[project @ 2005-02-03 11:21:03 by simonmar]
[ghc-hetmet.git] / ghc / rts / Schedule.c
index 869b2ab..1aac258 100644 (file)
@@ -1,79 +1,70 @@
 /* ---------------------------------------------------------------------------
- * $Id: Schedule.c,v 1.57 2000/03/20 09:42:50 andy Exp $
  *
- * (c) The GHC Team, 1998-2000
+ * (c) The GHC Team, 1998-2004
  *
  * Scheduler
  *
- * The main scheduling code in GranSim is quite different from that in std
- * (concurrent) Haskell: while concurrent Haskell just iterates over the
- * threads in the runnable queue, GranSim is event driven, i.e. it iterates
- * over the events in the global event queue.  -- HWL
+ * Different GHC ways use this scheduler quite differently (see comments below)
+ * Here is the global picture:
+ *
+ * WAY  Name     CPP flag  What's it for
+ * --------------------------------------
+ * mp   GUM      PAR          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)
+ *
  * --------------------------------------------------------------------------*/
 
-//@node Main scheduling code, , ,
-//@section Main scheduling code
-
-/* Version with scheduler monitor support for SMPs.
-
-   This design provides a high-level API to create and schedule threads etc.
-   as documented in the SMP design document.
-
-   It uses a monitor design controlled by a single mutex to exercise control
-   over accesses to shared data structures, and builds on the Posix threads
-   library.
-
-   The majority of state is shared.  In order to keep essential per-task state,
-   there is a Capability structure, which contains all the information
-   needed to run a thread: its STG registers, a pointer to its TSO, a
-   nursery etc.  During STG execution, a pointer to the capability is
-   kept in a register (BaseReg).
-
-   In a non-SMP build, there is one global capability, namely MainRegTable.
-
-   SDM & KH, 10/99
+/* 
+ * Version with support for distributed memory parallelism aka GUM (WAY=mp):
+
+   The main scheduling loop in GUM iterates until a finish message is received.
+   In that case a global flag @receivedFinish@ is set and this instance of
+   the RTS shuts down. See ghc/rts/parallel/HLComms.c:processMessages()
+   for the handling of incoming messages, such as PP_FINISH.
+   Note that in the parallel case we have a system manager that coordinates
+   different PEs, each of which are running one instance of the RTS.
+   See ghc/rts/parallel/SysMan.c for the main routine of the parallel program.
+   From this routine processes executing ghc/rts/Main.c are spawned. -- HWL
+
+ * Version with support for simulating parallel execution aka GranSim (WAY=mg):
+
+   The main scheduling code in GranSim is quite different from that in std
+   (concurrent) Haskell: while concurrent Haskell just iterates over the
+   threads in the runnable queue, GranSim is event driven, i.e. it iterates
+   over the events in the global event queue.  -- HWL
 */
 
-//@menu
-//* Includes::                 
-//* Variables and Data structures::  
-//* Prototypes::               
-//* Main scheduling loop::     
-//* Suspend and Resume::       
-//* Run queue code::           
-//* Garbage Collextion Routines::  
-//* Blocking Queue Routines::  
-//* Exception Handling Routines::  
-//* Debugging Routines::       
-//* Index::                    
-//@end menu
-
-//@node Includes, Variables and Data structures, Main scheduling code, Main scheduling code
-//@subsection Includes
-
+#include "PosixSource.h"
 #include "Rts.h"
 #include "SchedAPI.h"
 #include "RtsUtils.h"
 #include "RtsFlags.h"
+#include "BlockAlloc.h"
 #include "Storage.h"
 #include "StgRun.h"
-#include "StgStartup.h"
-#include "GC.h"
 #include "Hooks.h"
+#define COMPILING_SCHEDULER
 #include "Schedule.h"
 #include "StgMiscClosures.h"
 #include "Storage.h"
-#include "Evaluator.h"
+#include "Interpreter.h"
 #include "Exception.h"
 #include "Printer.h"
-#include "Main.h"
 #include "Signals.h"
-#include "Profiling.h"
 #include "Sanity.h"
 #include "Stats.h"
-#include "Sparks.h"
-#include "Itimer.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)
 # include "GranSimRts.h"
 # include "GranSim.h"
 # include "FetchMe.h"
 # include "HLC.h"
 #endif
+#include "Sparks.h"
+#include "Capability.h"
+#include "OSThreads.h"
+#include  "Task.h"
 
+#ifdef HAVE_SYS_TYPES_H
+#include <sys/types.h>
+#endif
+#ifdef HAVE_UNISTD_H
+#include <unistd.h>
+#endif
+
+#include <string.h>
+#include <stdlib.h>
 #include <stdarg.h>
 
-//@node Variables and Data structures, Prototypes, Includes, Main scheduling code
-//@subsection Variables and Data structures
+#ifdef HAVE_ERRNO_H
+#include <errno.h>
+#endif
+
+#ifdef THREADED_RTS
+#define USED_IN_THREADED_RTS
+#else
+#define USED_IN_THREADED_RTS STG_UNUSED
+#endif
 
-/* Main threads:
- *
- * These are the threads which clients have requested that we run.  
- *
- * In an SMP build, we might have several concurrent clients all
- * waiting for results, and each one will wait on a condition variable
- * until the result is available.
- *
- * In non-SMP, clients are strictly nested: the first client calls
- * into the RTS, which might call out again to C with a _ccall_GC, and
- * eventually re-enter the RTS.
- *
- * Main threads information is kept in a linked list:
- */
-//@cindex StgMainThread
-typedef struct StgMainThread_ {
-  StgTSO *         tso;
-  SchedulerStatus  stat;
-  StgClosure **    ret;
-#ifdef SMP
-  pthread_cond_t wakeup;
+#ifdef RTS_SUPPORTS_THREADS
+#define USED_WHEN_RTS_SUPPORTS_THREADS
+#else
+#define USED_WHEN_RTS_SUPPORTS_THREADS STG_UNUSED
 #endif
-  struct StgMainThread_ *link;
-} StgMainThread;
 
 /* Main thread queue.
  * Locks required: sched_mutex.
  */
-static StgMainThread *main_threads;
+StgMainThread *main_threads = NULL;
 
 /* Thread queues.
  * Locks required: sched_mutex.
@@ -128,7 +120,7 @@ StgTSO* ActiveTSO = NULL; /* for assigning system costs; GranSim-Light only */
 /* rtsTime TimeOfNextEvent, EndOfTimeSlice;            now in GranSim.c */
 
 /* 
-   In GranSim we have a runable and a blocked queue for each processor.
+   In GranSim we have a runnable and a blocked queue for each processor.
    In order to minimise code changes new arrays run_queue_hds/tls
    are created. run_queue_hd is then a short cut (macro) for
    run_queue_hds[CurrentProc] (see GranSim.h).
@@ -137,25 +129,33 @@ StgTSO* ActiveTSO = NULL; /* for assigning system costs; GranSim-Light only */
 StgTSO *run_queue_hds[MAX_PROC], *run_queue_tls[MAX_PROC];
 StgTSO *blocked_queue_hds[MAX_PROC], *blocked_queue_tls[MAX_PROC];
 StgTSO *ccalling_threadss[MAX_PROC];
-StgTSO *all_threadss[MAX_PROC];
+/* We use the same global list of threads (all_threads) in GranSim as in
+   the std RTS (i.e. we are cheating). However, we don't use this list in
+   the GranSim specific code at the moment (so we are only potentially
+   cheating).  */
 
 #else /* !GRAN */
 
-StgTSO *run_queue_hd, *run_queue_tl;
-StgTSO *blocked_queue_hd, *blocked_queue_tl;
+StgTSO *run_queue_hd = NULL;
+StgTSO *run_queue_tl = NULL;
+StgTSO *blocked_queue_hd = NULL;
+StgTSO *blocked_queue_tl = NULL;
+StgTSO *sleeping_queue = NULL;    /* perhaps replace with a hash table? */
+
+#endif
 
 /* Linked list of all threads.
  * Used for detecting garbage collected threads.
  */
-StgTSO *all_threads;
+StgTSO *all_threads = NULL;
 
-/* Threads suspended in _ccall_GC.
+/* When a thread performs a safe C call (_ccall_GC, using old
+ * terminology), it gets put on the suspended_ccalling_threads
+ * list. Used by the garbage collector.
  */
 static StgTSO *suspended_ccalling_threads;
 
-static void GetRoots(void);
 static StgTSO *threadStackOverflow(StgTSO *tso);
-#endif
 
 /* 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
@@ -163,18 +163,20 @@ static StgTSO *threadStackOverflow(StgTSO *tso);
 */
 
 /* flag set by signal handler to precipitate a context switch */
-//@cindex context_switch
-nat context_switch;
+int context_switch = 0;
 
 /* if this flag is set as well, give up execution */
-//@cindex interrupted
-rtsBool interrupted;
+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'.
+ */
+rtsBool in_haskell = rtsFalse;
 
 /* Next thread ID to allocate.
- * Locks required: sched_mutex
+ * Locks required: thread_id_mutex
  */
-//@cindex next_thread_id
-StgThreadID next_thread_id = 1;
+static StgThreadID next_thread_id = 1;
 
 /*
  * Pointers to the state of the current thread.
@@ -185,83 +187,115 @@ StgThreadID next_thread_id = 1;
 /* The smallest stack size that makes any sense is:
  *    RESERVED_STACK_WORDS    (so we can get back from the stack overflow)
  *  + sizeofW(StgStopFrame)   (the stg_stop_thread_info frame)
- *  + 1                       (the realworld token for an IO thread)
  *  + 1                       (the closure to enter)
+ *  + 1                              (stg_ap_v_ret)
+ *  + 1                              (spare slot req'd by stg_ap_v_ret)
  *
  * A thread with this stack will bomb immediately with a stack
  * overflow, which will increase its stack size.  
  */
 
-#define MIN_STACK_WORDS (RESERVED_STACK_WORDS + sizeofW(StgStopFrame) + 2)
+#define MIN_STACK_WORDS (RESERVED_STACK_WORDS + sizeofW(StgStopFrame) + 3)
 
-/* Free capability list.
- * Locks required: sched_mutex.
- */
-#ifdef SMP
-//@cindex free_capabilities
-//@cindex n_free_capabilities
-Capability *free_capabilities; /* Available capabilities for running threads */
-nat n_free_capabilities;       /* total number of available capabilities */
-#else
-//@cindex MainRegTable
-Capability MainRegTable;       /* for non-SMP, we have one global capability */
-#endif
 
 #if defined(GRAN)
-StgTSO      *CurrentTSOs[MAX_PROC];
-#else
-StgTSO      *CurrentTSO;
+StgTSO *CurrentTSO;
 #endif
 
-rtsBool ready_to_gc;
+/*  This is used in `TSO.h' and gcc 2.96 insists that this variable actually 
+ *  exists - earlier gccs apparently didn't.
+ *  -= chak
+ */
+StgTSO dummy_tso;
+
+static rtsBool ready_to_gc;
 
-/* All our current task ids, saved in case we need to kill them later.
+/*
+ * Set to TRUE when entering a shutdown state (via shutdownHaskellAndExit()) --
+ * in an MT setting, needed to signal that a worker thread shouldn't hang around
+ * in the scheduler when it is out of work.
  */
-#ifdef SMP
-//@cindex task_ids
-task_info *task_ids;
-#endif
+static rtsBool shutting_down_scheduler = rtsFalse;
 
 void            addToBlockedQueue ( StgTSO *tso );
 
-static void     schedule          ( void );
+static void     schedule          ( StgMainThread *mainThread, Capability *initialCapability );
        void     interruptStgRts   ( void );
-static StgTSO * createThread_     ( nat size, rtsBool have_lock );
 
-#ifdef DEBUG
-static void sched_belch(char *s, ...);
+#if !defined(PAR) && !defined(RTS_SUPPORTS_THREADS)
+static void     detectBlackHoles  ( void );
 #endif
 
-#ifdef SMP
-//@cindex sched_mutex
-//@cindex term_mutex
-//@cindex thread_ready_cond
-//@cindex gc_pending_cond
-pthread_mutex_t sched_mutex       = PTHREAD_MUTEX_INITIALIZER;
-pthread_mutex_t term_mutex        = PTHREAD_MUTEX_INITIALIZER;
-pthread_cond_t  thread_ready_cond = PTHREAD_COND_INITIALIZER;
-pthread_cond_t  gc_pending_cond   = PTHREAD_COND_INITIALIZER;
+static void     raiseAsync_(StgTSO *tso, StgClosure *exception, rtsBool stop_at_atomically);
 
-nat await_death;
-#endif
+#if defined(RTS_SUPPORTS_THREADS)
+/* ToDo: carefully document the invariants that go together
+ *       with these synchronisation objects.
+ */
+Mutex     sched_mutex       = INIT_MUTEX_VAR;
+Mutex     term_mutex        = INIT_MUTEX_VAR;
+
+#endif /* RTS_SUPPORTS_THREADS */
 
 #if defined(PAR)
 StgTSO *LastTSO;
 rtsTime TimeOfLastYield;
+rtsBool emitSchedule = rtsTrue;
 #endif
 
-/*
- * The thread state for the main thread.
-// ToDo: check whether not needed any more
-StgTSO   *MainTSO;
- */
+#if DEBUG
+static char *whatNext_strs[] = {
+  "(unknown)",
+  "ThreadRunGHC",
+  "ThreadInterpret",
+  "ThreadKilled",
+  "ThreadRelocated",
+  "ThreadComplete"
+};
+#endif
 
+#if defined(PAR)
+StgTSO * createSparkThread(rtsSpark spark);
+StgTSO * activateSpark (rtsSpark spark);  
+#endif
 
-//@node Prototypes, Main scheduling loop, Variables and Data structures, Main scheduling code
-//@subsection Prototypes
+/* ----------------------------------------------------------------------------
+ * Starting Tasks
+ * ------------------------------------------------------------------------- */
+
+#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);
+}
 
-//@node Main scheduling loop, Suspend and Resume, Prototypes, Main scheduling code
-//@subsection Main scheduling loop
+void
+startSchedulerTaskIfNecessary(void)
+{
+  if(run_queue_hd != END_TSO_QUEUE
+    || blocked_queue_hd != END_TSO_QUEUE
+    || sleeping_queue != END_TSO_QUEUE)
+  {
+    if(!startingWorkerThread)
+    { // we don't want to start another worker thread
+      // just because the last one hasn't yet reached the
+      // "waiting for capability" state
+      startingWorkerThread = rtsTrue;
+      if(!startTask(taskStart))
+      {
+        startingWorkerThread = rtsFalse;
+      }
+    }
+  }
+}
+#endif
 
 /* ---------------------------------------------------------------------------
    Main scheduling loop.
@@ -282,10 +316,25 @@ StgTSO   *MainTSO;
       * waiting for work, or
       * waiting for a GC to complete.
 
+   GRAN version:
+     In a GranSim setup this loop iterates over the global event queue.
+     This revolves around the global event queue, which determines what 
+     to do next. Therefore, it's more complicated than either the 
+     concurrent or the parallel (GUM) setup.
+
+   GUM version:
+     GUM iterates over incoming messages.
+     It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
+     and sends out a fish whenever it has nothing to do; in-between
+     doing the actual reductions (shared code below) it processes the
+     incoming messages and deals with delayed operations 
+     (see PendingFetches).
+     This is not the ugliest code you could imagine, but it's bloody close.
+
    ------------------------------------------------------------------------ */
-//@cindex schedule
 static void
-schedule( void )
+schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
+          Capability *initialCapability )
 {
   StgTSO *t;
   Capability *cap;
@@ -293,230 +342,378 @@ schedule( void )
 #if defined(GRAN)
   rtsEvent *event;
 #elif defined(PAR)
+  StgSparkPool *pool;
   rtsSpark spark;
   StgTSO *tso;
   GlobalTaskId pe;
+  rtsBool receivedFinish = rtsFalse;
+# if defined(DEBUG)
+  nat tp_size, sp_size; // stats only
+# endif
 #endif
   rtsBool was_interrupted = rtsFalse;
+  nat prev_what_next;
   
-  ACQUIRE_LOCK(&sched_mutex);
+  // Pre-condition: sched_mutex is held.
+  // We might have a capability, passed in as initialCapability.
+  cap = initialCapability;
+
+#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
+  // simply initialise it in the non-threaded case
+  grabCapability(&cap);
+#endif
 
 #if defined(GRAN)
-# error ToDo: implement GranSim scheduler
+  /* 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];
+  }      
+
+  event = get_next_event();
+
+  while (event!=(rtsEvent*)NULL) {
+    /* Choose the processor with the next event */
+    CurrentProc = event->proc;
+    CurrentTSO = event->tso;
+
 #elif defined(PAR)
-  while (!GlobalStopPending) {          /* GlobalStopPending set in par_exit */
 
-    if (PendingFetches != END_BF_QUEUE) {
-        processFetches();
-    }
-#else
+  while (!receivedFinish) {    /* set by processMessages */
+                               /* when receiving PP_FINISH message         */ 
+
+#else // everything except GRAN and PAR
+
   while (1) {
+
 #endif
 
-    IF_DEBUG(scheduler, printAllThreads());
+     IF_DEBUG(scheduler, printAllThreads());
 
-    /* If we're interrupted (the user pressed ^C, or some other
-     * termination condition occurred), kill all the currently running
-     * threads.
-     */
-    if (interrupted) {
-      IF_DEBUG(scheduler, sched_belch("interrupted"));
-      for (t = run_queue_hd; t != END_TSO_QUEUE; t = t->link) {
-       deleteThread(t);
+#if defined(RTS_SUPPORTS_THREADS)
+      // Yield the capability to higher-priority tasks if necessary.
+      //
+      if (cap != NULL) {
+         yieldCapability(&cap);
       }
-      for (t = blocked_queue_hd; t != END_TSO_QUEUE; t = t->link) {
-       deleteThread(t);
-      }
-      run_queue_hd = run_queue_tl = END_TSO_QUEUE;
-      blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
-      interrupted = rtsFalse;
-      was_interrupted = rtsTrue;
-    }
 
-    /* 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...
-     */
-#ifdef SMP
-    { 
-      StgMainThread *m, **prev;
-      prev = &main_threads;
-      for (m = main_threads; m != NULL; m = m->link) {
-       switch (m->tso->what_next) {
-       case ThreadComplete:
-         if (m->ret) {
-           *(m->ret) = (StgClosure *)m->tso->sp[0];
-         }
-         *prev = m->link;
-         m->stat = Success;
-         pthread_cond_broadcast(&m->wakeup);
-         break;
-       case ThreadKilled:
-         *prev = m->link;
-         if (was_interrupted) {
-           m->stat = Interrupted;
-         } else {
-           m->stat = Killed;
-         }
-         pthread_cond_broadcast(&m->wakeup);
-         break;
-       default:
-         break;
-       }
+      // If we do not currently hold a capability, we wait for one
+      //
+      if (cap == NULL) {
+         waitForCapability(&sched_mutex, &cap,
+                           mainThread ? &mainThread->bound_thread_cond : NULL);
       }
+
+      // 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);
     }
+
+    //
+    // If we're interrupted (the user pressed ^C, or some other
+    // termination condition occurred), kill all the currently running
+    // threads.
+    //
+    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.
+       errorBelch("interrupted");
+       releaseCapability(cap);
+       RELEASE_LOCK(&sched_mutex);
+       shutdownHaskellAndExit(EXIT_SUCCESS);
 #else
-    /* 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) {
-       main_threads = main_threads->link;
-       if (m->tso->what_next == ThreadComplete) {
-         /* we finished successfully, fill in the return value */
-         if (m->ret) { *(m->ret) = (StgClosure *)m->tso->sp[0]; };
-         m->stat = Success;
-         return;
-       } else {
-         if (was_interrupted) {
-           m->stat = Interrupted;
-         } else {
-           m->stat = Killed;
-         }
-         return;
-       }
-      }
+       deleteAllThreads();
+#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);
     }
 #endif
 
-    /* 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.
+    //
+    // 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() );
+    }
+    // we can be interrupted while waiting for I/O...
+    if (interrupted) continue;
+
+    /* 
+     * 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.
      */
-#if defined(SMP)
+#if !defined(PAR) && !defined(RTS_SUPPORTS_THREADS)
+    if (   EMPTY_THREAD_QUEUES() )
     {
-      nat n = n_free_capabilities;
-      StgTSO *tso = run_queue_hd;
+       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() ) { goto not_deadlocked; }
+
+#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..."));
 
-      /* Count the run queue */
-      while (n > 0 && tso != END_TSO_QUEUE) {
-       tso = tso->link;
-       n--;
-      }
+           awaitUserSignals();
 
-      for (; n > 0; n--) {
-       StgClosure *spark;
-       spark = findSpark();
-       if (spark == NULL) {
-         break; /* no more sparks in the pool */
-       } else {
-         /* I'd prefer this to be done in activateSpark -- HWL */
-         /* tricky - it needs to hold the scheduler lock and
-          * not try to re-acquire it -- SDM */
-         StgTSO *tso;
-         tso = createThread_(RtsFlags.GcFlags.initialStkSize, rtsTrue);
-         pushClosure(tso,spark);
-         PUSH_ON_RUN_QUEUE(tso);
-#ifdef PAR
-         advisory_thread_count++;
+           // we might be interrupted...
+           if (interrupted) { continue; }
+
+           if (signals_pending()) {
+               RELEASE_LOCK(&sched_mutex);
+               startSignalHandlers();
+               ACQUIRE_LOCK(&sched_mutex);
+           }
+           ASSERT(!EMPTY_RUN_QUEUE());
+           goto not_deadlocked;
+       }
 #endif
-         
-         IF_DEBUG(scheduler,
-                  sched_belch("turning spark of closure %p into a thread",
-                              (StgClosure *)spark));
+
+       /* 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);
+               break;
+           default:
+               barf("deadlock: main thread blocked in a strange way");
+           }
        }
-      }
-      /* We need to wake up the other tasks if we just created some
-       * work for them.
-       */
-      if (n_free_capabilities - n > 1) {
-         pthread_cond_signal(&thread_ready_cond);
-      }
     }
-#endif /* SMP */
+  not_deadlocked:
 
-    /* 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.
-     * ToDo: what if another client comes along & requests another
-     * main thread?
-     */
-    if (blocked_queue_hd != END_TSO_QUEUE) {
-      awaitEvent(
-          (run_queue_hd == END_TSO_QUEUE)
-#ifdef SMP
-       && (n_free_capabilities == RtsFlags.ParFlags.nNodes)
-#endif
-       );
-    }
-    
-    /* check for signals each time around the scheduler */
-#ifndef __MINGW32__
-    if (signals_pending()) {
-      start_signal_handlers();
-    }
+#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
 
-    /* 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.  Inform
-     * all the main threads.
+#if defined(RTS_SUPPORTS_THREADS) || defined(mingw32_HOST_OS)
+    /* win32: might be back here due to awaitEvent() being abandoned
+     * as a result of a console event having been delivered.
      */
-#ifdef SMP
-    if (blocked_queue_hd == END_TSO_QUEUE
-       && run_queue_hd == END_TSO_QUEUE
-       && (n_free_capabilities == RtsFlags.ParFlags.nNodes)
-       ) {
-      StgMainThread *m;
-      for (m = main_threads; m != NULL; m = m->link) {
-         m->ret = NULL;
-         m->stat = Deadlock;
-         pthread_cond_broadcast(&m->wakeup);
-      }
-      main_threads = NULL;
-    }
-#else /* ! SMP */
-    if (blocked_queue_hd == END_TSO_QUEUE
-       && run_queue_hd == END_TSO_QUEUE) {
-      StgMainThread *m = main_threads;
-      m->ret = NULL;
-      m->stat = Deadlock;
-      main_threads = m->link;
-      return;
+    if ( EMPTY_RUN_QUEUE() ) {
+       continue; // nothing to do
     }
 #endif
 
-#ifdef SMP
-    /* If there's a GC pending, don't do anything until it has
-     * completed.
-     */
-    if (ready_to_gc) {
-      IF_DEBUG(scheduler,sched_belch("waiting for GC"));
-      pthread_cond_wait(&gc_pending_cond, &sched_mutex);
-    }
+#if defined(GRAN)
+    if (RtsFlags.GranFlags.Light)
+      GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
+
+    /* adjust time based on time-stamp */
+    if (event->time > CurrentTime[CurrentProc] &&
+        event->evttype != ContinueThread)
+      CurrentTime[CurrentProc] = event->time;
     
-    /* block until we've got a thread on the run queue and a free
-     * capability.
-     */
-    while (run_queue_hd == END_TSO_QUEUE || free_capabilities == NULL) {
-      IF_DEBUG(scheduler, sched_belch("waiting for work"));
-      pthread_cond_wait(&thread_ready_cond, &sched_mutex);
-      IF_DEBUG(scheduler, sched_belch("work now available"));
-    }
-#endif
+    /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
+    if (!RtsFlags.GranFlags.Light)
+      handleIdlePEs();
+
+    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, debugBelch("GRAN: doing ContinueThread\n"));
+      /* ToDo: check assertion
+      ASSERT(run_queue_hd != (StgTSO*)NULL &&
+            run_queue_hd != END_TSO_QUEUE);
+      */
+      /* Ignore ContinueThreads for fetching threads (if synchr comm) */
+      if (!RtsFlags.GranFlags.DoAsyncFetch &&
+         procStatus[CurrentProc]==Fetching) {
+       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) {
+       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) { 
+       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]) { 
+       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
+      }
+      /*
+      new_event(proc, proc, CurrentTime[proc],
+               FindWork,
+               (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
+      goto next_thread; 
+      */ /* Catches superfluous CONTINUEs -- should be unnecessary */
+      break; // now actually run the thread; DaH Qu'vam yImuHbej 
+
+    case FetchNode:
+      do_the_fetchnode(event);
+      goto next_thread;             /* handle next event in event queue  */
+      
+    case GlobalBlock:
+      do_the_globalblock(event);
+      goto next_thread;             /* handle next event in event queue  */
+      
+    case FetchReply:
+      do_the_fetchreply(event);
+      goto next_thread;             /* handle next event in event queue  */
+      
+    case UnblockThread:   /* Move from the blocked queue to the tail of */
+      do_the_unblock(event);
+      goto next_thread;             /* handle next event in event queue  */
+      
+    case ResumeThread:  /* Move from the blocked queue to the tail of */
+      /* the runnable queue ( i.e. Qu' SImqa'lu') */ 
+      event->tso->gran.blocktime += 
+       CurrentTime[CurrentProc] - event->tso->gran.blockedat;
+      do_the_startthread(event);
+      goto next_thread;             /* handle next event in event queue  */
+      
+    case StartThread:
+      do_the_startthread(event);
+      goto next_thread;             /* handle next event in event queue  */
+      
+    case MoveThread:
+      do_the_movethread(event);
+      goto next_thread;             /* handle next event in event queue  */
+      
+    case MoveSpark:
+      do_the_movespark(event);
+      goto next_thread;             /* handle next event in event queue  */
+      
+    case FindWork:
+      do_the_findwork(event);
+      goto next_thread;             /* handle next event in event queue  */
+      
+    default:
+      barf("Illegal event type %u\n", event->evttype);
+    }  /* switch */
+    
+    /* This point was scheduler_loop in the old RTS */
+
+    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, debugBelch("GRAN: time of next event is: %ld\n", 
+                        TimeOfNextEvent));
+
+    if (RtsFlags.GranFlags.Light) 
+      GranSimLight_leave_system(event, &ActiveTSO); 
+
+    EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
+
+    IF_DEBUG(gran, 
+            debugBelch("GRAN: end of time-slice is %#lx\n", EndOfTimeSlice));
+
+    /* in a GranSim setup the TSO stays on the run queue */
+    t = CurrentTSO;
+    /* Take a thread from the run queue. */
+    POP_RUN_QUEUE(t); // take_off_run_queue(t);
+
+    IF_DEBUG(gran, 
+            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
+
+    IF_DEBUG(gran, 
+            DumpGranEvent(GR_SCHEDULE, t));
+
+    procStatus[CurrentProc] = Busy;
 
-#if defined(GRAN)
-# error ToDo: implement GranSim scheduler
 #elif defined(PAR)
+    if (PendingFetches != END_BF_QUEUE) {
+        processFetches();
+    }
+
     /* ToDo: phps merge with spark activation above */
     /* check whether we have local work and send requests if we have none */
-    if (run_queue_hd == END_TSO_QUEUE) {  /* no runnable threads */
+    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
       if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
-         (pending_sparks_hd[REQUIRED_POOL] < pending_sparks_tl[REQUIRED_POOL] ||
-          pending_sparks_hd[ADVISORY_POOL] < pending_sparks_tl[ADVISORY_POOL])) {
+         pool->hd < pool->tl) {
        /* 
         * ToDo: add GC code check that we really have enough heap afterwards!!
         * Old comment:
@@ -525,27 +722,31 @@ schedule( void )
         * to turn one of those pending sparks into a
         * thread... 
         */
-       
-       spark = findSpark();                /* get a spark */
+
+       spark = findSpark(rtsFalse);                /* get a spark */
        if (spark != (rtsSpark) NULL) {
          tso = activateSpark(spark);       /* turn the spark into a thread */
-         IF_PAR_DEBUG(verbose,
-                      belch("== [%x] schedule: Created TSO %p (%d); %d threads active",
-                            mytid, tso, tso->id, advisory_thread_count));
+         IF_PAR_DEBUG(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 */
-           belch("^^ failed to activate spark");
+           debugBelch("==^^ failed to activate spark\n");
            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(ADVISORY_POOL)));
+                      debugBelch("==^^ no local sparks (spark pool contains only NFs: %d)\n", 
+                            spark_queue_len(pool)));
          goto next_thread;
        }
-      } else  
+      }
+
+      /* 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
         * (there may be some out there already), and wait for
@@ -554,47 +755,75 @@ schedule( void )
         * we're hoping to see.  (Of course, we still have to
         * respond to other types of messages.)
         */
-       if (//!fishing &&  
-           outstandingFishes < RtsFlags.ParFlags.maxFishes ) { // &&
-         // (last_fish_arrived_at+FISH_DELAY < CURRENT_TIME)) {
-         /* fishing set in sendFish, processFish;
+       TIME now = msTime() /*CURRENT_TIME*/;
+       IF_PAR_DEBUG(verbose, 
+                    debugBelch("--  now=%ld\n", now));
+       IF_PAR_DEBUG(verbose,
+                    if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
+                        (last_fish_arrived_at!=0 &&
+                         last_fish_arrived_at+RtsFlags.ParFlags.fishDelay > now)) {
+                      debugBelch("--$$ delaying FISH until %ld (last fish %ld, delay %ld, now %ld)\n",
+                            last_fish_arrived_at+RtsFlags.ParFlags.fishDelay,
+                            last_fish_arrived_at,
+                            RtsFlags.ParFlags.fishDelay, now);
+                    });
+       
+       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++;
+         }
        }
-       
-       processMessages();
+      
+       receivedFinish = processMessages();
        goto next_thread;
-       // ReSchedule(0);
       }
     } else if (PacketsWaiting()) {  /* Look for incoming messages */
-      processMessages();
+      receivedFinish = processMessages();
     }
 
     /* 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 */
-    t = take_off_run_queue(END_TSO_QUEUE);
+    POP_RUN_QUEUE(t);  // take_off_run_queue(END_TSO_QUEUE);
+    IF_DEBUG(sanity,checkTSO(t));
 
     /* ToDo: write something to the log-file
     if (RTSflags.ParFlags.granSimStats && !sameThread)
         DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
-    */
 
     CurrentTSO = t;
+    */
+    /* the spark pool for the current PE */
+    pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
+
+    IF_DEBUG(scheduler, 
+            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_DEBUG(scheduler, belch("--^^ %d sparks on [%#x] (hd=%x; tl=%x; lim=%x)", 
-                             spark_queue_len(ADVISORY_POOL), CURRENT_PROC,
-                             pending_sparks_hd[ADVISORY_POOL], 
-                             pending_sparks_tl[ADVISORY_POOL], 
-                             pending_sparks_lim[ADVISORY_POOL]));
-
-    IF_DEBUG(scheduler, belch("--== %d threads on [%#x] (hd=%x; tl=%x)", 
-                             run_queue_len(), CURRENT_PROC,
-                             run_queue_hd, run_queue_tl));
-
-    if (t != LastTSO) {
+    if (RtsFlags.ParFlags.ParStats.Full && 
+       (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 
@@ -604,110 +833,259 @@ schedule( void )
       */
       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
-     */
-    t = POP_RUN_QUEUE();
-    IF_DEBUG(sanity,checkTSO(t));
+    // grab a thread from the run queue
+    ASSERT(run_queue_hd != END_TSO_QUEUE);
+    POP_RUN_QUEUE(t);
 
+    // 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
-    
-    /* grab a capability
-     */
-#ifdef SMP
-    cap = free_capabilities;
-    free_capabilities = cap->link;
-    n_free_capabilities--;
-#else
-    cap = &MainRegTable;
+
+#ifdef THREADED_RTS
+    {
+      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; 
+       }
+      }
+    }
 #endif
+
+    cap->r.rCurrentTSO = t;
     
-    cap->rCurrentTSO = t;
-    
-    /* set the context_switch flag
+    /* context switches are now initiated by the timer signal, unless
+     * the user specified "context switch as often as possible", with
+     * +RTS -C0
      */
-    if (run_queue_hd == END_TSO_QUEUE)
-      context_switch = 0;
-    else
-      context_switch = 1;
+    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;
+
+run_thread:
 
     RELEASE_LOCK(&sched_mutex);
-    
-    IF_DEBUG(scheduler,sched_belch("running thread %d", t->id));
+
+    IF_DEBUG(scheduler, sched_belch("-->> running thread %ld %s ...", 
+                             (long)t->id, whatNext_strs[t->what_next]));
+
+#ifdef PROFILING
+    startHeapProfTimer();
+#endif
 
     /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
     /* Run the current thread 
      */
-    switch (cap->rCurrentTSO->what_next) {
+    prev_what_next = t->what_next;
+
+    errno = t->saved_errno;
+    in_haskell = rtsTrue;
+
+    switch (prev_what_next) {
+
     case ThreadKilled:
     case ThreadComplete:
-      /* Thread already finished, return to scheduler. */
-      ret = ThreadFinished;
-      break;
-    case ThreadEnterGHC:
-      ret = StgRun((StgFunPtr) stg_enterStackTop, cap);
-      break;
+       /* Thread already finished, return to scheduler. */
+       ret = ThreadFinished;
+       break;
+
     case ThreadRunGHC:
-      ret = StgRun((StgFunPtr) stg_returnToStackTop, cap);
-      break;
-    case ThreadEnterHugs:
-#ifdef INTERPRETER
-      {
-         StgClosure* c;
-        IF_DEBUG(scheduler,sched_belch("entering Hugs"));
-        c = (StgClosure *)(cap->rCurrentTSO->sp[0]);
-        cap->rCurrentTSO->sp += 1;
-        ret = enter(cap,c);
-         break;
-      }
-#else
-      barf("Panic: entered a BCO but no bytecode interpreter in this build");
-#endif
+       ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
+       break;
+
+    case ThreadInterpret:
+       ret = interpretBCO(cap);
+       break;
+
     default:
       barf("schedule: invalid what_next field");
     }
+
+    in_haskell = rtsFalse;
+
+    // The TSO might have moved, 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 */
 #ifdef PROFILING
+    stopHeapProfTimer();
     CCCS = CCS_SYSTEM;
 #endif
     
     ACQUIRE_LOCK(&sched_mutex);
-
-#ifdef SMP
-    IF_DEBUG(scheduler,fprintf(stderr,"scheduler (task %ld): ", pthread_self()););
-#else
-    IF_DEBUG(scheduler,fprintf(stderr,"scheduler: "););
+    
+#ifdef RTS_SUPPORTS_THREADS
+    IF_DEBUG(scheduler,debugBelch("sched (task %p): ", osThreadId()););
+#elif !defined(GRAN) && !defined(PAR)
+    IF_DEBUG(scheduler,debugBelch("sched: "););
 #endif
-    t = cap->rCurrentTSO;
     
+#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
+       if it is the same one as before */
+    LastTSO = t; 
+    TimeOfLastYield = CURRENT_TIME;
+#endif
+
     switch (ret) {
     case HeapOverflow:
+#if defined(GRAN)
+      IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
+      globalGranStats.tot_heapover++;
+#elif defined(PAR)
+      globalParStats.tot_heapover++;
+#endif
+
+      // 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;
+         nat blocks;
+         
+         blocks = (nat)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);
+             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 stopped: HeapOverflow", t->id));
+      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(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 */
       break;
       
     case StackOverflow:
+#if defined(GRAN)
+      IF_DEBUG(gran, 
+              DumpGranEvent(GR_DESCHEDULE, t));
+      globalGranStats.tot_stackover++;
+#elif defined(PAR)
+      // IF_DEBUG(par, 
+      // DumpGranEvent(GR_DESCHEDULE, t);
+      globalParStats.tot_stackover++;
+#endif
+      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.
        */
-      IF_DEBUG(scheduler,belch("thread %ld stopped, StackOverflow", t->id));
       threadPaused(t);
       { 
-       StgMainThread *m;
        /* enlarge the stack */
        StgTSO *new_t = threadStackOverflow(t);
        
@@ -715,24 +1093,30 @@ schedule( void )
         * 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;
-         }
+       if (t->main != NULL) {
+           t->main->tso = new_t;
        }
-       threadPaused(new_t);
        PUSH_ON_RUN_QUEUE(new_t);
       }
       break;
 
     case ThreadYielding:
+      // 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;
+
 #if defined(GRAN)
       IF_DEBUG(gran, 
               DumpGranEvent(GR_DESCHEDULE, t));
       globalGranStats.tot_yields++;
 #elif defined(PAR)
-      IF_DEBUG(par, 
-              DumpGranEvent(GR_DESCHEDULE, t));
+      // 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
@@ -740,125 +1124,423 @@ schedule( void )
        * GC is finished.
        */
       IF_DEBUG(scheduler,
-              if (t->what_next == ThreadEnterHugs) {
-                /* ToDo: or maybe a timer expired when we were in Hugs?
-                 * or maybe someone hit ctrl-C
-                 */
-                belch("thread %ld stopped to switch to Hugs", t->id);
-              } else {
-                belch("thread %ld stopped, yielding", t->id);
-              }
-              );
+               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) {
+         goto run_thread;
+      }
+
       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
+
+#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, 
+              debugBelch("GRAN: eventq and runnableq after adding yielded thread to queue again:\n");
+              G_EVENTQ(0);
+              G_CURR_THREADQ(0));
+#endif /* GRAN */
       break;
-      
+
     case ThreadBlocked:
 #if defined(GRAN)
-# error ToDo: implement GranSim scheduler
-#elif defined(PAR)
-      IF_DEBUG(par, 
+      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)); 
-#else
-#endif
+      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.
        */
+      ASSERT(t->why_blocked != NotBlocked);
       IF_DEBUG(scheduler,
-              fprintf(stderr, "thread %d stopped, ", t->id);
+              debugBelch("--<< thread %d (%s) stopped: ", 
+                      t->id, whatNext_strs[t->what_next]);
               printThreadBlockage(t);
-              fprintf(stderr, "\n"));
+              debugBelch("\n"));
+
+      /* Only for dumping event to log file 
+        ToDo: do I need this in GranSim, too?
+      blockThread(t);
+      */
+#endif
       threadPaused(t);
       break;
-      
+
     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.
        */
-      IF_DEBUG(scheduler,belch("thread %ld finished", t->id));
-      t->what_next = ThreadComplete;
+      /* We also end up here if the thread kills itself with an
+       * uncaught exception, see Exception.hc.
+       */
+      IF_DEBUG(scheduler,debugBelch("--++ thread %d (%s) finished\n", 
+                              t->id, whatNext_strs[t->what_next]));
 #if defined(GRAN)
-      // ToDo: endThread(t, CurrentProc); // clean-up the thread
+      endThread(t, CurrentProc); // clean-up the thread
 #elif defined(PAR)
+      /* For now all are advisory -- HWL */
+      //if(t->priority==AdvisoryPriority) ??
       advisory_thread_count--;
-      if (RtsFlags.ParFlags.ParStats.Full) 
+      
+# ifdef DIST
+      if(t->dist.priority==RevalPriority)
+       FinishReval(t);
+# endif
+      
+      if (RtsFlags.ParFlags.ParStats.Full &&
+         !RtsFlags.ParFlags.ParStats.Suppressed) 
        DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
 #endif
+
+      //
+      // 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
+         )
+      {
+         // 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 (was_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;
+      }
+
+#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
       break;
-      
+
     default:
-      barf("doneThread: invalid thread return code");
+      barf("schedule: invalid thread return code %d", (int)ret);
+    }
+
+#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.
+    if (performHeapProfile ||
+       (RtsFlags.ProfFlags.profileInterval==0 &&
+        RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
+       GarbageCollect(GetRoots, rtsTrue);
+       heapCensus();
+       performHeapProfile = rtsFalse;
+       ready_to_gc = rtsFalse; // we already GC'd
     }
-    
-#ifdef SMP
-    cap->link = free_capabilities;
-    free_capabilities = cap;
-    n_free_capabilities++;
 #endif
 
-#ifdef SMP
-    if (ready_to_gc && n_free_capabilities == RtsFlags.ParFlags.nNodes) 
-#else
-    if (ready_to_gc) 
+    if (ready_to_gc) {
+      /* 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
-      {
+          }
+        }
+      }
+
       /* 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.
        */
-#ifdef SMP
+#if defined(RTS_SUPPORTS_THREADS)
       IF_DEBUG(scheduler,sched_belch("doing GC"));
 #endif
-      GarbageCollect(GetRoots);
+      GarbageCollect(GetRoots,rtsFalse);
       ready_to_gc = rtsFalse;
-#ifdef SMP
-      pthread_cond_broadcast(&gc_pending_cond);
-#endif
+#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, 
+              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 */
-
-#else /* GRAN */
-  /* not any more
-  next_thread:
-    t = take_off_run_queue(END_TSO_QUEUE);
-  */
 #endif /* GRAN */
+
   } /* end of while(1) */
+
+  IF_PAR_DEBUG(verbose,
+              debugBelch("== Leaving schedule() after having received Finish\n"));
+}
+
+/* ---------------------------------------------------------------------------
+ * rtsSupportsBoundThreads(): is the RTS built to support bound threads?
+ * used by Control.Concurrent for error checking.
+ * ------------------------------------------------------------------------- */
+StgBool
+rtsSupportsBoundThreads(void)
+{
+#ifdef THREADED_RTS
+  return rtsTrue;
+#else
+  return rtsFalse;
+#endif
 }
 
-/* A hack for Hugs concurrency support.  Needs sanitisation (?) */
-void deleteAllThreads ( void )
+/* ---------------------------------------------------------------------------
+ * isThreadBound(tso): check whether tso is bound to an OS thread.
+ * ------------------------------------------------------------------------- */
+StgBool
+isThreadBound(StgTSO* tso USED_IN_THREADED_RTS)
 {
-  StgTSO* t;
-  IF_DEBUG(scheduler,sched_belch("deleteAllThreads()"));
-  for (t = run_queue_hd; t != END_TSO_QUEUE; t = t->link) {
-    deleteThread(t);
-  }
-  for (t = blocked_queue_hd; t != END_TSO_QUEUE; t = t->link) {
-    deleteThread(t);
+#ifdef THREADED_RTS
+  return (tso->main != NULL);
+#endif
+  return rtsFalse;
+}
+
+/* ---------------------------------------------------------------------------
+ * 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
+#ifndef FORKPROCESS_PRIMOP_SUPPORTED
+           STG_UNUSED
+#endif
+           )
+{
+#ifdef FORKPROCESS_PRIMOP_SUPPORTED
+  pid_t pid;
+  StgTSO* t,*next;
+  StgMainThread *m;
+  SchedulerStatus rc;
+
+  IF_DEBUG(scheduler,sched_belch("forking!"));
+  rts_lock(); // This not only acquires sched_mutex, it also
+              // makes sure that no other threads are running
+
+  pid = fork();
+
+  if (pid) { /* parent */
+
+  /* just return the pid */
+    rts_unlock();
+    return pid;
+    
+  } else { /* child */
+    
+    
+      // delete all threads
+    run_queue_hd = run_queue_tl = END_TSO_QUEUE;
+    
+    for (t = all_threads; t != END_TSO_QUEUE; t = next) {
+      next = t->link;
+
+        // don't allow threads to catch the ThreadKilled exception
+      deleteThreadImmediately(t);
+    }
+    
+      // wipe the main thread list
+    while((m = main_threads) != NULL) {
+      main_threads = m->link;
+# ifdef THREADED_RTS
+      closeCondition(&m->bound_thread_cond);
+# endif
+      stgFree(m);
+    }
+    
+    rc = rts_evalStableIO(entry, NULL);  // run the action
+    rts_checkSchedStatus("forkProcess",rc);
+    
+    rts_unlock();
+    
+    hs_exit();                      // clean up and exit
+    stg_exit(0);
   }
-  run_queue_hd = run_queue_tl = END_TSO_QUEUE;
-  blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
+#else /* !FORKPROCESS_PRIMOP_SUPPORTED */
+  barf("forkProcess#: primop not supported, sorry!\n");
+  return -1;
+#endif
+}
+
+/* ---------------------------------------------------------------------------
+ * deleteAllThreads():  kill all the live threads.
+ *
+ * This is used when we catch a user interrupt (^C), before performing
+ * any necessary cleanups and running finalizers.
+ *
+ * Locks: sched_mutex held.
+ * ------------------------------------------------------------------------- */
+   
+void
+deleteAllThreads ( void )
+{
+  StgTSO* t, *next;
+  IF_DEBUG(scheduler,sched_belch("deleting all threads"));
+  for (t = all_threads; t != END_TSO_QUEUE; t = next) {
+      next = t->global_link;
+      deleteThread(t);
+  }      
+
+  // 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(sleeping_queue == END_TSO_QUEUE);
 }
 
 /* startThread and  insertThread are now in GranSim.c -- HWL */
 
-//@node Suspend and Resume, Run queue code, Main scheduling loop, Main scheduling code
-//@subsection Suspend and Resume
 
 /* ---------------------------------------------------------------------------
  * Suspending & resuming Haskell threads.
@@ -876,40 +1558,74 @@ void deleteAllThreads ( void )
  * ------------------------------------------------------------------------- */
    
 StgInt
-suspendThread( Capability *cap )
+suspendThread( StgRegTable *reg )
 {
   nat tok;
+  Capability *cap;
+  int saved_errno = errno;
+
+  /* assume that *reg is a pointer to the StgRegTable part
+   * of a Capability.
+   */
+  cap = (Capability *)((void *)((unsigned char*)reg - sizeof(StgFunTable)));
 
   ACQUIRE_LOCK(&sched_mutex);
 
   IF_DEBUG(scheduler,
-          sched_belch("thread %d did a _ccall_gc\n", cap->rCurrentTSO->id));
+          sched_belch("thread %d did a _ccall_gc", cap->r.rCurrentTSO->id));
 
-  threadPaused(cap->rCurrentTSO);
-  cap->rCurrentTSO->link = suspended_ccalling_threads;
-  suspended_ccalling_threads = cap->rCurrentTSO;
+  // XXX this might not be necessary --SDM
+  cap->r.rCurrentTSO->what_next = ThreadRunGHC;
+
+  threadPaused(cap->r.rCurrentTSO);
+  cap->r.rCurrentTSO->link = suspended_ccalling_threads;
+  suspended_ccalling_threads = cap->r.rCurrentTSO;
+
+  if(cap->r.rCurrentTSO->blocked_exceptions == NULL)  {
+      cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
+      cap->r.rCurrentTSO->blocked_exceptions = END_TSO_QUEUE;
+  } else {
+      cap->r.rCurrentTSO->why_blocked = BlockedOnCCall_NoUnblockExc;
+  }
 
   /* Use the thread ID as the token; it should be unique */
-  tok = cap->rCurrentTSO->id;
+  tok = cap->r.rCurrentTSO->id;
 
-#ifdef SMP
-  cap->link = free_capabilities;
-  free_capabilities = cap;
-  n_free_capabilities++;
+  /* Hand back capability */
+  releaseCapability(cap);
+  
+#if defined(RTS_SUPPORTS_THREADS)
+  /* Preparing to leave the RTS, so ensure there's a native thread/task
+     waiting to take over.
+  */
+  IF_DEBUG(scheduler, sched_belch("worker (token %d): leaving RTS", tok));
 #endif
 
   RELEASE_LOCK(&sched_mutex);
+  
+  errno = saved_errno;
+  in_haskell = rtsFalse;
   return tok; 
 }
 
-Capability *
+StgRegTable *
 resumeThread( StgInt tok )
 {
   StgTSO *tso, **prev;
   Capability *cap;
+  int saved_errno = errno;
 
+#if defined(RTS_SUPPORTS_THREADS)
+  /* Wait for permission to re-enter the RTS with the result. */
   ACQUIRE_LOCK(&sched_mutex);
+  waitForReturnCapability(&sched_mutex, &cap);
+
+  IF_DEBUG(scheduler, sched_belch("worker (token %d): re-entering RTS", tok));
+#else
+  grabCapability(&cap);
+#endif
 
+  /* Remove the thread off of the suspended list */
   prev = &suspended_ccalling_threads;
   for (tso = suspended_ccalling_threads; 
        tso != END_TSO_QUEUE; 
@@ -922,24 +1638,21 @@ resumeThread( StgInt tok )
   if (tso == END_TSO_QUEUE) {
     barf("resumeThread: thread not found");
   }
-
-#ifdef SMP
-  while (free_capabilities == NULL) {
-    IF_DEBUG(scheduler, sched_belch("waiting to resume"));
-    pthread_cond_wait(&thread_ready_cond, &sched_mutex);
-    IF_DEBUG(scheduler, sched_belch("resuming thread %d", tso->id));
+  tso->link = END_TSO_QUEUE;
+  
+  if(tso->why_blocked == BlockedOnCCall) {
+      awakenBlockedQueueNoLock(tso->blocked_exceptions);
+      tso->blocked_exceptions = NULL;
   }
-  cap = free_capabilities;
-  free_capabilities = cap->link;
-  n_free_capabilities--;
-#else  
-  cap = &MainRegTable;
-#endif
-
-  cap->rCurrentTSO = tso;
+  
+  /* Reset blocking status */
+  tso->why_blocked  = NotBlocked;
 
+  cap->r.rCurrentTSO = tso;
   RELEASE_LOCK(&sched_mutex);
-  return cap;
+  errno = saved_errno;
+  in_haskell = rtsTrue;
+  return &cap->r;
 }
 
 
@@ -955,10 +1668,11 @@ static void unblockThread(StgTSO *tso);
  * instances of Eq/Ord for ThreadIds.
  * ------------------------------------------------------------------------ */
 
-int cmp_thread(const StgTSO *tso1, const StgTSO *tso2) 
+int
+cmp_thread(StgPtr tso1, StgPtr tso2) 
 { 
-  StgThreadID id1 = tso1->id; 
-  StgThreadID id2 = tso2->id;
+  StgThreadID id1 = ((StgTSO *)tso1)->id; 
+  StgThreadID id2 = ((StgTSO *)tso2)->id;
  
   if (id1 < id2) return (-1);
   if (id1 > id2) return 1;
@@ -966,6 +1680,33 @@ int cmp_thread(const StgTSO *tso1, const StgTSO *tso2)
 }
 
 /* ---------------------------------------------------------------------------
+ * Fetching the ThreadID from an StgTSO.
+ *
+ * This is used in the implementation of Show for ThreadIds.
+ * ------------------------------------------------------------------------ */
+int
+rts_getThreadId(StgPtr tso) 
+{
+  return ((StgTSO *)tso)->id;
+}
+
+#ifdef DEBUG
+void
+labelThread(StgPtr tso, char *label)
+{
+  int len;
+  void *buf;
+
+  /* Caveat: Once set, you can only set the thread name to "" */
+  len = strlen(label)+1;
+  buf = stgMallocBytes(len * sizeof(char), "Schedule.c:labelThread()");
+  strncpy(buf,label,len);
+  /* Update will free the old memory for us */
+  updateThreadLabel(((StgTSO *)tso)->id,buf);
+}
+#endif /* DEBUG */
+
+/* ---------------------------------------------------------------------------
    Create a new thread.
 
    The new thread starts with the given stack size.  Before the
@@ -975,30 +1716,19 @@ int cmp_thread(const StgTSO *tso1, const StgTSO *tso2)
 
    createGenThread() and createIOThread() (in SchedAPI.h) are
    convenient packaged versions of this function.
+
+   currently pri (priority) is only used in a GRAN setup -- HWL
    ------------------------------------------------------------------------ */
-//@cindex createThread
 #if defined(GRAN)
-/* currently pri (priority) is only used in a GRAN setup -- HWL */
+/*   currently pri (priority) is only used in a GRAN setup -- HWL */
 StgTSO *
-createThread(nat stack_size, StgInt pri)
-{
-  return createThread_(stack_size, rtsFalse, pri);
-}
-
-static StgTSO *
-createThread_(nat size, rtsBool have_lock, StgInt pri)
-{
+createThread(nat size, StgInt pri)
 #else
 StgTSO *
-createThread(nat stack_size)
+createThread(nat size)
+#endif
 {
-  return createThread_(stack_size, rtsFalse);
-}
 
-static StgTSO *
-createThread_(nat size, rtsBool have_lock)
-{
-#endif
     StgTSO *tso;
     nat stack_size;
 
@@ -1007,7 +1737,7 @@ createThread_(nat size, rtsBool have_lock)
   /* 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;
   }
@@ -1028,46 +1758,41 @@ createThread_(nat size, rtsBool have_lock)
   stack_size = size - TSO_STRUCT_SIZEW;
 
   tso = (StgTSO *)allocate(size);
-  TICK_ALLOC_TSO(size-TSO_STRUCT_SIZEW, 0);
+  TICK_ALLOC_TSO(stack_size, 0);
 
-  SET_HDR(tso, &TSO_info, CCS_MAIN);
+  SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
 #if defined(GRAN)
   SET_GRAN_HDR(tso, ThisPE);
 #endif
-  tso->what_next     = ThreadEnterGHC;
 
-  /* tso->id needs to be unique.  For now we use a heavyweight mutex to
-   * protect the increment operation on next_thread_id.
-   * In future, we could use an atomic increment instead.
-   */
-  if (!have_lock) { ACQUIRE_LOCK(&sched_mutex); }
-  tso->id = next_thread_id++; 
-  if (!have_lock) { RELEASE_LOCK(&sched_mutex); }
+  // Always start with the compiled code evaluator
+  tso->what_next = ThreadRunGHC;
 
+  tso->id = next_thread_id++; 
   tso->why_blocked  = NotBlocked;
   tso->blocked_exceptions = NULL;
 
-  tso->splim        = (P_)&(tso->stack) + RESERVED_STACK_WORDS;
+  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
 
   /* put a stop frame on the stack */
   tso->sp -= sizeofW(StgStopFrame);
-  SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_MAIN);
-  tso->su = (StgUpdateFrame*)tso->sp;
-
-  IF_DEBUG(scheduler,belch("---- Initialised TSO %ld (%p), stack size = %lx words", 
-                          tso->id, tso, tso->stack_size));
+  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
@@ -1076,18 +1801,30 @@ createThread_(nat size, rtsBool have_lock)
    */
 #endif
 
+#if defined(GRAN) 
+  if (RtsFlags.GranFlags.GranSimStats.Full) 
+    DumpGranEvent(GR_START,tso);
+#elif defined(PAR)
+  if (RtsFlags.ParFlags.ParStats.Full) 
+    DumpGranEvent(GR_STARTQ,tso);
+  /* HACk to avoid SCHEDULE 
+     LastTSO = tso; */
+#endif
+
   /* Link the new thread on the global thread list.
    */
-#if defined(GRAN)
-#error ToDo
-#else
   tso->global_link = all_threads;
   all_threads = tso;
+
+#if defined(DIST)
+  tso->dist.priority = MandatoryPriority; //by default that is...
 #endif
 
 #if defined(GRAN)
   tso->gran.pri = pri;
+# if defined(DEBUG)
   tso->gran.magic = TSO_MAGIC; // debugging only
+# endif
   tso->gran.sparkname   = 0;
   tso->gran.startedat   = CURRENT_TIME; 
   tso->gran.exported    = 0;
@@ -1108,6 +1845,9 @@ createThread_(nat size, rtsBool have_lock)
 
   IF_DEBUG(gran,printTSO(tso));
 #elif defined(PAR)
+# if defined(DEBUG)
+  tso->par.magic = TSO_MAGIC; // debugging only
+# endif
   tso->par.sparkname   = 0;
   tso->par.startedat   = CURRENT_TIME; 
   tso->par.exported    = 0;
@@ -1128,12 +1868,90 @@ createThread_(nat size, rtsBool have_lock)
   globalGranStats.threads_created_on_PE[CurrentProc]++;
   globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
   globalGranStats.tot_sq_probes++;
+#elif defined(PAR)
+  // collect parallel global statistics (currently done together with GC stats)
+  if (RtsFlags.ParFlags.ParStats.Global &&
+      RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
+    //debugBelch("Creating thread %d @ %11.2f\n", tso->id, usertime()); 
+    globalParStats.tot_threads_created++;
+  }
 #endif 
 
+#if defined(GRAN)
+  IF_GRAN_DEBUG(pri,
+               sched_belch("==__ schedule: Created TSO %d (%p);",
+                     CurrentProc, tso, tso->id));
+#elif defined(PAR)
+    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;
+}
+
+#if defined(PAR)
+/* RFP:
+   all parallel thread creation calls should fall through the following routine.
+*/
+StgTSO *
+createSparkThread(rtsSpark spark) 
+{ StgTSO *tso;
+  ASSERT(spark != (rtsSpark)NULL);
+  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)",
+         RtsFlags.ParFlags.maxThreads, advisory_thread_count);    
+    return END_TSO_QUEUE;
+  }
+  else
+  { threadsCreated++;
+    tso = createThread(RtsFlags.GcFlags.initialStkSize);
+    if (tso==END_TSO_QUEUE)    
+      barf("createSparkThread: Cannot create TSO");
+#if defined(DIST)
+    tso->priority = AdvisoryPriority;
+#endif
+    pushClosure(tso,spark);
+    PUSH_ON_RUN_QUEUE(tso);
+    advisory_thread_count++;    
+  }
+  return tso;
+}
+#endif
+
+/*
+  Turn a spark into a thread.
+  ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
+*/
+#if defined(PAR)
+StgTSO *
+activateSpark (rtsSpark spark) 
+{
+  StgTSO *tso;
+
+  tso = createSparkThread(spark);
+  if (RtsFlags.ParFlags.ParStats.Full) {   
+    //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
+    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;
+  // tso->gran.locked =   !spark->global;
+  // tso->gran.sparkname = spark->name;
+
   return tso;
 }
+#endif
+
+static SchedulerStatus waitThread_(/*out*/StgMainThread* m,
+                                  Capability *initialCapability
+                                  );
+
 
 /* ---------------------------------------------------------------------------
  * scheduleThread()
@@ -1145,40 +1963,81 @@ createThread_(nat size, rtsBool have_lock)
  * on this thread's stack before the scheduler is invoked.
  * ------------------------------------------------------------------------ */
 
+static void scheduleThread_ (StgTSO* tso);
+
 void
-scheduleThread(StgTSO *tso)
+scheduleThread_(StgTSO *tso)
 {
-  ACQUIRE_LOCK(&sched_mutex);
-
-  /* Put the new thread on the head of the runnable queue.  The caller
-   * better push an appropriate closure on this thread's stack
-   * beforehand.  In the SMP case, the thread may start running as
-   * soon as we release the scheduler lock below.
-   */
-  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();
+}
 
-  IF_DEBUG(scheduler,printTSO(tso));
+void
+scheduleThread(StgTSO* tso)
+{
+  ACQUIRE_LOCK(&sched_mutex);
+  scheduleThread_(tso);
   RELEASE_LOCK(&sched_mutex);
 }
 
-/* ---------------------------------------------------------------------------
- * startTasks()
- *
- * Start up Posix threads to run each of the scheduler tasks.
- * I believe the task ids are not needed in the system as defined.
- *  KH @ 25/10/99
- * ------------------------------------------------------------------------ */
+#if defined(RTS_SUPPORTS_THREADS)
+static Condition bound_cond_cache;
+static int bound_cond_cache_full = 0;
+#endif
+
 
-#ifdef SMP
-static void *
-taskStart( void *arg STG_UNUSED )
+SchedulerStatus
+scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret,
+                  Capability *initialCapability)
 {
-  schedule();
-  return NULL;
-}
+    // Precondition: sched_mutex must be held
+    StgMainThread *m;
+
+    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_full) {
+       m->bound_thread_cond = bound_cond_cache;
+       bound_cond_cache_full = 0;
+    } else {
+       initCondition(&m->bound_thread_cond);
+    }
 #endif
 
+    /* Put the thread on the main-threads list prior to scheduling the TSO.
+       Failure to do so introduces a race condition in the MT case (as
+       identified by Wolfgang Thaller), whereby the new task/OS thread 
+       created by scheduleThread_() would complete prior to the thread
+       that spawned it managed to put 'itself' on the main-threads list.
+       The upshot of it all being that the worker thread wouldn't get to
+       signal the completion of the its work item for the main thread to
+       see (==> it got stuck waiting.)    -- sof 6/02.
+    */
+    IF_DEBUG(scheduler, sched_belch("waiting for thread (%d)", tso->id));
+    
+    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.
+
+    return waitThread_(m, initialCapability);
+}
+
 /* ---------------------------------------------------------------------------
  * initScheduler()
  *
@@ -1186,22 +2045,8 @@ taskStart( void *arg STG_UNUSED )
  * queues contained any threads, they'll be garbage collected at the
  * next pass.
  *
- * This now calls startTasks(), so should only be called once!  KH @ 25/10/99
  * ------------------------------------------------------------------------ */
 
-#ifdef SMP
-static void
-term_handler(int sig STG_UNUSED)
-{
-  stat_workerStop();
-  ACQUIRE_LOCK(&term_mutex);
-  await_death--;
-  RELEASE_LOCK(&term_mutex);
-  pthread_exit(NULL);
-}
-#endif
-
-//@cindex initScheduler
 void 
 initScheduler(void)
 {
@@ -1214,12 +2059,14 @@ initScheduler(void)
     blocked_queue_hds[i]  = END_TSO_QUEUE;
     blocked_queue_tls[i]  = END_TSO_QUEUE;
     ccalling_threadss[i]  = END_TSO_QUEUE;
+    sleeping_queue        = END_TSO_QUEUE;
   }
 #else
   run_queue_hd      = END_TSO_QUEUE;
   run_queue_tl      = END_TSO_QUEUE;
   blocked_queue_hd  = END_TSO_QUEUE;
   blocked_queue_tl  = END_TSO_QUEUE;
+  sleeping_queue    = END_TSO_QUEUE;
 #endif 
 
   suspended_ccalling_threads  = END_TSO_QUEUE;
@@ -1230,292 +2077,99 @@ initScheduler(void)
   context_switch = 0;
   interrupted    = 0;
 
-  enteredCAFs = END_CAF_LIST;
-
-  /* Install the SIGHUP handler */
-#ifdef SMP
-  {
-    struct sigaction action,oact;
-
-    action.sa_handler = term_handler;
-    sigemptyset(&action.sa_mask);
-    action.sa_flags = 0;
-    if (sigaction(SIGTERM, &action, &oact) != 0) {
-      barf("can't install TERM handler");
-    }
-  }
+  RtsFlags.ConcFlags.ctxtSwitchTicks =
+      RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
+      
+#if defined(RTS_SUPPORTS_THREADS)
+  /* Initialise the mutex and condition variables used by
+   * the scheduler. */
+  initMutex(&sched_mutex);
+  initMutex(&term_mutex);
 #endif
+  
+  ACQUIRE_LOCK(&sched_mutex);
 
-#ifdef SMP
-  /* Allocate N Capabilities */
-  {
-    nat i;
-    Capability *cap, *prev;
-    cap  = NULL;
-    prev = NULL;
-    for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
-      cap = stgMallocBytes(sizeof(Capability), "initScheduler:capabilities");
-      cap->link = prev;
-      prev = cap;
-    }
-    free_capabilities = cap;
-    n_free_capabilities = RtsFlags.ParFlags.nNodes;
-  }
-  IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Allocated %d capabilities\n",
-                            n_free_capabilities););
+  /* A capability holds the state a native thread needs in
+   * order to execute STG code. At least one capability is
+   * floating around (only SMP builds have more than one).
+   */
+  initCapabilities();
+  
+#if defined(RTS_SUPPORTS_THREADS)
+    /* start our haskell execution tasks */
+    startTaskManager(0,taskStart);
 #endif
 
-#if defined(SMP) || defined(PAR)
+#if /* defined(SMP) ||*/ defined(PAR)
   initSparkPools();
 #endif
-}
 
-#ifdef SMP
-void
-startTasks( void )
-{
-  nat i;
-  int r;
-  pthread_t tid;
-  
-  /* make some space for saving all the thread ids */
-  task_ids = stgMallocBytes(RtsFlags.ParFlags.nNodes * sizeof(task_info),
-                           "initScheduler:task_ids");
-  
-  /* and create all the threads */
-  for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
-    r = pthread_create(&tid,NULL,taskStart,NULL);
-    if (r != 0) {
-      barf("startTasks: Can't create new Posix thread");
-    }
-    task_ids[i].id = tid;
-    task_ids[i].mut_time = 0.0;
-    task_ids[i].mut_etime = 0.0;
-    task_ids[i].gc_time = 0.0;
-    task_ids[i].gc_etime = 0.0;
-    task_ids[i].elapsedtimestart = elapsedtime();
-    IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Started task: %ld\n",tid););
-  }
+  RELEASE_LOCK(&sched_mutex);
 }
-#endif
 
 void
 exitScheduler( void )
 {
-#ifdef SMP
-  nat i;
-
-  /* Don't want to use pthread_cancel, since we'd have to install
-   * these silly exception handlers (pthread_cleanup_{push,pop}) around
-   * all our locks.
-   */
-#if 0
-  /* Cancel all our tasks */
-  for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
-    pthread_cancel(task_ids[i].id);
-  }
-  
-  /* Wait for all the tasks to terminate */
-  for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
-    IF_DEBUG(scheduler,fprintf(stderr,"scheduler: waiting for task %ld\n", 
-                              task_ids[i].id));
-    pthread_join(task_ids[i].id, NULL);
-  }
-#endif
-
-  /* Send 'em all a SIGHUP.  That should shut 'em up.
-   */
-  await_death = RtsFlags.ParFlags.nNodes;
-  for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
-    pthread_kill(task_ids[i].id,SIGTERM);
-  }
-  while (await_death > 0) {
-    sched_yield();
-  }
+#if defined(RTS_SUPPORTS_THREADS)
+  stopTaskManager();
 #endif
+  shutting_down_scheduler = rtsTrue;
 }
 
-/* -----------------------------------------------------------------------------
+/* ----------------------------------------------------------------------------
    Managing the per-task allocation areas.
    
    Each capability comes with an allocation area.  These are
    fixed-length block lists into which allocation can be done.
 
    ToDo: no support for two-space collection at the moment???
-   -------------------------------------------------------------------------- */
-
-/* -----------------------------------------------------------------------------
- * waitThread is the external interface for running a new computataion
- * and waiting for the result.
- *
- * In the non-SMP case, we create a new main thread, push it on the 
- * main-thread stack, and invoke the scheduler to run it.  The
- * scheduler will return when the top main thread on the stack has
- * completed or died, and fill in the necessary fields of the
- * main_thread structure.
- *
- * In the SMP case, we create a main thread as before, but we then
- * create a new condition variable and sleep on it.  When our new
- * main thread has completed, we'll be woken up and the status/result
- * will be in the main_thread struct.
- * -------------------------------------------------------------------------- */
+   ------------------------------------------------------------------------- */
 
+static
 SchedulerStatus
-waitThread(StgTSO *tso, /*out*/StgClosure **ret)
+waitThread_(StgMainThread* m, Capability *initialCapability)
 {
-  StgMainThread *m;
   SchedulerStatus stat;
 
-  ACQUIRE_LOCK(&sched_mutex);
-  
-  m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
-
-  m->tso = tso;
-  m->ret = ret;
-  m->stat = NoStatus;
-#ifdef SMP
-  pthread_cond_init(&m->wakeup, NULL);
-#endif
-
-  m->link = main_threads;
-  main_threads = m;
+  // Precondition: sched_mutex must be held.
+  IF_DEBUG(scheduler, sched_belch("new main thread (%d)", m->tso->id));
 
-  IF_DEBUG(scheduler, fprintf(stderr, "scheduler: new main thread (%d)\n", 
-                             m->tso->id));
-
-#ifdef SMP
-  do {
-    pthread_cond_wait(&m->wakeup, &sched_mutex);
-  } while (m->stat == NoStatus);
+#if defined(GRAN)
+  /* GranSim specific init */
+  CurrentTSO = m->tso;                // the TSO to run
+  procStatus[MainProc] = Busy;        // status of main PE
+  CurrentProc = MainProc;             // PE to run it on
+  schedule(m,initialCapability);
 #else
-  schedule();
+  schedule(m,initialCapability);
   ASSERT(m->stat != NoStatus);
 #endif
 
   stat = m->stat;
 
-#ifdef SMP
-  pthread_cond_destroy(&m->wakeup);
-#endif
-
-  IF_DEBUG(scheduler, fprintf(stderr, "scheduler: main thread (%d) finished\n", 
-                             m->tso->id));
-  free(m);
-
-  RELEASE_LOCK(&sched_mutex);
-
-  return stat;
-}
-
-//@node Run queue code, Garbage Collextion Routines, Suspend and Resume, Main scheduling code
-//@subsection Run queue code 
-
-#if 0
-/* 
-   NB: In GranSim we have many run queues; run_queue_hd is actually a macro
-       unfolding to run_queue_hds[CurrentProc], thus CurrentProc is an
-       implicit global variable that has to be correct when calling these
-       fcts -- HWL 
-*/
-
-/* Put the new thread on the head of the runnable queue.
- * The caller of createThread better push an appropriate closure
- * on this thread's stack before the scheduler is invoked.
- */
-static /* inline */ void
-add_to_run_queue(tso)
-StgTSO* tso; 
-{
-  ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
-  tso->link = run_queue_hd;
-  run_queue_hd = tso;
-  if (run_queue_tl == END_TSO_QUEUE) {
-    run_queue_tl = tso;
-  }
-}
-
-/* Put the new thread at the end of the runnable queue. */
-static /* inline */ void
-push_on_run_queue(tso)
-StgTSO* tso; 
-{
-  ASSERT(get_itbl((StgClosure *)tso)->type == TSO);
-  ASSERT(run_queue_hd!=NULL && run_queue_tl!=NULL);
-  ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
-  if (run_queue_hd == END_TSO_QUEUE) {
-    run_queue_hd = tso;
+#if defined(RTS_SUPPORTS_THREADS)
+  // Free the condition variable, returning it to the cache if possible.
+  if (!bound_cond_cache_full) {
+      bound_cond_cache = m->bound_thread_cond;
+      bound_cond_cache_full = 1;
   } else {
-    run_queue_tl->link = tso;
+      closeCondition(&m->bound_thread_cond);
   }
-  run_queue_tl = tso;
-}
-
-/* 
-   Should be inlined because it's used very often in schedule.  The tso
-   argument is actually only needed in GranSim, where we want to have the
-   possibility to schedule *any* TSO on the run queue, irrespective of the
-   actual ordering. Therefore, if tso is not the nil TSO then we traverse
-   the run queue and dequeue the tso, adjusting the links in the queue. 
-*/
-//@cindex take_off_run_queue
-static /* inline */ StgTSO*
-take_off_run_queue(StgTSO *tso) {
-  StgTSO *t, *prev;
+#endif
 
-  /* 
-     qetlaHbogh Qu' ngaSbogh ghomDaQ {tso} yIteq!
+  IF_DEBUG(scheduler, sched_belch("main thread (%d) finished", m->tso->id));
+  stgFree(m);
 
-     if tso is specified, unlink that tso from the run_queue (doesn't have
-     to be at the beginning of the queue); GranSim only 
-  */
-  if (tso!=END_TSO_QUEUE) {
-    /* find tso in queue */
-    for (t=run_queue_hd, prev=END_TSO_QUEUE; 
-        t!=END_TSO_QUEUE && t!=tso;
-        prev=t, t=t->link) 
-      /* nothing */ ;
-    ASSERT(t==tso);
-    /* now actually dequeue the tso */
-    if (prev!=END_TSO_QUEUE) {
-      ASSERT(run_queue_hd!=t);
-      prev->link = t->link;
-    } else {
-      /* t is at beginning of thread queue */
-      ASSERT(run_queue_hd==t);
-      run_queue_hd = t->link;
-    }
-    /* t is at end of thread queue */
-    if (t->link==END_TSO_QUEUE) {
-      ASSERT(t==run_queue_tl);
-      run_queue_tl = prev;
-    } else {
-      ASSERT(run_queue_tl!=t);
-    }
-    t->link = END_TSO_QUEUE;
-  } else {
-    /* take tso from the beginning of the queue; std concurrent code */
-    t = run_queue_hd;
-    if (t != END_TSO_QUEUE) {
-      run_queue_hd = t->link;
-      t->link = END_TSO_QUEUE;
-      if (run_queue_hd == END_TSO_QUEUE) {
-       run_queue_tl = END_TSO_QUEUE;
-      }
-    }
-  }
-  return t;
+  // Postcondition: sched_mutex still held
+  return stat;
 }
 
-#endif /* 0 */
-
-//@node Garbage Collextion Routines, Blocking Queue Routines, Run queue code, Main scheduling code
-//@subsection Garbage Collextion Routines
-
 /* ---------------------------------------------------------------------------
    Where are the roots that we know about?
 
         - all the threads on the runnable queue
         - all the threads on the blocked queue
+        - all the threads on the sleeping queue
        - all the thread currently executing a _ccall_GC
         - all the "main threads"
      
@@ -1526,46 +2180,58 @@ take_off_run_queue(StgTSO *tso) {
        KH @ 25/10/99
 */
 
-static void GetRoots(void)
+void
+GetRoots( evac_fn evac )
 {
-  StgMainThread *m;
-
 #if defined(GRAN)
   {
     nat i;
     for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
       if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
-       run_queue_hds[i]    = (StgTSO *)MarkRoot((StgClosure *)run_queue_hds[i]);
+         evac((StgClosure **)&run_queue_hds[i]);
       if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
-       run_queue_tls[i]    = (StgTSO *)MarkRoot((StgClosure *)run_queue_tls[i]);
+         evac((StgClosure **)&run_queue_tls[i]);
       
       if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
-       blocked_queue_hds[i] = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hds[i]);
+         evac((StgClosure **)&blocked_queue_hds[i]);
       if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
-       blocked_queue_tls[i] = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tls[i]);
+         evac((StgClosure **)&blocked_queue_tls[i]);
       if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
-       ccalling_threadss[i] = (StgTSO *)MarkRoot((StgClosure *)ccalling_threadss[i]);
+         evac((StgClosure **)&ccalling_threads[i]);
     }
   }
 
   markEventQueue();
 
 #else /* !GRAN */
-  run_queue_hd      = (StgTSO *)MarkRoot((StgClosure *)run_queue_hd);
-  run_queue_tl      = (StgTSO *)MarkRoot((StgClosure *)run_queue_tl);
-
-  blocked_queue_hd  = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hd);
-  blocked_queue_tl  = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tl);
+  if (run_queue_hd != END_TSO_QUEUE) {
+      ASSERT(run_queue_tl != END_TSO_QUEUE);
+      evac((StgClosure **)&run_queue_hd);
+      evac((StgClosure **)&run_queue_tl);
+  }
+  
+  if (blocked_queue_hd != END_TSO_QUEUE) {
+      ASSERT(blocked_queue_tl != END_TSO_QUEUE);
+      evac((StgClosure **)&blocked_queue_hd);
+      evac((StgClosure **)&blocked_queue_tl);
+  }
+  
+  if (sleeping_queue != END_TSO_QUEUE) {
+      evac((StgClosure **)&sleeping_queue);
+  }
 #endif 
 
-  for (m = main_threads; m != NULL; m = m->link) {
-    m->tso = (StgTSO *)MarkRoot((StgClosure *)m->tso);
+  if (suspended_ccalling_threads != END_TSO_QUEUE) {
+      evac((StgClosure **)&suspended_ccalling_threads);
   }
-  suspended_ccalling_threads = 
-    (StgTSO *)MarkRoot((StgClosure *)suspended_ccalling_threads);
 
-#if defined(SMP) || defined(PAR) || defined(GRAN)
-  markSparkQueue();
+#if defined(PAR) || defined(GRAN)
+  markSparkQueue(evac);
+#endif
+
+#if defined(RTS_USER_SIGNALS)
+  // mark the signal handlers (signals should be already blocked)
+  markSignalHandlers(evac);
 #endif
 }
 
@@ -1582,27 +2248,39 @@ static void GetRoots(void)
    This needs to be protected by the GC condition variable above.  KH.
    -------------------------------------------------------------------------- */
 
-void (*extra_roots)(void);
+static void (*extra_roots)(evac_fn);
 
 void
 performGC(void)
 {
-  GarbageCollect(GetRoots);
+  /* Obligated to hold this lock upon entry */
+  ACQUIRE_LOCK(&sched_mutex);
+  GarbageCollect(GetRoots,rtsFalse);
+  RELEASE_LOCK(&sched_mutex);
+}
+
+void
+performMajorGC(void)
+{
+  ACQUIRE_LOCK(&sched_mutex);
+  GarbageCollect(GetRoots,rtsTrue);
+  RELEASE_LOCK(&sched_mutex);
 }
 
 static void
-AllRoots(void)
+AllRoots(evac_fn evac)
 {
-  GetRoots();                  /* the scheduler's roots */
-  extra_roots();               /* the user's roots */
+    GetRoots(evac);            // the scheduler's roots
+    extra_roots(evac);         // the user's roots
 }
 
 void
-performGCWithRoots(void (*get_roots)(void))
+performGCWithRoots(void (*get_roots)(evac_fn))
 {
+  ACQUIRE_LOCK(&sched_mutex);
   extra_roots = get_roots;
-
-  GarbageCollect(AllRoots);
+  GarbageCollect(AllRoots,rtsFalse);
+  RELEASE_LOCK(&sched_mutex);
 }
 
 /* -----------------------------------------------------------------------------
@@ -1617,24 +2295,22 @@ performGCWithRoots(void (*get_roots)(void))
 static StgTSO *
 threadStackOverflow(StgTSO *tso)
 {
-  nat new_stack_size, new_tso_size, diff, stack_words;
+  nat new_stack_size, new_tso_size, stack_words;
   StgPtr new_sp;
   StgTSO *dest;
 
   IF_DEBUG(sanity,checkTSO(tso));
   if (tso->stack_size >= tso->max_stack_size) {
-#if 0
-    /* 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));
-#endif
-#ifdef INTERPRETER
-    fprintf(stderr, "fatal: stack overflow in Hugs; aborting\n" );
-    exit(1);
-#else
+
+    IF_DEBUG(gc,
+            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)));
+
     /* Send this thread the StackOverflow exception */
     raiseAsync(tso, (StgClosure *)stackOverflow_closure);
-#endif
     return tso;
   }
 
@@ -1648,10 +2324,10 @@ threadStackOverflow(StgTSO *tso)
   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,"scheduler: 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_tso_size-sizeofW(StgTSO),0);
+  TICK_ALLOC_TSO(new_stack_size,0);
 
   /* copy the TSO block and the old stack into the new area */
   memcpy(dest,tso,TSO_STRUCT_SIZE);
@@ -1660,29 +2336,29 @@ threadStackOverflow(StgTSO *tso)
   memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
 
   /* relocate the stack pointers... */
-  diff = (P_)new_sp - (P_)tso->sp; /* In *words* */
-  dest->su    = (StgUpdateFrame *) ((P_)dest->su + diff);
-  dest->sp    = new_sp;
-  dest->splim = (P_)dest->splim + (nat)((P_)dest - (P_)tso);
+  dest->sp         = new_sp;
   dest->stack_size = new_stack_size;
        
-  /* and relocate the update frame list */
-  relocate_TSO(tso, dest);
-
   /* Mark the old TSO as relocated.  We have to check for relocated
    * TSOs in the garbage collector and any primops that deal with TSOs.
    *
-   * It's important to set the sp and su values to just beyond the end
+   * It's important to set the sp value to just beyond the end
    * of the stack, so we don't attempt to scavenge any part of the
    * dead TSO's stack.
    */
   tso->what_next = ThreadRelocated;
   tso->link = dest;
   tso->sp = (P_)&(tso->stack[tso->stack_size]);
-  tso->su = (StgUpdateFrame *)tso->sp;
   tso->why_blocked = NotBlocked;
   dest->mut_link = NULL;
 
+  IF_PAR_DEBUG(verbose,
+              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, 
+                                               tso->sp+64)));
+  
   IF_DEBUG(sanity,checkTSO(tso));
 #if 0
   IF_DEBUG(scheduler,printTSO(dest));
@@ -1691,22 +2367,17 @@ threadStackOverflow(StgTSO *tso)
   return dest;
 }
 
-//@node Blocking Queue Routines, Exception Handling Routines, Garbage Collextion Routines, Main scheduling code
-//@subsection Blocking Queue Routines
-
 /* ---------------------------------------------------------------------------
    Wake up a queue that was blocked on some resource.
    ------------------------------------------------------------------------ */
 
-/* ToDo: check push_on_run_queue vs. PUSH_ON_RUN_QUEUE */
-
 #if defined(GRAN)
-static inline void
+STATIC_INLINE void
 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
 {
 }
 #elif defined(PAR)
-static inline void
+STATIC_INLINE void
 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
 {
   /* write RESUME events to log file and
@@ -1715,6 +2386,8 @@ unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
     DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC, 
                     GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
                     0, 0 /* spark_queue_len(ADVISORY_POOL) */);
+    if (EMPTY_RUN_QUEUE())
+      emitSchedule = rtsTrue;
 
     switch (get_itbl(node)->type) {
        case FETCH_ME_BQ:
@@ -1725,6 +2398,10 @@ unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
        case BLACKHOLE_BQ:
          ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
          break;
+#ifdef DIST
+        case MVAR:
+          break;
+#endif   
        default:
          barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
        }
@@ -1736,63 +2413,41 @@ unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
 static StgBlockingQueueElement *
 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
 {
-    StgBlockingQueueElement *next;
+    StgTSO *tso;
     PEs node_loc, tso_loc;
 
     node_loc = where_is(node); // should be lifted out of loop
     tso = (StgTSO *)bqe;  // wastes an assignment to get the type right
-    tso_loc = where_is(tso);
+    tso_loc = where_is((StgClosure *)tso);
     if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
       /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
       ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
-      bq_processing_time += RtsFlags.GranFlags.Costs.lunblocktime;
+      CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
       // insertThread(tso, node_loc);
-      new_event(tso_loc, tso_loc,
-               CurrentTime[CurrentProc]+bq_processing_time,
+      new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
                ResumeThread,
                tso, node, (rtsSpark*)NULL);
       tso->link = END_TSO_QUEUE; // overwrite link just to be sure 
       // len_local++;
       // len++;
     } else { // TSO is remote (actually should be FMBQ)
-      bq_processing_time += RtsFlags.GranFlags.Costs.mpacktime;
-      bq_processing_time += RtsFlags.GranFlags.Costs.gunblocktime;
-      new_event(tso_loc, CurrentProc, 
-               CurrentTime[CurrentProc]+bq_processing_time+
-               RtsFlags.GranFlags.Costs.latency,
+      CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
+                                  RtsFlags.GranFlags.Costs.gunblocktime +
+                                 RtsFlags.GranFlags.Costs.latency;
+      new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
                UnblockThread,
                tso, node, (rtsSpark*)NULL);
       tso->link = END_TSO_QUEUE; // overwrite link just to be sure 
-      bq_processing_time += RtsFlags.GranFlags.Costs.mtidytime;
       // len++;
-    }      
+    }
     /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
     IF_GRAN_DEBUG(bq,
-                 fprintf(stderr," %s TSO %d (%p) [PE %d] (blocked_on=%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->blocked_on, tso->link))
-    tso->blocked_on = NULL;
-    IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)", 
+                         tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
+    tso->block_info.closure = NULL;
+    IF_DEBUG(scheduler,debugBelch("-- Waking up thread %ld (%p)\n", 
                             tso->id, tso));
-  }
-
-  /* if this is the BQ of an RBH, we have to put back the info ripped out of
-     the closure to make room for the anchor of the BQ */
-  if (next!=END_BQ_QUEUE) {
-    ASSERT(get_itbl(node)->type == RBH && get_itbl(next)->type == CONSTR);
-    /*
-    ASSERT((info_ptr==&RBH_Save_0_info) ||
-          (info_ptr==&RBH_Save_1_info) ||
-          (info_ptr==&RBH_Save_2_info));
-    */
-    /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
-    ((StgRBH *)node)->blocking_queue = ((StgRBHSave *)next)->payload[0];
-    ((StgRBH *)node)->mut_link       = ((StgRBHSave *)next)->payload[1];
-
-    IF_GRAN_DEBUG(bq,
-                 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
-                       node, info_type(node)));
-  }
 }
 #elif defined(PAR)
 static StgBlockingQueueElement *
@@ -1805,9 +2460,9 @@ 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;
@@ -1816,8 +2471,8 @@ unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
     case BLOCKED_FETCH:
       /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
       next = bqe->link;
-      bqe->link = PendingFetches;
-      PendingFetches = bqe;
+      bqe->link = (StgBlockingQueueElement *)PendingFetches;
+      PendingFetches = (StgBlockedFetch *)bqe;
       break;
 
 # if defined(DEBUG)
@@ -1825,9 +2480,9 @@ unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
         see comments on RBHSave closures above */
     case CONSTR:
       /* check that the closure is an RBHSave closure */
-      ASSERT(get_itbl((StgClosure *)bqe) == &RBH_Save_0_info ||
-            get_itbl((StgClosure *)bqe) == &RBH_Save_1_info ||
-            get_itbl((StgClosure *)bqe) == &RBH_Save_2_info);
+      ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
+            get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
+            get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
       break;
 
     default:
@@ -1836,7 +2491,7 @@ unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
           (StgClosure *)bqe);
 # endif
     }
-  // IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
+  IF_PAR_DEBUG(bq, debugBelch(", %p (%s)\n", bqe, info_type((StgClosure*)bqe)));
   return next;
 }
 
@@ -1850,24 +2505,25 @@ 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(PAR) || defined(GRAN)
-inline StgTSO *
-unblockOne(StgTSO *tso, StgClosure *node)
+#if defined(GRAN) || defined(PAR)
+INLINE_ME StgBlockingQueueElement *
+unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
 {
   ACQUIRE_LOCK(&sched_mutex);
-  tso = unblockOneLocked(tso, node);
+  bqe = unblockOneLocked(bqe, node);
   RELEASE_LOCK(&sched_mutex);
-  return tso;
+  return bqe;
 }
 #else
-inline StgTSO *
+INLINE_ME StgTSO *
 unblockOne(StgTSO *tso)
 {
   ACQUIRE_LOCK(&sched_mutex);
@@ -1881,20 +2537,19 @@ unblockOne(StgTSO *tso)
 void 
 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
 {
-  StgBlockingQueueElement *bqe, *next;
-  StgTSO *tso;
-  PEs node_loc, tso_loc;
-  rtsTime bq_processing_time = 0;
-  nat len = 0, len_local = 0;
+  StgBlockingQueueElement *bqe;
+  PEs node_loc;
+  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));
 
   node_loc = where_is(node);
 
-  ASSERT(get_itbl(q)->type == TSO ||   // q is either a TSO or an RBHSave
+  ASSERT(q == END_BQ_QUEUE ||
+        get_itbl(q)->type == TSO ||   // q is either a TSO or an RBHSave
         get_itbl(q)->type == CONSTR); // closure (type constructor)
   ASSERT(is_unique(node));
 
@@ -1904,13 +2559,13 @@ 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++;
@@ -1927,91 +2582,279 @@ awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
     */
     //tso = (StgTSO *)bqe;  // wastes an assignment to get the type right
     //tso_loc = where_is(tso);
+    len++;
     bqe = unblockOneLocked(bqe, node);
   }
 
+  /* if this is the BQ of an RBH, we have to put back the info ripped out of
+     the closure to make room for the anchor of the BQ */
+  if (bqe!=END_BQ_QUEUE) {
+    ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
+    /*
+    ASSERT((info_ptr==&RBH_Save_0_info) ||
+          (info_ptr==&RBH_Save_1_info) ||
+          (info_ptr==&RBH_Save_2_info));
+    */
+    /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
+    ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
+    ((StgRBH *)node)->mut_link       = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
+
+    IF_GRAN_DEBUG(bq,
+                 debugBelch("## Filled in RBH_Save for %p (%s) at end of AwBQ\n",
+                       node, info_type(node)));
+  }
+
   /* statistics gathering */
-  /* ToDo: fix counters
   if (RtsFlags.GranFlags.GranSimStats.Global) {
-    globalGranStats.tot_bq_processing_time += bq_processing_time;
+    // globalGranStats.tot_bq_processing_time += bq_processing_time;
     globalGranStats.tot_bq_len += len;      // total length of all bqs awakened
-    globalGranStats.tot_bq_len_local += len_local;  // same for local TSOs only
+    // globalGranStats.tot_bq_len_local += len_local;  // same for local TSOs only
     globalGranStats.tot_awbq++;             // total no. of bqs awakened
   }
   IF_GRAN_DEBUG(bq,
-               fprintf(stderr,"## BQ Stats of %p: [%d entries, %d local] %s\n",
-                       node, len, len_local, (next!=END_TSO_QUEUE) ? "RBH" : ""));
-  */
+               debugBelch("## BQ Stats of %p: [%d entries] %s\n",
+                       node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
 }
 #elif defined(PAR)
 void 
 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
 {
-  StgBlockingQueueElement *bqe, *next;
+  StgBlockingQueueElement *bqe;
 
   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));
-
-  ASSERT(get_itbl(q)->type == TSO ||           
+#ifdef DIST  
+  //RFP
+  if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
+    IF_PAR_DEBUG(verbose, debugBelch("## ... nothing to unblock so lets just return. RFP (BUG?)\n"));
+    return;
+  }
+#endif
+  
+  ASSERT(q == END_BQ_QUEUE ||
+        get_itbl(q)->type == TSO ||           
         get_itbl(q)->type == BLOCKED_FETCH || 
         get_itbl(q)->type == CONSTR); 
 
-  bqe = q;
-  while (get_itbl(bqe)->type==TSO || 
-        get_itbl(bqe)->type==BLOCKED_FETCH) {
-    bqe = unblockOneLocked(bqe, node);
-  }
-  RELEASE_LOCK(&sched_mutex);
-}
+  bqe = q;
+  while (get_itbl(bqe)->type==TSO || 
+        get_itbl(bqe)->type==BLOCKED_FETCH) {
+    bqe = unblockOneLocked(bqe, node);
+  }
+  RELEASE_LOCK(&sched_mutex);
+}
+
+#else   /* !GRAN && !PAR */
+
+void
+awakenBlockedQueueNoLock(StgTSO *tso)
+{
+  while (tso != END_TSO_QUEUE) {
+    tso = unblockOneLocked(tso);
+  }
+}
+
+void
+awakenBlockedQueue(StgTSO *tso)
+{
+  ACQUIRE_LOCK(&sched_mutex);
+  while (tso != END_TSO_QUEUE) {
+    tso = unblockOneLocked(tso);
+  }
+  RELEASE_LOCK(&sched_mutex);
+}
+#endif
+
+/* ---------------------------------------------------------------------------
+   Interrupt execution
+   - usually called inside a signal handler so it mustn't do anything fancy.   
+   ------------------------------------------------------------------------ */
+
+void
+interruptStgRts(void)
+{
+    interrupted    = 1;
+    context_switch = 1;
+}
+
+/* -----------------------------------------------------------------------------
+   Unblock a thread
+
+   This is for use when we raise an exception in another thread, which
+   may be blocked.
+   This has nothing to do with the UnblockThread event in GranSim. -- HWL
+   -------------------------------------------------------------------------- */
+
+#if defined(GRAN) || defined(PAR)
+/*
+  NB: only the type of the blocking queue is different in GranSim and GUM
+      the operations on the queue-elements are the same
+      long live polymorphism!
+
+  Locks: sched_mutex is held upon entry and exit.
+
+*/
+static void
+unblockThread(StgTSO *tso)
+{
+  StgBlockingQueueElement *t, **last;
+
+  switch (tso->why_blocked) {
+
+  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);
+    {
+      StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
+      StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
+
+      last = (StgBlockingQueueElement **)&mvar->head;
+      for (t = (StgBlockingQueueElement *)mvar->head; 
+          t != END_BQ_QUEUE; 
+          last = &t->link, last_tso = t, t = t->link) {
+       if (t == (StgBlockingQueueElement *)tso) {
+         *last = (StgBlockingQueueElement *)tso->link;
+         if (mvar->tail == tso) {
+           mvar->tail = (StgTSO *)last_tso;
+         }
+         goto done;
+       }
+      }
+      barf("unblockThread (MVAR): TSO not found");
+    }
+
+  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_BQ_QUEUE; 
+          last = &t->link, t = t->link) {
+       if (t == (StgBlockingQueueElement *)tso) {
+         *last = (StgBlockingQueueElement *)tso->link;
+         goto done;
+       }
+      }
+      barf("unblockThread (BLACKHOLE): TSO not found");
+    }
+
+  case BlockedOnException:
+    {
+      StgTSO *target  = tso->block_info.tso;
+
+      ASSERT(get_itbl(target)->type == TSO);
 
-#else   /* !GRAN && !PAR */
-void
-awakenBlockedQueue(StgTSO *tso)
-{
-  ACQUIRE_LOCK(&sched_mutex);
-  while (tso != END_TSO_QUEUE) {
-    tso = unblockOneLocked(tso);
-  }
-  RELEASE_LOCK(&sched_mutex);
-}
-#endif
+      if (target->what_next == ThreadRelocated) {
+         target = target->link;
+         ASSERT(get_itbl(target)->type == TSO);
+      }
 
-//@node Exception Handling Routines, Debugging Routines, Blocking Queue Routines, Main scheduling code
-//@subsection Exception Handling Routines
+      ASSERT(target->blocked_exceptions != NULL);
 
-/* ---------------------------------------------------------------------------
-   Interrupt execution
-   - usually called inside a signal handler so it mustn't do anything fancy.   
-   ------------------------------------------------------------------------ */
+      last = (StgBlockingQueueElement **)&target->blocked_exceptions;
+      for (t = (StgBlockingQueueElement *)target->blocked_exceptions; 
+          t != END_BQ_QUEUE; 
+          last = &t->link, t = t->link) {
+       ASSERT(get_itbl(t)->type == TSO);
+       if (t == (StgBlockingQueueElement *)tso) {
+         *last = (StgBlockingQueueElement *)tso->link;
+         goto done;
+       }
+      }
+      barf("unblockThread (Exception): TSO not found");
+    }
 
-void
-interruptStgRts(void)
-{
-    interrupted    = 1;
-    context_switch = 1;
-}
+  case BlockedOnRead:
+  case BlockedOnWrite:
+#if defined(mingw32_HOST_OS)
+  case BlockedOnDoProc:
+#endif
+    {
+      /* take TSO off blocked_queue */
+      StgBlockingQueueElement *prev = NULL;
+      for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE; 
+          prev = t, t = t->link) {
+       if (t == (StgBlockingQueueElement *)tso) {
+         if (prev == NULL) {
+           blocked_queue_hd = (StgTSO *)t->link;
+           if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
+             blocked_queue_tl = END_TSO_QUEUE;
+           }
+         } else {
+           prev->link = t->link;
+           if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
+             blocked_queue_tl = (StgTSO *)prev;
+           }
+         }
+         goto done;
+       }
+      }
+      barf("unblockThread (I/O): TSO not found");
+    }
 
-/* -----------------------------------------------------------------------------
-   Unblock a thread
+  case BlockedOnDelay:
+    {
+      /* take TSO off sleeping_queue */
+      StgBlockingQueueElement *prev = NULL;
+      for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE; 
+          prev = t, t = t->link) {
+       if (t == (StgBlockingQueueElement *)tso) {
+         if (prev == NULL) {
+           sleeping_queue = (StgTSO *)t->link;
+         } else {
+           prev->link = t->link;
+         }
+         goto done;
+       }
+      }
+      barf("unblockThread (delay): TSO not found");
+    }
 
-   This is for use when we raise an exception in another thread, which
-   may be blocked.
-   This has nothing to do with the UnblockThread event in GranSim. -- HWL
-   -------------------------------------------------------------------------- */
+  default:
+    barf("unblockThread");
+  }
 
+ done:
+  tso->link = END_TSO_QUEUE;
+  tso->why_blocked = NotBlocked;
+  tso->block_info.closure = NULL;
+  PUSH_ON_RUN_QUEUE(tso);
+}
+#else
 static void
 unblockThread(StgTSO *tso)
 {
   StgTSO *t, **last;
+  
+  /* To avoid locking unnecessarily. */
+  if (tso->why_blocked == NotBlocked) {
+    return;
+  }
 
-  ACQUIRE_LOCK(&sched_mutex);
   switch (tso->why_blocked) {
 
-  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);
@@ -2054,6 +2897,12 @@ unblockThread(StgTSO *tso)
       StgTSO *target  = tso->block_info.tso;
 
       ASSERT(get_itbl(target)->type == TSO);
+
+      while (target->what_next == ThreadRelocated) {
+         target = target->link;
+         ASSERT(get_itbl(target)->type == TSO);
+      }
+      
       ASSERT(target->blocked_exceptions != NULL);
 
       last = &target->blocked_exceptions;
@@ -2068,9 +2917,11 @@ unblockThread(StgTSO *tso)
       barf("unblockThread (Exception): TSO not found");
     }
 
-  case BlockedOnDelay:
   case BlockedOnRead:
   case BlockedOnWrite:
+#if defined(mingw32_HOST_OS)
+  case BlockedOnDoProc:
+#endif
     {
       StgTSO *prev = NULL;
       for (t = blocked_queue_hd; t != END_TSO_QUEUE; 
@@ -2093,6 +2944,23 @@ unblockThread(StgTSO *tso)
       barf("unblockThread (I/O): TSO not found");
     }
 
+  case BlockedOnDelay:
+    {
+      StgTSO *prev = NULL;
+      for (t = sleeping_queue; t != END_TSO_QUEUE; 
+          prev = t, t = t->link) {
+       if (t == tso) {
+         if (prev == NULL) {
+           sleeping_queue = t->link;
+         } else {
+           prev->link = t->link;
+         }
+         goto done;
+       }
+      }
+      barf("unblockThread (delay): TSO not found");
+    }
+
   default:
     barf("unblockThread");
   }
@@ -2101,9 +2969,9 @@ unblockThread(StgTSO *tso)
   tso->link = END_TSO_QUEUE;
   tso->why_blocked = NotBlocked;
   tso->block_info.closure = NULL;
-  PUSH_ON_RUN_QUEUE(tso);
-  RELEASE_LOCK(&sched_mutex);
+  APPEND_TO_RUN_QUEUE(tso);
 }
+#endif
 
 /* -----------------------------------------------------------------------------
  * raiseAsync()
@@ -2121,12 +2989,12 @@ unblockThread(StgTSO *tso)
  * the top of the stack.
  * 
  * How exactly do we save all the active computations?  We create an
- * AP_UPD for every UpdateFrame on the stack.  Entering one of these
- * AP_UPDs pushes everything from the corresponding update frame
+ * AP_STACK for every UpdateFrame on the stack.  Entering one of these
+ * AP_STACKs pushes everything from the corresponding update frame
  * upwards onto the stack.  (Actually, it pushes everything up to the
- * next update frame plus a pointer to the next AP_UPD object.
- * Entering the next AP_UPD object pushes more onto the stack until we
- * reach the last AP_UPD object - at which point the stack should look
+ * next update frame plus a pointer to the next AP_STACK object.
+ * Entering the next AP_STACK object pushes more onto the stack until we
+ * reach the last AP_STACK object - at which point the stack should look
  * exactly as it did when we killed the TSO and we can continue
  * execution by entering the closure on top of the stack.
  *
@@ -2135,204 +3003,388 @@ unblockThread(StgTSO *tso)
  * CATCH_FRAME on the stack.  In either case, we strip the entire
  * stack and replace the thread with a zombie.
  *
+ * Locks: sched_mutex held upon entry nor exit.
+ *
  * -------------------------------------------------------------------------- */
  
 void 
 deleteThread(StgTSO *tso)
 {
-  raiseAsync(tso,NULL);
+  if (tso->why_blocked != BlockedOnCCall &&
+      tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
+      raiseAsync(tso,NULL);
+  }
 }
 
-void
-raiseAsync(StgTSO *tso, StgClosure *exception)
-{
-  StgUpdateFrame* su = tso->su;
-  StgPtr          sp = tso->sp;
-  
-  /* Thread already dead? */
+#ifdef FORKPROCESS_PRIMOP_SUPPORTED
+static void 
+deleteThreadImmediately(StgTSO *tso)
+{ // for forkProcess only:
+  // delete thread without giving it a chance to catch the KillThread exception
+
   if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
-    return;
+      return;
   }
 
-  IF_DEBUG(scheduler, sched_belch("raising exception in thread %ld.", tso->id));
+  if (tso->why_blocked != BlockedOnCCall &&
+      tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
+    unblockThread(tso);
+  }
 
-  /* Remove it from any blocking queues */
-  unblockThread(tso);
+  tso->what_next = ThreadKilled;
+}
+#endif
 
-  /* The stack freezing code assumes there's a closure pointer on
-   * the top of the stack.  This isn't always the case with compiled
-   * code, so we have to push a dummy closure on the top which just
-   * returns to the next return address on the stack.
-   */
-  if ( LOOKS_LIKE_GHC_INFO((void*)*sp) ) {
-    *(--sp) = (W_)&dummy_ret_closure;
-  }
+void
+raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
+{
+  /* When raising async exs from contexts where sched_mutex isn't held;
+     use raiseAsyncWithLock(). */
+  ACQUIRE_LOCK(&sched_mutex);
+  raiseAsync(tso,exception);
+  RELEASE_LOCK(&sched_mutex);
+}
 
-  while (1) {
-    int words = ((P_)su - (P_)sp) - 1;
-    nat i;
-    StgAP_UPD * ap;
+void
+raiseAsync(StgTSO *tso, StgClosure *exception)
+{
+    raiseAsync_(tso, exception, rtsFalse);
+}
 
-    /* If we find a CATCH_FRAME, and we've got an exception to raise,
-     * then build PAP(handler,exception,realworld#), and leave it on
-     * top of the stack ready to enter.
-     */
-    if (get_itbl(su)->type == CATCH_FRAME && exception != NULL) {
-      StgCatchFrame *cf = (StgCatchFrame *)su;
-      /* we've got an exception to raise, so let's pass it to the
-       * handler in this frame.
-       */
-      ap = (StgAP_UPD *)allocate(sizeofW(StgPAP) + 2);
-      TICK_ALLOC_UPD_PAP(3,0);
-      SET_HDR(ap,&PAP_info,cf->header.prof.ccs);
-             
-      ap->n_args = 2;
-      ap->fun = cf->handler;   /* :: Exception -> IO a */
-      ap->payload[0] = (P_)exception;
-      ap->payload[1] = ARG_TAG(0); /* realworld token */
-
-      /* throw away the stack from Sp up to and including the
-       * CATCH_FRAME.
-       */
-      sp = (P_)su + sizeofW(StgCatchFrame) - 1; 
-      tso->su = cf->link;
-
-      /* Restore the blocked/unblocked state for asynchronous exceptions
-       * at the CATCH_FRAME.  
-       *
-       * If exceptions were unblocked at the catch, arrange that they
-       * are unblocked again after executing the handler by pushing an
-       * unblockAsyncExceptions_ret stack frame.
-       */
-      if (!cf->exceptions_blocked) {
-       *(sp--) = (W_)&unblockAsyncExceptionszh_ret_info;
-      }
-      
-      /* Ensure that async exceptions are blocked when running the handler.
-       */
-      if (tso->blocked_exceptions == NULL) {
-       tso->blocked_exceptions = END_TSO_QUEUE;
-      }
-      
-      /* Put the newly-built PAP on top of the stack, ready to execute
-       * when the thread restarts.
-       */
-      sp[0] = (W_)ap;
-      tso->sp = sp;
-      tso->what_next = ThreadEnterGHC;
-      return;
+static void
+raiseAsync_(StgTSO *tso, StgClosure *exception, rtsBool stop_at_atomically)
+{
+    StgRetInfoTable *info;
+    StgPtr sp;
+  
+    // Thread already dead?
+    if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
+       return;
     }
 
-    /* First build an AP_UPD consisting of the stack chunk above the
-     * current update frame, with the top word on the stack as the
-     * fun field.
-     */
-    ap = (StgAP_UPD *)allocate(AP_sizeW(words));
+    IF_DEBUG(scheduler, 
+            sched_belch("raising exception in thread %ld.", (long)tso->id));
     
-    ASSERT(words >= 0);
+    // Remove it from any blocking queues
+    unblockThread(tso);
+
+    sp = tso->sp;
     
-    ap->n_args = words;
-    ap->fun    = (StgClosure *)sp[0];
-    sp++;
-    for(i=0; i < (nat)words; ++i) {
-      ap->payload[i] = (P_)*sp++;
+    // The stack freezing code assumes there's a closure pointer on
+    // the top of the stack, so we have to arrange that this is the case...
+    //
+    if (sp[0] == (W_)&stg_enter_info) {
+       sp++;
+    } else {
+       sp--;
+       sp[0] = (W_)&stg_dummy_ret_closure;
     }
-    
-    switch (get_itbl(su)->type) {
-      
-    case UPDATE_FRAME:
-      {
-       SET_HDR(ap,&AP_UPD_info,su->header.prof.ccs /* ToDo */); 
-       TICK_ALLOC_UP_THK(words+1,0);
+
+    while (1) {
+       nat i;
+
+       // 1. Let the top of the stack be the "current closure"
+       //
+       // 2. Walk up the stack until we find either an UPDATE_FRAME or a
+       // CATCH_FRAME.
+       //
+       // 3. If it's an UPDATE_FRAME, then make an AP_STACK containing the
+       // current closure applied to the chunk of stack up to (but not
+       // including) the update frame.  This closure becomes the "current
+       // closure".  Go back to step 2.
+       //
+       // 4. If it's a CATCH_FRAME, then leave the exception handler on
+       // 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
+       
        
-       IF_DEBUG(scheduler,
-                fprintf(stderr,  "scheduler: Updating ");
-                printPtr((P_)su->updatee); 
-                fprintf(stderr,  " with ");
-                printObj((StgClosure *)ap);
-                );
+       StgPtr frame;
        
-       /* Replace the updatee with an indirection - happily
-        * this will also wake up any threads currently
-        * waiting on the result.
-        */
-       UPD_IND_NOLOCK(su->updatee,ap);  /* revert the black hole */
-       su = su->link;
-       sp += sizeofW(StgUpdateFrame) -1;
-       sp[0] = (W_)ap; /* push onto stack */
-       break;
-      }
-      
-    case CATCH_FRAME:
-      {
-       StgCatchFrame *cf = (StgCatchFrame *)su;
-       StgClosure* o;
+       frame = sp + 1;
+       info = get_ret_itbl((StgClosure *)frame);
        
-       /* We want a PAP, not an AP_UPD.  Fortunately, the
-        * layout's the same.
-        */
-       SET_HDR(ap,&PAP_info,su->header.prof.ccs /* ToDo */);
-       TICK_ALLOC_UPD_PAP(words+1,0);
+       while (info->i.type != UPDATE_FRAME
+              && (info->i.type != CATCH_FRAME || exception == NULL)
+              && 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);
+       }
        
-       /* now build o = FUN(catch,ap,handler) */
-       o = (StgClosure *)allocate(sizeofW(StgClosure)+2);
-       TICK_ALLOC_FUN(2,0);
-       SET_HDR(o,&catch_info,su->header.prof.ccs /* ToDo */);
-       o->payload[0] = (StgClosure *)ap;
-       o->payload[1] = cf->handler;
+       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
+           // top of the CATCH_FRAME ready to enter.
+           //
+       {
+#ifdef PROFILING
+           StgCatchFrame *cf = (StgCatchFrame *)frame;
+#endif
+           StgClosure *raise;
+           
+           // we've got an exception to raise, so let's pass it to the
+           // handler in this frame.
+           //
+           raise = (StgClosure *)allocate(sizeofW(StgClosure)+1);
+           TICK_ALLOC_SE_THK(1,0);
+           SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
+           raise->payload[0] = exception;
+           
+           // throw away the stack from Sp up to the CATCH_FRAME.
+           //
+           sp = frame - 1;
+           
+           /* Ensure that async excpetions are blocked now, so we don't get
+            * a surprise exception before we get around to executing the
+            * handler.
+            */
+           if (tso->blocked_exceptions == NULL) {
+               tso->blocked_exceptions = END_TSO_QUEUE;
+           }
+           
+           /* Put the newly-built THUNK on top of the stack, ready to execute
+            * when the thread restarts.
+            */
+           sp[0] = (W_)raise;
+           sp[-1] = (W_)&stg_enter_info;
+           tso->sp = sp-1;
+           tso->what_next = ThreadRunGHC;
+           IF_DEBUG(sanity, checkTSO(tso));
+           return;
+       }
        
-       IF_DEBUG(scheduler,
-                fprintf(stderr,  "scheduler: Built ");
-                printObj((StgClosure *)o);
-                );
+       case UPDATE_FRAME:
+       {
+           StgAP_STACK * ap;
+           nat words;
+           
+           // First build an AP_STACK consisting of the stack chunk above the
+           // current update frame, with the top word on the stack as the
+           // fun field.
+           //
+           words = frame - sp - 1;
+           ap = (StgAP_STACK *)allocate(PAP_sizeW(words));
+           
+           ap->size = words;
+           ap->fun  = (StgClosure *)sp[0];
+           sp++;
+           for(i=0; i < (nat)words; ++i) {
+               ap->payload[i] = (StgClosure *)*sp++;
+           }
+           
+           SET_HDR(ap,&stg_AP_STACK_info,
+                   ((StgClosure *)frame)->header.prof.ccs /* ToDo */); 
+           TICK_ALLOC_UP_THK(words+1,0);
+           
+           IF_DEBUG(scheduler,
+                    debugBelch("sched: Updating ");
+                    printPtr((P_)((StgUpdateFrame *)frame)->updatee); 
+                    debugBelch(" with ");
+                    printObj((StgClosure *)ap);
+               );
+
+           // Replace the updatee with an indirection - happily
+           // this will also wake up any threads currently
+           // waiting on the result.
+           //
+           // Warning: if we're in a loop, more than one update frame on
+           // the stack may point to the same object.  Be careful not to
+           // overwrite an IND_OLDGEN in this case, because we'll screw
+           // up the mutable lists.  To be on the safe side, don't
+           // overwrite any kind of indirection at all.  See also
+           // threadSqueezeStack in GC.c, where we have to make a similar
+           // check.
+           //
+           if (!closure_IND(((StgUpdateFrame *)frame)->updatee)) {
+               // revert the black hole
+               UPD_IND_NOLOCK(((StgUpdateFrame *)frame)->updatee,
+                              (StgClosure *)ap);
+           }
+           sp += sizeofW(StgUpdateFrame) - 1;
+           sp[0] = (W_)ap; // push onto stack
+           break;
+       }
        
-       /* pop the old handler and put o on the stack */
-       su = cf->link;
-       sp += sizeofW(StgCatchFrame) - 1;
-       sp[0] = (W_)o;
-       break;
-      }
+       case STOP_FRAME:
+           // We've stripped the entire stack, the thread is now dead.
+           sp += sizeofW(StgStopFrame);
+           tso->what_next = ThreadKilled;
+           tso->sp = sp;
+           return;
+           
+       default:
+           barf("raiseAsync");
+       }
+    }
+    barf("raiseAsync");
+}
+
+/* -----------------------------------------------------------------------------
+   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 SEQ_FRAME:
-      {
-       StgSeqFrame *sf = (StgSeqFrame *)su;
-       StgClosure* o;
-       
-       SET_HDR(ap,&PAP_info,su->header.prof.ccs /* ToDo */);
-       TICK_ALLOC_UPD_PAP(words+1,0);
-       
-       /* now build o = FUN(seq,ap) */
-       o = (StgClosure *)allocate(sizeofW(StgClosure)+1);
-       TICK_ALLOC_SE_THK(1,0);
-       SET_HDR(o,&seq_info,su->header.prof.ccs /* ToDo */);
-       o->payload[0] = (StgClosure *)ap;
-       
-       IF_DEBUG(scheduler,
-                fprintf(stderr,  "scheduler: Built ");
-                printObj((StgClosure *)o);
-                );
-       
-       /* pop the old handler and put o on the stack */
-       su = sf->link;
-       sp += sizeofW(StgSeqFrame) - 1;
-       sp[0] = (W_)o;
-       break;
-      }
+    case ATOMICALLY_FRAME:
+      IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p during retrry\n", p));
+      tso->sp = p;
+      return ATOMICALLY_FRAME;
       
-    case STOP_FRAME:
-      /* We've stripped the entire stack, the thread is now dead. */
-      sp += sizeofW(StgStopFrame) - 1;
-      sp[0] = (W_)exception;   /* save the exception */
-      tso->what_next = ThreadKilled;
-      tso->su = (StgUpdateFrame *)(sp+1);
-      tso->sp = sp;
-      return;
+    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:
-      barf("raiseAsync");
+      ASSERT(info->i.type != CATCH_FRAME);
+      ASSERT(info->i.type != STOP_FRAME);
+      p = next; 
+      continue;
     }
   }
-  barf("raiseAsync");
 }
 
 /* -----------------------------------------------------------------------------
@@ -2341,6 +3393,8 @@ raiseAsync(StgTSO *tso, StgClosure *exception)
    up and sent a signal: BlockedOnDeadMVar if the thread was blocked
    on an MVar, or NonTermination if the thread was blocked on a Black
    Hole.
+
+   Locks: sched_mutex isn't held upon entry nor exit.
    -------------------------------------------------------------------------- */
 
 void
@@ -2357,11 +3411,15 @@ resurrectThreads( StgTSO *threads )
     switch (tso->why_blocked) {
     case BlockedOnMVar:
     case BlockedOnException:
+      /* Called by GC - sched_mutex lock is currently held. */
       raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
       break;
     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
@@ -2374,63 +3432,79 @@ resurrectThreads( StgTSO *threads )
   }
 }
 
-//@node Debugging Routines, Index, Exception Handling Routines, Main scheduling code
-//@subsection Debugging Routines
-
-/* -----------------------------------------------------------------------------
-   Debugging: why is a thread blocked
-   -------------------------------------------------------------------------- */
-
-#ifdef DEBUG
+/* ----------------------------------------------------------------------------
+ * 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
 printThreadBlockage(StgTSO *tso)
 {
   switch (tso->why_blocked) {
   case BlockedOnRead:
-    fprintf(stderr,"blocked on read from fd %d", tso->block_info.fd);
+    debugBelch("is blocked on read from fd %d", tso->block_info.fd);
     break;
   case BlockedOnWrite:
-    fprintf(stderr,"blocked on write to fd %d", tso->block_info.fd);
+    debugBelch("is blocked on write to fd %d", tso->block_info.fd);
+    break;
+#if defined(mingw32_HOST_OS)
+    case BlockedOnDoProc:
+    debugBelch("is blocked on proc (request: %d)", tso->block_info.async_result->reqID);
     break;
-  case BlockedOnDelay:
-#if defined(HAVE_SETITIMER)
-    fprintf(stderr,"blocked on delay of %d ms", tso->block_info.delay);
-#else
-    fprintf(stderr,"blocked on delay of %d ms", 
-           tso->block_info.target - getourtimeofday());
 #endif
+  case BlockedOnDelay:
+    debugBelch("is blocked until %d", tso->block_info.target);
     break;
   case BlockedOnMVar:
-    fprintf(stderr,"blocked on an MVar");
+    debugBelch("is blocked on an MVar");
     break;
   case BlockedOnException:
-    fprintf(stderr,"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,"blocked on a black hole");
+    debugBelch("is blocked on a black hole");
     break;
   case NotBlocked:
-    fprintf(stderr,"not blocked");
+    debugBelch("is not blocked");
     break;
 #if defined(PAR)
   case BlockedOnGA:
-    fprintf(stderr,"blocked on global address");
+    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:
+    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
+  case BlockedOnCCall:
+    debugBelch("is blocked on an external call");
+    break;
+  case BlockedOnCCall_NoUnblockExc:
+    debugBelch("is blocked on an external call (exceptions were already blocked)");
+    break;
+  case BlockedOnSTM:
+    debugBelch("is blocked on an STM operation");
+    break;
+  default:
+    barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
+        tso->why_blocked, tso->id, tso);
   }
 }
 
+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);
@@ -2442,18 +3516,40 @@ printAllThreads(void)
 {
   StgTSO *t;
 
-  sched_belch("all threads:");
+# 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!*/);
+
+  debugBelch("all threads at [%s]:\n", time_string);
+# elif defined(PAR)
+  char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
+  ullong_format_string(CURRENT_TIME,
+                      time_string, rtsFalse/*no commas!*/);
+
+  debugBelch("all threads at [%s]:\n", time_string);
+# else
+  debugBelch("all threads:\n");
+# endif
+
   for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
-    fprintf(stderr, "\tthread %d is ", t->id);
+    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");
   }
 }
     
+#ifdef DEBUG
+
 /* 
    Print a whole blocking queue attached to node (debugging only).
 */
-//@cindex print_bq
 # if defined(PAR)
 void 
 print_bq (StgClosure *node)
@@ -2462,63 +3558,76 @@ 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 */
   ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
         get_itbl(node)->type == FETCH_ME_BQ ||
-        get_itbl(node)->type == RBH);
+        get_itbl(node)->type == RBH ||
+        get_itbl(node)->type == MVAR);
     
   ASSERT(node!=(StgClosure*)NULL);         // sanity check
+
+  print_bqe(((StgBlockingQueue*)node)->blocking_queue);
+}
+
+/* 
+   Print a whole blocking queue starting with the element bqe.
+*/
+void 
+print_bqe (StgBlockingQueueElement *bqe)
+{
+  rtsBool end;
+
   /* 
      NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
   */
-  for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
+  for (end = (bqe==END_BQ_QUEUE);
        !end; // iterate until bqe points to a CONSTR
-       end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
-    ASSERT(bqe != END_BQ_QUEUE);             // sanity check
-    ASSERT(bqe != (StgTSO*)NULL);            // sanity check
+       end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), 
+       bqe = end ? END_BQ_QUEUE : bqe->link) {
+    ASSERT(bqe != END_BQ_QUEUE);                               // sanity check
+    ASSERT(bqe != (StgBlockingQueueElement *)NULL);            // sanity check
     /* types of closures that may appear in a blocking queue */
     ASSERT(get_itbl(bqe)->type == TSO ||           
           get_itbl(bqe)->type == BLOCKED_FETCH || 
           get_itbl(bqe)->type == CONSTR); 
     /* only BQs of an RBH end with an RBH_Save closure */
-    ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
+    //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
 
     switch (get_itbl(bqe)->type) {
     case TSO:
-      fprintf(stderr," TSO %d (%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),",
-             (get_itbl(bqe) == &RBH_Save_0_info ? "RBH_Save_0" :
-              get_itbl(bqe) == &RBH_Save_1_info ? "RBH_Save_1" :
-              get_itbl(bqe) == &RBH_Save_2_info ? "RBH_Save_2" :
+      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" :
               "RBH_Save_?"), get_itbl(bqe));
       break;
     default:
-      barf("Unexpected closure type %s in blocking queue of %p (%s)",
-          info_type(bqe), node, info_type(node));
+      barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
+          info_type((StgClosure *)bqe)); // , node, info_type(node));
       break;
     }
   } /* for */
-  fputc('\n', stderr);
+  debugBelch("\n");
 }
 # elif defined(GRAN)
 void 
 print_bq (StgClosure *node)
 {
   StgBlockingQueueElement *bqe;
-  StgTSO *tso;
   PEs node_loc, tso_loc;
   rtsBool end;
 
@@ -2530,7 +3639,7 @@ 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);
 
   /* 
@@ -2540,7 +3649,7 @@ print_bq (StgClosure *node)
        !end; // iterate until bqe points to a CONSTR
        end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
     ASSERT(bqe != END_BQ_QUEUE);             // sanity check
-    ASSERT(bqe != (StgTSO*)NULL);            // sanity check
+    ASSERT(bqe != (StgBlockingQueueElement *)NULL);  // sanity check
     /* types of closures that may appear in a blocking queue */
     ASSERT(get_itbl(bqe)->type == TSO ||           
           get_itbl(bqe)->type == CONSTR); 
@@ -2550,23 +3659,23 @@ print_bq (StgClosure *node)
     tso_loc = where_is((StgClosure *)bqe);
     switch (get_itbl(bqe)->type) {
     case TSO:
-      fprintf(stderr," TSO %d (%x) on [PE %d],",
-             ((StgTSO *)bqe)->id, ((StgTSO *)bqe), tso_loc);
+      debugBelch(" TSO %d (%p) on [PE %d],",
+             ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
       break;
     case CONSTR:
-      fprintf(stderr," %s (IP %p),",
-             (get_itbl(bqe) == &RBH_Save_0_info ? "RBH_Save_0" :
-              get_itbl(bqe) == &RBH_Save_1_info ? "RBH_Save_1" :
-              get_itbl(bqe) == &RBH_Save_2_info ? "RBH_Save_2" :
+      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" :
               "RBH_Save_?"), get_itbl(bqe));
       break;
     default:
       barf("Unexpected closure type %s in blocking queue of %p (%s)",
-          info_type(bqe), node, info_type(node));
+          info_type((StgClosure *)bqe), node, info_type(node));
       break;
     }
   } /* for */
-  fputc('\n', stderr);
+  debugBelch("\n");
 }
 #else
 /* 
@@ -2583,9 +3692,9 @@ print_bq (StgClosure *node)
        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);
+    debugBelch(" TSO %d (%p),", tso->id, tso);
   }
-  fputc('\n', stderr);
+  debugBelch("\n");
 }
 # endif
 
@@ -2605,47 +3714,21 @@ run_queue_len(void)
 }
 #endif
 
-static void
+void
 sched_belch(char *s, ...)
 {
   va_list ap;
   va_start(ap,s);
-#ifdef SMP
-  fprintf(stderr, "scheduler (task %ld): ", pthread_self());
+#ifdef RTS_SUPPORTS_THREADS
+  debugBelch("sched (task %p): ", osThreadId());
+#elif defined(PAR)
+  debugBelch("== ");
 #else
-  fprintf(stderr, "scheduler: ");
+  debugBelch("sched: ");
 #endif
-  vfprintf(stderr, s, ap);
-  fprintf(stderr, "\n");
+  vdebugBelch(s, ap);
+  debugBelch("\n");
+  va_end(ap);
 }
 
 #endif /* DEBUG */
-
-
-//@node Index,  , Debugging Routines, Main scheduling code
-//@subsection Index
-
-//@index
-//* MainRegTable::  @cindex\s-+MainRegTable
-//* StgMainThread::  @cindex\s-+StgMainThread
-//* awaken_blocked_queue::  @cindex\s-+awaken_blocked_queue
-//* blocked_queue_hd::  @cindex\s-+blocked_queue_hd
-//* blocked_queue_tl::  @cindex\s-+blocked_queue_tl
-//* context_switch::  @cindex\s-+context_switch
-//* createThread::  @cindex\s-+createThread
-//* free_capabilities::  @cindex\s-+free_capabilities
-//* gc_pending_cond::  @cindex\s-+gc_pending_cond
-//* initScheduler::  @cindex\s-+initScheduler
-//* interrupted::  @cindex\s-+interrupted
-//* n_free_capabilities::  @cindex\s-+n_free_capabilities
-//* next_thread_id::  @cindex\s-+next_thread_id
-//* print_bq::  @cindex\s-+print_bq
-//* run_queue_hd::  @cindex\s-+run_queue_hd
-//* run_queue_tl::  @cindex\s-+run_queue_tl
-//* sched_mutex::  @cindex\s-+sched_mutex
-//* schedule::  @cindex\s-+schedule
-//* take_off_run_queue::  @cindex\s-+take_off_run_queue
-//* task_ids::  @cindex\s-+task_ids
-//* term_mutex::  @cindex\s-+term_mutex
-//* thread_ready_cond::  @cindex\s-+thread_ready_cond
-//@end index