %d", t->id));
+         // no, the current native thread is bound to a different
+         // Haskell thread, so pass it to any worker thread
+         PUSH_ON_RUN_QUEUE(t);
+         passCapabilityToWorker();
+         continue; 
+       }
+      }
     }
-    /* we can be interrupted while waiting for I/O... */
-    if (interrupted) continue;
+#endif
 
-    /* 
-     * Detect deadlock: when we have no threads to run, there are no
-     * threads waiting on I/O or sleeping, and all the other tasks are
-     * waiting for work, we must have a deadlock of some description.
-     *
-     * We first try to find threads blocked on themselves (ie. black
-     * holes), and generate NonTermination exceptions where necessary.
-     *
-     * If no threads are black holed, we have a deadlock situation, so
-     * inform all the main threads.
+    cap->r.rCurrentTSO = t;
+    
+    /* context switches are now initiated by the timer signal, unless
+     * the user specified "context switch as often as possible", with
+     * +RTS -C0
      */
-#if !defined(PAR) && !defined(RTS_SUPPORTS_THREADS)
-    if (   EMPTY_THREAD_QUEUES() )
-    {
-       IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
-       // Garbage collection can release some new threads due to
-       // either (a) finalizers or (b) threads resurrected because
-       // they are about to be send BlockedOnDeadMVar.  Any threads
-       // thus released will be immediately runnable.
-       GarbageCollect(GetRoots,rtsTrue);
+    if ((RtsFlags.ConcFlags.ctxtSwitchTicks == 0
+        && (run_queue_hd != END_TSO_QUEUE
+            || blocked_queue_hd != END_TSO_QUEUE
+            || sleeping_queue != END_TSO_QUEUE)))
+       context_switch = 1;
 
-       if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
+run_thread:
 
-       IF_DEBUG(scheduler, 
-                sched_belch("still deadlocked, checking for black holes..."));
-       detectBlackHoles();
+    RELEASE_LOCK(&sched_mutex);
 
-       if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
+    IF_DEBUG(scheduler, sched_belch("-->> running thread %ld %s ...", 
+                             (long)t->id, whatNext_strs[t->what_next]));
 
-#if defined(RTS_USER_SIGNALS)
-       /* If we have user-installed signal handlers, then wait
-        * for signals to arrive rather then bombing out with a
-        * deadlock.
-        */
-       if ( anyUserHandlers() ) {
-           IF_DEBUG(scheduler, 
-                    sched_belch("still deadlocked, waiting for signals..."));
+#if defined(PROFILING)
+    startHeapProfTimer();
+#endif
 
-           awaitUserSignals();
+    // ----------------------------------------------------------------------
+    // Run the current thread 
 
-           // we might be interrupted...
-           if (interrupted) { continue; }
+    prev_what_next = t->what_next;
 
-           if (signals_pending()) {
-               RELEASE_LOCK(&sched_mutex);
-               startSignalHandlers();
-               ACQUIRE_LOCK(&sched_mutex);
-           }
-           ASSERT(!EMPTY_RUN_QUEUE());
-           goto not_deadlocked;
-       }
-#endif
+    errno = t->saved_errno;
+    in_haskell = rtsTrue;
 
-       /* Probably a real deadlock.  Send the current main thread the
-        * Deadlock exception (or in the SMP build, send *all* main
-        * threads the deadlock exception, since none of them can make
-        * progress).
-        */
-       {
-           StgMainThread *m;
-           m = main_threads;
-           switch (m->tso->why_blocked) {
-           case BlockedOnBlackHole:
-               raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
-               break;
-           case BlockedOnException:
-           case BlockedOnMVar:
-               raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
-               break;
-           default:
-               barf("deadlock: main thread blocked in a strange way");
-           }
-       }
-    }
-  not_deadlocked:
+    switch (prev_what_next) {
 
-#elif defined(RTS_SUPPORTS_THREADS)
-    // ToDo: add deadlock detection in threaded RTS
-#elif defined(PAR)
-    // ToDo: add deadlock detection in GUM (similar to SMP) -- HWL
-#endif
+    case ThreadKilled:
+    case ThreadComplete:
+       /* Thread already finished, return to scheduler. */
+       ret = ThreadFinished;
+       break;
 
-#if defined(RTS_SUPPORTS_THREADS)
-    if ( EMPTY_RUN_QUEUE() ) {
-      continue; // nothing to do
+    case ThreadRunGHC:
+       ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
+       break;
+
+    case ThreadInterpret:
+       ret = interpretBCO(cap);
+       break;
+
+    default:
+      barf("schedule: invalid what_next field");
+    }
+
+    // We have run some Haskell code: there might be blackhole-blocked
+    // threads to wake up now.
+    if ( blackhole_queue != END_TSO_QUEUE ) {
+       blackholes_need_checking = rtsTrue;
+    }
+
+    in_haskell = rtsFalse;
+
+    // The TSO might have moved, eg. if it re-entered the RTS and a GC
+    // happened.  So find the new location:
+    t = cap->r.rCurrentTSO;
+
+    // And save the current errno in this thread.
+    t->saved_errno = errno;
+
+    // ----------------------------------------------------------------------
+    
+    /* Costs for the scheduler are assigned to CCS_SYSTEM */
+#if defined(PROFILING)
+    stopHeapProfTimer();
+    CCCS = CCS_SYSTEM;
+#endif
+    
+    ACQUIRE_LOCK(&sched_mutex);
+    
+#if defined(RTS_SUPPORTS_THREADS)
+    IF_DEBUG(scheduler,debugBelch("sched (task %p): ", osThreadId()););
+#elif !defined(GRAN) && !defined(PARALLEL_HASKELL)
+    IF_DEBUG(scheduler,debugBelch("sched: "););
+#endif
+    
+    schedulePostRunThread();
+
+    switch (ret) {
+    case HeapOverflow:
+       ready_to_gc = scheduleHandleHeapOverflow(cap,t);
+       break;
+
+    case StackOverflow:
+       scheduleHandleStackOverflow(t);
+       break;
+
+    case ThreadYielding:
+       if (scheduleHandleYield(t, prev_what_next)) {
+            // shortcut for switching between compiler/interpreter:
+           goto run_thread; 
+       }
+       break;
+
+    case ThreadBlocked:
+       scheduleHandleThreadBlocked(t);
+       threadPaused(t);
+       break;
+
+    case ThreadFinished:
+       if (scheduleHandleThreadFinished(mainThread, cap, t)) return;;
+       break;
+
+    default:
+      barf("schedule: invalid thread return code %d", (int)ret);
+    }
+
+    scheduleDoHeapProfile();
+    scheduleDoGC();
+  } /* end of while() */
+
+  IF_PAR_DEBUG(verbose,
+              debugBelch("== Leaving schedule() after having received Finish\n"));
+}
+
+/* ----------------------------------------------------------------------------
+ * Setting up the scheduler loop
+ * ASSUMES: sched_mutex
+ * ------------------------------------------------------------------------- */
+
+static void
+schedulePreLoop(void)
+{
+#if defined(GRAN) 
+    /* set up first event to get things going */
+    /* ToDo: assign costs for system setup and init MainTSO ! */
+    new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
+             ContinueThread, 
+             CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
+    
+    IF_DEBUG(gran,
+            debugBelch("GRAN: Init CurrentTSO (in schedule) = %p\n", 
+                       CurrentTSO);
+            G_TSO(CurrentTSO, 5));
+    
+    if (RtsFlags.GranFlags.Light) {
+       /* Save current time; GranSim Light only */
+       CurrentTSO->gran.clock = CurrentTime[CurrentProc];
+    }      
+#endif
+}
+
+/* ----------------------------------------------------------------------------
+ * Deal with the interrupt flag
+ * ASSUMES: sched_mutex
+ * ------------------------------------------------------------------------- */
+
+static
+void scheduleHandleInterrupt(void)
+{
+    //
+    // Test for interruption.  If interrupted==rtsTrue, then either
+    // we received a keyboard interrupt (^C), or the scheduler is
+    // trying to shut down all the tasks (shutting_down_scheduler) in
+    // the threaded RTS.
+    //
+    if (interrupted) {
+       if (shutting_down_scheduler) {
+           IF_DEBUG(scheduler, sched_belch("shutting down"));
+#if defined(RTS_SUPPORTS_THREADS)
+           shutdownThread();
+#endif
+       } else {
+           IF_DEBUG(scheduler, sched_belch("interrupted"));
+           deleteAllThreads();
+       }
+    }
+}
+
+/* ----------------------------------------------------------------------------
+ * Start any pending signal handlers
+ * ASSUMES: sched_mutex
+ * ------------------------------------------------------------------------- */
+
+static void
+scheduleStartSignalHandlers(void)
+{
+#if defined(RTS_USER_SIGNALS)
+    if (signals_pending()) {
+      RELEASE_LOCK(&sched_mutex); /* ToDo: kill */
+      startSignalHandlers();
+      ACQUIRE_LOCK(&sched_mutex);
+    }
+#endif
+}
+
+/* ----------------------------------------------------------------------------
+ * Check for blocked threads that can be woken up.
+ * ASSUMES: sched_mutex
+ * ------------------------------------------------------------------------- */
+
+static void
+scheduleCheckBlockedThreads(void)
+{
+    //
+    // Check whether any waiting threads need to be woken up.  If the
+    // run queue is empty, and there are no other tasks running, we
+    // can wait indefinitely for something to happen.
+    //
+    if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue) )
+    {
+#if defined(RTS_SUPPORTS_THREADS)
+       // We shouldn't be here...
+       barf("schedule: awaitEvent() in threaded RTS");
+#endif
+       awaitEvent( EMPTY_RUN_QUEUE() && !blackholes_need_checking );
+    }
+}
+
+
+/* ----------------------------------------------------------------------------
+ * Check for threads blocked on BLACKHOLEs that can be woken up
+ * ASSUMES: sched_mutex
+ * ------------------------------------------------------------------------- */
+static void
+scheduleCheckBlackHoles( void )
+{
+    if ( blackholes_need_checking )
+    {
+       checkBlackHoles();
+       blackholes_need_checking = rtsFalse;
     }
+}
+
+/* ----------------------------------------------------------------------------
+ * Detect deadlock conditions and attempt to resolve them.
+ * ASSUMES: sched_mutex
+ * ------------------------------------------------------------------------- */
+
+static void
+scheduleDetectDeadlock(void)
+{
+    /* 
+     * Detect deadlock: when we have no threads to run, there are no
+     * threads blocked, waiting for I/O, or sleeping, and all the
+     * other tasks are waiting for work, we must have a deadlock of
+     * some description.
+     */
+    if ( EMPTY_THREAD_QUEUES() )
+    {
+#if !defined(PARALLEL_HASKELL) && !defined(RTS_SUPPORTS_THREADS)
+       IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
+
+       // Garbage collection can release some new threads due to
+       // either (a) finalizers or (b) threads resurrected because
+       // they are unreachable and will therefore be sent an
+       // exception.  Any threads thus released will be immediately
+       // runnable.
+       GarbageCollect(GetRoots,rtsTrue);
+       if ( !EMPTY_RUN_QUEUE() ) return;
+
+#if defined(RTS_USER_SIGNALS)
+       /* If we have user-installed signal handlers, then wait
+        * for signals to arrive rather then bombing out with a
+        * deadlock.
+        */
+       if ( anyUserHandlers() ) {
+           IF_DEBUG(scheduler, 
+                    sched_belch("still deadlocked, waiting for signals..."));
+
+           awaitUserSignals();
+
+           if (signals_pending()) {
+               RELEASE_LOCK(&sched_mutex);
+               startSignalHandlers();
+               ACQUIRE_LOCK(&sched_mutex);
+           }
+
+           // either we have threads to run, or we were interrupted:
+           ASSERT(!EMPTY_RUN_QUEUE() || interrupted);
+       }
+#endif
+
+       /* Probably a real deadlock.  Send the current main thread the
+        * Deadlock exception (or in the SMP build, send *all* main
+        * threads the deadlock exception, since none of them can make
+        * progress).
+        */
+       {
+           StgMainThread *m;
+           m = main_threads;
+           switch (m->tso->why_blocked) {
+           case BlockedOnBlackHole:
+           case BlockedOnException:
+           case BlockedOnMVar:
+               raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
+               return;
+           default:
+               barf("deadlock: main thread blocked in a strange way");
+           }
+       }
+
+#elif defined(RTS_SUPPORTS_THREADS)
+    // ToDo: add deadlock detection in threaded RTS
+#elif defined(PARALLEL_HASKELL)
+    // ToDo: add deadlock detection in GUM (similar to SMP) -- HWL
 #endif
+    }
+}
+
+/* ----------------------------------------------------------------------------
+ * Process an event (GRAN only)
+ * ------------------------------------------------------------------------- */
 
 #if defined(GRAN)
+static StgTSO *
+scheduleProcessEvent(rtsEvent *event)
+{
+    StgTSO *t;
+
     if (RtsFlags.GranFlags.Light)
       GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
 
@@ -652,13 +971,13 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
     if (!RtsFlags.GranFlags.Light)
       handleIdlePEs();
 
-    IF_DEBUG(gran, fprintf(stderr, "GRAN: switch by event-type\n"));
+    IF_DEBUG(gran, debugBelch("GRAN: switch by event-type\n"));
 
     /* main event dispatcher in GranSim */
     switch (event->evttype) {
       /* Should just be continuing execution */
     case ContinueThread:
-      IF_DEBUG(gran, fprintf(stderr, "GRAN: doing ContinueThread\n"));
+      IF_DEBUG(gran, debugBelch("GRAN: doing ContinueThread\n"));
       /* ToDo: check assertion
       ASSERT(run_queue_hd != (StgTSO*)NULL &&
             run_queue_hd != END_TSO_QUEUE);
@@ -666,25 +985,25 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
       /* Ignore ContinueThreads for fetching threads (if synchr comm) */
       if (!RtsFlags.GranFlags.DoAsyncFetch &&
          procStatus[CurrentProc]==Fetching) {
-       belch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]",
+       debugBelch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]\n",
              CurrentTSO->id, CurrentTSO, CurrentProc);
        goto next_thread;
       }        
       /* Ignore ContinueThreads for completed threads */
       if (CurrentTSO->what_next == ThreadComplete) {
-       belch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)", 
+       debugBelch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)\n", 
              CurrentTSO->id, CurrentTSO, CurrentProc);
        goto next_thread;
       }        
       /* Ignore ContinueThreads for threads that are being migrated */
       if (PROCS(CurrentTSO)==Nowhere) { 
-       belch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)",
+       debugBelch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)\n",
              CurrentTSO->id, CurrentTSO, CurrentProc);
        goto next_thread;
       }
       /* The thread should be at the beginning of the run queue */
       if (CurrentTSO!=run_queue_hds[CurrentProc]) { 
-       belch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread",
+       debugBelch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread\n",
              CurrentTSO->id, CurrentTSO, CurrentProc);
        break; // run the thread anyway
       }
@@ -741,14 +1060,14 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
     
     /* This point was scheduler_loop in the old RTS */
 
-    IF_DEBUG(gran, belch("GRAN: after main switch"));
+    IF_DEBUG(gran, debugBelch("GRAN: after main switch\n"));
 
     TimeOfLastEvent = CurrentTime[CurrentProc];
     TimeOfNextEvent = get_time_of_next_event();
     IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
     // CurrentTSO = ThreadQueueHd;
 
-    IF_DEBUG(gran, belch("GRAN: time of next event is: %ld", 
+    IF_DEBUG(gran, debugBelch("GRAN: time of next event is: %ld\n", 
                         TimeOfNextEvent));
 
     if (RtsFlags.GranFlags.Light) 
@@ -757,7 +1076,7 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
     EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
 
     IF_DEBUG(gran, 
-            belch("GRAN: end of time-slice is %#lx", EndOfTimeSlice));
+            debugBelch("GRAN: end of time-slice is %#lx\n", EndOfTimeSlice));
 
     /* in a GranSim setup the TSO stays on the run queue */
     t = CurrentTSO;
@@ -765,7 +1084,7 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
     POP_RUN_QUEUE(t); // take_off_run_queue(t);
 
     IF_DEBUG(gran, 
-            fprintf(stderr, "GRAN: About to run current thread, which is\n");
+            debugBelch("GRAN: About to run current thread, which is\n");
             G_TSO(t,5));
 
     context_switch = 0; // turned on via GranYield, checking events and time slice
@@ -774,18 +1093,54 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
             DumpGranEvent(GR_SCHEDULE, t));
 
     procStatus[CurrentProc] = Busy;
+}
+#endif // GRAN
 
-#elif defined(PAR)
+/* ----------------------------------------------------------------------------
+ * Send pending messages (PARALLEL_HASKELL only)
+ * ------------------------------------------------------------------------- */
+
+#if defined(PARALLEL_HASKELL)
+static StgTSO *
+scheduleSendPendingMessages(void)
+{
+    StgSparkPool *pool;
+    rtsSpark spark;
+    StgTSO *t;
+
+# if defined(PAR) // global Mem.Mgmt., omit for now
     if (PendingFetches != END_BF_QUEUE) {
         processFetches();
     }
+# endif
+    
+    if (RtsFlags.ParFlags.BufferTime) {
+       // if we use message buffering, we must send away all message
+       // packets which have become too old...
+       sendOldBuffers(); 
+    }
+}
+#endif
+
+/* ----------------------------------------------------------------------------
+ * Activate spark threads (PARALLEL_HASKELL only)
+ * ------------------------------------------------------------------------- */
+
+#if defined(PARALLEL_HASKELL)
+static void
+scheduleActivateSpark(void)
+{
+#if defined(SPARKS)
+  ASSERT(EMPTY_RUN_QUEUE());
+/* We get here if the run queue is empty and want some work.
+   We try to turn a spark into a thread, and add it to the run queue,
+   from where it will be picked up in the next iteration of the scheduler
+   loop.
+*/
 
-    /* ToDo: phps merge with spark activation above */
-    /* check whether we have local work and send requests if we have none */
-    if (EMPTY_RUN_QUEUE()) {  /* no runnable threads */
       /* :-[  no local threads => look out for local sparks */
       /* the spark pool for the current PE */
-      pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
+      pool = &(cap.r.rSparks); // JB: cap = (old) MainCap
       if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
          pool->hd < pool->tl) {
        /* 
@@ -797,29 +1152,73 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
         * thread... 
         */
 
-       spark = findSpark(rtsFalse);                /* get a spark */
-       if (spark != (rtsSpark) NULL) {
-         tso = activateSpark(spark);       /* turn the spark into a thread */
-         IF_PAR_DEBUG(schedule,
-                      belch("==== schedule: Created TSO %d (%p); %d threads active",
-                            tso->id, tso, advisory_thread_count));
-
-         if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
-           belch("==^^ failed to activate spark");
-           goto next_thread;
-         }               /* otherwise fall through & pick-up new tso */
-       } else {
-         IF_PAR_DEBUG(verbose,
-                      belch("==^^ no local sparks (spark pool contains only NFs: %d)", 
-                            spark_queue_len(pool)));
-         goto next_thread;
+       spark = findSpark(rtsFalse);            /* get a spark */
+       if (spark != (rtsSpark) NULL) {
+         tso = createThreadFromSpark(spark);       /* turn the spark into a thread */
+         IF_PAR_DEBUG(fish, // schedule,
+                      debugBelch("==== schedule: Created TSO %d (%p); %d threads active\n",
+                            tso->id, tso, advisory_thread_count));
+
+         if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
+           IF_PAR_DEBUG(fish, // schedule,
+                        debugBelch("==^^ failed to create thread from spark @ %lx\n",
+                            spark));
+           return rtsFalse; /* failed to generate a thread */
+         }                  /* otherwise fall through & pick-up new tso */
+       } else {
+         IF_PAR_DEBUG(fish, // schedule,
+                      debugBelch("==^^ no local sparks (spark pool contains only NFs: %d)\n", 
+                            spark_queue_len(pool)));
+         return rtsFalse;  /* failed to generate a thread */
+       }
+       return rtsTrue;  /* success in generating a thread */
+  } else { /* no more threads permitted or pool empty */
+    return rtsFalse;  /* failed to generateThread */
+  }
+#else
+  tso = NULL; // avoid compiler warning only
+  return rtsFalse;  /* dummy in non-PAR setup */
+#endif // SPARKS
+}
+#endif // PARALLEL_HASKELL
+
+/* ----------------------------------------------------------------------------
+ * Get work from a remote node (PARALLEL_HASKELL only)
+ * ------------------------------------------------------------------------- */
+    
+#if defined(PARALLEL_HASKELL)
+static rtsBool
+scheduleGetRemoteWork(rtsBool *receivedFinish)
+{
+  ASSERT(EMPTY_RUN_QUEUE());
+
+  if (RtsFlags.ParFlags.BufferTime) {
+       IF_PAR_DEBUG(verbose, 
+               debugBelch("...send all pending data,"));
+        {
+         nat i;
+         for (i=1; i<=nPEs; i++)
+           sendImmediately(i); // send all messages away immediately
        }
-      }
+  }
+# ifndef SPARKS
+       //++EDEN++ idle() , i.e. send all buffers, wait for work
+       // suppress fishing in EDEN... just look for incoming messages
+       // (blocking receive)
+  IF_PAR_DEBUG(verbose, 
+              debugBelch("...wait for incoming messages...\n"));
+  *receivedFinish = processMessages(); // blocking receive...
+
+       // and reenter scheduling loop after having received something
+       // (return rtsFalse below)
+
+# else /* activate SPARKS machinery */
+/* We get here, if we have no work, tried to activate a local spark, but still
+   have no work. We try to get a remote spark, by sending a FISH message.
+   Thread migration should be added here, and triggered when a sequence of 
+   fishes returns without work. */
+       delay = (RtsFlags.ParFlags.fishDelay!=0ll ? RtsFlags.ParFlags.fishDelay : 0ll);
 
-      /* If we still have no work we need to send a FISH to get a spark
-        from another PE 
-      */
-      if (EMPTY_RUN_QUEUE()) {
       /* =8-[  no local sparks => look for work on other PEs */
        /*
         * We really have absolutely no work.  Send out a fish
@@ -829,48 +1228,106 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
         * we're hoping to see.  (Of course, we still have to
         * respond to other types of messages.)
         */
-       TIME now = msTime() /*CURRENT_TIME*/;
+       rtsTime now = msTime() /*CURRENT_TIME*/;
        IF_PAR_DEBUG(verbose, 
-                    belch("--  now=%ld", now));
-       IF_PAR_DEBUG(verbose,
-                    if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
-                        (last_fish_arrived_at!=0 &&
-                         last_fish_arrived_at+RtsFlags.ParFlags.fishDelay > now)) {
-                      belch("--$$ delaying FISH until %ld (last fish %ld, delay %ld, now %ld)",
-                            last_fish_arrived_at+RtsFlags.ParFlags.fishDelay,
-                            last_fish_arrived_at,
-                            RtsFlags.ParFlags.fishDelay, now);
-                    });
-       
+                    debugBelch("--  now=%ld\n", now));
+       IF_PAR_DEBUG(fish, // verbose,
+            if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
+                (last_fish_arrived_at!=0 &&
+                 last_fish_arrived_at+delay > now)) {
+              debugBelch("--$$ <%llu> delaying FISH until %llu (last fish %llu, delay %llu)\n",
+                    now, last_fish_arrived_at+delay, 
+                    last_fish_arrived_at,
+                    delay);
+            });
+  
        if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
-           (last_fish_arrived_at==0 ||
-            (last_fish_arrived_at+RtsFlags.ParFlags.fishDelay <= now))) {
-         /* outstandingFishes is set in sendFish, processFish;
-            avoid flooding system with fishes via delay */
-         pe = choosePE();
-         sendFish(pe, mytid, NEW_FISH_AGE, NEW_FISH_HISTORY, 
-                  NEW_FISH_HUNGER);
-
-         // Global statistics: count no. of fishes
-         if (RtsFlags.ParFlags.ParStats.Global &&
-             RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
-           globalParStats.tot_fish_mess++;
-         }
-       }
-      
-       receivedFinish = processMessages();
-       goto next_thread;
+           advisory_thread_count < RtsFlags.ParFlags.maxThreads) { // send a FISH, but when?
+         if (last_fish_arrived_at==0 ||
+             (last_fish_arrived_at+delay <= now)) {           // send FISH now!
+           /* outstandingFishes is set in sendFish, processFish;
+              avoid flooding system with fishes via delay */
+    next_fish_to_send_at = 0;  
+  } else {
+    /* ToDo: this should be done in the main scheduling loop to avoid the
+             busy wait here; not so bad if fish delay is very small  */
+    int iq = 0; // DEBUGGING -- HWL
+    next_fish_to_send_at = last_fish_arrived_at+delay; // remember when to send  
+    /* send a fish when ready, but process messages that arrive in the meantime */
+    do {
+      if (PacketsWaiting()) {
+        iq++; // DEBUGGING
+        *receivedFinish = processMessages();
       }
-    } else if (PacketsWaiting()) {  /* Look for incoming messages */
-      receivedFinish = processMessages();
-    }
+      now = msTime();
+    } while (!*receivedFinish || now<next_fish_to_send_at);
+    // JB: This means the fish could become obsolete, if we receive
+    // work. Better check for work again? 
+    // last line: while (!receivedFinish || !haveWork || now<...)
+    // next line: if (receivedFinish || haveWork )
+
+    if (*receivedFinish) // no need to send a FISH if we are finishing anyway
+      return rtsFalse;  // NB: this will leave scheduler loop
+                       // immediately after return!
+                         
+    IF_PAR_DEBUG(fish, // verbose,
+              debugBelch("--$$ <%llu> sent delayed fish (%d processMessages); active/total threads=%d/%d\n",now,iq,run_queue_len(),advisory_thread_count));
 
-    /* Now we are sure that we have some work available */
-    ASSERT(run_queue_hd != END_TSO_QUEUE);
+  }
 
-    /* Take a thread from the run queue, if we have work */
-    POP_RUN_QUEUE(t);  // take_off_run_queue(END_TSO_QUEUE);
-    IF_DEBUG(sanity,checkTSO(t));
+    // JB: IMHO, this should all be hidden inside sendFish(...)
+    /* pe = choosePE(); 
+       sendFish(pe, thisPE, NEW_FISH_AGE, NEW_FISH_HISTORY, 
+                NEW_FISH_HUNGER);
+
+    // Global statistics: count no. of fishes
+    if (RtsFlags.ParFlags.ParStats.Global &&
+         RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
+          globalParStats.tot_fish_mess++;
+          }
+    */ 
+
+  /* delayed fishes must have been sent by now! */
+  next_fish_to_send_at = 0;  
+  }
+      
+  *receivedFinish = processMessages();
+# endif /* SPARKS */
+
+ return rtsFalse;
+ /* NB: this function always returns rtsFalse, meaning the scheduler
+    loop continues with the next iteration; 
+    rationale: 
+      return code means success in finding work; we enter this function
+      if there is no local work, thus have to send a fish which takes
+      time until it arrives with work; in the meantime we should process
+      messages in the main loop;
+ */
+}
+#endif // PARALLEL_HASKELL
+
+/* ----------------------------------------------------------------------------
+ * PAR/GRAN: Report stats & debugging info(?)
+ * ------------------------------------------------------------------------- */
+
+#if defined(PAR) || defined(GRAN)
+static void
+scheduleGranParReport(void)
+{
+  ASSERT(run_queue_hd != END_TSO_QUEUE);
+
+  /* Take a thread from the run queue, if we have work */
+  POP_RUN_QUEUE(t);  // take_off_run_queue(END_TSO_QUEUE);
+
+    /* If this TSO has got its outport closed in the meantime, 
+     *   it mustn't be run. Instead, we have to clean it up as if it was finished.
+     * It has to be marked as TH_DEAD for this purpose.
+     * If it is TH_TERM instead, it is supposed to have finished in the normal way.
+
+JB: TODO: investigate wether state change field could be nuked
+     entirely and replaced by the normal tso state (whatnext
+     field). All we want to do is to kill tsos from outside.
+     */
 
     /* ToDo: write something to the log-file
     if (RTSflags.ParFlags.granSimStats && !sameThread)
@@ -879,25 +1336,20 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
     CurrentTSO = t;
     */
     /* the spark pool for the current PE */
-    pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
+    pool = &(cap.r.rSparks); //  cap = (old) MainCap
 
     IF_DEBUG(scheduler, 
-            belch("--=^ %d threads, %d sparks on [%#x]", 
+            debugBelch("--=^ %d threads, %d sparks on [%#x]\n", 
                   run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
 
-# if 1
-    if (0 && RtsFlags.ParFlags.ParStats.Full && 
-       t && LastTSO && t->id != LastTSO->id && 
-       LastTSO->why_blocked == NotBlocked && 
-       LastTSO->what_next != ThreadComplete) {
-      // if previously scheduled TSO not blocked we have to record the context switch
-      DumpVeryRawGranEvent(TimeOfLastYield, CURRENT_PROC, CURRENT_PROC,
-                          GR_DESCHEDULE, LastTSO, (StgClosure *)NULL, 0, 0);
-    }
+    IF_PAR_DEBUG(fish,
+            debugBelch("--=^ %d threads, %d sparks on [%#x]\n", 
+                  run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
 
     if (RtsFlags.ParFlags.ParStats.Full && 
-       (emitSchedule /* forced emit */ ||
-        (t && LastTSO && t->id != LastTSO->id))) {
+       (t->par.sparkname != (StgInt)0) && // only log spark generated threads
+       (emitSchedule || // forced emit
+         (t && LastTSO && t->id != LastTSO->id))) {
       /* 
         we are running a different TSO, so write a schedule event to log file
         NB: If we use fair scheduling we also have to write  a deschedule 
@@ -905,130 +1357,24 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
             previous tso has blocked whenever we switch to another tso, so
             we don't need it in GUM for now
       */
+      IF_PAR_DEBUG(fish, // schedule,
+                  debugBelch("____ scheduling spark generated thread %d (%lx) (%lx) via a forced emit\n",t->id,t,t->par.sparkname));
+
       DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
                       GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
       emitSchedule = rtsFalse;
     }
-     
-# endif
-#else /* !GRAN && !PAR */
-  
-    // grab a thread from the run queue
-    ASSERT(run_queue_hd != END_TSO_QUEUE);
-    POP_RUN_QUEUE(t);
+}     
+#endif
 
-    // Sanity check the thread we're about to run.  This can be
-    // expensive if there is lots of thread switching going on...
-    IF_DEBUG(sanity,checkTSO(t));
-#endif
-
-#ifdef THREADED_RTS
-    {
-      StgMainThread *m;
-      for(m = main_threads; m; m = m->link)
-      {
-       if(m->tso == t)
-         break;
-      }
-      
-      if(m)
-      {
-       if(m == mainThread)
-       {
-         IF_DEBUG(scheduler,
-           sched_belch("### Running thread %d in bound thread", t->id));
-         // yes, the Haskell thread is bound to the current native thread
-       }
-       else
-       {
-         IF_DEBUG(scheduler,
-           sched_belch("### thread %d bound to another OS thread", t->id));
-         // no, bound to a different Haskell thread: pass to that thread
-         PUSH_ON_RUN_QUEUE(t);
-         passCapability(&m->bound_thread_cond);
-         continue;
-       }
-      }
-      else
-      {
-       if(mainThread != NULL)
-        // The thread we want to run is bound.
-       {
-         IF_DEBUG(scheduler,
-           sched_belch("### this OS thread cannot run thread %d", t->id));
-         // no, the current native thread is bound to a different
-         // Haskell thread, so pass it to any worker thread
-         PUSH_ON_RUN_QUEUE(t);
-         passCapabilityToWorker();
-         continue; 
-       }
-      }
-    }
-#endif
-
-    cap->r.rCurrentTSO = t;
-    
-    /* context switches are now initiated by the timer signal, unless
-     * the user specified "context switch as often as possible", with
-     * +RTS -C0
-     */
-    if ((RtsFlags.ConcFlags.ctxtSwitchTicks == 0
-        && (run_queue_hd != END_TSO_QUEUE
-            || blocked_queue_hd != END_TSO_QUEUE
-            || sleeping_queue != END_TSO_QUEUE)))
-       context_switch = 1;
-    else
-       context_switch = 0;
-
-run_thread:
-
-    RELEASE_LOCK(&sched_mutex);
-
-    IF_DEBUG(scheduler, sched_belch("-->> running thread %ld %s ...", 
-                             t->id, whatNext_strs[t->what_next]));
-
-#ifdef PROFILING
-    startHeapProfTimer();
-#endif
+/* ----------------------------------------------------------------------------
+ * After running a thread...
+ * ASSUMES: sched_mutex
+ * ------------------------------------------------------------------------- */
 
-    /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
-    /* Run the current thread 
-     */
-    prev_what_next = t->what_next;
-    switch (prev_what_next) {
-    case ThreadKilled:
-    case ThreadComplete:
-       /* Thread already finished, return to scheduler. */
-       ret = ThreadFinished;
-       break;
-    case ThreadRunGHC:
-       errno = t->saved_errno;
-       ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
-       t->saved_errno = errno;
-       break;
-    case ThreadInterpret:
-       ret = interpretBCO(cap);
-       break;
-    default:
-      barf("schedule: invalid what_next field");
-    }
-    /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
-    
-    /* Costs for the scheduler are assigned to CCS_SYSTEM */
-#ifdef PROFILING
-    stopHeapProfTimer();
-    CCCS = CCS_SYSTEM;
-#endif
-    
-    ACQUIRE_LOCK(&sched_mutex);
-    
-#ifdef RTS_SUPPORTS_THREADS
-    IF_DEBUG(scheduler,fprintf(stderr,"sched (task %p): ", osThreadId()););
-#elif !defined(GRAN) && !defined(PAR)
-    IF_DEBUG(scheduler,fprintf(stderr,"sched: "););
-#endif
-    t = cap->r.rCurrentTSO;
-    
+static void
+schedulePostRunThread(void)
+{
 #if defined(PAR)
     /* HACK 675: if the last thread didn't yield, make sure to print a 
        SCHEDULE event to the log file when StgRunning the next thread, even
@@ -1037,218 +1383,52 @@ run_thread:
     TimeOfLastYield = CURRENT_TIME;
 #endif
 
-    switch (ret) {
+  /* some statistics gathering in the parallel case */
+
+#if defined(GRAN) || defined(PAR) || defined(EDEN)
+  switch (ret) {
     case HeapOverflow:
-#if defined(GRAN)
+# if defined(GRAN)
       IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
       globalGranStats.tot_heapover++;
-#elif defined(PAR)
+# elif defined(PAR)
       globalParStats.tot_heapover++;
-#endif
-
-      // did the task ask for a large block?
-      if (cap->r.rHpAlloc > BLOCK_SIZE_W) {
-         // if so, get one and push it on the front of the nursery.
-         bdescr *bd;
-         nat blocks;
-         
-         blocks = (nat)BLOCK_ROUND_UP(cap->r.rHpAlloc * sizeof(W_)) / BLOCK_SIZE;
-
-         IF_DEBUG(scheduler,belch("--<< thread %ld (%s) stopped: requesting a large block (size %d)", 
-                                  t->id, whatNext_strs[t->what_next], blocks));
-
-         // don't do this if it would push us over the
-         // alloc_blocks_lim limit; we'll GC first.
-         if (alloc_blocks + blocks < alloc_blocks_lim) {
-
-             alloc_blocks += blocks;
-             bd = allocGroup( blocks );
-
-             // link the new group into the list
-             bd->link = cap->r.rCurrentNursery;
-             bd->u.back = cap->r.rCurrentNursery->u.back;
-             if (cap->r.rCurrentNursery->u.back != NULL) {
-                 cap->r.rCurrentNursery->u.back->link = bd;
-             } else {
-                 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
-                        g0s0->blocks == cap->r.rNursery);
-                 cap->r.rNursery = g0s0->blocks = bd;
-             }           
-             cap->r.rCurrentNursery->u.back = bd;
-
-             // initialise it as a nursery block.  We initialise the
-             // step, gen_no, and flags field of *every* sub-block in
-             // this large block, because this is easier than making
-             // sure that we always find the block head of a large
-             // block whenever we call Bdescr() (eg. evacuate() and
-             // isAlive() in the GC would both have to do this, at
-             // least).
-             { 
-                 bdescr *x;
-                 for (x = bd; x < bd + blocks; x++) {
-                     x->step = g0s0;
-                     x->gen_no = 0;
-                     x->flags = 0;
-                 }
-             }
-
-             // don't forget to update the block count in g0s0.
-             g0s0->n_blocks += blocks;
-             // This assert can be a killer if the app is doing lots
-             // of large block allocations.
-             ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
-
-             // now update the nursery to point to the new block
-             cap->r.rCurrentNursery = bd;
-
-             // we might be unlucky and have another thread get on the
-             // run queue before us and steal the large block, but in that
-             // case the thread will just end up requesting another large
-             // block.
-             PUSH_ON_RUN_QUEUE(t);
-             break;
-         }
-      }
-
-      /* make all the running tasks block on a condition variable,
-       * maybe set context_switch and wait till they all pile in,
-       * then have them wait on a GC condition variable.
-       */
-      IF_DEBUG(scheduler,belch("--<< thread %ld (%s) stopped: HeapOverflow", 
-                              t->id, whatNext_strs[t->what_next]));
-      threadPaused(t);
-#if defined(GRAN)
-      ASSERT(!is_on_queue(t,CurrentProc));
-#elif defined(PAR)
-      /* Currently we emit a DESCHEDULE event before GC in GUM.
-         ToDo: either add separate event to distinguish SYSTEM time from rest
-              or just nuke this DESCHEDULE (and the following SCHEDULE) */
-      if (0 && RtsFlags.ParFlags.ParStats.Full) {
-       DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
-                        GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
-       emitSchedule = rtsTrue;
-      }
-#endif
-      
-      ready_to_gc = rtsTrue;
-      context_switch = 1;              /* stop other threads ASAP */
-      PUSH_ON_RUN_QUEUE(t);
-      /* actual GC is done at the end of the while loop */
+# endif
       break;
-      
-    case StackOverflow:
-#if defined(GRAN)
+
+     case StackOverflow:
+# if defined(GRAN)
       IF_DEBUG(gran, 
               DumpGranEvent(GR_DESCHEDULE, t));
       globalGranStats.tot_stackover++;
-#elif defined(PAR)
+# elif defined(PAR)
       // IF_DEBUG(par, 
       // DumpGranEvent(GR_DESCHEDULE, t);
       globalParStats.tot_stackover++;
-#endif
-      IF_DEBUG(scheduler,belch("--<< thread %ld (%s) stopped, StackOverflow", 
-                              t->id, whatNext_strs[t->what_next]));
-      /* just adjust the stack for this thread, then pop it back
-       * on the run queue.
-       */
-      threadPaused(t);
-      { 
-       StgMainThread *m;
-       /* enlarge the stack */
-       StgTSO *new_t = threadStackOverflow(t);
-       
-       /* This TSO has moved, so update any pointers to it from the
-        * main thread stack.  It better not be on any other queues...
-        * (it shouldn't be).
-        */
-       for (m = main_threads; m != NULL; m = m->link) {
-         if (m->tso == t) {
-           m->tso = new_t;
-         }
-       }
-       threadPaused(new_t);
-       PUSH_ON_RUN_QUEUE(new_t);
-      }
+# endif
       break;
 
     case ThreadYielding:
-#if defined(GRAN)
+# if defined(GRAN)
       IF_DEBUG(gran, 
               DumpGranEvent(GR_DESCHEDULE, t));
       globalGranStats.tot_yields++;
-#elif defined(PAR)
+# elif defined(PAR)
       // IF_DEBUG(par, 
       // DumpGranEvent(GR_DESCHEDULE, t);
       globalParStats.tot_yields++;
-#endif
-      /* put the thread back on the run queue.  Then, if we're ready to
-       * GC, check whether this is the last task to stop.  If so, wake
-       * up the GC thread.  getThread will block during a GC until the
-       * GC is finished.
-       */
-      IF_DEBUG(scheduler,
-               if (t->what_next != prev_what_next) {
-                  belch("--<< thread %ld (%s) stopped to switch evaluators", 
-                        t->id, whatNext_strs[t->what_next]);
-               } else {
-                   belch("--<< thread %ld (%s) stopped, yielding", 
-                        t->id, whatNext_strs[t->what_next]);
-               }
-               );
-
-      IF_DEBUG(sanity,
-              //belch("&& Doing sanity check on yielding TSO %ld.", t->id);
-              checkTSO(t));
-      ASSERT(t->link == END_TSO_QUEUE);
-
-      // Shortcut if we're just switching evaluators: don't bother
-      // doing stack squeezing (which can be expensive), just run the
-      // thread.
-      if (t->what_next != prev_what_next) {
-         goto run_thread;
-      }
-
-      threadPaused(t);
-
-#if defined(GRAN)
-      ASSERT(!is_on_queue(t,CurrentProc));
-
-      IF_DEBUG(sanity,
-              //belch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
-              checkThreadQsSanity(rtsTrue));
-#endif
-
-#if defined(PAR)
-      if (RtsFlags.ParFlags.doFairScheduling) { 
-       /* this does round-robin scheduling; good for concurrency */
-       APPEND_TO_RUN_QUEUE(t);
-      } else {
-       /* this does unfair scheduling; good for parallelism */
-       PUSH_ON_RUN_QUEUE(t);
-      }
-#else
-      // this does round-robin scheduling; good for concurrency
-      APPEND_TO_RUN_QUEUE(t);
-#endif
-
-#if defined(GRAN)
-      /* add a ContinueThread event to actually process the thread */
-      new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
-               ContinueThread,
-               t, (StgClosure*)NULL, (rtsSpark*)NULL);
-      IF_GRAN_DEBUG(bq, 
-              belch("GRAN: eventq and runnableq after adding yielded thread to queue again:");
-              G_EVENTQ(0);
-              G_CURR_THREADQ(0));
-#endif /* GRAN */
-      break;
+# endif
+      break; 
 
     case ThreadBlocked:
-#if defined(GRAN)
+# if defined(GRAN)
       IF_DEBUG(scheduler,
-              belch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ", 
-                              t->id, t, whatNext_strs[t->what_next], t->block_info.closure, (t->block_info.closure==(StgClosure*)NULL ? 99 : where_is(t->block_info.closure)));
-              if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
+              debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ", 
+                         t->id, t, whatNext_strs[t->what_next], t->block_info.closure, 
+                         (t->block_info.closure==(StgClosure*)NULL ? 99 : where_is(t->block_info.closure)));
+              if (t->block_info.closure!=(StgClosure*)NULL)
+                print_bq(t->block_info.closure);
+              debugBelch("\n"));
 
       // ??? needed; should emit block before
       IF_DEBUG(gran, 
@@ -1264,75 +1444,422 @@ run_thread:
            procStatus[CurrentProc]==Fetching)) 
        procStatus[CurrentProc] = Idle;
       */
-#elif defined(PAR)
-      IF_DEBUG(scheduler,
-              belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ", 
-                    t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
-      IF_PAR_DEBUG(bq,
+# elif defined(PAR)
+//++PAR++  blockThread() writes the event (change?)
+# endif
+    break;
+
+  case ThreadFinished:
+    break;
 
-                  if (t->block_info.closure!=(StgClosure*)NULL) 
-                    print_bq(t->block_info.closure));
+  default:
+    barf("parGlobalStats: unknown return code");
+    break;
+    }
+#endif
+}
 
-      /* Send a fetch (if BlockedOnGA) and dump event to log file */
-      blockThread(t);
+/* -----------------------------------------------------------------------------
+ * Handle a thread that returned to the scheduler with ThreadHeepOverflow
+ * ASSUMES: sched_mutex
+ * -------------------------------------------------------------------------- */
 
-      /* whatever we schedule next, we must log that schedule */
-      emitSchedule = rtsTrue;
+static rtsBool
+scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
+{
+    // did the task ask for a large block?
+    if (cap->r.rHpAlloc > BLOCK_SIZE) {
+       // if so, get one and push it on the front of the nursery.
+       bdescr *bd;
+       lnat blocks;
+       
+       blocks = (lnat)BLOCK_ROUND_UP(cap->r.rHpAlloc) / BLOCK_SIZE;
+       
+       IF_DEBUG(scheduler,
+                debugBelch("--<< thread %ld (%s) stopped: requesting a large block (size %d)\n", 
+                           (long)t->id, whatNext_strs[t->what_next], blocks));
+       
+       // don't do this if it would push us over the
+       // alloc_blocks_lim limit; we'll GC first.
+       if (alloc_blocks + blocks < alloc_blocks_lim) {
+           
+           alloc_blocks += blocks;
+           bd = allocGroup( blocks );
+           
+           // link the new group into the list
+           bd->link = cap->r.rCurrentNursery;
+           bd->u.back = cap->r.rCurrentNursery->u.back;
+           if (cap->r.rCurrentNursery->u.back != NULL) {
+               cap->r.rCurrentNursery->u.back->link = bd;
+           } else {
+               ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
+                      g0s0->blocks == cap->r.rNursery);
+               cap->r.rNursery = g0s0->blocks = bd;
+           }             
+           cap->r.rCurrentNursery->u.back = bd;
+           
+           // initialise it as a nursery block.  We initialise the
+           // step, gen_no, and flags field of *every* sub-block in
+           // this large block, because this is easier than making
+           // sure that we always find the block head of a large
+           // block whenever we call Bdescr() (eg. evacuate() and
+           // isAlive() in the GC would both have to do this, at
+           // least).
+           { 
+               bdescr *x;
+               for (x = bd; x < bd + blocks; x++) {
+                   x->step = g0s0;
+                   x->gen_no = 0;
+                   x->flags = 0;
+               }
+           }
+           
+           // don't forget to update the block count in g0s0.
+           g0s0->n_blocks += blocks;
+           // This assert can be a killer if the app is doing lots
+           // of large block allocations.
+           ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
+           
+           // now update the nursery to point to the new block
+           cap->r.rCurrentNursery = bd;
+           
+           // we might be unlucky and have another thread get on the
+           // run queue before us and steal the large block, but in that
+           // case the thread will just end up requesting another large
+           // block.
+           PUSH_ON_RUN_QUEUE(t);
+           return rtsFalse;  /* not actually GC'ing */
+       }
+    }
+    
+    /* make all the running tasks block on a condition variable,
+     * maybe set context_switch and wait till they all pile in,
+     * then have them wait on a GC condition variable.
+     */
+    IF_DEBUG(scheduler,
+            debugBelch("--<< thread %ld (%s) stopped: HeapOverflow\n", 
+                       (long)t->id, whatNext_strs[t->what_next]));
+    threadPaused(t);
+#if defined(GRAN)
+    ASSERT(!is_on_queue(t,CurrentProc));
+#elif defined(PARALLEL_HASKELL)
+    /* Currently we emit a DESCHEDULE event before GC in GUM.
+       ToDo: either add separate event to distinguish SYSTEM time from rest
+       or just nuke this DESCHEDULE (and the following SCHEDULE) */
+    if (0 && RtsFlags.ParFlags.ParStats.Full) {
+       DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
+                        GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
+       emitSchedule = rtsTrue;
+    }
+#endif
+      
+    PUSH_ON_RUN_QUEUE(t);
+    return rtsTrue;
+    /* actual GC is done at the end of the while loop in schedule() */
+}
+
+/* -----------------------------------------------------------------------------
+ * Handle a thread that returned to the scheduler with ThreadStackOverflow
+ * ASSUMES: sched_mutex
+ * -------------------------------------------------------------------------- */
+
+static void
+scheduleHandleStackOverflow( StgTSO *t)
+{
+    IF_DEBUG(scheduler,debugBelch("--<< thread %ld (%s) stopped, StackOverflow\n", 
+                                 (long)t->id, whatNext_strs[t->what_next]));
+    /* just adjust the stack for this thread, then pop it back
+     * on the run queue.
+     */
+    threadPaused(t);
+    { 
+       /* enlarge the stack */
+       StgTSO *new_t = threadStackOverflow(t);
+       
+       /* This TSO has moved, so update any pointers to it from the
+        * main thread stack.  It better not be on any other queues...
+        * (it shouldn't be).
+        */
+       if (t->main != NULL) {
+           t->main->tso = new_t;
+       }
+       PUSH_ON_RUN_QUEUE(new_t);
+    }
+}
+
+/* -----------------------------------------------------------------------------
+ * Handle a thread that returned to the scheduler with ThreadYielding
+ * ASSUMES: sched_mutex
+ * -------------------------------------------------------------------------- */
+
+static rtsBool
+scheduleHandleYield( StgTSO *t, nat prev_what_next )
+{
+    // Reset the context switch flag.  We don't do this just before
+    // running the thread, because that would mean we would lose ticks
+    // during GC, which can lead to unfair scheduling (a thread hogs
+    // the CPU because the tick always arrives during GC).  This way
+    // penalises threads that do a lot of allocation, but that seems
+    // better than the alternative.
+    context_switch = 0;
+    
+    /* put the thread back on the run queue.  Then, if we're ready to
+     * GC, check whether this is the last task to stop.  If so, wake
+     * up the GC thread.  getThread will block during a GC until the
+     * GC is finished.
+     */
+    IF_DEBUG(scheduler,
+            if (t->what_next != prev_what_next) {
+                debugBelch("--<< thread %ld (%s) stopped to switch evaluators\n", 
+                           (long)t->id, whatNext_strs[t->what_next]);
+            } else {
+                debugBelch("--<< thread %ld (%s) stopped, yielding\n",
+                           (long)t->id, whatNext_strs[t->what_next]);
+            }
+       );
+    
+    IF_DEBUG(sanity,
+            //debugBelch("&& Doing sanity check on yielding TSO %ld.", t->id);
+            checkTSO(t));
+    ASSERT(t->link == END_TSO_QUEUE);
+    
+    // Shortcut if we're just switching evaluators: don't bother
+    // doing stack squeezing (which can be expensive), just run the
+    // thread.
+    if (t->what_next != prev_what_next) {
+       return rtsTrue;
+    }
+    
+    threadPaused(t);
+    
+#if defined(GRAN)
+    ASSERT(!is_on_queue(t,CurrentProc));
+      
+    IF_DEBUG(sanity,
+            //debugBelch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
+            checkThreadQsSanity(rtsTrue));
+
+#endif
+
+    addToRunQueue(t);
 
+#if defined(GRAN)
+    /* add a ContinueThread event to actually process the thread */
+    new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
+             ContinueThread,
+             t, (StgClosure*)NULL, (rtsSpark*)NULL);
+    IF_GRAN_DEBUG(bq, 
+                 debugBelch("GRAN: eventq and runnableq after adding yielded thread to queue again:\n");
+                 G_EVENTQ(0);
+                 G_CURR_THREADQ(0));
+#endif
+    return rtsFalse;
+}
+
+/* -----------------------------------------------------------------------------
+ * Handle a thread that returned to the scheduler with ThreadBlocked
+ * ASSUMES: sched_mutex
+ * -------------------------------------------------------------------------- */
+
+static void
+scheduleHandleThreadBlocked( StgTSO *t
+#if !defined(GRAN) && !defined(DEBUG)
+    STG_UNUSED
+#endif
+    )
+{
+#if defined(GRAN)
+    IF_DEBUG(scheduler,
+            debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: \n", 
+                       t->id, t, whatNext_strs[t->what_next], t->block_info.closure, (t->block_info.closure==(StgClosure*)NULL ? 99 : where_is(t->block_info.closure)));
+            if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
+    
+    // ??? needed; should emit block before
+    IF_DEBUG(gran, 
+            DumpGranEvent(GR_DESCHEDULE, t)); 
+    prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
+    /*
+      ngoq Dogh!
+      ASSERT(procStatus[CurrentProc]==Busy || 
+      ((procStatus[CurrentProc]==Fetching) && 
+      (t->block_info.closure!=(StgClosure*)NULL)));
+      if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
+      !(!RtsFlags.GranFlags.DoAsyncFetch &&
+      procStatus[CurrentProc]==Fetching)) 
+      procStatus[CurrentProc] = Idle;
+    */
+#elif defined(PAR)
+    IF_DEBUG(scheduler,
+            debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: \n", 
+                       t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
+    IF_PAR_DEBUG(bq,
+                
+                if (t->block_info.closure!=(StgClosure*)NULL) 
+                print_bq(t->block_info.closure));
+    
+    /* Send a fetch (if BlockedOnGA) and dump event to log file */
+    blockThread(t);
+    
+    /* whatever we schedule next, we must log that schedule */
+    emitSchedule = rtsTrue;
+    
 #else /* !GRAN */
       /* don't need to do anything.  Either the thread is blocked on
        * I/O, in which case we'll have called addToBlockedQueue
        * previously, or it's blocked on an MVar or Blackhole, in which
        * case it'll be on the relevant queue already.
        */
-      IF_DEBUG(scheduler,
-              fprintf(stderr, "--<< thread %d (%s) stopped: ", 
-                      t->id, whatNext_strs[t->what_next]);
-              printThreadBlockage(t);
-              fprintf(stderr, "\n"));
-      fflush(stderr);
-
-      /* Only for dumping event to log file 
-        ToDo: do I need this in GranSim, too?
-      blockThread(t);
-      */
+    ASSERT(t->why_blocked != NotBlocked);
+    IF_DEBUG(scheduler,
+            debugBelch("--<< thread %d (%s) stopped: ", 
+                       t->id, whatNext_strs[t->what_next]);
+            printThreadBlockage(t);
+            debugBelch("\n"));
+    
+    /* Only for dumping event to log file 
+       ToDo: do I need this in GranSim, too?
+       blockThread(t);
+    */
 #endif
-      threadPaused(t);
-      break;
+}
+
+/* -----------------------------------------------------------------------------
+ * Handle a thread that returned to the scheduler with ThreadFinished
+ * ASSUMES: sched_mutex
+ * -------------------------------------------------------------------------- */
+
+static rtsBool
+scheduleHandleThreadFinished( StgMainThread *mainThread
+                             USED_WHEN_RTS_SUPPORTS_THREADS,
+                             Capability *cap,
+                             StgTSO *t )
+{
+    /* Need to check whether this was a main thread, and if so,
+     * return with the return value.
+     *
+     * We also end up here if the thread kills itself with an
+     * uncaught exception, see Exception.cmm.
+     */
+    IF_DEBUG(scheduler,debugBelch("--++ thread %d (%s) finished\n", 
+                                 t->id, whatNext_strs[t->what_next]));
 
-    case ThreadFinished:
-      /* Need to check whether this was a main thread, and if so, signal
-       * the task that started it with the return value.  If we have no
-       * more main threads, we probably need to stop all the tasks until
-       * we get a new one.
-       */
-      /* We also end up here if the thread kills itself with an
-       * uncaught exception, see Exception.hc.
-       */
-      IF_DEBUG(scheduler,belch("--++ thread %d (%s) finished", 
-                              t->id, whatNext_strs[t->what_next]));
 #if defined(GRAN)
       endThread(t, CurrentProc); // clean-up the thread
-#elif defined(PAR)
+#elif defined(PARALLEL_HASKELL)
       /* For now all are advisory -- HWL */
       //if(t->priority==AdvisoryPriority) ??
-      advisory_thread_count--;
+      advisory_thread_count--; // JB: Caution with this counter, buggy!
       
-# ifdef DIST
+# if defined(DIST)
       if(t->dist.priority==RevalPriority)
        FinishReval(t);
 # endif
-      
+    
+# if defined(EDENOLD)
+      // the thread could still have an outport... (BUG)
+      if (t->eden.outport != -1) {
+      // delete the outport for the tso which has finished...
+       IF_PAR_DEBUG(eden_ports,
+                  debugBelch("WARNING: Scheduler removes outport %d for TSO %d.\n",
+                             t->eden.outport, t->id));
+       deleteOPT(t);
+      }
+      // thread still in the process (HEAVY BUG! since outport has just been closed...)
+      if (t->eden.epid != -1) {
+       IF_PAR_DEBUG(eden_ports,
+                  debugBelch("WARNING: Scheduler removes TSO %d from process %d .\n",
+                          t->id, t->eden.epid));
+       removeTSOfromProcess(t);
+      }
+# endif 
+
+# if defined(PAR)
       if (RtsFlags.ParFlags.ParStats.Full &&
          !RtsFlags.ParFlags.ParStats.Suppressed) 
        DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
+
+      //  t->par only contains statistics: left out for now...
+      IF_PAR_DEBUG(fish,
+                  debugBelch("**** end thread: ended sparked thread %d (%lx); sparkname: %lx\n",
+                             t->id,t,t->par.sparkname));
+# endif
+#endif // PARALLEL_HASKELL
+
+      //
+      // Check whether the thread that just completed was a main
+      // thread, and if so return with the result.  
+      //
+      // There is an assumption here that all thread completion goes
+      // through this point; we need to make sure that if a thread
+      // ends up in the ThreadKilled state, that it stays on the run
+      // queue so it can be dealt with here.
+      //
+      if (
+#if defined(RTS_SUPPORTS_THREADS)
+         mainThread != NULL
+#else
+         mainThread->tso == t
 #endif
-      break;
-      
-    default:
-      barf("schedule: invalid thread return code %d", (int)ret);
-    }
+         )
+      {
+         // We are a bound thread: this must be our thread that just
+         // completed.
+         ASSERT(mainThread->tso == t);
+
+         if (t->what_next == ThreadComplete) {
+             if (mainThread->ret) {
+                 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
+                 *(mainThread->ret) = (StgClosure *)mainThread->tso->sp[1]; 
+             }
+             mainThread->stat = Success;
+         } else {
+             if (mainThread->ret) {
+                 *(mainThread->ret) = NULL;
+             }
+             if (interrupted) {
+                 mainThread->stat = Interrupted;
+             } else {
+                 mainThread->stat = Killed;
+             }
+         }
+#ifdef DEBUG
+         removeThreadLabel((StgWord)mainThread->tso->id);
+#endif
+         if (mainThread->prev == NULL) {
+             main_threads = mainThread->link;
+         } else {
+             mainThread->prev->link = mainThread->link;
+         }
+         if (mainThread->link != NULL) {
+             mainThread->link->prev = NULL;
+         }
+         releaseCapability(cap);
+         return rtsTrue; // tells schedule() to return
+      }
 
+#ifdef RTS_SUPPORTS_THREADS
+      ASSERT(t->main == NULL);
+#else
+      if (t->main != NULL) {
+         // Must be a main thread that is not the topmost one.  Leave
+         // it on the run queue until the stack has unwound to the
+         // point where we can deal with this.  Leaving it on the run
+         // queue also ensures that the garbage collector knows about
+         // this thread and its return value (it gets dropped from the
+         // all_threads list so there's no other way to find it).
+         APPEND_TO_RUN_QUEUE(t);
+      }
+#endif
+      return rtsFalse;
+}
+
+/* -----------------------------------------------------------------------------
+ * Perform a heap census, if PROFILING
+ * -------------------------------------------------------------------------- */
+
+static void
+scheduleDoHeapProfile(void)
+{
 #ifdef PROFILING
     // When we have +RTS -i0 and we're heap profiling, do a census at
     // every GC.  This lets us get repeatable runs for debugging.
@@ -1345,45 +1872,74 @@ run_thread:
        ready_to_gc = rtsFalse; // we already GC'd
     }
 #endif
+}
 
-    if (ready_to_gc) {
-      /* everybody back, start the GC.
-       * Could do it in this thread, or signal a condition var
-       * to do it in another thread.  Either way, we need to
-       * broadcast on gc_pending_cond afterward.
-       */
+/* -----------------------------------------------------------------------------
+ * Perform a garbage collection if necessary
+ * ASSUMES: sched_mutex
+ * -------------------------------------------------------------------------- */
+
+static void
+scheduleDoGC(void)
+{
+    StgTSO *t;
+
+#ifdef SMP
+    // The last task to stop actually gets to do the GC.  The rest
+    // of the tasks release their capabilities and wait gc_pending_cond.
+    if (ready_to_gc && allFreeCapabilities())
+#else
+    if (ready_to_gc)
+#endif
+    {
+       /* Kick any transactions which are invalid back to their
+        * atomically frames.  When next scheduled they will try to
+        * commit, this commit will fail and they will retry.
+        */
+       for (t = all_threads; t != END_TSO_QUEUE; t = t -> link) {
+           if (t -> what_next != ThreadRelocated && t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
+               if (!stmValidateTransaction (t -> trec)) {
+                   IF_DEBUG(stm, sched_belch("trec %p found wasting its time", t));
+                   
+                   // strip the stack back to the ATOMICALLY_FRAME, aborting
+                   // the (nested) transaction, and saving the stack of any
+                   // partially-evaluated thunks on the heap.
+                   raiseAsync_(t, NULL, rtsTrue);
+                   
+#ifdef REG_R1
+                   ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
+#endif
+               }
+           }
+       }
+
+       // so this happens periodically:
+       scheduleCheckBlackHoles();
+
+       /* everybody back, start the GC.
+        * Could do it in this thread, or signal a condition var
+        * to do it in another thread.  Either way, we need to
+        * broadcast on gc_pending_cond afterward.
+        */
 #if defined(RTS_SUPPORTS_THREADS)
-      IF_DEBUG(scheduler,sched_belch("doing GC"));
+       IF_DEBUG(scheduler,sched_belch("doing GC"));
+#endif
+       GarbageCollect(GetRoots,rtsFalse);
+       ready_to_gc = rtsFalse;
+#if defined(SMP)
+       broadcastCondition(&gc_pending_cond);
 #endif
-      GarbageCollect(GetRoots,rtsFalse);
-      ready_to_gc = rtsFalse;
 #if defined(GRAN)
-      /* add a ContinueThread event to continue execution of current thread */
-      new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
-               ContinueThread,
-               t, (StgClosure*)NULL, (rtsSpark*)NULL);
-      IF_GRAN_DEBUG(bq, 
-              fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n");
-              G_EVENTQ(0);
-              G_CURR_THREADQ(0));
+       /* add a ContinueThread event to continue execution of current thread */
+       new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
+                 ContinueThread,
+                 t, (StgClosure*)NULL, (rtsSpark*)NULL);
+       IF_GRAN_DEBUG(bq, 
+                     debugBelch("GRAN: eventq and runnableq after Garbage collection:\n\n");
+                     G_EVENTQ(0);
+                     G_CURR_THREADQ(0));
 #endif /* GRAN */
     }
-
-#if defined(GRAN)
-  next_thread:
-    IF_GRAN_DEBUG(unused,
-                 print_eventq(EventHd));
-
-    event = get_next_event();
-#elif defined(PAR)
-  next_thread:
-    /* ToDo: wait for next message to arrive rather than busy wait */
-#endif /* GRAN */
-
-  } /* end of while(1) */
-
-  IF_PAR_DEBUG(verbose,
-              belch("== Leaving schedule() after having received Finish"));
 }
 
 /* ---------------------------------------------------------------------------
@@ -1409,12 +1965,7 @@ StgBool
 isThreadBound(StgTSO* tso USED_IN_THREADED_RTS)
 {
 #ifdef THREADED_RTS
-  StgMainThread *m;
-  for(m = main_threads; m; m = m->link)
-  {
-    if(m->tso == tso)
-      return rtsTrue;
-  }
+  return (tso->main != NULL);
 #endif
   return rtsFalse;
 }
@@ -1423,13 +1974,22 @@ isThreadBound(StgTSO* tso USED_IN_THREADED_RTS)
  * Singleton fork(). Do not copy any running threads.
  * ------------------------------------------------------------------------- */
 
+#ifndef mingw32_HOST_OS
+#define FORKPROCESS_PRIMOP_SUPPORTED
+#endif
+
+#ifdef FORKPROCESS_PRIMOP_SUPPORTED
 static void 
 deleteThreadImmediately(StgTSO *tso);
-
+#endif
 StgInt
-forkProcess(HsStablePtr *entry)
+forkProcess(HsStablePtr *entry
+#ifndef FORKPROCESS_PRIMOP_SUPPORTED
+           STG_UNUSED
+#endif
+           )
 {
-#ifndef mingw32_TARGET_OS
+#ifdef FORKPROCESS_PRIMOP_SUPPORTED
   pid_t pid;
   StgTSO* t,*next;
   StgMainThread *m;
@@ -1463,18 +2023,12 @@ forkProcess(HsStablePtr *entry)
       // wipe the main thread list
     while((m = main_threads) != NULL) {
       main_threads = m->link;
-#ifdef THREADED_RTS
+# ifdef THREADED_RTS
       closeCondition(&m->bound_thread_cond);
-#endif
+# endif
       stgFree(m);
     }
     
-#ifdef RTS_SUPPORTS_THREADS
-    resetTaskManagerAfterFork();      // tell startTask() and friends that
-    startingWorkerThread = rtsFalse;  // we have no worker threads any more
-    resetWorkerWakeupPipeAfterFork();
-#endif
-    
     rc = rts_evalStableIO(entry, NULL);  // run the action
     rts_checkSchedStatus("forkProcess",rc);
     
@@ -1483,10 +2037,10 @@ forkProcess(HsStablePtr *entry)
     hs_exit();                      // clean up and exit
     stg_exit(0);
   }
-#else /* mingw32 */
-  barf("forkProcess#: primop not implemented for mingw32, sorry!\n");
+#else /* !FORKPROCESS_PRIMOP_SUPPORTED */
+  barf("forkProcess#: primop not supported, sorry!\n");
   return -1;
-#endif /* mingw32 */
+#endif
 }
 
 /* ---------------------------------------------------------------------------
@@ -1507,9 +2061,16 @@ deleteAllThreads ( void )
       next = t->global_link;
       deleteThread(t);
   }      
-  run_queue_hd = run_queue_tl = END_TSO_QUEUE;
-  blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
-  sleeping_queue = END_TSO_QUEUE;
+
+  // The run queue now contains a bunch of ThreadKilled threads.  We
+  // must not throw these away: the main thread(s) will be in there
+  // somewhere, and the main scheduler loop has to deal with it.
+  // Also, the run queue is the only thing keeping these threads from
+  // being GC'd, and we don't want the "main thread has been GC'd" panic.
+
+  ASSERT(blocked_queue_hd == END_TSO_QUEUE);
+  ASSERT(blackhole_queue == END_TSO_QUEUE);
+  ASSERT(sleeping_queue == END_TSO_QUEUE);
 }
 
 /* startThread and  insertThread are now in GranSim.c -- HWL */
@@ -1531,12 +2092,7 @@ deleteAllThreads ( void )
  * ------------------------------------------------------------------------- */
    
 StgInt
-suspendThread( StgRegTable *reg, 
-              rtsBool concCall
-#if !defined(DEBUG)
-              STG_UNUSED
-#endif
-              )
+suspendThread( StgRegTable *reg )
 {
   nat tok;
   Capability *cap;
@@ -1550,7 +2106,7 @@ suspendThread( StgRegTable *reg,
   ACQUIRE_LOCK(&sched_mutex);
 
   IF_DEBUG(scheduler,
-          sched_belch("thread %d did a _ccall_gc (is_concurrent: %d)", cap->r.rCurrentTSO->id,concCall));
+          sched_belch("thread %d did a _ccall_gc", cap->r.rCurrentTSO->id));
 
   // XXX this might not be necessary --SDM
   cap->r.rCurrentTSO->what_next = ThreadRunGHC;
@@ -1559,17 +2115,12 @@ suspendThread( StgRegTable *reg,
   cap->r.rCurrentTSO->link = suspended_ccalling_threads;
   suspended_ccalling_threads = cap->r.rCurrentTSO;
 
-#if defined(RTS_SUPPORTS_THREADS)
-  if(cap->r.rCurrentTSO->blocked_exceptions == NULL)
-  {
+  if(cap->r.rCurrentTSO->blocked_exceptions == NULL)  {
       cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
       cap->r.rCurrentTSO->blocked_exceptions = END_TSO_QUEUE;
-  }
-  else
-  {
+  } else {
       cap->r.rCurrentTSO->why_blocked = BlockedOnCCall_NoUnblockExc;
   }
-#endif
 
   /* Use the thread ID as the token; it should be unique */
   tok = cap->r.rCurrentTSO->id;
@@ -1584,8 +2135,7 @@ suspendThread( StgRegTable *reg,
   IF_DEBUG(scheduler, sched_belch("worker (token %d): leaving RTS", tok));
 #endif
 
-  /* Other threads _might_ be available for execution; signal this */
-  THREAD_RUNNABLE();
+  in_haskell = rtsFalse;
   RELEASE_LOCK(&sched_mutex);
   
   errno = saved_errno;
@@ -1593,8 +2143,7 @@ suspendThread( StgRegTable *reg,
 }
 
 StgRegTable *
-resumeThread( StgInt tok,
-             rtsBool concCall STG_UNUSED )
+resumeThread( StgInt tok )
 {
   StgTSO *tso, **prev;
   Capability *cap;
@@ -1625,29 +2174,21 @@ resumeThread( StgInt tok,
   }
   tso->link = END_TSO_QUEUE;
   
-#if defined(RTS_SUPPORTS_THREADS)
-  if(tso->why_blocked == BlockedOnCCall)
-  {
+  if(tso->why_blocked == BlockedOnCCall) {
       awakenBlockedQueueNoLock(tso->blocked_exceptions);
       tso->blocked_exceptions = NULL;
   }
-#endif
   
   /* Reset blocking status */
   tso->why_blocked  = NotBlocked;
 
   cap->r.rCurrentTSO = tso;
+  in_haskell = rtsTrue;
   RELEASE_LOCK(&sched_mutex);
   errno = saved_errno;
   return &cap->r;
 }
 
-
-/* ---------------------------------------------------------------------------
- * Static functions
- * ------------------------------------------------------------------------ */
-static void unblockThread(StgTSO *tso);
-
 /* ---------------------------------------------------------------------------
  * Comparing Thread ids.
  *
@@ -1720,11 +2261,11 @@ createThread(nat size)
     nat stack_size;
 
     /* First check whether we should create a thread at all */
-#if defined(PAR)
+#if defined(PARALLEL_HASKELL)
   /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
   if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
     threadsIgnored++;
-    belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
+    debugBelch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)\n",
          RtsFlags.ParFlags.maxThreads, advisory_thread_count);
     return END_TSO_QUEUE;
   }
@@ -1760,12 +2301,15 @@ createThread(nat size)
   tso->blocked_exceptions = NULL;
 
   tso->saved_errno = 0;
+  tso->main = NULL;
   
   tso->stack_size   = stack_size;
   tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize) 
                               - TSO_STRUCT_SIZEW;
   tso->sp           = (P_)&(tso->stack) + stack_size;
 
+  tso->trec = NO_TREC;
+
 #ifdef PROFILING
   tso->prof.CCCS = CCS_MAIN;
 #endif
@@ -1773,9 +2317,10 @@ createThread(nat size)
   /* put a stop frame on the stack */
   tso->sp -= sizeofW(StgStopFrame);
   SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
+  tso->link = END_TSO_QUEUE;
+
   // ToDo: check this
 #if defined(GRAN)
-  tso->link = END_TSO_QUEUE;
   /* uses more flexible routine in GranSim */
   insertThread(tso, CurrentProc);
 #else
@@ -1787,7 +2332,7 @@ createThread(nat size)
 #if defined(GRAN) 
   if (RtsFlags.GranFlags.GranSimStats.Full) 
     DumpGranEvent(GR_START,tso);
-#elif defined(PAR)
+#elif defined(PARALLEL_HASKELL)
   if (RtsFlags.ParFlags.ParStats.Full) 
     DumpGranEvent(GR_STARTQ,tso);
   /* HACk to avoid SCHEDULE 
@@ -1827,7 +2372,7 @@ createThread(nat size)
     tso->gran.clock  = 0;
 
   IF_DEBUG(gran,printTSO(tso));
-#elif defined(PAR)
+#elif defined(PARALLEL_HASKELL)
 # if defined(DEBUG)
   tso->par.magic = TSO_MAGIC; // debugging only
 # endif
@@ -1851,26 +2396,26 @@ createThread(nat size)
   globalGranStats.threads_created_on_PE[CurrentProc]++;
   globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
   globalGranStats.tot_sq_probes++;
-#elif defined(PAR)
+#elif defined(PARALLEL_HASKELL)
   // collect parallel global statistics (currently done together with GC stats)
   if (RtsFlags.ParFlags.ParStats.Global &&
       RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
-    //fprintf(stderr, "Creating thread %d @ %11.2f\n", tso->id, usertime()); 
+    //debugBelch("Creating thread %d @ %11.2f\n", tso->id, usertime()); 
     globalParStats.tot_threads_created++;
   }
 #endif 
 
 #if defined(GRAN)
   IF_GRAN_DEBUG(pri,
-               belch("==__ schedule: Created TSO %d (%p);",
+               sched_belch("==__ schedule: Created TSO %d (%p);",
                      CurrentProc, tso, tso->id));
-#elif defined(PAR)
-    IF_PAR_DEBUG(verbose,
-                belch("==__ schedule: Created TSO %d (%p); %d threads active",
-                      tso->id, tso, advisory_thread_count));
+#elif defined(PARALLEL_HASKELL)
+  IF_PAR_DEBUG(verbose,
+              sched_belch("==__ schedule: Created TSO %d (%p); %d threads active",
+                          (long)tso->id, tso, advisory_thread_count));
 #else
   IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words", 
-                                tso->id, tso->stack_size));
+                                (long)tso->id, (long)tso->stack_size));
 #endif    
   return tso;
 }
@@ -1880,9 +2425,10 @@ createThread(nat size)
    all parallel thread creation calls should fall through the following routine.
 */
 StgTSO *
-createSparkThread(rtsSpark spark) 
+createThreadFromSpark(rtsSpark spark) 
 { StgTSO *tso;
   ASSERT(spark != (rtsSpark)NULL);
+// JB: TAKE CARE OF THIS COUNTER! BUGGY
   if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) 
   { threadsIgnored++;
     barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
@@ -1898,8 +2444,8 @@ createSparkThread(rtsSpark spark)
     tso->priority = AdvisoryPriority;
 #endif
     pushClosure(tso,spark);
-    PUSH_ON_RUN_QUEUE(tso);
-    advisory_thread_count++;    
+    addToRunQueue(tso);
+    advisory_thread_count++;  // JB: TAKE CARE OF THIS COUNTER! BUGGY
   }
   return tso;
 }
@@ -1909,7 +2455,7 @@ createSparkThread(rtsSpark spark)
   Turn a spark into a thread.
   ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
 */
-#if defined(PAR)
+#if 0
 StgTSO *
 activateSpark (rtsSpark spark) 
 {
@@ -1918,9 +2464,9 @@ activateSpark (rtsSpark spark)
   tso = createSparkThread(spark);
   if (RtsFlags.ParFlags.ParStats.Full) {   
     //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
-    IF_PAR_DEBUG(verbose,
-                belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
-                      (StgClosure *)spark, info_type((StgClosure *)spark)));
+      IF_PAR_DEBUG(verbose,
+                  debugBelch("==^^ activateSpark: turning spark of closure %p (%s) into a thread\n",
+                             (StgClosure *)spark, info_type((StgClosure *)spark)));
   }
   // ToDo: fwd info on local/global spark to thread -- HWL
   // tso->gran.exported =  spark->exported;
@@ -1931,11 +2477,6 @@ activateSpark (rtsSpark spark)
 }
 #endif
 
-static SchedulerStatus waitThread_(/*out*/StgMainThread* m,
-                                  Capability *initialCapability
-                                  );
-
-
 /* ---------------------------------------------------------------------------
  * scheduleThread()
  *
@@ -1946,14 +2487,13 @@ static SchedulerStatus waitThread_(/*out*/StgMainThread* m,
  * on this thread's stack before the scheduler is invoked.
  * ------------------------------------------------------------------------ */
 
-static void scheduleThread_ (StgTSO* tso);
-
-void
+static void
 scheduleThread_(StgTSO *tso)
 {
-  // Precondition: sched_mutex must be held.
-  PUSH_ON_RUN_QUEUE(tso);
-  THREAD_RUNNABLE();
+  // The thread goes at the *end* of the run-queue, to avoid possible
+  // starvation of any threads already on the queue.
+  APPEND_TO_RUN_QUEUE(tso);
+  threadRunnable();
 }
 
 void
@@ -1965,9 +2505,11 @@ scheduleThread(StgTSO* tso)
 }
 
 #if defined(RTS_SUPPORTS_THREADS)
-static Condition *bound_cond_cache = NULL;
+static Condition bound_cond_cache;
+static int bound_cond_cache_full = 0;
 #endif
 
+
 SchedulerStatus
 scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret,
                   Capability *initialCapability)
@@ -1977,15 +2519,23 @@ scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret,
 
     m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
     m->tso = tso;
+    tso->main = m;
     m->ret = ret;
     m->stat = NoStatus;
+    m->link = main_threads;
+    m->prev = NULL;
+    if (main_threads != NULL) {
+       main_threads->prev = m;
+    }
+    main_threads = m;
+
 #if defined(RTS_SUPPORTS_THREADS)
     // Allocating a new condition for each thread is expensive, so we
     // cache one.  This is a pretty feeble hack, but it helps speed up
     // consecutive call-ins quite a bit.
-    if (bound_cond_cache != NULL) {
-       m->bound_thread_cond = *bound_cond_cache;
-       bound_cond_cache = NULL;
+    if (bound_cond_cache_full) {
+       m->bound_thread_cond = bound_cond_cache;
+       bound_cond_cache_full = 0;
     } else {
        initCondition(&m->bound_thread_cond);
     }
@@ -2002,11 +2552,8 @@ scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret,
     */
     IF_DEBUG(scheduler, sched_belch("waiting for thread (%d)", tso->id));
     
-    m->link = main_threads;
-    main_threads = m;
-    
-    PUSH_ON_RUN_QUEUE(tso);
-    // NB. Don't call THREAD_RUNNABLE() here, because the thread is
+    APPEND_TO_RUN_QUEUE(tso);
+    // NB. Don't call threadRunnable() here, because the thread is
     // bound and only runnable by *this* OS thread, so waking up other
     // workers will just slow things down.
 
@@ -2034,6 +2581,7 @@ initScheduler(void)
     blocked_queue_hds[i]  = END_TSO_QUEUE;
     blocked_queue_tls[i]  = END_TSO_QUEUE;
     ccalling_threadss[i]  = END_TSO_QUEUE;
+    blackhole_queue[i]    = END_TSO_QUEUE;
     sleeping_queue        = END_TSO_QUEUE;
   }
 #else
@@ -2041,6 +2589,7 @@ initScheduler(void)
   run_queue_tl      = END_TSO_QUEUE;
   blocked_queue_hd  = END_TSO_QUEUE;
   blocked_queue_tl  = END_TSO_QUEUE;
+  blackhole_queue   = END_TSO_QUEUE;
   sleeping_queue    = END_TSO_QUEUE;
 #endif 
 
@@ -2075,12 +2624,11 @@ initScheduler(void)
     startTaskManager(0,taskStart);
 #endif
 
-#if /* defined(SMP) ||*/ defined(PAR)
+#if /* defined(SMP) ||*/ defined(PARALLEL_HASKELL)
   initSparkPools();
 #endif
 
   RELEASE_LOCK(&sched_mutex);
-
 }
 
 void
@@ -2089,6 +2637,7 @@ exitScheduler( void )
 #if defined(RTS_SUPPORTS_THREADS)
   stopTaskManager();
 #endif
+  interrupted = rtsTrue;
   shutting_down_scheduler = rtsTrue;
 }
 
@@ -2101,8 +2650,7 @@ exitScheduler( void )
    ToDo: no support for two-space collection at the moment???
    ------------------------------------------------------------------------- */
 
-static
-SchedulerStatus
+static SchedulerStatus
 waitThread_(StgMainThread* m, Capability *initialCapability)
 {
   SchedulerStatus stat;
@@ -2125,8 +2673,9 @@ waitThread_(StgMainThread* m, Capability *initialCapability)
 
 #if defined(RTS_SUPPORTS_THREADS)
   // Free the condition variable, returning it to the cache if possible.
-  if (bound_cond_cache == NULL) {
-      *bound_cond_cache = m->bound_thread_cond;
+  if (!bound_cond_cache_full) {
+      bound_cond_cache = m->bound_thread_cond;
+      bound_cond_cache_full = 1;
   } else {
       closeCondition(&m->bound_thread_cond);
   }
@@ -2196,11 +2745,15 @@ GetRoots( evac_fn evac )
   }
 #endif 
 
+  if (blackhole_queue != END_TSO_QUEUE) {
+      evac((StgClosure **)&blackhole_queue);
+  }
+
   if (suspended_ccalling_threads != END_TSO_QUEUE) {
       evac((StgClosure **)&suspended_ccalling_threads);
   }
 
-#if defined(PAR) || defined(GRAN)
+#if defined(PARALLEL_HASKELL) || defined(GRAN)
   markSparkQueue(evac);
 #endif
 
@@ -2208,25 +2761,6 @@ GetRoots( evac_fn evac )
   // mark the signal handlers (signals should be already blocked)
   markSignalHandlers(evac);
 #endif
-
-  // main threads which have completed need to be retained until they
-  // are dealt with in the main scheduler loop.  They won't be
-  // retained any other way: the GC will drop them from the
-  // all_threads list, so we have to be careful to treat them as roots
-  // here.
-  { 
-      StgMainThread *m;
-      for (m = main_threads; m != NULL; m = m->link) {
-         switch (m->tso->what_next) {
-         case ThreadComplete:
-         case ThreadKilled:
-             evac((StgClosure **)&m->tso);
-             break;
-         default:
-             break;
-         }
-      }
-  }
 }
 
 /* -----------------------------------------------------------------------------
@@ -2289,7 +2823,8 @@ performGCWithRoots(void (*get_roots)(evac_fn))
 static StgTSO *
 threadStackOverflow(StgTSO *tso)
 {
-  nat new_stack_size, new_tso_size, stack_words;
+  nat new_stack_size, stack_words;
+  lnat new_tso_size;
   StgPtr new_sp;
   StgTSO *dest;
 
@@ -2297,8 +2832,8 @@ threadStackOverflow(StgTSO *tso)
   if (tso->stack_size >= tso->max_stack_size) {
 
     IF_DEBUG(gc,
-            belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld)",
-                  tso->id, tso, tso->stack_size, tso->max_stack_size);
+            debugBelch("@@ threadStackOverflow of TSO %ld (%p): stack too large (now %ld; max is %ld)\n",
+                  (long)tso->id, tso, (long)tso->stack_size, (long)tso->max_stack_size);
             /* If we're debugging, just print out the top of the stack */
             printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size, 
                                              tso->sp+64)));
@@ -2313,12 +2848,12 @@ threadStackOverflow(StgTSO *tso)
    * Finally round up so the TSO ends up as a whole number of blocks.
    */
   new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
-  new_tso_size   = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) + 
+  new_tso_size   = (lnat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) + 
                                       TSO_STRUCT_SIZE)/sizeof(W_);
   new_tso_size = round_to_mblocks(new_tso_size);  /* Be MBLOCK-friendly */
   new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
 
-  IF_DEBUG(scheduler, fprintf(stderr,"== sched: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
+  IF_DEBUG(scheduler, debugBelch("== sched: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
 
   dest = (StgTSO *)allocate(new_tso_size);
   TICK_ALLOC_TSO(new_stack_size,0);
@@ -2344,10 +2879,9 @@ threadStackOverflow(StgTSO *tso)
   tso->link = dest;
   tso->sp = (P_)&(tso->stack[tso->stack_size]);
   tso->why_blocked = NotBlocked;
-  dest->mut_link = NULL;
 
   IF_PAR_DEBUG(verbose,
-              belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld",
+              debugBelch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld\n",
                     tso->id, tso, tso->stack_size);
               /* If we're debugging, just print out the top of the stack */
               printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size, 
@@ -2370,7 +2904,7 @@ STATIC_INLINE void
 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
 {
 }
-#elif defined(PAR)
+#elif defined(PARALLEL_HASKELL)
 STATIC_INLINE void
 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
 {
@@ -2436,14 +2970,14 @@ unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
     }
     /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
     IF_GRAN_DEBUG(bq,
-                 fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
+                 debugBelch(" %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
                          (node_loc==tso_loc ? "Local" : "Global"), 
                          tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
     tso->block_info.closure = NULL;
-    IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)", 
+    IF_DEBUG(scheduler,debugBelch("-- Waking up thread %ld (%p)\n", 
                             tso->id, tso));
 }
-#elif defined(PAR)
+#elif defined(PARALLEL_HASKELL)
 static StgBlockingQueueElement *
 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
 {
@@ -2454,9 +2988,9 @@ unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
       ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
       /* if it's a TSO just push it onto the run_queue */
       next = bqe->link;
-      // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
-      PUSH_ON_RUN_QUEUE((StgTSO *)bqe); 
-      THREAD_RUNNABLE();
+      ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
+      APPEND_TO_RUN_QUEUE((StgTSO *)bqe); 
+      threadRunnable();
       unblockCount(bqe, node);
       /* reset blocking status after dumping event */
       ((StgTSO *)bqe)->why_blocked = NotBlocked;
@@ -2485,11 +3019,11 @@ unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
           (StgClosure *)bqe);
 # endif
     }
-  IF_PAR_DEBUG(bq, fprintf(stderr, ", %p (%s)", bqe, info_type((StgClosure*)bqe)));
+  IF_PAR_DEBUG(bq, debugBelch(", %p (%s)\n", bqe, info_type((StgClosure*)bqe)));
   return next;
 }
 
-#else /* !GRAN && !PAR */
+#else /* !GRAN && !PARALLEL_HASKELL */
 static StgTSO *
 unblockOneLocked(StgTSO *tso)
 {
@@ -2499,14 +3033,15 @@ unblockOneLocked(StgTSO *tso)
   ASSERT(tso->why_blocked != NotBlocked);
   tso->why_blocked = NotBlocked;
   next = tso->link;
-  PUSH_ON_RUN_QUEUE(tso);
-  THREAD_RUNNABLE();
-  IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
+  tso->link = END_TSO_QUEUE;
+  APPEND_TO_RUN_QUEUE(tso);
+  threadRunnable();
+  IF_DEBUG(scheduler,sched_belch("waking up thread %ld", (long)tso->id));
   return next;
 }
 #endif
 
-#if defined(GRAN) || defined(PAR)
+#if defined(GRAN) || defined(PARALLEL_HASKELL)
 INLINE_ME StgBlockingQueueElement *
 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
 {
@@ -2535,7 +3070,7 @@ awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
   nat len = 0; 
 
   IF_GRAN_DEBUG(bq, 
-               belch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
+               debugBelch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): \n", \
                      node, CurrentProc, CurrentTime[CurrentProc], 
                      CurrentTSO->id, CurrentTSO));
 
@@ -2552,13 +3087,13 @@ awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
   */
   if (CurrentProc!=node_loc) {
     IF_GRAN_DEBUG(bq, 
-                 belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)",
+                 debugBelch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)\n",
                        node, node_loc, CurrentProc, CurrentTSO->id, 
                        // CurrentTSO, where_is(CurrentTSO),
                        node->header.gran.procs));
     node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
     IF_GRAN_DEBUG(bq, 
-                 belch("## new bitmask of node %p is %#x",
+                 debugBelch("## new bitmask of node %p is %#x\n",
                        node, node->header.gran.procs));
     if (RtsFlags.GranFlags.GranSimStats.Global) {
       globalGranStats.tot_fake_fetches++;
@@ -2593,7 +3128,7 @@ awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
     ((StgRBH *)node)->mut_link       = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
 
     IF_GRAN_DEBUG(bq,
-                 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
+                 debugBelch("## Filled in RBH_Save for %p (%s) at end of AwBQ\n",
                        node, info_type(node)));
   }
 
@@ -2605,10 +3140,10 @@ awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
     globalGranStats.tot_awbq++;             // total no. of bqs awakened
   }
   IF_GRAN_DEBUG(bq,
-               fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n",
+               debugBelch("## BQ Stats of %p: [%d entries] %s\n",
                        node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
 }
-#elif defined(PAR)
+#elif defined(PARALLEL_HASKELL)
 void 
 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
 {
@@ -2617,12 +3152,12 @@ awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
   ACQUIRE_LOCK(&sched_mutex);
 
   IF_PAR_DEBUG(verbose, 
-              belch("##-_ AwBQ for node %p on [%x]: ",
+              debugBelch("##-_ AwBQ for node %p on [%x]: \n",
                     node, mytid));
 #ifdef DIST  
   //RFP
   if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
-    IF_PAR_DEBUG(verbose, belch("## ... nothing to unblock so lets just return. RFP (BUG?)"));
+    IF_PAR_DEBUG(verbose, debugBelch("## ... nothing to unblock so lets just return. RFP (BUG?)\n"));
     return;
   }
 #endif
@@ -2640,9 +3175,8 @@ awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
   RELEASE_LOCK(&sched_mutex);
 }
 
-#else   /* !GRAN && !PAR */
+#else   /* !GRAN && !PARALLEL_HASKELL */
 
-#ifdef RTS_SUPPORTS_THREADS
 void
 awakenBlockedQueueNoLock(StgTSO *tso)
 {
@@ -2650,7 +3184,6 @@ awakenBlockedQueueNoLock(StgTSO *tso)
     tso = unblockOneLocked(tso);
   }
 }
-#endif
 
 void
 awakenBlockedQueue(StgTSO *tso)
@@ -2673,9 +3206,6 @@ interruptStgRts(void)
 {
     interrupted    = 1;
     context_switch = 1;
-#ifdef RTS_SUPPORTS_THREADS
-    wakeBlockedWorkerThread();
-#endif
 }
 
 /* -----------------------------------------------------------------------------
@@ -2686,7 +3216,7 @@ interruptStgRts(void)
    This has nothing to do with the UnblockThread event in GranSim. -- HWL
    -------------------------------------------------------------------------- */
 
-#if defined(GRAN) || defined(PAR)
+#if defined(GRAN) || defined(PARALLEL_HASKELL)
 /*
   NB: only the type of the blocking queue is different in GranSim and GUM
       the operations on the queue-elements are the same
@@ -2705,6 +3235,14 @@ unblockThread(StgTSO *tso)
   case NotBlocked:
     return;  /* not blocked */
 
+  case BlockedOnSTM:
+    // Be careful: nothing to do here!  We tell the scheduler that the thread
+    // is runnable and we leave it to the stack-walking code to abort the 
+    // transaction while unwinding the stack.  We should perhaps have a debugging
+    // test to make sure that this really happens and that the 'zombie' transaction
+    // does not get committed.
+    goto done;
+
   case BlockedOnMVar:
     ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
     {
@@ -2771,7 +3309,7 @@ unblockThread(StgTSO *tso)
 
   case BlockedOnRead:
   case BlockedOnWrite:
-#if defined(mingw32_TARGET_OS)
+#if defined(mingw32_HOST_OS)
   case BlockedOnDoProc:
 #endif
     {
@@ -2838,6 +3376,14 @@ unblockThread(StgTSO *tso)
 
   switch (tso->why_blocked) {
 
+  case BlockedOnSTM:
+    // Be careful: nothing to do here!  We tell the scheduler that the thread
+    // is runnable and we leave it to the stack-walking code to abort the 
+    // transaction while unwinding the stack.  We should perhaps have a debugging
+    // test to make sure that this really happens and that the 'zombie' transaction
+    // does not get committed.
+    goto done;
+
   case BlockedOnMVar:
     ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
     {
@@ -2859,12 +3405,9 @@ unblockThread(StgTSO *tso)
     }
 
   case BlockedOnBlackHole:
-    ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
     {
-      StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
-
-      last = &bq->blocking_queue;
-      for (t = bq->blocking_queue; t != END_TSO_QUEUE; 
+      last = &blackhole_queue;
+      for (t = blackhole_queue; t != END_TSO_QUEUE; 
           last = &t->link, t = t->link) {
        if (t == tso) {
          *last = tso->link;
@@ -2901,7 +3444,7 @@ unblockThread(StgTSO *tso)
 
   case BlockedOnRead:
   case BlockedOnWrite:
-#if defined(mingw32_TARGET_OS)
+#if defined(mingw32_HOST_OS)
   case BlockedOnDoProc:
 #endif
     {
@@ -2951,11 +3494,54 @@ unblockThread(StgTSO *tso)
   tso->link = END_TSO_QUEUE;
   tso->why_blocked = NotBlocked;
   tso->block_info.closure = NULL;
-  PUSH_ON_RUN_QUEUE(tso);
+  APPEND_TO_RUN_QUEUE(tso);
 }
 #endif
 
 /* -----------------------------------------------------------------------------
+ * checkBlackHoles()
+ *
+ * Check the blackhole_queue for threads that can be woken up.  We do
+ * this periodically: before every GC, and whenever the run queue is
+ * empty.
+ *
+ * An elegant solution might be to just wake up all the blocked
+ * threads with awakenBlockedQueue occasionally: they'll go back to
+ * sleep again if the object is still a BLACKHOLE.  Unfortunately this
+ * doesn't give us a way to tell whether we've actually managed to
+ * wake up any threads, so we would be busy-waiting.
+ *
+ * -------------------------------------------------------------------------- */
+
+static rtsBool
+checkBlackHoles( void )
+{
+    StgTSO **prev, *t;
+    rtsBool any_woke_up = rtsFalse;
+    StgHalfWord type;
+
+    IF_DEBUG(scheduler, sched_belch("checking threads blocked on black holes"));
+
+    // ASSUMES: sched_mutex
+    prev = &blackhole_queue;
+    t = blackhole_queue;
+    while (t != END_TSO_QUEUE) {
+       ASSERT(t->why_blocked == BlockedOnBlackHole);
+       type = get_itbl(t->block_info.closure)->type;
+       if (type != BLACKHOLE && type != CAF_BLACKHOLE) {
+           t = unblockOneLocked(t);
+           *prev = t;
+           any_woke_up = rtsTrue;
+       } else {
+           prev = &t->link;
+           t = t->link;
+       }
+    }
+
+    return any_woke_up;
+}
+
+/* -----------------------------------------------------------------------------
  * raiseAsync()
  *
  * The following function implements the magic for raising an
@@ -2992,9 +3578,13 @@ unblockThread(StgTSO *tso)
 void 
 deleteThread(StgTSO *tso)
 {
-  raiseAsync(tso,NULL);
+  if (tso->why_blocked != BlockedOnCCall &&
+      tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
+      raiseAsync(tso,NULL);
+  }
 }
 
+#ifdef FORKPROCESS_PRIMOP_SUPPORTED
 static void 
 deleteThreadImmediately(StgTSO *tso)
 { // for forkProcess only:
@@ -3003,13 +3593,15 @@ deleteThreadImmediately(StgTSO *tso)
   if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
       return;
   }
-#if defined(RTS_SUPPORTS_THREADS)
-  if (tso->why_blocked != BlockedOnCCall
-      && tso->why_blocked != BlockedOnCCall_NoUnblockExc)
-#endif
+
+  if (tso->why_blocked != BlockedOnCCall &&
+      tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
     unblockThread(tso);
+  }
+
   tso->what_next = ThreadKilled;
 }
+#endif
 
 void
 raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
@@ -3024,6 +3616,12 @@ raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
 void
 raiseAsync(StgTSO *tso, StgClosure *exception)
 {
+    raiseAsync_(tso, exception, rtsFalse);
+}
+
+static void
+raiseAsync_(StgTSO *tso, StgClosure *exception, rtsBool stop_at_atomically)
+{
     StgRetInfoTable *info;
     StgPtr sp;
   
@@ -3033,7 +3631,7 @@ raiseAsync(StgTSO *tso, StgClosure *exception)
     }
 
     IF_DEBUG(scheduler, 
-            sched_belch("raising exception in thread %ld.", tso->id));
+            sched_belch("raising exception in thread %ld.", (long)tso->id));
     
     // Remove it from any blocking queues
     unblockThread(tso);
@@ -3067,6 +3665,10 @@ raiseAsync(StgTSO *tso, StgClosure *exception)
        // top of the stack applied to the exception.
        // 
        // 5. If it's a STOP_FRAME, then kill the thread.
+        // 
+        // NB: if we pass an ATOMICALLY_FRAME then abort the associated 
+        // transaction
+       
        
        StgPtr frame;
        
@@ -3075,13 +3677,45 @@ raiseAsync(StgTSO *tso, StgClosure *exception)
        
        while (info->i.type != UPDATE_FRAME
               && (info->i.type != CATCH_FRAME || exception == NULL)
-              && info->i.type != STOP_FRAME) {
+              && info->i.type != STOP_FRAME
+              && (info->i.type != ATOMICALLY_FRAME || stop_at_atomically == rtsFalse))
+       {
+            if (info->i.type == CATCH_RETRY_FRAME || info->i.type == ATOMICALLY_FRAME) {
+              // IF we find an ATOMICALLY_FRAME then we abort the
+              // current transaction and propagate the exception.  In
+              // this case (unlike ordinary exceptions) we do not care
+              // whether the transaction is valid or not because its
+              // possible validity cannot have caused the exception
+              // and will not be visible after the abort.
+              IF_DEBUG(stm,
+                       debugBelch("Found atomically block delivering async exception\n"));
+              stmAbortTransaction(tso -> trec);
+              tso -> trec = stmGetEnclosingTRec(tso -> trec);
+            }
            frame += stack_frame_sizeW((StgClosure *)frame);
            info = get_ret_itbl((StgClosure *)frame);
        }
        
        switch (info->i.type) {
            
+       case ATOMICALLY_FRAME:
+           ASSERT(stop_at_atomically);
+           ASSERT(stmGetEnclosingTRec(tso->trec) == NO_TREC);
+           stmCondemnTransaction(tso -> trec);
+#ifdef REG_R1
+           tso->sp = frame;
+#else
+           // R1 is not a register: the return convention for IO in
+           // this case puts the return value on the stack, so we
+           // need to set up the stack to return to the atomically
+           // frame properly...
+           tso->sp = frame - 2;
+           tso->sp[1] = (StgWord) &stg_NO_FINALIZER_closure; // why not?
+           tso->sp[0] = (StgWord) &stg_ut_1_0_unreg_info;
+#endif
+           tso->what_next = ThreadRunGHC;
+           return;
+
        case CATCH_FRAME:
            // If we find a CATCH_FRAME, and we've got an exception to raise,
            // then build the THUNK raise(exception), and leave it on
@@ -3148,9 +3782,9 @@ raiseAsync(StgTSO *tso, StgClosure *exception)
            TICK_ALLOC_UP_THK(words+1,0);
            
            IF_DEBUG(scheduler,
-                    fprintf(stderr,  "sched: Updating ");
+                    debugBelch("sched: Updating ");
                     printPtr((P_)((StgUpdateFrame *)frame)->updatee); 
-                    fprintf(stderr,  " with ");
+                    debugBelch(" with ");
                     printObj((StgClosure *)ap);
                );
 
@@ -3168,7 +3802,8 @@ raiseAsync(StgTSO *tso, StgClosure *exception)
            //
            if (!closure_IND(((StgUpdateFrame *)frame)->updatee)) {
                // revert the black hole
-               UPD_IND_NOLOCK(((StgUpdateFrame *)frame)->updatee,ap);
+               UPD_IND_NOLOCK(((StgUpdateFrame *)frame)->updatee,
+                              (StgClosure *)ap);
            }
            sp += sizeofW(StgUpdateFrame) - 1;
            sp[0] = (W_)ap; // push onto stack
@@ -3190,6 +3825,137 @@ raiseAsync(StgTSO *tso, StgClosure *exception)
 }
 
 /* -----------------------------------------------------------------------------
+   raiseExceptionHelper
+   
+   This function is called by the raise# primitve, just so that we can
+   move some of the tricky bits of raising an exception from C-- into
+   C.  Who knows, it might be a useful re-useable thing here too.
+   -------------------------------------------------------------------------- */
+
+StgWord
+raiseExceptionHelper (StgTSO *tso, StgClosure *exception)
+{
+    StgClosure *raise_closure = NULL;
+    StgPtr p, next;
+    StgRetInfoTable *info;
+    //
+    // This closure represents the expression 'raise# E' where E
+    // is the exception raise.  It is used to overwrite all the
+    // thunks which are currently under evaluataion.
+    //
+
+    //    
+    // LDV profiling: stg_raise_info has THUNK as its closure
+    // type. Since a THUNK takes at least MIN_UPD_SIZE words in its
+    // payload, MIN_UPD_SIZE is more approprate than 1.  It seems that
+    // 1 does not cause any problem unless profiling is performed.
+    // However, when LDV profiling goes on, we need to linearly scan
+    // small object pool, where raise_closure is stored, so we should
+    // use MIN_UPD_SIZE.
+    //
+    // raise_closure = (StgClosure *)RET_STGCALL1(P_,allocate,
+    //                                        sizeofW(StgClosure)+1);
+    //
+
+    //
+    // Walk up the stack, looking for the catch frame.  On the way,
+    // we update any closures pointed to from update frames with the
+    // raise closure that we just built.
+    //
+    p = tso->sp;
+    while(1) {
+       info = get_ret_itbl((StgClosure *)p);
+       next = p + stack_frame_sizeW((StgClosure *)p);
+       switch (info->i.type) {
+           
+       case UPDATE_FRAME:
+           // Only create raise_closure if we need to.
+           if (raise_closure == NULL) {
+               raise_closure = 
+                   (StgClosure *)allocate(sizeofW(StgClosure)+MIN_UPD_SIZE);
+               SET_HDR(raise_closure, &stg_raise_info, CCCS);
+               raise_closure->payload[0] = exception;
+           }
+           UPD_IND(((StgUpdateFrame *)p)->updatee,raise_closure);
+           p = next;
+           continue;
+
+        case ATOMICALLY_FRAME:
+            IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p\n", p));
+            tso->sp = p;
+            return ATOMICALLY_FRAME;
+           
+       case CATCH_FRAME:
+           tso->sp = p;
+           return CATCH_FRAME;
+
+        case CATCH_STM_FRAME:
+            IF_DEBUG(stm, debugBelch("Found CATCH_STM_FRAME at %p\n", p));
+            tso->sp = p;
+            return CATCH_STM_FRAME;
+           
+       case STOP_FRAME:
+           tso->sp = p;
+           return STOP_FRAME;
+
+        case CATCH_RETRY_FRAME:
+       default:
+           p = next; 
+           continue;
+       }
+    }
+}
+
+
+/* -----------------------------------------------------------------------------
+   findRetryFrameHelper
+
+   This function is called by the retry# primitive.  It traverses the stack
+   leaving tso->sp referring to the frame which should handle the retry.  
+
+   This should either be a CATCH_RETRY_FRAME (if the retry# is within an orElse#) 
+   or should be a ATOMICALLY_FRAME (if the retry# reaches the top level).  
+
+   We skip CATCH_STM_FRAMEs because retries are not considered to be exceptions,
+   despite the similar implementation.
+
+   We should not expect to see CATCH_FRAME or STOP_FRAME because those should
+   not be created within memory transactions.
+   -------------------------------------------------------------------------- */
+
+StgWord
+findRetryFrameHelper (StgTSO *tso)
+{
+  StgPtr           p, next;
+  StgRetInfoTable *info;
+
+  p = tso -> sp;
+  while (1) {
+    info = get_ret_itbl((StgClosure *)p);
+    next = p + stack_frame_sizeW((StgClosure *)p);
+    switch (info->i.type) {
+      
+    case ATOMICALLY_FRAME:
+      IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p during retrry\n", p));
+      tso->sp = p;
+      return ATOMICALLY_FRAME;
+      
+    case CATCH_RETRY_FRAME:
+      IF_DEBUG(stm, debugBelch("Found CATCH_RETRY_FRAME at %p during retrry\n", p));
+      tso->sp = p;
+      return CATCH_RETRY_FRAME;
+      
+    case CATCH_STM_FRAME:
+    default:
+      ASSERT(info->i.type != CATCH_FRAME);
+      ASSERT(info->i.type != STOP_FRAME);
+      p = next; 
+      continue;
+    }
+  }
+}
+
+/* -----------------------------------------------------------------------------
    resurrectThreads is called after garbage collection on the list of
    threads found to be garbage.  Each of these threads will be woken
    up and sent a signal: BlockedOnDeadMVar if the thread was blocked
@@ -3219,6 +3985,9 @@ resurrectThreads( StgTSO *threads )
     case BlockedOnBlackHole:
       raiseAsync(tso,(StgClosure *)NonTermination_closure);
       break;
+    case BlockedOnSTM:
+      raiseAsync(tso,(StgClosure *)BlockedIndefinitely_closure);
+      break;
     case NotBlocked:
       /* This might happen if the thread was blocked on a black hole
        * belonging to a thread that we've just woken up (raiseAsync
@@ -3231,140 +4000,77 @@ resurrectThreads( StgTSO *threads )
   }
 }
 
-/* -----------------------------------------------------------------------------
- * Blackhole detection: if we reach a deadlock, test whether any
- * threads are blocked on themselves.  Any threads which are found to
- * be self-blocked get sent a NonTermination exception.
- *
- * This is only done in a deadlock situation in order to avoid
- * performance overhead in the normal case.
- *
- * Locks: sched_mutex is held upon entry and exit.
- * -------------------------------------------------------------------------- */
-
-static void
-detectBlackHoles( void )
-{
-    StgTSO *tso = all_threads;
-    StgClosure *frame;
-    StgClosure *blocked_on;
-    StgRetInfoTable *info;
-
-    for (tso = all_threads; tso != END_TSO_QUEUE; tso = tso->global_link) {
-
-       while (tso->what_next == ThreadRelocated) {
-           tso = tso->link;
-           ASSERT(get_itbl(tso)->type == TSO);
-       }
-      
-       if (tso->why_blocked != BlockedOnBlackHole) {
-           continue;
-       }
-       blocked_on = tso->block_info.closure;
-
-       frame = (StgClosure *)tso->sp;
-
-       while(1) {
-           info = get_ret_itbl(frame);
-           switch (info->i.type) {
-           case UPDATE_FRAME:
-               if (((StgUpdateFrame *)frame)->updatee == blocked_on) {
-                   /* We are blocking on one of our own computations, so
-                    * send this thread the NonTermination exception.  
-                    */
-                   IF_DEBUG(scheduler, 
-                            sched_belch("thread %d is blocked on itself", tso->id));
-                   raiseAsync(tso, (StgClosure *)NonTermination_closure);
-                   goto done;
-               }
-               
-               frame = (StgClosure *) ((StgUpdateFrame *)frame + 1);
-               continue;
-
-           case STOP_FRAME:
-               goto done;
-
-               // normal stack frames; do nothing except advance the pointer
-           default:
-               (StgPtr)frame += stack_frame_sizeW(frame);
-           }
-       }   
-       done: ;
-    }
-}
-
 /* ----------------------------------------------------------------------------
  * Debugging: why is a thread blocked
  * [Also provides useful information when debugging threaded programs
  *  at the Haskell source code level, so enable outside of DEBUG. --sof 7/02]
    ------------------------------------------------------------------------- */
 
-static
-void
+static void
 printThreadBlockage(StgTSO *tso)
 {
   switch (tso->why_blocked) {
   case BlockedOnRead:
-    fprintf(stderr,"is blocked on read from fd %d", tso->block_info.fd);
+    debugBelch("is blocked on read from fd %ld", tso->block_info.fd);
     break;
   case BlockedOnWrite:
-    fprintf(stderr,"is blocked on write to fd %d", tso->block_info.fd);
+    debugBelch("is blocked on write to fd %ld", tso->block_info.fd);
     break;
-#if defined(mingw32_TARGET_OS)
+#if defined(mingw32_HOST_OS)
     case BlockedOnDoProc:
-    fprintf(stderr,"is blocked on proc (request: %d)", tso->block_info.async_result->reqID);
+    debugBelch("is blocked on proc (request: %ld)", tso->block_info.async_result->reqID);
     break;
 #endif
   case BlockedOnDelay:
-    fprintf(stderr,"is blocked until %d", tso->block_info.target);
+    debugBelch("is blocked until %ld", tso->block_info.target);
     break;
   case BlockedOnMVar:
-    fprintf(stderr,"is blocked on an MVar");
+    debugBelch("is blocked on an MVar");
     break;
   case BlockedOnException:
-    fprintf(stderr,"is blocked on delivering an exception to thread %d",
+    debugBelch("is blocked on delivering an exception to thread %d",
            tso->block_info.tso->id);
     break;
   case BlockedOnBlackHole:
-    fprintf(stderr,"is blocked on a black hole");
+    debugBelch("is blocked on a black hole");
     break;
   case NotBlocked:
-    fprintf(stderr,"is not blocked");
+    debugBelch("is not blocked");
     break;
-#if defined(PAR)
+#if defined(PARALLEL_HASKELL)
   case BlockedOnGA:
-    fprintf(stderr,"is blocked on global address; local FM_BQ is %p (%s)",
+    debugBelch("is blocked on global address; local FM_BQ is %p (%s)",
            tso->block_info.closure, info_type(tso->block_info.closure));
     break;
   case BlockedOnGA_NoSend:
-    fprintf(stderr,"is blocked on global address (no send); local FM_BQ is %p (%s)",
+    debugBelch("is blocked on global address (no send); local FM_BQ is %p (%s)",
            tso->block_info.closure, info_type(tso->block_info.closure));
     break;
 #endif
-#if defined(RTS_SUPPORTS_THREADS)
   case BlockedOnCCall:
-    fprintf(stderr,"is blocked on an external call");
+    debugBelch("is blocked on an external call");
     break;
   case BlockedOnCCall_NoUnblockExc:
-    fprintf(stderr,"is blocked on an external call (exceptions were already blocked)");
+    debugBelch("is blocked on an external call (exceptions were already blocked)");
+    break;
+  case BlockedOnSTM:
+    debugBelch("is blocked on an STM operation");
     break;
-#endif
   default:
     barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
         tso->why_blocked, tso->id, tso);
   }
 }
 
-static
-void
+static void
 printThreadStatus(StgTSO *tso)
 {
   switch (tso->what_next) {
   case ThreadKilled:
-    fprintf(stderr,"has been killed");
+    debugBelch("has been killed");
     break;
   case ThreadComplete:
-    fprintf(stderr,"has completed");
+    debugBelch("has completed");
     break;
   default:
     printThreadBlockage(tso);
@@ -3375,30 +4081,33 @@ void
 printAllThreads(void)
 {
   StgTSO *t;
-  void *label;
 
 # if defined(GRAN)
   char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
   ullong_format_string(TIME_ON_PROC(CurrentProc), 
                       time_string, rtsFalse/*no commas!*/);
 
-  fprintf(stderr, "all threads at [%s]:\n", time_string);
-# elif defined(PAR)
+  debugBelch("all threads at [%s]:\n", time_string);
+# elif defined(PARALLEL_HASKELL)
   char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
   ullong_format_string(CURRENT_TIME,
                       time_string, rtsFalse/*no commas!*/);
 
-  fprintf(stderr,"all threads at [%s]:\n", time_string);
+  debugBelch("all threads at [%s]:\n", time_string);
 # else
-  fprintf(stderr,"all threads:\n");
+  debugBelch("all threads:\n");
 # endif
 
   for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
-    fprintf(stderr, "\tthread %d @ %p ", t->id, (void *)t);
-    label = lookupThreadLabel(t->id);
-    if (label) fprintf(stderr,"[\"%s\"] ",(char *)label);
+    debugBelch("\tthread %d @ %p ", t->id, (void *)t);
+#if defined(DEBUG)
+    {
+      void *label = lookupThreadLabel(t->id);
+      if (label) debugBelch("[\"%s\"] ",(char *)label);
+    }
+#endif
     printThreadStatus(t);
-    fprintf(stderr,"\n");
+    debugBelch("\n");
   }
 }
     
@@ -3407,7 +4116,7 @@ printAllThreads(void)
 /* 
    Print a whole blocking queue attached to node (debugging only).
 */
-# if defined(PAR)
+# if defined(PARALLEL_HASKELL)
 void 
 print_bq (StgClosure *node)
 {
@@ -3415,7 +4124,7 @@ print_bq (StgClosure *node)
   StgTSO *tso;
   rtsBool end;
 
-  fprintf(stderr,"## BQ of closure %p (%s): ",
+  debugBelch("## BQ of closure %p (%s): ",
          node, info_type(node));
 
   /* should cover all closures that may have a blocking queue */
@@ -3455,18 +4164,18 @@ print_bqe (StgBlockingQueueElement *bqe)
 
     switch (get_itbl(bqe)->type) {
     case TSO:
-      fprintf(stderr," TSO %u (%x),",
+      debugBelch(" TSO %u (%x),",
              ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
       break;
     case BLOCKED_FETCH:
-      fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
+      debugBelch(" BF (node=%p, ga=((%x, %d, %x)),",
              ((StgBlockedFetch *)bqe)->node, 
              ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
              ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
              ((StgBlockedFetch *)bqe)->ga.weight);
       break;
     case CONSTR:
-      fprintf(stderr," %s (IP %p),",
+      debugBelch(" %s (IP %p),",
              (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
               get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
               get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
@@ -3478,7 +4187,7 @@ print_bqe (StgBlockingQueueElement *bqe)
       break;
     }
   } /* for */
-  fputc('\n', stderr);
+  debugBelch("\n");
 }
 # elif defined(GRAN)
 void 
@@ -3496,7 +4205,7 @@ print_bq (StgClosure *node)
   ASSERT(node!=(StgClosure*)NULL);         // sanity check
   node_loc = where_is(node);
 
-  fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ",
+  debugBelch("## BQ of closure %p (%s) on [PE %d]: ",
          node, info_type(node), node_loc);
 
   /* 
@@ -3516,11 +4225,11 @@ print_bq (StgClosure *node)
     tso_loc = where_is((StgClosure *)bqe);
     switch (get_itbl(bqe)->type) {
     case TSO:
-      fprintf(stderr," TSO %d (%p) on [PE %d],",
+      debugBelch(" TSO %d (%p) on [PE %d],",
              ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
       break;
     case CONSTR:
-      fprintf(stderr," %s (IP %p),",
+      debugBelch(" %s (IP %p),",
              (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
               get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
               get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
@@ -3532,30 +4241,11 @@ print_bq (StgClosure *node)
       break;
     }
   } /* for */
-  fputc('\n', stderr);
-}
-#else
-/* 
-   Nice and easy: only TSOs on the blocking queue
-*/
-void 
-print_bq (StgClosure *node)
-{
-  StgTSO *tso;
-
-  ASSERT(node!=(StgClosure*)NULL);         // sanity check
-  for (tso = ((StgBlockingQueue*)node)->blocking_queue;
-       tso != END_TSO_QUEUE; 
-       tso=tso->link) {
-    ASSERT(tso!=NULL && tso!=END_TSO_QUEUE);   // sanity check
-    ASSERT(get_itbl(tso)->type == TSO);  // guess what, sanity check
-    fprintf(stderr," TSO %d (%p),", tso->id, tso);
-  }
-  fputc('\n', stderr);
+  debugBelch("\n");
 }
 # endif
 
-#if defined(PAR)
+#if defined(PARALLEL_HASKELL)
 static nat
 run_queue_len(void)
 {
@@ -3577,15 +4267,14 @@ sched_belch(char *s, ...)
   va_list ap;
   va_start(ap,s);
 #ifdef RTS_SUPPORTS_THREADS
-  fprintf(stderr, "sched (task %p): ", osThreadId());
-#elif defined(PAR)
-  fprintf(stderr, "== ");
+  debugBelch("sched (task %p): ", osThreadId());
+#elif defined(PARALLEL_HASKELL)
+  debugBelch("== ");
 #else
-  fprintf(stderr, "sched: ");
+  debugBelch("sched: ");
 #endif
-  vfprintf(stderr, s, ap);
-  fprintf(stderr, "\n");
-  fflush(stderr);
+  vdebugBelch(s, ap);
+  debugBelch("\n");
   va_end(ap);
 }