remove accidental extra in previous patch
[ghc-hetmet.git] / ghc / rts / Schedule.c
index f40e457..bbc6a8b 100644 (file)
 /* ---------------------------------------------------------------------------
- * $Id: Schedule.c,v 1.134 2002/03/12 13:57:11 simonmar Exp $
  *
- * (c) The GHC Team, 1998-2000
+ * (c) The GHC Team, 1998-2005
  *
- * Scheduler
- *
- * 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 distributed 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)
+ * The scheduler and thread-related functionality
  *
  * --------------------------------------------------------------------------*/
 
-//@node Main scheduling code, , ,
-//@section Main scheduling code
-
-/* 
- * Version with scheduler monitor support for SMPs (WAY=s):
-
-   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::  
-//* 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 "OSThreads.h"
 #include "Storage.h"
 #include "StgRun.h"
-#include "StgStartup.h"
 #include "Hooks.h"
 #include "Schedule.h"
 #include "StgMiscClosures.h"
-#include "Storage.h"
 #include "Interpreter.h"
 #include "Exception.h"
 #include "Printer.h"
-#include "Main.h"
-#include "Signals.h"
+#include "RtsSignals.h"
 #include "Sanity.h"
 #include "Stats.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)
+#if defined(GRAN) || defined(PARALLEL_HASKELL)
 # include "GranSimRts.h"
 # include "GranSim.h"
 # include "ParallelRts.h"
 #endif
 #include "Sparks.h"
 #include "Capability.h"
-#include "OSThreads.h"
-#include  "Task.h"
+#include "Task.h"
+#include "AwaitEvent.h"
+#if defined(mingw32_HOST_OS)
+#include "win32/IOManager.h"
+#endif
+
+#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
 
-/* Main thread queue.
- * Locks required: sched_mutex.
- */
-StgMainThread *main_threads;
+// Turn off inlining when debugging - it obfuscates things
+#ifdef DEBUG
+# undef  STATIC_INLINE
+# define STATIC_INLINE static
+#endif
+
+/* -----------------------------------------------------------------------------
+ * Global variables
+ * -------------------------------------------------------------------------- */
 
-/* Thread queues.
- * Locks required: sched_mutex.
- */
 #if defined(GRAN)
 
 StgTSO* ActiveTSO = NULL; /* for assigning system costs; GranSim-Light only */
@@ -149,62 +98,63 @@ StgTSO *ccalling_threadss[MAX_PROC];
 
 #else /* !GRAN */
 
-StgTSO *run_queue_hd, *run_queue_tl;
-StgTSO *blocked_queue_hd, *blocked_queue_tl;
-StgTSO *sleeping_queue;                /* perhaps replace with a hash table? */
+#if !defined(THREADED_RTS)
+// Blocked/sleeping thrads
+StgTSO *blocked_queue_hd = NULL;
+StgTSO *blocked_queue_tl = NULL;
+StgTSO *sleeping_queue = NULL;    // perhaps replace with a hash table?
+#endif
 
+/* Threads blocked on blackholes.
+ * LOCK: sched_mutex+capability, or all capabilities
+ */
+StgTSO *blackhole_queue = NULL;
 #endif
 
+/* The blackhole_queue should be checked for threads to wake up.  See
+ * Schedule.h for more thorough comment.
+ * LOCK: none (doesn't matter if we miss an update)
+ */
+rtsBool blackholes_need_checking = rtsFalse;
+
 /* Linked list of all threads.
  * Used for detecting garbage collected threads.
+ * LOCK: sched_mutex+capability, or all capabilities
  */
-StgTSO *all_threads;
+StgTSO *all_threads = NULL;
 
-/* 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.
+/* flag set by signal handler to precipitate a context switch
+ * LOCK: none (just an advisory flag)
  */
-static StgTSO *suspended_ccalling_threads;
-
-static StgTSO *threadStackOverflow(StgTSO *tso);
-
-/* KH: The following two flags are shared memory locations.  There is no need
-       to lock them, since they are only unset at the end of a scheduler
-       operation.
-*/
+int context_switch = 0;
 
-/* flag set by signal handler to precipitate a context switch */
-//@cindex context_switch
-nat context_switch;
+/* flag that tracks whether we have done any execution in this time slice.
+ * LOCK: currently none, perhaps we should lock (but needs to be
+ * updated in the fast path of the scheduler).
+ */
+nat recent_activity = ACTIVITY_YES;
 
-/* if this flag is set as well, give up execution */
-//@cindex interrupted
-rtsBool interrupted;
+/* if this flag is set as well, give up execution
+ * LOCK: none (changes once, from false->true)
+ */
+rtsBool sched_state = SCHED_RUNNING;
 
 /* Next thread ID to allocate.
- * Locks required: sched_mutex
+ * LOCK: sched_mutex
  */
-//@cindex next_thread_id
-StgThreadID next_thread_id = 1;
+static StgThreadID next_thread_id = 1;
 
-/*
- * Pointers to the state of the current thread.
- * Rule of thumb: if CurrentTSO != NULL, then we're running a Haskell
- * thread.  If CurrentTSO == NULL, then we're at the scheduler level.
- */
 /* 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)
 
 #if defined(GRAN)
 StgTSO *CurrentTSO;
@@ -216,87 +166,125 @@ StgTSO *CurrentTSO;
  */
 StgTSO dummy_tso;
 
-rtsBool ready_to_gc;
-
-void            addToBlockedQueue ( StgTSO *tso );
+/*
+ * 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.
+ */
+rtsBool shutting_down_scheduler = rtsFalse;
 
-static void     schedule          ( void );
-       void     interruptStgRts   ( void );
-#if defined(GRAN)
-static StgTSO * createThread_     ( nat size, rtsBool have_lock, StgInt pri );
-#else
-static StgTSO * createThread_     ( nat size, rtsBool have_lock );
+/*
+ * This mutex protects most of the global scheduler data in
+ * the THREADED_RTS runtime.
+ */
+#if defined(THREADED_RTS)
+Mutex sched_mutex;
 #endif
 
-static void     detectBlackHoles  ( void );
-
-#ifdef DEBUG
-static void sched_belch(char *s, ...);
+#if defined(PARALLEL_HASKELL)
+StgTSO *LastTSO;
+rtsTime TimeOfLastYield;
+rtsBool emitSchedule = rtsTrue;
 #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;
+/* -----------------------------------------------------------------------------
+ * static function prototypes
+ * -------------------------------------------------------------------------- */
 
-# if defined(SMP)
-static Condition gc_pending_cond = INIT_COND_VAR;
-nat await_death;
-# endif
+static Capability *schedule (Capability *initialCapability, Task *task);
 
-#endif /* RTS_SUPPORTS_THREADS */
+//
+// These function all encapsulate parts of the scheduler loop, and are
+// abstracted only to make the structure and control flow of the
+// scheduler clearer.
+//
+static void schedulePreLoop (void);
+#if defined(THREADED_RTS)
+static void schedulePushWork(Capability *cap, Task *task);
+#endif
+static void scheduleStartSignalHandlers (Capability *cap);
+static void scheduleCheckBlockedThreads (Capability *cap);
+static void scheduleCheckBlackHoles (Capability *cap);
+static void scheduleDetectDeadlock (Capability *cap, Task *task);
+#if defined(GRAN)
+static StgTSO *scheduleProcessEvent(rtsEvent *event);
+#endif
+#if defined(PARALLEL_HASKELL)
+static StgTSO *scheduleSendPendingMessages(void);
+static void scheduleActivateSpark(void);
+static rtsBool scheduleGetRemoteWork(rtsBool *receivedFinish);
+#endif
+#if defined(PAR) || defined(GRAN)
+static void scheduleGranParReport(void);
+#endif
+static void schedulePostRunThread(void);
+static rtsBool scheduleHandleHeapOverflow( Capability *cap, StgTSO *t );
+static void scheduleHandleStackOverflow( Capability *cap, Task *task, 
+                                        StgTSO *t);
+static rtsBool scheduleHandleYield( Capability *cap, StgTSO *t, 
+                                   nat prev_what_next );
+static void scheduleHandleThreadBlocked( StgTSO *t );
+static rtsBool scheduleHandleThreadFinished( Capability *cap, Task *task,
+                                            StgTSO *t );
+static rtsBool scheduleDoHeapProfile(rtsBool ready_to_gc);
+static Capability *scheduleDoGC(Capability *cap, Task *task,
+                               rtsBool force_major, 
+                               void (*get_roots)(evac_fn));
+
+static void unblockThread(Capability *cap, StgTSO *tso);
+static rtsBool checkBlackHoles(Capability *cap);
+static void AllRoots(evac_fn evac);
+
+static StgTSO *threadStackOverflow(Capability *cap, StgTSO *tso);
+
+static void raiseAsync_(Capability *cap, StgTSO *tso, StgClosure *exception, 
+                       rtsBool stop_at_atomically, StgPtr stop_here);
+
+static void deleteThread (Capability *cap, StgTSO *tso);
+static void deleteAllThreads (Capability *cap);
 
-#if defined(PAR)
-StgTSO *LastTSO;
-rtsTime TimeOfLastYield;
-rtsBool emitSchedule = rtsTrue;
+#ifdef DEBUG
+static void printThreadBlockage(StgTSO *tso);
+static void printThreadStatus(StgTSO *tso);
+void printThreadQueue(StgTSO *tso);
 #endif
 
-#if DEBUG
-char *whatNext_strs[] = {
-  "ThreadEnterGHC",
+#if defined(PARALLEL_HASKELL)
+StgTSO * createSparkThread(rtsSpark spark);
+StgTSO * activateSpark (rtsSpark spark);  
+#endif
+
+#ifdef DEBUG
+static char *whatNext_strs[] = {
+  "(unknown)",
   "ThreadRunGHC",
-  "ThreadEnterInterp",
+  "ThreadInterpret",
   "ThreadKilled",
+  "ThreadRelocated",
   "ThreadComplete"
 };
-
-char *threadReturnCode_strs[] = {
-  "HeapOverflow",                      /* might also be StackOverflow */
-  "StackOverflow",
-  "ThreadYielding",
-  "ThreadBlocked",
-  "ThreadFinished"
-};
-#endif
-
-#if defined(PAR)
-StgTSO * createSparkThread(rtsSpark spark);
-StgTSO * activateSpark (rtsSpark spark);  
 #endif
 
-/*
- * The thread state for the main thread.
-// ToDo: check whether not needed any more
-StgTSO   *MainTSO;
- */
+/* -----------------------------------------------------------------------------
+ * Putting a thread on the run queue: different scheduling policies
+ * -------------------------------------------------------------------------- */
 
-#if defined(PAR) || defined(RTS_SUPPORTS_THREADS)
-static void taskStart(void);
-static void
-taskStart(void)
+STATIC_INLINE void
+addToRunQueue( Capability *cap, StgTSO *t )
 {
-  schedule();
-}
+#if defined(PARALLEL_HASKELL)
+    if (RtsFlags.ParFlags.doFairScheduling) { 
+       // this does round-robin scheduling; good for concurrency
+       appendToRunQueue(cap,t);
+    } else {
+       // this does unfair scheduling; good for parallelism
+       pushOnRunQueue(cap,t);
+    }
+#else
+    // this does round-robin scheduling; good for concurrency
+    appendToRunQueue(cap,t);
 #endif
-
-
-
-
-//@node Main scheduling loop, Suspend and Resume, Prototypes, Main scheduling code
-//@subsection Main scheduling loop
+}
 
 /* ---------------------------------------------------------------------------
    Main scheduling loop.
@@ -310,13 +298,6 @@ taskStart(void)
       * thread ends
       * stack overflow
 
-   Locking notes:  we acquire the scheduler lock once at the beginning
-   of the scheduler loop, and release it when
-    
-      * running a thread, or
-      * 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 
@@ -333,18 +314,16 @@ taskStart(void)
      This is not the ugliest code you could imagine, but it's bloody close.
 
    ------------------------------------------------------------------------ */
-//@cindex schedule
-static void
-schedule( void )
+
+static Capability *
+schedule (Capability *initialCapability, Task *task)
 {
   StgTSO *t;
   Capability *cap;
   StgThreadReturnCode ret;
 #if defined(GRAN)
   rtsEvent *event;
-#elif defined(PAR)
-  StgSparkPool *pool;
-  rtsSpark spark;
+#elif defined(PARALLEL_HASKELL)
   StgTSO *tso;
   GlobalTaskId pe;
   rtsBool receivedFinish = rtsFalse;
@@ -352,358 +331,712 @@ schedule( void )
   nat tp_size, sp_size; // stats only
 # endif
 #endif
-  rtsBool was_interrupted = rtsFalse;
-  
-  ACQUIRE_LOCK(&sched_mutex);
-#if defined(RTS_SUPPORTS_THREADS)
-  /* Check to see whether there are any worker threads
-     waiting to deposit external call results. If so,
-     yield our capability */
-  yieldToReturningWorker(&sched_mutex, cap);
-
-  waitForWorkCapability(&sched_mutex, &cap, rtsFalse);
+  nat prev_what_next;
+  rtsBool ready_to_gc;
+#if defined(THREADED_RTS)
+  rtsBool first = rtsTrue;
 #endif
+  
+  cap = initialCapability;
 
-#if defined(GRAN)
-  /* set up first event to get things going */
-  /* ToDo: assign costs for system setup and init MainTSO ! */
-  new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
-           ContinueThread, 
-           CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
-
-  IF_DEBUG(gran,
-          fprintf(stderr, "GRAN: Init CurrentTSO (in schedule) = %p\n", CurrentTSO);
-          G_TSO(CurrentTSO, 5));
-
-  if (RtsFlags.GranFlags.Light) {
-    /* Save current time; GranSim Light only */
-    CurrentTSO->gran.clock = CurrentTime[CurrentProc];
-  }      
+  // Pre-condition: this task owns initialCapability.
+  // The sched_mutex is *NOT* held
+  // NB. on return, we still hold a capability.
 
-  event = get_next_event();
+  IF_DEBUG(scheduler,
+          sched_belch("### NEW SCHEDULER LOOP (task: %p, cap: %p)",
+                      task, initialCapability);
+      );
 
-  while (event!=(rtsEvent*)NULL) {
-    /* Choose the processor with the next event */
-    CurrentProc = event->proc;
-    CurrentTSO = event->tso;
+  schedulePreLoop();
 
-#elif defined(PAR)
+  // -----------------------------------------------------------
+  // Scheduler loop starts here:
 
-  while (!receivedFinish) {    /* set by processMessages */
-                               /* when receiving PP_FINISH message         */ 
+#if defined(PARALLEL_HASKELL)
+#define TERMINATION_CONDITION        (!receivedFinish)
+#elif defined(GRAN)
+#define TERMINATION_CONDITION        ((event = get_next_event()) != (rtsEvent*)NULL) 
 #else
+#define TERMINATION_CONDITION        rtsTrue
+#endif
 
-  while (1) {
+  while (TERMINATION_CONDITION) {
 
+#if defined(GRAN)
+      /* Choose the processor with the next event */
+      CurrentProc = event->proc;
+      CurrentTSO = event->tso;
 #endif
 
-    IF_DEBUG(scheduler, printAllThreads());
+#if defined(THREADED_RTS)
+      if (first) {
+         // don't yield the first time, we want a chance to run this
+         // thread for a bit, even if there are others banging at the
+         // door.
+         first = rtsFalse;
+         ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
+      } else {
+         // Yield the capability to higher-priority tasks if necessary.
+         yieldCapability(&cap, task);
+      }
+#endif
+      
+#if defined(THREADED_RTS)
+      schedulePushWork(cap,task);
+#endif
 
-    /* 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"));
-      deleteAllThreads();
-      interrupted = rtsFalse;
-      was_interrupted = rtsTrue;
+    // Check whether we have re-entered the RTS from Haskell without
+    // going via suspendThread()/resumeThread (i.e. a 'safe' foreign
+    // call).
+    if (cap->in_haskell) {
+         errorBelch("schedule: re-entered unsafely.\n"
+                    "   Perhaps a 'foreign import unsafe' should be 'safe'?");
+         stg_exit(EXIT_FAILURE);
     }
 
-    /* Go through the list of main threads and wake up any
-     * clients whose computations have finished.  ToDo: this
-     * should be done more efficiently without a linear scan
-     * of the main threads list, somehow...
-     */
-#if defined(RTS_SUPPORTS_THREADS)
-    { 
-      StgMainThread *m, **prev;
-      prev = &main_threads;
-      for (m = main_threads; m != NULL; 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;
-         broadcastCondition(&m->wakeup);
-         break;
-       case ThreadKilled:
-         if (m->ret) *(m->ret) = NULL;
-         *prev = m->link;
-         if (was_interrupted) {
-           m->stat = Interrupted;
-         } else {
-           m->stat = Killed;
-         }
-         broadcastCondition(&m->wakeup);
-         break;
-       default:
-         break;
+    // The interruption / shutdown sequence.
+    // 
+    // In order to cleanly shut down the runtime, we want to:
+    //   * make sure that all main threads return to their callers
+    //     with the state 'Interrupted'.
+    //   * clean up all OS threads assocated with the runtime
+    //   * free all memory etc.
+    //
+    // So the sequence for ^C goes like this:
+    //
+    //   * ^C handler sets sched_state := SCHED_INTERRUPTING and
+    //     arranges for some Capability to wake up
+    //
+    //   * all threads in the system are halted, and the zombies are
+    //     placed on the run queue for cleaning up.  We acquire all
+    //     the capabilities in order to delete the threads, this is
+    //     done by scheduleDoGC() for convenience (because GC already
+    //     needs to acquire all the capabilities).  We can't kill
+    //     threads involved in foreign calls.
+    // 
+    //   * sched_state := SCHED_INTERRUPTED
+    //
+    //   * somebody calls shutdownHaskell(), which calls exitScheduler()
+    //
+    //   * sched_state := SCHED_SHUTTING_DOWN
+    //
+    //   * all workers exit when the run queue on their capability
+    //     drains.  All main threads will also exit when their TSO
+    //     reaches the head of the run queue and they can return.
+    //
+    //   * eventually all Capabilities will shut down, and the RTS can
+    //     exit.
+    //
+    //   * We might be left with threads blocked in foreign calls, 
+    //     we should really attempt to kill these somehow (TODO);
+    
+    switch (sched_state) {
+    case SCHED_RUNNING:
+       break;
+    case SCHED_INTERRUPTING:
+       IF_DEBUG(scheduler, sched_belch("SCHED_INTERRUPTING"));
+#if defined(THREADED_RTS)
+       discardSparksCap(cap);
+#endif
+       /* scheduleDoGC() deletes all the threads */
+       cap = scheduleDoGC(cap,task,rtsFalse,GetRoots);
+       break;
+    case SCHED_INTERRUPTED:
+       IF_DEBUG(scheduler, sched_belch("SCHED_INTERRUPTED"));
+       break;
+    case SCHED_SHUTTING_DOWN:
+       IF_DEBUG(scheduler, sched_belch("SCHED_SHUTTING_DOWN"));
+       // If we are a worker, just exit.  If we're a bound thread
+       // then we will exit below when we've removed our TSO from
+       // the run queue.
+       if (task->tso == NULL && emptyRunQueue(cap)) {
+           return cap;
        }
-      }
+       break;
+    default:
+       barf("sched_state: %d", sched_state);
     }
 
-#else /* not threaded */
-
-# if defined(PAR)
-    /* in GUM do this only on the Main PE */
-    if (IAmMainThread)
-# endif
-    /* If our main thread has finished or been killed, return.
-     */
+#if defined(THREADED_RTS)
+    // If the run queue is empty, take a spark and turn it into a thread.
     {
-      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 (m->ret) { *(m->ret) = NULL; };
-         if (was_interrupted) {
-           m->stat = Interrupted;
-         } else {
-           m->stat = Killed;
-         }
-         return;
+       if (emptyRunQueue(cap)) {
+           StgClosure *spark;
+           spark = findSpark(cap);
+           if (spark != NULL) {
+               IF_DEBUG(scheduler,
+                        sched_belch("turning spark of closure %p into a thread",
+                                    (StgClosure *)spark));
+               createSparkThread(cap,spark);     
+           }
        }
-      }
     }
-#endif
+#endif // THREADED_RTS
 
-    /* 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.
-     *
-     * Disable spark support in SMP for now, non-essential & requires
-     * a little bit of work to make it compile cleanly. -- sof 1/02.
-     */
-#if 0 /* defined(SMP) */
-    {
-      nat n = getFreeCapabilities();
-      StgTSO *tso = run_queue_hd;
+    scheduleStartSignalHandlers(cap);
 
-      /* Count the run queue */
-      while (n > 0 && tso != END_TSO_QUEUE) {
-       tso = tso->link;
-       n--;
-      }
+    // Only check the black holes here if we've nothing else to do.
+    // During normal execution, the black hole list only gets checked
+    // at GC time, to avoid repeatedly traversing this possibly long
+    // list each time around the scheduler.
+    if (emptyRunQueue(cap)) { scheduleCheckBlackHoles(cap); }
 
-      for (; n > 0; n--) {
-       StgClosure *spark;
-       spark = findSpark(rtsFalse);
-       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 */
-         createSparkThread(spark);       
-         IF_DEBUG(scheduler,
-                  sched_belch("==^^ turning spark of closure %p into a thread",
-                              (StgClosure *)spark));
-       }
-      }
-      /* We need to wake up the other tasks if we just created some
-       * work for them.
-       */
-      if (getFreeCapabilities() - n > 1) {
-         signalCondition( &thread_ready_cond );
-      }
-    }
-#endif // SMP
+    scheduleCheckBlockedThreads(cap);
 
-    /* check for signals each time around the scheduler */
-#ifndef mingw32_TARGET_OS
-    if (signals_pending()) {
-      startSignalHandlers();
-    }
+    scheduleDetectDeadlock(cap,task);
+#if defined(THREADED_RTS)
+    cap = task->cap;    // reload cap, it might have changed
 #endif
 
-    /* Check whether any waiting threads need to be woken up.  If the
-     * run queue is empty, and there are no other tasks running, we
-     * can wait indefinitely for something to happen.
-     * ToDo: what if another client comes along & requests another
-     * main thread?
-     */
-    if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue) ) {
-      awaitEvent( EMPTY_RUN_QUEUE()
-#if defined(SMP)
-       && allFreeCapabilities()
+    // Normally, the only way we can get here with no threads to
+    // run is if a keyboard interrupt received during 
+    // scheduleCheckBlockedThreads() or scheduleDetectDeadlock().
+    // Additionally, it is not fatal for the
+    // threaded RTS to reach here with no threads to run.
+    //
+    // win32: might be here due to awaitEvent() being abandoned
+    // as a result of a console event having been delivered.
+    if ( emptyRunQueue(cap) ) {
+#if !defined(THREADED_RTS) && !defined(mingw32_HOST_OS)
+       ASSERT(sched_state >= SCHED_INTERRUPTING);
 #endif
-       );
+       continue; // nothing to do
     }
-    /* 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.
-     */
-#ifndef PAR
-    if (   EMPTY_THREAD_QUEUES()
-#if defined(RTS_SUPPORTS_THREADS)
-       && EMPTY_QUEUE(suspended_ccalling_threads)
-#endif
-#ifdef SMP
-       && allFreeCapabilities()
-#endif
-       )
-    {
-       IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
-#if defined(THREADED_RTS)
-       /* and SMP mode ..? */
-       releaseCapability(cap);
-#endif
-       // Garbage collection can release some new threads due to
-       // either (a) finalizers or (b) threads resurrected because
-       // they are about to be send BlockedOnDeadMVar.  Any threads
-       // thus released will be immediately runnable.
-       GarbageCollect(GetRoots,rtsTrue);
-
-       if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
 
-       IF_DEBUG(scheduler, 
-                sched_belch("still deadlocked, checking for black holes..."));
-       detectBlackHoles();
+#if defined(PARALLEL_HASKELL)
+    scheduleSendPendingMessages();
+    if (emptyRunQueue(cap) && scheduleActivateSpark()) 
+       continue;
 
-       if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
+#if defined(SPARKS)
+    ASSERT(next_fish_to_send_at==0);  // i.e. no delayed fishes left!
+#endif
 
-#ifndef mingw32_TARGET_OS
-       /* If we have user-installed signal handlers, then wait
-        * for signals to arrive rather then bombing out with a
-        * deadlock.
-        */
-       if ( anyUserHandlers() ) {
-           IF_DEBUG(scheduler, 
-                    sched_belch("still deadlocked, waiting for signals..."));
+    /* If we still have no work we need to send a FISH to get a spark
+       from another PE */
+    if (emptyRunQueue(cap)) {
+       if (!scheduleGetRemoteWork(&receivedFinish)) continue;
+       ASSERT(rtsFalse); // should not happen at the moment
+    }
+    // from here: non-empty run queue.
+    //  TODO: merge above case with this, only one call processMessages() !
+    if (PacketsWaiting()) {  /* process incoming messages, if
+                               any pending...  only in else
+                               because getRemoteWork waits for
+                               messages as well */
+       receivedFinish = processMessages();
+    }
+#endif
 
-           awaitUserSignals();
+#if defined(GRAN)
+    scheduleProcessEvent(event);
+#endif
 
-           // we might be interrupted...
-           if (interrupted) { continue; }
+    // 
+    // Get a thread to run
+    //
+    t = popRunQueue(cap);
 
-           if (signals_pending()) {
-               startSignalHandlers();
-           }
-           ASSERT(!EMPTY_RUN_QUEUE());
-           goto not_deadlocked;
-       }
+#if defined(GRAN) || defined(PAR)
+    scheduleGranParReport(); // some kind of debuging output
+#else
+    // Sanity check the thread we're about to run.  This can be
+    // expensive if there is lots of thread switching going on...
+    IF_DEBUG(sanity,checkTSO(t));
 #endif
 
-       /* 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;
-#if defined(RTS_SUPPORTS_THREADS)
-           for (m = main_threads; m != NULL; m = m->link) {
-               switch (m->tso->why_blocked) {
-               case BlockedOnBlackHole:
-                   raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
-                   break;
-               case BlockedOnException:
-               case BlockedOnMVar:
-                   raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
-                   break;
-               default:
-                   barf("deadlock: main thread blocked in a strange way");
-               }
+#if defined(THREADED_RTS)
+    // Check whether we can run this thread in the current task.
+    // If not, we have to pass our capability to the right task.
+    {
+       Task *bound = t->bound;
+      
+       if (bound) {
+           if (bound == task) {
+               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
+               pushOnRunQueue(cap,t);
+               continue;
            }
-#else
-           m = main_threads;
-           switch (m->tso->why_blocked) {
-           case BlockedOnBlackHole:
-               raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
-               break;
-           case BlockedOnException:
-           case BlockedOnMVar:
-               raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
-               break;
-           default:
-               barf("deadlock: main thread blocked in a strange way");
+       } else {
+           // The thread we want to run is unbound.
+           if (task->tso) { 
+               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
+               pushOnRunQueue(cap,t);
+               continue; 
            }
-#endif
        }
-
-#if defined(RTS_SUPPORTS_THREADS)
-       /* ToDo: revisit conditions (and mechanism) for shutting
-          down a multi-threaded world  */
-       IF_DEBUG(scheduler, sched_belch("all done, i think...shutting down."));
-       shutdownHaskellAndExit(0);
-#endif
     }
-  not_deadlocked:
-
-#elif defined(PAR)
-    /* ToDo: add deadlock detection in GUM (similar to SMP) -- HWL */
 #endif
 
-#if defined(SMP)
-    /* If there's a GC pending, don't do anything until it has
-     * completed.
+    cap->r.rCurrentTSO = t;
+    
+    /* context switches are initiated by the timer signal, unless
+     * the user specified "context switch as often as possible", with
+     * +RTS -C0
      */
-    if (ready_to_gc) {
-      IF_DEBUG(scheduler,sched_belch("waiting for GC"));
-      waitCondition( &gc_pending_cond, &sched_mutex );
+    if (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
+       && !emptyThreadQueues(cap)) {
+       context_switch = 1;
     }
-#endif    
+        
+run_thread:
 
-#if defined(RTS_SUPPORTS_THREADS)
-    /* block until we've got a thread on the run queue and a free
-     * capability.
-     *
-     */
-    if ( EMPTY_RUN_QUEUE() ) {
-      /* Give up our capability */
-      releaseCapability(cap);
-      IF_DEBUG(scheduler, sched_belch("thread %d: waiting for work", osThreadId()));
-      waitForWorkCapability(&sched_mutex, &cap, rtsTrue);
-      IF_DEBUG(scheduler, sched_belch("thread %d: work now available", osThreadId()));
-#if 0
-      while ( EMPTY_RUN_QUEUE() ) {
-       waitForWorkCapability(&sched_mutex, &cap);
-       IF_DEBUG(scheduler, sched_belch("thread %d: work now available", osThreadId()));
-      }
-#endif
-    }
+    IF_DEBUG(scheduler, sched_belch("-->> running thread %ld %s ...", 
+                             (long)t->id, whatNext_strs[t->what_next]));
+
+#if defined(PROFILING)
+    startHeapProfTimer();
 #endif
 
-#if defined(GRAN)
-    if (RtsFlags.GranFlags.Light)
-      GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
+    // ----------------------------------------------------------------------
+    // Run the current thread 
 
-    /* adjust time based on time-stamp */
-    if (event->time > CurrentTime[CurrentProc] &&
-        event->evttype != ContinueThread)
-      CurrentTime[CurrentProc] = event->time;
-    
+    ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
+
+    prev_what_next = t->what_next;
+
+    errno = t->saved_errno;
+    cap->in_haskell = rtsTrue;
+
+    dirtyTSO(t);
+
+    recent_activity = ACTIVITY_YES;
+
+    switch (prev_what_next) {
+       
+    case ThreadKilled:
+    case ThreadComplete:
+       /* Thread already finished, return to scheduler. */
+       ret = ThreadFinished;
+       break;
+       
+    case ThreadRunGHC:
+    {
+       StgRegTable *r;
+       r = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
+       cap = regTableToCapability(r);
+       ret = r->rRet;
+       break;
+    }
+    
+    case ThreadInterpret:
+       cap = interpretBCO(cap);
+       ret = cap->r.rRet;
+       break;
+       
+    default:
+       barf("schedule: invalid what_next field");
+    }
+
+    cap->in_haskell = rtsFalse;
+
+    // The TSO might have moved, eg. if it re-entered the RTS and a GC
+    // happened.  So find the new location:
+    t = cap->r.rCurrentTSO;
+
+    // We have run some Haskell code: there might be blackhole-blocked
+    // threads to wake up now.
+    // Lock-free test here should be ok, we're just setting a flag.
+    if ( blackhole_queue != END_TSO_QUEUE ) {
+       blackholes_need_checking = rtsTrue;
+    }
+    
+    // And save the current errno in this thread.
+    // XXX: possibly bogus for SMP because this thread might already
+    // be running again, see code below.
+    t->saved_errno = errno;
+
+#if defined(THREADED_RTS)
+    // If ret is ThreadBlocked, and this Task is bound to the TSO that
+    // blocked, we are in limbo - the TSO is now owned by whatever it
+    // is blocked on, and may in fact already have been woken up,
+    // perhaps even on a different Capability.  It may be the case
+    // that task->cap != cap.  We better yield this Capability
+    // immediately and return to normaility.
+    if (ret == ThreadBlocked) {
+       IF_DEBUG(scheduler,
+                sched_belch("--<< thread %d (%s) stopped: blocked\n",
+                            t->id, whatNext_strs[t->what_next]));
+       continue;
+    }
+#endif
+
+    ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
+
+    // ----------------------------------------------------------------------
+    
+    // Costs for the scheduler are assigned to CCS_SYSTEM
+#if defined(PROFILING)
+    stopHeapProfTimer();
+    CCCS = CCS_SYSTEM;
+#endif
+    
+#if defined(THREADED_RTS)
+    IF_DEBUG(scheduler,debugBelch("sched (task %p): ", (void *)(unsigned long)(unsigned int)osThreadId()););
+#elif !defined(GRAN) && !defined(PARALLEL_HASKELL)
+    IF_DEBUG(scheduler,debugBelch("sched: "););
+#endif
+    
+    schedulePostRunThread();
+
+    ready_to_gc = rtsFalse;
+
+    switch (ret) {
+    case HeapOverflow:
+       ready_to_gc = scheduleHandleHeapOverflow(cap,t);
+       break;
+
+    case StackOverflow:
+       scheduleHandleStackOverflow(cap,task,t);
+       break;
+
+    case ThreadYielding:
+       if (scheduleHandleYield(cap, t, prev_what_next)) {
+            // shortcut for switching between compiler/interpreter:
+           goto run_thread; 
+       }
+       break;
+
+    case ThreadBlocked:
+       scheduleHandleThreadBlocked(t);
+       break;
+
+    case ThreadFinished:
+       if (scheduleHandleThreadFinished(cap, task, t)) return cap;
+       ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
+       break;
+
+    default:
+      barf("schedule: invalid thread return code %d", (int)ret);
+    }
+
+    if (scheduleDoHeapProfile(ready_to_gc)) { ready_to_gc = rtsFalse; }
+    if (ready_to_gc) {
+      cap = scheduleDoGC(cap,task,rtsFalse,GetRoots);
+    }
+  } /* end of while() */
+
+  IF_PAR_DEBUG(verbose,
+              debugBelch("== Leaving schedule() after having received Finish\n"));
+}
+
+/* ----------------------------------------------------------------------------
+ * Setting up the scheduler loop
+ * ------------------------------------------------------------------------- */
+
+static void
+schedulePreLoop(void)
+{
+#if defined(GRAN) 
+    /* set up first event to get things going */
+    /* ToDo: assign costs for system setup and init MainTSO ! */
+    new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
+             ContinueThread, 
+             CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
+    
+    IF_DEBUG(gran,
+            debugBelch("GRAN: Init CurrentTSO (in schedule) = %p\n", 
+                       CurrentTSO);
+            G_TSO(CurrentTSO, 5));
+    
+    if (RtsFlags.GranFlags.Light) {
+       /* Save current time; GranSim Light only */
+       CurrentTSO->gran.clock = CurrentTime[CurrentProc];
+    }      
+#endif
+}
+
+/* -----------------------------------------------------------------------------
+ * schedulePushWork()
+ *
+ * Push work to other Capabilities if we have some.
+ * -------------------------------------------------------------------------- */
+
+#if defined(THREADED_RTS)
+static void
+schedulePushWork(Capability *cap USED_IF_THREADS, 
+                Task *task      USED_IF_THREADS)
+{
+    Capability *free_caps[n_capabilities], *cap0;
+    nat i, n_free_caps;
+
+    // Check whether we have more threads on our run queue, or sparks
+    // in our pool, that we could hand to another Capability.
+    if ((emptyRunQueue(cap) || cap->run_queue_hd->link == END_TSO_QUEUE)
+       && sparkPoolSizeCap(cap) < 2) {
+       return;
+    }
+
+    // First grab as many free Capabilities as we can.
+    for (i=0, n_free_caps=0; i < n_capabilities; i++) {
+       cap0 = &capabilities[i];
+       if (cap != cap0 && tryGrabCapability(cap0,task)) {
+           if (!emptyRunQueue(cap0) || cap->returning_tasks_hd != NULL) {
+               // it already has some work, we just grabbed it at 
+               // the wrong moment.  Or maybe it's deadlocked!
+               releaseCapability(cap0);
+           } else {
+               free_caps[n_free_caps++] = cap0;
+           }
+       }
+    }
+
+    // we now have n_free_caps free capabilities stashed in
+    // free_caps[].  Share our run queue equally with them.  This is
+    // probably the simplest thing we could do; improvements we might
+    // want to do include:
+    //
+    //   - giving high priority to moving relatively new threads, on 
+    //     the gournds that they haven't had time to build up a
+    //     working set in the cache on this CPU/Capability.
+    //
+    //   - giving low priority to moving long-lived threads
+
+    if (n_free_caps > 0) {
+       StgTSO *prev, *t, *next;
+       rtsBool pushed_to_all;
+
+       IF_DEBUG(scheduler, sched_belch("excess threads on run queue and %d free capabilities, sharing...", n_free_caps));
+
+       i = 0;
+       pushed_to_all = rtsFalse;
+
+       if (cap->run_queue_hd != END_TSO_QUEUE) {
+           prev = cap->run_queue_hd;
+           t = prev->link;
+           prev->link = END_TSO_QUEUE;
+           for (; t != END_TSO_QUEUE; t = next) {
+               next = t->link;
+               t->link = END_TSO_QUEUE;
+               if (t->what_next == ThreadRelocated
+                   || t->bound == task) { // don't move my bound thread
+                   prev->link = t;
+                   prev = t;
+               } else if (i == n_free_caps) {
+                   pushed_to_all = rtsTrue;
+                   i = 0;
+                   // keep one for us
+                   prev->link = t;
+                   prev = t;
+               } else {
+                   IF_DEBUG(scheduler, sched_belch("pushing thread %d to capability %d", t->id, free_caps[i]->no));
+                   appendToRunQueue(free_caps[i],t);
+                   if (t->bound) { t->bound->cap = free_caps[i]; }
+                   i++;
+               }
+           }
+           cap->run_queue_tl = prev;
+       }
+
+       // If there are some free capabilities that we didn't push any
+       // threads to, then try to push a spark to each one.
+       if (!pushed_to_all) {
+           StgClosure *spark;
+           // i is the next free capability to push to
+           for (; i < n_free_caps; i++) {
+               if (emptySparkPoolCap(free_caps[i])) {
+                   spark = findSpark(cap);
+                   if (spark != NULL) {
+                       IF_DEBUG(scheduler, sched_belch("pushing spark %p to capability %d", spark, free_caps[i]->no));
+                       newSpark(&(free_caps[i]->r), spark);
+                   }
+               }
+           }
+       }
+
+       // release the capabilities
+       for (i = 0; i < n_free_caps; i++) {
+           task->cap = free_caps[i];
+           releaseCapability(free_caps[i]);
+       }
+    }
+    task->cap = cap; // reset to point to our Capability.
+}
+#endif
+
+/* ----------------------------------------------------------------------------
+ * Start any pending signal handlers
+ * ------------------------------------------------------------------------- */
+
+#if defined(RTS_USER_SIGNALS) && (!defined(THREADED_RTS) || defined(mingw32_HOST_OS))
+static void
+scheduleStartSignalHandlers(Capability *cap)
+{
+    if (signals_pending()) { // safe outside the lock
+       startSignalHandlers(cap);
+    }
+}
+#else
+static void
+scheduleStartSignalHandlers(Capability *cap STG_UNUSED)
+{
+}
+#endif
+
+/* ----------------------------------------------------------------------------
+ * Check for blocked threads that can be woken up.
+ * ------------------------------------------------------------------------- */
+
+static void
+scheduleCheckBlockedThreads(Capability *cap USED_IF_NOT_THREADS)
+{
+#if !defined(THREADED_RTS)
+    //
+    // 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 ( !emptyQueue(blocked_queue_hd) || !emptyQueue(sleeping_queue) )
+    {
+       awaitEvent( emptyRunQueue(cap) && !blackholes_need_checking );
+    }
+#endif
+}
+
+
+/* ----------------------------------------------------------------------------
+ * Check for threads blocked on BLACKHOLEs that can be woken up
+ * ------------------------------------------------------------------------- */
+static void
+scheduleCheckBlackHoles (Capability *cap)
+{
+    if ( blackholes_need_checking ) // check without the lock first
+    {
+       ACQUIRE_LOCK(&sched_mutex);
+       if ( blackholes_need_checking ) {
+           checkBlackHoles(cap);
+           blackholes_need_checking = rtsFalse;
+       }
+       RELEASE_LOCK(&sched_mutex);
+    }
+}
+
+/* ----------------------------------------------------------------------------
+ * Detect deadlock conditions and attempt to resolve them.
+ * ------------------------------------------------------------------------- */
+
+static void
+scheduleDetectDeadlock (Capability *cap, Task *task)
+{
+
+#if defined(PARALLEL_HASKELL)
+    // ToDo: add deadlock detection in GUM (similar to THREADED_RTS) -- HWL
+    return;
+#endif
+
+    /* 
+     * Detect deadlock: when we have no threads to run, there are no
+     * threads blocked, waiting for I/O, or sleeping, and all the
+     * other tasks are waiting for work, we must have a deadlock of
+     * some description.
+     */
+    if ( emptyThreadQueues(cap) )
+    {
+#if defined(THREADED_RTS)
+       /* 
+        * In the threaded RTS, we only check for deadlock if there
+        * has been no activity in a complete timeslice.  This means
+        * we won't eagerly start a full GC just because we don't have
+        * any threads to run currently.
+        */
+       if (recent_activity != ACTIVITY_INACTIVE) return;
+#endif
+
+       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.
+       cap = scheduleDoGC (cap, task, rtsTrue/*force  major GC*/, GetRoots);
+
+       recent_activity = ACTIVITY_DONE_GC;
+       
+       if ( !emptyRunQueue(cap) ) return;
+
+#if defined(RTS_USER_SIGNALS) && (!defined(THREADED_RTS) || defined(mingw32_HOST_OS))
+       /* If we have user-installed signal handlers, then wait
+        * for signals to arrive rather then bombing out with a
+        * deadlock.
+        */
+       if ( anyUserHandlers() ) {
+           IF_DEBUG(scheduler, 
+                    sched_belch("still deadlocked, waiting for signals..."));
+
+           awaitUserSignals();
+
+           if (signals_pending()) {
+               startSignalHandlers(cap);
+           }
+
+           // either we have threads to run, or we were interrupted:
+           ASSERT(!emptyRunQueue(cap) || sched_state >= SCHED_INTERRUPTING);
+       }
+#endif
+
+#if !defined(THREADED_RTS)
+       /* Probably a real deadlock.  Send the current main thread the
+        * Deadlock exception.
+        */
+       if (task->tso) {
+           switch (task->tso->why_blocked) {
+           case BlockedOnSTM:
+           case BlockedOnBlackHole:
+           case BlockedOnException:
+           case BlockedOnMVar:
+               raiseAsync(cap, task->tso, (StgClosure *)NonTermination_closure);
+               return;
+           default:
+               barf("deadlock: main thread blocked in a strange way");
+           }
+       }
+       return;
+#endif
+    }
+}
+
+/* ----------------------------------------------------------------------------
+ * Process an event (GRAN only)
+ * ------------------------------------------------------------------------- */
+
+#if defined(GRAN)
+static StgTSO *
+scheduleProcessEvent(rtsEvent *event)
+{
+    StgTSO *t;
+
+    if (RtsFlags.GranFlags.Light)
+      GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
+
+    /* adjust time based on time-stamp */
+    if (event->time > CurrentTime[CurrentProc] &&
+        event->evttype != ContinueThread)
+      CurrentTime[CurrentProc] = event->time;
+    
     /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
     if (!RtsFlags.GranFlags.Light)
       handleIdlePEs();
 
-    IF_DEBUG(gran, fprintf(stderr, "GRAN: switch by event-type\n"));
+    IF_DEBUG(gran, debugBelch("GRAN: switch by event-type\n"));
 
     /* main event dispatcher in GranSim */
     switch (event->evttype) {
       /* Should just be continuing execution */
     case ContinueThread:
-      IF_DEBUG(gran, fprintf(stderr, "GRAN: doing ContinueThread\n"));
+      IF_DEBUG(gran, debugBelch("GRAN: doing ContinueThread\n"));
       /* ToDo: check assertion
       ASSERT(run_queue_hd != (StgTSO*)NULL &&
             run_queue_hd != END_TSO_QUEUE);
@@ -711,25 +1044,25 @@ schedule( void )
       /* Ignore ContinueThreads for fetching threads (if synchr comm) */
       if (!RtsFlags.GranFlags.DoAsyncFetch &&
          procStatus[CurrentProc]==Fetching) {
-       belch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]",
+       debugBelch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]\n",
              CurrentTSO->id, CurrentTSO, CurrentProc);
        goto next_thread;
       }        
       /* Ignore ContinueThreads for completed threads */
       if (CurrentTSO->what_next == ThreadComplete) {
-       belch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)", 
+       debugBelch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)\n", 
              CurrentTSO->id, CurrentTSO, CurrentProc);
        goto next_thread;
       }        
       /* Ignore ContinueThreads for threads that are being migrated */
       if (PROCS(CurrentTSO)==Nowhere) { 
-       belch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)",
+       debugBelch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)\n",
              CurrentTSO->id, CurrentTSO, CurrentProc);
        goto next_thread;
       }
       /* The thread should be at the beginning of the run queue */
       if (CurrentTSO!=run_queue_hds[CurrentProc]) { 
-       belch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread",
+       debugBelch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread\n",
              CurrentTSO->id, CurrentTSO, CurrentProc);
        break; // run the thread anyway
       }
@@ -786,14 +1119,14 @@ schedule( void )
     
     /* This point was scheduler_loop in the old RTS */
 
-    IF_DEBUG(gran, belch("GRAN: after main switch"));
+    IF_DEBUG(gran, debugBelch("GRAN: after main switch\n"));
 
     TimeOfLastEvent = CurrentTime[CurrentProc];
     TimeOfNextEvent = get_time_of_next_event();
     IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
     // CurrentTSO = ThreadQueueHd;
 
-    IF_DEBUG(gran, belch("GRAN: time of next event is: %ld", 
+    IF_DEBUG(gran, debugBelch("GRAN: time of next event is: %ld\n", 
                         TimeOfNextEvent));
 
     if (RtsFlags.GranFlags.Light) 
@@ -802,15 +1135,15 @@ schedule( void )
     EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
 
     IF_DEBUG(gran, 
-            belch("GRAN: end of time-slice is %#lx", EndOfTimeSlice));
+            debugBelch("GRAN: end of time-slice is %#lx\n", EndOfTimeSlice));
 
     /* in a GranSim setup the TSO stays on the run queue */
     t = CurrentTSO;
     /* Take a thread from the run queue. */
-    t = POP_RUN_QUEUE(); // take_off_run_queue(t);
+    POP_RUN_QUEUE(t); // take_off_run_queue(t);
 
     IF_DEBUG(gran, 
-            fprintf(stderr, "GRAN: About to run current thread, which is\n");
+            debugBelch("GRAN: About to run current thread, which is\n");
             G_TSO(t,5));
 
     context_switch = 0; // turned on via GranYield, checking events and time slice
@@ -819,18 +1152,54 @@ schedule( void )
             DumpGranEvent(GR_SCHEDULE, t));
 
     procStatus[CurrentProc] = Busy;
+}
+#endif // GRAN
 
-#elif defined(PAR)
+/* ----------------------------------------------------------------------------
+ * Send pending messages (PARALLEL_HASKELL only)
+ * ------------------------------------------------------------------------- */
+
+#if defined(PARALLEL_HASKELL)
+static StgTSO *
+scheduleSendPendingMessages(void)
+{
+    StgSparkPool *pool;
+    rtsSpark spark;
+    StgTSO *t;
+
+# if defined(PAR) // global Mem.Mgmt., omit for now
     if (PendingFetches != END_BF_QUEUE) {
         processFetches();
     }
+# endif
+    
+    if (RtsFlags.ParFlags.BufferTime) {
+       // if we use message buffering, we must send away all message
+       // packets which have become too old...
+       sendOldBuffers(); 
+    }
+}
+#endif
+
+/* ----------------------------------------------------------------------------
+ * Activate spark threads (PARALLEL_HASKELL only)
+ * ------------------------------------------------------------------------- */
+
+#if defined(PARALLEL_HASKELL)
+static void
+scheduleActivateSpark(void)
+{
+#if defined(SPARKS)
+  ASSERT(emptyRunQueue());
+/* We get here if the run queue is empty and want some work.
+   We try to turn a spark into a thread, and add it to the run queue,
+   from where it will be picked up in the next iteration of the scheduler
+   loop.
+*/
 
-    /* ToDo: phps merge with spark activation above */
-    /* check whether we have local work and send requests if we have none */
-    if (EMPTY_RUN_QUEUE()) {  /* no runnable threads */
       /* :-[  no local threads => look out for local sparks */
       /* the spark pool for the current PE */
-      pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
+      pool = &(cap.r.rSparks); // JB: cap = (old) MainCap
       if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
          pool->hd < pool->tl) {
        /* 
@@ -842,29 +1211,73 @@ schedule( void )
         * thread... 
         */
 
-       spark = findSpark(rtsFalse);                /* get a spark */
-       if (spark != (rtsSpark) NULL) {
-         tso = activateSpark(spark);       /* turn the spark into a thread */
-         IF_PAR_DEBUG(schedule,
-                      belch("==== schedule: Created TSO %d (%p); %d threads active",
-                            tso->id, tso, advisory_thread_count));
-
-         if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
-           belch("==^^ failed to activate spark");
-           goto next_thread;
-         }               /* otherwise fall through & pick-up new tso */
-       } else {
-         IF_PAR_DEBUG(verbose,
-                      belch("==^^ no local sparks (spark pool contains only NFs: %d)", 
-                            spark_queue_len(pool)));
-         goto next_thread;
+       spark = findSpark(rtsFalse);            /* get a spark */
+       if (spark != (rtsSpark) NULL) {
+         tso = createThreadFromSpark(spark);       /* turn the spark into a thread */
+         IF_PAR_DEBUG(fish, // schedule,
+                      debugBelch("==== schedule: Created TSO %d (%p); %d threads active\n",
+                            tso->id, tso, advisory_thread_count));
+
+         if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
+           IF_PAR_DEBUG(fish, // schedule,
+                        debugBelch("==^^ failed to create thread from spark @ %lx\n",
+                            spark));
+           return rtsFalse; /* failed to generate a thread */
+         }                  /* otherwise fall through & pick-up new tso */
+       } else {
+         IF_PAR_DEBUG(fish, // schedule,
+                      debugBelch("==^^ no local sparks (spark pool contains only NFs: %d)\n", 
+                            spark_queue_len(pool)));
+         return rtsFalse;  /* failed to generate a thread */
+       }
+       return rtsTrue;  /* success in generating a thread */
+  } else { /* no more threads permitted or pool empty */
+    return rtsFalse;  /* failed to generateThread */
+  }
+#else
+  tso = NULL; // avoid compiler warning only
+  return rtsFalse;  /* dummy in non-PAR setup */
+#endif // SPARKS
+}
+#endif // PARALLEL_HASKELL
+
+/* ----------------------------------------------------------------------------
+ * Get work from a remote node (PARALLEL_HASKELL only)
+ * ------------------------------------------------------------------------- */
+    
+#if defined(PARALLEL_HASKELL)
+static rtsBool
+scheduleGetRemoteWork(rtsBool *receivedFinish)
+{
+  ASSERT(emptyRunQueue());
+
+  if (RtsFlags.ParFlags.BufferTime) {
+       IF_PAR_DEBUG(verbose, 
+               debugBelch("...send all pending data,"));
+        {
+         nat i;
+         for (i=1; i<=nPEs; i++)
+           sendImmediately(i); // send all messages away immediately
        }
-      }
+  }
+# ifndef SPARKS
+       //++EDEN++ idle() , i.e. send all buffers, wait for work
+       // suppress fishing in EDEN... just look for incoming messages
+       // (blocking receive)
+  IF_PAR_DEBUG(verbose, 
+              debugBelch("...wait for incoming messages...\n"));
+  *receivedFinish = processMessages(); // blocking receive...
+
+       // and reenter scheduling loop after having received something
+       // (return rtsFalse below)
+
+# else /* activate SPARKS machinery */
+/* We get here, if we have no work, tried to activate a local spark, but still
+   have no work. We try to get a remote spark, by sending a FISH message.
+   Thread migration should be added here, and triggered when a sequence of 
+   fishes returns without work. */
+       delay = (RtsFlags.ParFlags.fishDelay!=0ll ? RtsFlags.ParFlags.fishDelay : 0ll);
 
-      /* If we still have no work we need to send a FISH to get a spark
-        from another PE 
-      */
-      if (EMPTY_RUN_QUEUE()) {
       /* =8-[  no local sparks => look for work on other PEs */
        /*
         * We really have absolutely no work.  Send out a fish
@@ -874,48 +1287,106 @@ schedule( void )
         * we're hoping to see.  (Of course, we still have to
         * respond to other types of messages.)
         */
-       TIME now = msTime() /*CURRENT_TIME*/;
+       rtsTime now = msTime() /*CURRENT_TIME*/;
        IF_PAR_DEBUG(verbose, 
-                    belch("--  now=%ld", now));
-       IF_PAR_DEBUG(verbose,
-                    if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
-                        (last_fish_arrived_at!=0 &&
-                         last_fish_arrived_at+RtsFlags.ParFlags.fishDelay > now)) {
-                      belch("--$$ delaying FISH until %ld (last fish %ld, delay %ld, now %ld)",
-                            last_fish_arrived_at+RtsFlags.ParFlags.fishDelay,
-                            last_fish_arrived_at,
-                            RtsFlags.ParFlags.fishDelay, now);
-                    });
-       
+                    debugBelch("--  now=%ld\n", now));
+       IF_PAR_DEBUG(fish, // verbose,
+            if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
+                (last_fish_arrived_at!=0 &&
+                 last_fish_arrived_at+delay > now)) {
+              debugBelch("--$$ <%llu> delaying FISH until %llu (last fish %llu, delay %llu)\n",
+                    now, last_fish_arrived_at+delay, 
+                    last_fish_arrived_at,
+                    delay);
+            });
+  
        if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
-           (last_fish_arrived_at==0 ||
-            (last_fish_arrived_at+RtsFlags.ParFlags.fishDelay <= now))) {
-         /* outstandingFishes is set in sendFish, processFish;
-            avoid flooding system with fishes via delay */
-         pe = choosePE();
-         sendFish(pe, mytid, NEW_FISH_AGE, NEW_FISH_HISTORY, 
-                  NEW_FISH_HUNGER);
-
-         // Global statistics: count no. of fishes
-         if (RtsFlags.ParFlags.ParStats.Global &&
-             RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
-           globalParStats.tot_fish_mess++;
-         }
-       }
-      
-       receivedFinish = processMessages();
-       goto next_thread;
+           advisory_thread_count < RtsFlags.ParFlags.maxThreads) { // send a FISH, but when?
+         if (last_fish_arrived_at==0 ||
+             (last_fish_arrived_at+delay <= now)) {           // send FISH now!
+           /* outstandingFishes is set in sendFish, processFish;
+              avoid flooding system with fishes via delay */
+    next_fish_to_send_at = 0;  
+  } else {
+    /* ToDo: this should be done in the main scheduling loop to avoid the
+             busy wait here; not so bad if fish delay is very small  */
+    int iq = 0; // DEBUGGING -- HWL
+    next_fish_to_send_at = last_fish_arrived_at+delay; // remember when to send  
+    /* send a fish when ready, but process messages that arrive in the meantime */
+    do {
+      if (PacketsWaiting()) {
+        iq++; // DEBUGGING
+        *receivedFinish = processMessages();
       }
-    } else if (PacketsWaiting()) {  /* Look for incoming messages */
-      receivedFinish = processMessages();
-    }
+      now = msTime();
+    } while (!*receivedFinish || now<next_fish_to_send_at);
+    // JB: This means the fish could become obsolete, if we receive
+    // work. Better check for work again? 
+    // last line: while (!receivedFinish || !haveWork || now<...)
+    // next line: if (receivedFinish || haveWork )
+
+    if (*receivedFinish) // no need to send a FISH if we are finishing anyway
+      return rtsFalse;  // NB: this will leave scheduler loop
+                       // immediately after return!
+                         
+    IF_PAR_DEBUG(fish, // verbose,
+              debugBelch("--$$ <%llu> sent delayed fish (%d processMessages); active/total threads=%d/%d\n",now,iq,run_queue_len(),advisory_thread_count));
 
-    /* Now we are sure that we have some work available */
-    ASSERT(run_queue_hd != END_TSO_QUEUE);
+  }
 
-    /* Take a thread from the run queue, if we have work */
-    t = POP_RUN_QUEUE();  // take_off_run_queue(END_TSO_QUEUE);
-    IF_DEBUG(sanity,checkTSO(t));
+    // JB: IMHO, this should all be hidden inside sendFish(...)
+    /* pe = choosePE(); 
+       sendFish(pe, thisPE, NEW_FISH_AGE, NEW_FISH_HISTORY, 
+                NEW_FISH_HUNGER);
+
+    // Global statistics: count no. of fishes
+    if (RtsFlags.ParFlags.ParStats.Global &&
+         RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
+          globalParStats.tot_fish_mess++;
+          }
+    */ 
+
+  /* delayed fishes must have been sent by now! */
+  next_fish_to_send_at = 0;  
+  }
+      
+  *receivedFinish = processMessages();
+# endif /* SPARKS */
+
+ return rtsFalse;
+ /* NB: this function always returns rtsFalse, meaning the scheduler
+    loop continues with the next iteration; 
+    rationale: 
+      return code means success in finding work; we enter this function
+      if there is no local work, thus have to send a fish which takes
+      time until it arrives with work; in the meantime we should process
+      messages in the main loop;
+ */
+}
+#endif // PARALLEL_HASKELL
+
+/* ----------------------------------------------------------------------------
+ * PAR/GRAN: Report stats & debugging info(?)
+ * ------------------------------------------------------------------------- */
+
+#if defined(PAR) || defined(GRAN)
+static void
+scheduleGranParReport(void)
+{
+  ASSERT(run_queue_hd != END_TSO_QUEUE);
+
+  /* Take a thread from the run queue, if we have work */
+  POP_RUN_QUEUE(t);  // take_off_run_queue(END_TSO_QUEUE);
+
+    /* If this TSO has got its outport closed in the meantime, 
+     *   it mustn't be run. Instead, we have to clean it up as if it was finished.
+     * It has to be marked as TH_DEAD for this purpose.
+     * If it is TH_TERM instead, it is supposed to have finished in the normal way.
+
+JB: TODO: investigate wether state change field could be nuked
+     entirely and replaced by the normal tso state (whatnext
+     field). All we want to do is to kill tsos from outside.
+     */
 
     /* ToDo: write something to the log-file
     if (RTSflags.ParFlags.granSimStats && !sameThread)
@@ -924,25 +1395,20 @@ schedule( void )
     CurrentTSO = t;
     */
     /* the spark pool for the current PE */
-    pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
+    pool = &(cap.r.rSparks); //  cap = (old) MainCap
 
     IF_DEBUG(scheduler, 
-            belch("--=^ %d threads, %d sparks on [%#x]", 
+            debugBelch("--=^ %d threads, %d sparks on [%#x]\n", 
                   run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
 
-# if 1
-    if (0 && RtsFlags.ParFlags.ParStats.Full && 
-       t && LastTSO && t->id != LastTSO->id && 
-       LastTSO->why_blocked == NotBlocked && 
-       LastTSO->what_next != ThreadComplete) {
-      // if previously scheduled TSO not blocked we have to record the context switch
-      DumpVeryRawGranEvent(TimeOfLastYield, CURRENT_PROC, CURRENT_PROC,
-                          GR_DESCHEDULE, LastTSO, (StgClosure *)NULL, 0, 0);
-    }
+    IF_PAR_DEBUG(fish,
+            debugBelch("--=^ %d threads, %d sparks on [%#x]\n", 
+                  run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
 
     if (RtsFlags.ParFlags.ParStats.Full && 
-       (emitSchedule /* forced emit */ ||
-        (t && LastTSO && t->id != LastTSO->id))) {
+       (t->par.sparkname != (StgInt)0) && // only log spark generated threads
+       (emitSchedule || // forced emit
+         (t && LastTSO && t->id != LastTSO->id))) {
       /* 
         we are running a different TSO, so write a schedule event to log file
         NB: If we use fair scheduling we also have to write  a deschedule 
@@ -950,88 +1416,23 @@ schedule( void )
             previous tso has blocked whenever we switch to another tso, so
             we don't need it in GUM for now
       */
+      IF_PAR_DEBUG(fish, // schedule,
+                  debugBelch("____ scheduling spark generated thread %d (%lx) (%lx) via a forced emit\n",t->id,t,t->par.sparkname));
+
       DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
                       GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
       emitSchedule = rtsFalse;
     }
-     
-# endif
-#else /* !GRAN && !PAR */
-  
-    /* grab a thread from the run queue */
-    ASSERT(run_queue_hd != END_TSO_QUEUE);
-    t = POP_RUN_QUEUE();
-    // 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
-    
-    grabCapability(&cap);
-    cap->r.rCurrentTSO = t;
-    
-    /* context switches are now initiated by the timer signal, unless
-     * the user specified "context switch as often as possible", with
-     * +RTS -C0
-     */
-    if (
-#ifdef PROFILING
-       RtsFlags.ProfFlags.profileInterval == 0 ||
-#endif
-       (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
-        && (run_queue_hd != END_TSO_QUEUE
-            || blocked_queue_hd != END_TSO_QUEUE
-            || sleeping_queue != END_TSO_QUEUE)))
-       context_switch = 1;
-    else
-       context_switch = 0;
-
-    RELEASE_LOCK(&sched_mutex);
 
-    IF_DEBUG(scheduler, sched_belch("-->> Running TSO %ld (%p) %s ...", 
-                             t->id, t, whatNext_strs[t->what_next]));
-
-#ifdef PROFILING
-    startHeapProfTimer();
-#endif
-
-    /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
-    /* Run the current thread 
-     */
-    switch (cap->r.rCurrentTSO->what_next) {
-    case ThreadKilled:
-    case ThreadComplete:
-       /* Thread already finished, return to scheduler. */
-       ret = ThreadFinished;
-       break;
-    case ThreadEnterGHC:
-       ret = StgRun((StgFunPtr) stg_enterStackTop, &cap->r);
-       break;
-    case ThreadRunGHC:
-       ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
-       break;
-    case ThreadEnterInterp:
-       ret = interpretBCO(cap);
-       break;
-    default:
-      barf("schedule: invalid what_next field");
-    }
-    /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
-    
-    /* Costs for the scheduler are assigned to CCS_SYSTEM */
-#ifdef PROFILING
-    stopHeapProfTimer();
-    CCCS = CCS_SYSTEM;
-#endif
-    
-    ACQUIRE_LOCK(&sched_mutex);
+/* ----------------------------------------------------------------------------
+ * After running a thread...
+ * ------------------------------------------------------------------------- */
 
-#ifdef SMP
-    IF_DEBUG(scheduler,fprintf(stderr,"scheduler (task %ld): ", osThreadId()););
-#elif !defined(GRAN) && !defined(PAR)
-    IF_DEBUG(scheduler,fprintf(stderr,"scheduler: "););
-#endif
-    t = cap->r.rCurrentTSO;
-    
+static void
+schedulePostRunThread(void)
+{
 #if defined(PAR)
     /* HACK 675: if the last thread didn't yield, make sure to print a 
        SCHEDULE event to the log file when StgRunning the next thread, even
@@ -1040,200 +1441,52 @@ schedule( void )
     TimeOfLastYield = CURRENT_TIME;
 #endif
 
-    switch (ret) {
+  /* some statistics gathering in the parallel case */
+
+#if defined(GRAN) || defined(PAR) || defined(EDEN)
+  switch (ret) {
     case HeapOverflow:
-#if defined(GRAN)
+# if defined(GRAN)
       IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
       globalGranStats.tot_heapover++;
-#elif defined(PAR)
+# elif defined(PAR)
       globalParStats.tot_heapover++;
-#endif
-
-      // did the task ask for a large block?
-      if (cap->r.rHpAlloc > BLOCK_SIZE_W) {
-         // if so, get one and push it on the front of the nursery.
-         bdescr *bd;
-         nat blocks;
-         
-         blocks = (nat)BLOCK_ROUND_UP(cap->r.rHpAlloc * sizeof(W_)) / BLOCK_SIZE;
-
-         IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: requesting a large block (size %d)", 
-                                  t->id, t,
-                                  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
-             bd->step = g0s0;
-             bd->gen_no = 0;
-             bd->flags = 0;
-             bd->free = bd->start;
-
-             // don't forget to update the block count in g0s0.
-             g0s0->n_blocks += blocks;
-             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 (%p; %s) stopped: HeapOverflow", 
-                              t->id, t, whatNext_strs[t->what_next]));
-      threadPaused(t);
-#if defined(GRAN)
-      ASSERT(!is_on_queue(t,CurrentProc));
-#elif defined(PAR)
-      /* Currently we emit a DESCHEDULE event before GC in GUM.
-         ToDo: either add separate event to distinguish SYSTEM time from rest
-              or just nuke this DESCHEDULE (and the following SCHEDULE) */
-      if (0 && RtsFlags.ParFlags.ParStats.Full) {
-       DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
-                        GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
-       emitSchedule = rtsTrue;
-      }
-#endif
-      
-      ready_to_gc = rtsTrue;
-      context_switch = 1;              /* stop other threads ASAP */
-      PUSH_ON_RUN_QUEUE(t);
-      /* actual GC is done at the end of the while loop */
+# endif
       break;
-      
-    case StackOverflow:
-#if defined(GRAN)
+
+     case StackOverflow:
+# if defined(GRAN)
       IF_DEBUG(gran, 
               DumpGranEvent(GR_DESCHEDULE, t));
       globalGranStats.tot_stackover++;
-#elif defined(PAR)
+# elif defined(PAR)
       // IF_DEBUG(par, 
       // DumpGranEvent(GR_DESCHEDULE, t);
       globalParStats.tot_stackover++;
-#endif
-      IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped, StackOverflow", 
-                              t->id, t, whatNext_strs[t->what_next]));
-      /* just adjust the stack for this thread, then pop it back
-       * on the run queue.
-       */
-      threadPaused(t);
-      { 
-       StgMainThread *m;
-       /* enlarge the stack */
-       StgTSO *new_t = threadStackOverflow(t);
-       
-       /* This TSO has moved, so update any pointers to it from the
-        * main thread stack.  It better not be on any other queues...
-        * (it shouldn't be).
-        */
-       for (m = main_threads; m != NULL; m = m->link) {
-         if (m->tso == t) {
-           m->tso = new_t;
-         }
-       }
-       threadPaused(new_t);
-       PUSH_ON_RUN_QUEUE(new_t);
-      }
+# endif
       break;
 
     case ThreadYielding:
-#if defined(GRAN)
+# if defined(GRAN)
       IF_DEBUG(gran, 
               DumpGranEvent(GR_DESCHEDULE, t));
       globalGranStats.tot_yields++;
-#elif defined(PAR)
+# elif defined(PAR)
       // IF_DEBUG(par, 
       // DumpGranEvent(GR_DESCHEDULE, t);
       globalParStats.tot_yields++;
-#endif
-      /* put the thread back on the run queue.  Then, if we're ready to
-       * GC, check whether this is the last task to stop.  If so, wake
-       * up the GC thread.  getThread will block during a GC until the
-       * GC is finished.
-       */
-      IF_DEBUG(scheduler,
-               if (t->what_next == ThreadEnterInterp) {
-                  /* ToDo: or maybe a timer expired when we were in Hugs?
-                   * or maybe someone hit ctrl-C
-                    */
-                   belch("--<< thread %ld (%p; %s) stopped to switch to Hugs", 
-                        t->id, t, whatNext_strs[t->what_next]);
-               } else {
-                   belch("--<< thread %ld (%p; %s) stopped, yielding", 
-                        t->id, t, whatNext_strs[t->what_next]);
-               }
-               );
-
-      threadPaused(t);
-
-      IF_DEBUG(sanity,
-              //belch("&& Doing sanity check on yielding TSO %ld.", t->id);
-              checkTSO(t));
-      ASSERT(t->link == END_TSO_QUEUE);
-#if defined(GRAN)
-      ASSERT(!is_on_queue(t,CurrentProc));
+# endif
+      break; 
 
-      IF_DEBUG(sanity,
-              //belch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
-              checkThreadQsSanity(rtsTrue));
-#endif
-#if defined(PAR)
-      if (RtsFlags.ParFlags.doFairScheduling) { 
-       /* this does round-robin scheduling; good for concurrency */
-       APPEND_TO_RUN_QUEUE(t);
-      } else {
-       /* this does unfair scheduling; good for parallelism */
-       PUSH_ON_RUN_QUEUE(t);
-      }
-#else
-      /* this does round-robin scheduling; good for concurrency */
-      APPEND_TO_RUN_QUEUE(t);
-#endif
-#if defined(GRAN)
-      /* add a ContinueThread event to actually process the thread */
-      new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
-               ContinueThread,
-               t, (StgClosure*)NULL, (rtsSpark*)NULL);
-      IF_GRAN_DEBUG(bq, 
-              belch("GRAN: eventq and runnableq after adding yielded thread to queue again:");
-              G_EVENTQ(0);
-              G_CURR_THREADQ(0));
-#endif /* GRAN */
-      break;
-      
     case ThreadBlocked:
-#if defined(GRAN)
+# if defined(GRAN)
       IF_DEBUG(scheduler,
-              belch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ", 
-                              t->id, t, whatNext_strs[t->what_next], t->block_info.closure, (t->block_info.closure==(StgClosure*)NULL ? 99 : where_is(t->block_info.closure)));
-              if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
+              debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ", 
+                         t->id, t, whatNext_strs[t->what_next], t->block_info.closure, 
+                         (t->block_info.closure==(StgClosure*)NULL ? 99 : where_is(t->block_info.closure)));
+              if (t->block_info.closure!=(StgClosure*)NULL)
+                print_bq(t->block_info.closure);
+              debugBelch("\n"));
 
       // ??? needed; should emit block before
       IF_DEBUG(gran, 
@@ -1245,162 +1498,770 @@ schedule( void )
              ((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;
-      */
+         !(!RtsFlags.GranFlags.DoAsyncFetch &&
+           procStatus[CurrentProc]==Fetching)) 
+       procStatus[CurrentProc] = Idle;
+      */
+# elif defined(PAR)
+//++PAR++  blockThread() writes the event (change?)
+# endif
+    break;
+
+  case ThreadFinished:
+    break;
+
+  default:
+    barf("parGlobalStats: unknown return code");
+    break;
+    }
+#endif
+}
+
+/* -----------------------------------------------------------------------------
+ * Handle a thread that returned to the scheduler with ThreadHeepOverflow
+ * -------------------------------------------------------------------------- */
+
+static rtsBool
+scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
+{
+    // did the task ask for a large block?
+    if (cap->r.rHpAlloc > BLOCK_SIZE) {
+       // if so, get one and push it on the front of the nursery.
+       bdescr *bd;
+       lnat blocks;
+       
+       blocks = (lnat)BLOCK_ROUND_UP(cap->r.rHpAlloc) / BLOCK_SIZE;
+       
+       IF_DEBUG(scheduler,
+                debugBelch("--<< thread %ld (%s) stopped: requesting a large block (size %ld)\n", 
+                           (long)t->id, whatNext_strs[t->what_next], blocks));
+       
+       // don't do this if the nursery is (nearly) full, we'll GC first.
+       if (cap->r.rCurrentNursery->link != NULL ||
+           cap->r.rNursery->n_blocks == 1) {  // paranoia to prevent infinite loop
+                                              // if the nursery has only one block.
+           
+           ACQUIRE_SM_LOCK
+           bd = allocGroup( blocks );
+           RELEASE_SM_LOCK
+           cap->r.rNursery->n_blocks += 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 {
+#if !defined(THREADED_RTS)
+               ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
+                      g0s0 == cap->r.rNursery);
+#endif
+               cap->r.rNursery->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 = cap->r.rNursery;
+                   x->gen_no = 0;
+                   x->flags = 0;
+               }
+           }
+           
+           // This assert can be a killer if the app is doing lots
+           // of large block allocations.
+           IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
+           
+           // 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.
+           pushOnRunQueue(cap,t);
+           return rtsFalse;  /* not actually GC'ing */
+       }
+    }
+    
+    IF_DEBUG(scheduler,
+            debugBelch("--<< thread %ld (%s) stopped: HeapOverflow\n", 
+                       (long)t->id, whatNext_strs[t->what_next]));
+#if defined(GRAN)
+    ASSERT(!is_on_queue(t,CurrentProc));
+#elif defined(PARALLEL_HASKELL)
+    /* Currently we emit a DESCHEDULE event before GC in GUM.
+       ToDo: either add separate event to distinguish SYSTEM time from rest
+       or just nuke this DESCHEDULE (and the following SCHEDULE) */
+    if (0 && RtsFlags.ParFlags.ParStats.Full) {
+       DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
+                        GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
+       emitSchedule = rtsTrue;
+    }
+#endif
+      
+    pushOnRunQueue(cap,t);
+    return rtsTrue;
+    /* actual GC is done at the end of the while loop in schedule() */
+}
+
+/* -----------------------------------------------------------------------------
+ * Handle a thread that returned to the scheduler with ThreadStackOverflow
+ * -------------------------------------------------------------------------- */
+
+static void
+scheduleHandleStackOverflow (Capability *cap, Task *task, StgTSO *t)
+{
+    IF_DEBUG(scheduler,debugBelch("--<< thread %ld (%s) stopped, StackOverflow\n", 
+                                 (long)t->id, whatNext_strs[t->what_next]));
+    /* just adjust the stack for this thread, then pop it back
+     * on the run queue.
+     */
+    { 
+       /* enlarge the stack */
+       StgTSO *new_t = threadStackOverflow(cap, t);
+       
+       /* The TSO attached to this Task may have moved, so update the
+        * pointer to it.
+        */
+       if (task->tso == t) {
+           task->tso = new_t;
+       }
+       pushOnRunQueue(cap,new_t);
+    }
+}
+
+/* -----------------------------------------------------------------------------
+ * Handle a thread that returned to the scheduler with ThreadYielding
+ * -------------------------------------------------------------------------- */
+
+static rtsBool
+scheduleHandleYield( Capability *cap, StgTSO *t, nat prev_what_next )
+{
+    // Reset the context switch flag.  We don't do this just before
+    // running the thread, because that would mean we would lose ticks
+    // during GC, which can lead to unfair scheduling (a thread hogs
+    // the CPU because the tick always arrives during GC).  This way
+    // penalises threads that do a lot of allocation, but that seems
+    // better than the alternative.
+    context_switch = 0;
+    
+    /* put the thread back on the run queue.  Then, if we're ready to
+     * GC, check whether this is the last task to stop.  If so, wake
+     * up the GC thread.  getThread will block during a GC until the
+     * GC is finished.
+     */
+    IF_DEBUG(scheduler,
+            if (t->what_next != prev_what_next) {
+                debugBelch("--<< thread %ld (%s) stopped to switch evaluators\n", 
+                           (long)t->id, whatNext_strs[t->what_next]);
+            } else {
+                debugBelch("--<< thread %ld (%s) stopped, yielding\n",
+                           (long)t->id, whatNext_strs[t->what_next]);
+            }
+       );
+    
+    IF_DEBUG(sanity,
+            //debugBelch("&& Doing sanity check on yielding TSO %ld.", t->id);
+            checkTSO(t));
+    ASSERT(t->link == END_TSO_QUEUE);
+    
+    // Shortcut if we're just switching evaluators: don't bother
+    // doing stack squeezing (which can be expensive), just run the
+    // thread.
+    if (t->what_next != prev_what_next) {
+       return rtsTrue;
+    }
+    
+#if defined(GRAN)
+    ASSERT(!is_on_queue(t,CurrentProc));
+      
+    IF_DEBUG(sanity,
+            //debugBelch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
+            checkThreadQsSanity(rtsTrue));
+
+#endif
+
+    addToRunQueue(cap,t);
+
+#if defined(GRAN)
+    /* add a ContinueThread event to actually process the thread */
+    new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
+             ContinueThread,
+             t, (StgClosure*)NULL, (rtsSpark*)NULL);
+    IF_GRAN_DEBUG(bq, 
+                 debugBelch("GRAN: eventq and runnableq after adding yielded thread to queue again:\n");
+                 G_EVENTQ(0);
+                 G_CURR_THREADQ(0));
+#endif
+    return rtsFalse;
+}
+
+/* -----------------------------------------------------------------------------
+ * Handle a thread that returned to the scheduler with ThreadBlocked
+ * -------------------------------------------------------------------------- */
+
+static void
+scheduleHandleThreadBlocked( StgTSO *t
+#if !defined(GRAN) && !defined(DEBUG)
+    STG_UNUSED
+#endif
+    )
+{
+#if defined(GRAN)
+    IF_DEBUG(scheduler,
+            debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: \n", 
+                       t->id, t, whatNext_strs[t->what_next], t->block_info.closure, (t->block_info.closure==(StgClosure*)NULL ? 99 : where_is(t->block_info.closure)));
+            if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
+    
+    // ??? needed; should emit block before
+    IF_DEBUG(gran, 
+            DumpGranEvent(GR_DESCHEDULE, t)); 
+    prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
+    /*
+      ngoq Dogh!
+      ASSERT(procStatus[CurrentProc]==Busy || 
+      ((procStatus[CurrentProc]==Fetching) && 
+      (t->block_info.closure!=(StgClosure*)NULL)));
+      if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
+      !(!RtsFlags.GranFlags.DoAsyncFetch &&
+      procStatus[CurrentProc]==Fetching)) 
+      procStatus[CurrentProc] = Idle;
+    */
 #elif defined(PAR)
-      IF_DEBUG(scheduler,
-              belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ", 
-                    t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
-      IF_PAR_DEBUG(bq,
+    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 */
 
-                  if (t->block_info.closure!=(StgClosure*)NULL) 
-                    print_bq(t->block_info.closure));
+      // We don't need to do anything.  The thread is blocked, and it
+      // has tidied up its stack and placed itself on whatever queue
+      // it needs to be on.
 
-      /* Send a fetch (if BlockedOnGA) and dump event to log file */
-      blockThread(t);
+#if !defined(THREADED_RTS)
+    ASSERT(t->why_blocked != NotBlocked);
+            // This might not be true under THREADED_RTS: we don't have
+            // exclusive access to this TSO, so someone might have
+            // woken it up by now.  This actually happens: try
+            // conc023 +RTS -N2.
+#endif
 
-      /* whatever we schedule next, we must log that schedule */
-      emitSchedule = rtsTrue;
+    IF_DEBUG(scheduler,
+            debugBelch("--<< thread %d (%s) stopped: ", 
+                       t->id, whatNext_strs[t->what_next]);
+            printThreadBlockage(t);
+            debugBelch("\n"));
+    
+    /* Only for dumping event to log file 
+       ToDo: do I need this in GranSim, too?
+       blockThread(t);
+    */
+#endif
+}
 
-#else /* !GRAN */
-      /* don't need to do anything.  Either the thread is blocked on
-       * I/O, in which case we'll have called addToBlockedQueue
-       * previously, or it's blocked on an MVar or Blackhole, in which
-       * case it'll be on the relevant queue already.
-       */
-      IF_DEBUG(scheduler,
-              fprintf(stderr, "--<< thread %d (%p) stopped: ", t->id, t);
-              printThreadBlockage(t);
-              fprintf(stderr, "\n"));
+/* -----------------------------------------------------------------------------
+ * Handle a thread that returned to the scheduler with ThreadFinished
+ * -------------------------------------------------------------------------- */
+
+static rtsBool
+scheduleHandleThreadFinished (Capability *cap STG_UNUSED, Task *task, StgTSO *t)
+{
+    /* Need to check whether this was a main thread, and if so,
+     * return with the return value.
+     *
+     * We also end up here if the thread kills itself with an
+     * uncaught exception, see Exception.cmm.
+     */
+    IF_DEBUG(scheduler,debugBelch("--++ thread %d (%s) finished\n", 
+                                 t->id, whatNext_strs[t->what_next]));
 
-      /* 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.
-       */
-      /* We also end up here if the thread kills itself with an
-       * uncaught exception, see Exception.hc.
-       */
-      IF_DEBUG(scheduler,belch("--++ thread %d (%p) finished", t->id, t));
 #if defined(GRAN)
       endThread(t, CurrentProc); // clean-up the thread
-#elif defined(PAR)
+#elif defined(PARALLEL_HASKELL)
       /* For now all are advisory -- HWL */
       //if(t->priority==AdvisoryPriority) ??
-      advisory_thread_count--;
+      advisory_thread_count--; // JB: Caution with this counter, buggy!
       
-# ifdef DIST
+# if defined(DIST)
       if(t->dist.priority==RevalPriority)
        FinishReval(t);
 # endif
-      
+    
+# if defined(EDENOLD)
+      // the thread could still have an outport... (BUG)
+      if (t->eden.outport != -1) {
+      // delete the outport for the tso which has finished...
+       IF_PAR_DEBUG(eden_ports,
+                  debugBelch("WARNING: Scheduler removes outport %d for TSO %d.\n",
+                             t->eden.outport, t->id));
+       deleteOPT(t);
+      }
+      // thread still in the process (HEAVY BUG! since outport has just been closed...)
+      if (t->eden.epid != -1) {
+       IF_PAR_DEBUG(eden_ports,
+                  debugBelch("WARNING: Scheduler removes TSO %d from process %d .\n",
+                          t->id, t->eden.epid));
+       removeTSOfromProcess(t);
+      }
+# endif 
+
+# if defined(PAR)
       if (RtsFlags.ParFlags.ParStats.Full &&
          !RtsFlags.ParFlags.ParStats.Suppressed) 
        DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
+
+      //  t->par only contains statistics: left out for now...
+      IF_PAR_DEBUG(fish,
+                  debugBelch("**** end thread: ended sparked thread %d (%lx); sparkname: %lx\n",
+                             t->id,t,t->par.sparkname));
+# endif
+#endif // PARALLEL_HASKELL
+
+      //
+      // Check whether the thread that just completed was a bound
+      // 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 (t->bound) {
+
+         if (t->bound != task) {
+#if !defined(THREADED_RTS)
+             // Must be a bound 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).
+             appendToRunQueue(cap,t);
+             return rtsFalse;
+#else
+             // this cannot happen in the threaded RTS, because a
+             // bound thread can only be run by the appropriate Task.
+             barf("finished bound thread that isn't mine");
 #endif
-      break;
-      
-    default:
-      barf("schedule: invalid thread return code %d", (int)ret);
-    }
-    
-#if defined(RTS_SUPPORTS_THREADS)
-    /* I don't understand what this re-grab is doing -- sof */
-    grabCapability(&cap);
+         }
+
+         ASSERT(task->tso == t);
+
+         if (t->what_next == ThreadComplete) {
+             if (task->ret) {
+                 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
+                 *(task->ret) = (StgClosure *)task->tso->sp[1]; 
+             }
+             task->stat = Success;
+         } else {
+             if (task->ret) {
+                 *(task->ret) = NULL;
+             }
+             if (sched_state >= SCHED_INTERRUPTING) {
+                 task->stat = Interrupted;
+             } else {
+                 task->stat = Killed;
+             }
+         }
+#ifdef DEBUG
+         removeThreadLabel((StgWord)task->tso->id);
 #endif
+         return rtsTrue; // tells schedule() to return
+      }
 
-#ifdef PROFILING
-    if (RtsFlags.ProfFlags.profileInterval==0 || performHeapProfile) {
+      return rtsFalse;
+}
+
+/* -----------------------------------------------------------------------------
+ * Perform a heap census, if PROFILING
+ * -------------------------------------------------------------------------- */
+
+static rtsBool
+scheduleDoHeapProfile( rtsBool ready_to_gc STG_UNUSED )
+{
+#if defined(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)) {
+
+       // checking black holes is necessary before GC, otherwise
+       // there may be threads that are unreachable except by the
+       // blackhole queue, which the GC will consider to be
+       // deadlocked.
+       scheduleCheckBlackHoles(&MainCapability);
+
+       IF_DEBUG(scheduler, sched_belch("garbage collecting before heap census"));
        GarbageCollect(GetRoots, rtsTrue);
+
+       IF_DEBUG(scheduler, sched_belch("performing heap census"));
        heapCensus();
+
        performHeapProfile = rtsFalse;
-       ready_to_gc = rtsFalse; // we already GC'd
+       return rtsTrue;  // true <=> we already GC'd
     }
 #endif
+    return rtsFalse;
+}
+
+/* -----------------------------------------------------------------------------
+ * Perform a garbage collection if necessary
+ * -------------------------------------------------------------------------- */
 
-    if (ready_to_gc 
-#ifdef SMP
-       && allFreeCapabilities() 
+static Capability *
+scheduleDoGC (Capability *cap, Task *task USED_IF_THREADS,
+             rtsBool force_major, void (*get_roots)(evac_fn))
+{
+    StgTSO *t;
+#ifdef THREADED_RTS
+    static volatile StgWord waiting_for_gc;
+    rtsBool was_waiting;
+    nat i;
 #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.
-       */
-#if defined(RTS_SUPPORTS_THREADS)
-      IF_DEBUG(scheduler,sched_belch("doing GC"));
+
+#ifdef THREADED_RTS
+    // In order to GC, there must be no threads running Haskell code.
+    // Therefore, the GC thread needs to hold *all* the capabilities,
+    // and release them after the GC has completed.  
+    //
+    // This seems to be the simplest way: previous attempts involved
+    // making all the threads with capabilities give up their
+    // capabilities and sleep except for the *last* one, which
+    // actually did the GC.  But it's quite hard to arrange for all
+    // the other tasks to sleep and stay asleep.
+    //
+       
+    was_waiting = cas(&waiting_for_gc, 0, 1);
+    if (was_waiting) {
+       do {
+           IF_DEBUG(scheduler, sched_belch("someone else is trying to GC..."));
+           if (cap) yieldCapability(&cap,task);
+       } while (waiting_for_gc);
+       return cap;  // NOTE: task->cap might have changed here
+    }
+
+    for (i=0; i < n_capabilities; i++) {
+       IF_DEBUG(scheduler, sched_belch("ready_to_gc, grabbing all the capabilies (%d/%d)", i, n_capabilities));
+       if (cap != &capabilities[i]) {
+           Capability *pcap = &capabilities[i];
+           // we better hope this task doesn't get migrated to
+           // another Capability while we're waiting for this one.
+           // It won't, because load balancing happens while we have
+           // all the Capabilities, but even so it's a slightly
+           // unsavoury invariant.
+           task->cap = pcap;
+           context_switch = 1;
+           waitForReturnCapability(&pcap, task);
+           if (pcap != &capabilities[i]) {
+               barf("scheduleDoGC: got the wrong capability");
+           }
+       }
+    }
+
+    waiting_for_gc = rtsFalse;
 #endif
-      GarbageCollect(GetRoots,rtsFalse);
-      ready_to_gc = rtsFalse;
-#ifdef SMP
-      broadcastCondition(&gc_pending_cond);
+
+    /* 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.
+     */
+    { 
+       StgTSO *next;
+
+       for (t = all_threads; t != END_TSO_QUEUE; t = next) {
+           if (t->what_next == ThreadRelocated) {
+               next = t->link;
+           } else {
+               next = t->global_link;
+               if (t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
+                   if (!stmValidateNestOfTransactions (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_(&capabilities[0], t, NULL, rtsTrue, NULL);
+                       
+#ifdef REG_R1
+                       ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
 #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, 
-              fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n");
-              G_EVENTQ(0);
-              G_CURR_THREADQ(0));
-#endif /* GRAN */
+                   }
+               }
+           }
+       }
     }
+    
+    // so this happens periodically:
+    if (cap) scheduleCheckBlackHoles(cap);
+    
+    IF_DEBUG(scheduler, printAllThreads());
 
-#if defined(GRAN)
-  next_thread:
-    IF_GRAN_DEBUG(unused,
-                 print_eventq(EventHd));
+    /*
+     * We now have all the capabilities; if we're in an interrupting
+     * state, then we should take the opportunity to delete all the
+     * threads in the system.
+     */
+    if (sched_state >= SCHED_INTERRUPTING) {
+       deleteAllThreads(&capabilities[0]);
+       sched_state = SCHED_INTERRUPTED;
+    }
 
-    event = get_next_event();
-#elif defined(PAR)
-  next_thread:
-    /* ToDo: wait for next message to arrive rather than busy wait */
+    /* everybody back, start the GC.
+     * Could do it in this thread, or signal a condition var
+     * to do it in another thread.  Either way, we need to
+     * broadcast on gc_pending_cond afterward.
+     */
+#if defined(THREADED_RTS)
+    IF_DEBUG(scheduler,sched_belch("doing GC"));
+#endif
+    GarbageCollect(get_roots, force_major);
+    
+#if defined(THREADED_RTS)
+    // release our stash of capabilities.
+    for (i = 0; i < n_capabilities; i++) {
+       if (cap != &capabilities[i]) {
+           task->cap = &capabilities[i];
+           releaseCapability(&capabilities[i]);
+       }
+    }
+    if (cap) {
+       task->cap = cap;
+    } else {
+       task->cap = NULL;
+    }
+#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 */
 
-  } /* end of while(1) */
+    return cap;
+}
 
-  IF_PAR_DEBUG(verbose,
-              belch("== Leaving schedule() after having received Finish"));
+/* ---------------------------------------------------------------------------
+ * rtsSupportsBoundThreads(): is the RTS built to support bound threads?
+ * used by Control.Concurrent for error checking.
+ * ------------------------------------------------------------------------- */
+StgBool
+rtsSupportsBoundThreads(void)
+{
+#if defined(THREADED_RTS)
+  return rtsTrue;
+#else
+  return rtsFalse;
+#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.
+ * isThreadBound(tso): check whether tso is bound to an OS thread.
+ * ------------------------------------------------------------------------- */
+StgBool
+isThreadBound(StgTSO* tso USED_IF_THREADS)
+{
+#if defined(THREADED_RTS)
+  return (tso->bound != NULL);
+#endif
+  return rtsFalse;
+}
+
+/* ---------------------------------------------------------------------------
+ * Singleton fork(). Do not copy any running threads.
+ * ------------------------------------------------------------------------- */
+
+#if !defined(mingw32_HOST_OS)
+#define FORKPROCESS_PRIMOP_SUPPORTED
+#endif
+
+#ifdef FORKPROCESS_PRIMOP_SUPPORTED
+static void 
+deleteThread_(Capability *cap, StgTSO *tso);
+#endif
+StgInt
+forkProcess(HsStablePtr *entry
+#ifndef FORKPROCESS_PRIMOP_SUPPORTED
+           STG_UNUSED
+#endif
+           )
+{
+#ifdef FORKPROCESS_PRIMOP_SUPPORTED
+    Task *task;
+    pid_t pid;
+    StgTSO* t,*next;
+    Capability *cap;
+    
+#if defined(THREADED_RTS)
+    if (RtsFlags.ParFlags.nNodes > 1) {
+       errorBelch("forking not supported with +RTS -N<n> greater than 1");
+       stg_exit(EXIT_FAILURE);
+    }
+#endif
+
+    IF_DEBUG(scheduler,sched_belch("forking!"));
+    
+    // ToDo: for SMP, we should probably acquire *all* the capabilities
+    cap = rts_lock();
+    
+    pid = fork();
+    
+    if (pid) { // parent
+       
+       // just return the pid
+       rts_unlock(cap);
+       return pid;
+       
+    } else { // child
+       
+       // Now, all OS threads except the thread that forked are
+       // stopped.  We need to stop all Haskell threads, including
+       // those involved in foreign calls.  Also we need to delete
+       // all Tasks, because they correspond to OS threads that are
+       // now gone.
+
+       for (t = all_threads; t != END_TSO_QUEUE; t = next) {
+           next = t->global_link;
+           // don't allow threads to catch the ThreadKilled
+           // exception, but we do want to raiseAsync() because these
+           // threads may be evaluating thunks that we need later.
+           deleteThread_(cap,t);
+       }
+       
+       // Empty the run queue.  It seems tempting to let all the
+       // killed threads stay on the run queue as zombies to be
+       // cleaned up later, but some of them correspond to bound
+       // threads for which the corresponding Task does not exist.
+       cap->run_queue_hd = END_TSO_QUEUE;
+       cap->run_queue_tl = END_TSO_QUEUE;
+
+       // Any suspended C-calling Tasks are no more, their OS threads
+       // don't exist now:
+       cap->suspended_ccalling_tasks = NULL;
+
+       // Empty the all_threads list.  Otherwise, the garbage
+       // collector may attempt to resurrect some of these threads.
+       all_threads = END_TSO_QUEUE;
+
+       // Wipe the task list, except the current Task.
+       ACQUIRE_LOCK(&sched_mutex);
+       for (task = all_tasks; task != NULL; task=task->all_link) {
+           if (task != cap->running_task) {
+               discardTask(task);
+           }
+       }
+       RELEASE_LOCK(&sched_mutex);
+
+#if defined(THREADED_RTS)
+       // Wipe our spare workers list, they no longer exist.  New
+       // workers will be created if necessary.
+       cap->spare_workers = NULL;
+       cap->returning_tasks_hd = NULL;
+       cap->returning_tasks_tl = NULL;
+#endif
+
+       cap = rts_evalStableIO(cap, entry, NULL);  // run the action
+       rts_checkSchedStatus("forkProcess",cap);
+       
+       rts_unlock(cap);
+       hs_exit();                      // clean up and exit
+       stg_exit(EXIT_SUCCESS);
+    }
+#else /* !FORKPROCESS_PRIMOP_SUPPORTED */
+    barf("forkProcess#: primop not supported on this platform, sorry!\n");
+    return -1;
+#endif
+}
+
+/* ---------------------------------------------------------------------------
+ * Delete all the threads in the system
  * ------------------------------------------------------------------------- */
    
-void deleteAllThreads ( void )
+static void
+deleteAllThreads ( Capability *cap )
 {
   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);
+      if (t->what_next == ThreadRelocated) {
+         next = t->link;
+      } else {
+         next = t->global_link;
+         deleteThread(cap,t);
+      }
   }      
-  run_queue_hd = run_queue_tl = END_TSO_QUEUE;
-  blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
-  sleeping_queue = END_TSO_QUEUE;
+
+  // The run queue now contains a bunch of ThreadKilled threads.  We
+  // must not throw these away: the main thread(s) will be in there
+  // somewhere, and the main scheduler loop has to deal with it.
+  // Also, the run queue is the only thing keeping these threads from
+  // being GC'd, and we don't want the "main thread has been GC'd" panic.
+
+#if !defined(THREADED_RTS)
+  ASSERT(blocked_queue_hd == END_TSO_QUEUE);
+  ASSERT(sleeping_queue == END_TSO_QUEUE);
+#endif
 }
 
-/* startThread and  insertThread are now in GranSim.c -- HWL */
+/* -----------------------------------------------------------------------------
+   Managing the suspended_ccalling_tasks list.
+   Locks required: sched_mutex
+   -------------------------------------------------------------------------- */
 
+STATIC_INLINE void
+suspendTask (Capability *cap, Task *task)
+{
+    ASSERT(task->next == NULL && task->prev == NULL);
+    task->next = cap->suspended_ccalling_tasks;
+    task->prev = NULL;
+    if (cap->suspended_ccalling_tasks) {
+       cap->suspended_ccalling_tasks->prev = task;
+    }
+    cap->suspended_ccalling_tasks = task;
+}
 
-//@node Suspend and Resume, Run queue code, Main scheduling loop, Main scheduling code
-//@subsection Suspend and Resume
+STATIC_INLINE void
+recoverSuspendedTask (Capability *cap, Task *task)
+{
+    if (task->prev) {
+       task->prev->next = task->next;
+    } else {
+       ASSERT(cap->suspended_ccalling_tasks == task);
+       cap->suspended_ccalling_tasks = task->next;
+    }
+    if (task->next) {
+       task->next->prev = task->prev;
+    }
+    task->next = task->prev = NULL;
+}
 
 /* ---------------------------------------------------------------------------
  * Suspending & resuming Haskell threads.
@@ -1417,111 +2278,100 @@ void deleteAllThreads ( void )
  * on return from the C function.
  * ------------------------------------------------------------------------- */
    
-StgInt
-suspendThread( StgRegTable *reg, 
-              rtsBool concCall
-#if !defined(RTS_SUPPORTS_THREADS)
-              STG_UNUSED
-#endif
-              )
+void *
+suspendThread (StgRegTable *reg)
 {
-  nat tok;
   Capability *cap;
+  int saved_errno = errno;
+  StgTSO *tso;
+  Task *task;
 
-  /* assume that *reg is a pointer to the StgRegTable part
-   * of a Capability.
+  /* assume that *reg is a pointer to the StgRegTable part of a Capability.
    */
-  cap = (Capability *)((void *)reg - sizeof(StgFunTable));
+  cap = regTableToCapability(reg);
 
-  ACQUIRE_LOCK(&sched_mutex);
+  task = cap->running_task;
+  tso = cap->r.rCurrentTSO;
 
   IF_DEBUG(scheduler,
-          sched_belch("thread %d did a _ccall_gc", cap->r.rCurrentTSO->id));
+          sched_belch("thread %d did a safe foreign call", cap->r.rCurrentTSO->id));
 
-  threadPaused(cap->r.rCurrentTSO);
-  cap->r.rCurrentTSO->link = suspended_ccalling_threads;
-  suspended_ccalling_threads = cap->r.rCurrentTSO;
+  // XXX this might not be necessary --SDM
+  tso->what_next = ThreadRunGHC;
 
-#if defined(RTS_SUPPORTS_THREADS)
-  cap->r.rCurrentTSO->why_blocked  = BlockedOnCCall;
-#endif
+  threadPaused(cap,tso);
 
-  /* Use the thread ID as the token; it should be unique */
-  tok = cap->r.rCurrentTSO->id;
+  if(tso->blocked_exceptions == NULL)  {
+      tso->why_blocked = BlockedOnCCall;
+      tso->blocked_exceptions = END_TSO_QUEUE;
+  } else {
+      tso->why_blocked = BlockedOnCCall_NoUnblockExc;
+  }
+
+  // Hand back capability
+  task->suspended_tso = tso;
+
+  ACQUIRE_LOCK(&cap->lock);
 
-  /* Hand back capability */
-  releaseCapability(cap);
+  suspendTask(cap,task);
+  cap->in_haskell = rtsFalse;
+  releaseCapability_(cap);
   
-#if defined(RTS_SUPPORTS_THREADS)
+  RELEASE_LOCK(&cap->lock);
+
+#if defined(THREADED_RTS)
   /* Preparing to leave the RTS, so ensure there's a native thread/task
      waiting to take over.
-     
-     ToDo: optimise this and only create a new task if there's a need
-     for one (i.e., if there's only one Concurrent Haskell thread alive,
-     there's no need to create a new task).
   */
-  IF_DEBUG(scheduler, sched_belch("worker thread (%d): leaving RTS", tok));
-  if (concCall) {
-    startTask(taskStart);
-  }
+  IF_DEBUG(scheduler, sched_belch("thread %d: leaving RTS", tso->id));
 #endif
 
-  /* Other threads _might_ be available for execution; signal this */
-  THREAD_RUNNABLE();
-  RELEASE_LOCK(&sched_mutex);
-  return tok; 
+  errno = saved_errno;
+  return task;
 }
 
 StgRegTable *
-resumeThread( StgInt tok,
-             rtsBool concCall
-#if !defined(RTS_SUPPORTS_THREADS)
-              STG_UNUSED
-#endif
-             )
+resumeThread (void *task_)
 {
-  StgTSO *tso, **prev;
-  Capability *cap;
-
-#if defined(RTS_SUPPORTS_THREADS)
-  /* Wait for permission to re-enter the RTS with the result. */
-  if ( concCall ) {
-    grabReturnCapability(&sched_mutex, &cap);
-  } else {
-    grabCapability(&cap);
-  }
-#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; 
-       prev = &tso->link, tso = tso->link) {
-    if (tso->id == (StgThreadID)tok) {
-      *prev = tso->link;
-      break;
+    StgTSO *tso;
+    Capability *cap;
+    int saved_errno = errno;
+    Task *task = task_;
+
+    cap = task->cap;
+    // Wait for permission to re-enter the RTS with the result.
+    waitForReturnCapability(&cap,task);
+    // we might be on a different capability now... but if so, our
+    // entry on the suspended_ccalling_tasks list will also have been
+    // migrated.
+
+    // Remove the thread from the suspended list
+    recoverSuspendedTask(cap,task);
+
+    tso = task->suspended_tso;
+    task->suspended_tso = NULL;
+    tso->link = END_TSO_QUEUE;
+    IF_DEBUG(scheduler, sched_belch("thread %d: re-entering RTS", tso->id));
+    
+    if (tso->why_blocked == BlockedOnCCall) {
+       awakenBlockedQueue(cap,tso->blocked_exceptions);
+       tso->blocked_exceptions = NULL;
     }
-  }
-  if (tso == END_TSO_QUEUE) {
-    barf("resumeThread: thread not found");
-  }
-  tso->link = END_TSO_QUEUE;
-  /* Reset blocking status */
-  tso->why_blocked  = NotBlocked;
-
-  RELEASE_LOCK(&sched_mutex);
+    
+    /* Reset blocking status */
+    tso->why_blocked  = NotBlocked;
+    
+    cap->r.rCurrentTSO = tso;
+    cap->in_haskell = rtsTrue;
+    errno = saved_errno;
 
-  cap->r.rCurrentTSO = tso;
-  return &cap->r;
-}
+    /* We might have GC'd, mark the TSO dirty again */
+    dirtyTSO(tso);
 
+    IF_DEBUG(sanity, checkTSO(tso));
 
-/* ---------------------------------------------------------------------------
- * Static functions
- * ------------------------------------------------------------------------ */
-static void unblockThread(StgTSO *tso);
+    return &cap->r;
+}
 
 /* ---------------------------------------------------------------------------
  * Comparing Thread ids.
@@ -1530,10 +2380,11 @@ 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;
@@ -1545,10 +2396,27 @@ int cmp_thread(const StgTSO *tso1, const StgTSO *tso2)
  *
  * This is used in the implementation of Show for ThreadIds.
  * ------------------------------------------------------------------------ */
-int rts_getThreadId(const StgTSO *tso) 
+int
+rts_getThreadId(StgPtr tso) 
+{
+  return ((StgTSO *)tso)->id;
+}
+
+#ifdef DEBUG
+void
+labelThread(StgPtr tso, char *label)
 {
-  return tso->id;
+  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.
@@ -1563,192 +2431,181 @@ int rts_getThreadId(const StgTSO *tso)
 
    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 */
 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)
-{
-  return createThread_(stack_size, rtsFalse);
-}
-
-static StgTSO *
-createThread_(nat size, rtsBool have_lock)
-{
+createThread(Capability *cap, nat size)
 #endif
-
+{
     StgTSO *tso;
     nat stack_size;
 
+    /* sched_mutex is *not* required */
+
     /* First check whether we should create a thread at all */
-#if defined(PAR)
-  /* 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)",
-         RtsFlags.ParFlags.maxThreads, advisory_thread_count);
-    return END_TSO_QUEUE;
-  }
-  threadsCreated++;
+#if defined(PARALLEL_HASKELL)
+    /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
+    if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
+       threadsIgnored++;
+       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;
+    }
+    threadsCreated++;
 #endif
 
 #if defined(GRAN)
-  ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
+    ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
 #endif
 
-  // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
-
-  /* catch ridiculously small stack sizes */
-  if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
-    size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
-  }
+    // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
 
-  stack_size = size - TSO_STRUCT_SIZEW;
+    /* catch ridiculously small stack sizes */
+    if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
+       size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
+    }
 
-  tso = (StgTSO *)allocate(size);
-  TICK_ALLOC_TSO(stack_size, 0);
+    stack_size = size - TSO_STRUCT_SIZEW;
+    
+    tso = (StgTSO *)allocateLocal(cap, size);
+    TICK_ALLOC_TSO(stack_size, 0);
 
-  SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
+    SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
 #if defined(GRAN)
-  SET_GRAN_HDR(tso, ThisPE);
+    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); }
 
-  tso->why_blocked  = NotBlocked;
-  tso->blocked_exceptions = NULL;
+    // Always start with the compiled code evaluator
+    tso->what_next = ThreadRunGHC;
 
-  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->why_blocked  = NotBlocked;
+    tso->blocked_exceptions = NULL;
+    tso->flags = TSO_DIRTY;
+    
+    tso->saved_errno = 0;
+    tso->bound = 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;
+    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_SYSTEM);
-  tso->su = (StgUpdateFrame*)tso->sp;
-
+    tso->sp -= sizeofW(StgStopFrame);
+    SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
+    tso->link = END_TSO_QUEUE;
+    
   // ToDo: check this
 #if defined(GRAN)
-  tso->link = END_TSO_QUEUE;
-  /* uses more flexible routine in GranSim */
-  insertThread(tso, CurrentProc);
+    /* uses more flexible routine in GranSim */
+    insertThread(tso, CurrentProc);
 #else
-  /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
-   * from its creation
-   */
+    /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
+     * from its creation
+     */
 #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; */
+    if (RtsFlags.GranFlags.GranSimStats.Full) 
+       DumpGranEvent(GR_START,tso);
+#elif defined(PARALLEL_HASKELL)
+    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.
-   */
-  tso->global_link = all_threads;
-  all_threads = tso;
-
+    
+    /* Link the new thread on the global thread list.
+     */
+    ACQUIRE_LOCK(&sched_mutex);
+    tso->id = next_thread_id++;  // while we have the mutex
+    tso->global_link = all_threads;
+    all_threads = tso;
+    RELEASE_LOCK(&sched_mutex);
+    
 #if defined(DIST)
-  tso->dist.priority = MandatoryPriority; //by default that is...
+    tso->dist.priority = MandatoryPriority; //by default that is...
 #endif
-
+    
 #if defined(GRAN)
-  tso->gran.pri = pri;
+    tso->gran.pri = pri;
 # if defined(DEBUG)
-  tso->gran.magic = TSO_MAGIC; // debugging only
+    tso->gran.magic = TSO_MAGIC; // debugging only
 # endif
-  tso->gran.sparkname   = 0;
-  tso->gran.startedat   = CURRENT_TIME; 
-  tso->gran.exported    = 0;
-  tso->gran.basicblocks = 0;
-  tso->gran.allocs      = 0;
-  tso->gran.exectime    = 0;
-  tso->gran.fetchtime   = 0;
-  tso->gran.fetchcount  = 0;
-  tso->gran.blocktime   = 0;
-  tso->gran.blockcount  = 0;
-  tso->gran.blockedat   = 0;
-  tso->gran.globalsparks = 0;
-  tso->gran.localsparks  = 0;
-  if (RtsFlags.GranFlags.Light)
-    tso->gran.clock  = Now; /* local clock */
-  else
-    tso->gran.clock  = 0;
-
-  IF_DEBUG(gran,printTSO(tso));
-#elif defined(PAR)
+    tso->gran.sparkname   = 0;
+    tso->gran.startedat   = CURRENT_TIME; 
+    tso->gran.exported    = 0;
+    tso->gran.basicblocks = 0;
+    tso->gran.allocs      = 0;
+    tso->gran.exectime    = 0;
+    tso->gran.fetchtime   = 0;
+    tso->gran.fetchcount  = 0;
+    tso->gran.blocktime   = 0;
+    tso->gran.blockcount  = 0;
+    tso->gran.blockedat   = 0;
+    tso->gran.globalsparks = 0;
+    tso->gran.localsparks  = 0;
+    if (RtsFlags.GranFlags.Light)
+       tso->gran.clock  = Now; /* local clock */
+    else
+       tso->gran.clock  = 0;
+    
+    IF_DEBUG(gran,printTSO(tso));
+#elif defined(PARALLEL_HASKELL)
 # if defined(DEBUG)
-  tso->par.magic = TSO_MAGIC; // debugging only
+    tso->par.magic = TSO_MAGIC; // debugging only
 # endif
-  tso->par.sparkname   = 0;
-  tso->par.startedat   = CURRENT_TIME; 
-  tso->par.exported    = 0;
-  tso->par.basicblocks = 0;
-  tso->par.allocs      = 0;
-  tso->par.exectime    = 0;
-  tso->par.fetchtime   = 0;
-  tso->par.fetchcount  = 0;
-  tso->par.blocktime   = 0;
-  tso->par.blockcount  = 0;
-  tso->par.blockedat   = 0;
-  tso->par.globalsparks = 0;
-  tso->par.localsparks  = 0;
+    tso->par.sparkname   = 0;
+    tso->par.startedat   = CURRENT_TIME; 
+    tso->par.exported    = 0;
+    tso->par.basicblocks = 0;
+    tso->par.allocs      = 0;
+    tso->par.exectime    = 0;
+    tso->par.fetchtime   = 0;
+    tso->par.fetchcount  = 0;
+    tso->par.blocktime   = 0;
+    tso->par.blockcount  = 0;
+    tso->par.blockedat   = 0;
+    tso->par.globalsparks = 0;
+    tso->par.localsparks  = 0;
 #endif
-
+    
 #if defined(GRAN)
-  globalGranStats.tot_threads_created++;
-  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) {
-    //fprintf(stderr, "Creating thread %d @ %11.2f\n", tso->id, usertime()); 
-    globalParStats.tot_threads_created++;
-  }
+    globalGranStats.tot_threads_created++;
+    globalGranStats.threads_created_on_PE[CurrentProc]++;
+    globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
+    globalGranStats.tot_sq_probes++;
+#elif defined(PARALLEL_HASKELL)
+    // collect parallel global statistics (currently done together with GC stats)
+    if (RtsFlags.ParFlags.ParStats.Global &&
+       RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
+       //debugBelch("Creating thread %d @ %11.2f\n", tso->id, usertime()); 
+       globalParStats.tot_threads_created++;
+    }
 #endif 
-
+    
 #if defined(GRAN)
-  IF_GRAN_DEBUG(pri,
-               belch("==__ schedule: Created TSO %d (%p);",
-                     CurrentProc, tso, tso->id));
-#elif defined(PAR)
+    IF_GRAN_DEBUG(pri,
+                 sched_belch("==__ schedule: Created TSO %d (%p);",
+                             CurrentProc, tso, tso->id));
+#elif defined(PARALLEL_HASKELL)
     IF_PAR_DEBUG(verbose,
-                belch("==__ schedule: Created TSO %d (%p); %d threads active",
-                      tso->id, tso, advisory_thread_count));
+                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));
+    IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words", 
+                                  (long)tso->id, (long)tso->stack_size));
 #endif    
-  return tso;
+    return tso;
 }
 
 #if defined(PAR)
@@ -1756,9 +2613,10 @@ createThread_(nat size, rtsBool have_lock)
    all parallel thread creation calls should fall through the following routine.
 */
 StgTSO *
-createSparkThread(rtsSpark spark) 
+createThreadFromSpark(rtsSpark spark) 
 { StgTSO *tso;
   ASSERT(spark != (rtsSpark)NULL);
+// JB: TAKE CARE OF THIS COUNTER! BUGGY
   if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) 
   { threadsIgnored++;
     barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
@@ -1767,15 +2625,15 @@ createSparkThread(rtsSpark spark)
   }
   else
   { threadsCreated++;
-    tso = createThread_(RtsFlags.GcFlags.initialStkSize, rtsTrue);
+    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++;    
+    addToRunQueue(tso);
+    advisory_thread_count++;  // JB: TAKE CARE OF THIS COUNTER! BUGGY
   }
   return tso;
 }
@@ -1785,8 +2643,7 @@ createSparkThread(rtsSpark spark)
   Turn a spark into a thread.
   ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
 */
-#if defined(PAR)
-//@cindex activateSpark
+#if 0
 StgTSO *
 activateSpark (rtsSpark spark) 
 {
@@ -1795,9 +2652,9 @@ activateSpark (rtsSpark spark)
   tso = createSparkThread(spark);
   if (RtsFlags.ParFlags.ParStats.Full) {   
     //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
-    IF_PAR_DEBUG(verbose,
-                belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
-                      (StgClosure *)spark, info_type((StgClosure *)spark)));
+      IF_PAR_DEBUG(verbose,
+                  debugBelch("==^^ activateSpark: turning spark of closure %p (%s) into a thread\n",
+                             (StgClosure *)spark, info_type((StgClosure *)spark)));
   }
   // ToDo: fwd info on local/global spark to thread -- HWL
   // tso->gran.exported =  spark->exported;
@@ -1811,56 +2668,83 @@ activateSpark (rtsSpark spark)
 /* ---------------------------------------------------------------------------
  * scheduleThread()
  *
- * scheduleThread puts a thread on the head of the runnable queue.
+ * scheduleThread puts a thread on the end  of the runnable queue.
  * This will usually be done immediately after a thread is created.
  * The caller of scheduleThread must create the thread using e.g.
  * createThread and push an appropriate closure
  * on this thread's stack before the scheduler is invoked.
  * ------------------------------------------------------------------------ */
 
-static void scheduleThread_ (StgTSO* tso, rtsBool createTask);
-
 void
-scheduleThread_(StgTSO *tso
-              , rtsBool createTask
-#if !defined(THREADED_RTS)
-                STG_UNUSED
-#endif
-             )
+scheduleThread(Capability *cap, StgTSO *tso)
 {
-  ACQUIRE_LOCK(&sched_mutex);
+    // The thread goes at the *end* of the run-queue, to avoid possible
+    // starvation of any threads already on the queue.
+    appendToRunQueue(cap,tso);
+}
 
-  /* 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);
-#if defined(THREADED_RTS)
-  /* If main() is scheduling a thread, don't bother creating a 
-   * new task.
-   */
-  if ( createTask ) {
-    startTask(taskStart);
-  }
-#endif
-  THREAD_RUNNABLE();
+Capability *
+scheduleWaitThread (StgTSO* tso, /*[out]*/HaskellObj* ret, Capability *cap)
+{
+    Task *task;
 
-#if 0
-  IF_DEBUG(scheduler,printTSO(tso));
+    // We already created/initialised the Task
+    task = cap->running_task;
+
+    // This TSO is now a bound thread; make the Task and TSO
+    // point to each other.
+    tso->bound = task;
+
+    task->tso = tso;
+    task->ret = ret;
+    task->stat = NoStatus;
+
+    appendToRunQueue(cap,tso);
+
+    IF_DEBUG(scheduler, sched_belch("new bound thread (%d)", tso->id));
+
+#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
 #endif
-  RELEASE_LOCK(&sched_mutex);
-}
 
-void scheduleThread(StgTSO* tso)
-{
-  return scheduleThread_(tso, rtsFalse);
+    cap = schedule(cap,task);
+
+    ASSERT(task->stat != NoStatus);
+    ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
+
+    IF_DEBUG(scheduler, sched_belch("bound thread (%d) finished", task->tso->id));
+    return cap;
 }
 
-void scheduleExtThread(StgTSO* tso)
+/* ----------------------------------------------------------------------------
+ * Starting Tasks
+ * ------------------------------------------------------------------------- */
+
+#if defined(THREADED_RTS)
+void
+workerStart(Task *task)
 {
-  return scheduleThread_(tso, rtsTrue);
+    Capability *cap;
+
+    // See startWorkerTask().
+    ACQUIRE_LOCK(&task->lock);
+    cap = task->cap;
+    RELEASE_LOCK(&task->lock);
+
+    // set the thread-local pointer to the Task:
+    taskEnter(task);
+
+    // schedule() runs without a lock.
+    cap = schedule(cap,task);
+
+    // On exit from schedule(), we have a Capability.
+    releaseCapability(cap);
+    taskStop(task);
 }
+#endif
 
 /* ---------------------------------------------------------------------------
  * initScheduler()
@@ -1871,364 +2755,108 @@ void scheduleExtThread(StgTSO* tso)
  *
  * ------------------------------------------------------------------------ */
 
-#ifdef SMP
-static void
-term_handler(int sig STG_UNUSED)
-{
-  stat_workerStop();
-  ACQUIRE_LOCK(&term_mutex);
-  await_death--;
-  RELEASE_LOCK(&term_mutex);
-  shutdownThread();
-}
-#endif
-
 void 
 initScheduler(void)
 {
 #if defined(GRAN)
   nat i;
-
   for (i=0; i<=MAX_PROC; i++) {
     run_queue_hds[i]      = END_TSO_QUEUE;
     run_queue_tls[i]      = END_TSO_QUEUE;
     blocked_queue_hds[i]  = END_TSO_QUEUE;
     blocked_queue_tls[i]  = END_TSO_QUEUE;
     ccalling_threadss[i]  = END_TSO_QUEUE;
+    blackhole_queue[i]    = END_TSO_QUEUE;
     sleeping_queue        = END_TSO_QUEUE;
   }
-#else
-  run_queue_hd      = END_TSO_QUEUE;
-  run_queue_tl      = END_TSO_QUEUE;
+#elif !defined(THREADED_RTS)
   blocked_queue_hd  = END_TSO_QUEUE;
   blocked_queue_tl  = END_TSO_QUEUE;
   sleeping_queue    = END_TSO_QUEUE;
-#endif 
-
-  suspended_ccalling_threads  = END_TSO_QUEUE;
+#endif
 
-  main_threads = NULL;
-  all_threads  = END_TSO_QUEUE;
+  blackhole_queue   = END_TSO_QUEUE;
+  all_threads       = END_TSO_QUEUE;
 
   context_switch = 0;
-  interrupted    = 0;
+  sched_state    = SCHED_RUNNING;
 
   RtsFlags.ConcFlags.ctxtSwitchTicks =
       RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
       
-#if defined(RTS_SUPPORTS_THREADS)
+#if defined(THREADED_RTS)
   /* Initialise the mutex and condition variables used by
    * the scheduler. */
   initMutex(&sched_mutex);
-  initMutex(&term_mutex);
-
-  initCondition(&thread_ready_cond);
 #endif
   
-#if defined(SMP)
-  initCondition(&gc_pending_cond);
-#endif
-
-#if defined(RTS_SUPPORTS_THREADS)
   ACQUIRE_LOCK(&sched_mutex);
-#endif
-
-  /* Install the SIGHUP handler */
-#if defined(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");
-    }
-  }
-#endif
 
   /* 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).
+   * floating around (only THREADED_RTS builds have more than one).
    */
   initCapabilities();
-  
-#if defined(RTS_SUPPORTS_THREADS)
-    /* start our haskell execution tasks */
-# if defined(SMP)
-    startTaskManager(RtsFlags.ParFlags.nNodes, taskStart);
-# else
-    startTaskManager(0,taskStart);
-# endif
-#endif
 
-#if /* defined(SMP) ||*/ defined(PAR)
+  initTaskManager();
+
+#if defined(THREADED_RTS) || defined(PARALLEL_HASKELL)
   initSparkPools();
 #endif
 
-#if defined(RTS_SUPPORTS_THREADS)
-  RELEASE_LOCK(&sched_mutex);
+#if defined(THREADED_RTS)
+  /*
+   * Eagerly start one worker to run each Capability, except for
+   * Capability 0.  The idea is that we're probably going to start a
+   * bound thread on Capability 0 pretty soon, so we don't want a
+   * worker task hogging it.
+   */
+  { 
+      nat i;
+      Capability *cap;
+      for (i = 1; i < n_capabilities; i++) {
+         cap = &capabilities[i];
+         ACQUIRE_LOCK(&cap->lock);
+         startWorkerTask(cap, workerStart);
+         RELEASE_LOCK(&cap->lock);
+      }
+  }
 #endif
 
+  RELEASE_LOCK(&sched_mutex);
 }
 
 void
 exitScheduler( void )
 {
-#if defined(RTS_SUPPORTS_THREADS)
-  stopTaskManager();
-#endif
-}
-
-/* -----------------------------------------------------------------------------
-   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 computation
- * 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.
- * -------------------------------------------------------------------------- */
-
-int 
-howManyThreadsAvail ( void )
-{
-   int i = 0;
-   StgTSO* q;
-   for (q = run_queue_hd; q != END_TSO_QUEUE; q = q->link)
-      i++;
-   for (q = blocked_queue_hd; q != END_TSO_QUEUE; q = q->link)
-      i++;
-   for (q = sleeping_queue; q != END_TSO_QUEUE; q = q->link)
-      i++;
-   return i;
-}
-
-void
-finishAllThreads ( void )
-{
-   do {
-      while (run_queue_hd != END_TSO_QUEUE) {
-         waitThread ( run_queue_hd, NULL);
-      }
-      while (blocked_queue_hd != END_TSO_QUEUE) {
-         waitThread ( blocked_queue_hd, NULL);
-      }
-      while (sleeping_queue != END_TSO_QUEUE) {
-         waitThread ( blocked_queue_hd, NULL);
-      }
-   } while 
-      (blocked_queue_hd != END_TSO_QUEUE || 
-       run_queue_hd     != END_TSO_QUEUE ||
-       sleeping_queue   != END_TSO_QUEUE);
-}
-
-SchedulerStatus
-waitThread(StgTSO *tso, /*out*/StgClosure **ret)
-{ 
-#if defined(THREADED_RTS)
-  return waitThread_(tso,ret, rtsFalse);
-#else
-  return waitThread_(tso,ret);
-#endif
-}
+    Task *task = NULL;
 
-SchedulerStatus
-waitThread_(StgTSO *tso,
-           /*out*/StgClosure **ret
 #if defined(THREADED_RTS)
-           , rtsBool blockWaiting
-#endif
-          )
-{
-  StgMainThread *m;
-  SchedulerStatus stat;
-
-  ACQUIRE_LOCK(&sched_mutex);
-  
-  m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
-
-  m->tso = tso;
-  m->ret = ret;
-  m->stat = NoStatus;
-#if defined(RTS_SUPPORTS_THREADS)
-  initCondition(&m->wakeup);
-#endif
-
-  m->link = main_threads;
-  main_threads = m;
-
-  IF_DEBUG(scheduler, sched_belch("== scheduler: new main thread (%d)\n", m->tso->id));
-
-#if defined(RTS_SUPPORTS_THREADS)
-
-# if defined(THREADED_RTS)
-  if (!blockWaiting) {
-    /* In the threaded case, the OS thread that called main()
-     * gets to enter the RTS directly without going via another
-     * task/thread.
-     */
+    ACQUIRE_LOCK(&sched_mutex);
+    task = newBoundTask();
     RELEASE_LOCK(&sched_mutex);
-    schedule();
-    ASSERT(m->stat != NoStatus);
-  } else 
-# endif
-  {
-    IF_DEBUG(scheduler, sched_belch("sfoo"));
-    do {
-      waitCondition(&m->wakeup, &sched_mutex);
-    } while (m->stat == NoStatus);
-  }
-#elif 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();
-#else
-  RELEASE_LOCK(&sched_mutex);
-  schedule();
-  ASSERT(m->stat != NoStatus);
-#endif
-
-  stat = m->stat;
-
-#if defined(RTS_SUPPORTS_THREADS)
-  closeCondition(&m->wakeup);
 #endif
 
-  IF_DEBUG(scheduler, fprintf(stderr, "== scheduler: main thread (%d) finished\n", 
-                             m->tso->id));
-  free(m);
+    // If we haven't killed all the threads yet, do it now.
+    if (sched_state < SCHED_INTERRUPTED) {
+       sched_state = SCHED_INTERRUPTING;
+       scheduleDoGC(NULL,task,rtsFalse,GetRoots);    
+    }
+    sched_state = SCHED_SHUTTING_DOWN;
 
 #if defined(THREADED_RTS)
-  if (blockWaiting) 
-#endif
-    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;
-  } else {
-    run_queue_tl->link = tso;
-  }
-  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;
-
-  /* 
-     qetlaHbogh Qu' ngaSbogh ghomDaQ {tso} yIteq!
-
-     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;
-      }
+    { 
+       nat i;
+       
+       for (i = 0; i < n_capabilities; i++) {
+           shutdownCapability(&capabilities[i], task);
+       }
+       boundTaskExiting(task);
+       stopTaskManager();
     }
-  }
-  return t;
+#endif
 }
 
-#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?
 
@@ -2246,54 +2874,59 @@ take_off_run_queue(StgTSO *tso) {
 */
 
 void
-GetRoots(evac_fn evac)
+GetRoots( evac_fn evac )
 {
-  StgMainThread *m;
+    nat i;
+    Capability *cap;
+    Task *task;
 
 #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)))
-         evac((StgClosure **)&run_queue_hds[i]);
-      if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
-         evac((StgClosure **)&run_queue_tls[i]);
-      
-      if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
-         evac((StgClosure **)&blocked_queue_hds[i]);
-      if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
-         evac((StgClosure **)&blocked_queue_tls[i]);
-      if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
-         evac((StgClosure **)&ccalling_threads[i]);
+       if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
+           evac((StgClosure **)&run_queue_hds[i]);
+       if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
+           evac((StgClosure **)&run_queue_tls[i]);
+       
+       if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
+           evac((StgClosure **)&blocked_queue_hds[i]);
+       if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
+           evac((StgClosure **)&blocked_queue_tls[i]);
+       if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
+           evac((StgClosure **)&ccalling_threads[i]);
     }
-  }
 
-  markEventQueue();
+    markEventQueue();
 
 #else /* !GRAN */
-  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);
-  }
+
+    for (i = 0; i < n_capabilities; i++) {
+       cap = &capabilities[i];
+       evac((StgClosure **)&cap->run_queue_hd);
+       evac((StgClosure **)&cap->run_queue_tl);
+       
+       for (task = cap->suspended_ccalling_tasks; task != NULL; 
+            task=task->next) {
+           IF_DEBUG(scheduler,sched_belch("evac'ing suspended TSO %d", task->suspended_tso->id));
+           evac((StgClosure **)&task->suspended_tso);
+       }
+    }
+    
+#if !defined(THREADED_RTS)
+    evac((StgClosure **)(void *)&blocked_queue_hd);
+    evac((StgClosure **)(void *)&blocked_queue_tl);
+    evac((StgClosure **)(void *)&sleeping_queue);
 #endif 
+#endif
 
-  if (suspended_ccalling_threads != END_TSO_QUEUE) {
-      evac((StgClosure **)&suspended_ccalling_threads);
-  }
+    // evac((StgClosure **)&blackhole_queue);
 
-#if defined(PAR) || defined(GRAN)
-  markSparkQueue(evac);
+#if defined(THREADED_RTS) || defined(PARALLEL_HASKELL) || defined(GRAN)
+    markSparkQueue(evac);
+#endif
+    
+#if defined(RTS_USER_SIGNALS)
+    // mark the signal handlers (signals should be already blocked)
+    markSignalHandlers(evac);
 #endif
 }
 
@@ -2310,23 +2943,34 @@ GetRoots(evac_fn evac)
    This needs to be protected by the GC condition variable above.  KH.
    -------------------------------------------------------------------------- */
 
-void (*extra_roots)(evac_fn);
+static void (*extra_roots)(evac_fn);
+
+static void
+performGC_(rtsBool force_major, void (*get_roots)(evac_fn))
+{
+    Task *task = myTask();
+
+    if (task == NULL) {
+       ACQUIRE_LOCK(&sched_mutex);
+       task = newBoundTask();
+       RELEASE_LOCK(&sched_mutex);
+       scheduleDoGC(NULL,task,force_major, get_roots);
+       boundTaskExiting(task);
+    } else {
+       scheduleDoGC(NULL,task,force_major, get_roots);
+    }
+}
 
 void
 performGC(void)
 {
-  /* Obligated to hold this lock upon entry */
-  ACQUIRE_LOCK(&sched_mutex);
-  GarbageCollect(GetRoots,rtsFalse);
-  RELEASE_LOCK(&sched_mutex);
+    performGC_(rtsFalse, GetRoots);
 }
 
 void
 performMajorGC(void)
 {
-  ACQUIRE_LOCK(&sched_mutex);
-  GarbageCollect(GetRoots,rtsTrue);
-  RELEASE_LOCK(&sched_mutex);
+    performGC_(rtsTrue, GetRoots);
 }
 
 static void
@@ -2339,10 +2983,8 @@ AllRoots(evac_fn evac)
 void
 performGCWithRoots(void (*get_roots)(evac_fn))
 {
-  ACQUIRE_LOCK(&sched_mutex);
-  extra_roots = get_roots;
-  GarbageCollect(AllRoots,rtsFalse);
-  RELEASE_LOCK(&sched_mutex);
+    extra_roots = get_roots;
+    performGC_(rtsFalse, AllRoots);
 }
 
 /* -----------------------------------------------------------------------------
@@ -2355,9 +2997,10 @@ performGCWithRoots(void (*get_roots)(evac_fn))
    -------------------------------------------------------------------------- */
 
 static StgTSO *
-threadStackOverflow(StgTSO *tso)
+threadStackOverflow(Capability *cap, StgTSO *tso)
 {
-  nat new_stack_size, new_tso_size, diff, stack_words;
+  nat new_stack_size, stack_words;
+  lnat new_tso_size;
   StgPtr new_sp;
   StgTSO *dest;
 
@@ -2365,14 +3008,14 @@ threadStackOverflow(StgTSO *tso)
   if (tso->stack_size >= tso->max_stack_size) {
 
     IF_DEBUG(gc,
-            belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld",
-                  tso->id, tso, tso->stack_size, tso->max_stack_size);
+            debugBelch("@@ threadStackOverflow of TSO %ld (%p): stack too large (now %ld; max is %ld)\n",
+                  (long)tso->id, tso, (long)tso->stack_size, (long)tso->max_stack_size);
             /* If we're debugging, just print out the top of the stack */
             printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size, 
                                              tso->sp+64)));
 
     /* Send this thread the StackOverflow exception */
-    raiseAsync(tso, (StgClosure *)stackOverflow_closure);
+    raiseAsync(cap, tso, (StgClosure *)stackOverflow_closure);
     return tso;
   }
 
@@ -2381,12 +3024,12 @@ threadStackOverflow(StgTSO *tso)
    * Finally round up so the TSO ends up as a whole number of blocks.
    */
   new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
-  new_tso_size   = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) + 
+  new_tso_size   = (lnat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) + 
                                       TSO_STRUCT_SIZE)/sizeof(W_);
   new_tso_size = round_to_mblocks(new_tso_size);  /* Be MBLOCK-friendly */
   new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
 
-  IF_DEBUG(scheduler, fprintf(stderr,"== scheduler: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
+  IF_DEBUG(scheduler, sched_belch("increasing stack size from %ld words to %d.\n", (long)tso->stack_size, new_stack_size));
 
   dest = (StgTSO *)allocate(new_tso_size);
   TICK_ALLOC_TSO(new_stack_size,0);
@@ -2398,30 +3041,23 @@ 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->sp         = new_sp;
   dest->stack_size = new_stack_size;
        
-  /* and relocate the update frame list */
-  relocate_stack(dest, diff);
-
   /* 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,
-              belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld",
+              debugBelch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld\n",
                     tso->id, tso, tso->stack_size);
               /* If we're debugging, just print out the top of the stack */
               printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size, 
@@ -2435,20 +3071,17 @@ 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.
    ------------------------------------------------------------------------ */
 
 #if defined(GRAN)
-static inline void
+STATIC_INLINE void
 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
 {
 }
-#elif defined(PAR)
-static inline void
+#elif defined(PARALLEL_HASKELL)
+STATIC_INLINE void
 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
 {
   /* write RESUME events to log file and
@@ -2457,7 +3090,7 @@ 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())
+    if (emptyRunQueue())
       emitSchedule = rtsTrue;
 
     switch (get_itbl(node)->type) {
@@ -2474,15 +3107,15 @@ unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
           break;
 #endif   
        default:
-         barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
+         barf("{unblockOne}Daq Qagh: unexpected closure in blocking queue");
        }
       }
 }
 #endif
 
 #if defined(GRAN)
-static StgBlockingQueueElement *
-unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
+StgBlockingQueueElement *
+unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
 {
     StgTSO *tso;
     PEs node_loc, tso_loc;
@@ -2513,16 +3146,16 @@ unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
     }
     /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
     IF_GRAN_DEBUG(bq,
-                 fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
+                 debugBelch(" %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
                          (node_loc==tso_loc ? "Local" : "Global"), 
                          tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
     tso->block_info.closure = NULL;
-    IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)", 
+    IF_DEBUG(scheduler,debugBelch("-- Waking up thread %ld (%p)\n", 
                             tso->id, tso));
 }
-#elif defined(PAR)
-static StgBlockingQueueElement *
-unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
+#elif defined(PARALLEL_HASKELL)
+StgBlockingQueueElement *
+unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
 {
     StgBlockingQueueElement *next;
 
@@ -2531,9 +3164,9 @@ 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;
@@ -2557,18 +3190,18 @@ unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
       break;
 
     default:
-      barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
+      barf("{unblockOne}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
           get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe), 
           (StgClosure *)bqe);
 # endif
     }
-  IF_PAR_DEBUG(bq, fprintf(stderr, ", %p (%s)", bqe, info_type((StgClosure*)bqe)));
+  IF_PAR_DEBUG(bq, debugBelch(", %p (%s)\n", bqe, info_type((StgClosure*)bqe)));
   return next;
 }
+#endif
 
-#else /* !GRAN && !PAR */
-static StgTSO *
-unblockOneLocked(StgTSO *tso)
+StgTSO *
+unblockOne(Capability *cap, StgTSO *tso)
 {
   StgTSO *next;
 
@@ -2576,32 +3209,22 @@ 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;
+
+  // We might have just migrated this TSO to our Capability:
+  if (tso->bound) {
+      tso->bound->cap = cap;
+  }
+
+  appendToRunQueue(cap,tso);
+
+  // we're holding a newly woken thread, make sure we context switch
+  // quickly so we can migrate it if necessary.
+  context_switch = 1;
+  IF_DEBUG(scheduler,sched_belch("waking up thread %ld", (long)tso->id));
   return next;
 }
-#endif
 
-#if defined(GRAN) || defined(PAR)
-inline StgBlockingQueueElement *
-unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
-{
-  ACQUIRE_LOCK(&sched_mutex);
-  bqe = unblockOneLocked(bqe, node);
-  RELEASE_LOCK(&sched_mutex);
-  return bqe;
-}
-#else
-inline StgTSO *
-unblockOne(StgTSO *tso)
-{
-  ACQUIRE_LOCK(&sched_mutex);
-  tso = unblockOneLocked(tso);
-  RELEASE_LOCK(&sched_mutex);
-  return tso;
-}
-#endif
 
 #if defined(GRAN)
 void 
@@ -2612,7 +3235,7 @@ awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
   nat len = 0; 
 
   IF_GRAN_DEBUG(bq, 
-               belch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
+               debugBelch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): \n", \
                      node, CurrentProc, CurrentTime[CurrentProc], 
                      CurrentTSO->id, CurrentTSO));
 
@@ -2629,13 +3252,13 @@ 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++;
@@ -2653,7 +3276,7 @@ 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);
+    bqe = unblockOne(bqe, node);
   }
 
   /* if this is the BQ of an RBH, we have to put back the info ripped out of
@@ -2670,7 +3293,7 @@ awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
     ((StgRBH *)node)->mut_link       = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
 
     IF_GRAN_DEBUG(bq,
-                 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
+                 debugBelch("## Filled in RBH_Save for %p (%s) at end of AwBQ\n",
                        node, info_type(node)));
   }
 
@@ -2682,24 +3305,22 @@ awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
     globalGranStats.tot_awbq++;             // total no. of bqs awakened
   }
   IF_GRAN_DEBUG(bq,
-               fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n",
+               debugBelch("## BQ Stats of %p: [%d entries] %s\n",
                        node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
 }
-#elif defined(PAR)
+#elif defined(PARALLEL_HASKELL)
 void 
 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
 {
   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));
 #ifdef DIST  
   //RFP
   if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
-    IF_PAR_DEBUG(verbose, belch("## ... nothing to unblock so lets just return. RFP (BUG?)"));
+    IF_PAR_DEBUG(verbose, debugBelch("## ... nothing to unblock so lets just return. RFP (BUG?)\n"));
     return;
   }
 #endif
@@ -2712,26 +3333,23 @@ awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
   bqe = q;
   while (get_itbl(bqe)->type==TSO || 
         get_itbl(bqe)->type==BLOCKED_FETCH) {
-    bqe = unblockOneLocked(bqe, node);
+    bqe = unblockOne(bqe, node);
   }
-  RELEASE_LOCK(&sched_mutex);
 }
 
-#else   /* !GRAN && !PAR */
+#else   /* !GRAN && !PARALLEL_HASKELL */
+
 void
-awakenBlockedQueue(StgTSO *tso)
+awakenBlockedQueue(Capability *cap, StgTSO *tso)
 {
-  ACQUIRE_LOCK(&sched_mutex);
-  while (tso != END_TSO_QUEUE) {
-    tso = unblockOneLocked(tso);
-  }
-  RELEASE_LOCK(&sched_mutex);
+    if (tso == NULL) return; // hack; see bug #1235728, and comments in
+                            // Exception.cmm
+    while (tso != END_TSO_QUEUE) {
+       tso = unblockOne(cap,tso);
+    }
 }
 #endif
 
-//@node Exception Handling Routines, Debugging Routines, Blocking Queue Routines, Main scheduling code
-//@subsection Exception Handling Routines
-
 /* ---------------------------------------------------------------------------
    Interrupt execution
    - usually called inside a signal handler so it mustn't do anything fancy.   
@@ -2740,8 +3358,11 @@ awakenBlockedQueue(StgTSO *tso)
 void
 interruptStgRts(void)
 {
-    interrupted    = 1;
+    sched_state = SCHED_INTERRUPTING;
     context_switch = 1;
+#if defined(THREADED_RTS)
+    prodAllCapabilities();
+#endif
 }
 
 /* -----------------------------------------------------------------------------
@@ -2752,23 +3373,33 @@ interruptStgRts(void)
    This has nothing to do with the UnblockThread event in GranSim. -- HWL
    -------------------------------------------------------------------------- */
 
-#if defined(GRAN) || defined(PAR)
+#if defined(GRAN) || defined(PARALLEL_HASKELL)
 /*
   NB: only the type of the blocking queue is different in GranSim and GUM
       the operations on the queue-elements are the same
       long live polymorphism!
+
+  Locks: sched_mutex is held upon entry and exit.
+
 */
 static void
-unblockThread(StgTSO *tso)
+unblockThread(Capability *cap, StgTSO *tso)
 {
   StgBlockingQueueElement *t, **last;
 
-  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);
     {
@@ -2835,6 +3466,9 @@ unblockThread(StgTSO *tso)
 
   case BlockedOnRead:
   case BlockedOnWrite:
+#if defined(mingw32_HOST_OS)
+  case BlockedOnDoProc:
+#endif
     {
       /* take TSO off blocked_queue */
       StgBlockingQueueElement *prev = NULL;
@@ -2852,6 +3486,12 @@ unblockThread(StgTSO *tso)
              blocked_queue_tl = (StgTSO *)prev;
            }
          }
+#if defined(mingw32_HOST_OS)
+         /* (Cooperatively) signal that the worker thread should abort
+          * the request.
+          */
+         abandonWorkRequest(tso->block_info.async_result->reqID);
+#endif
          goto done;
        }
       }
@@ -2873,7 +3513,7 @@ unblockThread(StgTSO *tso)
          goto done;
        }
       }
-      barf("unblockThread (I/O): TSO not found");
+      barf("unblockThread (delay): TSO not found");
     }
 
   default:
@@ -2884,20 +3524,28 @@ 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);
+  pushOnRunQueue(cap,tso);
 }
 #else
 static void
-unblockThread(StgTSO *tso)
+unblockThread(Capability *cap, 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);
@@ -2920,12 +3568,9 @@ unblockThread(StgTSO *tso)
     }
 
   case BlockedOnBlackHole:
-    ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
     {
-      StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
-
-      last = &bq->blocking_queue;
-      for (t = bq->blocking_queue; t != END_TSO_QUEUE; 
+      last = &blackhole_queue;
+      for (t = blackhole_queue; t != END_TSO_QUEUE; 
           last = &t->link, t = t->link) {
        if (t == tso) {
          *last = tso->link;
@@ -2960,8 +3605,12 @@ unblockThread(StgTSO *tso)
       barf("unblockThread (Exception): TSO not found");
     }
 
+#if !defined(THREADED_RTS)
   case BlockedOnRead:
   case BlockedOnWrite:
+#if defined(mingw32_HOST_OS)
+  case BlockedOnDoProc:
+#endif
     {
       StgTSO *prev = NULL;
       for (t = blocked_queue_hd; t != END_TSO_QUEUE; 
@@ -2978,6 +3627,12 @@ unblockThread(StgTSO *tso)
              blocked_queue_tl = prev;
            }
          }
+#if defined(mingw32_HOST_OS)
+         /* (Cooperatively) signal that the worker thread should abort
+          * the request.
+          */
+         abandonWorkRequest(tso->block_info.async_result->reqID);
+#endif
          goto done;
        }
       }
@@ -2998,8 +3653,9 @@ unblockThread(StgTSO *tso)
          goto done;
        }
       }
-      barf("unblockThread (I/O): TSO not found");
+      barf("unblockThread (delay): TSO not found");
     }
+#endif
 
   default:
     barf("unblockThread");
@@ -3009,12 +3665,65 @@ 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);
+  appendToRunQueue(cap,tso);
+
+  // We might have just migrated this TSO to our Capability:
+  if (tso->bound) {
+      tso->bound->cap = cap;
+  }
 }
 #endif
 
 /* -----------------------------------------------------------------------------
+ * checkBlackHoles()
+ *
+ * Check the blackhole_queue for threads that can be woken up.  We do
+ * this periodically: before every GC, and whenever the run queue is
+ * empty.
+ *
+ * An elegant solution might be to just wake up all the blocked
+ * threads with awakenBlockedQueue occasionally: they'll go back to
+ * sleep again if the object is still a BLACKHOLE.  Unfortunately this
+ * doesn't give us a way to tell whether we've actually managed to
+ * wake up any threads, so we would be busy-waiting.
+ *
+ * -------------------------------------------------------------------------- */
+
+static rtsBool
+checkBlackHoles (Capability *cap)
+{
+    StgTSO **prev, *t;
+    rtsBool any_woke_up = rtsFalse;
+    StgHalfWord type;
+
+    // blackhole_queue is global:
+    ASSERT_LOCK_HELD(&sched_mutex);
+
+    IF_DEBUG(scheduler, sched_belch("checking threads blocked on black holes"));
+
+    // ASSUMES: sched_mutex
+    prev = &blackhole_queue;
+    t = blackhole_queue;
+    while (t != END_TSO_QUEUE) {
+       ASSERT(t->why_blocked == BlockedOnBlackHole);
+       type = get_itbl(t->block_info.closure)->type;
+       if (type != BLACKHOLE && type != CAF_BLACKHOLE) {
+           IF_DEBUG(sanity,checkTSO(t));
+           t = unblockOne(cap, t);
+           // urk, the threads migrate to the current capability
+           // here, but we'd like to keep them on the original one.
+           *prev = t;
+           any_woke_up = rtsTrue;
+       } else {
+           prev = &t->link;
+           t = t->link;
+       }
+    }
+
+    return any_woke_up;
+}
+
+/* -----------------------------------------------------------------------------
  * raiseAsync()
  *
  * The following function implements the magic for raising an
@@ -3030,12 +3739,12 @@ 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.
  *
@@ -3044,354 +3753,509 @@ unblockThread(StgTSO *tso)
  * CATCH_FRAME on the stack.  In either case, we strip the entire
  * stack and replace the thread with a zombie.
  *
+ * ToDo: in THREADED_RTS mode, this function is only safe if either
+ * (a) we hold all the Capabilities (eg. in GC, or if there is only
+ * one Capability), or (b) we own the Capability that the TSO is
+ * currently blocked on or on the run queue of.
+ *
  * -------------------------------------------------------------------------- */
  
-void 
-deleteThread(StgTSO *tso)
+void
+raiseAsync(Capability *cap, StgTSO *tso, StgClosure *exception)
 {
-  raiseAsync(tso,NULL);
+    raiseAsync_(cap, tso, exception, rtsFalse, NULL);
 }
 
 void
-raiseAsync(StgTSO *tso, StgClosure *exception)
+suspendComputation(Capability *cap, StgTSO *tso, StgPtr stop_here)
 {
-  StgUpdateFrame* su = tso->su;
-  StgPtr          sp = tso->sp;
-  
-  /* Thread already dead? */
-  if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
-    return;
-  }
-
-  IF_DEBUG(scheduler, sched_belch("raising exception in thread %ld.", tso->id));
-
-  /* Remove it from any blocking queues */
-  unblockThread(tso);
-
-  /* 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_)&stg_dummy_ret_closure;
-  }
+    raiseAsync_(cap, tso, NULL, rtsFalse, stop_here);
+}
 
-  while (1) {
-    nat words = ((P_)su - (P_)sp) - 1;
+static void
+raiseAsync_(Capability *cap, StgTSO *tso, StgClosure *exception, 
+           rtsBool stop_at_atomically, StgPtr stop_here)
+{
+    StgRetInfoTable *info;
+    StgPtr sp, frame;
     nat i;
-    StgAP_UPD * ap;
-
-    /* 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.
-     */
-    if (get_itbl(su)->type == CATCH_FRAME && exception != NULL) {
-      StgCatchFrame *cf = (StgCatchFrame *)su;
-      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 = (P_)su - 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;
-      tso->sp = sp;
-      tso->su = su;
-      tso->what_next = ThreadEnterGHC;
-      IF_DEBUG(sanity, checkTSO(tso));
-      return;
+  
+    // 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(cap,tso);
+
+    // mark it dirty; we're about to change its stack.
+    dirtyTSO(tso);
+
+    sp = tso->sp;
     
-    ap->n_args = words;
-    ap->fun    = (StgClosure *)sp[0];
-    sp++;
-    for(i=0; i < (nat)words; ++i) {
-      ap->payload[i] = (StgClosure *)*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,&stg_AP_UPD_info,su->header.prof.ccs /* ToDo */); 
-       TICK_ALLOC_UP_THK(words+1,0);
-       
-       IF_DEBUG(scheduler,
-                fprintf(stderr,  "scheduler: Updating ");
-                printPtr((P_)su->updatee); 
-                fprintf(stderr,  " 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(su->updatee)) {
-           UPD_IND_NOLOCK(su->updatee,ap);  /* revert the black hole */
+
+    frame = sp + 1;
+    while (stop_here == NULL || frame < stop_here) {
+
+       // 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
+       
+       info = get_ret_itbl((StgClosure *)frame);
+
+       switch (info->i.type) {
+
+       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 *)allocateLocal(cap,AP_STACK_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
+           //
+           // 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
+           frame = sp + 1;
+           continue; //no need to bump frame
        }
-       su = su->link;
-       sp += sizeofW(StgUpdateFrame) -1;
-       sp[0] = (W_)ap; /* push onto stack */
-       break;
-      }
 
-    case CATCH_FRAME:
-      {
-       StgCatchFrame *cf = (StgCatchFrame *)su;
-       StgClosure* o;
-       
-       /* We want a PAP, not an AP_UPD.  Fortunately, the
-        * layout's the same.
-        */
-       SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
-       TICK_ALLOC_UPD_PAP(words+1,0);
-       
-       /* now build o = FUN(catch,ap,handler) */
-       o = (StgClosure *)allocate(sizeofW(StgClosure)+2);
-       TICK_ALLOC_FUN(2,0);
-       SET_HDR(o,&stg_catch_info,su->header.prof.ccs /* ToDo */);
-       o->payload[0] = (StgClosure *)ap;
-       o->payload[1] = cf->handler;
-       
-       IF_DEBUG(scheduler,
-                fprintf(stderr,  "scheduler: Built ");
-                printObj((StgClosure *)o);
-                );
-       
-       /* pop the old handler and put o on the stack */
-       su = cf->link;
-       sp += sizeofW(StgCatchFrame) - 1;
-       sp[0] = (W_)o;
-       break;
-      }
-      
-    case SEQ_FRAME:
-      {
-       StgSeqFrame *sf = (StgSeqFrame *)su;
-       StgClosure* o;
-       
-       SET_HDR(ap,&stg_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,&stg_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 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 STOP_FRAME:
+           // We've stripped the entire stack, the thread is now dead.
+           tso->what_next = ThreadKilled;
+           tso->sp = frame + sizeofW(StgStopFrame);
+           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
+           StgThunk *raise;
+           
+           if (exception == NULL) break;
+
+           // we've got an exception to raise, so let's pass it to the
+           // handler in this frame.
+           //
+           raise = (StgThunk *)allocateLocal(cap,sizeofW(StgThunk)+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;
+           }
 
-    default:
-      barf("raiseAsync");
+           /* 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;
+       }
+           
+       case ATOMICALLY_FRAME:
+           if (stop_at_atomically) {
+               ASSERT(stmGetEnclosingTRec(tso->trec) == NO_TREC);
+               stmCondemnTransaction(cap, 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;
+           }
+           // Not stop_at_atomically... fall through and abort the
+           // transaction.
+           
+       case CATCH_RETRY_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"));
+            StgTRecHeader *trec = tso -> trec;
+            StgTRecHeader *outer = stmGetEnclosingTRec(trec);
+            stmAbortTransaction(cap, trec);
+            tso -> trec = outer;
+           break;
+           
+       default:
+           break;
+       }
+
+       // move on to the next stack frame
+       frame += stack_frame_sizeW((StgClosure *)frame);
     }
-  }
-  barf("raiseAsync");
+
+    // if we got here, then we stopped at stop_here
+    ASSERT(stop_here != NULL);
 }
 
 /* -----------------------------------------------------------------------------
-   resurrectThreads is called after garbage collection on the list of
-   threads found to be garbage.  Each of these threads will be woken
-   up and sent a signal: BlockedOnDeadMVar if the thread was blocked
-   on an MVar, or NonTermination if the thread was blocked on a Black
-   Hole.
+   Deleting threads
+
+   This is used for interruption (^C) and forking, and corresponds to
+   raising an exception but without letting the thread catch the
+   exception.
    -------------------------------------------------------------------------- */
 
-void
-resurrectThreads( StgTSO *threads )
+static void 
+deleteThread (Capability *cap, StgTSO *tso)
 {
-  StgTSO *tso, *next;
+  if (tso->why_blocked != BlockedOnCCall &&
+      tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
+      raiseAsync(cap,tso,NULL);
+  }
+}
 
-  for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
-    next = tso->global_link;
-    tso->global_link = all_threads;
-    all_threads = tso;
-    IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
+#ifdef FORKPROCESS_PRIMOP_SUPPORTED
+static void 
+deleteThread_(Capability *cap, StgTSO *tso)
+{ // for forkProcess only:
+  // like deleteThread(), but we delete threads in foreign calls, too.
 
-    switch (tso->why_blocked) {
-    case BlockedOnMVar:
-    case BlockedOnException:
-      raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
-      break;
-    case BlockedOnBlackHole:
-      raiseAsync(tso,(StgClosure *)NonTermination_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
-       * can wake up threads, remember...).
-       */
-      continue;
-    default:
-      barf("resurrectThreads: thread blocked in a strange way");
+    if (tso->why_blocked == BlockedOnCCall ||
+       tso->why_blocked == BlockedOnCCall_NoUnblockExc) {
+       unblockOne(cap,tso);
+       tso->what_next = ThreadKilled;
+    } else {
+       deleteThread(cap,tso);
     }
-  }
 }
+#endif
 
 /* -----------------------------------------------------------------------------
- * Blackhole detection: if we reach a deadlock, test whether any
- * threads are blocked on themselves.  Any threads which are found to
- * be self-blocked get sent a NonTermination exception.
- *
- * This is only done in a deadlock situation in order to avoid
- * performance overhead in the normal case.
- * -------------------------------------------------------------------------- */
+   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.
+   -------------------------------------------------------------------------- */
 
-static void
-detectBlackHoles( void )
+StgWord
+raiseExceptionHelper (StgRegTable *reg, StgTSO *tso, StgClosure *exception)
 {
-    StgTSO *t = all_threads;
-    StgUpdateFrame *frame;
-    StgClosure *blocked_on;
-
-    for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
+    Capability *cap = regTableToCapability(reg);
+    StgThunk *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.
+    //
+
+    // OLD COMMENT (we don't have MIN_UPD_SIZE now):
+    // 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 = 
+                   (StgThunk *)allocateLocal(cap,sizeofW(StgThunk)+1);
+               SET_HDR(raise_closure, &stg_raise_info, CCCS);
+               raise_closure->payload[0] = exception;
+           }
+           UPD_IND(((StgUpdateFrame *)p)->updatee,(StgClosure *)raise_closure);
+           p = next;
+           continue;
 
-       while (t->what_next == ThreadRelocated) {
-           t = t->link;
-           ASSERT(get_itbl(t)->type == TSO);
-       }
-      
-       if (t->why_blocked != BlockedOnBlackHole) {
+        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;
        }
+    }
+}
 
-       blocked_on = t->block_info.closure;
 
-       for (frame = t->su; ; frame = frame->link) {
-           switch (get_itbl(frame)->type) {
+/* -----------------------------------------------------------------------------
+   findRetryFrameHelper
 
-           case UPDATE_FRAME:
-               if (frame->updatee == blocked_on) {
-                   /* We are blocking on one of our own computations, so
-                    * send this thread the NonTermination exception.  
-                    */
-                   IF_DEBUG(scheduler, 
-                            sched_belch("thread %d is blocked on itself", t->id));
-                   raiseAsync(t, (StgClosure *)NonTermination_closure);
-                   goto done;
-               }
-               else {
-                   continue;
-               }
+   This function is called by the retry# primitive.  It traverses the stack
+   leaving tso->sp referring to the frame which should handle the retry.  
 
-           case CATCH_FRAME:
-           case SEQ_FRAME:
-               continue;
-               
-           case STOP_FRAME:
-               break;
-           }
-           break;
-       }
+   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).  
 
-    done: ;
-    }   
-}
+   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;
 
-//@node Debugging Routines, Index, Exception Handling Routines, Main scheduling code
-//@subsection Debugging Routines
+  p = tso -> sp;
+  while (1) {
+    info = get_ret_itbl((StgClosure *)p);
+    next = p + stack_frame_sizeW((StgClosure *)p);
+    switch (info->i.type) {
+      
+    case ATOMICALLY_FRAME:
+      IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p during retrry\n", p));
+      tso->sp = p;
+      return ATOMICALLY_FRAME;
+      
+    case CATCH_RETRY_FRAME:
+      IF_DEBUG(stm, debugBelch("Found CATCH_RETRY_FRAME at %p during retrry\n", p));
+      tso->sp = p;
+      return CATCH_RETRY_FRAME;
+      
+    case CATCH_STM_FRAME:
+    default:
+      ASSERT(info->i.type != CATCH_FRAME);
+      ASSERT(info->i.type != STOP_FRAME);
+      p = next; 
+      continue;
+    }
+  }
+}
 
 /* -----------------------------------------------------------------------------
-   Debugging: why is a thread blocked
-   -------------------------------------------------------------------------- */
+   resurrectThreads is called after garbage collection on the list of
+   threads found to be garbage.  Each of these threads will be woken
+   up and sent a signal: BlockedOnDeadMVar if the thread was blocked
+   on an MVar, or NonTermination if the thread was blocked on a Black
+   Hole.
 
-#ifdef DEBUG
+   Locks: assumes we hold *all* the capabilities.
+   -------------------------------------------------------------------------- */
 
 void
+resurrectThreads (StgTSO *threads)
+{
+    StgTSO *tso, *next;
+    Capability *cap;
+
+    for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
+       next = tso->global_link;
+       tso->global_link = all_threads;
+       all_threads = tso;
+       IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
+       
+       // Wake up the thread on the Capability it was last on for a
+       // bound thread, or last_free_capability otherwise.
+       if (tso->bound) {
+           cap = tso->bound->cap;
+       } else {
+           cap = last_free_capability;
+       }
+       
+       switch (tso->why_blocked) {
+       case BlockedOnMVar:
+       case BlockedOnException:
+           /* Called by GC - sched_mutex lock is currently held. */
+           raiseAsync(cap, tso,(StgClosure *)BlockedOnDeadMVar_closure);
+           break;
+       case BlockedOnBlackHole:
+           raiseAsync(cap, tso,(StgClosure *)NonTermination_closure);
+           break;
+       case BlockedOnSTM:
+           raiseAsync(cap, 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
+            * can wake up threads, remember...).
+            */
+           continue;
+       default:
+           barf("resurrectThreads: thread blocked in a strange way");
+       }
+    }
+}
+
+/* ----------------------------------------------------------------------------
+ * 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]
+   ------------------------------------------------------------------------- */
+
+#if DEBUG
+static void
 printThreadBlockage(StgTSO *tso)
 {
   switch (tso->why_blocked) {
   case BlockedOnRead:
-    fprintf(stderr,"is blocked on read from fd %d", tso->block_info.fd);
+    debugBelch("is blocked on read from fd %d", (int)(tso->block_info.fd));
     break;
   case BlockedOnWrite:
-    fprintf(stderr,"is blocked on write to fd %d", tso->block_info.fd);
+    debugBelch("is blocked on write to fd %d", (int)(tso->block_info.fd));
+    break;
+#if defined(mingw32_HOST_OS)
+    case BlockedOnDoProc:
+    debugBelch("is blocked on proc (request: %ld)", tso->block_info.async_result->reqID);
     break;
+#endif
   case BlockedOnDelay:
-    fprintf(stderr,"is blocked until %d", tso->block_info.target);
+    debugBelch("is blocked until %ld", (long)(tso->block_info.target));
     break;
   case BlockedOnMVar:
-    fprintf(stderr,"is blocked on an MVar");
+    debugBelch("is blocked on an MVar @ %p", tso->block_info.closure);
     break;
   case BlockedOnException:
-    fprintf(stderr,"is blocked on delivering an exception to thread %d",
+    debugBelch("is blocked on delivering an exception to thread %d",
            tso->block_info.tso->id);
     break;
   case BlockedOnBlackHole:
-    fprintf(stderr,"is blocked on a black hole");
+    debugBelch("is blocked on a black hole");
     break;
   case NotBlocked:
-    fprintf(stderr,"is not blocked");
+    debugBelch("is not blocked");
     break;
-#if defined(PAR)
+#if defined(PARALLEL_HASKELL)
   case BlockedOnGA:
-    fprintf(stderr,"is blocked on global address; local FM_BQ is %p (%s)",
+    debugBelch("is blocked on global address; local FM_BQ is %p (%s)",
            tso->block_info.closure, info_type(tso->block_info.closure));
     break;
   case BlockedOnGA_NoSend:
-    fprintf(stderr,"is blocked on global address (no send); local FM_BQ is %p (%s)",
+    debugBelch("is blocked on global address (no send); local FM_BQ is %p (%s)",
            tso->block_info.closure, info_type(tso->block_info.closure));
     break;
 #endif
-#if defined(RTS_SUPPORTS_THREADS)
   case BlockedOnCCall:
-    fprintf(stderr,"is blocked on an external call");
+    debugBelch("is blocked on an external call");
+    break;
+  case BlockedOnCCall_NoUnblockExc:
+    debugBelch("is blocked on an external call (exceptions were already blocked)");
+    break;
+  case BlockedOnSTM:
+    debugBelch("is blocked on an STM operation");
     break;
-#endif
   default:
     barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
         tso->why_blocked, tso->id, tso);
@@ -3399,53 +4263,90 @@ printThreadBlockage(StgTSO *tso)
 }
 
 void
-printThreadStatus(StgTSO *tso)
+printThreadStatus(StgTSO *t)
 {
-  switch (tso->what_next) {
-  case ThreadKilled:
-    fprintf(stderr,"has been killed");
-    break;
-  case ThreadComplete:
-    fprintf(stderr,"has completed");
-    break;
-  default:
-    printThreadBlockage(tso);
-  }
+    debugBelch("\tthread %4d @ %p ", t->id, (void *)t);
+    {
+      void *label = lookupThreadLabel(t->id);
+      if (label) debugBelch("[\"%s\"] ",(char *)label);
+    }
+    if (t->what_next == ThreadRelocated) {
+       debugBelch("has been relocated...\n");
+    } else {
+       switch (t->what_next) {
+       case ThreadKilled:
+           debugBelch("has been killed");
+           break;
+       case ThreadComplete:
+           debugBelch("has completed");
+           break;
+       default:
+           printThreadBlockage(t);
+       }
+       debugBelch("\n");
+    }
 }
 
 void
 printAllThreads(void)
 {
-  StgTSO *t;
+  StgTSO *t, *next;
+  nat i;
+  Capability *cap;
 
 # 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!*/);
 
-  sched_belch("all threads at [%s]:", time_string);
-# elif defined(PAR)
+  debugBelch("all threads at [%s]:\n", time_string);
+# elif defined(PARALLEL_HASKELL)
   char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
   ullong_format_string(CURRENT_TIME,
                       time_string, rtsFalse/*no commas!*/);
 
-  sched_belch("all threads at [%s]:", time_string);
+  debugBelch("all threads at [%s]:\n", time_string);
 # else
-  sched_belch("all threads:");
+  debugBelch("all threads:\n");
 # endif
 
-  for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
-    fprintf(stderr, "\tthread %d ", t->id);
-    printThreadStatus(t);
-    fprintf(stderr,"\n");
+  for (i = 0; i < n_capabilities; i++) {
+      cap = &capabilities[i];
+      debugBelch("threads on capability %d:\n", cap->no);
+      for (t = cap->run_queue_hd; t != END_TSO_QUEUE; t = t->link) {
+         printThreadStatus(t);
+      }
+  }
+
+  debugBelch("other threads:\n");
+  for (t = all_threads; t != END_TSO_QUEUE; t = next) {
+      if (t->why_blocked != NotBlocked) {
+         printThreadStatus(t);
+      }
+      if (t->what_next == ThreadRelocated) {
+         next = t->link;
+      } else {
+         next = t->global_link;
+      }
   }
 }
-    
+
+// useful from gdb
+void 
+printThreadQueue(StgTSO *t)
+{
+    nat i = 0;
+    for (; t != END_TSO_QUEUE; t = t->link) {
+       printThreadStatus(t);
+       i++;
+    }
+    debugBelch("%d threads on queue\n", i);
+}
+
 /* 
    Print a whole blocking queue attached to node (debugging only).
 */
-//@cindex print_bq
-# if defined(PAR)
+# if defined(PARALLEL_HASKELL)
 void 
 print_bq (StgClosure *node)
 {
@@ -3453,7 +4354,7 @@ print_bq (StgClosure *node)
   StgTSO *tso;
   rtsBool end;
 
-  fprintf(stderr,"## BQ of closure %p (%s): ",
+  debugBelch("## BQ of closure %p (%s): ",
          node, info_type(node));
 
   /* should cover all closures that may have a blocking queue */
@@ -3493,18 +4394,18 @@ print_bqe (StgBlockingQueueElement *bqe)
 
     switch (get_itbl(bqe)->type) {
     case TSO:
-      fprintf(stderr," TSO %u (%x),",
+      debugBelch(" TSO %u (%x),",
              ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
       break;
     case BLOCKED_FETCH:
-      fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
+      debugBelch(" BF (node=%p, ga=((%x, %d, %x)),",
              ((StgBlockedFetch *)bqe)->node, 
              ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
              ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
              ((StgBlockedFetch *)bqe)->ga.weight);
       break;
     case CONSTR:
-      fprintf(stderr," %s (IP %p),",
+      debugBelch(" %s (IP %p),",
              (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
               get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
               get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
@@ -3516,7 +4417,7 @@ print_bqe (StgBlockingQueueElement *bqe)
       break;
     }
   } /* for */
-  fputc('\n', stderr);
+  debugBelch("\n");
 }
 # elif defined(GRAN)
 void 
@@ -3534,7 +4435,7 @@ 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);
 
   /* 
@@ -3554,11 +4455,11 @@ print_bq (StgClosure *node)
     tso_loc = where_is((StgClosure *)bqe);
     switch (get_itbl(bqe)->type) {
     case TSO:
-      fprintf(stderr," TSO %d (%p) on [PE %d],",
+      debugBelch(" TSO %d (%p) on [PE %d],",
              ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
       break;
     case CONSTR:
-      fprintf(stderr," %s (IP %p),",
+      debugBelch(" %s (IP %p),",
              (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
               get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
               get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
@@ -3570,83 +4471,42 @@ print_bq (StgClosure *node)
       break;
     }
   } /* for */
-  fputc('\n', stderr);
-}
-#else
-/* 
-   Nice and easy: only TSOs on the blocking queue
-*/
-void 
-print_bq (StgClosure *node)
-{
-  StgTSO *tso;
-
-  ASSERT(node!=(StgClosure*)NULL);         // sanity check
-  for (tso = ((StgBlockingQueue*)node)->blocking_queue;
-       tso != END_TSO_QUEUE; 
-       tso=tso->link) {
-    ASSERT(tso!=NULL && tso!=END_TSO_QUEUE);   // sanity check
-    ASSERT(get_itbl(tso)->type == TSO);  // guess what, sanity check
-    fprintf(stderr," TSO %d (%p),", tso->id, tso);
-  }
-  fputc('\n', stderr);
+  debugBelch("\n");
 }
 # endif
 
-#if defined(PAR)
+#if defined(PARALLEL_HASKELL)
 static nat
 run_queue_len(void)
 {
-  nat i;
-  StgTSO *tso;
-
-  for (i=0, tso=run_queue_hd; 
-       tso != END_TSO_QUEUE;
-       i++, tso=tso->link)
-    /* nothing */
-
-  return i;
+    nat i;
+    StgTSO *tso;
+    
+    for (i=0, tso=run_queue_hd; 
+        tso != END_TSO_QUEUE;
+        i++, tso=tso->link) {
+       /* nothing */
+    }
+       
+    return i;
 }
 #endif
 
-static void
+void
 sched_belch(char *s, ...)
 {
-  va_list ap;
-  va_start(ap,s);
-#ifdef SMP
-  fprintf(stderr, "scheduler (task %ld): ", osThreadId());
-#elif defined(PAR)
-  fprintf(stderr, "== ");
+    va_list ap;
+    va_start(ap,s);
+#ifdef THREADED_RTS
+    debugBelch("sched (task %p): ", (void *)(unsigned long)(unsigned int)osThreadId());
+#elif defined(PARALLEL_HASKELL)
+    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
-//* 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
-//* gc_pending_cond::  @cindex\s-+gc_pending_cond
-//* initScheduler::  @cindex\s-+initScheduler
-//* interrupted::  @cindex\s-+interrupted
-//* 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
-//* term_mutex::  @cindex\s-+term_mutex
-//@end index