Merge branch 'rickard/timer-improvements'

OTP-14356

* rickard/timer-improvements:
  Fix of later timer wheel
  Minimum timeout position in each timer wheel
  Manage timers to trigger at once in a slot similar to other timers
  Introduce timer slot range counters
  Timer wheel divided into a "soon wheel" and a "later wheel"
  Remove unnecessary cancel callback from timer-wheel timers
  Rearrange timer struct fields in order to simplify
  Use timer wheel for short BIF timers
  Use magic refs for BIF timers
  Remove accessor BIF timer implementation
  Fix aux-work timer implementation

11 files changed, 2353 insertions, 1165 deletions

diff --git a/erts/emulator/beam/break.c b/erts/emulator/beam/break.c
index 0b40d70cb7..538ab0d947 100644
--- a/erts/emulator/beam/break.c
+++ b/erts/emulator/beam/break.c
@@ -590,7 +590,7 @@ do_break(void)
 #endif
 #ifdef DEBUG
 	case 't':
-	    erts_p_slpq();
+	    /* erts_p_slpq(); */
 	    return;
 	case 'b':
 	    bin_check();
diff --git a/erts/emulator/beam/erl_alloc.c b/erts/emulator/beam/erl_alloc.c
index a374593c5d..71957b2259 100644
--- a/erts/emulator/beam/erl_alloc.c
+++ b/erts/emulator/beam/erl_alloc.c
@@ -678,10 +678,6 @@ erts_alloc_init(int *argc, char **argv, ErtsAllocInitOpts *eaiop)
 	= erts_timer_type_size(ERTS_ALC_T_HL_PTIMER);
     fix_type_sizes[ERTS_ALC_FIX_TYPE_IX(ERTS_ALC_T_BIF_TIMER)]
 	= erts_timer_type_size(ERTS_ALC_T_BIF_TIMER);
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-    fix_type_sizes[ERTS_ALC_FIX_TYPE_IX(ERTS_ALC_T_ABIF_TIMER)]
-	= erts_timer_type_size(ERTS_ALC_T_ABIF_TIMER);
-#endif
     fix_type_sizes[ERTS_ALC_FIX_TYPE_IX(ERTS_ALC_T_NIF_EXP_TRACE)]
 	= sizeof(NifExportTrace);
     fix_type_sizes[ERTS_ALC_FIX_TYPE_IX(ERTS_ALC_T_MREF_NSCHED_ENT)]
@@ -2440,12 +2436,6 @@ erts_memory(fmtfn_t *print_to_p, void *print_to_arg, void *proc, Eterm earg)
 		       &size.processes_used,
 		       fi,
 		       ERTS_ALC_T_BIF_TIMER);
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-	add_fix_values(&size.processes,
-		       &size.processes_used,
-		       fi,
-		       ERTS_ALC_T_ABIF_TIMER);
-#endif
 	add_fix_values(&size.processes,
 		       &size.processes_used,
 		       fi,
diff --git a/erts/emulator/beam/erl_alloc.types b/erts/emulator/beam/erl_alloc.types
index 43f43f9034..f296a98125 100644
--- a/erts/emulator/beam/erl_alloc.types
+++ b/erts/emulator/beam/erl_alloc.types
@@ -168,7 +168,6 @@ type	TIMER_SERVICE	LONG_LIVED	SYSTEM		timer_service
 type    LL_PTIMER	FIXED_SIZE	PROCESSES	ll_ptimer
 type    HL_PTIMER	FIXED_SIZE	PROCESSES	hl_ptimer
 type    BIF_TIMER	FIXED_SIZE	PROCESSES	bif_timer
-# type    ABIF_TIMER	FIXED_SIZE	PROCESSES	accessor_bif_timer
 type    TIMER_REQUEST	SHORT_LIVED	PROCESSES	timer_request
 type    BTM_YIELD_STATE	SHORT_LIVED	PROCESSES	btm_yield_state
 type	REG_TABLE	STANDARD	SYSTEM		reg_tab
diff --git a/erts/emulator/beam/erl_hl_timer.c b/erts/emulator/beam/erl_hl_timer.c
index 26be8c7edf..13d6136672 100644
--- a/erts/emulator/beam/erl_hl_timer.c
+++ b/erts/emulator/beam/erl_hl_timer.c
@@ -29,6 +29,8 @@
 #  include "config.h"
 #endif
 
+/* #define ERTS_MAGIC_REF_BIF_TIMERS */
+
 #include "sys.h"
 #include "global.h"
 #include "bif.h"
@@ -36,6 +38,9 @@
 #define ERTS_WANT_TIMER_WHEEL_API
 #include "erl_time.h"
 #include "erl_hl_timer.h"
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+#include "erl_binary.h"
+#endif
 
 #define ERTS_TMR_CHECK_CANCEL_ON_CREATE 0
 
@@ -106,9 +111,6 @@ typedef enum {
 #define ERTS_TMR_ROFLG_PROC		(((Uint32) 1) << 14)
 #define ERTS_TMR_ROFLG_PORT		(((Uint32) 1) << 15)
 #define ERTS_TMR_ROFLG_CALLBACK		(((Uint32) 1) << 16)
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-#define ERTS_TMR_ROFLG_ABIF_TMR		(((Uint32) 1) << 17)
-#endif
 
 #define ERTS_TMR_ROFLG_SID_MASK	\
     (ERTS_TMR_ROFLG_HLT - (Uint32) 1)
@@ -127,6 +129,13 @@ typedef struct ErtsHLTimer_ ErtsHLTimer;
 
 #define ERTS_HLT_PFIELD_NOT_IN_TABLE	(~((UWord) 0))
 
+typedef struct ErtsBifTimer_ ErtsBifTimer;
+
+typedef struct {
+    ErtsBifTimer *next;
+    ErtsBifTimer *prev;    
+} ErtsBifTimerList;
+
 typedef struct {
     UWord parent; /* parent pointer and flags... */
     union {
@@ -144,9 +153,9 @@ typedef struct {
 
 typedef struct {
     UWord parent; /* parent pointer and flags... */
-    ErtsHLTimer *right;
-    ErtsHLTimer *left;
-} ErtsHLTimerTree;
+    ErtsBifTimer *right;
+    ErtsBifTimer *left;
+} ErtsBifTimerTree;
 
 typedef struct {
     Uint32 roflgs;
@@ -155,67 +164,75 @@ typedef struct {
 	void *arg;
 	erts_atomic_t next;
     } u;
+    union {
+	Process *proc;
+	Port *port;
+	Eterm name;
+	void (*callback)(void *);
+    } receiver;
 } ErtsTmrHead;
 
 struct ErtsHLTimer_ {
     ErtsTmrHead head; /* NEED to be first! */
+    ErtsMonotonicTime timeout;
     union {
 	ErtsThrPrgrLaterOp cleanup;
 	ErtsHLTimerTimeTree tree;
     } time;
-    ErtsMonotonicTime timeout;
-    union {
-	Process *proc;
-	Port *port;
-	Eterm name;
-	void (*callback)(void *);
-    } receiver;
 
 #ifdef ERTS_HLT_HARD_DEBUG
     int pending_timeout;
 #endif
-
-    erts_smp_atomic32_t state;
-
-    /* BIF timer only fields follow... */
-    struct {
-	Uint32 refn[ERTS_REF_NUMBERS];
-	ErtsHLTimerTree proc_tree;
-	ErtsHLTimerTree tree;
-	Eterm message;
-	ErlHeapFragment *bp;
-    } btm;
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-    struct {
-	Eterm accessor;
-	ErtsHLTimerTree tree;
-    } abtm;
-#endif
 };
 
-#define ERTS_HL_PTIMER_SIZE offsetof(ErtsHLTimer, btm)
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-#define ERTS_BIF_TIMER_SIZE offsetof(ErtsHLTimer, abtm)
-#define ERTS_ABIF_TIMER_SIZE sizeof(ErtsHLTimer)
-#else
-#define ERTS_BIF_TIMER_SIZE sizeof(ErtsHLTimer)
-#endif
-
 typedef struct {
     ErtsTmrHead head; /* NEED to be first! */
     union {
-	void *p;
-	void (*callback)(void *);
+        ErtsTWheelTimer tw_tmr;
+	ErtsThrPrgrLaterOp cleanup;
     } u;
-    ErtsTWheelTimer tw_tmr;
 } ErtsTWTimer;
 
+struct ErtsBifTimer_ {
+    union {
+        ErtsTmrHead head;
+        ErtsHLTimer hlt;
+        ErtsTWTimer twt;
+    } type;
+    struct {
+        erts_smp_atomic32_t state;
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+        ErtsMagicBinary *mbin;
+        ErtsHLTimerList proc_list;
+#else
+	Uint32 refn[ERTS_REF_NUMBERS];
+	ErtsBifTimerTree proc_tree;
+	ErtsBifTimerTree tree;
+#endif
+        Eterm message;
+        ErlHeapFragment *bp;
+    } btm;
+};
+
 typedef union {
     ErtsTmrHead head;
     ErtsHLTimer hlt;
     ErtsTWTimer twt;
+    ErtsBifTimer btm;
 } ErtsTimer;
 
+typedef ErtsTimer *(*ErtsCreateTimerFunc)(ErtsSchedulerData *esdp,
+                                          ErtsMonotonicTime timeout_pos,
+                                          int short_time, ErtsTmrType type,
+                                          void *rcvrp, Eterm rcvr,
+                                          Eterm msg,
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+                                          ErtsMagicBinary *mbin,
+#else
+                                          Uint32 *refn,
+#endif
+                                          void (*callback)(void *), void *arg);
+
 #ifdef SMALL_MEMORY
 #define BIF_TIMER_PREALC_SZ	10
 #define PTIMER_PREALC_SZ	10
@@ -225,7 +242,7 @@ typedef union {
 #endif
 
 ERTS_SCHED_PREF_PALLOC_IMPL(bif_timer_pre,
-			    ErtsHLTimer,
+			    ErtsBifTimer,
 			    BIF_TIMER_PREALC_SZ)
 
 ERTS_SCHED_PREF_QUICK_ALLOC_IMPL(tw_timer,
@@ -296,12 +313,16 @@ struct ErtsHLTimerService_ {
     ErtsHLTCncldTmrQ canceled_queue;
 #endif
     ErtsHLTimer *time_tree;
-    ErtsHLTimer *btm_tree;
+#ifndef ERTS_MAGIC_REF_BIF_TIMERS
+    ErtsBifTimer *btm_tree;
+#endif
     ErtsHLTimer *next_timeout;
     ErtsYieldingTimeoutState yield;
     ErtsTWheelTimer service_timer;
 };
 
+#ifndef ERTS_MAGIC_REF_BIF_TIMERS
+
 static ERTS_INLINE int
 refn_is_lt(Uint32 *x, Uint32 *y)
 {
@@ -317,6 +338,14 @@ refn_is_lt(Uint32 *x, Uint32 *y)
     return x[0] < y[0];
 }
 
+static ERTS_INLINE int
+refn_is_eq(Uint32 *x, Uint32 *y)
+{
+    return (x[0] == y[0]) & (x[1] == y[1]) & (x[2] == y[2]);
+}
+
+#endif
+
 #define ERTS_RBT_PREFIX time
 #define ERTS_RBT_T ErtsHLTimer
 #define ERTS_RBT_KEY_T ErtsMonotonicTime
@@ -506,8 +535,16 @@ same_time_list_lookup(ErtsHLTimer *root, ErtsHLTimer *x)
 
 #endif /* ERTS_HLT_HARD_DEBUG */
 
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+#define ERTS_BTM_HLT2REFN(T) ((T)->btm.mbin->refn)
+#else
+#define ERTS_BTM_HLT2REFN(T) ((T)->btm.refn)
+#endif
+
+#ifndef ERTS_MAGIC_REF_BIF_TIMERS
+
 #define ERTS_RBT_PREFIX btm
-#define ERTS_RBT_T ErtsHLTimer
+#define ERTS_RBT_T ErtsBifTimer
 #define ERTS_RBT_KEY_T Uint32 *
 #define ERTS_RBT_FLAGS_T UWord
 #define ERTS_RBT_INIT_EMPTY_TNODE(T)					\
@@ -533,7 +570,7 @@ same_time_list_lookup(ErtsHLTimer *root, ErtsHLTimer *x)
 	(T)->btm.tree.parent |= (F);					\
     } while (0)
 #define ERTS_RBT_GET_PARENT(T)						\
-    ((ErtsHLTimer *) ((T)->btm.tree.parent & ~ERTS_HLT_PFLGS_MASK))
+    ((ErtsBifTimer *) ((T)->btm.tree.parent & ~ERTS_HLT_PFLGS_MASK))
 #define ERTS_RBT_SET_PARENT(T, P)					\
     do {								\
 	ERTS_HLT_ASSERT((((UWord) (P)) & ERTS_HLT_PFLGS_MASK) == 0);	\
@@ -544,20 +581,94 @@ same_time_list_lookup(ErtsHLTimer *root, ErtsHLTimer *x)
 #define ERTS_RBT_SET_RIGHT(T, R) ((T)->btm.tree.right = (R))
 #define ERTS_RBT_GET_LEFT(T) ((T)->btm.tree.left)
 #define ERTS_RBT_SET_LEFT(T, L) ((T)->btm.tree.left = (L))
-#define ERTS_RBT_GET_KEY(T) ((T)->btm.refn)
+#define ERTS_RBT_GET_KEY(T) ERTS_BTM_HLT2REFN((T))
 #define ERTS_RBT_IS_LT(KX, KY) refn_is_lt((KX), (KY))
-#define ERTS_RBT_IS_EQ(KX, KY) \
-    (((KX)[0] == (KY)[0]) & ((KX)[1] == (KY)[1]) & ((KX)[2] == (KY)[2]))
+#define ERTS_RBT_IS_EQ(KX, KY) refn_is_eq((KX), (KY))
 #define ERTS_RBT_WANT_DELETE
 #define ERTS_RBT_WANT_INSERT
+#ifndef ERTS_MAGIC_REF_BIF_TIMERS
 #define ERTS_RBT_WANT_LOOKUP
+#endif
 #define ERTS_RBT_WANT_FOREACH
 #define ERTS_RBT_UNDEF
 
 #include "erl_rbtree.h"
 
+#endif /* !ERTS_MAGIC_REF_BIF_TIMERS */
+
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+
+static ERTS_INLINE void
+proc_btm_list_insert(ErtsBifTimer **list, ErtsBifTimer *x)
+{
+    ErtsBifTimer *y = *list;
+    if (!y) {
+        x->btm.proc_list.next = x;
+        x->btm.proc_list.prev = x;
+        *list = x;
+    }
+    else {
+        ERTS_HLT_ASSERT(y->btm.proc_list.prev->btm.proc_list.next == y);
+        x->btm.proc_list.next = y;
+        x->btm.proc_list.prev = y->btm.proc_list.prev;
+        y->btm.proc_list.prev->btm.proc_list.next = x;
+        y->btm.proc_list.prev = x;
+    }
+}
+
+static ERTS_INLINE void
+proc_btm_list_delete(ErtsBifTimer **list, ErtsBifTimer *x)
+{
+    ErtsBifTimer *y = *list;
+    if (y == x && x->btm.proc_list.next == x) {
+        ERTS_HLT_ASSERT(x->btm.proc_list.prev == x);
+        *list = NULL;
+    }
+    else {
+        if (y == x)
+            *list = x->btm.proc_list.next;
+        ERTS_HLT_ASSERT(x->btm.proc_list.prev->btm.proc_list.next == x);
+        ERTS_HLT_ASSERT(x->btm.proc_list.next->btm.proc_list.prev == x);
+        x->btm.proc_list.prev->btm.proc_list.next = x->btm.proc_list.next;
+        x->btm.proc_list.next->btm.proc_list.prev = x->btm.proc_list.prev;
+    }
+    x->btm.proc_list.next = NULL;
+}
+
+static ERTS_INLINE int
+proc_btm_list_foreach_destroy_yielding(ErtsBifTimer **list,
+                                       void (*destroy)(ErtsBifTimer *, void *),
+                                       void *arg,
+                                       int limit)
+{
+    int i;
+    ErtsBifTimer *first, *last;
+
+    first = *list;
+    if (!first)
+        return 0;
+
+    last = first->btm.proc_list.prev;
+    for (i = 0; i < limit; i++) {
+        ErtsBifTimer *x = last;
+        last = last->btm.proc_list.prev;
+        (*destroy)(x, arg);
+        x->btm.proc_list.next = NULL;
+        if (x == first) {
+            *list = NULL;
+            return 0;
+        }
+    }
+
+    last->btm.proc_list.next = first;
+    first->btm.proc_list.prev = last;
+    return 1;
+}
+
+#else /* !ERTS_MAGIC_REF_BIF_TIMERS */
+
 #define ERTS_RBT_PREFIX proc_btm
-#define ERTS_RBT_T ErtsHLTimer
+#define ERTS_RBT_T ErtsBifTimer
 #define ERTS_RBT_KEY_T Uint32 *
 #define ERTS_RBT_FLAGS_T UWord
 #define ERTS_RBT_INIT_EMPTY_TNODE(T)					\
@@ -583,7 +694,7 @@ same_time_list_lookup(ErtsHLTimer *root, ErtsHLTimer *x)
 	(T)->btm.proc_tree.parent |= (F);				\
     } while (0)
 #define ERTS_RBT_GET_PARENT(T)						\
-    ((ErtsHLTimer *) ((T)->btm.proc_tree.parent & ~ERTS_HLT_PFLGS_MASK))
+    ((ErtsBifTimer *) ((T)->btm.proc_tree.parent & ~ERTS_HLT_PFLGS_MASK))
 #define ERTS_RBT_SET_PARENT(T, P)					\
     do {								\
 	ERTS_HLT_ASSERT((((UWord) (P)) & ERTS_HLT_PFLGS_MASK) == 0);	\
@@ -594,71 +705,20 @@ same_time_list_lookup(ErtsHLTimer *root, ErtsHLTimer *x)
 #define ERTS_RBT_SET_RIGHT(T, R) ((T)->btm.proc_tree.right = (R))
 #define ERTS_RBT_GET_LEFT(T) ((T)->btm.proc_tree.left)
 #define ERTS_RBT_SET_LEFT(T, L) ((T)->btm.proc_tree.left = (L))
-#define ERTS_RBT_GET_KEY(T) ((T)->btm.refn)
+#define ERTS_RBT_GET_KEY(T) ERTS_BTM_HLT2REFN((T))
 #define ERTS_RBT_IS_LT(KX, KY) refn_is_lt((KX), (KY))
-#define ERTS_RBT_IS_EQ(KX, KY) \
-    (((KX)[0] == (KY)[0]) & ((KX)[1] == (KY)[1]) & ((KX)[2] == (KY)[2]))
-#define ERTS_RBT_WANT_DELETE
-#define ERTS_RBT_WANT_INSERT
-#define ERTS_RBT_WANT_LOOKUP
-#define ERTS_RBT_WANT_FOREACH_DESTROY_YIELDING
-#define ERTS_RBT_UNDEF
-
-#include "erl_rbtree.h"
-
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-
-#define ERTS_RBT_PREFIX abtm
-#define ERTS_RBT_T ErtsHLTimer
-#define ERTS_RBT_KEY_T Uint32 *
-#define ERTS_RBT_FLAGS_T UWord
-#define ERTS_RBT_INIT_EMPTY_TNODE(T)					\
-    do {								\
-	(T)->abtm.tree.parent = (UWord) NULL;				\
-	(T)->abtm.tree.right = NULL;					\
-	(T)->abtm.tree.left = NULL;					\
-    } while (0)
-#define ERTS_RBT_IS_RED(T)						\
-    ((int) ((T)->abtm.tree.parent & ERTS_HLT_PFLG_RED))
-#define ERTS_RBT_SET_RED(T)						\
-    ((T)->abtm.tree.parent |= ERTS_HLT_PFLG_RED)
-#define ERTS_RBT_IS_BLACK(T)						\
-    (!ERTS_RBT_IS_RED((T)))
-#define ERTS_RBT_SET_BLACK(T)						\
-    ((T)->abtm.tree.parent &= ~ERTS_HLT_PFLG_RED)
-#define ERTS_RBT_GET_FLAGS(T)						\
-    ((T)->abtm.tree.parent & ERTS_HLT_PFLGS_MASK)
-#define ERTS_RBT_SET_FLAGS(T, F)					\
-    do {								\
-	ERTS_HLT_ASSERT((((UWord) (F)) & ~ERTS_HLT_PFLGS_MASK) == 0);	\
-	(T)->abtm.tree.parent &= ~ERTS_HLT_PFLGS_MASK;			\
-	(T)->abtm.tree.parent |= (F);					\
-    } while (0)
-#define ERTS_RBT_GET_PARENT(T)						\
-    ((ErtsHLTimer *) ((T)->abtm.tree.parent & ~ERTS_HLT_PFLGS_MASK))
-#define ERTS_RBT_SET_PARENT(T, P)					\
-    do {								\
-	ERTS_HLT_ASSERT((((UWord) (P)) & ERTS_HLT_PFLGS_MASK) == 0);	\
-	(T)->abtm.tree.parent &= ERTS_HLT_PFLGS_MASK;			\
-	(T)->abtm.tree.parent |= (UWord) (P);				\
-    } while (0)
-#define ERTS_RBT_GET_RIGHT(T) ((T)->abtm.tree.right)
-#define ERTS_RBT_SET_RIGHT(T, R) ((T)->abtm.tree.right = (R))
-#define ERTS_RBT_GET_LEFT(T) ((T)->abtm.tree.left)
-#define ERTS_RBT_SET_LEFT(T, L) ((T)->abtm.tree.left = (L))
-#define ERTS_RBT_GET_KEY(T) ((T)->btm.refn)
-#define ERTS_RBT_IS_LT(KX, KY) refn_is_lt((KX), (KY))
-#define ERTS_RBT_IS_EQ(KX, KY) \
-    (((KX)[0] == (KY)[0]) & ((KX)[1] == (KY)[1]) & ((KX)[2] == (KY)[2]))
+#define ERTS_RBT_IS_EQ(KX, KY) refn_is_eq((KX), (KY))
 #define ERTS_RBT_WANT_DELETE
 #define ERTS_RBT_WANT_INSERT
+#ifndef ERTS_MAGIC_REF_BIF_TIMERS
 #define ERTS_RBT_WANT_LOOKUP
+#endif
 #define ERTS_RBT_WANT_FOREACH_DESTROY_YIELDING
 #define ERTS_RBT_UNDEF
 
 #include "erl_rbtree.h"
 
-#endif /* ERTS_BTM_ACCESSOR_SUPPORT */
+#endif /* !ERTS_MAGIC_REF_BIF_TIMERS */
 
 #ifdef ERTS_SMP
 static void init_canceled_queue(ErtsHLTCncldTmrQ *cq);
@@ -680,7 +740,9 @@ erts_create_timer_service(void)
     srv = erts_alloc_permanent_cache_aligned(ERTS_ALC_T_TIMER_SERVICE,
 					     sizeof(ErtsHLTimerService));
     srv->time_tree = NULL;
+#ifndef ERTS_MAGIC_REF_BIF_TIMERS
     srv->btm_tree = NULL;
+#endif
     srv->next_timeout = NULL;
     srv->yield = init_yield;
     erts_twheel_init_timer(&srv->service_timer);
@@ -697,11 +759,8 @@ erts_timer_type_size(ErtsAlcType_t type)
 {
     switch (type) {
     case ERTS_ALC_T_LL_PTIMER: return sizeof(ErtsTWTimer);
-    case ERTS_ALC_T_HL_PTIMER: return ERTS_HL_PTIMER_SIZE;
-    case ERTS_ALC_T_BIF_TIMER: return ERTS_BIF_TIMER_SIZE;
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-    case ERTS_ALC_T_ABIF_TIMER: return ERTS_ABIF_TIMER_SIZE;
-#endif
+    case ERTS_ALC_T_HL_PTIMER: return sizeof(ErtsHLTimer);
+    case ERTS_ALC_T_BIF_TIMER: return sizeof(ErtsBifTimer);
     default: ERTS_INTERNAL_ERROR("Unknown type");
     }
     return 0;
@@ -760,6 +819,111 @@ port_timeout_common(Port *port, void *tmr)
     return 0;
 }
 
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+
+static erts_smp_atomic_t *
+mbin_to_btmref__(ErtsMagicBinary *mbin)
+{
+    return erts_smp_binary_to_magic_indirection((Binary *) mbin);
+}
+
+static ERTS_INLINE void
+magic_binary_init(ErtsMagicBinary *mbin, ErtsBifTimer *tmr)
+{
+    erts_smp_atomic_t *aptr = mbin_to_btmref__(mbin);
+    erts_smp_atomic_init_nob(aptr, (erts_aint_t) tmr);
+}
+
+static ERTS_INLINE ErtsBifTimer *
+magic_binary_to_btm(ErtsMagicBinary *mbin)
+{
+    erts_smp_atomic_t *aptr = mbin_to_btmref__(mbin);
+    ErtsBifTimer *tmr = (ErtsBifTimer *) erts_smp_atomic_read_nob(aptr);
+    ERTS_HLT_ASSERT(!tmr || tmr->btm.mbin == mbin);
+    return tmr;
+}
+
+#endif /* ERTS_MAGIC_REF_BIF_TIMERS */
+
+static ERTS_INLINE erts_aint_t
+init_btm_specifics(ErtsSchedulerData *esdp,
+                   ErtsBifTimer *tmr, Eterm msg,
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+                   ErtsMagicBinary *mbin
+#else
+                   Uint32 *refn
+#endif
+    )
+{
+    Uint hsz = is_immed(msg) ? ((Uint) 0) : size_object(msg);
+    int refc;
+    if (!hsz) {
+        tmr->btm.message = msg;
+        tmr->btm.bp = NULL;
+    }
+    else {
+        ErlHeapFragment *bp = new_message_buffer(hsz);
+        Eterm *hp = bp->mem;
+        tmr->btm.message = copy_struct(msg, hsz, &hp, &bp->off_heap);
+        tmr->btm.bp = bp;
+    }
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+    refc = 1;
+    tmr->btm.mbin = mbin;
+    erts_refc_inc(&mbin->refc, 1);
+    magic_binary_init(mbin, tmr);
+    tmr->btm.proc_list.next = NULL;
+#else
+    refc = 0;
+    tmr->btm.refn[0] = refn[0];
+    tmr->btm.refn[1] = refn[1];
+    tmr->btm.refn[2] = refn[2];
+
+    tmr->btm.proc_tree.parent = ERTS_HLT_PFIELD_NOT_IN_TABLE;
+
+    btm_rbt_insert(&esdp->timer_service->btm_tree, tmr);
+#endif
+
+    erts_smp_atomic32_init_nob(&tmr->btm.state, ERTS_TMR_STATE_ACTIVE);
+    return refc; /* refc from magic binary... */
+}
+
+static void tw_bif_timer_timeout(void *vbtmp);
+
+static ERTS_INLINE void
+timer_destroy(ErtsTimer *tmr, int twt, int btm)
+{
+    if (!btm) {
+        if (twt)
+            tw_timer_free(&tmr->twt);
+        else
+            erts_free(ERTS_ALC_T_HL_PTIMER, tmr);
+    }
+    else {
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+        Binary *bp = (Binary *) tmr->btm.btm.mbin;
+        if (erts_refc_dectest(&bp->refc, 0) == 0)
+            erts_bin_free(bp);
+#endif
+	if (tmr->head.roflgs & ERTS_TMR_ROFLG_PRE_ALC)
+	    bif_timer_pre_free(&tmr->btm);
+	else
+	    erts_free(ERTS_ALC_T_BIF_TIMER, &tmr->btm);
+    }
+}
+
+static ERTS_INLINE void
+timer_pre_dec_refc(ErtsTimer *tmr)
+{
+#ifdef ERTS_HLT_DEBUG
+    erts_aint_t refc;
+    refc = erts_smp_atomic32_dec_read_nob(&tmr->head.refc);
+    ERTS_HLT_ASSERT(refc > 0);
+#else
+    erts_smp_atomic32_dec_nob(&tmr->head.refc);
+#endif
+}
+
 /*
  * Basic timer wheel timer stuff
  */
@@ -767,26 +931,39 @@ port_timeout_common(Port *port, void *tmr)
 static void
 scheduled_tw_timer_destroy(void *vtmr)
 {
-    tw_timer_free((ErtsTWTimer *) vtmr);
+    ErtsTimer * tmr = (ErtsTimer *) vtmr;
+    int btm = !!(tmr->head.roflgs & ERTS_TMR_ROFLG_BIF_TMR);
+    timer_destroy((ErtsTimer *) vtmr, 1, btm);
 }
 
 static void
 schedule_tw_timer_destroy(ErtsTWTimer *tmr)
 {
+    Uint size;
     /*
      * Reference to process/port can be
      * dropped at once...
      */
     if (tmr->head.roflgs & ERTS_TMR_ROFLG_PROC)
-	erts_proc_dec_refc((Process *) tmr->u.p);
+	erts_proc_dec_refc(tmr->head.receiver.proc);
     else if (tmr->head.roflgs & ERTS_TMR_ROFLG_PORT)
-	erts_port_dec_refc((Port *) tmr->u.p);
+	erts_port_dec_refc(tmr->head.receiver.port);
+
+    if (!(tmr->head.roflgs & ERTS_TMR_ROFLG_BIF_TMR))
+	size = sizeof(ErtsHLTimer);
+    else {
+	/* Message buffer already dropped... */
+	size = sizeof(ErtsBifTimer);
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+        size += sizeof(ErtsMagicIndirectionWord);
+#endif
+    }
 
     erts_schedule_thr_prgr_later_cleanup_op(
 	scheduled_tw_timer_destroy,
 	(void *) tmr,
-	&tmr->tw_tmr.u.cleanup,
-	sizeof(ErtsTWTimer));
+	&tmr->u.cleanup,
+	size);
 }
 
 static ERTS_INLINE void
@@ -802,7 +979,7 @@ static void
 tw_proc_timeout(void *vtwtp)
 {
     ErtsTWTimer *twtp = (ErtsTWTimer *) vtwtp;
-    Process *proc = (Process *) twtp->u.p;
+    Process *proc = twtp->head.receiver.proc;
     if (proc_timeout_common(proc, vtwtp))
 	tw_timer_dec_refc(twtp);
     tw_timer_dec_refc(twtp);
@@ -812,84 +989,126 @@ static void
 tw_port_timeout(void *vtwtp)
 {
     ErtsTWTimer *twtp = (ErtsTWTimer *) vtwtp;
-    Port *port = (Port *) twtp->u.p;
+    Port *port = twtp->head.receiver.port;
     if (port_timeout_common(port, vtwtp))
 	tw_timer_dec_refc(twtp);
     tw_timer_dec_refc(twtp);
 }
 
 static void
-tw_ptimer_cancel(void *vtwtp)
-{
-    tw_timer_dec_refc((ErtsTWTimer *) vtwtp);    
-}
-
-static void
 cancel_tw_timer(ErtsSchedulerData *esdp, ErtsTWTimer *tmr)
 {
     ERTS_HLT_ASSERT((tmr->head.roflgs & ERTS_TMR_ROFLG_SID_MASK)
 		    == (Uint32) esdp->no);
-    erts_twheel_cancel_timer(esdp->timer_wheel, &tmr->tw_tmr);
+    erts_twheel_cancel_timer(esdp->timer_wheel, &tmr->u.tw_tmr);
+    tw_timer_dec_refc(tmr);
 }
 
 static void
 tw_callback_timeout(void *vtwtp)
 {
     ErtsTWTimer *twtp = (ErtsTWTimer *) vtwtp;
-    void (*callback)(void *) = twtp->u.callback;
+    void (*callback)(void *) = twtp->head.receiver.callback;
     void *arg = twtp->head.u.arg;
     tw_timer_dec_refc(twtp);
     (*callback)(arg);
 }
 
-static ErtsTWTimer *
-create_tw_timer(ErtsSchedulerData *esdp, 
-		ErtsTmrType type, void *p,
-		void (*callback)(void *), void *arg,
-		ErtsMonotonicTime timeout_pos)
+static ErtsTimer *
+create_tw_timer(ErtsSchedulerData *esdp,
+		ErtsMonotonicTime timeout_pos,
+		int short_time, ErtsTmrType type,
+		void *rcvrp, Eterm rcvr,
+		Eterm msg,
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+                ErtsMagicBinary *mbin,
+#else
+                Uint32 *refn,
+#endif
+		void (*callback)(void *), void *arg)
 {
     ErtsTWTimer *tmr;
     void (*timeout_func)(void *);
-    void (*cancel_func)(void *);
     erts_aint32_t refc;
 
-    tmr = tw_timer_alloc();
-    erts_twheel_init_timer(&tmr->tw_tmr);
-
-    tmr->head.roflgs = (Uint32) esdp->no;
-    ERTS_HLT_ASSERT((tmr->head.roflgs
-		     & ~ERTS_TMR_ROFLG_SID_MASK) == 0);
+    if (type != ERTS_TMR_BIF) {
+        tmr = tw_timer_alloc();
+        tmr->head.roflgs = 0;
+    }
+    else {
+        if (short_time) {
+            tmr = (ErtsTWTimer *) bif_timer_pre_alloc();
+            if (!tmr)
+                goto alloc_bif_timer;
+            tmr->head.roflgs = (ERTS_TMR_ROFLG_BIF_TMR
+                                | ERTS_TMR_ROFLG_PRE_ALC);
+        }
+        else {
+        alloc_bif_timer:
+            tmr = (ErtsTWTimer *) erts_alloc(ERTS_ALC_T_BIF_TIMER,
+                                             sizeof(ErtsBifTimer));
+            tmr->head.roflgs = ERTS_TMR_ROFLG_BIF_TMR;
+        }
+    }
+
+    erts_twheel_init_timer(&tmr->u.tw_tmr);
+    tmr->head.roflgs |= (Uint32) esdp->no;
+    ERTS_HLT_ASSERT((((Uint32) esdp->no)
+                     & ~ERTS_TMR_ROFLG_SID_MASK) == 0);
 
     switch (type) {
 
     case ERTS_TMR_PROC:
-	tmr->u.p = p;
+	tmr->head.receiver.proc = (Process *) rcvrp;
 	tmr->head.roflgs |= ERTS_TMR_ROFLG_PROC;
 	timeout_func = tw_proc_timeout;
-	cancel_func = tw_ptimer_cancel;
-	erts_proc_inc_refc((Process *) p);
+	erts_proc_inc_refc((Process *) rcvrp);
 	refc = 2;
 	break;
 
     case ERTS_TMR_PORT:
-	tmr->u.p = p;
+	tmr->head.receiver.port = (Port *) rcvrp;
 	tmr->head.roflgs |= ERTS_TMR_ROFLG_PORT;
 	timeout_func = tw_port_timeout;
-	cancel_func = tw_ptimer_cancel;
-	erts_port_inc_refc((Port *) p);
+	erts_port_inc_refc((Port *) rcvrp);
 	refc = 2;
 	break;
 
     case ERTS_TMR_CALLBACK:
 	tmr->head.u.arg = arg;
-	tmr->u.callback = callback;
+	tmr->head.receiver.callback = callback;
 
 	tmr->head.roflgs |= ERTS_TMR_ROFLG_CALLBACK;
 	timeout_func = tw_callback_timeout;
-	cancel_func = NULL;
 	refc = 1;
 	break;
 
+    case ERTS_TMR_BIF:
+
+	timeout_func = tw_bif_timer_timeout;
+	if (is_internal_pid(rcvr)) {
+	    tmr->head.roflgs |= ERTS_TMR_ROFLG_PROC;
+	    tmr->head.receiver.proc = (Process *) rcvrp;
+	    refc = 2;
+	}
+	else {
+	    ERTS_HLT_ASSERT(is_atom(rcvr));
+	    tmr->head.roflgs |= ERTS_TMR_ROFLG_REG_NAME;
+	    tmr->head.receiver.name = (Eterm) rcvr;
+	    refc = 1;
+	}
+
+        refc += init_btm_specifics(esdp,
+                                   (ErtsBifTimer *) tmr,
+                                   msg,
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+                                   mbin
+#else
+                                   refn
+#endif
+            );
+        break;
+
     default:
 	ERTS_INTERNAL_ERROR("Unsupported timer type");
 	return NULL;
@@ -898,41 +1117,24 @@ create_tw_timer(ErtsSchedulerData *esdp,
     erts_smp_atomic32_init_nob(&tmr->head.refc, refc);
 
     erts_twheel_set_timer(esdp->timer_wheel,
-			  &tmr->tw_tmr,
+			  &tmr->u.tw_tmr,
 			  timeout_func,
-			  cancel_func,
 			  tmr,
 			  timeout_pos);
 
-    return tmr;
+    return (ErtsTimer *) tmr;
 }
 
 /*
  * Basic high level timer stuff
  */
 
-static ERTS_INLINE void
-hl_timer_destroy(ErtsHLTimer *tmr)
-{
-    Uint32 roflgs = tmr->head.roflgs;
-    if (!(roflgs & ERTS_TMR_ROFLG_BIF_TMR))
-	erts_free(ERTS_ALC_T_HL_PTIMER, tmr);
-    else {
-	if (roflgs & ERTS_TMR_ROFLG_PRE_ALC)
-	    bif_timer_pre_free(tmr);
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-	else if (roflgs & ERTS_TMR_ROFLG_ABIF_TMR)
-	    erts_free(ERTS_ALC_T_ABIF_TIMER, tmr);
-#endif
-	else
-	    erts_free(ERTS_ALC_T_BIF_TIMER, tmr);
-    }
-}
-
 static void
 scheduled_hl_timer_destroy(void *vtmr)
 {
-    hl_timer_destroy((ErtsHLTimer *) vtmr);
+    ErtsTimer * tmr = (ErtsTimer *) vtmr;
+    int btm = !!(tmr->head.roflgs & ERTS_TMR_ROFLG_BIF_TMR);
+    timer_destroy((ErtsTimer *) vtmr, 0, btm);
 }
 
 static void
@@ -948,25 +1150,25 @@ schedule_hl_timer_destroy(ErtsHLTimer *tmr, Uint32 roflgs)
     ERTS_HLT_ASSERT(erts_smp_atomic32_read_nob(&tmr->head.refc) == 0);
 
     if (roflgs & ERTS_TMR_ROFLG_REG_NAME) {
-	ERTS_HLT_ASSERT(is_atom(tmr->receiver.name));
+	ERTS_HLT_ASSERT(is_atom(tmr->head.receiver.name));
     }
     else if (roflgs & ERTS_TMR_ROFLG_PROC) {
-	ERTS_HLT_ASSERT(tmr->receiver.proc);
-	erts_proc_dec_refc(tmr->receiver.proc);
+	ERTS_HLT_ASSERT(tmr->head.receiver.proc);
+	erts_proc_dec_refc(tmr->head.receiver.proc);
     }
     else if (roflgs & ERTS_TMR_ROFLG_PORT) {
-	ERTS_HLT_ASSERT(tmr->receiver.port);
-	erts_port_dec_refc(tmr->receiver.port);
+	ERTS_HLT_ASSERT(tmr->head.receiver.port);
+	erts_port_dec_refc(tmr->head.receiver.port);
     }
 
     if (!(roflgs & ERTS_TMR_ROFLG_BIF_TMR))
-	size = ERTS_HL_PTIMER_SIZE;
-    else {
-	/*
-	 * Message buffer can be dropped at
-	 * once...
-	 */
 	size = sizeof(ErtsHLTimer);
+    else {
+	/* Message buffer already dropped... */
+	size = sizeof(ErtsBifTimer);
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+        size += sizeof(ErtsMagicIndirectionWord);
+#endif
     }
 
     erts_schedule_thr_prgr_later_cleanup_op(
@@ -975,18 +1177,6 @@ schedule_hl_timer_destroy(ErtsHLTimer *tmr, Uint32 roflgs)
 }
 
 static ERTS_INLINE void
-hl_timer_pre_dec_refc(ErtsHLTimer *tmr)
-{
-#ifdef ERTS_HLT_DEBUG
-    erts_aint_t refc;
-    refc = erts_smp_atomic32_dec_read_nob(&tmr->head.refc);
-    ERTS_HLT_ASSERT(refc > 0);
-#else
-    erts_smp_atomic32_dec_nob(&tmr->head.refc);
-#endif
-}
-
-static ERTS_INLINE void
 hl_timer_dec_refc(ErtsHLTimer *tmr, Uint32 roflgs)
 {
     if (erts_smp_atomic32_dec_read_relb(&tmr->head.refc) == 0) {
@@ -1018,39 +1208,135 @@ check_canceled_queue(ErtsSchedulerData *esdp, ErtsHLTimerService *srv)
 #endif
 }
 
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
 
-static void
-hlt_delete_abtm(ErtsHLTimer *tmr)
+static int
+bif_timer_ref_destructor(Binary *unused)
 {
-    Process *proc;
+    return 1;
+}
 
-    ERTS_HLT_ASSERT(tmr->head.roflgs & ERTS_TMR_ROFLG_ABIF_TMR);
+static ERTS_INLINE void
+btm_clear_magic_binary(ErtsBifTimer *tmr)
+{
+    erts_smp_atomic_t *aptr = mbin_to_btmref__(tmr->btm.mbin);
+    Uint32 roflgs = tmr->type.head.roflgs;
+#ifdef ERTS_HLT_DEBUG
+    erts_aint_t tval = erts_smp_atomic_xchg_nob(aptr,
+                                                (erts_aint_t) NULL);
+    ERTS_HLT_ASSERT(tval == (erts_aint_t) tmr);
+#else
+    erts_smp_atomic_set_nob(aptr, (erts_aint_t) NULL);
+#endif
+    if (roflgs & ERTS_TMR_ROFLG_HLT)
+        hl_timer_dec_refc(&tmr->type.hlt, roflgs);
+    else
+        tw_timer_dec_refc(&tmr->type.twt);
+}
 
-    proc = erts_proc_lookup(tmr->abtm.accessor);
+#endif /* ERTS_MAGIC_REF_BIF_TIMERS */
 
-    if (proc) {
-	int deref = 0;
-	erts_smp_proc_lock(proc, ERTS_PROC_LOCK_BTM);
-	if (tmr->abtm.tree.parent != ERTS_HLT_PFIELD_NOT_IN_TABLE) {
-	    abtm_rbt_delete(&proc->accessor_bif_timers, tmr);
-	    deref = 1;
-	    tmr->abtm.tree.parent = ERTS_HLT_PFIELD_NOT_IN_TABLE;
-	}
-	erts_smp_proc_unlock(proc, ERTS_PROC_LOCK_BTM);
-	if (deref)
-	    hl_timer_pre_dec_refc(tmr);
+static ERTS_INLINE void
+bif_timer_timeout(ErtsHLTimerService *srv,
+                  ErtsBifTimer *tmr,
+                  Uint32 roflgs)
+{
+    erts_aint32_t state;
+
+    ERTS_HLT_ASSERT(tmr->type.head.roflgs == roflgs);    
+    ERTS_HLT_ASSERT(roflgs & ERTS_TMR_ROFLG_BIF_TMR);
+
+    state = erts_smp_atomic32_cmpxchg_acqb(&tmr->btm.state,
+					   ERTS_TMR_STATE_TIMED_OUT,
+					   ERTS_TMR_STATE_ACTIVE);
+
+    ERTS_HLT_ASSERT(state == ERTS_TMR_STATE_CANCELED
+		    || state == ERTS_TMR_STATE_ACTIVE);
+
+    if (state == ERTS_TMR_STATE_ACTIVE) {
+        Process *proc;
+
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+        btm_clear_magic_binary(tmr);
+#endif
+
+        if (roflgs & ERTS_TMR_ROFLG_REG_NAME) {
+            Eterm term;
+            term = tmr->type.head.receiver.name;
+            ERTS_HLT_ASSERT(is_atom(term));
+            term = erts_whereis_name_to_id(NULL, term);
+            proc = erts_proc_lookup(term);
+        }
+        else {
+            ERTS_HLT_ASSERT(roflgs & ERTS_TMR_ROFLG_PROC);
+            proc = tmr->type.head.receiver.proc;
+            ERTS_HLT_ASSERT(proc);
+        }
+        if (proc) {
+            if (!ERTS_PROC_IS_EXITING(proc)) {
+                int dec_refc = 0;
+                ErtsMessage *mp = erts_alloc_message(0, NULL);
+                mp->data.heap_frag = tmr->btm.bp;
+                tmr->btm.bp = NULL;
+                erts_queue_message(proc, 0, mp, tmr->btm.message,
+                                   am_clock_service);
+                erts_smp_proc_lock(proc, ERTS_PROC_LOCK_BTM);
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+                if (tmr->btm.proc_list.next) {
+                    proc_btm_list_delete(&proc->bif_timers, tmr);
+                    dec_refc = 1;
+                }
+#else
+                if (tmr->btm.proc_tree.parent != ERTS_HLT_PFIELD_NOT_IN_TABLE) {
+                    proc_btm_rbt_delete(&proc->bif_timers, tmr);
+                    tmr->btm.proc_tree.parent = ERTS_HLT_PFIELD_NOT_IN_TABLE;
+                    dec_refc = 1;
+                }
+#endif
+                erts_smp_proc_unlock(proc, ERTS_PROC_LOCK_BTM);
+                if (dec_refc)
+                    timer_pre_dec_refc((ErtsTimer *) tmr);
+            }
+        }
+        if (tmr->btm.bp)
+            free_message_buffer(tmr->btm.bp);
     }
+
+#ifndef ERTS_MAGIC_REF_BIF_TIMERS
+    if (tmr->btm.tree.parent != ERTS_HLT_PFIELD_NOT_IN_TABLE) {
+	btm_rbt_delete(&srv->btm_tree, tmr);
+	tmr->btm.tree.parent = ERTS_HLT_PFIELD_NOT_IN_TABLE;
+    }
+#endif
+
+
 }
 
+static void
+tw_bif_timer_timeout(void *vbtmp)
+{
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+    ErtsHLTimerService *srv = NULL;
+#else
+    ErtsSchedulerData *esdp = erts_get_scheduler_data();
+    ErtsHLTimerService *srv = esdp->timer_service;
 #endif
+    ErtsBifTimer *btmp = (ErtsBifTimer *) vbtmp;
+    bif_timer_timeout(srv, btmp, btmp->type.head.roflgs);
+    tw_timer_dec_refc(&btmp->type.twt);
+}
 
-static ErtsHLTimer *
+static ErtsTimer *
 create_hl_timer(ErtsSchedulerData *esdp,
 		ErtsMonotonicTime timeout_pos,
 		int short_time, ErtsTmrType type,
-		void *rcvrp, Eterm rcvr, Eterm acsr,
-		Eterm msg, Uint32 *refn,
+		void *rcvrp, Eterm rcvr,
+		Eterm msg,
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+                ErtsMagicBinary *mbin,
+#else
+                Uint32 *refn,
+#endif
 		void (*callback)(void *), void *arg)
 {
     ErtsHLTimerService *srv = esdp->timer_service;
@@ -1069,7 +1355,7 @@ create_hl_timer(ErtsSchedulerData *esdp,
     if (type != ERTS_TMR_BIF) {
 
 	tmr = erts_alloc(ERTS_ALC_T_HL_PTIMER,
-			 ERTS_HL_PTIMER_SIZE);
+                         sizeof(ErtsHLTimer));
 	tmr->timeout = timeout_pos;
 
 	switch (type) {
@@ -1078,7 +1364,7 @@ create_hl_timer(ErtsSchedulerData *esdp,
 	    ERTS_HLT_ASSERT(is_internal_pid(rcvr));
 
 	    erts_proc_inc_refc((Process *) rcvrp);
-	    tmr->receiver.proc = (Process *) rcvrp;
+	    tmr->head.receiver.proc = (Process *) rcvrp;
 	    roflgs |= ERTS_TMR_ROFLG_PROC;
 	    refc = 2;
 	    break;
@@ -1086,14 +1372,14 @@ create_hl_timer(ErtsSchedulerData *esdp,
 	case ERTS_TMR_PORT:
 	    ERTS_HLT_ASSERT(is_internal_port(rcvr));
 	    erts_port_inc_refc((Port *) rcvrp);
-	    tmr->receiver.port = (Port *) rcvrp;
+	    tmr->head.receiver.port = (Port *) rcvrp;
 	    roflgs |= ERTS_TMR_ROFLG_PORT;
 	    refc = 2;
 	    break;
 
 	case ERTS_TMR_CALLBACK:
 	    roflgs |= ERTS_TMR_ROFLG_CALLBACK;
-	    tmr->receiver.callback = callback;
+	    tmr->head.receiver.callback = callback;
 	    tmr->head.u.arg = arg;
 	    refc = 1;
 	    break;
@@ -1105,84 +1391,47 @@ create_hl_timer(ErtsSchedulerData *esdp,
 
     }
     else { /* ERTS_TMR_BIF */
-	Uint hsz;
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-	int is_abif_tmr = is_value(acsr) && acsr != rcvr;
-#endif
 
 	if (short_time) {
-	    tmr = bif_timer_pre_alloc();
+	    tmr = (ErtsHLTimer *) bif_timer_pre_alloc();
 	    if (!tmr)
 		goto alloc_bif_timer;
 	    roflgs |= ERTS_TMR_ROFLG_PRE_ALC;
 	}
 	else {
 	alloc_bif_timer:
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-	    if (is_abif_tmr)
-		tmr = erts_alloc(ERTS_ALC_T_ABIF_TIMER,
-				 ERTS_ABIF_TIMER_SIZE);
-	    else
-#endif
-		tmr = erts_alloc(ERTS_ALC_T_BIF_TIMER,
-				 ERTS_BIF_TIMER_SIZE);
-	}
+            tmr = (ErtsHLTimer *) erts_alloc(ERTS_ALC_T_BIF_TIMER,
+                                             sizeof(ErtsBifTimer));
+        }
 
 	tmr->timeout = timeout_pos;
 
 	roflgs |= ERTS_TMR_ROFLG_BIF_TMR;
 	if (is_internal_pid(rcvr)) {
 	    roflgs |= ERTS_TMR_ROFLG_PROC;
-	    tmr->receiver.proc = (Process *) rcvrp;
+	    tmr->head.receiver.proc = (Process *) rcvrp;
 	    refc = 2;
 	}
 	else {
 	    ERTS_HLT_ASSERT(is_atom(rcvr));
 	    roflgs |= ERTS_TMR_ROFLG_REG_NAME;
-	    tmr->receiver.name = rcvr;
+	    tmr->head.receiver.name = rcvr;
 	    refc = 1;
 	}
 
-	hsz = is_immed(msg) ? ((Uint) 0) : size_object(msg);
-	if (!hsz) {
-	    tmr->btm.message = msg;
-	    tmr->btm.bp = NULL;
-	}
-	else {
-	    ErlHeapFragment *bp = new_message_buffer(hsz);
-	    Eterm *hp = bp->mem;
-	    tmr->btm.message = copy_struct(msg, hsz, &hp, &bp->off_heap);
-	    tmr->btm.bp = bp;
-	}
-	tmr->btm.refn[0] = refn[0];
-	tmr->btm.refn[1] = refn[1];
-	tmr->btm.refn[2] = refn[2];
-
-	tmr->btm.proc_tree.parent = ERTS_HLT_PFIELD_NOT_IN_TABLE;
-
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-	if (is_abif_tmr) {
-	    Process *aproc;
-	    roflgs |= ERTS_TMR_ROFLG_ABIF_TMR;
-	    tmr->abtm.accessor = acsr;
-	    aproc = erts_proc_lookup(acsr);
-	    if (!aproc)
-		tmr->abtm.tree.parent = ERTS_HLT_PFIELD_NOT_IN_TABLE;
-	    else {
-		refc++;
-		erts_smp_proc_lock(aproc, ERTS_PROC_LOCK_BTM);
-		abtm_rbt_insert(&aproc->accessor_bif_timers, tmr);
-		erts_smp_proc_unlock(aproc, ERTS_PROC_LOCK_BTM);
-	    }
-	}
+        refc += init_btm_specifics(esdp,
+                                   (ErtsBifTimer *) tmr,
+                                   msg,
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+                                   mbin
+#else
+                                   refn
 #endif
-
-	btm_rbt_insert(&srv->btm_tree, tmr);
+            );
     }
 
     tmr->head.roflgs = roflgs;
     erts_smp_atomic32_init_nob(&tmr->head.refc, refc);
-    erts_smp_atomic32_init_nob(&tmr->state, ERTS_TMR_STATE_ACTIVE);
 
     if (!srv->next_timeout
 	|| tmr->timeout < srv->next_timeout->timeout) {
@@ -1192,7 +1441,6 @@ create_hl_timer(ErtsSchedulerData *esdp,
 	erts_twheel_set_timer(esdp->timer_wheel,
 			      &srv->service_timer,
 			      hlt_service_timeout,
-			      NULL,
 			      (void *) esdp,
 			      tmr->timeout);
 	srv->next_timeout = tmr;
@@ -1209,79 +1457,20 @@ create_hl_timer(ErtsSchedulerData *esdp,
 
     ERTS_HLT_HDBG_CHK_SRV(srv);
 
-    return tmr;
-}
-
-static ERTS_INLINE void
-hlt_bif_timer_timeout(ErtsHLTimer *tmr, Uint32 roflgs)
-{
-    ErtsProcLocks proc_locks = ERTS_PROC_LOCKS_MSG_SEND;
-    Process *proc;
-    int queued_message = 0;
-    int dec_refc = 0;
-    Uint32 is_reg_name = (roflgs & ERTS_TMR_ROFLG_REG_NAME);
-    ERTS_HLT_ASSERT(roflgs & ERTS_TMR_ROFLG_BIF_TMR);
-
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-    if (tmr->head.roflgs & ERTS_TMR_ROFLG_ABIF_TMR)
-	hlt_delete_abtm(tmr);
-#endif
-
-    if (is_reg_name) {
-	Eterm pid;
-	ERTS_HLT_ASSERT(is_atom(tmr->receiver.name));
-	pid = erts_whereis_name_to_id(NULL, tmr->receiver.name);
-	proc = erts_proc_lookup(pid);
-    }
-    else {
-	ERTS_HLT_ASSERT(roflgs & ERTS_TMR_ROFLG_PROC);
-	ERTS_HLT_ASSERT(tmr->receiver.proc);
-
-	proc = tmr->receiver.proc;
-	proc_locks |= ERTS_PROC_LOCK_BTM;
-    }
-    if (proc) {
-	erts_smp_proc_lock(proc, proc_locks);
-	/*
-	 * If process is exiting, let it clean up
-	 * the btm tree by itself (it may be in
-	 * the middle of tree destruction).
-	 */
-	if (!ERTS_PROC_IS_EXITING(proc)) {
-	    ErtsMessage *mp = erts_alloc_message(0, NULL);
-	    mp->data.heap_frag = tmr->btm.bp;
-	    erts_queue_message(proc, proc_locks, mp,
-			       tmr->btm.message, am_clock_service);
-	    erts_smp_proc_unlock(proc, ERTS_PROC_LOCKS_MSG_SEND);
-	    queued_message = 1;
-	    proc_locks &= ~ERTS_PROC_LOCKS_MSG_SEND;
-	    tmr->btm.bp = NULL;
-	    if (tmr->btm.proc_tree.parent != ERTS_HLT_PFIELD_NOT_IN_TABLE) {
-		proc_btm_rbt_delete(&proc->bif_timers, tmr);
-		tmr->btm.proc_tree.parent = ERTS_HLT_PFIELD_NOT_IN_TABLE;
-		dec_refc = 1;
-	    }
-	}
-	if (proc_locks)
-	    erts_smp_proc_unlock(proc, proc_locks);
-	if (dec_refc)
-	    hl_timer_pre_dec_refc(tmr);
-    }
-    if (!queued_message && tmr->btm.bp)
-	free_message_buffer(tmr->btm.bp);
+    return (ErtsTimer *) tmr;
 }
 
 static ERTS_INLINE void
 hlt_proc_timeout(ErtsHLTimer *tmr)
 {
-    if (proc_timeout_common(tmr->receiver.proc, (void *) tmr))
+    if (proc_timeout_common(tmr->head.receiver.proc, (void *) tmr))
 	hl_timer_dec_refc(tmr, tmr->head.roflgs);
 }
 
 static ERTS_INLINE void
 hlt_port_timeout(ErtsHLTimer *tmr)
 {
-    if (port_timeout_common(tmr->receiver.port, (void *) tmr))
+    if (port_timeout_common(tmr->head.receiver.port, (void *) tmr))
 	hl_timer_dec_refc(tmr, tmr->head.roflgs);
 }
 
@@ -1289,41 +1478,24 @@ static void hlt_timeout(ErtsHLTimer *tmr, void *vsrv)
 {
     ErtsHLTimerService *srv = (ErtsHLTimerService *) vsrv;
     Uint32 roflgs;
-    erts_aint32_t state;
 
     ERTS_HLT_HDBG_CHK_SRV(srv);
 
     roflgs = tmr->head.roflgs;
     ERTS_HLT_ASSERT(roflgs & ERTS_TMR_ROFLG_HLT);
 
-    state = erts_smp_atomic32_cmpxchg_acqb(&tmr->state,
-					   ERTS_TMR_STATE_TIMED_OUT,
-					   ERTS_TMR_STATE_ACTIVE);
-
-    ERTS_HLT_ASSERT(state == ERTS_TMR_STATE_CANCELED
-		    || state == ERTS_TMR_STATE_ACTIVE);
-
-    if (state == ERTS_TMR_STATE_ACTIVE) {
-
-	if (roflgs & ERTS_TMR_ROFLG_BIF_TMR)
-	    hlt_bif_timer_timeout(tmr, roflgs);
-	else if (roflgs & ERTS_TMR_ROFLG_PROC)
-	    hlt_proc_timeout(tmr);
-	else if (roflgs & ERTS_TMR_ROFLG_PORT)
-	    hlt_port_timeout(tmr);
-	else {
-	    ERTS_HLT_ASSERT(roflgs & ERTS_TMR_ROFLG_CALLBACK);
-	    (*tmr->receiver.callback)(tmr->head.u.arg);
-	}
-
+    if (roflgs & ERTS_TMR_ROFLG_BIF_TMR)
+        bif_timer_timeout(srv, (ErtsBifTimer *) tmr, roflgs);
+    else if (roflgs & ERTS_TMR_ROFLG_PROC)
+        hlt_proc_timeout(tmr);
+    else if (roflgs & ERTS_TMR_ROFLG_PORT)
+        hlt_port_timeout(tmr);
+    else {
+        ERTS_HLT_ASSERT(roflgs & ERTS_TMR_ROFLG_CALLBACK);
+        (*tmr->head.receiver.callback)(tmr->head.u.arg);
     }
 
     tmr->time.tree.parent = ERTS_HLT_PFIELD_NOT_IN_TABLE;
-    if ((roflgs & ERTS_TMR_ROFLG_BIF_TMR)
-	&& tmr->btm.tree.parent != ERTS_HLT_PFIELD_NOT_IN_TABLE) {
-	btm_rbt_delete(&srv->btm_tree, tmr);
-	tmr->btm.tree.parent = ERTS_HLT_PFIELD_NOT_IN_TABLE;
-    }
 
     ERTS_HLT_HDBG_CHK_SRV(srv);
 
@@ -1390,7 +1562,6 @@ hlt_service_timeout(void *vesdp)
 	erts_twheel_set_timer(esdp->timer_wheel,
 			      &srv->service_timer,
 			      hlt_service_timeout,
-			      NULL,
 			      vesdp,
 			      tmr->timeout);
 }
@@ -1402,19 +1573,6 @@ hlt_delete_timer(ErtsSchedulerData *esdp, ErtsHLTimer *tmr)
 
     ERTS_HLT_HDBG_CHK_SRV(srv);
 
-    if (tmr->head.roflgs & ERTS_TMR_ROFLG_BIF_TMR) {
-
-	if (tmr->btm.tree.parent != ERTS_HLT_PFIELD_NOT_IN_TABLE) {
-	    btm_rbt_delete(&srv->btm_tree, tmr);
-	    tmr->btm.tree.parent = ERTS_HLT_PFIELD_NOT_IN_TABLE;
-	}
-
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-	if (tmr->head.roflgs & ERTS_TMR_ROFLG_ABIF_TMR)
-	    hlt_delete_abtm(tmr);
-#endif
-    }
-
     if (tmr->time.tree.parent == ERTS_HLT_PFIELD_NOT_IN_TABLE) {
 	/* Already removed... */
 	ERTS_HLT_HDBG_CHK_SRV(srv);
@@ -1460,7 +1618,6 @@ hlt_delete_timer(ErtsSchedulerData *esdp, ErtsHLTimer *tmr)
 		erts_twheel_set_timer(esdp->timer_wheel,
 				      &srv->service_timer,
 				      hlt_service_timeout,
-				      NULL,
 				      (void *) esdp,
 				      smlst->timeout);
 	    }
@@ -1485,6 +1642,17 @@ cleanup_sched_local_canceled_timer(ErtsSchedulerData *esdp,
     ERTS_HLT_ASSERT(esdp == erts_get_scheduler_data());
     ERTS_HLT_ASSERT((tmr->head.roflgs & ERTS_TMR_ROFLG_SID_MASK)
 		    == (Uint32) esdp->no);
+
+#ifndef ERTS_MAGIC_REF_BIF_TIMERS
+    if (roflgs & ERTS_TMR_ROFLG_BIF_TMR) {
+        ErtsBifTimer *btm = (ErtsBifTimer *) tmr;
+	if (btm->btm.tree.parent != ERTS_HLT_PFIELD_NOT_IN_TABLE) {
+	    btm_rbt_delete(&esdp->timer_service->btm_tree, btm);
+	    btm->btm.tree.parent = ERTS_HLT_PFIELD_NOT_IN_TABLE;
+	}
+    }
+#endif
+
     if (roflgs & ERTS_TMR_ROFLG_HLT) {
 	hlt_delete_timer(esdp, &tmr->hlt);
 	hl_timer_dec_refc(&tmr->hlt, roflgs);
@@ -1750,57 +1918,86 @@ continue_cancel_ptimer(ErtsSchedulerData *esdp, ErtsTimer *tmr)
  * BIF timer specific
  */
 
+
 Uint erts_bif_timer_memory_size(void)
 {
     return (Uint) 0;
 }
 
 static BIF_RETTYPE
-setup_bif_timer(Process *c_p, ErtsMonotonicTime timeout_pos,
-		int short_time, Eterm rcvr, Eterm acsr,
-		Eterm msg, int wrap)
+setup_bif_timer(Process *c_p, int twheel, ErtsMonotonicTime timeout_pos,
+		int short_time, Eterm rcvr, Eterm msg, int wrap)
 {
     BIF_RETTYPE ret;
     Eterm ref, tmo_msg, *hp;
-    ErtsHLTimer *tmr;
+    ErtsBifTimer *tmr;
     ErtsSchedulerData *esdp;
-    DeclareTmpHeap(tmp_hp, 4, c_p);
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+    Binary *mbin;
+#endif
+    Eterm tmp_hp[4];
+    ErtsCreateTimerFunc create_timer;
 
     if (is_not_internal_pid(rcvr) && is_not_atom(rcvr))
 	goto badarg;
 
     esdp = erts_proc_sched_data(c_p);
 
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+    mbin = erts_create_magic_indirection(bif_timer_ref_destructor);
+    hp = HAlloc(c_p, ERTS_MAGIC_REF_THING_SIZE);
+    ref = erts_mk_magic_ref(&hp, &c_p->off_heap, mbin);
+    ASSERT(erts_get_ref_numbers_thr_id(((ErtsMagicBinary *)mbin)->refn)
+           == (Uint32) esdp->no);
+#else
     hp = HAlloc(c_p, ERTS_REF_THING_SIZE);
     ref = erts_sched_make_ref_in_buffer(esdp, hp);
-
-    ASSERT(erts_get_ref_numbers_thr_id(
-	       internal_ref_numbers(ref)) == (Uint32) esdp->no);
-
-    UseTmpHeap(4, c_p);
+    ASSERT(erts_get_ref_numbers_thr_id(internal_ordinary_ref_numbers(ref))
+           == (Uint32) esdp->no);
+#endif
 
     tmo_msg = wrap ? TUPLE3(tmp_hp, am_timeout, ref, msg) : msg;
 
-    tmr = create_hl_timer(esdp, timeout_pos, short_time,
-			  ERTS_TMR_BIF, NULL,  rcvr, acsr, tmo_msg,
-			  internal_ref_numbers(ref), NULL, NULL);
-
-    UnUseTmpHeap(4, c_p);
+    create_timer = twheel ? create_tw_timer : create_hl_timer;
+    tmr = (ErtsBifTimer *) create_timer(esdp, timeout_pos,
+                                        short_time, ERTS_TMR_BIF,
+                                        NULL, rcvr, tmo_msg,
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+                                        (ErtsMagicBinary *) mbin,
+#else
+                                        internal_ordinary_ref_numbers(ref),
+#endif
+                                        NULL, NULL);
 
     if (is_internal_pid(rcvr)) {
 	Process *proc = erts_pid2proc_opt(c_p, ERTS_PROC_LOCK_MAIN,
 					  rcvr, ERTS_PROC_LOCK_BTM,
 					  ERTS_P2P_FLG_INC_REFC);
 	if (!proc) {
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+            btm_clear_magic_binary(tmr);
+#else
+            if (tmr->btm.tree.parent != ERTS_HLT_PFIELD_NOT_IN_TABLE) {
+                btm_rbt_delete(&esdp->timer_service->btm_tree, tmr);
+                tmr->btm.tree.parent = ERTS_HLT_PFIELD_NOT_IN_TABLE;
+            }
+#endif
 	    if (tmr->btm.bp)
 		free_message_buffer(tmr->btm.bp);
-	    hlt_delete_timer(esdp, tmr);
-	    hl_timer_destroy(tmr);
+            if (twheel)
+                cancel_tw_timer(esdp, &tmr->type.twt);
+            else
+                hlt_delete_timer(esdp, &tmr->type.hlt);
+            timer_destroy((ErtsTimer *) tmr, twheel, 1);
 	}
 	else {
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+            proc_btm_list_insert(&proc->bif_timers, tmr);
+#else
 	    proc_btm_rbt_insert(&proc->bif_timers, tmr);
+#endif
 	    erts_smp_proc_unlock(proc, ERTS_PROC_LOCK_BTM);
-	    tmr->receiver.proc = proc;
+            tmr->type.head.receiver.proc = proc;
 	}
     }
 
@@ -1814,27 +2011,33 @@ badarg:
 }
 
 static int
-cancel_bif_timer(ErtsHLTimer *tmr)
+cancel_bif_timer(ErtsBifTimer *tmr)
 {
     erts_aint_t state;
     Uint32 roflgs;
     int res;
 
-    state = erts_smp_atomic32_cmpxchg_acqb(&tmr->state,
+    state = erts_smp_atomic32_cmpxchg_acqb(&tmr->btm.state,
 					   ERTS_TMR_STATE_CANCELED,
 					   ERTS_TMR_STATE_ACTIVE);
     if (state != ERTS_TMR_STATE_ACTIVE)
 	return 0;
 
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+    btm_clear_magic_binary(tmr);
+#endif
+
     if (tmr->btm.bp)
 	free_message_buffer(tmr->btm.bp);
 
     res = -1;
 
-    roflgs = tmr->head.roflgs;
+    roflgs = tmr->type.head.roflgs;
     if (roflgs & ERTS_TMR_ROFLG_PROC) {
-	Process *proc = tmr->receiver.proc;
-	ERTS_HLT_ASSERT(!(tmr->head.roflgs & ERTS_TMR_ROFLG_REG_NAME));
+	Process *proc;
+
+        proc = tmr->type.head.receiver.proc;
+	ERTS_HLT_ASSERT(!(tmr->type.head.roflgs & ERTS_TMR_ROFLG_REG_NAME));
 
 	erts_smp_proc_lock(proc, ERTS_PROC_LOCK_BTM);
 	/*
@@ -1842,29 +2045,238 @@ cancel_bif_timer(ErtsHLTimer *tmr)
 	 * the btm tree by itself (it may be in
 	 * the middle of tree destruction).
 	 */
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+        if (!ERTS_PROC_IS_EXITING(proc) && tmr->btm.proc_list.next) {
+            proc_btm_list_delete(&proc->bif_timers, tmr);
+            res = 1;
+        }
+#else
 	if (!ERTS_PROC_IS_EXITING(proc)
 	    && tmr->btm.proc_tree.parent != ERTS_HLT_PFIELD_NOT_IN_TABLE) {
 	    proc_btm_rbt_delete(&proc->bif_timers, tmr);
 	    tmr->btm.proc_tree.parent = ERTS_HLT_PFIELD_NOT_IN_TABLE;
 	    res = 1;
 	}
+#endif
 	erts_smp_proc_unlock(proc, ERTS_PROC_LOCK_BTM);
     }
 
     return res;
 }
 
+static ERTS_INLINE Sint64
+access_btm(ErtsBifTimer *tmr, Uint32 sid, ErtsSchedulerData *esdp, int cancel)
+{
+    int cncl_res;
+    Sint64 time_left;
+    ErtsMonotonicTime timeout;
+    int is_hlt;
+
+    if (!tmr)
+        return -1;
+
+    is_hlt = !!(tmr->type.head.roflgs & ERTS_TMR_ROFLG_HLT);
+    timeout = (is_hlt
+               ? tmr->type.hlt.timeout
+               : erts_tweel_read_timeout(&tmr->type.twt.u.tw_tmr));
+
+    if (!cancel) {
+        erts_aint32_t state = erts_smp_atomic32_read_acqb(&tmr->btm.state);
+        if (state == ERTS_TMR_STATE_ACTIVE)
+            return get_time_left(esdp, timeout);
+        return -1;
+    }
+
+    cncl_res = cancel_bif_timer(tmr);
+    if (!cncl_res)
+        return -1;
+
+    time_left = get_time_left(esdp, timeout);
+
+    if (sid != (Uint32) esdp->no) {
+        if (cncl_res > 0)
+            queue_canceled_timer(esdp, sid, (ErtsTimer *) tmr);
+    }
+    else {
+#ifndef ERTS_MAGIC_REF_BIF_TIMERS
+        if (tmr->btm.tree.parent != ERTS_HLT_PFIELD_NOT_IN_TABLE) {
+	    btm_rbt_delete(&esdp->timer_service->btm_tree, tmr);
+	    tmr->btm.tree.parent = ERTS_HLT_PFIELD_NOT_IN_TABLE;
+	}
+#endif
+        if (is_hlt) {
+            if (cncl_res > 0)
+                hl_timer_dec_refc(&tmr->type.hlt, tmr->type.hlt.head.roflgs);
+            hlt_delete_timer(esdp, &tmr->type.hlt);
+        }
+        else {
+            if (cncl_res > 0)
+                tw_timer_dec_refc(&tmr->type.twt);
+            cancel_tw_timer(esdp, &tmr->type.twt);
+        }
+    }
+
+    return time_left;
+}
+
+static ERTS_INLINE Eterm
+return_info(Process *c_p, Sint64 time_left)
+{
+    Uint hsz;
+    Eterm *hp;
+
+    if (time_left < 0)
+        return am_false;
+
+    if (time_left <= (Sint64) MAX_SMALL)
+        return make_small((Sint) time_left);
+
+    hsz = ERTS_SINT64_HEAP_SIZE(time_left);
+    hp = HAlloc(c_p, hsz);
+    return erts_sint64_to_big(time_left, &hp);
+}
+
+static ERTS_INLINE Eterm
+send_async_info(Process *proc, ErtsProcLocks initial_locks,
+                Eterm tref, int cancel, Sint64 time_left)
+{
+    ErtsProcLocks locks = initial_locks;
+    ErtsMessage *mp;
+    Eterm tag, res, msg, ref;
+    Uint hsz;
+    Eterm *hp;
+    ErlOffHeap *ohp;
+
+    hsz = 4;
+    hsz += NC_HEAP_SIZE(tref);
+
+    if (time_left > (Sint64) MAX_SMALL)
+        hsz += ERTS_SINT64_HEAP_SIZE(time_left);
+
+    mp = erts_alloc_message_heap(proc, &locks, hsz, &hp, &ohp);
+
+    if (cancel)
+        tag = am_cancel_timer;
+    else
+        tag = am_read_timer;
+
+    ref = STORE_NC(&hp, ohp, tref);
+
+    if (time_left < 0)
+        res = am_false;
+    else if (time_left <= (Sint64) MAX_SMALL)
+        res = make_small((Sint) time_left);
+    else
+        res = erts_sint64_to_big(time_left, &hp);
+
+    msg = TUPLE3(hp, tag, ref, res);
+
+    erts_queue_message(proc, locks, mp, msg, am_clock_service);
+
+    locks &= ~initial_locks;
+    if (locks)
+        erts_smp_proc_unlock(proc, locks);
+
+    return am_ok;
+}
+
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+
+static BIF_RETTYPE
+access_bif_timer(Process *c_p, Eterm tref, int cancel, int async, int info)
+{
+    BIF_RETTYPE ret;
+    Eterm res;
+    Sint64 time_left;
+
+    if (!is_internal_magic_ref(tref)) {
+	if (is_not_ref(tref)) {
+            ERTS_BIF_PREP_ERROR(ret, c_p, BADARG);
+            return ret;
+        }
+        time_left = -1;
+    }
+    else {
+        ErtsMagicBinary *mbin;
+        mbin = (ErtsMagicBinary *) erts_magic_ref2bin(tref);
+        if (mbin->destructor != bif_timer_ref_destructor)
+            time_left = -1;
+        else {
+            ErtsBifTimer *tmr;
+            Uint32 sid;
+            tmr = magic_binary_to_btm(mbin);
+            sid = erts_get_ref_numbers_thr_id(internal_magic_ref_numbers(tref));
+            ASSERT(1 <= sid && sid <= erts_no_schedulers);
+            time_left = access_btm(tmr, sid, erts_proc_sched_data(c_p), cancel);
+        }
+    }
+
+    if (!info)
+        res = am_ok;
+    else if (!async)
+        res = return_info(c_p, time_left);
+    else
+        res = send_async_info(c_p, ERTS_PROC_LOCK_MAIN,
+                              tref, cancel, time_left);
+
+    ERTS_BIF_PREP_RET(ret, res);
+
+    return ret;
+}
+
+#else /* !ERTS_MAGIC_REF_BIF_TIMERS */
+
+static ERTS_INLINE Eterm
+send_sync_info(Process *proc, ErtsProcLocks initial_locks,
+               Uint32 *refn, int cancel, Sint64 time_left)
+{
+    ErtsProcLocks locks = initial_locks;
+    ErtsMessage *mp;
+    Eterm res, msg, ref;
+    Uint hsz;
+    Eterm *hp;
+    ErlOffHeap *ohp;
+
+    hsz = 3 + ERTS_REF_THING_SIZE;
+
+    if (time_left > (Sint64) MAX_SMALL)
+        hsz += ERTS_SINT64_HEAP_SIZE(time_left);
+
+    mp = erts_alloc_message_heap(proc, &locks, hsz, &hp, &ohp);
+
+    write_ref_thing(hp, refn[0], refn[1], refn[2]);
+    ref = make_internal_ref(hp);
+    hp += ERTS_REF_THING_SIZE;
+
+    if (time_left < 0)
+        res = am_false;
+    else if (time_left <= (Sint64) MAX_SMALL)
+        res = make_small((Sint) time_left);
+    else
+        res = erts_sint64_to_big(time_left, &hp);
+
+    msg = TUPLE2(hp, ref, res);
+
+    erts_queue_message(proc, locks, mp, msg, am_clock_service);
+
+    locks &= ~initial_locks;
+    if (locks)
+        erts_smp_proc_unlock(proc, locks);
+
+    return am_ok;
+}
+
 static ERTS_INLINE Eterm
 access_sched_local_btm(Process *c_p, Eterm pid,
-		       Eterm tref, Uint32 *trefn,
-		       Uint32 *rrefn,
-		       int async, int cancel,
-		       int return_res,
-		       int info)
+                       Eterm tref, Uint32 *trefn,
+                       Uint32 *rrefn,
+                       int async, int cancel,
+                       int return_res,
+                       int info)
 {
     ErtsSchedulerData *esdp;
     ErtsHLTimerService *srv;
-    ErtsHLTimer *tmr;
+    ErtsBifTimer *tmr;
     Sint64 time_left;
     Process *proc;
     ErtsProcLocks proc_locks;
@@ -1884,111 +2296,40 @@ access_sched_local_btm(Process *c_p, Eterm pid,
     srv = esdp->timer_service;
 
     tmr = btm_rbt_lookup(srv->btm_tree, trefn);
-    if (tmr) {
-	if (!cancel) {
-	    erts_aint32_t state = erts_smp_atomic32_read_acqb(&tmr->state);
-	    if (state == ERTS_TMR_STATE_ACTIVE)
-		time_left = get_time_left(esdp, tmr->timeout);
-	}
-	else {
-	    int cncl_res = cancel_bif_timer(tmr);
-	    if (cncl_res) {
-
-		time_left = get_time_left(esdp, tmr->timeout);
 
-		if (cncl_res > 0)
-		    hl_timer_dec_refc(tmr, tmr->head.roflgs);
-
-		hlt_delete_timer(esdp, tmr);
-	    }
-	}
-    }
+    time_left = access_btm(tmr, (Uint32) esdp->no, esdp, cancel);
 
     if (!info)
-	return am_ok;
-
-    if (return_res) {
-	ERTS_HLT_ASSERT(c_p);
-	if (time_left < 0)
-	    return am_false;
-	else if (time_left <= (Sint64) MAX_SMALL)
-	    return make_small((Sint) time_left);
-	else {
-	    Uint hsz = ERTS_SINT64_HEAP_SIZE(time_left);
-	    Eterm *hp = HAlloc(c_p, hsz);
-	    return erts_sint64_to_big(time_left, &hp);
-	}
-    }
+        return am_ok;
 
     if (c_p) {
-	proc = c_p;
-	proc_locks = ERTS_PROC_LOCK_MAIN;
+        proc = c_p;
+        proc_locks = ERTS_PROC_LOCK_MAIN;
     }
     else {
-	proc = erts_proc_lookup(pid);
-	proc_locks = 0;
+        proc = erts_proc_lookup(pid);
+        proc_locks = 0;
     }
 
-    if (proc) {
-	Uint hsz;
-	ErtsMessage *mp;
-	Eterm *hp, msg, ref, result;
-	ErlOffHeap *ohp;
-	Uint32 *refn;
-#ifdef ERTS_HLT_DEBUG
-	Eterm *hp_end;
-#endif
-
-	hsz = ERTS_REF_THING_SIZE;
-	if (async) {
-	    refn = trefn; /* timer ref */
-	    hsz += 4; /* 3-tuple */
-	}
-	else {
-	    refn = rrefn; /* request ref */
-	    hsz += 3; /* 2-tuple */
-	}
-
-	ERTS_HLT_ASSERT(refn);
-
-	if (time_left > (Sint64) MAX_SMALL)
-	    hsz += ERTS_SINT64_HEAP_SIZE(time_left);
-
-	mp = erts_alloc_message_heap(proc, &proc_locks,
-				     hsz, &hp, &ohp);
-
-#ifdef ERTS_HLT_DEBUG
-	hp_end = hp + hsz;
-#endif
-
-	if (time_left < 0)
-	    result = am_false;
-	else if (time_left <= (Sint64) MAX_SMALL)
-	    result = make_small((Sint) time_left);
-	else
-	    result = erts_sint64_to_big(time_left, &hp);
-
-	write_ref_thing(hp,
-			refn[0],
-			refn[1],
-			refn[2]);
-	ref = make_internal_ref(hp);
-	hp += ERTS_REF_THING_SIZE;
-
-	msg = (async
-	       ? TUPLE3(hp, (cancel
-			     ? am_cancel_timer
-			     : am_read_timer), ref, result)
-	       : TUPLE2(hp, ref, result));
-
-	ERTS_HLT_ASSERT(hp + (async ? 4 : 3) == hp_end);
-
-	erts_queue_message(proc, proc_locks, mp, msg, am_clock_service);
-
-	if (c_p)
-	    proc_locks &= ~ERTS_PROC_LOCK_MAIN;
-	if (proc_locks)
-	    erts_smp_proc_unlock(proc, proc_locks);
+    if (!async) {
+        if (c_p)
+            return return_info(c_p, time_left);
+
+        if (proc)
+            return send_sync_info(proc, proc_locks,
+                                  rrefn, cancel, time_left);
+    }
+    else if (proc) {
+        Eterm ref;
+        Eterm heap[ERTS_REF_THING_SIZE];
+        if (is_value(tref))
+            ref = tref;
+        else {
+            write_ref_thing(&heap[0], trefn[0], trefn[1], trefn[2]);
+            ref = make_internal_ref(&heap[0]);
+        }
+        return send_async_info(proc, proc_locks,
+                               ref, cancel, time_left);
     }
 
     return am_ok;
@@ -2021,108 +2362,64 @@ bif_timer_access_request(void *vreq)
 static int
 try_access_sched_remote_btm(ErtsSchedulerData *esdp,
 			    Process *c_p, Uint32 sid,
-			    Uint32 *trefn,
+			    Eterm tref, Uint32 *trefn,
 			    int async, int cancel,
 			    int info, Eterm *resp)
 {
-    ErtsHLTimer *tmr;
+    ErtsBifTimer *tmr;
     Sint64 time_left;
 
     ERTS_HLT_ASSERT(c_p);
 
     /*
      * Check if the timer is aimed at current
-     * process of if this process is an accessor
-     * of the timer...
+     * process...
      */
     erts_smp_proc_lock(c_p, ERTS_PROC_LOCK_BTM);
     tmr = proc_btm_rbt_lookup(c_p->bif_timers, trefn);
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-    if (!tmr)
-	tmr = abtm_rbt_lookup(c_p->accessor_bif_timers, trefn);
-#endif
     erts_smp_proc_unlock(c_p, ERTS_PROC_LOCK_BTM);
     if (!tmr)
 	return 0;
 
-    if (!cancel) {
-	erts_aint32_t state = erts_smp_atomic32_read_acqb(&tmr->state);
-	if (state == ERTS_TMR_STATE_ACTIVE)
-	    time_left = get_time_left(esdp, tmr->timeout);
-	else
-	    time_left = -1;
-    }
-    else {
-	int cncl_res = cancel_bif_timer(tmr);
-	if (!cncl_res)
-	    time_left = -1;
-	else {
-	    time_left = get_time_left(esdp, tmr->timeout);
-	    if (cncl_res > 0)
-		queue_canceled_timer(esdp, sid, (ErtsTimer *) tmr);
-	}
-    }
+    time_left = access_btm(tmr, sid, esdp, cancel);
 
-    if (!info) {
+    if (!info)
 	*resp = am_ok;
-	return 1;
-    }
-
-    if (!async) {
-	if (time_left < 0)
-	    *resp = am_false;
-	else if (time_left <= (Sint64) MAX_SMALL)
-	    *resp = make_small((Sint) time_left);
-	else {
-	    Uint hsz = ERTS_SINT64_HEAP_SIZE(time_left);
-	    Eterm *hp = HAlloc(c_p, hsz);
-	    *resp = erts_sint64_to_big(time_left, &hp);
-	}
-    }
-    else {
-	ErtsMessage *mp;
-	Eterm tag, res, msg, tref;
-	Uint hsz;
-	Eterm *hp;
-	ErtsProcLocks proc_locks = ERTS_PROC_LOCK_MAIN;
-	ErlOffHeap *ohp;
-
-	hsz = 4 + ERTS_REF_THING_SIZE;
-	if (time_left > (Sint64) MAX_SMALL)
-	    hsz += ERTS_SINT64_HEAP_SIZE(time_left);
-
-	mp = erts_alloc_message_heap(c_p, &proc_locks,
-				     hsz, &hp, &ohp);
-	if (cancel)
-	    tag = am_cancel_timer;
-	else
-	    tag = am_read_timer;
-
-	write_ref_thing(hp,
-			trefn[0],
-			trefn[1],
-			trefn[2]);
-	tref = make_internal_ref(hp);
-	hp += ERTS_REF_THING_SIZE;
-
-	if (time_left < 0)
-	    res = am_false;
-	else if (time_left <= (Sint64) MAX_SMALL)
-	    res = make_small((Sint) time_left);
-	else
-	    res = erts_sint64_to_big(time_left, &hp);
-
-	msg = TUPLE3(hp, tag, tref, res);
+    else if (!async)
+        *resp = return_info(c_p, time_left);
+    else
+        *resp = send_async_info(c_p, ERTS_PROC_LOCK_MAIN,
+                                tref, cancel, time_left);
 
-        erts_queue_message(c_p, proc_locks, mp, msg, am_clock_service);
+    return 1;
+}
 
-	proc_locks &= ~ERTS_PROC_LOCK_MAIN;
-	if (proc_locks)
-	    erts_smp_proc_unlock(c_p, proc_locks);
+static Eterm
+no_timer_result(Process *c_p, Eterm tref, int cancel, int async, int info)
+{
+    ErtsMessage *mp;
+    Uint hsz;
+    Eterm *hp, msg, ref, tag;
+    ErlOffHeap *ohp;
+    ErtsProcLocks locks;
 
-	*resp = am_ok;
-    }
-    return 1;
+    if (!async)
+        return am_false;
+    if (!info)
+        return am_ok;
+
+    hsz = 4;
+    hsz += NC_HEAP_SIZE(tref);
+    locks = ERTS_PROC_LOCK_MAIN;
+    mp = erts_alloc_message_heap(c_p, &locks, hsz, &hp, &ohp);
+    ref = STORE_NC(&hp, ohp, tref);
+    tag = cancel ? am_cancel_timer : am_read_timer;
+    msg = TUPLE3(hp, tag, ref, am_false);
+    erts_queue_message(c_p, locks, mp, msg, am_clock_service);
+    locks &= ~ERTS_PROC_LOCK_MAIN;
+    if (locks)
+        erts_smp_proc_unlock(c_p, locks);
+    return am_ok;
 }
 
 static BIF_RETTYPE
@@ -2156,7 +2453,7 @@ access_bif_timer(Process *c_p, Eterm tref, int cancel, int async, int info)
 	ERTS_BIF_PREP_RET(ret, res);
     }
     else if (try_access_sched_remote_btm(esdp, c_p,
-					 sid, trefn,
+					 sid, tref, trefn,
 					 async, cancel,
 					 info, &res)) {
 	ERTS_BIF_PREP_RET(ret, res);
@@ -2235,11 +2532,11 @@ badarg:
     return ret;
 
 no_timer:
-    ERTS_BIF_PREP_RET(ret, am_false);
-    return ret;
-    
+    return no_timer_result(c_p, tref, cancel, async, info);
 }
 
+#endif /* !ERTS_MAGIC_REF_BIF_TIMERS */
+
 static ERTS_INLINE int
 bool_arg(Eterm val, int *argp)
 {
@@ -2251,8 +2548,8 @@ bool_arg(Eterm val, int *argp)
 }
 
 static ERTS_INLINE int
-parse_bif_timer_options(Eterm option_list, int *async, int *info,
-			int *abs, Eterm *accessor)
+parse_bif_timer_options(Eterm option_list, int *async,
+                        int *info, int *abs)
 {
     Eterm list = option_list;
 
@@ -2262,8 +2559,6 @@ parse_bif_timer_options(Eterm option_list, int *async, int *info,
 	*info = 1;
     if (abs)
 	*abs = 0;
-    if (accessor)
-	*accessor = THE_NON_VALUE;
 
     while (is_list(list)) {
 	Eterm *consp, *tp, opt;
@@ -2290,13 +2585,6 @@ parse_bif_timer_options(Eterm option_list, int *async, int *info,
 	    if (!abs || !bool_arg(tp[2], abs))
 		return 0;
 	    break;
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-	case am_accessor:
-	    if (!accessor || is_not_internal_pid(tp[2]))
-		return 0;
-	    *accessor = tp[2];
-	    break;
-#endif
 	default:
 	    return 0;
 	}
@@ -2310,42 +2598,57 @@ parse_bif_timer_options(Eterm option_list, int *async, int *info,
 }
 
 static void
-exit_cancel_bif_timer(ErtsHLTimer *tmr, void *vesdp)
+exit_cancel_bif_timer(ErtsBifTimer *tmr, void *vesdp)
 {
     ErtsSchedulerData *esdp = (ErtsSchedulerData *) vesdp;
     Uint32 sid, roflgs;
     erts_aint_t state;
+    int is_hlt;
 
-    state = erts_smp_atomic32_cmpxchg_acqb(&tmr->state,
+    state = erts_smp_atomic32_cmpxchg_acqb(&tmr->btm.state,
 					   ERTS_TMR_STATE_CANCELED,
 					   ERTS_TMR_STATE_ACTIVE);
 
-    roflgs = tmr->head.roflgs;
+    roflgs = tmr->type.head.roflgs;
     sid = roflgs & ERTS_TMR_ROFLG_SID_MASK;
+    is_hlt = !!(roflgs & ERTS_TMR_ROFLG_HLT);
 
-    ERTS_HLT_ASSERT(sid == erts_get_ref_numbers_thr_id(tmr->btm.refn));
+    ERTS_HLT_ASSERT(sid == erts_get_ref_numbers_thr_id(ERTS_BTM_HLT2REFN(tmr)));
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+    ERTS_HLT_ASSERT(tmr->btm.proc_list.next);
+#else
     ERTS_HLT_ASSERT(tmr->btm.proc_tree.parent
 		    != ERTS_HLT_PFIELD_NOT_IN_TABLE);
-
     tmr->btm.proc_tree.parent = ERTS_HLT_PFIELD_NOT_IN_TABLE;
+#endif
 
-    if (sid == (Uint32) esdp->no) {
-	if (state == ERTS_TMR_STATE_ACTIVE) {
-	    if (tmr->btm.bp)
-		free_message_buffer(tmr->btm.bp);
-	    hlt_delete_timer(esdp, tmr);
-	}
-	hl_timer_dec_refc(tmr, roflgs);
-    }
-    else {
-	if (state == ERTS_TMR_STATE_ACTIVE) {
-	    if (tmr->btm.bp)
-		free_message_buffer(tmr->btm.bp);
-	    queue_canceled_timer(esdp, sid, (ErtsTimer *) tmr);
+    if (state == ERTS_TMR_STATE_ACTIVE) {
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+        btm_clear_magic_binary(tmr);
+#endif
+        if (tmr->btm.bp)
+            free_message_buffer(tmr->btm.bp);
+
+        if (sid != (Uint32) esdp->no) {
+            queue_canceled_timer(esdp, sid, (ErtsTimer *) tmr);
+            return;
+        }
+
+#ifndef ERTS_MAGIC_REF_BIF_TIMERS
+        if (tmr->btm.tree.parent != ERTS_HLT_PFIELD_NOT_IN_TABLE) {
+	    btm_rbt_delete(&esdp->timer_service->btm_tree, tmr);
+	    tmr->btm.tree.parent = ERTS_HLT_PFIELD_NOT_IN_TABLE;
 	}
-	else
-	    hl_timer_dec_refc(tmr, roflgs);
+#endif
+        if (is_hlt)
+            hlt_delete_timer(esdp, &tmr->type.hlt);
+        else
+            cancel_tw_timer(esdp, &tmr->type.twt);
     }
+    if (is_hlt)
+        hl_timer_dec_refc(&tmr->type.hlt, roflgs);
+    else
+        tw_timer_dec_refc(&tmr->type.twt);
 }
 
 #ifdef ERTS_HLT_DEBUG
@@ -2354,20 +2657,29 @@ exit_cancel_bif_timer(ErtsHLTimer *tmr, void *vesdp)
 #  define ERTS_BTM_MAX_DESTROY_LIMIT 50
 #endif
 
+#ifndef ERTS_MAGIC_REF_BIF_TIMERS
 typedef struct {
     ErtsBifTimers *bif_timers;
     union {
 	proc_btm_rbt_yield_state_t proc_btm_yield_state;
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-	abtm_rbt_yield_state_t abtm_yield_state;
-#endif
     } u;
 } ErtsBifTimerYieldState;
+#endif
 
-int erts_cancel_bif_timers(Process *p, ErtsBifTimers *btm, void **vyspp)
+int erts_cancel_bif_timers(Process *p, ErtsBifTimers **btm, void **vyspp)
 {
     ErtsSchedulerData *esdp = erts_proc_sched_data(p);
-    ErtsBifTimerYieldState ys = {btm, {ERTS_RBT_YIELD_STAT_INITER}};
+
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+
+    return proc_btm_list_foreach_destroy_yielding(btm,
+                                                  exit_cancel_bif_timer,
+                                                  (void *) esdp,
+                                                  ERTS_BTM_MAX_DESTROY_LIMIT);
+
+#else /* !ERTS_MAGIC_REF_BIF_TIMERS */
+
+    ErtsBifTimerYieldState ys = {*btm, {ERTS_RBT_YIELD_STAT_INITER}};
     ErtsBifTimerYieldState *ysp;
     int res;
 
@@ -2399,63 +2711,18 @@ int erts_cancel_bif_timers(Process *p, ErtsBifTimers *btm, void **vyspp)
     }
 
     return res;
-}
-
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-
-static void
-detach_bif_timer(ErtsHLTimer *tmr, void *vesdp)
-{
-    tmr->abtm.tree.parent = ERTS_HLT_PFIELD_NOT_IN_TABLE;
-    hl_timer_dec_refc(tmr, tmr->head.roflgs);
-}
-
-int erts_detach_accessor_bif_timers(Process *p, ErtsBifTimers *btm, void **vyspp)
-{
-    ErtsSchedulerData *esdp = erts_proc_sched_data(p);
-    ErtsBifTimerYieldState ys = {btm, {ERTS_RBT_YIELD_STAT_INITER}};
-    ErtsBifTimerYieldState *ysp;
-    int res;
-
-    ysp = (ErtsBifTimerYieldState *) *vyspp;
-    if (!ysp)
-	ysp = &ys;
-
-    res = abtm_rbt_foreach_destroy_yielding(&ysp->bif_timers,
-					    detach_bif_timer,
-					    (void *) esdp,
-					    &ysp->u.abtm_yield_state,
-					    ERTS_BTM_MAX_DESTROY_LIMIT);
-
-    if (res == 0) {
-	if (ysp != &ys)
-	    erts_free(ERTS_ALC_T_BTM_YIELD_STATE, ysp);
-	*vyspp = NULL;
-    }
-    else {
-
-	if (ysp == &ys) {
-	    ysp = erts_alloc(ERTS_ALC_T_BTM_YIELD_STATE,
-			     sizeof(ErtsBifTimerYieldState));
-	    sys_memcpy((void *) ysp, (void *) &ys,
-		       sizeof(ErtsBifTimerYieldState));
-	}
-
-	*vyspp = (void *) ysp;
-    }
 
-    return res;
+#endif /* !ERTS_MAGIC_REF_BIF_TIMERS */
 }
 
-#endif /* ERTS_BTM_ACCESSOR_SUPPORT */
-
 static ERTS_INLINE int
 parse_timeout_pos(ErtsSchedulerData *esdp, Eterm arg,
 		  ErtsMonotonicTime *conv_arg, int abs,
-		  ErtsMonotonicTime *tposp, int *stimep)
+		  ErtsMonotonicTime *tposp, int *stimep,
+                  ErtsMonotonicTime *msp)
 {
-    ErtsMonotonicTime t;
-
+    ErtsMonotonicTime t, now;
+    
     if (!term_to_Sint64(arg, &t)) {
 	ERTS_HLT_ASSERT(!is_small(arg));
 	if (!is_big(arg))
@@ -2470,22 +2737,30 @@ parse_timeout_pos(ErtsSchedulerData *esdp, Eterm arg,
     if (conv_arg)
 	*conv_arg = t;
 
+    now = erts_get_monotonic_time(esdp);
+
     if (abs) {
 	t += -1*ERTS_MONOTONIC_OFFSET_MSEC; /* external to internal */
 	if (t < ERTS_MONOTONIC_TO_MSEC(ERTS_MONOTONIC_BEGIN))
 	    return 1;
 	if (t > ERTS_MONOTONIC_TO_MSEC(ERTS_MONOTONIC_END))
 	    return 1;
+        if (msp)
+            *msp = t - ERTS_MONOTONIC_TO_MSEC(now);
+
 	*stimep = (t - ERTS_MONOTONIC_TO_MSEC(esdp->last_monotonic_time)
 		   < ERTS_BIF_TIMER_SHORT_TIME);
 	*tposp = ERTS_MSEC_TO_CLKTCKS(t);
     }
     else {
-	ErtsMonotonicTime now, ticks;
+	ErtsMonotonicTime ticks;
 	
 	if (t < 0)
 	    return -1;
 
+        if (msp)
+            *msp = t;
+
 	ticks = ERTS_MSEC_TO_CLKTCKS(t);
 
 	if (ERTS_CLKTCK_RESOLUTION > 1000 && ticks < 0)
@@ -2493,7 +2768,6 @@ parse_timeout_pos(ErtsSchedulerData *esdp, Eterm arg,
 
 	ERTS_HLT_ASSERT(ticks >= 0);
 
-	now = erts_get_monotonic_time(esdp);
 	ticks += ERTS_MONOTONIC_TO_CLKTCKS(now-1);
 	ticks += 1;
 
@@ -2516,66 +2790,68 @@ parse_timeout_pos(ErtsSchedulerData *esdp, Eterm arg,
 
 BIF_RETTYPE send_after_3(BIF_ALIST_3)
 {
-    ErtsMonotonicTime timeout_pos;
+    ErtsMonotonicTime timeout_pos, tmo;
     int short_time, tres;
 
-    tres = parse_timeout_pos(erts_proc_sched_data(BIF_P), BIF_ARG_1, NULL,
-			     0, &timeout_pos, &short_time);
+    tres = parse_timeout_pos(erts_proc_sched_data(BIF_P), BIF_ARG_1,
+			     NULL, 0, &timeout_pos, &short_time, &tmo);
     if (tres != 0)
 	BIF_ERROR(BIF_P, BADARG);
 
-    return setup_bif_timer(BIF_P, timeout_pos, short_time,
-			   BIF_ARG_2, BIF_ARG_2, BIF_ARG_3, 0);
+    return setup_bif_timer(BIF_P, tmo < ERTS_TIMER_WHEEL_MSEC,
+                           timeout_pos, short_time, BIF_ARG_2,
+                           BIF_ARG_3, 0);
 }
 
 BIF_RETTYPE send_after_4(BIF_ALIST_4)
 {
-    ErtsMonotonicTime timeout_pos;
-    Eterm accessor;
+    ErtsMonotonicTime timeout_pos, tmo;
     int short_time, abs, tres;
 
-    if (!parse_bif_timer_options(BIF_ARG_4, NULL, NULL, &abs, &accessor))
+    if (!parse_bif_timer_options(BIF_ARG_4, NULL, NULL, &abs))
 	BIF_ERROR(BIF_P, BADARG);
     
     tres = parse_timeout_pos(erts_proc_sched_data(BIF_P), BIF_ARG_1, NULL,
-			     abs, &timeout_pos, &short_time);
+			     abs, &timeout_pos, &short_time, &tmo);
     if (tres != 0)
 	BIF_ERROR(BIF_P, BADARG);
 
-    return setup_bif_timer(BIF_P, timeout_pos, short_time,
-			   BIF_ARG_2, accessor, BIF_ARG_3, 0);
+    return setup_bif_timer(BIF_P, tmo < ERTS_TIMER_WHEEL_MSEC,
+                           timeout_pos, short_time, BIF_ARG_2,
+                           BIF_ARG_3, 0);
 }
 
 BIF_RETTYPE start_timer_3(BIF_ALIST_3)
 {
-    ErtsMonotonicTime timeout_pos;
+    ErtsMonotonicTime timeout_pos, tmo;
     int short_time, tres;
 
     tres = parse_timeout_pos(erts_proc_sched_data(BIF_P), BIF_ARG_1, NULL,
-			     0, &timeout_pos, &short_time);
+			     0, &timeout_pos, &short_time, &tmo);
     if (tres != 0)
 	BIF_ERROR(BIF_P, BADARG);
 
-    return setup_bif_timer(BIF_P, timeout_pos, short_time,
-			   BIF_ARG_2, BIF_ARG_2, BIF_ARG_3, !0);
+    return setup_bif_timer(BIF_P, tmo < ERTS_TIMER_WHEEL_MSEC,
+                           timeout_pos, short_time, BIF_ARG_2,
+                           BIF_ARG_3, !0);
 }
 
 BIF_RETTYPE start_timer_4(BIF_ALIST_4)
 {
-    ErtsMonotonicTime timeout_pos;
-    Eterm accessor;
+    ErtsMonotonicTime timeout_pos, tmo;
     int short_time, abs, tres;
 
-    if (!parse_bif_timer_options(BIF_ARG_4, NULL, NULL, &abs, &accessor))
+    if (!parse_bif_timer_options(BIF_ARG_4, NULL, NULL, &abs))
 	BIF_ERROR(BIF_P, BADARG);
 
     tres = parse_timeout_pos(erts_proc_sched_data(BIF_P), BIF_ARG_1, NULL,
-			     abs, &timeout_pos, &short_time);
+			     abs, &timeout_pos, &short_time, &tmo);
     if (tres != 0)
 	BIF_ERROR(BIF_P, BADARG);
 
-    return setup_bif_timer(BIF_P, timeout_pos, short_time,
-			   BIF_ARG_2, accessor, BIF_ARG_3, !0);
+    return setup_bif_timer(BIF_P, tmo < ERTS_TIMER_WHEEL_MSEC,
+                           timeout_pos, short_time, BIF_ARG_2,
+                           BIF_ARG_3, !0);
 }
 
 BIF_RETTYPE cancel_timer_1(BIF_ALIST_1)
@@ -2588,7 +2864,7 @@ BIF_RETTYPE cancel_timer_2(BIF_ALIST_2)
     BIF_RETTYPE ret;
     int async, info;
 
-    if (parse_bif_timer_options(BIF_ARG_2, &async, &info, NULL, NULL))
+    if (parse_bif_timer_options(BIF_ARG_2, &async, &info, NULL))
 	return access_bif_timer(BIF_P, BIF_ARG_1, 1, async, info);
 
     ERTS_BIF_PREP_ERROR(ret, BIF_P, BADARG);
@@ -2605,7 +2881,7 @@ BIF_RETTYPE read_timer_2(BIF_ALIST_2)
     BIF_RETTYPE ret;
     int async;
 
-    if (parse_bif_timer_options(BIF_ARG_2, &async, NULL, NULL, NULL))
+    if (parse_bif_timer_options(BIF_ARG_2, &async, NULL, NULL))
 	return access_bif_timer(BIF_P, BIF_ARG_1, 0, async, 1);
 
     ERTS_BIF_PREP_ERROR(ret, BIF_P, BADARG);
@@ -2620,14 +2896,13 @@ start_callback_timer(ErtsSchedulerData *esdp,
 		     void *arg)
 		     
 {
-    if (twt)
-	create_tw_timer(esdp, ERTS_TMR_CALLBACK, NULL,
-			callback, arg, timeout_pos);
-    else
-	create_hl_timer(esdp, timeout_pos, 0,
-			ERTS_TMR_CALLBACK, NULL,
-			NIL, THE_NON_VALUE, NIL,
-			NULL, callback, arg);
+    ErtsCreateTimerFunc create_timer = (twt
+                                        ? create_tw_timer
+                                        : create_hl_timer);
+    (void) create_timer(esdp, timeout_pos, 0,
+                        ERTS_TMR_CALLBACK, NULL,
+                        NIL, THE_NON_VALUE, NULL,
+                        callback, arg);
 }
 
 typedef struct {
@@ -2704,18 +2979,18 @@ set_proc_timer_common(Process *c_p, ErtsSchedulerData *esdp, Sint64 tmo,
     if (tmo == 0)
 	c_p->flags |= F_TIMO;
     else {
+        ErtsCreateTimerFunc create_timer;
 
 	c_p->flags |= F_INSLPQUEUE;
 	c_p->flags &= ~F_TIMO;
 
-	if (tmo < ERTS_TIMER_WHEEL_MSEC)
-	    tmr = (void *) create_tw_timer(esdp, ERTS_TMR_PROC, (void *) c_p,
-					   NULL, NULL, timeout_pos);
-	else
-	    tmr = (void *) create_hl_timer(esdp, timeout_pos, short_time,
-					   ERTS_TMR_PROC, (void *) c_p,
-					   c_p->common.id, THE_NON_VALUE,
-					   NIL, NULL, NULL, NULL);
+        create_timer = (tmo < ERTS_TIMER_WHEEL_MSEC
+                        ? create_tw_timer
+                        : create_hl_timer);
+        tmr = (void *) create_timer(esdp, timeout_pos, short_time,
+                                    ERTS_TMR_PROC, (void *) c_p,
+                                    c_p->common.id, THE_NON_VALUE,
+                                    NULL, NULL, NULL);
 	erts_smp_atomic_set_relb(&c_p->common.timer, (erts_aint_t) tmr);
     }
 }
@@ -2731,7 +3006,7 @@ erts_set_proc_timer_term(Process *c_p, Eterm etmo)
 		    == ERTS_PTMR_NONE);
 
     tres = parse_timeout_pos(esdp, etmo, &tmo, 0,
-			     &timeout_pos, &short_time);
+			     &timeout_pos, &short_time, NULL);
     if (tres != 0)
 	return tres;
 
@@ -2789,6 +3064,7 @@ erts_set_port_timer(Port *c_prt, Sint64 tmo)
     void *tmr;
     ErtsSchedulerData *esdp = erts_get_scheduler_data();
     ErtsMonotonicTime timeout_pos;
+    ErtsCreateTimerFunc create_timer;
 
     if (erts_smp_atomic_read_nob(&c_prt->common.timer) != ERTS_PTMR_NONE)
 	erts_cancel_port_timer(c_prt);
@@ -2797,13 +3073,12 @@ erts_set_port_timer(Port *c_prt, Sint64 tmo)
 
     timeout_pos = get_timeout_pos(erts_get_monotonic_time(esdp), tmo);
 
-    if (tmo < ERTS_TIMER_WHEEL_MSEC)
-	tmr = (void *) create_tw_timer(esdp, ERTS_TMR_PORT, (void *) c_prt,
-				       NULL, NULL, timeout_pos);
-    else
-	tmr = (void *) create_hl_timer(esdp, timeout_pos, 0, ERTS_TMR_PORT,
-				       (void *) c_prt, c_prt->common.id,
-				       THE_NON_VALUE, NIL, NULL, NULL, NULL);
+    create_timer = (tmo < ERTS_TIMER_WHEEL_MSEC
+                    ? create_tw_timer
+                    : create_hl_timer);
+    tmr = (void *) create_timer(esdp, timeout_pos, 0, ERTS_TMR_PORT,
+                                (void *) c_prt, c_prt->common.id,
+                                THE_NON_VALUE, NULL, NULL, NULL);
     erts_smp_atomic_set_relb(&c_prt->common.timer, (erts_aint_t) tmr);
 }
 
@@ -2842,7 +3117,7 @@ erts_read_port_timer(Port *c_prt)
     if (tmr->head.roflgs & ERTS_TMR_ROFLG_HLT)
 	timeout_pos = tmr->hlt.timeout;
     else
-	timeout_pos = tmr->twt.tw_tmr.timeout_pos;
+	timeout_pos = erts_tweel_read_timeout(&tmr->twt.u.tw_tmr);
     return get_time_left(NULL, timeout_pos);
 }
 
@@ -2857,20 +3132,35 @@ typedef struct {
 } ErtsBTMPrint;
 
 static void
-btm_print(ErtsHLTimer *tmr, void *vbtmp)
+btm_print(ErtsBifTimer *tmr, void *vbtmp, ErtsMonotonicTime tpos, int is_hlt)
 {
     ErtsBTMPrint *btmp = (ErtsBTMPrint *) vbtmp;
     ErtsMonotonicTime left;
     Eterm receiver;
 
-    if (tmr->timeout <= btmp->now)
-	left = 0;
-    else
-        left = ERTS_CLKTCKS_TO_MSEC(tmr->timeout - btmp->now);
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+    if (!(tmr->type.head.roflgs & ERTS_TMR_ROFLG_BIF_TMR))
+        return;
+#endif
 
-    receiver = ((tmr->head.roflgs & ERTS_TMR_ROFLG_REG_NAME)
-		? tmr->receiver.name
-		: tmr->receiver.proc->common.id);
+    if (is_hlt) {
+        ERTS_HLT_ASSERT(tmr->type.head.roflgs & ERTS_TMR_ROFLG_HLT);
+        if (tmr->type.hlt.timeout <= btmp->now)
+            left = 0;
+        else
+            left = ERTS_CLKTCKS_TO_MSEC(tmr->type.hlt.timeout - btmp->now);
+    }
+    else {
+        ERTS_HLT_ASSERT(!(tmr->type.head.roflgs & ERTS_TMR_ROFLG_HLT));
+        if (tpos <= btmp->now)
+            left = 0;
+        else
+            left = ERTS_CLKTCKS_TO_MSEC(tpos - btmp->now);
+    }
+
+    receiver = ((tmr->type.head.roflgs & ERTS_TMR_ROFLG_REG_NAME)
+                ? tmr->type.head.receiver.name
+                : tmr->type.head.receiver.proc->common.id);
 
     erts_print(btmp->to, btmp->to_arg,
 	       "=timer:%T\n"
@@ -2881,6 +3171,36 @@ btm_print(ErtsHLTimer *tmr, void *vbtmp)
 	       (Sint64) left);
 }
 
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+
+static void
+hlt_btm_print(ErtsHLTimer *tmr, void *vbtmp)
+{
+    btm_print((ErtsBifTimer *) tmr, vbtmp, 0, 1);
+}
+
+static void
+twt_btm_print(void *vbtmp, ErtsMonotonicTime tpos, void *vtwtp)
+{
+    btm_print((ErtsBifTimer *) vtwtp, vbtmp, tpos, 0);
+}
+
+#else
+
+static void
+btm_tree_print(ErtsBifTimer *tmr, void *vbtmp)
+{
+    int is_hlt = !!(tmr->type.head.roflgs & ERTS_TMR_ROFLG_HLT);
+    ErtsMonotonicTime tpos;
+    if (is_hlt)
+        tpos = 0;
+    else
+        tpos = erts_tweel_read_timeout(&tmr->type.twt.u.tw_tmr);
+    btm_print(tmr, vbtmp, tpos, is_hlt);
+}
+
+#endif
+
 void
 erts_print_bif_timer_info(fmtfn_t to, void *to_arg)
 {
@@ -2898,7 +3218,15 @@ erts_print_bif_timer_info(fmtfn_t to, void *to_arg)
     for (six = 0; six < erts_no_schedulers; six++) {
 	ErtsHLTimerService *srv =
 	    erts_aligned_scheduler_data[six].esd.timer_service;
-	btm_rbt_foreach(srv->btm_tree, btm_print, (void *) &btmp);
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+	ErtsTimerWheel *twheel =
+	    erts_aligned_scheduler_data[six].esd.timer_wheel;
+	erts_twheel_debug_foreach(twheel, tw_bif_timer_timeout,
+				  twt_btm_print, (void *) &btmp);
+	time_rbt_foreach(srv->time_tree, hlt_btm_print, (void *) &btmp);
+#else
+	btm_rbt_foreach(srv->btm_tree, btm_tree_print, (void *) &btmp);
+#endif
     }
 }
 
@@ -2911,19 +3239,37 @@ typedef struct {
 } ErtsBTMForeachDebug;
 
 static void
-debug_btm_foreach(ErtsHLTimer *tmr, void *vbtmfd)
+debug_btm_foreach(ErtsBifTimer *tmr, void *vbtmfd)
 {
-    if (erts_smp_atomic32_read_nob(&tmr->state) == ERTS_TMR_STATE_ACTIVE) {
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+    if (!(tmr->type.head.roflgs & ERTS_TMR_ROFLG_BIF_TMR))
+        return;
+#endif
+    if (erts_smp_atomic32_read_nob(&tmr->btm.state) == ERTS_TMR_STATE_ACTIVE) {
 	ErtsBTMForeachDebug *btmfd = (ErtsBTMForeachDebug *) vbtmfd;
-	(*btmfd->func)(((tmr->head.roflgs & ERTS_TMR_ROFLG_REG_NAME)
-			? tmr->receiver.name
-			: tmr->receiver.proc->common.id),
-		       tmr->btm.message,
-		       tmr->btm.bp,
-		       btmfd->arg);
+        Eterm id = ((tmr->type.head.roflgs & ERTS_TMR_ROFLG_REG_NAME)
+                    ? tmr->type.head.receiver.name
+                    : tmr->type.head.receiver.proc->common.id);
+	(*btmfd->func)(id, tmr->btm.message, tmr->btm.bp, btmfd->arg);
     }
 }
 
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+
+static void
+hlt_debug_btm_foreach(ErtsHLTimer *tmr, void *vbtmfd)
+{
+    debug_btm_foreach((ErtsBifTimer *) tmr, vbtmfd);
+}
+
+static void
+twt_debug_btm_foreach(void *vbtmfd, ErtsMonotonicTime tpos, void *vtwtp)
+{
+    debug_btm_foreach((ErtsBifTimer *) vtwtp, vbtmfd);
+}
+
+#endif
+
 void
 erts_debug_bif_timer_foreach(void (*func)(Eterm,
 					  Eterm,
@@ -2943,9 +3289,20 @@ erts_debug_bif_timer_foreach(void (*func)(Eterm,
     for (six = 0; six < erts_no_schedulers; six++) {
 	ErtsHLTimerService *srv =
 	    erts_aligned_scheduler_data[six].esd.timer_service;
+#ifdef ERTS_MAGIC_REF_BIF_TIMERS
+	ErtsTimerWheel *twheel =
+	    erts_aligned_scheduler_data[six].esd.timer_wheel;            
+	erts_twheel_debug_foreach(twheel, tw_bif_timer_timeout,
+				  twt_debug_btm_foreach,
+                                  (void *) &btmfd);
+        time_rbt_foreach(srv->time_tree,
+                         hlt_debug_btm_foreach,
+                         (void *) &btmfd);
+#else
 	btm_rbt_foreach(srv->btm_tree,
 			debug_btm_foreach,
 			(void *) &btmfd);
+#endif
     }
 }
 
@@ -2964,7 +3321,7 @@ debug_callback_timer_foreach_list(ErtsHLTimer *tmr, void *vdfct)
 	= (ErtsDebugForeachCallbackTimer *) vdfct;
 
     if ((tmr->head.roflgs & ERTS_TMR_ROFLG_CALLBACK)
-	&& (tmr->receiver.callback == dfct->tclbk))
+	&& (tmr->head.receiver.callback == dfct->tclbk))
 	(*dfct->func)(dfct->arg,
 		      tmr->timeout,
 		      tmr->head.u.arg);
@@ -2982,7 +3339,7 @@ debug_callback_timer_foreach(ErtsHLTimer *tmr, void *vdfct)
 			       vdfct);
 
     if ((tmr->head.roflgs & ERTS_TMR_ROFLG_CALLBACK)
-	&& (tmr->receiver.callback == dfct->tclbk))
+	&& (tmr->head.receiver.callback == dfct->tclbk))
 	(*dfct->func)(dfct->arg,
 		      tmr->timeout,
 		      tmr->head.u.arg);
@@ -2997,7 +3354,7 @@ debug_tw_callback_timer(void *vdfct,
     ErtsDebugForeachCallbackTimer *dfct
 	= (ErtsDebugForeachCallbackTimer *) vdfct;
 
-    if (twtp->u.callback == dfct->tclbk)
+    if (twtp->head.receiver.callback == dfct->tclbk)
 	(*dfct->func)(dfct->arg,
 		      timeout_pos,
 		      twtp->head.u.arg);
@@ -3068,7 +3425,9 @@ st_hdbg_func(ErtsHLTimer *tmr, void *vhdbg)
     }
     ERTS_HLT_ASSERT(tmr->time.tree.u.l.next->time.tree.u.l.prev == tmr);
     ERTS_HLT_ASSERT(tmr->time.tree.u.l.prev->time.tree.u.l.next == tmr);
-    ERTS_HLT_ASSERT(btm_rbt_lookup(hdbg->srv->btm_tree, tmr->btm.refn) == tmr);
+#ifndef ERTS_MAGIC_REF_BIF_TIMERS
+    ERTS_HLT_ASSERT(btm_rbt_lookup(hdbg->srv->btm_tree, ERTS_BTM_HLT2REFN(tmr)) == tmr);
+#endif
 }
 
 static void
@@ -3097,8 +3456,10 @@ tt_hdbg_func(ErtsHLTimer *tmr, void *vhdbg)
 				& ~ERTS_HLT_PFLGS_MASK);
 	ERTS_HLT_ASSERT(tmr == prnt);
     }
+#ifndef ERTS_MAGIC_REF_BIF_TIMERS
     if (tmr->head.roflgs & ERTS_TMR_ROFLG_BIF_TMR)
-	ERTS_HLT_ASSERT(btm_rbt_lookup(hdbg->srv->btm_tree, tmr->btm.refn) == tmr);
+	ERTS_HLT_ASSERT(btm_rbt_lookup(hdbg->srv->btm_tree, ERTS_BTM_HLT2REFN(tmr)) == tmr);
+#endif
     if (tmr->time.tree.same_time) {
 	ErtsHdbgHLT st_hdbg;
 	st_hdbg.srv = hdbg->srv;
@@ -3164,6 +3525,7 @@ hdbg_chk_srv(ErtsHLTimerService *srv)
 	time_rbt_foreach(srv->time_tree, tt_hdbg_func, (void *) &hdbg);
 	ERTS_HLT_ASSERT(hdbg.found_root);
     }
+#ifndef ERTS_MAGIC_REF_BIF_TIMERS
     if (srv->btm_tree) {
 	ErtsHdbgHLT hdbg;
 	hdbg.srv = srv;
@@ -3172,6 +3534,7 @@ hdbg_chk_srv(ErtsHLTimerService *srv)
 	btm_rbt_foreach(srv->btm_tree, bt_hdbg_func, (void *) &hdbg);
 	ERTS_HLT_ASSERT(hdbg.found_root);
     }
+#endif
 }
 
 #endif /* ERTS_HLT_HARD_DEBUG */
diff --git a/erts/emulator/beam/erl_hl_timer.h b/erts/emulator/beam/erl_hl_timer.h
index 9cdcd581a0..f70fcdd1c0 100644
--- a/erts/emulator/beam/erl_hl_timer.h
+++ b/erts/emulator/beam/erl_hl_timer.h
@@ -21,7 +21,7 @@
 #ifndef ERL_HL_TIMER_H__
 #define ERL_HL_TIMER_H__
 
-typedef struct ErtsHLTimer_ ErtsBifTimers;
+typedef struct ErtsBifTimer_ ErtsBifTimers;
 typedef struct ErtsHLTimerService_ ErtsHLTimerService;
 
 #include "sys.h"
@@ -56,7 +56,7 @@ void erts_cancel_proc_timer(Process *);
 void erts_set_port_timer(Port *, Sint64);
 void erts_cancel_port_timer(Port *);
 Sint64 erts_read_port_timer(Port *);
-int erts_cancel_bif_timers(Process *, ErtsBifTimers *, void **);
+int erts_cancel_bif_timers(Process *, ErtsBifTimers **, void **);
 int erts_detach_accessor_bif_timers(Process *, ErtsBifTimers *, void **);
 ErtsHLTimerService *erts_create_timer_service(void);
 void erts_hl_timer_init(void);
diff --git a/erts/emulator/beam/erl_init.c b/erts/emulator/beam/erl_init.c
index 541bfec532..ac0324d846 100644
--- a/erts/emulator/beam/erl_init.c
+++ b/erts/emulator/beam/erl_init.c
@@ -2334,6 +2334,7 @@ erl_start(int argc, char **argv)
 	set_main_stack_size();
 	erts_sched_init_time_sup(esdp);
         erts_ets_sched_spec_data_init(esdp);
+        erts_aux_work_timeout_late_init(esdp);
 	process_main(esdp->x_reg_array, esdp->f_reg_array);
     }
 #endif
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index 894e0ee582..a7c90867f0 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -562,7 +562,6 @@ static int stack_element_dump(fmtfn_t to, void *to_arg, Eterm* sp, int yreg);
 
 static void aux_work_timeout(void *unused);
 static void aux_work_timeout_early_init(int no_schedulers);
-static void aux_work_timeout_late_init(void);
 static void setup_aux_work_timer(ErtsSchedulerData *esdp);
 
 static int execute_sys_tasks(Process *c_p,
@@ -2792,6 +2791,9 @@ typedef struct {
 
     int initialized;
     erts_atomic32_t refc;
+#ifdef DEBUG
+    erts_atomic32_t used;
+#endif
     erts_atomic32_t type[1];
 } ErtsAuxWorkTmo;
 
@@ -2801,6 +2803,13 @@ static ERTS_INLINE void
 start_aux_work_timer(ErtsSchedulerData *esdp)
 {
     ErtsMonotonicTime tmo = erts_get_monotonic_time(esdp);
+#ifdef DEBUG
+    Uint no = (Uint) erts_atomic32_xchg_mb(&aux_work_tmo->used,
+                                           (erts_aint32_t) esdp->no);
+    ASSERT(esdp->type == ERTS_SCHED_NORMAL);
+    ASSERT(!no);
+#endif
+
     tmo = ERTS_MONOTONIC_TO_CLKTCKS(tmo-1);
     tmo += ERTS_MSEC_TO_CLKTCKS(1000) + 1;
     erts_twheel_init_timer(&aux_work_tmo->timer.data);
@@ -2808,7 +2817,6 @@ start_aux_work_timer(ErtsSchedulerData *esdp)
     erts_twheel_set_timer(esdp->timer_wheel,
 			  &aux_work_tmo->timer.data,
 			  aux_work_timeout,
-			  NULL,
 			  (void *) esdp,
 			  tmo);
 }
@@ -2837,16 +2845,19 @@ aux_work_timeout_early_init(int no_schedulers)
     aux_work_tmo = (ErtsAuxWorkTmo *) p;
     aux_work_tmo->initialized = 0;
     erts_atomic32_init_nob(&aux_work_tmo->refc, 0);
+#ifdef DEBUG
+    erts_atomic32_init_nob(&aux_work_tmo->used, 0);
+#endif
     for (i = 0; i <= no_schedulers; i++)
 	erts_atomic32_init_nob(&aux_work_tmo->type[i], 0);
 }
 
 void
-aux_work_timeout_late_init(void)
+erts_aux_work_timeout_late_init(ErtsSchedulerData *esdp)
 {
     aux_work_tmo->initialized = 1;
-    if (erts_atomic32_read_nob(&aux_work_tmo->refc))
-	start_aux_work_timer(erts_get_scheduler_data());
+    if (erts_atomic32_read_acqb(&aux_work_tmo->refc))
+	start_aux_work_timer(esdp);
 }
 
 static void
@@ -2854,6 +2865,13 @@ aux_work_timeout(void *vesdp)
 {
     erts_aint32_t refc;
     int i;
+#ifdef DEBUG
+    ErtsSchedulerData *esdp = erts_get_scheduler_data();
+    Uint no = (Uint) erts_atomic32_xchg_mb(&aux_work_tmo->used, 0);
+    ASSERT(no == esdp->no);
+    ASSERT(esdp == (ErtsSchedulerData *) vesdp);
+#endif
+
 #ifdef ERTS_SMP
     i = 0;
 #else
@@ -6467,8 +6485,6 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online
     /* init port tasks */
     erts_port_task_init();
 
-    aux_work_timeout_late_init();
-
 #ifndef ERTS_SMP
 #ifdef ERTS_DO_VERIFY_UNUSED_TEMP_ALLOC
     erts_scheduler_data->verify_unused_temp_alloc
@@ -8754,6 +8770,9 @@ sched_thread_func(void *vesdp)
 
     erts_sched_init_time_sup(esdp);
 
+    if (no == 1)
+        erts_aux_work_timeout_late_init(esdp);
+
     (void) ERTS_RUNQ_FLGS_SET_NOB(esdp->run_queue,
 				  ERTS_RUNQ_FLG_EXEC);
 
@@ -12429,9 +12448,6 @@ erl_create_process(Process* parent, /* Parent of process (default group leader).
     p->msg_inq.len = 0;
 #endif
     p->bif_timers = NULL;
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-    p->accessor_bif_timers = NULL;
-#endif
     p->mbuf = NULL;
     p->msg_frag = NULL;
     p->mbuf_sz = 0;
@@ -12630,9 +12646,6 @@ void erts_init_empty_process(Process *p)
     p->msg.save = &p->msg.first;
     p->msg.len = 0;
     p->bif_timers = NULL;
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-    p->accessor_bif_timers = NULL;
-#endif
     p->dictionary = NULL;
     p->seq_trace_clock = 0;
     p->seq_trace_lastcnt = 0;
@@ -12733,9 +12746,6 @@ erts_debug_verify_clean_empty_process(Process* p)
     ASSERT(p->msg.first == NULL);
     ASSERT(p->msg.len == 0);
     ASSERT(p->bif_timers == NULL);
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-    ASSERT(p->accessor_bif_timers == NULL);
-#endif
     ASSERT(p->dictionary == NULL);
     ASSERT(p->catches == 0);
     ASSERT(p->cp == NULL);
@@ -13801,7 +13811,7 @@ erts_continue_exit_process(Process *p)
 
     ASSERT(erts_proc_read_refc(p) > 0);
     if (p->bif_timers) {
-	if (erts_cancel_bif_timers(p, p->bif_timers, &p->u.terminate)) {
+	if (erts_cancel_bif_timers(p, &p->bif_timers, &p->u.terminate)) {
 	    ASSERT(erts_proc_read_refc(p) > 0);
 	    goto yield;
 	}
@@ -13809,19 +13819,6 @@ erts_continue_exit_process(Process *p)
 	p->bif_timers = NULL;
     }
 
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-    if (p->accessor_bif_timers) {
-	if (erts_detach_accessor_bif_timers(p,
-					    p->accessor_bif_timers,
-					    &p->u.terminate)) {
-	    ASSERT(erts_proc_read_refc(p) > 0);
-	    goto yield;
-	}
-	ASSERT(erts_proc_read_refc(p) > 0);
-	p->accessor_bif_timers = NULL;
-    }
-#endif
-
 #ifdef ERTS_SMP
     if (p->flags & F_SCHDLR_ONLN_WAITQ)
 	abort_sched_onln_chng_waitq(p);
diff --git a/erts/emulator/beam/erl_process.h b/erts/emulator/beam/erl_process.h
index 883d9f2a4c..b21597d63b 100644
--- a/erts/emulator/beam/erl_process.h
+++ b/erts/emulator/beam/erl_process.h
@@ -1026,9 +1026,6 @@ struct process {
     ErlMessageQueue msg;	/* Message queue */
 
     ErtsBifTimers *bif_timers;	/* Bif timers aiming at this process */
-#ifdef ERTS_BTM_ACCESSOR_SUPPORT
-    ErtsBifTimers *accessor_bif_timers;	/* Accessor bif timers */
-#endif
 
     ProcDict  *dictionary;       /* Process dictionary, may be NULL */
 
@@ -1830,6 +1827,7 @@ void erts_schedule_multi_misc_aux_work(int ignore_self,
 				       void (*func)(void *),
 				       void *arg);
 erts_aint32_t erts_set_aux_work_timeout(int, erts_aint32_t, int);
+void erts_aux_work_timeout_late_init(ErtsSchedulerData *esdp);
 void erts_sched_notify_check_cpu_bind(void);
 Uint erts_active_schedulers(void);
 void erts_init_process(int, int, int);
diff --git a/erts/emulator/beam/erl_time.h b/erts/emulator/beam/erl_time.h
index a1c4220633..46d6da6448 100644
--- a/erts/emulator/beam/erl_time.h
+++ b/erts/emulator/beam/erl_time.h
@@ -21,19 +21,52 @@
 #ifndef ERL_TIME_H__
 #define ERL_TIME_H__
 
-/* timer wheel size NEED to be a power of 2 */
-#ifdef SMALL_MEMORY
-#define ERTS_TIW_SIZE (1 << 13)
-#else
-#define ERTS_TIW_SIZE (1 << 16)
+#if 0
+#  define ERTS_TW_DEBUG
+#endif
+#if defined(DEBUG) && !defined(ERTS_TW_DEBUG)
+#  define ERTS_TW_DEBUG
 #endif
 
-#if defined(DEBUG) || 0
+#if defined(ERTS_TW_DEBUG)
 #define ERTS_TIME_ASSERT(B) ERTS_ASSERT(B)
 #else
 #define ERTS_TIME_ASSERT(B) ((void) 1)
 #endif
 
+#ifdef ERTS_TW_DEBUG
+/*
+ * Soon wheel will handle about 1 seconds
+ * Later wheel will handle about 8 minutes
+ */
+#  define ERTS_TW_SOON_WHEEL_BITS 10
+#  define ERTS_TW_LATER_WHEEL_BITS 10
+#else
+#  ifdef SMALL_MEMORY
+/*
+ * Soon wheel will handle about 4 seconds
+ * Later wheel will handle about 2 hours and 19 minutes
+ */
+#    define ERTS_TW_SOON_WHEEL_BITS 12
+#    define ERTS_TW_LATER_WHEEL_BITS 12
+#  else
+/*
+ * Soon wheel will handle about 16 seconds
+ * Later wheel will handle about 37 hours and 16 minutes
+ */
+#    define ERTS_TW_SOON_WHEEL_BITS 14
+#    define ERTS_TW_LATER_WHEEL_BITS 14
+#  endif
+#endif
+
+/*
+ * Number of slots in each timer wheel...
+ *
+ * These *need* to be a power of 2
+ */
+#define ERTS_TW_SOON_WHEEL_SIZE (1 << ERTS_TW_SOON_WHEEL_BITS)
+#define ERTS_TW_LATER_WHEEL_SIZE (1 << ERTS_TW_LATER_WHEEL_BITS)
+
 typedef enum {
     ERTS_NO_TIME_WARP_MODE,
     ERTS_SINGLE_TIME_WARP_MODE,
@@ -103,7 +136,10 @@ Eterm erts_system_time_source(struct process*c_p);
 #define ERTS_CLKTCK_RESOLUTION (erts_time_sup__.r.o.clktck_resolution)
 #endif
 
-#define ERTS_TIMER_WHEEL_MSEC (ERTS_TIW_SIZE/(ERTS_CLKTCK_RESOLUTION/1000))
+#define ERTS_TW_SOON_WHEEL_MSEC (ERTS_TW_SOON_WHEEL_SIZE/(ERTS_CLKTCK_RESOLUTION/1000))
+#define ERTS_TW_LATER_WHEEL_MSEC (ERTS_TW_LATER_WHEEL_SIZE*ERTS_TW_SOON_WHEEL_MSEC/2)
+
+#define ERTS_TIMER_WHEEL_MSEC ERTS_TW_LATER_WHEEL_MSEC
 
 struct erts_time_sup_read_only__ {
     ErtsMonotonicTime monotonic_time_unit;
@@ -412,34 +448,25 @@ erts_time_unit_conversion(Uint64 value,
 void erts_sched_init_time_sup(ErtsSchedulerData *esdp);
 
 
-#define ERTS_TWHEEL_SLOT_AT_ONCE -1
-#define ERTS_TWHEEL_SLOT_INACTIVE -2
+#define ERTS_TW_SLOT_INACTIVE (-2)
 
 /*
 ** Timer entry:
 */
 typedef struct erl_timer {
-    struct erl_timer* next;	/* next entry tiw slot or chain */
-    struct erl_timer* prev;	/* prev entry tiw slot or chain */
-    union {
-	struct {
-	    void (*timeout)(void*); /* called when timeout */
-	    void (*cancel)(void*);  /* called when cancel (may be NULL) */
-	    void* arg;              /* argument to timeout/cancel procs */
-	} func;
-	ErtsThrPrgrLaterOp cleanup;
-    } u;
     ErtsMonotonicTime timeout_pos; /* Timeout in absolute clock ticks */
+    struct erl_timer* next;     /* next entry tiw slot or chain */
+    struct erl_timer* prev;	/* prev entry tiw slot or chain */
+    void (*timeout)(void*); /* called when timeout */
+    void* arg;              /* argument to timeout/cancel procs */
     int slot;
 } ErtsTWheelTimer;
 
 typedef void (*ErlTimeoutProc)(void*);
-typedef void (*ErlCancelProc)(void*);
 
 void erts_twheel_set_timer(ErtsTimerWheel *tiw,
 			   ErtsTWheelTimer *p, ErlTimeoutProc timeout,
-			   ErlCancelProc cancel, void *arg,
-			   ErtsMonotonicTime timeout_pos);
+			   void *arg, ErtsMonotonicTime timeout_pos);
 void erts_twheel_cancel_timer(ErtsTimerWheel *tiw, ErtsTWheelTimer *p);
 ErtsTimerWheel *erts_create_timer_wheel(ErtsSchedulerData *esdp);
 
@@ -447,12 +474,13 @@ ErtsMonotonicTime erts_check_next_timeout_time(ErtsSchedulerData *);
 
 ERTS_GLB_INLINE void erts_twheel_init_timer(ErtsTWheelTimer *p);
 ERTS_GLB_INLINE ErtsMonotonicTime erts_next_timeout_time(ErtsNextTimeoutRef);
+ERTS_GLB_INLINE ErtsMonotonicTime erts_tweel_read_timeout(ErtsTWheelTimer *twt);
 
 #if ERTS_GLB_INLINE_INCL_FUNC_DEF
 
 ERTS_GLB_INLINE void erts_twheel_init_timer(ErtsTWheelTimer *p)
 {
-    p->slot = ERTS_TWHEEL_SLOT_INACTIVE;
+    p->slot = ERTS_TW_SLOT_INACTIVE;
 }
 
 ERTS_GLB_INLINE ErtsMonotonicTime erts_next_timeout_time(ErtsNextTimeoutRef nxt_tmo_ref)
@@ -460,6 +488,12 @@ ERTS_GLB_INLINE ErtsMonotonicTime erts_next_timeout_time(ErtsNextTimeoutRef nxt_
     return *((ErtsMonotonicTime *) nxt_tmo_ref);
 }
 
+ERTS_GLB_INLINE ErtsMonotonicTime
+erts_tweel_read_timeout(ErtsTWheelTimer *twt)
+{
+    return twt->timeout_pos;
+}
+
 #endif /* ERTS_GLB_INLINE_INCL_FUNC_DEF */
 
 void
diff --git a/erts/emulator/beam/erl_time_sup.c b/erts/emulator/beam/erl_time_sup.c
index cf9d3adc86..c69fec3c80 100644
--- a/erts/emulator/beam/erl_time_sup.c
+++ b/erts/emulator/beam/erl_time_sup.c
@@ -678,7 +678,6 @@ check_time_correction(void *vesdp)
     erts_twheel_set_timer(esdp->timer_wheel,
 			  &time_sup.inf.c.parmon.timer,
 			  check_time_correction,
-			  NULL,
 			  (void *) esdp,
 			  timeout_pos);
 }
@@ -729,7 +728,6 @@ check_time_offset(void *vesdp)
     erts_twheel_set_timer(esdp->timer_wheel,
 			  &time_sup.inf.c.parmon.timer,
 			  check_time_offset,
-			  NULL,
 			  vesdp,
 			  timeout_pos);
 }
@@ -836,7 +834,6 @@ late_init_time_correction(ErtsSchedulerData *esdp)
     erts_twheel_set_timer(esdp->timer_wheel,
 			  &time_sup.inf.c.parmon.timer,
 			  check_func,
-			  NULL,
 			  (quick_init_drift_adj
 			   ? NULL
 			   : esdp),
diff --git a/erts/emulator/beam/time.c b/erts/emulator/beam/time.c
index 6f15082130..cee3cb619f 100644
--- a/erts/emulator/beam/time.c
+++ b/erts/emulator/beam/time.c
@@ -17,57 +17,157 @@
  * 
  * %CopyrightEnd%
  */
-
+ 
 /*
- * TIMING WHEEL
+ * TIMER WHEEL
+ *
+ *
+ * The time scale used for timers is Erlang monotonic time. The
+ * time unit used is ERTS specific clock ticks. A clock tick is
+ * currently defined to 1 millisecond. That is, the resolution of
+ * timers triggered by the runtime system is 1 millisecond.
  * 
- * Timeouts kept in an wheel. A timeout is measured relative to the
- * current slot (tiw_pos) in the wheel, and inserted at slot 
- * (tiw_pos + timeout) % TIW_SIZE. Each timeout also has a count
- * equal to timeout/TIW_SIZE, which is needed since the time axis
- * is wrapped arount the wheel. 
+ * When a timer is set, it is determined at what Erlang monotonic
+ * time, in clock ticks, it should be triggered.
  *
- * Several slots may be processed in one operation. If the number of
- * slots is greater that the wheel size, the wheel is only traversed
- * once,
+ * The 'pos' field of the wheel corresponds to current time of
+ * the wheel. That is, it corresponds to Erlang monotonic time in
+ * clock tick time unit. The 'pos' field of the wheel is
+ * monotonically increased when erts_bump_timers() is called. All
+ * timers in the wheel that have a time less than or equal to
+ * 'pos' are triggered by the bump operation. The bump operation
+ * may however be spread over multiple calls to erts_bump_timers()
+ * if there are a lots of timers to trigger.
  *
- * The following example shows a time axis where there is one timeout
- * at each "tick", and where 1, 2, 3 ... wheel slots are released in
- * one operation. The notation "<x" means "release all items with
- * counts less than x". 
+ * Each scheduler thread maintains its own timer wheel. The timer
+ * wheel of a scheduler, however, actually consists of two wheels.
+ * A soon wheel and a later wheel.
+ *
+ *
+ * -- The Soon Wheel --
+ *
+ * The soon wheel contain timers that should be triggered soon.
+ * That is, they are soon to be triggered. Each slot in the soon
+ * wheel is 1 clock tick wide. The number of slots in the soon
+ * wheel is currently 2¹⁴. That is, it contains timers in the
+ * range ('pos', 'pos' + 2¹⁴] which corresponds to a bit more
+ * than 16 seconds.
+ *
+ * When the bump operation is started, 'pos' is moved forward to a
+ * position that corresponds to current Erlang monotonic time. Then
+ * all timers that are in the range (old 'pos', new 'pos'] are
+ * triggered. During a bump operation, the soon wheel may contain
+ * timers in the two, possibly overlapping, ranges (old 'pos',
+ * old 'pos' + 2¹⁴], and (new 'pos', new 'pos' + 2¹⁴]. This may
+ * occur even if the bump operation doesn't yield, due to timeout
+ * callbacks inserting new timers.
+ *
+ *
+ * -- The Later Wheel --
+ *
+ * The later wheel contain timers that are further away from 'pos'
+ * than the width of the soon timer wheel. That is, currently
+ * timers further away from 'pos' than 2¹⁴ clock ticks. The width
+ * of each slot in the later wheel is half the width of the soon
+ * wheel. That is, each slot is currently 2¹³ clock ticks wide
+ * which corresponds to about 8 seconds. If three timers of the
+ * times 'pos' + 17000, 'pos' + 18000, and 'pos' + 19000 are
+ * inserted, they will all end up in the same slot in the later
+ * wheel.
+ *
+ * The number of slots in the later wheel is currently the same as
+ * in the soon wheel, i.e. 2¹⁴. That is, one revolution of the later
+ * wheel currently corresponds to 2¹⁴×2¹³ clock ticks which is
+ * almost 37 ½ hour. Timers even further away than that are put in
+ * the later slot identified by their time modulo the size of the later
+ * wheel. Such timers are however very uncommon. Most timers used
+ * by the runtime system will utilize the high level timer API.
+ * The high level timer implementation will not insert timers
+ * further away then one revolution into the later wheel. It will
+ * instead keep such timers in a tree of very long timers. The
+ * high level timer implementation utilize one timer wheel timer
+ * for the management of this tree of timers. This timer is set to
+ * the closest timeout in the tree. This timer may however be
+ * further away than one revolution in the later wheel.
+ *
+ * The 'later.pos' field identifies next position in the later wheel.
+ * 'later.pos' is always increased by the width of a later wheel slot.
+ * That is, currently 2¹³ clock ticks. When 'pos' is moved (during
+ * a bump operation) closer to 'later.pos' than the width of a later
+ * wheel slot, i.e. currently when 'pos' + 2¹³ ≥ 'later.pos', we
+ * inspect the slot identified by 'later.pos' and then move 'later.pos'
+ * forward. When inspecting the later slot we move all timers in the
+ * slot, that are in the soon wheel range, from the later wheel to
+ * the soon wheel. Timers one or more revolutions of the later wheel
+ * away are kept in the slot.
+ *
+ * During normal operation, timers originally located in the later
+ * wheel will currently be moved into the soon wheel about 8 to
+ * 16 seconds before they should be triggered. During extremely
+ * heavy load, the scheduler might however be heavily delayed, so
+ * the code must be prepared for situations where time for
+ * triggering the timer has passed when we inspect the later wheel
+ * slot, and then trigger the timer immediately. We must also be
+ * prepared to inspect multiple later wheel slots at once due to the
+ * delay.
+ *
+ *
+ * -- Slot Management --
+ *
+ * All timers of a slot are placed in a circular double linked
+ * list. This makes insertion and removal of a timer O(1).
+ *
+ * While bumping timers in a slot, we move the circular list
+ * away from the slot, and refer to it from the 'sentinel'
+ * field. The list will stay there until we are done with it
+ * even if the bump operation should yield. The cancel operation
+ * can remove the timer from this position as well as from the
+ * slot position by just removing it from the circular double
+ * linked list that it is in.
+ *
+ * -- At Once Slot --
+ *
+ * If a timer is set that has a time earlier or equal to 'pos',
+ * it is not inserted into the wheel. It is instead inserted,
+ * into a circular double linked list referred to by the "at
+ * once" slot. When the bump operation is performed these timers
+ * will be triggered at once. The circular list of the slot will
+ * be moved to the 'sentinel' field while bumping these timers
+ * as when bumping an ordinary wheel slot. A yielding bump
+ * operation and cancelation of timers is handled the same way
+ * as if the timer was in a wheel slot.
+ *
+ * -- Searching for Next Timeout --
+ *
+ * In order to limit the amount of work needed in order to find
+ * next timeout, we keep track of total amount of timers in the
+ * wheels, total amount of timers in the later wheel, total amount
+ * of timers in soon wheel, and the total amount of timers in
+ * each range of slots. Each slot range currently contain 512
+ * slots.
+ *
+ * When next timeout is less than the soon wheel width away we
+ * determine the exact timeout. Due to the timer counts of
+ * slot ranges, we currently at most need to search 1024 slots
+ * in the soon wheel. This besides inspecting slot range counts
+ * and two slots in the later wheel which potentially might trigger
+ * timeouts for moving timers from the later wheel to the soon wheel
+ * earlier than timeouts in the soon wheel. We also keep track
+ * of latest known minimum timeout position in each wheel which
+ * makes it possible to avoid scanning from current position
+ * each time.
+ *
+ * When next timeout is further away than the soon wheel width
+ * we settle for the earliest possible timeout in the first
+ * non-empty slot range. The further away the next timeout is, the
+ * more likely it is that the next timeout change before we
+ * actually get there. That is, a change due to another timer is
+ * set to an earlier time and/or the timer is cancelled. It is
+ * therefore in this case no point determining next timeout
+ * exactly. If the state should not change, we will wake up a bit
+ * early and do a recalculation of next timeout and eventually
+ * we will be so close to it that we determine it exactly.
  *
- * Size of wheel: 4
- * 
- *   --|----|----|----|----|----|----|----|----|----|----|----|----|----
- *    0.0  0.1  0.2  0.3  1.0  1.1  1.2  1.3  2.0  2.1  2.2  2.3  3.0
- * 
- * 1   [    )
- *     <1  0.1  0.2  0.3  0.0  1.1  1.2  1.3  1.0  2.1  2.2  2.3  2.0
- * 
- * 2   [         )
- *     <1   <1  0.2  0.3  0.0  0.1  1.2  1.3  1.0  1.1  2.2  2.3  2.0
- * 
- * 3   [              )
- *     <1   <1   <1  0.3  0.0  0.1  0.2  1.3  1.0  1.1  1.2  2.3  2.0
- * 
- * 4   [                   )
- *     <1   <1   <1   <1  0.0  0.1  0.2  0.3  1.0  1.1  1.2  1.3  2.0
- * 
- * 5   [                        )
- *     <2   <1   <1   <1.      0.1  0.2  0.3  0.0  1.1  1.2  1.3  1.0
- * 
- * 6   [                             )
- *     <2   <2   <1   <1.           0.2  0.3  0.0  0.1  1.2  1.3  1.0
- * 
- * 7   [                                  )
- *     <2   <2   <2   <1.                0.3  0.0  0.1  0.2  1.3  1.0
- * 
- * 8   [                                       )   
- *     <2   <2   <2   <2.                     0.0  0.1  0.2  0.3  1.0
- * 
- * 9   [                                            )
- *     <3   <2   <2   <2.                          0.1  0.2  0.3  0.0
- * 
  */
 
 #ifdef HAVE_CONFIG_H
@@ -80,8 +180,11 @@
 #define ERTS_WANT_TIMER_WHEEL_API
 #include "erl_time.h"
 
-#define ERTS_MONOTONIC_DAY ERTS_SEC_TO_MONOTONIC(60*60*24)
-#define ERTS_CLKTCKS_DAY ERTS_MONOTONIC_TO_CLKTCKS(ERTS_MONOTONIC_DAY)
+#define ERTS_MAX_CLKTCKS \
+    ERTS_MONOTONIC_TO_CLKTCKS(ERTS_MONOTONIC_TIME_MAX)
+
+#define ERTS_CLKTCKS_WEEK \
+    ERTS_MONOTONIC_TO_CLKTCKS(ERTS_SEC_TO_MONOTONIC(7*60*60*24))
 
 #ifdef ERTS_ENABLE_LOCK_CHECK
 #define ASSERT_NO_LOCKED_LOCKS		erts_lc_check_exact(NULL, 0)
@@ -90,6 +193,10 @@
 #endif
 
 #if 0
+#  define ERTS_TW_HARD_DEBUG
+#endif
+
+#if defined(ERTS_TW_HARD_DEBUG) && !defined(ERTS_TW_DEBUG)
 #  define ERTS_TW_DEBUG
 #endif
 #if defined(DEBUG) && !defined(ERTS_TW_DEBUG)
@@ -97,16 +204,62 @@
 #endif
 
 #undef ERTS_TW_ASSERT
-#if defined(ERTS_TW_DEBUG)
+#if defined(ERTS_TW_DEBUG) 
 #  define ERTS_TW_ASSERT(E) ERTS_ASSERT(E)
 #else
 #  define ERTS_TW_ASSERT(E) ((void) 1)
 #endif
 
 #ifdef ERTS_TW_DEBUG
-#  define ERTS_TWHEEL_BUMP_YIELD_LIMIT 5
+#  define ERTS_TWHEEL_BUMP_YIELD_LIMIT        500
 #else
-#  define ERTS_TWHEEL_BUMP_YIELD_LIMIT 100
+#  define ERTS_TWHEEL_BUMP_YIELD_LIMIT        10000
+#endif
+#define ERTS_TW_COST_SLOT                     1
+#define ERTS_TW_COST_SLOT_MOVE                5
+#define ERTS_TW_COST_TIMEOUT                  100
+
+/*
+ * Every slot in the soon wheel is a clock tick (as defined
+ * by ERTS) wide. A clock tick is currently 1 milli second.
+ */
+
+#define ERTS_TW_SOON_WHEEL_FIRST_SLOT 0
+#define ERTS_TW_SOON_WHEEL_END_SLOT \
+    (ERTS_TW_SOON_WHEEL_FIRST_SLOT + ERTS_TW_SOON_WHEEL_SIZE)
+
+#define ERTS_TW_SOON_WHEEL_MASK (ERTS_TW_SOON_WHEEL_SIZE-1)
+
+/*
+ * Every slot in the later wheel is as wide as half the size
+ * of the soon wheel.
+ */
+
+#define ERTS_TW_LATER_WHEEL_SHIFT (ERTS_TW_SOON_WHEEL_BITS - 1)
+#define ERTS_TW_LATER_WHEEL_SLOT_SIZE \
+    ((ErtsMonotonicTime) (1 << ERTS_TW_LATER_WHEEL_SHIFT))
+#define ERTS_TW_LATER_WHEEL_POS_MASK \
+    (~((ErtsMonotonicTime) (1 << ERTS_TW_LATER_WHEEL_SHIFT)-1))
+
+#define ERTS_TW_LATER_WHEEL_FIRST_SLOT ERTS_TW_SOON_WHEEL_SIZE
+#define ERTS_TW_LATER_WHEEL_END_SLOT \
+    (ERTS_TW_LATER_WHEEL_FIRST_SLOT + ERTS_TW_LATER_WHEEL_SIZE)
+
+#define ERTS_TW_LATER_WHEEL_MASK (ERTS_TW_LATER_WHEEL_SIZE-1)
+
+#define ERTS_TW_SCNT_BITS 9
+#define ERTS_TW_SCNT_SHIFT 
+#define ERTS_TW_SCNT_SIZE \
+    ((ERTS_TW_SOON_WHEEL_SIZE + ERTS_TW_LATER_WHEEL_SIZE) \
+     >> ERTS_TW_SCNT_BITS)
+
+#ifdef __GNUC__
+#if ERTS_TW_SOON_WHEEL_BITS < ERTS_TW_SCNT_BITS
+#  warning Consider larger soon timer wheel
+#endif
+#if ERTS_TW_SOON_WHEEL_BITS < ERTS_TW_SCNT_BITS
+#  warning Consider larger later timer wheel
+#endif
 #endif
 
 /* Actual interval time chosen by sys_init_time() */
@@ -119,95 +272,360 @@ static int tiw_itime; /* Constant after init */
 #  define TIW_ITIME tiw_itime
 #endif
 
+const int etp_tw_soon_wheel_size = ERTS_TW_SOON_WHEEL_SIZE;
+const ErtsMonotonicTime etp_tw_soon_wheel_mask = ERTS_TW_SOON_WHEEL_MASK;
+const int etp_tw_soon_wheel_first_slot = ERTS_TW_SOON_WHEEL_FIRST_SLOT;
+
+const int etp_tw_later_wheel_size = ERTS_TW_LATER_WHEEL_SIZE;
+const ErtsMonotonicTime etp_tw_later_wheel_slot_size = ERTS_TW_LATER_WHEEL_SLOT_SIZE;
+const int etp_tw_later_wheel_shift = ERTS_TW_LATER_WHEEL_SHIFT;
+const ErtsMonotonicTime etp_tw_later_wheel_mask = ERTS_TW_LATER_WHEEL_MASK;
+const ErtsMonotonicTime etp_tw_later_wheel_pos_mask = ERTS_TW_LATER_WHEEL_POS_MASK;
+const int etp_tw_later_wheel_first_slot = ERTS_TW_LATER_WHEEL_FIRST_SLOT;
+
 struct ErtsTimerWheel_ {
-    ErtsTWheelTimer *w[ERTS_TIW_SIZE];
+    ErtsTWheelTimer *slots[1                            /* At Once Slot */
+                           + ERTS_TW_SOON_WHEEL_SIZE    /* Soon Wheel Slots */
+                           + ERTS_TW_LATER_WHEEL_SIZE]; /* Later Wheel Slots */
+    ErtsTWheelTimer **w;
+    Sint scnt[ERTS_TW_SCNT_SIZE];
+    Sint bump_scnt[ERTS_TW_SCNT_SIZE];
     ErtsMonotonicTime pos;
     Uint nto;
     struct {
-	ErtsTWheelTimer *head;
-	ErtsTWheelTimer *tail;
 	Uint nto;
     } at_once;
+    struct {
+        ErtsMonotonicTime min_tpos;
+        Uint nto;
+    } soon;
+    struct {
+        ErtsMonotonicTime min_tpos;
+        int min_tpos_slot;
+        ErtsMonotonicTime pos;
+        Uint nto;
+    } later;
     int yield_slot;
     int yield_slots_left;
-    int yield_start_pos;
     ErtsTWheelTimer sentinel;
     int true_next_timeout_time;
+    ErtsMonotonicTime next_timeout_pos;
     ErtsMonotonicTime next_timeout_time;
 };
 
-static ERTS_INLINE ErtsMonotonicTime
-find_next_timeout(ErtsSchedulerData *esdp,
-		  ErtsTimerWheel *tiw,
-		  int search_all,
-		  ErtsMonotonicTime curr_time,       /* When !search_all */
-		  ErtsMonotonicTime max_search_time) /* When !search_all */
+#define ERTS_TW_SLOT_AT_ONCE (-1)
+
+#define ERTS_TW_BUMP_LATER_WHEEL(TIW) \
+    ((tiw)->pos + ERTS_TW_LATER_WHEEL_SLOT_SIZE >= (TIW)->later.pos)
+
+static int bump_later_wheel(ErtsTimerWheel *tiw, int *yield_count_p);
+
+#ifdef ERTS_TW_DEBUG
+#define ERTS_TW_DBG_VERIFY_EMPTY_SOON_SLOTS(TIW, TO_POS) \
+    dbg_verify_empty_soon_slots((TIW), (TO_POS))
+#define ERTS_TW_DBG_VERIFY_EMPTY_LATER_SLOTS(TIW, TO_POS) \
+    dbg_verify_empty_later_slots((TIW), (TO_POS))
+void dbg_verify_empty_soon_slots(ErtsTimerWheel *, ErtsMonotonicTime);
+void dbg_verify_empty_later_slots(ErtsTimerWheel *, ErtsMonotonicTime);
+#else
+#define ERTS_TW_DBG_VERIFY_EMPTY_SOON_SLOTS(TIW, TO_POS)
+#define ERTS_TW_DBG_VERIFY_EMPTY_LATER_SLOTS(TIW, TO_POS)
+#endif
+
+static ERTS_INLINE int
+scnt_get_ix(int slot)
 {
-    int start_ix, tiw_pos_ix;
-    ErtsTWheelTimer *p;
+    return slot >> ERTS_TW_SCNT_BITS;
+}
+
+static ERTS_INLINE void
+scnt_inc(Sint *scnt, int slot)
+{
+    scnt[slot >> ERTS_TW_SCNT_BITS]++;
+}
+
+#ifdef ERTS_TW_HARD_DEBUG
+
+static ERTS_INLINE void
+scnt_ix_inc(Sint *scnt, int six)
+{
+    scnt[six]++;
+}
+
+#endif
+
+static ERTS_INLINE void
+scnt_dec(Sint *scnt, int slot)
+{
+    scnt[slot >> ERTS_TW_SCNT_BITS]--;
+    ERTS_TW_ASSERT(scnt[slot >> ERTS_TW_SCNT_BITS] >= 0);
+}
+
+static ERTS_INLINE void
+scnt_ix_dec(Sint *scnt, int six)
+{
+    scnt[six]--;
+    ERTS_TW_ASSERT(scnt[six] >= 0);
+}
+
+static ERTS_INLINE void
+scnt_wheel_next(int *slotp, int *leftp, ErtsMonotonicTime *posp,
+                int *sixp, Sint *scnt, int first_slot,
+                int end_slot, ErtsMonotonicTime slot_sz)
+{
+    int slot = *slotp;
+    int left = *leftp;
+    int ix;
+
+    ERTS_TW_ASSERT(*leftp >= 0);
+
+    left--;
+    slot++;
+    if (slot == end_slot)
+        slot = first_slot;
+    ix = slot >> ERTS_TW_SCNT_BITS;
+
+    while (!scnt[ix] && left > 0) {
+        int diff, old_slot = slot;
+        ix++;
+        slot = (ix << ERTS_TW_SCNT_BITS);
+        diff = slot - old_slot;
+        if (left < diff) {
+            slot = old_slot + left;
+            diff = left;
+        }
+        if (slot < end_slot)
+            left -= diff;
+        else {
+            left -= end_slot - old_slot;
+            slot = first_slot;
+            ix = slot >> ERTS_TW_SCNT_BITS;
+        }
+    }
+
+    ERTS_TW_ASSERT(left >= -1);
+
+    if (posp)
+        *posp += slot_sz * ((ErtsMonotonicTime) (*leftp - left));
+    if (sixp)
+        *sixp = slot >> ERTS_TW_SCNT_BITS;
+    *leftp = left;
+    *slotp = slot;
+}
+
+
+static ERTS_INLINE void
+scnt_soon_wheel_next(int *slotp, int *leftp, ErtsMonotonicTime *posp,
+                    int *sixp, Sint *scnt)
+{
+    scnt_wheel_next(slotp, leftp, posp, sixp, scnt,
+                    ERTS_TW_SOON_WHEEL_FIRST_SLOT,
+                    ERTS_TW_SOON_WHEEL_END_SLOT, 1);
+}
+
+static ERTS_INLINE void
+scnt_later_wheel_next(int *slotp, int *leftp, ErtsMonotonicTime *posp,
+                    int *sixp, Sint *scnt)
+{
+    scnt_wheel_next(slotp, leftp, posp, sixp, scnt,
+                    ERTS_TW_LATER_WHEEL_FIRST_SLOT,
+                    ERTS_TW_LATER_WHEEL_END_SLOT,
+                    ERTS_TW_LATER_WHEEL_SLOT_SIZE);
+}
+
+
+static ERTS_INLINE int
+soon_slot(ErtsMonotonicTime soon_pos)
+{
+    ErtsMonotonicTime slot = soon_pos;
+    slot &= ERTS_TW_SOON_WHEEL_MASK;
+
+    ERTS_TW_ASSERT(ERTS_TW_SOON_WHEEL_FIRST_SLOT <= slot);
+    ERTS_TW_ASSERT(slot < ERTS_TW_SOON_WHEEL_END_SLOT);
+
+    return (int) slot;
+}
+
+static ERTS_INLINE int
+later_slot(ErtsMonotonicTime later_pos)
+{
+    ErtsMonotonicTime slot = later_pos;
+    slot >>= ERTS_TW_LATER_WHEEL_SHIFT;
+    slot &= ERTS_TW_LATER_WHEEL_MASK;
+    slot += ERTS_TW_LATER_WHEEL_FIRST_SLOT;
+
+    ERTS_TW_ASSERT(ERTS_TW_LATER_WHEEL_FIRST_SLOT <= slot);
+    ERTS_TW_ASSERT(slot < ERTS_TW_LATER_WHEEL_END_SLOT);
+
+    return (int) slot;
+}
+
+#ifdef ERTS_TW_HARD_DEBUG
+#define ERTS_HARD_DBG_CHK_WHEELS(TIW, CHK_MIN_TPOS) \
+    hrd_dbg_check_wheels((TIW), (CHK_MIN_TPOS))
+static void hrd_dbg_check_wheels(ErtsTimerWheel *tiw, int check_min_tpos);
+#else
+#define ERTS_HARD_DBG_CHK_WHEELS(TIW, CHK_MIN_TPOS)
+#endif
+
+static ErtsMonotonicTime
+find_next_timeout(ErtsSchedulerData *esdp, ErtsTimerWheel *tiw)
+{
+    int slot, slots;
     int true_min_timeout = 0;
-    ErtsMonotonicTime min_timeout, min_timeout_pos, slot_timeout_pos;
+    ErtsMonotonicTime min_timeout_pos;
+
+    ERTS_TW_ASSERT(tiw->pos + ERTS_TW_LATER_WHEEL_SLOT_SIZE < tiw->later.pos
+                   && tiw->later.pos <= tiw->pos + ERTS_TW_SOON_WHEEL_SIZE);
+
+    ERTS_HARD_DBG_CHK_WHEELS(tiw, 0);
+
+    ERTS_TW_ASSERT(tiw->yield_slot == ERTS_TW_SLOT_INACTIVE);
 
     if (tiw->nto == 0) { /* no timeouts in wheel */
-	if (!search_all)
-	    min_timeout_pos = tiw->pos;
-	else {
-	    curr_time = erts_get_monotonic_time(esdp);
-	    tiw->pos = min_timeout_pos = ERTS_MONOTONIC_TO_CLKTCKS(curr_time);
-	}
-	min_timeout_pos += ERTS_MONOTONIC_TO_CLKTCKS(ERTS_MONOTONIC_DAY);
-	goto found_next;
+        ErtsMonotonicTime curr_time = erts_get_monotonic_time(esdp);
+        tiw->pos = min_timeout_pos = ERTS_MONOTONIC_TO_CLKTCKS(curr_time);
+        tiw->later.pos = min_timeout_pos + ERTS_TW_SOON_WHEEL_SIZE;
+        tiw->later.pos &= ERTS_TW_LATER_WHEEL_POS_MASK;
+	min_timeout_pos += ERTS_CLKTCKS_WEEK;
+	goto done;
     }
 
-    slot_timeout_pos = min_timeout_pos = tiw->pos;
-    if (search_all)
-	min_timeout_pos += ERTS_MONOTONIC_TO_CLKTCKS(ERTS_MONOTONIC_DAY);
-    else
-	min_timeout_pos = ERTS_MONOTONIC_TO_CLKTCKS(curr_time + max_search_time);
+    ERTS_TW_ASSERT(tiw->soon.nto || tiw->later.nto);
 
-    start_ix = tiw_pos_ix = (int) (tiw->pos & (ERTS_TIW_SIZE-1));
+    if (!tiw->soon.nto) {
+        ErtsMonotonicTime tpos, min_tpos;
 
-    do {
-	if (++slot_timeout_pos >= min_timeout_pos)
-	    break;
-
-	p = tiw->w[tiw_pos_ix];
-
-	if (p) {
-	    ErtsTWheelTimer *end = p;
-
-	    do  {
-		ErtsMonotonicTime timeout_pos;
-		timeout_pos = p->timeout_pos;
-		if (min_timeout_pos > timeout_pos) {
-		    true_min_timeout = 1;
-		    min_timeout_pos = timeout_pos;
-		    if (min_timeout_pos <= slot_timeout_pos)
-			goto found_next;
-		}
-		p = p->next;
-	    } while (p != end);
-	}
+        /* Search later wheel... */
+
+        min_tpos = tiw->later.min_tpos & ERTS_TW_LATER_WHEEL_POS_MASK;
+
+        if (min_tpos <= tiw->later.pos) {
+            tpos = tiw->later.pos;
+            slots = ERTS_TW_LATER_WHEEL_SIZE;
+        }
+        else {
+            ErtsMonotonicTime tmp;
+            /* Don't inspect slots we know are empty... */
+            tmp = min_tpos - tiw->later.pos;
+            tmp /= ERTS_TW_LATER_WHEEL_SLOT_SIZE;
+            if (tmp >= ERTS_TW_LATER_WHEEL_SIZE) {
+                /* Timeout more than one revolution ahead... */
+
+                /* Pre-timeout for move from later to soon wheel... */
+                min_timeout_pos = min_tpos - ERTS_TW_LATER_WHEEL_SLOT_SIZE;
+                goto done;
+            }
+            tpos = min_tpos;
+            ERTS_TW_DBG_VERIFY_EMPTY_LATER_SLOTS(tiw, min_tpos);
+            slots = ERTS_TW_LATER_WHEEL_SIZE - ((int) tmp);
+        }
+
+        slot = later_slot(tpos);
+
+        /*
+         * We never search for an exact timeout in the
+         * later wheel, but instead settle for the first
+         * scnt range used.
+         */
+        if (tiw->w[slot])
+            true_min_timeout = 1;
+        else
+            scnt_later_wheel_next(&slot, &slots, &tpos, NULL, tiw->scnt);
+
+        tiw->later.min_tpos = tpos;
+        tiw->later.min_tpos_slot = slot;
+        ERTS_TW_ASSERT(slot == later_slot(tpos));
+
+        /* Pre-timeout for move from later to soon wheel... */
+        tpos -= ERTS_TW_LATER_WHEEL_SLOT_SIZE;
+        min_timeout_pos = tpos;
+    }
+    else {
+        ErtsMonotonicTime tpos;
+        /* Search soon wheel... */
+
+        min_timeout_pos = tiw->pos + ERTS_TW_SOON_WHEEL_SIZE;
+
+        /*
+         * Besides inspecting the soon wheel we
+         * may also have to inspect two slots in the
+         * later wheel which potentially can trigger
+         * timeouts before timeouts in soon wheel...
+         */
+        if (tiw->later.min_tpos > (tiw->later.pos
+                                   + 2*ERTS_TW_LATER_WHEEL_SLOT_SIZE)) {
+            ERTS_TW_DBG_VERIFY_EMPTY_LATER_SLOTS(
+                tiw, 2*ERTS_TW_LATER_WHEEL_SLOT_SIZE);
+        }
+        else {
+            int fslot;
+            tpos = tiw->later.pos;
+            tpos -= ERTS_TW_LATER_WHEEL_SLOT_SIZE;
+            fslot = later_slot(tiw->later.pos);
+            if (tiw->w[fslot])
+                min_timeout_pos = tpos;
+            else {
+                tpos += ERTS_TW_LATER_WHEEL_SLOT_SIZE;
+                if (tpos < min_timeout_pos) {
+                    fslot++;
+                    if (fslot == ERTS_TW_LATER_WHEEL_END_SLOT)
+                        fslot = ERTS_TW_LATER_WHEEL_FIRST_SLOT;
+                    if (tiw->w[fslot])
+                        min_timeout_pos = tpos;
+                }
+            }
+        }
+
+        if (tiw->soon.min_tpos <= tiw->pos) {
+            tpos = tiw->pos;
+            slots = ERTS_TW_SOON_WHEEL_SIZE;
+        }
+        else {
+            ErtsMonotonicTime tmp;
+            /* Don't inspect slots we know are empty... */
+            tmp = tiw->soon.min_tpos - tiw->pos;
+            ERTS_TW_ASSERT(ERTS_TW_SOON_WHEEL_SIZE > tmp);
+            ERTS_TW_DBG_VERIFY_EMPTY_SOON_SLOTS(tiw, tiw->soon.min_tpos);
+            slots = ERTS_TW_SOON_WHEEL_SIZE - ((int) tmp);
+            tpos = tiw->soon.min_tpos;
+        }
+
+        slot = soon_slot(tpos);
+
+        /* find next non-empty slot */
+        while (tpos < min_timeout_pos) {
+            if (tiw->w[slot]) {
+                ERTS_TW_ASSERT(tiw->w[slot]->timeout_pos == tpos);
+                min_timeout_pos = tpos;
+                break;
+            }
+            scnt_soon_wheel_next(&slot, &slots, &tpos, NULL, tiw->scnt);
+        }
 
-	tiw_pos_ix++;
-	if (tiw_pos_ix == ERTS_TIW_SIZE)
-	    tiw_pos_ix = 0;
-    } while (start_ix != tiw_pos_ix);
+        tiw->soon.min_tpos = min_timeout_pos;
+        true_min_timeout = 1;
+    }
+
+done: {
+        ErtsMonotonicTime min_timeout;
 
-found_next:
+        min_timeout = ERTS_CLKTCKS_TO_MONOTONIC(min_timeout_pos);
+        tiw->next_timeout_pos = min_timeout_pos;
+        tiw->next_timeout_time = min_timeout;
+        tiw->true_next_timeout_time = true_min_timeout;
 
-    min_timeout = ERTS_CLKTCKS_TO_MONOTONIC(min_timeout_pos);
-    tiw->next_timeout_time = min_timeout;
-    tiw->true_next_timeout_time = true_min_timeout;
+        ERTS_HARD_DBG_CHK_WHEELS(tiw, 1);
 
-    return min_timeout;
+        return min_timeout;
+    }
 }
 
 static ERTS_INLINE void
 insert_timer_into_slot(ErtsTimerWheel *tiw, int slot, ErtsTWheelTimer *p)
 {
-    ERTS_TW_ASSERT(slot >= 0);
-    ERTS_TW_ASSERT(slot < ERTS_TIW_SIZE);
+    ERTS_TW_ASSERT(ERTS_TW_SLOT_AT_ONCE <= slot
+                   && slot < ERTS_TW_LATER_WHEEL_END_SLOT);
     p->slot = slot;
     if (!tiw->w[slot]) {
 	tiw->w[slot] = p;
@@ -223,55 +641,89 @@ insert_timer_into_slot(ErtsTimerWheel *tiw, int slot, ErtsTWheelTimer *p)
 	prev->next = p;
 	next->prev = p;
     }
+    if (slot == ERTS_TW_SLOT_AT_ONCE)
+	tiw->at_once.nto++;
+    else {
+        ErtsMonotonicTime tpos = p->timeout_pos;
+        if (slot < ERTS_TW_SOON_WHEEL_END_SLOT) {
+            ERTS_TW_ASSERT(p->timeout_pos < tiw->pos + ERTS_TW_SOON_WHEEL_SIZE);
+            tiw->soon.nto++;
+            if (tiw->soon.min_tpos > tpos)
+                tiw->soon.min_tpos = tpos;
+        }
+        else {
+            ERTS_TW_ASSERT(p->timeout_pos >= tiw->pos + ERTS_TW_SOON_WHEEL_SIZE);
+            tiw->later.nto++;
+            if (tiw->later.min_tpos > tpos) {
+                tiw->later.min_tpos = tpos;
+                tiw->later.min_tpos_slot = slot;
+            }
+        }
+        scnt_inc(tiw->scnt, slot);
+    }
 }
 
 static ERTS_INLINE void
 remove_timer(ErtsTimerWheel *tiw, ErtsTWheelTimer *p)
 {
     int slot = p->slot;
-    ERTS_TW_ASSERT(slot != ERTS_TWHEEL_SLOT_INACTIVE);
-
-    if (slot >= 0) {
-	/*
-	 * Timer in wheel or in circular
-	 * list of timers currently beeing
-	 * triggered (referred by sentinel).
-	 */
-	ERTS_TW_ASSERT(slot < ERTS_TIW_SIZE);
-
-	if (p->next == p) {
-	    ERTS_TW_ASSERT(tiw->w[slot] == p);
-	    tiw->w[slot] = NULL;
-	}
-	else {
-	    if (tiw->w[slot] == p)
-		tiw->w[slot] = p->next;
-	    p->prev->next = p->next;
-	    p->next->prev = p->prev;
-	}
+    int empty_slot;
+    ERTS_TW_ASSERT(slot != ERTS_TW_SLOT_INACTIVE);
+
+    /*
+     * Timer is in circular list either referred to
+     * by at once slot, slot in soon wheel, slot
+     * in later wheel, or by sentinel (timers currently
+     * being triggered).
+     */
+    ERTS_TW_ASSERT(ERTS_TW_SLOT_AT_ONCE <= slot
+                   && slot < ERTS_TW_LATER_WHEEL_END_SLOT);
+
+    if (p->next == p) {
+        /* Cannot be referred by sentinel, i.e. must be referred by slot... */
+        ERTS_TW_ASSERT(tiw->w[slot] == p);
+        tiw->w[slot] = NULL;
+        empty_slot = 1;
     }
     else {
-	/* Timer in "at once" queue... */
-	ERTS_TW_ASSERT(slot == ERTS_TWHEEL_SLOT_AT_ONCE);
-	if (p->prev)
-	    p->prev->next = p->next;
-	else {
-	    ERTS_TW_ASSERT(tiw->at_once.head == p);
-	    tiw->at_once.head = p->next;
-	}
-	if (p->next)
-	    p->next->prev = p->prev;
-	else {
-	    ERTS_TW_ASSERT(tiw->at_once.tail == p);
-	    tiw->at_once.tail = p->prev;
-	}
+        if (tiw->w[slot] == p)
+            tiw->w[slot] = p->next;
+        p->prev->next = p->next;
+        p->next->prev = p->prev;
+        empty_slot = 0;
+    }
+    if (slot == ERTS_TW_SLOT_AT_ONCE) {
 	ERTS_TW_ASSERT(tiw->at_once.nto > 0);
 	tiw->at_once.nto--;
     }
-
-    p->slot = ERTS_TWHEEL_SLOT_INACTIVE;
-
-    tiw->nto--;
+    else {
+        scnt_dec(tiw->scnt, slot);
+        if (slot < ERTS_TW_SOON_WHEEL_END_SLOT) {
+            if (empty_slot
+                && tiw->true_next_timeout_time
+                && p->timeout_pos == tiw->next_timeout_pos) {
+                tiw->true_next_timeout_time = 0;
+            }
+            if (--tiw->soon.nto == 0)
+                tiw->soon.min_tpos = ERTS_MAX_CLKTCKS;
+        }
+        else {
+            if (empty_slot
+                && tiw->true_next_timeout_time
+                && tiw->later.min_tpos_slot == slot) {
+                ErtsMonotonicTime tpos = tiw->later.min_tpos;
+                tpos &= ERTS_TW_LATER_WHEEL_POS_MASK;
+                tpos -= ERTS_TW_LATER_WHEEL_SLOT_SIZE;
+                if (tpos == tiw->next_timeout_pos)
+                    tiw->true_next_timeout_time = 0;
+            }
+            if (--tiw->later.nto == 0) {
+                tiw->later.min_tpos = ERTS_MAX_CLKTCKS;
+                tiw->later.min_tpos_slot = ERTS_TW_LATER_WHEEL_END_SLOT;
+            }
+        }
+    }
+    p->slot = ERTS_TW_SLOT_INACTIVE;
 }
 
 ErtsMonotonicTime
@@ -280,58 +732,26 @@ erts_check_next_timeout_time(ErtsSchedulerData *esdp)
     ErtsTimerWheel *tiw = esdp->timer_wheel;
     ErtsMonotonicTime time;
     ERTS_MSACC_DECLARE_CACHE_X();
+    ERTS_TW_ASSERT(tiw->next_timeout_time
+                   == ERTS_CLKTCKS_TO_MONOTONIC(tiw->next_timeout_pos));
     if (tiw->true_next_timeout_time)
-	return tiw->next_timeout_time;
+	return tiw->next_timeout_time; /* known timeout... */
+    if (tiw->next_timeout_pos > tiw->pos + ERTS_TW_SOON_WHEEL_SIZE)
+        return tiw->next_timeout_time; /* sufficiently later away... */
     ERTS_MSACC_PUSH_AND_SET_STATE_CACHED_X(ERTS_MSACC_STATE_TIMERS);
-    time = find_next_timeout(esdp, tiw, 1, 0, 0);
+    time = find_next_timeout(esdp, tiw);
     ERTS_MSACC_POP_STATE_M_X();
     return time;
 }
 
-#ifndef ERTS_TW_DEBUG
-#define ERTS_DBG_CHK_SAFE_TO_SKIP_TO(TIW, TO) ((void) 0)
-#else
-#define ERTS_DBG_CHK_SAFE_TO_SKIP_TO(TIW, TO) debug_check_safe_to_skip_to((TIW), (TO))
-static void
-debug_check_safe_to_skip_to(ErtsTimerWheel *tiw, ErtsMonotonicTime skip_to_pos)
-{
-    int slots, ix;
-    ErtsTWheelTimer *tmr;
-    ErtsMonotonicTime tmp;
-
-    ix = (int) (tiw->pos & (ERTS_TIW_SIZE-1));
-    tmp = skip_to_pos - tiw->pos;
-    ERTS_TW_ASSERT(tmp >= 0);
-    if (tmp < (ErtsMonotonicTime) ERTS_TIW_SIZE)
-	slots = (int) tmp;
-    else
-	slots = ERTS_TIW_SIZE;
-
-     while (slots > 0) {
-	 tmr = tiw->w[ix];
-	 if (tmr) {
-	     ErtsTWheelTimer *end = tmr;
-	     do {
-		 ERTS_TW_ASSERT(tmr->timeout_pos > skip_to_pos);
-		 tmr = tmr->next;
-	     } while (tmr != end);
-	 }
-	 ix++;
-	 if (ix == ERTS_TIW_SIZE)
-	     ix = 0;
-	 slots--;
-    }
-}
-#endif
-
 static ERTS_INLINE void
 timeout_timer(ErtsTWheelTimer *p)
 {
     ErlTimeoutProc timeout;
     void *arg;
-    p->slot = ERTS_TWHEEL_SLOT_INACTIVE;
-    timeout = p->u.func.timeout;
-    arg = p->u.func.arg;
+    p->slot = ERTS_TW_SLOT_INACTIVE;
+    timeout = p->timeout;
+    arg = p->arg;
     (*timeout)(arg);
     ASSERT_NO_LOCKED_LOCKS;
 }
@@ -339,73 +759,108 @@ timeout_timer(ErtsTWheelTimer *p)
 void
 erts_bump_timers(ErtsTimerWheel *tiw, ErtsMonotonicTime curr_time)
 {
-    int tiw_pos_ix, slots, yielded_slot_restarted, yield_count;
-    ErtsMonotonicTime bump_to, tmp_slots, old_pos;
+    int slot, restarted, yield_count, slots, scnt_ix;
+    ErtsMonotonicTime bump_to;
+    Sint *scnt, *bump_scnt;
     ERTS_MSACC_PUSH_AND_SET_STATE_M_X(ERTS_MSACC_STATE_TIMERS);
 
     yield_count = ERTS_TWHEEL_BUMP_YIELD_LIMIT;
 
+    scnt = &tiw->scnt[0];
+    bump_scnt = &tiw->bump_scnt[0];
+
     /*
      * In order to be fair we always continue with work
      * where we left off when restarting after a yield.
      */
 
-    if (tiw->yield_slot >= 0) {
-	yielded_slot_restarted = 1;
-	tiw_pos_ix = tiw->yield_slot;
-	slots = tiw->yield_slots_left;
+    slot = tiw->yield_slot;
+    restarted = slot != ERTS_TW_SLOT_INACTIVE;
+    if (restarted) {
 	bump_to = tiw->pos;
-	old_pos = tiw->yield_start_pos;
-	goto restart_yielded_slot;
+        if (slot >= ERTS_TW_LATER_WHEEL_FIRST_SLOT)
+            goto restart_yielded_later_slot;
+        tiw->yield_slot = ERTS_TW_SLOT_INACTIVE;
+        if (slot == ERTS_TW_SLOT_AT_ONCE)
+            goto restart_yielded_at_once_slot;
+        scnt_ix = scnt_get_ix(slot);
+	slots = tiw->yield_slots_left;
+        ASSERT(0 <= slots && slots <= ERTS_TW_SOON_WHEEL_SIZE);
+        goto restart_yielded_soon_slot;
     }
 
     do {
 
-	yielded_slot_restarted = 0;
-
+        restarted = 0;
 	bump_to = ERTS_MONOTONIC_TO_CLKTCKS(curr_time);
+        tiw->true_next_timeout_time = 1;
+        tiw->next_timeout_pos = bump_to;
+        tiw->next_timeout_time = ERTS_CLKTCKS_TO_MONOTONIC(bump_to);
 
 	while (1) {
 	    ErtsTWheelTimer *p;
 
-	    old_pos = tiw->pos;
-
 	    if (tiw->nto == 0) {
 	    empty_wheel:
-		ERTS_DBG_CHK_SAFE_TO_SKIP_TO(tiw, bump_to);
+                ERTS_TW_DBG_VERIFY_EMPTY_SOON_SLOTS(tiw, bump_to);
+                ERTS_TW_DBG_VERIFY_EMPTY_LATER_SLOTS(tiw, bump_to);
 		tiw->true_next_timeout_time = 0;
-		tiw->next_timeout_time = curr_time + ERTS_MONOTONIC_DAY;
+                tiw->next_timeout_pos = bump_to + ERTS_CLKTCKS_WEEK;
+		tiw->next_timeout_time = ERTS_CLKTCKS_TO_MONOTONIC(tiw->next_timeout_pos);;
 		tiw->pos = bump_to;
-		tiw->yield_slot = ERTS_TWHEEL_SLOT_INACTIVE;
+                tiw->later.pos = bump_to + ERTS_TW_SOON_WHEEL_SIZE;
+                tiw->later.pos &= ERTS_TW_LATER_WHEEL_POS_MASK;
+		tiw->yield_slot = ERTS_TW_SLOT_INACTIVE;
                 ERTS_MSACC_POP_STATE_M_X();
 		return;
 	    }
 
-	    p = tiw->at_once.head;
-	    while (p) {
-		if (--yield_count <= 0) {
-		    ERTS_TW_ASSERT(tiw->nto > 0);
-		    ERTS_TW_ASSERT(tiw->at_once.nto > 0);
-		    tiw->yield_slot = ERTS_TWHEEL_SLOT_AT_ONCE;
-		    tiw->true_next_timeout_time = 1;
-		    tiw->next_timeout_time = ERTS_CLKTCKS_TO_MONOTONIC(old_pos);
-                    ERTS_MSACC_POP_STATE_M_X();
-		    return;
-		}
+	    p = tiw->w[ERTS_TW_SLOT_AT_ONCE];
+
+	    if (p) {
+
+                if (p->next == p) {
+                    ERTS_TW_ASSERT(tiw->sentinel.next == &tiw->sentinel);
+                    ERTS_TW_ASSERT(tiw->sentinel.prev == &tiw->sentinel);
+                }
+                else {
+                    tiw->sentinel.next = p->next;
+                    tiw->sentinel.prev = p->prev;
+                    tiw->sentinel.next->prev = &tiw->sentinel;
+                    tiw->sentinel.prev->next = &tiw->sentinel;
+                }
+                tiw->w[ERTS_TW_SLOT_AT_ONCE] = NULL;
+
+                while (1) {
+                    ERTS_TW_ASSERT(tiw->nto > 0);
+                    ERTS_TW_ASSERT(tiw->at_once.nto > 0);
+                    tiw->nto--;
+                    tiw->at_once.nto--;
+
+                    timeout_timer(p);
+
+                    yield_count -= ERTS_TW_COST_TIMEOUT;
 
-		ERTS_TW_ASSERT(tiw->nto > 0);
-		ERTS_TW_ASSERT(tiw->at_once.nto > 0);
-		tiw->nto--;
-		tiw->at_once.nto--;
-		tiw->at_once.head = p->next;
-		if (p->next)
-		    p->next->prev = NULL;
-		else
-		    tiw->at_once.tail = NULL;
+                restart_yielded_at_once_slot:
 
-		timeout_timer(p);
+                    p = tiw->sentinel.next;
+                    if (p == &tiw->sentinel) {
+                        ERTS_TW_ASSERT(tiw->sentinel.prev == &tiw->sentinel);
+                        break;
+                    }
+
+                    if (yield_count <= 0) {
+                        ERTS_TW_ASSERT(tiw->nto > 0);
+                        ERTS_TW_ASSERT(tiw->at_once.nto > 0);
+                        tiw->yield_slot = ERTS_TW_SLOT_AT_ONCE;
+                        ERTS_MSACC_POP_STATE_M_X();
+                        return; /* Yield! */
+                    }
+
+                    tiw->sentinel.next = p->next;
+                    p->next->prev = &tiw->sentinel;
+                }
 
-		p = tiw->at_once.head;
 	    }
 
 	    if (tiw->pos >= bump_to) {
@@ -416,39 +871,66 @@ erts_bump_timers(ErtsTimerWheel *tiw, ErtsMonotonicTime curr_time)
 	    if (tiw->nto == 0)
 		goto empty_wheel;
 
-	    if (tiw->true_next_timeout_time) {
-		ErtsMonotonicTime skip_until_pos;
+            /*
+             * Save slot counts in bump operation local
+             * array.
+             *
+             * The amount of timers to trigger (or move)
+             * will only decrease from now until we have
+             * completed this bump operation (even if we
+             * yield in the middle of it).
+             *
+             * The amount of timers in the wheels may
+             * however increase due to timers being set
+             * by timeout callbacks.
+             */
+            sys_memcpy((void *) bump_scnt, (void *) scnt,
+                       sizeof(Sint) * ERTS_TW_SCNT_SIZE);
+
+	    if (tiw->soon.min_tpos > tiw->pos) {
+		ErtsMonotonicTime skip_until_pos = tiw->soon.min_tpos;
+
 		/*
 		 * No need inspecting slots where we know no timeouts
 		 * to trigger should reside.
 		 */
 
-		skip_until_pos = ERTS_MONOTONIC_TO_CLKTCKS(tiw->next_timeout_time);
 		if (skip_until_pos > bump_to)
 		    skip_until_pos = bump_to;
 
 		skip_until_pos--;
 
 		if (skip_until_pos > tiw->pos) {
-		    ERTS_DBG_CHK_SAFE_TO_SKIP_TO(tiw, skip_until_pos);
-
+                    ERTS_TW_DBG_VERIFY_EMPTY_SOON_SLOTS(tiw, skip_until_pos);
 		    tiw->pos = skip_until_pos;
 		}
 	    }
 
-	    tiw_pos_ix = (int) ((tiw->pos+1) & (ERTS_TIW_SIZE-1));
-	    tmp_slots = (bump_to - tiw->pos);
-	    if (tmp_slots < (ErtsMonotonicTime) ERTS_TIW_SIZE)
-		slots = (int) tmp_slots;
-	    else
-		slots = ERTS_TIW_SIZE;
+            {
+                ErtsMonotonicTime tmp_slots = bump_to - tiw->pos;
+                tmp_slots = (bump_to - tiw->pos);
+                if (tmp_slots < ERTS_TW_SOON_WHEEL_SIZE)
+                    slots = (int) tmp_slots;
+                else
+                    slots = ERTS_TW_SOON_WHEEL_SIZE;
+            }
 
+            slot = soon_slot(tiw->pos+1);
 	    tiw->pos = bump_to;
 
+            tiw->next_timeout_pos = bump_to;
+            tiw->next_timeout_time = ERTS_CLKTCKS_TO_MONOTONIC(bump_to);
+
+            scnt_ix = scnt_get_ix(slot);
+
+            /* Timeout timers in soon wheel */
 	    while (slots > 0) {
 
-		p = tiw->w[tiw_pos_ix];
+                yield_count -= ERTS_TW_COST_SLOT;
+
+		p = tiw->w[slot];
 		if (p) {
+                    /* timeout callback need tiw->pos to be up to date */
 		    if (p->next == p) {
 			ERTS_TW_ASSERT(tiw->sentinel.next == &tiw->sentinel);
 			ERTS_TW_ASSERT(tiw->sentinel.prev == &tiw->sentinel);
@@ -459,22 +941,28 @@ erts_bump_timers(ErtsTimerWheel *tiw, ErtsMonotonicTime curr_time)
 			tiw->sentinel.next->prev = &tiw->sentinel;
 			tiw->sentinel.prev->next = &tiw->sentinel;
 		    }
-		    tiw->w[tiw_pos_ix] = NULL;
+		    tiw->w[slot] = NULL;
 
 		    while (1) {
 
-			if (p->timeout_pos > bump_to) {
-			    /* Very unusual case... */
-			    ++yield_count;
-			    insert_timer_into_slot(tiw, tiw_pos_ix, p);
-			}
-			else {
-			    /* Normal case... */
-			    timeout_timer(p);
-			    tiw->nto--;
-			}
-
-		    restart_yielded_slot:
+                        ERTS_TW_ASSERT(ERTS_TW_SOON_WHEEL_FIRST_SLOT <= p->slot
+                                       && p->slot < ERTS_TW_SOON_WHEEL_END_SLOT);
+                        if (--tiw->soon.nto == 0)
+                            tiw->soon.min_tpos = ERTS_MAX_CLKTCKS;
+                        scnt_ix_dec(scnt, scnt_ix);
+                        if (p->timeout_pos <= bump_to) {
+                            timeout_timer(p);
+                            tiw->nto--;
+                            scnt_ix_dec(bump_scnt, scnt_ix);
+                            yield_count -= ERTS_TW_COST_TIMEOUT;
+                        }
+                        else {
+                            /* uncommon case */
+                            insert_timer_into_slot(tiw, slot, p);
+                            yield_count -= ERTS_TW_COST_SLOT_MOVE;
+                        }
+
+		    restart_yielded_soon_slot:
 
 			p = tiw->sentinel.next;
 			if (p == &tiw->sentinel) {
@@ -482,12 +970,9 @@ erts_bump_timers(ErtsTimerWheel *tiw, ErtsMonotonicTime curr_time)
 			    break;
 			}
 
-			if (--yield_count <= 0) {
-			    tiw->true_next_timeout_time = 1;
-			    tiw->next_timeout_time = ERTS_CLKTCKS_TO_MONOTONIC(old_pos);
-			    tiw->yield_slot = tiw_pos_ix;
+			if (yield_count <= 0) {
+			    tiw->yield_slot = slot;
 			    tiw->yield_slots_left = slots;
-			    tiw->yield_start_pos = old_pos;
                             ERTS_MSACC_POP_STATE_M_X();
 			    return; /* Yield! */
 			}
@@ -496,24 +981,166 @@ erts_bump_timers(ErtsTimerWheel *tiw, ErtsMonotonicTime curr_time)
 			p->next->prev = &tiw->sentinel;
 		    }
 		}
-		tiw_pos_ix++;
-		if (tiw_pos_ix == ERTS_TIW_SIZE)
-		    tiw_pos_ix = 0;
-		slots--;
+
+                scnt_soon_wheel_next(&slot, &slots, NULL, &scnt_ix, bump_scnt);
 	    }
+
+            if (ERTS_TW_BUMP_LATER_WHEEL(tiw)) {
+            restart_yielded_later_slot:
+                if (bump_later_wheel(tiw, &yield_count))
+                    return; /* Yield! */
+            }
 	}
 
-    } while (yielded_slot_restarted);
+    } while (restarted);
 
-    tiw->yield_slot = ERTS_TWHEEL_SLOT_INACTIVE;
     tiw->true_next_timeout_time = 0;
-    tiw->next_timeout_time = curr_time + ERTS_MONOTONIC_DAY;
+    ERTS_TW_ASSERT(tiw->next_timeout_pos == bump_to);
 
-    /* Search at most two seconds ahead... */
-    (void) find_next_timeout(NULL, tiw, 0, curr_time, ERTS_SEC_TO_MONOTONIC(2));
+    (void) find_next_timeout(NULL, tiw);
     ERTS_MSACC_POP_STATE_M_X();
 }
 
+static int
+bump_later_wheel(ErtsTimerWheel *tiw, int *ycount_p)
+{
+    ErtsMonotonicTime cpos = tiw->pos;
+    ErtsMonotonicTime later_pos = tiw->later.pos;
+    int ycount = *ycount_p;
+    int slots, fslot, scnt_ix;
+    Sint *scnt, *bump_scnt;
+
+    scnt = &tiw->scnt[0];
+    bump_scnt = &tiw->bump_scnt[0];
+
+    ERTS_HARD_DBG_CHK_WHEELS(tiw, 0);
+
+    if (tiw->yield_slot >= ERTS_TW_LATER_WHEEL_FIRST_SLOT) {
+        fslot = tiw->yield_slot;
+        scnt_ix = scnt_get_ix(fslot);
+        slots = tiw->yield_slots_left;
+        ASSERT(0 <= slots && slots <= ERTS_TW_LATER_WHEEL_SIZE);
+        tiw->yield_slot = ERTS_TW_SLOT_INACTIVE;
+        goto restart_yielded_slot;
+    }
+    else {
+        ErtsMonotonicTime end_later_pos, tmp_slots, min_tpos;
+
+        min_tpos = tiw->later.min_tpos & ERTS_TW_LATER_WHEEL_POS_MASK;
+        end_later_pos = cpos + ERTS_TW_SOON_WHEEL_SIZE;
+        end_later_pos &= ERTS_TW_LATER_WHEEL_POS_MASK;
+
+        /* Skip known empty slots... */
+        if (min_tpos > later_pos) {
+            if (min_tpos > end_later_pos) {
+                ERTS_TW_DBG_VERIFY_EMPTY_LATER_SLOTS(tiw, end_later_pos);
+                tiw->later.pos = end_later_pos;
+                goto done;
+            }
+            later_pos = min_tpos;
+            ERTS_TW_DBG_VERIFY_EMPTY_LATER_SLOTS(tiw, later_pos);
+        }
+
+        tmp_slots = end_later_pos;
+        tmp_slots -= later_pos;
+        tmp_slots /= ERTS_TW_LATER_WHEEL_SLOT_SIZE;
+        if (tmp_slots < ERTS_TW_LATER_WHEEL_SIZE)
+            slots = (int) tmp_slots;
+        else
+            slots = ERTS_TW_LATER_WHEEL_SIZE;
+
+        fslot = later_slot(later_pos);
+        scnt_ix = scnt_get_ix(fslot);
+
+        tiw->later.pos = end_later_pos;
+    }
+
+    while (slots > 0) {
+        ErtsTWheelTimer *p;
+
+        ycount -= ERTS_TW_COST_SLOT;
+
+        p = tiw->w[fslot];
+
+        if (p) {
+
+            if (p->next == p) {
+                ERTS_TW_ASSERT(tiw->sentinel.next == &tiw->sentinel);
+                ERTS_TW_ASSERT(tiw->sentinel.prev == &tiw->sentinel);
+            }
+            else {
+                tiw->sentinel.next = p->next;
+                tiw->sentinel.prev = p->prev;
+                tiw->sentinel.next->prev = &tiw->sentinel;
+                tiw->sentinel.prev->next = &tiw->sentinel;
+            }
+            tiw->w[fslot] = NULL;
+
+            while (1) {
+                ErtsMonotonicTime tpos = p->timeout_pos;
+
+                ERTS_TW_ASSERT(p->slot == fslot);
+
+                if (--tiw->later.nto == 0) {
+                    tiw->later.min_tpos = ERTS_MAX_CLKTCKS;
+                    tiw->later.min_tpos_slot = ERTS_TW_LATER_WHEEL_END_SLOT;
+                }
+                scnt_ix_dec(scnt, scnt_ix);
+
+                if (tpos >= tiw->later.pos + ERTS_TW_LATER_WHEEL_SLOT_SIZE) {
+                    /* keep in later slot; very uncommon... */
+                    insert_timer_into_slot(tiw, fslot, p);
+                    ycount -= ERTS_TW_COST_SLOT_MOVE;
+                }
+                else {
+                    scnt_ix_dec(bump_scnt, scnt_ix);
+                    ERTS_TW_ASSERT(tpos < cpos + ERTS_TW_SOON_WHEEL_SIZE);
+                    if (tpos > cpos) {
+                        /* move into soon wheel */
+                        insert_timer_into_slot(tiw, soon_slot(tpos), p);
+                        ycount -= ERTS_TW_COST_SLOT_MOVE;
+                    }
+                    else {
+                        /* trigger at once */
+                        timeout_timer(p);
+                        tiw->nto--;
+                        ycount -= ERTS_TW_COST_TIMEOUT;
+                    }
+                }
+
+            restart_yielded_slot:
+
+                p = tiw->sentinel.next;
+                if (p == &tiw->sentinel) {
+                    ERTS_TW_ASSERT(tiw->sentinel.prev == &tiw->sentinel);
+                    break;
+                }
+
+                if (ycount < 0) {
+                    tiw->yield_slot = fslot;
+                    tiw->yield_slots_left = slots;
+                    *ycount_p = 0;
+                    ERTS_HARD_DBG_CHK_WHEELS(tiw, 0);
+                    return 1; /* Yield! */
+                }
+
+                tiw->sentinel.next = p->next;
+                p->next->prev = &tiw->sentinel;
+            }
+        }
+
+        scnt_later_wheel_next(&fslot, &slots, NULL, &scnt_ix, bump_scnt);
+    }
+
+done:
+
+    ERTS_HARD_DBG_CHK_WHEELS(tiw, 0);
+
+    *ycount_p = ycount;
+
+    return 0;
+}
+
 Uint
 erts_timer_wheel_memory_size(void)
 {
@@ -526,25 +1153,51 @@ erts_create_timer_wheel(ErtsSchedulerData *esdp)
     ErtsMonotonicTime mtime;
     int i;
     ErtsTimerWheel *tiw;
+
+    /* Some compile time sanity checks... */
+
+    /* Slots... */
+    ERTS_CT_ASSERT(ERTS_TW_SLOT_AT_ONCE == -1);
+    ERTS_CT_ASSERT(ERTS_TW_SLOT_INACTIVE < ERTS_TW_SLOT_AT_ONCE);
+    ERTS_CT_ASSERT(ERTS_TW_SLOT_AT_ONCE + 1 == ERTS_TW_SOON_WHEEL_FIRST_SLOT);
+    ERTS_CT_ASSERT(ERTS_TW_SOON_WHEEL_FIRST_SLOT < ERTS_TW_SOON_WHEEL_END_SLOT);
+    ERTS_CT_ASSERT(ERTS_TW_SOON_WHEEL_END_SLOT == ERTS_TW_LATER_WHEEL_FIRST_SLOT);    
+    ERTS_CT_ASSERT(ERTS_TW_LATER_WHEEL_FIRST_SLOT < ERTS_TW_LATER_WHEEL_END_SLOT);
+
+    /* Both wheel sizes should be a powers of 2 */
+    ERTS_CT_ASSERT(ERTS_TW_SOON_WHEEL_SIZE
+                   && !(ERTS_TW_SOON_WHEEL_SIZE & (ERTS_TW_SOON_WHEEL_SIZE-1)));
+    ERTS_CT_ASSERT(ERTS_TW_LATER_WHEEL_SIZE
+                   && !(ERTS_TW_LATER_WHEEL_SIZE & (ERTS_TW_LATER_WHEEL_SIZE-1)));
+
     tiw = erts_alloc_permanent_cache_aligned(ERTS_ALC_T_TIMER_WHEEL,
 					     sizeof(ErtsTimerWheel));
-    for(i = 0; i < ERTS_TIW_SIZE; i++)
+    tiw->w = &tiw->slots[1];
+    for(i = ERTS_TW_SLOT_AT_ONCE; i < ERTS_TW_LATER_WHEEL_END_SLOT; i++)
 	tiw->w[i] = NULL;
 
+    for (i = 0; i < ERTS_TW_SCNT_SIZE; i++)
+        tiw->scnt[i] = 0;
+
     mtime = erts_get_monotonic_time(esdp);
     tiw->pos = ERTS_MONOTONIC_TO_CLKTCKS(mtime);
     tiw->nto = 0;
-    tiw->at_once.head = NULL;
-    tiw->at_once.tail = NULL;
     tiw->at_once.nto = 0;
-    tiw->yield_slot = ERTS_TWHEEL_SLOT_INACTIVE;
+    tiw->soon.min_tpos = ERTS_MAX_CLKTCKS;
+    tiw->soon.nto = 0;
+    tiw->later.min_tpos = ERTS_MAX_CLKTCKS;
+    tiw->later.min_tpos_slot = ERTS_TW_LATER_WHEEL_END_SLOT;
+    tiw->later.pos = tiw->pos + ERTS_TW_SOON_WHEEL_SIZE;
+    tiw->later.pos &= ERTS_TW_LATER_WHEEL_POS_MASK;
+    tiw->later.nto = 0;
+    tiw->yield_slot = ERTS_TW_SLOT_INACTIVE;
     tiw->true_next_timeout_time = 0;
-    tiw->next_timeout_time = mtime + ERTS_MONOTONIC_DAY;
+    tiw->next_timeout_pos = tiw->pos + ERTS_CLKTCKS_WEEK;
+    tiw->next_timeout_time = ERTS_CLKTCKS_TO_MONOTONIC(tiw->next_timeout_pos);
     tiw->sentinel.next = &tiw->sentinel;
     tiw->sentinel.prev = &tiw->sentinel;
-    tiw->sentinel.u.func.timeout = NULL;
-    tiw->sentinel.u.func.cancel = NULL;
-    tiw->sentinel.u.func.arg = NULL;
+    tiw->sentinel.timeout = NULL;
+    tiw->sentinel.arg = NULL;
     return tiw;
 }
 
@@ -577,53 +1230,56 @@ erts_init_time(int time_correction, ErtsTimeWarpMode time_warp_mode)
 void
 erts_twheel_set_timer(ErtsTimerWheel *tiw,
 		      ErtsTWheelTimer *p, ErlTimeoutProc timeout,
-		      ErlCancelProc cancel, void *arg,
-		      ErtsMonotonicTime timeout_pos)
+		      void *arg, ErtsMonotonicTime timeout_pos)
 {
-    ErtsMonotonicTime timeout_time;
+    int slot;
     ERTS_MSACC_PUSH_AND_SET_STATE_M_X(ERTS_MSACC_STATE_TIMERS);
 
-    p->u.func.timeout = timeout;
-    p->u.func.cancel = cancel;
-    p->u.func.arg = arg;
+    p->timeout = timeout;
+    p->arg = arg;
+
+    ERTS_TW_ASSERT(p->slot == ERTS_TW_SLOT_INACTIVE);
 
-    ERTS_TW_ASSERT(p->slot == ERTS_TWHEEL_SLOT_INACTIVE);
+    tiw->nto++;
 
+    /* calculate slot */
     if (timeout_pos <= tiw->pos) {
-	tiw->nto++;
-	tiw->at_once.nto++;
-	p->next = NULL;
-	p->prev = tiw->at_once.tail;
-	if (tiw->at_once.tail) {
-	    ERTS_TW_ASSERT(tiw->at_once.head);
-	    tiw->at_once.tail->next = p;
-	}
-	else {	
-	    ERTS_TW_ASSERT(!tiw->at_once.head);
-	    tiw->at_once.head = p;
-	}
-	tiw->at_once.tail = p;
-	p->timeout_pos = tiw->pos;
-	p->slot = ERTS_TWHEEL_SLOT_AT_ONCE;
-	timeout_time = ERTS_CLKTCKS_TO_MONOTONIC(tiw->pos);
+        /* at once */
+        p->timeout_pos = timeout_pos = tiw->pos;
+        slot = ERTS_TW_SLOT_AT_ONCE;
+    }
+    else if (timeout_pos < tiw->pos + ERTS_TW_SOON_WHEEL_SIZE) {
+        /* soon wheel */
+        p->timeout_pos = timeout_pos;
+        slot = soon_slot(timeout_pos);
+        if (tiw->soon.min_tpos > timeout_pos)
+            tiw->soon.min_tpos = timeout_pos;
     }
     else {
-	int slot;
-
-	/* calculate slot */
-	slot = (int) (timeout_pos & (ERTS_TIW_SIZE-1));
-
-	insert_timer_into_slot(tiw, slot, p);
-
-	tiw->nto++;
-
-	timeout_time = ERTS_CLKTCKS_TO_MONOTONIC(timeout_pos);
-	p->timeout_pos = timeout_pos;
+        /* later wheel */
+        p->timeout_pos = timeout_pos;
+        slot = later_slot(timeout_pos);
+
+        /*
+         * Next timeout due to this timeout
+         * should be in good time before the
+         * actual timeout (one later wheel slot
+         * size). This, in order to move it
+         * from the later wheel to the soon
+         * wheel.
+         */
+        timeout_pos &= ERTS_TW_LATER_WHEEL_POS_MASK;
+        timeout_pos -= ERTS_TW_LATER_WHEEL_SLOT_SIZE;
     }
 
-    if (timeout_time < tiw->next_timeout_time) {
+    insert_timer_into_slot(tiw, slot, p);
+
+    if (timeout_pos <= tiw->next_timeout_pos) {
 	tiw->true_next_timeout_time = 1;
-	tiw->next_timeout_time = timeout_time;
+        if (timeout_pos < tiw->next_timeout_pos) {
+            tiw->next_timeout_pos = timeout_pos;
+            tiw->next_timeout_time = ERTS_CLKTCKS_TO_MONOTONIC(timeout_pos);
+        }
     }
     ERTS_MSACC_POP_STATE_M_X();
 }
@@ -631,15 +1287,10 @@ erts_twheel_set_timer(ErtsTimerWheel *tiw,
 void
 erts_twheel_cancel_timer(ErtsTimerWheel *tiw, ErtsTWheelTimer *p)
 {
-    if (p->slot != ERTS_TWHEEL_SLOT_INACTIVE) {
-	ErlCancelProc cancel;
-	void *arg;
+    if (p->slot != ERTS_TW_SLOT_INACTIVE) {
         ERTS_MSACC_PUSH_AND_SET_STATE_M_X(ERTS_MSACC_STATE_TIMERS);
 	remove_timer(tiw, p);
-	cancel = p->u.func.cancel;
-	arg = p->u.func.arg;
-	if (cancel)
-	    (*cancel)(arg);
+        tiw->nto--;
         ERTS_MSACC_POP_STATE_M_X();
     }
 }
@@ -657,22 +1308,17 @@ erts_twheel_debug_foreach(ErtsTimerWheel *tiw,
 
     tmr = tiw->sentinel.next;
     while (tmr != &tiw->sentinel) {
-	if (tmr->u.func.timeout == tclbk)
-	    (*func)(arg, tmr->timeout_pos, tmr->u.func.arg);
+	if (tmr->timeout == tclbk)
+	    (*func)(arg, tmr->timeout_pos, tmr->arg);
 	tmr = tmr->next;
     }
 
-    for (tmr = tiw->at_once.head; tmr; tmr = tmr->next) {
-	if (tmr->u.func.timeout == tclbk)
-	    (*func)(arg, tmr->timeout_pos, tmr->u.func.arg);
-    }
-
-    for (ix = 0; ix < ERTS_TIW_SIZE; ix++) {
+    for (ix = ERTS_TW_SLOT_AT_ONCE; ix < ERTS_TW_LATER_WHEEL_END_SLOT; ix++) {
 	tmr = tiw->w[ix];
 	if (tmr) {
 	    do {
-		if (tmr->u.func.timeout == tclbk)
-		    (*func)(arg, tmr->timeout_pos, tmr->u.func.arg);
+		if (tmr->timeout == tclbk)
+		    (*func)(arg, tmr->timeout_pos, tmr->arg);
 		tmr = tmr->next;
 	    } while (tmr != tiw->w[ix]);
 	}
@@ -680,35 +1326,198 @@ erts_twheel_debug_foreach(ErtsTimerWheel *tiw,
 }
 
 #ifdef ERTS_TW_DEBUG
-void erts_p_slpq(void)
+
+void
+dbg_verify_empty_soon_slots(ErtsTimerWheel *tiw, ErtsMonotonicTime to_pos)
 {
-    erts_printf("Not yet implemented...\n");
-#if 0
-    ErtsMonotonicTime current_time = erts_get_monotonic_time(NULL);
-    int i;
-    ErtsTWheelTimer* p;
-  
-    /* print the whole wheel, starting at the current position */
-    erts_printf("\ncurrent time = %bps tiw_pos = %d tiw_nto %d\n",
-		current_time, tiw->pos, tiw->nto);
-    i = tiw->pos;
-    if (tiw->w[i] != NULL) {
-	erts_printf("%d:\n", i);
-	for(p = tiw->w[i]; p != NULL; p = p->next) {
-	    erts_printf(" (timeout time %bps, slot %d)\n",
-			ERTS_CLKTCKS_TO_MONOTONIC(p->timeout_pos),
-			p->slot);
-	}
+    int ix;
+    ErtsMonotonicTime tmp;
+
+    ix = soon_slot(tiw->pos);
+    tmp = to_pos;
+    if (tmp > tiw->pos) {
+        int slots;
+        tmp -= tiw->pos;
+        ERTS_TW_ASSERT(tmp > 0);
+        if (tmp < (ErtsMonotonicTime) ERTS_TW_SOON_WHEEL_SIZE)
+            slots = (int) tmp;
+        else
+            slots = ERTS_TW_SOON_WHEEL_SIZE;
+
+        while (slots > 0) {
+            ERTS_TW_ASSERT(!tiw->w[ix]);
+            ix++;
+            if (ix == ERTS_TW_SOON_WHEEL_END_SLOT)
+                ix = ERTS_TW_SOON_WHEEL_FIRST_SLOT;
+            slots--;
+        }
+    }    
+}
+
+void
+dbg_verify_empty_later_slots(ErtsTimerWheel *tiw, ErtsMonotonicTime to_pos)
+{
+    int ix;
+    ErtsMonotonicTime tmp;
+
+    ix = later_slot(tiw->later.pos);
+    tmp = to_pos;
+    tmp &= ERTS_TW_LATER_WHEEL_POS_MASK;
+    if (tmp > tiw->later.pos) {
+        int slots;
+        tmp -= tiw->later.pos;
+        tmp /= ERTS_TW_LATER_WHEEL_SLOT_SIZE;
+        ERTS_TW_ASSERT(tmp > 0);
+        if (tmp < (ErtsMonotonicTime) ERTS_TW_LATER_WHEEL_SIZE)
+            slots = (int) tmp;
+        else
+            slots = ERTS_TW_LATER_WHEEL_SIZE;
+
+        while (slots > 0) {
+            ErtsTWheelTimer *tmr = tiw->w[ix];
+            if (tmr) {
+                ErtsTWheelTimer *end = tmr;
+                do {
+                    ERTS_TW_ASSERT(tmr->timeout_pos > to_pos);
+                    tmr = tmr->next;
+                } while (tmr != end);
+            }
+            ix++;
+            if (ix == ERTS_TW_LATER_WHEEL_END_SLOT)
+                ix = ERTS_TW_LATER_WHEEL_FIRST_SLOT;
+            slots--;
+        }
+    }    
+}
+
+#endif /* ERTS_TW_DEBUG */
+
+#ifdef ERTS_TW_HARD_DEBUG
+
+static void
+hrd_dbg_check_wheels(ErtsTimerWheel *tiw, int check_min_tpos)
+{
+    int ix, six, soon_tmo, later_tmo, at_once_tmo,
+        scnt_slot, scnt_slots, scnt_six;
+    ErtsMonotonicTime min_tpos;
+    Sint scnt[ERTS_TW_SCNT_SIZE];
+    ErtsTWheelTimer *p;
+
+    for (six = 0; six < ERTS_TW_SCNT_SIZE; six++)
+        scnt[six] = 0;
+
+    min_tpos = ERTS_MONOTONIC_TO_CLKTCKS(tiw->next_timeout_time);
+
+    at_once_tmo = 0;
+    p = tiw->w[ERTS_TW_SLOT_AT_ONCE];
+    if (p) {
+        ErtsTWheelTimer *first = p;
+        do {
+            at_once_tmo++;
+            ERTS_TW_ASSERT(p->slot == ERTS_TW_SLOT_AT_ONCE);
+            ERTS_TW_ASSERT(p->timeout_pos <= tiw->pos);
+            ERTS_TW_ASSERT(!check_min_tpos || tiw->pos >= min_tpos);
+            ERTS_TW_ASSERT(p->next->prev == p);
+            p = p->next;
+        } while (p != first);
     }
-    for(i = ((i+1) & (ERTS_TIW_SIZE-1)); i != (tiw->pos & (ERTS_TIW_SIZE-1)); i = ((i+1) & (ERTS_TIW_SIZE-1))) {
-	if (tiw->w[i] != NULL) {
-	    erts_printf("%d:\n", i);
-	    for(p = tiw->w[i]; p != NULL; p = p->next) {
-		erts_printf(" (timeout time %bps, slot %d)\n",
-			    ERTS_CLKTCKS_TO_MONOTONIC(p->timeout_pos), p->slot);
-	    }
-	}
+
+    soon_tmo = 0;
+    scnt_slot = ERTS_TW_SOON_WHEEL_END_SLOT-1;
+    scnt_slots = ERTS_TW_SOON_WHEEL_SIZE;
+    scnt_six = 0;
+    scnt_soon_wheel_next(&scnt_slot, &scnt_slots,
+                         NULL, &scnt_six, tiw->scnt);
+    for (ix = ERTS_TW_SOON_WHEEL_FIRST_SLOT;
+         ix < ERTS_TW_SOON_WHEEL_END_SLOT;
+         ix++) {
+        p = tiw->w[ix];
+        six = scnt_get_ix(ix);
+        ERTS_TW_ASSERT(!p || six == scnt_six);
+        if (p) {
+            ErtsTWheelTimer *first = p;
+            do {
+                ErtsMonotonicTime tpos = p->timeout_pos;
+                soon_tmo++;
+                scnt_ix_inc(scnt, six);
+                ERTS_TW_ASSERT(p->slot == ix);
+                ERTS_TW_ASSERT(ix == soon_slot(tpos));
+                ERTS_TW_ASSERT(p->timeout_pos < tiw->pos + ERTS_TW_SOON_WHEEL_SIZE);
+                ERTS_TW_ASSERT(!check_min_tpos || tpos >= min_tpos);
+                ERTS_TW_ASSERT(p->next->prev == p);
+                p = p->next;
+            } while (p != first);
+        }
+        if (ix == scnt_slot)
+            scnt_soon_wheel_next(&scnt_slot, &scnt_slots,
+                                 NULL, &scnt_six, tiw->scnt);
     }
-#endif
+
+    later_tmo = 0;
+    scnt_slot = ERTS_TW_SOON_WHEEL_END_SLOT-1;
+    scnt_slots = ERTS_TW_SOON_WHEEL_SIZE;
+    scnt_six = 0;
+    scnt_later_wheel_next(&scnt_slot, &scnt_slots,
+                         NULL, &scnt_six, tiw->scnt);
+    for (ix = ERTS_TW_LATER_WHEEL_FIRST_SLOT;
+         ix < ERTS_TW_LATER_WHEEL_END_SLOT;
+         ix++) {
+        p = tiw->w[ix];
+        six = scnt_get_ix(ix);
+        ERTS_TW_ASSERT(!p || six == scnt_six);
+        if (p) {
+            ErtsTWheelTimer *first = p;
+            six = scnt_get_ix(ix);
+            do {
+                ErtsMonotonicTime tpos = p->timeout_pos;
+                later_tmo++;
+                scnt_ix_inc(scnt, six);
+                ERTS_TW_ASSERT(p->slot == ix);
+                ERTS_TW_ASSERT(later_slot(tpos) == ix);
+                tpos &= ERTS_TW_LATER_WHEEL_POS_MASK;
+                tpos -= ERTS_TW_LATER_WHEEL_SLOT_SIZE;
+                ERTS_TW_ASSERT(!check_min_tpos || tpos >= min_tpos);
+                ERTS_TW_ASSERT(p->next->prev == p);
+                p = p->next;
+            } while (p != first);
+        }
+        if (ix == scnt_slot)
+            scnt_later_wheel_next(&scnt_slot, &scnt_slots,
+                                NULL, &scnt_six, tiw->scnt);
+    }
+
+    if (tiw->yield_slot != ERTS_TW_SLOT_INACTIVE) {
+        p = tiw->sentinel.next;
+        ix = tiw->yield_slot;
+        while (p != &tiw->sentinel) {
+            ErtsMonotonicTime tpos = p->timeout_pos;
+            ERTS_TW_ASSERT(ix == p->slot);
+            if (ix == ERTS_TW_SLOT_AT_ONCE)
+                at_once_tmo++;
+            else {
+                scnt_inc(scnt, ix);
+                if (ix >= ERTS_TW_LATER_WHEEL_FIRST_SLOT) {
+                    later_tmo++;
+                    ERTS_TW_ASSERT(ix == later_slot(tpos));
+                }
+                else {
+                    soon_tmo++;
+                    ERTS_TW_ASSERT(ix == (tpos & ERTS_TW_SOON_WHEEL_MASK));
+                    ERTS_TW_ASSERT(tpos < tiw->pos + ERTS_TW_SOON_WHEEL_SIZE);
+                }
+                p = p->next;
+            }
+        }
+    }
+
+
+    ERTS_TW_ASSERT(tiw->at_once.nto == at_once_tmo);
+    ERTS_TW_ASSERT(tiw->soon.nto == soon_tmo);
+    ERTS_TW_ASSERT(tiw->later.nto == later_tmo);
+    ERTS_TW_ASSERT(tiw->nto == soon_tmo + later_tmo + at_once_tmo);
+
+    for (six = 0; six < ERTS_TW_SCNT_SIZE; six++)
+        ERTS_TW_ASSERT(scnt[six] == tiw->scnt[six]);
 }
-#endif /* ERTS_TW_DEBUG */
+
+#endif /* ERTS_TW_HARD_DEBUG */