Merge branch 'bjorn/erts/breakpoints/OTP-10122'

* bjorn/erts/breakpoints/OTP-10122:
  Strengthen trace test cases
  Don't go to single-scheduler mode when managing breakpoints
  Refactor the code for managing breakpoints
  Change the data structures for breakpoints
  Add trace_local_SUITE:concurrency/1
  Make the lock check in erts_is_code_ix_locked() stronger

19 files changed, 1729 insertions, 1699 deletions

diff --git a/erts/emulator/beam/beam_bif_load.c b/erts/emulator/beam/beam_bif_load.c
index 9c65769b86..94f8edf165 100644
--- a/erts/emulator/beam/beam_bif_load.c
+++ b/erts/emulator/beam/beam_bif_load.c
@@ -672,11 +672,11 @@ set_default_trace_pattern(Eterm module)
     if (trace_pattern_is_on) {
 	Eterm mfa[1];
 	mfa[0] = module;
-	(void) erts_set_trace_pattern(mfa, 1,
+	(void) erts_set_trace_pattern(0, mfa, 1,
 				      match_spec,
 				      meta_match_spec,
 				      1, trace_pattern_flags,
-				      meta_tracer_pid);
+				      meta_tracer_pid, 1);
     }
 }
 
@@ -1006,12 +1006,11 @@ delete_code(Module* modp)
 		if (ep->code[3] == (BeamInstr) em_apply_bif) {
 		    continue;
 		}
-		else if (ep->code[3] == (BeamInstr) em_call_traced_function) {
+		else if (ep->code[3] ==
+			 (BeamInstr) BeamOp(op_i_generic_breakpoint)) {
 		    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
 		    ASSERT(modp->curr.num_traced_exports > 0);
-		    --modp->curr.num_traced_exports;
-		    MatchSetUnref(ep->match_prog_set);
-		    ep->match_prog_set = NULL;
+		    erts_clear_export_break(modp, ep->code+3);
 		}
 		else ASSERT(ep->code[3] == (BeamInstr) em_call_error_handler
 			    || !erts_initialized);
@@ -1019,7 +1018,6 @@ delete_code(Module* modp)
 	    ep->addressv[code_ix] = ep->code+3;
 	    ep->code[3] = (BeamInstr) em_call_error_handler;
 	    ep->code[4] = 0;
-	    ASSERT(ep->match_prog_set == NULL);
 	}
     }
 
diff --git a/erts/emulator/beam/beam_bp.c b/erts/emulator/beam/beam_bp.c
index 26dadfbbc0..50d18b0347 100644
--- a/erts/emulator/beam/beam_bp.c
+++ b/erts/emulator/beam/beam_bp.c
@@ -44,67 +44,34 @@
 #define ReAlloc(P, SIZ)		erts_realloc(ERTS_ALC_T_BPD, (P), (SZ))
 #define Free(P)			erts_free(ERTS_ALC_T_BPD, (P))
 
-/*
-** Doubly linked ring macros
-*/
-
-#define BpInit(a,i)        \
-do {                       \
-    (a)->orig_instr = (i); \
-    (a)->next = (a);       \
-    (a)->prev = (a);       \
-} while (0)
-
-#define BpSpliceNext(a,b)                   \
-do {                                        \
-    register BpData *c = (a), *d = (b), *e; \
-    e             = c->next->prev;          \
-    c->next->prev = d->next->prev;          \
-    d->next->prev = e;                      \
-    e       = c->next;                      \
-    c->next = d->next;                      \
-    d->next = e;                            \
-} while (0)
-
-#define BpSplicePrev(a,b)                   \
-do {                                        \
-    register BpData *c = (a), *d = (b), *e; \
-    e             = c->prev->next;          \
-    c->prev->next = d->prev->next;          \
-    d->prev->next = e;                      \
-    e       = c->prev;                      \
-    c->prev = d->prev;                      \
-    d->prev = e;                            \
-} while (0)
-
-#ifdef DEBUG
-#  define BpSingleton(a) ((a)->next == (a) && (a)->prev == (a))
+#if defined(ERTS_ENABLE_LOCK_CHECK) && defined(ERTS_SMP)
+#  define ERTS_SMP_REQ_PROC_MAIN_LOCK(P) \
+      if ((P)) erts_proc_lc_require_lock((P), ERTS_PROC_LOCK_MAIN)
+#  define ERTS_SMP_UNREQ_PROC_MAIN_LOCK(P) \
+      if ((P)) erts_proc_lc_unrequire_lock((P), ERTS_PROC_LOCK_MAIN)
 #else
-#  define BpSingleton(a) ((a)->next == (a))
+#  define ERTS_SMP_REQ_PROC_MAIN_LOCK(P)
+#  define ERTS_SMP_UNREQ_PROC_MAIN_LOCK(P)
 #endif
 
-#define BpInitAndSpliceNext(a,i,b) \
-do {                               \
-    (a)->orig_instr = (i);         \
-    (a)->prev       = (b);         \
-    (b)->next->prev = (a);         \
-    (a)->next = (b)->next;         \
-    (b)->next = (a);               \
-} while (0)
+#define ERTS_BPF_LOCAL_TRACE       0x01
+#define ERTS_BPF_META_TRACE        0x02
+#define ERTS_BPF_COUNT             0x04
+#define ERTS_BPF_COUNT_ACTIVE      0x08
+#define ERTS_BPF_DEBUG             0x10
+#define ERTS_BPF_TIME_TRACE        0x20
+#define ERTS_BPF_TIME_TRACE_ACTIVE 0x40
+#define ERTS_BPF_GLOBAL_TRACE      0x80
 
-#define BpInitAndSplicePrev(a,i,b) \
-do {                               \
-    (a)->orig_instr = (i);         \
-    (a)->next       = (b);         \
-    (b)->prev->next = (a);         \
-    (a)->prev = (b)->prev;         \
-    (b)->prev = (a);               \
-} while (0)
+#define ERTS_BPF_ALL               0xFF
 
+extern Eterm beam_return_to_trace[1];   /* OpCode(i_return_to_trace) */
+extern Eterm beam_return_trace[1];      /* OpCode(i_return_trace) */
+extern Eterm beam_exception_trace[1];   /* OpCode(i_exception_trace) */
+extern Eterm beam_return_time_trace[1]; /* OpCode(i_return_time_trace) */
 
-#define BREAK_IS_BIF (1)
-#define BREAK_IS_ERL (0)
-
+erts_smp_atomic32_t erts_active_bp_index;
+erts_smp_atomic32_t erts_staging_bp_index;
 
 /* *************************************************************************
 ** Local prototypes
@@ -113,26 +80,30 @@ do {                               \
 /*
 ** Helpers
 */
-
-static int set_break(Eterm mfa[3], int specified,
-		     Binary *match_spec, BeamInstr break_op,
-		     enum erts_break_op count_op, Eterm tracer_pid);
-static int set_module_break(Module *modp, Eterm mfa[3], int specified,
-			    Binary *match_spec, BeamInstr break_op,
-			    enum erts_break_op count_op, Eterm tracer_pid);
-static int set_function_break(Module *modp, BeamInstr *pc, int bif,
-			      Binary *match_spec, BeamInstr break_op,
-			      enum erts_break_op count_op, Eterm tracer_pid); 
-
-static int clear_break(Eterm mfa[3], int specified, 
-		       BeamInstr break_op);
-static int clear_module_break(Module *modp, Eterm mfa[3], int specified, 
-			      BeamInstr break_op);
-static int clear_function_break(Module *modp, BeamInstr *pc, int bif,
-				BeamInstr break_op);
-
-static BpData *is_break(BeamInstr *pc, BeamInstr break_op);
-static BpData *get_break(Process *p, BeamInstr *pc, BeamInstr break_op);
+static Eterm do_call_trace(Process* c_p, BeamInstr* I, Eterm* reg,
+			   int local, Binary* ms, Eterm tracer_pid);
+static void set_break(BpFunctions* f, Binary *match_spec, Uint break_flags,
+		      enum erts_break_op count_op, Eterm tracer_pid);
+static void set_function_break(BeamInstr *pc,
+			       Binary *match_spec,
+			       Uint break_flags,
+			       enum erts_break_op count_op,
+			       Eterm tracer_pid);
+
+static void clear_break(BpFunctions* f, Uint break_flags);
+static int clear_function_break(BeamInstr *pc, Uint break_flags);
+
+static BpDataTime* get_time_break(BeamInstr *pc);
+static GenericBpData* check_break(BeamInstr *pc, Uint break_flags);
+static void bp_time_diff(bp_data_time_item_t *item,
+			 process_breakpoint_time_t *pbt,
+			 Uint ms, Uint s, Uint us);
+
+static void bp_meta_unref(BpMetaPid* bmp);
+static void bp_count_unref(BpCount* bcp);
+static void bp_time_unref(BpDataTime* bdt);
+static void consolidate_bp_data(Module* modp, BeamInstr* pc, int local);
+static void uninstall_breakpoint(BeamInstr* pc);
 
 /* bp_hash */
 #define BP_TIME_ADD(pi0, pi1)                       \
@@ -152,240 +123,996 @@ static ERTS_INLINE bp_data_time_item_t * bp_hash_get(bp_time_hash_t *hash, bp_da
 static ERTS_INLINE bp_data_time_item_t * bp_hash_put(bp_time_hash_t *hash, bp_data_time_item_t *sitem);
 static void bp_hash_delete(bp_time_hash_t *hash);
 
-
 /* *************************************************************************
 ** External interfaces
 */
 
-erts_smp_spinlock_t erts_bp_lock;
-
 void 
 erts_bp_init(void) {
-    erts_smp_spinlock_init(&erts_bp_lock, "breakpoints");
+    erts_smp_atomic32_init_nob(&erts_active_bp_index, 0);
+    erts_smp_atomic32_init_nob(&erts_staging_bp_index, 1);
 }
 
-int 
-erts_set_trace_break(Eterm mfa[3], int specified, Binary *match_spec,
-		     Eterm tracer_pid) {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
-    return set_break(mfa, specified, match_spec,
-		     (BeamInstr) BeamOp(op_i_trace_breakpoint), 0, tracer_pid);
+
+void
+erts_bp_match_functions(BpFunctions* f, Eterm mfa[3], int specified)
+{
+    ErtsCodeIndex code_ix = erts_active_code_ix();
+    Uint max_funcs = 0;
+    int current;
+    int max_modules = module_code_size(code_ix);
+    int num_modules = 0;
+    Module* modp;
+    Module** module;
+    Uint i;
+
+    module = (Module **) Alloc(max_modules*sizeof(Module *));
+    num_modules = 0;
+    for (current = 0; current < max_modules; current++) {
+	modp = module_code(current, code_ix);
+	if (modp->curr.code) {
+	    max_funcs += modp->curr.code[MI_NUM_FUNCTIONS];
+	    module[num_modules++] = modp;
+	}
+    }
+
+    f->matching = (BpFunction *) Alloc(max_funcs*sizeof(BpFunction));
+    i = 0;
+    for (current = 0; current < num_modules; current++) {
+	BeamInstr** code_base = (BeamInstr **) module[current]->curr.code;
+	BeamInstr* code;
+	Uint num_functions = (Uint) code_base[MI_NUM_FUNCTIONS];
+	Uint fi;
+
+	if (specified > 0) {
+	    if (mfa[0] != make_atom(module[current]->module)) {
+		/* Wrong module name */
+		continue;
+	    }
+	}
+
+	for (fi = 0; fi < num_functions; fi++) {
+	    Eterm* pc;
+	    int wi;
+
+	    code = code_base[MI_FUNCTIONS+fi];
+	    ASSERT(code[0] == (BeamInstr) BeamOp(op_i_func_info_IaaI));
+	    pc = code+5;
+	    if (erts_is_native_break(pc)) {
+		continue;
+	    }
+	    if (is_nil(code[3])) { /* Ignore BIF stub */
+		continue;
+	    }
+	    for (wi = 0;
+		 wi < specified && (Eterm) code[2+wi] == mfa[wi];
+		 wi++) {
+		/* Empty loop body */
+	    }
+	    if (wi == specified) {
+		/* Store match */
+		f->matching[i].pc = pc;
+		f->matching[i].mod = module[current];
+		i++;
+	    }
+	}
+    }
+    f->matched = i;
+    Free(module);
 }
 
-int 
-erts_set_mtrace_break(Eterm mfa[3], int specified, Binary *match_spec,
-		      Eterm tracer_pid) {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
-    return set_break(mfa, specified, match_spec,
-		     (BeamInstr) BeamOp(op_i_mtrace_breakpoint), 0, tracer_pid);
+void
+erts_bp_match_export(BpFunctions* f, Eterm mfa[3], int specified)
+{
+    ErtsCodeIndex code_ix = erts_active_code_ix();
+    int i;
+    int num_exps = export_list_size(code_ix);
+    int ne;
+
+    f->matching = (BpFunction *) Alloc(num_exps*sizeof(BpFunction));
+    ne = 0;
+    for (i = 0; i < num_exps; i++) {
+	Export* ep = export_list(i, code_ix);
+	BeamInstr* pc;
+	int j;
+
+	for (j = 0; j < specified && mfa[j] == ep->code[j]; j++) {
+	    /* Empty loop body */
+	}
+	if (j < specified) {
+	    continue;
+	}
+	pc = ep->code+3;
+	if (ep->addressv[code_ix] == pc) {
+	    if ((*pc == (BeamInstr) em_apply_bif ||
+		 *pc == (BeamInstr) em_call_error_handler)) {
+		continue;
+	    }
+	    ASSERT(*pc == (BeamInstr) BeamOp(op_i_generic_breakpoint));
+	} else if (erts_is_native_break(ep->addressv[code_ix])) {
+	    continue;
+	}
+
+	f->matching[ne].pc = pc;
+	f->matching[ne].mod = erts_get_module(ep->code[0], code_ix);
+	ne++;
+
+    }
+    f->matched = ne;
 }
 
-/* set breakpoint data for on exported bif entry */
+void
+erts_bp_free_matched_functions(BpFunctions* f)
+{
+    Free(f->matching);
+}
 
 void
-erts_set_mtrace_bif(BeamInstr *pc, Binary *match_spec, Eterm tracer_pid) {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
-    set_function_break(NULL, pc, BREAK_IS_BIF, match_spec, (BeamInstr) BeamOp(op_i_mtrace_breakpoint), 0, tracer_pid);
+erts_consolidate_bp_data(BpFunctions* f, int local)
+{
+    BpFunction* fs = f->matching;
+    Uint i;
+    Uint n = f->matched;
+
+    ERTS_SMP_LC_ASSERT(erts_is_code_ix_locked());
+
+    for (i = 0; i < n; i++) {
+	consolidate_bp_data(fs[i].mod, fs[i].pc, local);
+    }
 }
 
-void erts_set_time_trace_bif(BeamInstr *pc, enum erts_break_op count_op) {
-    set_function_break(NULL, pc, BREAK_IS_BIF, NULL, (BeamInstr) BeamOp(op_i_time_breakpoint), count_op, NIL);
+void
+erts_consolidate_bif_bp_data(void)
+{
+    int i;
+
+    ERTS_SMP_LC_ASSERT(erts_is_code_ix_locked());
+    for (i = 0; i < BIF_SIZE; i++) {
+	Export *ep = bif_export[i];
+	consolidate_bp_data(0, ep->code+3, 0);
+    }
 }
 
-void erts_clear_time_trace_bif(BeamInstr *pc) {
-    clear_function_break(NULL, pc, BREAK_IS_BIF, (BeamInstr) BeamOp(op_i_time_breakpoint));
+static void
+consolidate_bp_data(Module* modp, BeamInstr* pc, int local)
+{
+    GenericBp* g = (GenericBp *) pc[-4];
+    GenericBpData* src;
+    GenericBpData* dst;
+    Uint flags;
+
+    if (g == 0) {
+	return;
+    }
+
+    src = &g->data[erts_active_bp_ix()];
+    dst = &g->data[erts_staging_bp_ix()];
+
+    /*
+     * The contents of the staging area may be out of date.
+     * Decrement all reference pointers.
+     */
+
+    flags = dst->flags;
+    if (flags & (ERTS_BPF_LOCAL_TRACE|ERTS_BPF_GLOBAL_TRACE)) {
+	MatchSetUnref(dst->local_ms);
+    }
+    if (flags & ERTS_BPF_META_TRACE) {
+	bp_meta_unref(dst->meta_pid);
+	MatchSetUnref(dst->meta_ms);
+    }
+    if (flags & ERTS_BPF_COUNT) {
+	bp_count_unref(dst->count);
+    }
+    if (flags & ERTS_BPF_TIME_TRACE) {
+	bp_time_unref(dst->time);
+    }
+
+    /*
+     * If all flags are zero, deallocate all breakpoint data.
+     */
+
+    flags = dst->flags = src->flags;
+    if (flags == 0) {
+	if (modp) {
+	    if (local) {
+		modp->curr.num_breakpoints--;
+	    } else {
+		modp->curr.num_traced_exports--;
+	    }
+	    ASSERT(modp->curr.num_breakpoints >= 0);
+	    ASSERT(modp->curr.num_traced_exports >= 0);
+	    ASSERT(*pc != (BeamInstr) BeamOp(op_i_generic_breakpoint));
+	}
+	pc[-4] = 0;
+	Free(g);
+	return;
+    }
+
+    /*
+     * Copy the active data to the staging area (making it ready
+     * for the next time it will be used).
+     */
+
+    if (flags & (ERTS_BPF_LOCAL_TRACE|ERTS_BPF_GLOBAL_TRACE)) {
+	dst->local_ms = src->local_ms;
+	MatchSetRef(dst->local_ms);
+    }
+    if (flags & ERTS_BPF_META_TRACE) {
+	dst->meta_pid = src->meta_pid;
+	erts_refc_inc(&dst->meta_pid->refc, 1);
+	dst->meta_ms = src->meta_ms;
+	MatchSetRef(dst->meta_ms);
+    }
+    if (flags & ERTS_BPF_COUNT) {
+	dst->count = src->count;
+	erts_refc_inc(&dst->count->refc, 1);
+    }
+    if (flags & ERTS_BPF_TIME_TRACE) {
+	dst->time = src->time;
+	erts_refc_inc(&dst->time->refc, 1);
+	ASSERT(dst->time->hash);
+    }
 }
 
-int 
-erts_set_debug_break(Eterm mfa[3], int specified) {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
-    return set_break(mfa, specified, NULL, 
-		     (BeamInstr) BeamOp(op_i_debug_breakpoint), 0, NIL);
+void
+erts_commit_staged_bp(void)
+{
+    ErtsBpIndex staging = erts_staging_bp_ix();
+    ErtsBpIndex active = erts_active_bp_ix();
+
+    erts_smp_atomic32_set_nob(&erts_active_bp_index, staging);
+    erts_smp_atomic32_set_nob(&erts_staging_bp_index, active);
 }
 
-int 
-erts_set_count_break(Eterm mfa[3], int specified, enum erts_break_op count_op) {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
-    return set_break(mfa, specified, NULL, 
-		     (BeamInstr) BeamOp(op_i_count_breakpoint), count_op, NIL);
+void
+erts_install_breakpoints(BpFunctions* f)
+{
+    Uint i;
+    Uint n = f->matched;
+    BeamInstr br = (BeamInstr) BeamOp(op_i_generic_breakpoint);
+
+    for (i = 0; i < n; i++) {
+	BeamInstr* pc = f->matching[i].pc;
+	GenericBp* g = (GenericBp *) pc[-4];
+	if (*pc != br && g) {
+	    Module* modp = f->matching[i].mod;
+
+	    /*
+	     * The breakpoint must be disabled in the active data
+	     * (it will enabled later by switching bp indices),
+	     * and enabled in the staging data.
+	     */
+	    ASSERT(g->data[erts_active_bp_ix()].flags == 0);
+	    ASSERT(g->data[erts_staging_bp_ix()].flags != 0);
+
+	    /*
+	     * The following write is not protected by any lock. We
+	     * assume that the hardware guarantees that a write of an
+	     * aligned word-size (or half-word) writes is atomic
+	     * (i.e. that other processes executing this code will not
+	     * see a half pointer).
+	     */
+	    *pc = br;
+	    modp->curr.num_breakpoints++;
+	}
+    }
 }
 
-int
-erts_set_time_break(Eterm mfa[3], int specified, enum erts_break_op count_op) {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
-    return set_break(mfa, specified, NULL,
-		     (BeamInstr) BeamOp(op_i_time_breakpoint), count_op, NIL);
+void
+erts_uninstall_breakpoints(BpFunctions* f)
+{
+    Uint i;
+    Uint n = f->matched;
+
+    for (i = 0; i < n; i++) {
+	BeamInstr* pc = f->matching[i].pc;
+	uninstall_breakpoint(pc);
+    }
 }
 
-int
-erts_clear_trace_break(Eterm mfa[3], int specified) {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
-    return clear_break(mfa, specified, 
-		       (BeamInstr) BeamOp(op_i_trace_breakpoint));
+static void
+uninstall_breakpoint(BeamInstr* pc)
+{
+    if (*pc == (BeamInstr) BeamOp(op_i_generic_breakpoint)) {
+	GenericBp* g = (GenericBp *) pc[-4];
+	if (g->data[erts_active_bp_ix()].flags == 0) {
+	    /*
+	     * The following write is not protected by any lock. We
+	     * assume that the hardware guarantees that a write of an
+	     * aligned word-size (or half-word) writes is atomic
+	     * (i.e. that other processes executing this code will not
+	     * see a half pointer).
+	     */
+	    *pc = g->orig_instr;
+	}
+    }
 }
 
-int
-erts_clear_mtrace_break(Eterm mfa[3], int specified) {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
-    return clear_break(mfa, specified, 
-		       (BeamInstr) BeamOp(op_i_mtrace_breakpoint));
+void
+erts_set_trace_break(BpFunctions* f, Binary *match_spec)
+{
+    set_break(f, match_spec, ERTS_BPF_LOCAL_TRACE, 0, am_true);
 }
 
 void
-erts_clear_mtrace_bif(BeamInstr *pc) {
-    clear_function_break(NULL, pc, BREAK_IS_BIF, (BeamInstr) BeamOp(op_i_mtrace_breakpoint));
+erts_set_mtrace_break(BpFunctions* f, Binary *match_spec, Eterm tracer_pid)
+{
+    set_break(f, match_spec, ERTS_BPF_META_TRACE, 0, tracer_pid);
 }
 
-int
-erts_clear_debug_break(Eterm mfa[3], int specified) {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
-    return clear_break(mfa, specified, 
-		       (BeamInstr) BeamOp(op_i_debug_breakpoint));
+void
+erts_set_call_trace_bif(BeamInstr *pc, Binary *match_spec, int local)
+{
+    Uint flags = local ? ERTS_BPF_LOCAL_TRACE : ERTS_BPF_GLOBAL_TRACE;
+
+    set_function_break(pc, match_spec, flags, 0, NIL);
 }
 
-int
-erts_clear_count_break(Eterm mfa[3], int specified) {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
-    return clear_break(mfa, specified, 
-		       (BeamInstr) BeamOp(op_i_count_breakpoint));
+void
+erts_set_mtrace_bif(BeamInstr *pc, Binary *match_spec, Eterm tracer_pid)
+{
+    set_function_break(pc, match_spec, ERTS_BPF_META_TRACE, 0, tracer_pid);
 }
 
-int
-erts_clear_time_break(Eterm mfa[3], int specified) {
-    ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
-    return clear_break(mfa, specified,
-		       (BeamInstr) BeamOp(op_i_time_breakpoint));
+void
+erts_set_time_trace_bif(BeamInstr *pc, enum erts_break_op count_op)
+{
+    set_function_break(pc, NULL,
+		       ERTS_BPF_TIME_TRACE|ERTS_BPF_TIME_TRACE_ACTIVE,
+		       count_op, NIL);
 }
 
-int
-erts_clear_break(Eterm mfa[3], int specified) {
+void
+erts_clear_time_trace_bif(BeamInstr *pc) {
+    clear_function_break(pc, ERTS_BPF_TIME_TRACE|ERTS_BPF_TIME_TRACE_ACTIVE);
+}
+
+void
+erts_set_debug_break(BpFunctions* f) {
+    set_break(f, NULL, ERTS_BPF_DEBUG, 0, NIL);
+}
+
+void
+erts_set_count_break(BpFunctions* f, enum erts_break_op count_op)
+{
+    set_break(f, 0, ERTS_BPF_COUNT|ERTS_BPF_COUNT_ACTIVE,
+	      count_op, NIL);
+}
+
+void
+erts_set_time_break(BpFunctions* f, enum erts_break_op count_op)
+{
+    set_break(f, 0, ERTS_BPF_TIME_TRACE|ERTS_BPF_TIME_TRACE_ACTIVE,
+	      count_op, NIL);
+}
+
+void
+erts_clear_trace_break(BpFunctions* f)
+{
+    clear_break(f, ERTS_BPF_LOCAL_TRACE);
+}
+
+void
+erts_clear_call_trace_bif(BeamInstr *pc, int local)
+{
+    GenericBp* g = (GenericBp *) pc[-4];
+
+    if (g) {
+	Uint flags = local ? ERTS_BPF_LOCAL_TRACE : ERTS_BPF_GLOBAL_TRACE;
+	if (g->data[erts_staging_bp_ix()].flags & flags) {
+	    clear_function_break(pc, flags);
+	}
+    }
+}
+
+void
+erts_clear_mtrace_break(BpFunctions* f)
+{
+    clear_break(f, ERTS_BPF_META_TRACE);
+}
+
+void
+erts_clear_mtrace_bif(BeamInstr *pc)
+{
+    clear_function_break(pc, ERTS_BPF_META_TRACE);
+}
+
+void
+erts_clear_debug_break(BpFunctions* f)
+{
     ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
-    return clear_break(mfa, specified, 0);
+    clear_break(f, ERTS_BPF_DEBUG);
+}
+
+void
+erts_clear_count_break(BpFunctions* f)
+{
+    clear_break(f, ERTS_BPF_COUNT|ERTS_BPF_COUNT_ACTIVE);
+}
+
+void
+erts_clear_time_break(BpFunctions* f)
+{
+    clear_break(f, ERTS_BPF_TIME_TRACE|ERTS_BPF_TIME_TRACE_ACTIVE);
+}
+
+void
+erts_clear_all_breaks(BpFunctions* f)
+{
+    clear_break(f, ERTS_BPF_ALL);
 }
 
 int 
 erts_clear_module_break(Module *modp) {
+    BeamInstr** code_base;
+    Uint n;
+    Uint i;
+
     ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
     ASSERT(modp);
-    return clear_module_break(modp, NULL, 0, 0);
+    code_base = (BeamInstr **) modp->curr.code;
+    if (code_base == NULL) {
+	return 0;
+    }
+    n = (Uint) code_base[MI_NUM_FUNCTIONS];
+    for (i = 0; i < n; ++i) {
+	BeamInstr* pc;
+
+	pc = code_base[MI_FUNCTIONS+i] + 5;
+	if (erts_is_native_break(pc)) {
+	    continue;
+	}
+	clear_function_break(pc, ERTS_BPF_ALL);
+    }
+
+    erts_commit_staged_bp();
+
+    for (i = 0; i < n; ++i) {
+	BeamInstr* pc;
+
+	pc = code_base[MI_FUNCTIONS+i] + 5;
+	if (erts_is_native_break(pc)) {
+	    continue;
+	}
+	uninstall_breakpoint(pc);
+	consolidate_bp_data(modp, pc, 1);
+	ASSERT(pc[-4] == 0);
+    }
+    return n;
 }
 
-int
-erts_clear_function_break(Module *modp, BeamInstr *pc) {
+void
+erts_clear_export_break(Module* modp, BeamInstr* pc)
+{
     ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
-    ASSERT(modp);
-    return clear_function_break(modp, pc, BREAK_IS_ERL, 0);
+
+    clear_function_break(pc, ERTS_BPF_ALL);
+    erts_commit_staged_bp();
+    *pc = (BeamInstr) 0;
+    consolidate_bp_data(modp, pc, 0);
+    ASSERT(pc[-4] == 0);
 }
 
+BeamInstr
+erts_generic_breakpoint(Process* c_p, BeamInstr* I, Eterm* reg)
+{
+    GenericBp* g;
+    GenericBpData* bp;
+    Uint bp_flags;
+    ErtsBpIndex ix = erts_active_bp_ix();
+
+    g = (GenericBp *) I[-4];
+    bp = &g->data[ix];
+    bp_flags = bp->flags;
+    ASSERT((bp_flags & ~ERTS_BPF_ALL) == 0);
+    if (bp_flags & (ERTS_BPF_LOCAL_TRACE|
+		    ERTS_BPF_GLOBAL_TRACE|
+		    ERTS_BPF_TIME_TRACE_ACTIVE) &&
+	!IS_TRACED_FL(c_p, F_TRACE_CALLS)) {
+	bp_flags &= ~(ERTS_BPF_LOCAL_TRACE|
+		      ERTS_BPF_GLOBAL_TRACE|
+		      ERTS_BPF_TIME_TRACE|
+		      ERTS_BPF_TIME_TRACE_ACTIVE);
+	if (bp_flags == 0) {	/* Quick exit */
+	    return g->orig_instr;
+	}
+    }
+
+    if (bp_flags & ERTS_BPF_LOCAL_TRACE) {
+	ASSERT((bp_flags & ERTS_BPF_GLOBAL_TRACE) == 0);
+	(void) do_call_trace(c_p, I, reg, 1, bp->local_ms, am_true);
+    } else if (bp_flags & ERTS_BPF_GLOBAL_TRACE) {
+	(void) do_call_trace(c_p, I, reg, 0, bp->local_ms, am_true);
+    }
+
+    if (bp_flags & ERTS_BPF_META_TRACE) {
+	Eterm old_pid;
+	Eterm new_pid;
+
+	old_pid = (Eterm) erts_smp_atomic_read_nob(&bp->meta_pid->pid);
+	new_pid = do_call_trace(c_p, I, reg, 1, bp->meta_ms, old_pid);
+	if (new_pid != old_pid) {
+	    erts_smp_atomic_set_nob(&bp->meta_pid->pid, new_pid);
+	}
+    }
+
+    if (bp_flags & ERTS_BPF_COUNT_ACTIVE) {
+	erts_smp_atomic_inc_nob(&bp->count->acount);
+    }
+
+    if (bp_flags & ERTS_BPF_TIME_TRACE_ACTIVE) {
+	Eterm w;
+	erts_trace_time_call(c_p, I, bp->time);
+	w = (BeamInstr) *c_p->cp;
+	if (! (w == (BeamInstr) BeamOp(op_i_return_time_trace) ||
+	       w == (BeamInstr) BeamOp(op_return_trace) ||
+	       w == (BeamInstr) BeamOp(op_i_return_to_trace)) ) {
+	    Eterm* E = c_p->stop;
+	    ASSERT(c_p->htop <= E && E <= c_p->hend);
+	    if (E - 2 < c_p->htop) {
+		(void) erts_garbage_collect(c_p, 2, reg, I[-1]);
+		ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
+	    }
+	    E = c_p->stop;
 
+	    ASSERT(c_p->htop <= E && E <= c_p->hend);
+
+	    E -= 2;
+	    E[0] = make_cp(I);
+	    E[1] = make_cp(c_p->cp);     /* original return address */
+	    c_p->cp = beam_return_time_trace;
+	    c_p->stop = E;
+	}
+    }
+
+    if (bp_flags & ERTS_BPF_DEBUG) {
+	return (BeamInstr) BeamOp(op_i_debug_breakpoint);
+    } else {
+	return g->orig_instr;
+    }
+}
 
 /*
- * SMP NOTE: Process p may have become exiting on return!
+ * Entry point called by the trace wrap functions in erl_bif_wrap.c
+ *
+ * The trace wrap functions are themselves called through the export
+ * entries instead of the original BIF functions.
  */
-BeamInstr
-erts_trace_break(Process *p, BeamInstr *pc, Eterm *args,
-		 Uint32 *ret_flags, Eterm *tracer_pid) {
-    Eterm tpid1, tpid2;
-    BpData **bds = (BpData **) (pc)[-4];
-    BpDataTrace *bdt = NULL;
+Eterm
+erts_bif_trace(int bif_index, Process* p, Eterm* args, BeamInstr* I)
+{
+    Eterm result;
+    Eterm (*func)(Process*, Eterm*, BeamInstr*);
+    Export* ep = bif_export[bif_index];
+    Uint32 flags = 0, flags_meta = 0;
+    Eterm meta_tracer_pid = NIL;
+    int applying = (I == &(ep->code[3])); /* Yup, the apply code for a bif
+					   * is actually in the
+					   * export entry */
+    BeamInstr *cp = p->cp;
+    GenericBp* g;
+    GenericBpData* bp;
+    Uint bp_flags = 0;
+
+    ERTS_SMP_CHK_HAVE_ONLY_MAIN_PROC_LOCK(p);
+
+    g = (GenericBp *) ep->fake_op_func_info_for_hipe[1];
+    if (g) {
+	bp = &g->data[erts_active_bp_ix()];
+	bp_flags = bp->flags;
+    }
 
-    ASSERT(bds);
-    ASSERT(pc[-5] == (BeamInstr) BeamOp(op_i_func_info_IaaI));
-    bdt = (BpDataTrace *) bds[bp_sched2ix_proc(p)];
-    ASSERT(bdt);
-    bdt = (BpDataTrace *) bdt->next;
-    ASSERT(bdt);
-    ASSERT(ret_flags);
-    ASSERT(tracer_pid);
-
-    ErtsSmpBPLock(bdt);
-    tpid1 = tpid2 = bdt->tracer_pid;
-    ErtsSmpBPUnlock(bdt);
-
-    *ret_flags = erts_call_trace(p, pc-3/*mfa*/, bdt->match_spec, args,
-				 1, &tpid2);
-    *tracer_pid = tpid2;
-    if (tpid1 != tpid2) {
-	ErtsSmpBPLock(bdt);
-	bdt->tracer_pid = tpid2;
-	ErtsSmpBPUnlock(bdt);
-    }
-    bds[bp_sched2ix_proc(p)] = (BpData *) bdt;
-    return bdt->orig_instr;
+    /*
+     * Make continuation pointer OK, it is not during direct BIF calls,
+     * but it is correct during apply of bif.
+     */
+    if (!applying) {
+	p->cp = I;
+    }
+    if (bp_flags & (ERTS_BPF_LOCAL_TRACE|ERTS_BPF_GLOBAL_TRACE) &&
+	IS_TRACED_FL(p, F_TRACE_CALLS)) {
+	int local = !!(bp_flags & ERTS_BPF_LOCAL_TRACE);
+	flags = erts_call_trace(p, ep->code, bp->local_ms, args,
+				local, &p->tracer_proc);
+    }
+    if (bp_flags & ERTS_BPF_META_TRACE) {
+	Eterm tpid1, tpid2;
+
+	tpid1 = tpid2 =
+	    (Eterm) erts_smp_atomic_read_nob(&bp->meta_pid->pid);
+	flags_meta = erts_call_trace(p, ep->code, bp->meta_ms, args,
+				     0, &tpid2);
+	meta_tracer_pid = tpid2;
+	if (tpid1 != tpid2) {
+	    erts_smp_atomic_set_nob(&bp->meta_pid->pid, tpid2);
+	}
+    }
+    if (bp_flags & ERTS_BPF_TIME_TRACE_ACTIVE &&
+	IS_TRACED_FL(p, F_TRACE_CALLS)) {
+	BeamInstr *pc = (BeamInstr *)ep->code+3;
+	erts_trace_time_call(p, pc, bp->time);
+    }
+
+    /* Restore original continuation pointer (if changed). */
+    p->cp = cp;
+
+    func = bif_table[bif_index].f;
+
+    result = func(p, args, I);
+
+    if (applying && (flags & MATCH_SET_RETURN_TO_TRACE)) {
+	BeamInstr i_return_trace      = beam_return_trace[0];
+	BeamInstr i_return_to_trace   = beam_return_to_trace[0];
+	BeamInstr i_return_time_trace = beam_return_time_trace[0];
+	Eterm *cpp;
+	/* Maybe advance cp to skip trace stack frames */
+	for (cpp = p->stop;  ;  cp = cp_val(*cpp++)) {
+	    if (*cp == i_return_trace) {
+		/* Skip stack frame variables */
+		while (is_not_CP(*cpp)) cpp++;
+		cpp += 2; /* Skip return_trace parameters */
+	    } else if (*cp == i_return_time_trace) {
+		/* Skip stack frame variables */
+		while (is_not_CP(*cpp)) cpp++;
+		cpp += 1; /* Skip return_time_trace parameters */
+	    } else if (*cp == i_return_to_trace) {
+		/* A return_to trace message is going to be generated
+		 * by normal means, so we do not have to.
+		 */
+		cp = NULL;
+		break;
+	    } else break;
+	}
+    }
+
+    /* Try to get these in the order
+     * they usually appear in normal code... */
+    if (is_non_value(result)) {
+	Uint reason = p->freason;
+	if (reason != TRAP) {
+	    Eterm class;
+	    Eterm value = p->fvalue;
+	    DeclareTmpHeapNoproc(nocatch,3);
+	    UseTmpHeapNoproc(3);
+	    /* Expand error value like in handle_error() */
+	    if (reason & EXF_ARGLIST) {
+		Eterm *tp;
+		ASSERT(is_tuple(value));
+		tp = tuple_val(value);
+		value = tp[1];
+	    }
+	    if ((reason & EXF_THROWN) && (p->catches <= 0)) {
+		value = TUPLE2(nocatch, am_nocatch, value);
+		reason = EXC_ERROR;
+	    }
+	    /* Note: expand_error_value() could theoretically
+	     * allocate on the heap, but not for any error
+	     * returned by a BIF, and it would do no harm,
+	     * just be annoying.
+	     */
+	    value = expand_error_value(p, reason, value);
+	    class = exception_tag[GET_EXC_CLASS(reason)];
+
+	    if (flags_meta & MATCH_SET_EXCEPTION_TRACE) {
+		erts_trace_exception(p, ep->code, class, value,
+				     &meta_tracer_pid);
+	    }
+	    if (flags & MATCH_SET_EXCEPTION_TRACE) {
+		erts_trace_exception(p, ep->code, class, value,
+				     &p->tracer_proc);
+	    }
+	    if ((flags & MATCH_SET_RETURN_TO_TRACE) && p->catches > 0) {
+		/* can only happen if(local)*/
+		Eterm *ptr = p->stop;
+		ASSERT(is_CP(*ptr));
+		ASSERT(ptr <= STACK_START(p));
+		/* Search the nearest stack frame for a catch */
+		while (++ptr < STACK_START(p)) {
+		    if (is_CP(*ptr)) break;
+		    if (is_catch(*ptr)) {
+			if (applying) {
+			    /* Apply of BIF, cp is in calling function */
+			    if (cp) erts_trace_return_to(p, cp);
+			} else {
+			    /* Direct bif call, I points into
+			     * calling function */
+			    erts_trace_return_to(p, I);
+			}
+		    }
+		}
+	    }
+	    UnUseTmpHeapNoproc(3);
+	    if ((flags_meta|flags) & MATCH_SET_EXCEPTION_TRACE) {
+		erts_smp_proc_lock(p, ERTS_PROC_LOCKS_ALL_MINOR);
+		p->trace_flags |= F_EXCEPTION_TRACE;
+		erts_smp_proc_unlock(p, ERTS_PROC_LOCKS_ALL_MINOR);
+	    }
+	}
+    } else {
+	if (flags_meta & MATCH_SET_RX_TRACE) {
+	    erts_trace_return(p, ep->code, result, &meta_tracer_pid);
+	}
+	/* MATCH_SET_RETURN_TO_TRACE cannot occur if(meta) */
+	if (flags & MATCH_SET_RX_TRACE) {
+	    erts_trace_return(p, ep->code, result, &p->tracer_proc);
+	}
+	if (flags & MATCH_SET_RETURN_TO_TRACE) {
+	    /* can only happen if(local)*/
+	    if (applying) {
+		/* Apply of BIF, cp is in calling function */
+		if (cp) erts_trace_return_to(p, cp);
+	    } else {
+		/* Direct bif call, I points into calling function */
+		erts_trace_return_to(p, I);
+	    }
+	}
+    }
+    ERTS_SMP_CHK_HAVE_ONLY_MAIN_PROC_LOCK(p);
+    return result;
+}
+
+static Eterm
+do_call_trace(Process* c_p, BeamInstr* I, Eterm* reg,
+	      int local, Binary* ms, Eterm tracer_pid)
+{
+    Eterm* cpp;
+    int return_to_trace = 0;
+    BeamInstr w;
+    BeamInstr *cp_save;
+    Uint32 flags;
+    Uint need = 0;
+    Eterm* E = c_p->stop;
+
+    w = *c_p->cp;
+    if (w == (BeamInstr) BeamOp(op_return_trace)) {
+	cpp = &E[2];
+    } else if (w == (BeamInstr) BeamOp(op_i_return_to_trace)) {
+	return_to_trace = 1;
+	cpp = &E[0];
+    } else if (w == (BeamInstr) BeamOp(op_i_return_time_trace)) {
+	cpp = &E[0];
+    } else {
+	cpp = NULL;
+    }
+    if (cpp) {
+	for (;;) {
+	    BeamInstr w = *cp_val(*cpp);
+	    if (w == (BeamInstr) BeamOp(op_return_trace)) {
+		cpp += 3;
+	    } else if (w == (BeamInstr) BeamOp(op_i_return_to_trace)) {
+		return_to_trace = 1;
+		cpp += 1;
+	    } else if (w == (BeamInstr) BeamOp(op_i_return_time_trace)) {
+		cpp += 2;
+	    } else {
+		break;
+	    }
+	}
+	cp_save = c_p->cp;
+	c_p->cp = (BeamInstr *) cp_val(*cpp);
+	ASSERT(is_CP(*cpp));
+    }
+    ERTS_SMP_UNREQ_PROC_MAIN_LOCK(c_p);
+    flags = erts_call_trace(c_p, I-3, ms, reg, local, &tracer_pid);
+    ERTS_SMP_REQ_PROC_MAIN_LOCK(c_p);
+    if (cpp) {
+	c_p->cp = cp_save;
+    }
+
+    ASSERT(!ERTS_PROC_IS_EXITING(c_p));
+    if ((flags & MATCH_SET_RETURN_TO_TRACE) && !return_to_trace) {
+	need += 1;
+    }
+    if (flags & MATCH_SET_RX_TRACE) {
+	need += 3;
+    }
+    if (need) {
+	ASSERT(c_p->htop <= E && E <= c_p->hend);
+	if (E - need < c_p->htop) {
+	    (void) erts_garbage_collect(c_p, need, reg, I[-1]);
+	    ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
+	    E = c_p->stop;
+	}
+    }
+    if (flags & MATCH_SET_RETURN_TO_TRACE && !return_to_trace) {
+	E -= 1;
+	ASSERT(c_p->htop <= E && E <= c_p->hend);
+	E[0] = make_cp(c_p->cp);
+	c_p->cp = (BeamInstr *) beam_return_to_trace;
+    }
+    if (flags & MATCH_SET_RX_TRACE) {
+	E -= 3;
+	ASSERT(c_p->htop <= E && E <= c_p->hend);
+	ASSERT(is_CP((Eterm) (UWord) (I - 3)));
+	ASSERT(am_true == tracer_pid ||
+	       is_internal_pid(tracer_pid) || is_internal_port(tracer_pid));
+	E[2] = make_cp(c_p->cp);
+	E[1] = tracer_pid;
+	E[0] = make_cp(I - 3); /* We ARE at the beginning of an
+				  instruction,
+				  the funcinfo is above i. */
+	c_p->cp = (flags & MATCH_SET_EXCEPTION_TRACE) ?
+	    beam_exception_trace : beam_return_trace;
+	erts_smp_proc_lock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
+	c_p->trace_flags |= F_EXCEPTION_TRACE;
+	erts_smp_proc_unlock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
+    }
+    c_p->stop = E;
+    return tracer_pid;
 }
 
+void
+erts_trace_time_call(Process* c_p, BeamInstr* I, BpDataTime* bdt)
+{
+    Uint ms,s,us;
+    process_breakpoint_time_t *pbt = NULL;
+    bp_data_time_item_t sitem, *item = NULL;
+    bp_time_hash_t *h = NULL;
+    BpDataTime *pbdt = NULL;
 
+    ASSERT(c_p);
+    ASSERT(erts_smp_atomic32_read_acqb(&c_p->state) & ERTS_PSFLG_RUNNING);
 
-/*
- * SMP NOTE: Process p may have become exiting on return!
- */
-Uint32
-erts_bif_mtrace(Process *p, BeamInstr *pc, Eterm *args, int local,
-		Eterm *tracer_pid) {
-    BpData **bds = (BpData **) (pc)[-4];
-    BpDataTrace *bdt = NULL;
+    /* get previous timestamp and breakpoint
+     * from the process psd  */
 
+    pbt = ERTS_PROC_GET_CALL_TIME(c_p);
+    get_sys_now(&ms, &s, &us);
 
-    ASSERT(tracer_pid);
-    if (bds) {
-	Eterm tpid1, tpid2;
-	Uint32 flags;
-	bdt = (BpDataTrace *)bds[bp_sched2ix_proc(p)];
+    /* get pbt
+     * timestamp = t0
+     * lookup bdt from code
+     * set ts0 to pbt
+     * add call count here?
+     */
+    if (pbt == 0) {
+	/* First call of process to instrumented function */
+	pbt = Alloc(sizeof(process_breakpoint_time_t));
+	(void *) ERTS_PROC_SET_CALL_TIME(c_p, ERTS_PROC_LOCK_MAIN, pbt);
+    } else {
+	ASSERT(pbt->pc);
+	/* add time to previous code */
+	bp_time_diff(&sitem, pbt, ms, s, us);
+	sitem.pid = c_p->id;
+	sitem.count = 0;
 
-	ErtsSmpBPLock(bdt);
-	tpid1 = tpid2 = bdt->tracer_pid;
-	ErtsSmpBPUnlock(bdt);
+	/* previous breakpoint */
+	pbdt = get_time_break(pbt->pc);
 
-	flags = erts_call_trace(p, pc-3/*mfa*/, bdt->match_spec, args,
-				local, &tpid2);
-	*tracer_pid = tpid2;
-	if (tpid1 != tpid2) {
-	    ErtsSmpBPLock(bdt);
-	    bdt->tracer_pid = tpid2;
-	    ErtsSmpBPUnlock(bdt);
+	/* if null then the breakpoint was removed */
+	if (pbdt) {
+	    h = &(pbdt->hash[bp_sched2ix_proc(c_p)]);
+
+	    ASSERT(h);
+	    ASSERT(h->item);
+
+	    item = bp_hash_get(h, &sitem);
+	    if (!item) {
+		item = bp_hash_put(h, &sitem);
+	    } else {
+		BP_TIME_ADD(item, &sitem);
+	    }
 	}
-	return flags;
     }
-    *tracer_pid = NIL;
-    return 0;
+
+    /* Add count to this code */
+    sitem.pid     = c_p->id;
+    sitem.count   = 1;
+    sitem.s_time  = 0;
+    sitem.us_time = 0;
+
+    /* this breakpoint */
+    ASSERT(bdt);
+    h = &(bdt->hash[bp_sched2ix_proc(c_p)]);
+
+    ASSERT(h);
+    ASSERT(h->item);
+
+    item = bp_hash_get(h, &sitem);
+    if (!item) {
+	item = bp_hash_put(h, &sitem);
+    } else {
+	BP_TIME_ADD(item, &sitem);
+    }
+
+    pbt->pc = I;
+    pbt->ms = ms;
+    pbt->s  = s;
+    pbt->us = us;
 }
 
+void
+erts_trace_time_return(Process *p, BeamInstr *pc)
+{
+    Uint ms,s,us;
+    process_breakpoint_time_t *pbt = NULL;
+    bp_data_time_item_t sitem, *item = NULL;
+    bp_time_hash_t *h = NULL;
+    BpDataTime *pbdt = NULL;
+
+    ASSERT(p);
+    ASSERT(erts_smp_atomic32_read_acqb(&p->state) & ERTS_PSFLG_RUNNING);
+
+    /* get previous timestamp and breakpoint
+     * from the process psd  */
+
+    pbt = ERTS_PROC_GET_CALL_TIME(p);
+    get_sys_now(&ms,&s,&us);
+
+    /* get pbt
+     * lookup bdt from code
+     * timestamp = t1
+     * get ts0 from pbt
+     * get item from bdt->hash[bp_hash(p->id)]
+     * ack diff (t1, t0) to item
+     */
+
+    if (pbt) {
+	/* might have been removed due to
+	 * trace_pattern(false)
+	 */
+	ASSERT(pbt->pc);
+
+	bp_time_diff(&sitem, pbt, ms, s, us);
+	sitem.pid   = p->id;
+	sitem.count = 0;
+
+	/* previous breakpoint */
+	pbdt = get_time_break(pbt->pc);
+
+	/* beware, the trace_pattern might have been removed */
+	if (pbdt) {
+	    h = &(pbdt->hash[bp_sched2ix_proc(p)]);
 
+	    ASSERT(h);
+	    ASSERT(h->item);
+
+	    item = bp_hash_get(h, &sitem);
+	    if (!item) {
+		item = bp_hash_put(h, &sitem);
+	    } else {
+		BP_TIME_ADD(item, &sitem);
+	    }
+	}
+
+	pbt->pc = pc;
+	pbt->ms = ms;
+	pbt->s  = s;
+	pbt->us = us;
+    }
+}
 
 int 
-erts_is_trace_break(BeamInstr *pc, Binary **match_spec_ret, Eterm *tracer_pid_ret) {
-    BpDataTrace *bdt = 
-	(BpDataTrace *) is_break(pc, (BeamInstr) BeamOp(op_i_trace_breakpoint));
-    
-    if (bdt) {
+erts_is_trace_break(BeamInstr *pc, Binary **match_spec_ret, int local)
+{
+    Uint flags = local ? ERTS_BPF_LOCAL_TRACE : ERTS_BPF_GLOBAL_TRACE;
+    GenericBpData* bp = check_break(pc, flags);
+
+    if (bp) {
 	if (match_spec_ret) {
-	    *match_spec_ret = bdt->match_spec;
-	}
-	if (tracer_pid_ret) {
-	    ErtsSmpBPLock(bdt);
-	    *tracer_pid_ret = bdt->tracer_pid;
-	    ErtsSmpBPUnlock(bdt);
+	    *match_spec_ret = bp->local_ms;
 	}
-	return !0;
+	return 1;
     }
     return 0;
 }
 
 int 
-erts_is_mtrace_break(BeamInstr *pc, Binary **match_spec_ret, Eterm *tracer_pid_ret) {
-    BpDataTrace *bdt = 
-	(BpDataTrace *) is_break(pc, (BeamInstr) BeamOp(op_i_mtrace_breakpoint));
+erts_is_mtrace_break(BeamInstr *pc, Binary **match_spec_ret,
+		     Eterm *tracer_pid_ret)
+{
+    GenericBpData* bp = check_break(pc, ERTS_BPF_META_TRACE);
     
-    if (bdt) {
+    if (bp) {
 	if (match_spec_ret) {
-	    *match_spec_ret = bdt->match_spec;
+	    *match_spec_ret = bp->meta_ms;
 	}
 	if (tracer_pid_ret) {
-	    ErtsSmpBPLock(bdt);
-	    *tracer_pid_ret = bdt->tracer_pid;
-	    ErtsSmpBPUnlock(bdt);
+	    *tracer_pid_ret =
+		(Eterm) erts_smp_atomic_read_nob(&bp->meta_pid->pid);
 	}
-	return !0;
+	return 1;
     }
     return 0;
 }
@@ -402,15 +1129,15 @@ erts_is_native_break(BeamInstr *pc) {
 }
 
 int 
-erts_is_count_break(BeamInstr *pc, Sint *count_ret) {
-    BpDataCount *bdc = 
-	(BpDataCount *) is_break(pc, (BeamInstr) BeamOp(op_i_count_breakpoint));
+erts_is_count_break(BeamInstr *pc, Uint *count_ret)
+{
+    GenericBpData* bp = check_break(pc, ERTS_BPF_COUNT);
     
-    if (bdc) {
+    if (bp) {
 	if (count_ret) {
-	    *count_ret = (Sint) erts_smp_atomic_read_nob(&bdc->acount);
+	    *count_ret = (Uint) erts_smp_atomic_read_nob(&bp->count->acount);
 	}
-	return !0;
+	return 1;
     }
     return 0;
 }
@@ -421,7 +1148,7 @@ int erts_is_time_break(Process *p, BeamInstr *pc, Eterm *retval) {
     Uint size;
     Eterm *hp, t;
     bp_data_time_item_t *item = NULL;
-    BpDataTime *bdt = (BpDataTime *) is_break(pc, (BeamInstr) BeamOp(op_i_time_breakpoint));
+    BpDataTime *bdt = get_time_break(pc);
 
     if (bdt) {
 	if (retval) {
@@ -464,7 +1191,7 @@ int erts_is_time_break(Process *p, BeamInstr *pc, Eterm *retval) {
 	    }
 	    bp_hash_delete(&hash);
 	}
-	return !0;
+	return 1;
     }
 
     return 0;
@@ -655,7 +1382,7 @@ void erts_schedule_time_break(Process *p, Uint schedule) {
 	     * the previous breakpoint.
 	     */
 
-	    pbdt = (BpDataTime *) get_break(p, pbt->pc, (BeamInstr) BeamOp(op_i_time_breakpoint));
+	    pbdt = get_time_break(pbt->pc);
 	    if (pbdt) {
 		get_sys_now(&ms,&s,&us);
 		bp_time_diff(&sitem, pbt, ms, s, us);
@@ -693,671 +1420,259 @@ void erts_schedule_time_break(Process *p, Uint schedule) {
     } /* pbt */
 }
 
-/* call_time breakpoint
- * Accumulated times are added to the previous bp,
- * not the current one. The current one is saved
- * for future reference.
- * The previous breakpoint is stored in the process it self, the psd.
- * We do not need to store in a stack frame.
- * There is no need for locking, each thread has its own
- * area in each bp to save data.
- * Since we need to diffrentiate between processes for each bp,
- * every bp has a hash (per thread) to process-bp statistics.
- * - egil
- */
-
-void erts_trace_time_break(Process *p, BeamInstr *pc, BpDataTime *bdt, Uint type) {
-    Uint ms,s,us;
-    process_breakpoint_time_t *pbt = NULL;
-    bp_data_time_item_t sitem, *item = NULL;
-    bp_time_hash_t *h = NULL;
-    BpDataTime *pbdt = NULL;
-
-    ASSERT(p);
-    ASSERT(ERTS_PSFLG_RUNNING & erts_smp_atomic32_read_acqb(&p->state));
-
-    /* get previous timestamp and breakpoint
-     * from the process psd  */
-
-    pbt = ERTS_PROC_GET_CALL_TIME(p);
-    get_sys_now(&ms,&s,&us);
-
-    switch(type) {
-	    /* get pbt
-	     * timestamp = t0
-	     * lookup bdt from code
-	     * set ts0 to pbt
-	     * add call count here?
-	     */
-	case ERTS_BP_CALL_TIME_CALL:
-	case ERTS_BP_CALL_TIME_TAIL_CALL:
-
-	    if (pbt) {
-		ASSERT(pbt->pc);
-		/* add time to previous code */
-		bp_time_diff(&sitem, pbt, ms, s, us);
-		sitem.pid   = p->id;
-		sitem.count = 0;
-
-		/* previous breakpoint */
-		pbdt = (BpDataTime *) get_break(p, pbt->pc, (BeamInstr) BeamOp(op_i_time_breakpoint));
-
-		/* if null then the breakpoint was removed */
-		if (pbdt) {
-		    h = &(pbdt->hash[bp_sched2ix_proc(p)]);
-
-		    ASSERT(h);
-		    ASSERT(h->item);
-
-		    item = bp_hash_get(h, &sitem);
-		    if (!item) {
-			item = bp_hash_put(h, &sitem);
-		    } else {
-			BP_TIME_ADD(item, &sitem);
-		    }
-		}
-
-	    } else {
-		/* first call of process to instrumented function */
-		pbt = Alloc(sizeof(process_breakpoint_time_t));
-		(void *) ERTS_PROC_SET_CALL_TIME(p, ERTS_PROC_LOCK_MAIN, pbt);
-	    }
-	    /* add count to this code */
-	    sitem.pid     = p->id;
-	    sitem.count   = 1;
-	    sitem.s_time  = 0;
-	    sitem.us_time = 0;
-
-	    /* this breakpoint */
-	    ASSERT(bdt);
-	    h = &(bdt->hash[bp_sched2ix_proc(p)]);
-
-	    ASSERT(h);
-	    ASSERT(h->item);
-
-	    item = bp_hash_get(h, &sitem);
-	    if (!item) {
-		item = bp_hash_put(h, &sitem);
-	    } else {
-		BP_TIME_ADD(item, &sitem);
-	    }
-
-	    pbt->pc = pc;
-	    pbt->ms = ms;
-	    pbt->s  = s;
-	    pbt->us = us;
-	    break;
-
-	case ERTS_BP_CALL_TIME_RETURN:
-	    /* get pbt
-	     * lookup bdt from code
-	     * timestamp = t1
-	     * get ts0 from pbt
-	     * get item from bdt->hash[bp_hash(p->id)]
-	     * ack diff (t1, t0) to item
-	     */
-
-	    if(pbt) {
-		/* might have been removed due to
-		 * trace_pattern(false)
-		 */
-		ASSERT(pbt->pc);
-
-		bp_time_diff(&sitem, pbt, ms, s, us);
-		sitem.pid   = p->id;
-		sitem.count = 0;
-
-		/* previous breakpoint */
-		pbdt = (BpDataTime *) get_break(p, pbt->pc, (BeamInstr) BeamOp(op_i_time_breakpoint));
-
-		/* beware, the trace_pattern might have been removed */
-		if (pbdt) {
-		    h = &(pbdt->hash[bp_sched2ix_proc(p)]);
-
-		    ASSERT(h);
-		    ASSERT(h->item);
-
-		    item = bp_hash_get(h, &sitem);
-		    if (!item) {
-			item = bp_hash_put(h, &sitem);
-		    } else {
-			BP_TIME_ADD(item, &sitem);
-		    }
-		}
-
-		pbt->pc = pc;
-		pbt->ms = ms;
-		pbt->s  = s;
-		pbt->us = us;
-	    }
-	    break;
-	default :
-	    ASSERT(0);
-		/* will never happen */
-	    break;
-    }
-}
-
-
 /* *************************************************************************
 ** Local helpers
 */
 
 
-static int set_break(Eterm mfa[3], int specified, 
-		     Binary *match_spec, BeamInstr break_op,
-		     enum erts_break_op count_op, Eterm tracer_pid)
+static void
+set_break(BpFunctions* f, Binary *match_spec, Uint break_flags,
+	  enum erts_break_op count_op, Eterm tracer_pid)
 {
-    Module *modp;
-    int num_processed = 0;
-    ErtsCodeIndex code_ix = erts_active_code_ix();
-    if (!specified) {
-	/* Find and process all modules in the system... */
-	int current;
-	int last = module_code_size(code_ix);
-	for (current = 0; current < last; current++) {
-	    modp = module_code(current, code_ix);
-	    ASSERT(modp != NULL);
-	    num_processed += 
-		set_module_break(modp, mfa, specified, 
-				 match_spec, break_op, count_op, 
-				 tracer_pid);
-	}
-    } else {
-	/* Process a single module */
-	if ((modp = erts_get_module(mfa[0], code_ix)) != NULL) {
-	    num_processed += 
-		set_module_break(modp, mfa, specified, 
-				 match_spec, break_op, count_op, 
-				 tracer_pid);
-	}	
-    }
-    return num_processed;
-}
-
-static int set_module_break(Module *modp, Eterm mfa[3], int specified,
-			    Binary *match_spec, BeamInstr break_op,
-			    enum erts_break_op count_op, Eterm tracer_pid) {
-    BeamInstr** code_base;
-    BeamInstr* code_ptr;
-    int num_processed = 0;
-    Uint i,n;
+    Uint i;
+    Uint n;
 
-    ASSERT(break_op);
-    ASSERT(modp);
-    code_base = (BeamInstr **) modp->curr.code;
-    if (code_base == NULL) {
-	return 0;
+    n = f->matched;
+    for (i = 0; i < n; i++) {
+	BeamInstr* pc = f->matching[i].pc;
+	set_function_break(pc, match_spec, break_flags,
+			   count_op, tracer_pid);
     }
-    n = (BeamInstr) code_base[MI_NUM_FUNCTIONS];
-    for (i = 0; i < n; ++i) {
-	code_ptr = code_base[MI_FUNCTIONS+i];
-	ASSERT(code_ptr[0] == (BeamInstr) BeamOp(op_i_func_info_IaaI));
-	if ((specified < 2 || mfa[1] == ((Eterm) code_ptr[3])) &&
-	    (specified < 3 || ((int) mfa[2]) == ((int) code_ptr[4]))) {
-	    BeamInstr *pc = code_ptr+5;
-	    
-	    num_processed +=
-		set_function_break(modp, pc, BREAK_IS_ERL, match_spec,
-				   break_op, count_op, tracer_pid);
-	}
-    }
-    return num_processed;
 }
 
-static int set_function_break(Module *modp, BeamInstr *pc, int bif,
-			      Binary *match_spec, BeamInstr break_op,
-			      enum erts_break_op count_op, Eterm tracer_pid) {
-
-    BeamInstr **code_base = NULL;
-    BpData *bd, **r, ***rs;
-    size_t size;
-    Uint ix = 0;
-    
-    if (bif == BREAK_IS_ERL) {
-	code_base = (BeamInstr **)modp->curr.code;
-	ASSERT(code_base);
-	ASSERT(code_base <= (BeamInstr **)pc);
-	ASSERT((BeamInstr **)pc < code_base + (modp->curr.code_length/sizeof(BeamInstr *)));
-    } else {
-	ASSERT(*pc == (BeamInstr) em_apply_bif);
-	ASSERT(modp == NULL);
+static void
+set_function_break(BeamInstr *pc, Binary *match_spec, Uint break_flags,
+		   enum erts_break_op count_op, Eterm tracer_pid)
+{
+    GenericBp* g;
+    GenericBpData* bp;
+    Uint common;
+    ErtsBpIndex ix = erts_staging_bp_ix();
+
+    ERTS_SMP_LC_ASSERT(erts_is_code_ix_locked());
+    g = (GenericBp *) pc[-4];
+    if (g == 0) {
+	int i;
+	if (count_op == erts_break_reset || count_op == erts_break_stop) {
+	    /* Do not insert a new breakpoint */
+	    return;
+	}
+	g = Alloc(sizeof(GenericBp));
+	g->orig_instr = *pc;
+	for (i = 0; i < ERTS_NUM_BP_IX; i++) {
+	    g->data[i].flags = 0;
+	}
+	pc[-4] = (BeamInstr) g;
     }
+    bp = &g->data[ix];
 
     /*
-     * Currently no trace support for native code.
+     * If we are changing an existing breakpoint, clean up old data.
      */
-    if (erts_is_native_break(pc)) {
-	return 0;
-    }
-    /* Do not allow two breakpoints of the same kind */
-    if ( (bd = is_break(pc, break_op))) {
-	if (break_op == (BeamInstr) BeamOp(op_i_trace_breakpoint)
-	    || break_op == (BeamInstr) BeamOp(op_i_mtrace_breakpoint)) {
-
-	    BpDataTrace *bdt = (BpDataTrace *) bd;
-	    Binary *old_match_spec;
-	    
-	    /* Update match spec and tracer */
-	    MatchSetRef(match_spec);
-	    ErtsSmpBPLock(bdt);
-	    old_match_spec = bdt->match_spec;
-	    bdt->match_spec = match_spec;
-	    bdt->tracer_pid = tracer_pid;
-	    ErtsSmpBPUnlock(bdt);
-	    MatchSetUnref(old_match_spec);
-	} else {
-	    BpDataCount *bdc = (BpDataCount *) bd;
-	    erts_aint_t count = 0;
-	    erts_aint_t res   = 0;
-
-	    ASSERT(! match_spec);
-	    ASSERT(is_nil(tracer_pid));
-		
-	    if (break_op == (BeamInstr) BeamOp(op_i_count_breakpoint)) {
-		if (count_op == erts_break_stop) {
-		    count = erts_smp_atomic_read_nob(&bdc->acount);
-		    if (count >= 0) {
-			while(1) {
-			    res = erts_smp_atomic_cmpxchg_nob(&bdc->acount, -count - 1, count);
-			    if ((res == count) || count < 0) break;
-			    count = res;
-			}
-		    }
-		} else {
-		    /* Reset call counter */
-		    erts_smp_atomic_set_nob(&bdc->acount, 0);
-		}
-
-	    } else if (break_op == (BeamInstr) BeamOp(op_i_time_breakpoint)) {
-		BpDataTime *bdt = (BpDataTime *) bd;
-		Uint i = 0;
-
-		ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
-
-		if (count_op == erts_break_stop) {
-		    bdt->pause = 1;
-		} else {
-		    bdt->pause = 0;
-		    for (i = 0; i < bdt->n; i++) {
-			bp_hash_delete(&(bdt->hash[i]));
-			bp_hash_init(&(bdt->hash[i]), 32);
-		    }
-		}
 
-	    } else {
-		ASSERT (! count_op);
-	    }
-	}
-	return 1;
-    }
-    if (break_op == (BeamInstr) BeamOp(op_i_trace_breakpoint) ||
-	break_op == (BeamInstr) BeamOp(op_i_mtrace_breakpoint)) {
-	size = sizeof(BpDataTrace);
-    } else {
-	ASSERT(! match_spec);
-	ASSERT(is_nil(tracer_pid));
-	if (break_op == (BeamInstr) BeamOp(op_i_count_breakpoint)) {
-	    if (count_op == erts_break_reset || count_op == erts_break_stop) {
-		/* Do not insert a new breakpoint */
-		return 1;
-	    }
-	    size = sizeof(BpDataCount);
-	} else if (break_op == (BeamInstr) BeamOp(op_i_time_breakpoint))  {
-	    if (count_op == erts_break_reset || count_op == erts_break_stop) {
-		/* Do not insert a new breakpoint */
-		return 1;
-	    }
-	    size = sizeof(BpDataTime);
+    common = break_flags & bp->flags;
+    if (common & (ERTS_BPF_LOCAL_TRACE|ERTS_BPF_GLOBAL_TRACE)) {
+	MatchSetUnref(bp->local_ms);
+    } else if (common & ERTS_BPF_META_TRACE) {
+	MatchSetUnref(bp->meta_ms);
+	bp_meta_unref(bp->meta_pid);
+    } else if (common & ERTS_BPF_COUNT) {
+	if (count_op == erts_break_stop) {
+	    bp->flags &= ~ERTS_BPF_COUNT_ACTIVE;
 	} else {
-	    ASSERT(! count_op);
-	    ASSERT(break_op == (BeamInstr) BeamOp(op_i_debug_breakpoint));
-	    size = sizeof(BpDataDebug);
+	    bp->flags |= ERTS_BPF_COUNT_ACTIVE;
+	    erts_smp_atomic_set_nob(&bp->count->acount, 0);
 	}
-    }
-    rs = (BpData ***) (pc-4);
-    if (! *rs) {
-	size_t ssize = sizeof(BeamInstr) * erts_no_schedulers;
-	*rs = (BpData **) Alloc(ssize);
-	sys_memzero(*rs, ssize);
-    }
-
-    r = &((*rs)[0]);
-
-    if (! *r) {
-	ASSERT(*pc != (BeamInstr) BeamOp(op_i_trace_breakpoint));
-	ASSERT(*pc != (BeamInstr) BeamOp(op_i_mtrace_breakpoint));
-	ASSERT(*pc != (BeamInstr) BeamOp(op_i_debug_breakpoint));
-	ASSERT(*pc != (BeamInstr) BeamOp(op_i_count_breakpoint));
-	ASSERT(*pc != (BeamInstr) BeamOp(op_i_time_breakpoint));
-	/* First breakpoint; create singleton ring */
-	bd = Alloc(size);
-	BpInit(bd, *pc);
-	*r = bd;
-	if (bif == BREAK_IS_ERL) {
-	    *pc = break_op;
-	}
-    } else {
-	ASSERT(*pc == (BeamInstr) BeamOp(op_i_trace_breakpoint) ||
-	       *pc == (BeamInstr) BeamOp(op_i_mtrace_breakpoint) ||
-	       *pc == (BeamInstr) BeamOp(op_i_debug_breakpoint) ||
-	       *pc == (BeamInstr) BeamOp(op_i_time_breakpoint) ||
-	       *pc == (BeamInstr) BeamOp(op_i_count_breakpoint) ||
-	       *pc == (BeamInstr) em_apply_bif);
-	if (*pc == (BeamInstr) BeamOp(op_i_debug_breakpoint)) {
-	    /* Debug bp must be last, so if it is also first; 
-	     * it must be singleton. */
-	    ASSERT(BpSingleton(*r));
-	    /* Insert new bp first in the ring, i.e second to last. */
-	    bd = Alloc(size);
-	    BpInitAndSpliceNext(bd, *pc, *r);
-	    if (bif == BREAK_IS_ERL) {
-		*pc = break_op;
-	    }
-	} else if ((*r)->prev->orig_instr
-		   == (BeamInstr) BeamOp(op_i_debug_breakpoint)) {
-	    /* Debug bp last in the ring; insert new second to last. */
-	    bd = Alloc(size);
-	    BpInitAndSplicePrev(bd, (*r)->prev->orig_instr, *r);
-	    (*r)->prev->orig_instr = break_op;
+	ASSERT((bp->flags & ~ERTS_BPF_ALL) == 0);
+	return;
+    } else if (common & ERTS_BPF_TIME_TRACE) {
+	BpDataTime* bdt = bp->time;
+	Uint i = 0;
+
+	if (count_op == erts_break_stop) {
+	    bp->flags &= ~ERTS_BPF_TIME_TRACE_ACTIVE;
 	} else {
-	    /* Just insert last in the ring */
-	    bd = Alloc(size);
-	    BpInitAndSpliceNext(bd, (*r)->orig_instr, *r);
-	    (*r)->orig_instr = break_op;
-	    *r = bd;
+	    bp->flags |= ERTS_BPF_TIME_TRACE_ACTIVE;
+	    for (i = 0; i < bdt->n; i++) {
+		bp_hash_delete(&(bdt->hash[i]));
+		bp_hash_init(&(bdt->hash[i]), 32);
+	    }
 	}
-    }
-    for (ix = 1; ix < erts_no_schedulers; ++ix) {
-	(*rs)[ix] = (*rs)[0];
+	ASSERT((bp->flags & ~ERTS_BPF_ALL) == 0);
+	return;
     }
 
-    bd->this_instr = break_op;
-    /* Init the bp type specific data */
-    if (break_op == (BeamInstr) BeamOp(op_i_trace_breakpoint) ||
-	break_op == (BeamInstr) BeamOp(op_i_mtrace_breakpoint)) {
-		
-	BpDataTrace *bdt = (BpDataTrace *) bd;
-		
-	MatchSetRef(match_spec);
-	bdt->match_spec = match_spec;
-	bdt->tracer_pid = tracer_pid;
-    } else if (break_op == (BeamInstr) BeamOp(op_i_time_breakpoint)) {
-	BpDataTime *bdt = (BpDataTime *) bd;
-	Uint i = 0;
-
-	bdt->pause = 0;
-	bdt->n     = erts_no_schedulers;
-	bdt->hash  = Alloc(sizeof(bp_time_hash_t)*(bdt->n));
+    /*
+     * Initialize the new breakpoint data.
+     */
 
+    if (break_flags & (ERTS_BPF_LOCAL_TRACE|ERTS_BPF_GLOBAL_TRACE)) {
+	MatchSetRef(match_spec);
+	bp->local_ms = match_spec;
+    } else if (break_flags & ERTS_BPF_META_TRACE) {
+	BpMetaPid* bmp;
+	MatchSetRef(match_spec);
+	bp->meta_ms = match_spec;
+	bmp = Alloc(sizeof(BpMetaPid));
+	erts_refc_init(&bmp->refc, 1);
+	erts_smp_atomic_init_nob(&bmp->pid, tracer_pid);
+	bp->meta_pid = bmp;
+    } else if (break_flags & ERTS_BPF_COUNT) {
+	BpCount* bcp;
+
+	ASSERT((bp->flags & ERTS_BPF_COUNT) == 0);
+	bcp = Alloc(sizeof(BpCount));
+	erts_refc_init(&bcp->refc, 1);
+	erts_smp_atomic_init_nob(&bcp->acount, 0);
+	bp->count = bcp;
+    } else if (break_flags & ERTS_BPF_TIME_TRACE) {
+	BpDataTime* bdt;
+	int i;
+
+	ASSERT((bp->flags & ERTS_BPF_TIME_TRACE) == 0);
+	bdt = Alloc(sizeof(BpDataTime));
+	erts_refc_init(&bdt->refc, 1);
+	bdt->n = erts_no_schedulers;
+	bdt->hash = Alloc(sizeof(bp_time_hash_t)*(bdt->n));
 	for (i = 0; i < bdt->n; i++) {
 	    bp_hash_init(&(bdt->hash[i]), 32);
 	}
-    } else if (break_op == (BeamInstr) BeamOp(op_i_count_breakpoint)) {
-	BpDataCount *bdc = (BpDataCount *) bd;
-	erts_smp_atomic_init_nob(&bdc->acount, 0);
+	bp->time = bdt;
     }
 
-    if (bif == BREAK_IS_ERL) {
-	++modp->curr.num_breakpoints;
+    bp->flags |= break_flags;
+    ASSERT((bp->flags & ~ERTS_BPF_ALL) == 0);
+}
+
+static void
+clear_break(BpFunctions* f, Uint break_flags)
+{
+    Uint i;
+    Uint n;
+
+    n = f->matched;
+    for (i = 0; i < n; i++) {
+	BeamInstr* pc = f->matching[i].pc;
+	clear_function_break(pc, break_flags);
     }
-    return 1;
 }
 
-static int clear_break(Eterm mfa[3], int specified, BeamInstr break_op)
+static int
+clear_function_break(BeamInstr *pc, Uint break_flags)
 {
-    ErtsCodeIndex code_ix = erts_active_code_ix();
-    int num_processed = 0;
-    Module *modp;
+    GenericBp* g;
+    GenericBpData* bp;
+    Uint common;
+    ErtsBpIndex ix = erts_staging_bp_ix();
 
-    if (!specified) {
-	/* Iterate over all modules */
-	int current;
-	int last = module_code_size(code_ix);
+    ERTS_SMP_LC_ASSERT(erts_is_code_ix_locked());
 
-	for (current = 0; current < last; current++) {
-	    modp = module_code(current, code_ix);
-	    ASSERT(modp != NULL);
-	    num_processed += clear_module_break(modp, mfa, specified, break_op);
-	}
-    } else {
-	/* Process a single module */
-	if ((modp = erts_get_module(mfa[0], code_ix)) != NULL) {
-	    num_processed += 
-		clear_module_break(modp, mfa, specified, break_op);
-	}	
+    if ((g = (GenericBp *) pc[-4]) == 0) {
+	return 1;
     }
-    return num_processed;
-}
 
-static int clear_module_break(Module *m, Eterm mfa[3], int specified, 
-			      BeamInstr break_op) {
-    BeamInstr** code_base;
-    BeamInstr* code_ptr;
-    int num_processed = 0;
-    Uint i;
-    BeamInstr n;
-    
-    ASSERT(m);
-    code_base = (BeamInstr **) m->curr.code;
-    if (code_base == NULL) {
-	return 0;
+    bp = &g->data[ix];
+    ASSERT((bp->flags & ~ERTS_BPF_ALL) == 0);
+    common = bp->flags & break_flags;
+    bp->flags &= ~break_flags;
+    if (common & (ERTS_BPF_LOCAL_TRACE|ERTS_BPF_GLOBAL_TRACE)) {
+	MatchSetUnref(bp->local_ms);
     }
-    n = (BeamInstr) code_base[MI_NUM_FUNCTIONS];
-    for (i = 0; i < n; ++i) {
-	code_ptr = code_base[MI_FUNCTIONS+i];
-	if ((specified < 2 || mfa[1] == ((Eterm) code_ptr[3])) &&
-	    (specified < 3 || ((int) mfa[2]) == ((int) code_ptr[4]))) {
-	    BeamInstr *pc = code_ptr + 5;
-	    
-	    num_processed += 
-		clear_function_break(m, pc, BREAK_IS_ERL, break_op);
-	}
+    if (common & ERTS_BPF_META_TRACE) {
+	MatchSetUnref(bp->meta_ms);
+    }
+    if (common & ERTS_BPF_COUNT) {
+	ASSERT((bp->flags & ERTS_BPF_COUNT_ACTIVE) == 0);
+	bp_count_unref(bp->count);
+    }
+    if (common & ERTS_BPF_TIME_TRACE) {
+	ASSERT((bp->flags & ERTS_BPF_TIME_TRACE_ACTIVE) == 0);
+	bp_time_unref(bp->time);
     }
-    return num_processed;
-}
 
-static int clear_function_break(Module *m, BeamInstr *pc, int bif, BeamInstr break_op) {
-    BpData *bd;
-    Uint ix = 0;
-    BeamInstr **code_base = NULL;
+    ASSERT((bp->flags & ~ERTS_BPF_ALL) == 0);
+    return 1;
+}
 
-    if (bif == BREAK_IS_ERL) {
-	code_base = (BeamInstr **)m->curr.code;
-	ASSERT(code_base);
-	ASSERT(code_base <= (BeamInstr **)pc);
-	ASSERT((BeamInstr **)pc < code_base + (m->curr.code_length/sizeof(BeamInstr *)));
-    } else {
-	ASSERT(*pc == (BeamInstr) em_apply_bif);
-	ASSERT(m == NULL);
+static void
+bp_meta_unref(BpMetaPid* bmp)
+{
+    if (erts_refc_dectest(&bmp->refc, 0) <= 0) {
+	Free(bmp);
     }
+}
 
-    /*
-     * Currently no trace support for native code.
-     */
-    if (erts_is_native_break(pc)) {
-	return 0;
+static void
+bp_count_unref(BpCount* bcp)
+{
+    if (erts_refc_dectest(&bcp->refc, 0) <= 0) {
+	Free(bcp);
     }
+}
 
-    while ( (bd = is_break(pc, break_op))) {
-	/* Remove all breakpoints of this type.
-	 * There should be only one of each type, 
-	 * but break_op may be 0 which matches any type. 
+static void
+bp_time_unref(BpDataTime* bdt)
+{
+    if (erts_refc_dectest(&bdt->refc, 0) <= 0) {
+	Uint i = 0;
+	Uint j = 0;
+	Process *h_p = NULL;
+	bp_data_time_item_t* item = NULL;
+	process_breakpoint_time_t* pbt = NULL;
+
+	/* remove all psd associated with the hash
+	 * and then delete the hash.
+	 * ... sigh ...
 	 */
-	BeamInstr op;
-	BpData ***rs = (BpData ***) (pc - 4);
-	BpData   **r = NULL;
 
-#ifdef DEBUG
-	for (ix = 1; ix < erts_no_schedulers; ++ix) {
-	    ASSERT((*rs)[ix] == (*rs)[0]);
-	}
-#endif
-	
-	r = &((*rs)[0]);
-
-	ASSERT(*r);
-	/* Find opcode for this breakpoint */
-	if (break_op) {
-	    op = break_op;
-	} else {
-	    if (bd == (*r)->next) {
-		/* First breakpoint in ring */
-		op = *pc;
-	    } else {
-		op = bd->prev->orig_instr;
-	    }
-	}
-	if (BpSingleton(bd)) {
-	    ASSERT(*r == bd);
-	    /* Only one breakpoint to remove */
-	    if (bif == BREAK_IS_ERL) {
-		*pc = bd->orig_instr;
-	    }
-	    Free(*rs);
-	    *rs = NULL;
-	} else {
-	    BpData *bd_prev = bd->prev;
-	    
-	    BpSpliceNext(bd, bd_prev);
-	    ASSERT(BpSingleton(bd));
-	    if (bd == *r) {
-		/* We removed the last breakpoint in the ring */
-		*r = bd_prev;
-		bd_prev->orig_instr = bd->orig_instr;
-	    } else if (bd_prev == *r) {
-		/* We removed the first breakpoint in the ring */
-		if (bif == BREAK_IS_ERL) {
-		    *pc = bd->orig_instr;
-		}
-	    } else {
-		bd_prev->orig_instr = bd->orig_instr;
-	    }
-	}
-	if (op == (BeamInstr) BeamOp(op_i_trace_breakpoint) ||
-	    op == (BeamInstr) BeamOp(op_i_mtrace_breakpoint)) {
-	    
-	    BpDataTrace *bdt = (BpDataTrace *) bd;
-	    MatchSetUnref(bdt->match_spec);
-	}
-	if (op == (BeamInstr) BeamOp(op_i_time_breakpoint)) {
-	    BpDataTime *bdt = (BpDataTime *) bd;
-	    Uint i = 0;
-	    Uint j = 0;
-	    Process *h_p = NULL;
-	    bp_data_time_item_t *item = NULL;
-	    process_breakpoint_time_t *pbt = NULL;
-
-	    /* remove all psd associated with the hash
-	     * and then delete the hash.
-	     * ... sigh ...
-	     */
-
-	    for( i = 0; i < bdt->n; ++i) {
-		if (bdt->hash[i].used) {
-		    for (j = 0; j < bdt->hash[i].n; ++j) {
-			item = &(bdt->hash[i].item[j]);
-			if (item->pid != NIL) {
-			    h_p = erts_proc_lookup(item->pid);
-			    if (h_p) {
-				pbt = ERTS_PROC_SET_CALL_TIME(h_p, ERTS_PROC_LOCK_MAIN, NULL);
-				if (pbt) {
-				    Free(pbt);
-				}
+	for (i = 0; i < bdt->n; ++i) {
+	    if (bdt->hash[i].used) {
+		for (j = 0; j < bdt->hash[i].n; ++j) {
+		    item = &(bdt->hash[i].item[j]);
+		    if (item->pid != NIL) {
+			h_p = erts_pid2proc(NULL, 0, item->pid,
+					    ERTS_PROC_LOCK_MAIN);
+			if (h_p) {
+			    pbt = ERTS_PROC_SET_CALL_TIME(h_p,
+							  ERTS_PROC_LOCK_MAIN,
+							  NULL);
+			    if (pbt) {
+				Free(pbt);
 			    }
+			    erts_smp_proc_unlock(h_p, ERTS_PROC_LOCK_MAIN);
 			}
 		    }
 		}
-		bp_hash_delete(&(bdt->hash[i]));
 	    }
-	    Free(bdt->hash);
-	    bdt->hash = NULL;
-	    bdt->n = 0;
+	    bp_hash_delete(&(bdt->hash[i]));
 	}
-	Free(bd);
-	if (bif == BREAK_IS_ERL) {
-	    ASSERT(m->curr.num_breakpoints > 0);
-	    --m->curr.num_breakpoints;
-	}
-	if (*rs) {
-	    for (ix = 1; ix < erts_no_schedulers; ++ix) {
-		(*rs)[ix] = (*rs)[0];
-	    }
-	}
-    } /* while bd != NULL */
-    return 1;
+	Free(bdt->hash);
+	Free(bdt);
+    }
 }
 
-
-
-/*
-** Searches (linear forward) the breakpoint ring for a specified opcode
-** and returns a pointer to the breakpoint data structure or NULL if
-** not found. If the specified opcode is 0, the last breakpoint is 
-** returned. The program counter must point to the first executable
-** (breakpoint) instruction of the function.
-*/
-
-BpData *erts_get_time_break(Process *p, BeamInstr *pc) {
-    return get_break(p, pc, (BeamInstr) BeamOp(op_i_time_breakpoint));
+static BpDataTime*
+get_time_break(BeamInstr *pc)
+{
+    GenericBpData* bp = check_break(pc, ERTS_BPF_TIME_TRACE);
+    return bp ? bp->time : 0;
 }
 
-static BpData *get_break(Process *p, BeamInstr *pc, BeamInstr break_op) {
-    ASSERT(pc[-5] == (BeamInstr) BeamOp(op_i_func_info_IaaI));
-    if (! erts_is_native_break(pc)) {
-	BpData **rs = (BpData **) pc[-4];
-	BpData  *bd = NULL, *ebd = NULL;
-
-	if (! rs) {
-	    return NULL;
-	}
-
-	bd = ebd = rs[bp_sched2ix_proc(p)];
-	ASSERT(bd);
-	if (bd->this_instr == break_op) {
-	    return bd;
-	}
-
-	bd = bd->next;
-	while (bd != ebd) {
-	    ASSERT(bd);
-	    if (bd->this_instr == break_op) {
-		ASSERT(bd);
-		return bd;
-	    }
-	    bd = bd->next;
-	}
-    }
-    return NULL;
-}
+static GenericBpData*
+check_break(BeamInstr *pc, Uint break_flags)
+{
+    GenericBp* g = (GenericBp *) pc[-4];
 
-static BpData *is_break(BeamInstr *pc, BeamInstr break_op) {
-    BpData **rs;
-    BpData  *bd = NULL, *ebd = NULL;
     ASSERT(pc[-5] == (BeamInstr) BeamOp(op_i_func_info_IaaI));
-
     if (erts_is_native_break(pc)) {
-	return NULL;
-    }
-    rs = (BpData **) pc[-4];
-    if (! rs) {
-	return NULL;
-    }
-
-    bd = ebd = rs[erts_bp_sched2ix()];
-    ASSERT(bd);
-    if ( (break_op == 0) || (bd->this_instr == break_op)) {
-	return bd;
+	return 0;
     }
-
-    bd = bd->next;
-    while (bd != ebd) {
-	ASSERT(bd);
-	if (bd->this_instr == break_op) {
-	    ASSERT(bd);
-	    return bd;
+    if (g) {
+	GenericBpData* bp = &g->data[erts_active_bp_ix()];
+	ASSERT((bp->flags & ~ERTS_BPF_ALL) == 0);
+	if (bp->flags & break_flags) {
+	    return bp;
 	}
-	bd = bd->next;
     }
-    return NULL;
+    return 0;
 }
diff --git a/erts/emulator/beam/beam_bp.h b/erts/emulator/beam/beam_bp.h
index 167069552f..28aaaa462a 100644
--- a/erts/emulator/beam/beam_bp.h
+++ b/erts/emulator/beam/beam_bp.h
@@ -25,77 +25,6 @@
 #include "erl_vm.h"
 #include "global.h"
 
-
-
-/* A couple of gotchas:
- *
- * The breakpoint structure from BeamInstr,
- * In beam_emu where the instruction counter pointer, I (or pc),
- * points to the *current* instruction. At that time, if the instruction
- * is a breakpoint instruction the pc looks like the following,
- *
- * I[-5]   | op_i_func_info_IaaI |        scheduler specific entries
- * I[-4]   |    BpData** bpa     |  --> | BpData * bdas1 | ... | BpData * bdasN |
- * I[-3]   |  Tagged Module      |            |                    |
- * I[-2]   |  Tagged Function    |            V                    V
- * I[-1]   |  Arity              |          BpData -> BpData -> BpData -> BpData
- * I[0]    | The bp instruction  |            ^       * the bp wheel *      |
- *                                            |------------------------------
- *
- * Common struct to all bp_data_*
- *
- * 1) The type of bp_data structure in the ring is deduced from the
- *    orig_instr field of the structure _before_ in the ring, except for
- *    the first structure in the ring that has its instruction in
- *    pc[0] of the code to execute.
- *    This is valid as long as you don't search for the function while it is
- *    being executed by something else. Or is in the middle of its rotation for
- *    any other reason.
- *    A key, the bp beam instruction, is included for this reason.
- *
- * 2) pc[-4][sched_id - 1] points to the _last_ structure in the ring before the
- *    breakpoints are being executed.
- *
- * So, as an example, when a breakpointed function starts to execute,
- * the first instruction that is a breakpoint instruction at pc[0] finds
- * its data at ((BpData **) pc[-4][sched_id - 1])->next and has to cast that pointer
- * to the correct bp_data type.
-*/
-
-typedef struct bp_data {
-    struct bp_data *next; /* Doubly linked ring pointers */
-    struct bp_data *prev; /* -"-                         */
-    BeamInstr orig_instr;      /* The original instruction to execute */
-    BeamInstr this_instr;      /* key */
-} BpData;
-/*
-** All the following bp_data_.. structs must begin the same way
-*/
-
-typedef struct bp_data_trace {
-    struct bp_data *next;
-    struct bp_data *prev;
-    BeamInstr       orig_instr;
-    BeamInstr       this_instr;   /* key */
-    Binary         *match_spec;
-    Eterm           tracer_pid;
-} BpDataTrace;
-
-typedef struct bp_data_debug {
-    struct bp_data *next;
-    struct bp_data *prev;
-    BeamInstr       orig_instr;
-    BeamInstr       this_instr;   /* key */
-} BpDataDebug;
-
-typedef struct bp_data_count {    /* Call count */
-    struct bp_data *next;
-    struct bp_data *prev;
-    BeamInstr       orig_instr;
-    BeamInstr       this_instr;   /* key */
-    erts_smp_atomic_t acount;
-} BpDataCount;
-
 typedef struct {
     Eterm pid;
     Sint  count;
@@ -110,13 +39,9 @@ typedef struct {
 } bp_time_hash_t;
 
 typedef struct bp_data_time {     /* Call time */
-    struct bp_data *next;
-    struct bp_data *prev;
-    BeamInstr       orig_instr;
-    BeamInstr       this_instr;   /* key */
-    Uint	    pause;
-    Uint	    n;
-    bp_time_hash_t  *hash;
+    Uint n;
+    bp_time_hash_t *hash;
+    erts_refc_t refc;
 } BpDataTime;
 
 typedef struct {
@@ -126,64 +51,42 @@ typedef struct {
     BeamInstr *pc;
 } process_breakpoint_time_t; /* used within psd */
 
-extern erts_smp_spinlock_t erts_bp_lock;
+typedef struct {
+    erts_smp_atomic_t acount;
+    erts_refc_t refc;
+} BpCount;
+
+typedef struct {
+    erts_smp_atomic_t pid;
+    erts_refc_t refc;
+} BpMetaPid;
+
+typedef struct generic_bp_data {
+    Uint flags;
+    Binary* local_ms;		/* Match spec for local call trace */
+    Binary* meta_ms;		/* Match spec for meta trace */
+    BpMetaPid* meta_pid;	/* Meta trace pid */
+    BpCount* count;		/* For call count */
+    BpDataTime* time;		/* For time trace */
+} GenericBpData;
+
+#define ERTS_NUM_BP_IX 2
+
+typedef struct generic_bp {
+    BeamInstr orig_instr;
+    GenericBpData data[ERTS_NUM_BP_IX];
+} GenericBp;
 
 #define ERTS_BP_CALL_TIME_SCHEDULE_IN      (0)
 #define ERTS_BP_CALL_TIME_SCHEDULE_OUT     (1)
 #define ERTS_BP_CALL_TIME_SCHEDULE_EXITING (2)
 
-#define ERTS_BP_CALL_TIME_CALL      (0)
-#define ERTS_BP_CALL_TIME_RETURN    (1)
-#define ERTS_BP_CALL_TIME_TAIL_CALL (2)
-
-#ifdef ERTS_SMP
-#define ErtsSmpBPLock(BDC) erts_smp_spin_lock(&erts_bp_lock)
-#define ErtsSmpBPUnlock(BDC) erts_smp_spin_unlock(&erts_bp_lock)
-#else
-#define ErtsSmpBPLock(BDC)
-#define ErtsSmpBPUnlock(BDC)
-#endif
-
 #ifdef ERTS_SMP
 #define bp_sched2ix_proc(p) ((p)->scheduler_data->no - 1)
 #else
 #define bp_sched2ix_proc(p) (0)
 #endif
 
-#define ErtsCountBreak(p, pc,instr_result)                       \
-do {                                                             \
-    BpData **bds = (BpData **) (pc)[-4];                         \
-    BpDataCount *bdc = NULL;                                     \
-    Uint ix = bp_sched2ix_proc( (p) );                           \
-    erts_aint_t count = 0;                                       \
-                                                                 \
-    ASSERT((pc)[-5] == (BeamInstr) BeamOp(op_i_func_info_IaaI)); \
-    ASSERT(bds);                                                 \
-    bdc = (BpDataCount *) bds[ix];                               \
-    bdc = (BpDataCount *) bdc->next;                             \
-    ASSERT(bdc);                                                 \
-    bds[ix] = (BpData *) bdc;                                    \
-    count = erts_smp_atomic_read_nob(&bdc->acount);                  \
-    if (count >= 0)  erts_smp_atomic_inc_nob(&bdc->acount);      \
-    *(instr_result) = bdc->orig_instr;                           \
-} while (0)
-
-#define ErtsBreakSkip(p, pc,instr_result)                        \
-do {                                                             \
-    BpData **bds = (BpData **) (pc)[-4];                         \
-    BpData *bd = NULL;                                           \
-    Uint ix = bp_sched2ix_proc( (p) );                           \
-                                                                 \
-    ASSERT((pc)[-5] == (BeamInstr) BeamOp(op_i_func_info_IaaI)); \
-    ASSERT(bds);                                                 \
-    bd = bds[ix];                                                \
-    ASSERT(bd);                                                  \
-    bd = bd->next;                                               \
-    ASSERT(bd);                                                  \
-    bds[ix] = bd;                                                \
-    *(instr_result) = bd->orig_instr;                            \
-} while (0)
-
 enum erts_break_op{
     erts_break_nop   =  0, /* Must be false */
     erts_break_set   = !0, /* Must be true */
@@ -191,7 +94,17 @@ enum erts_break_op{
     erts_break_stop
 };
 
+typedef Uint32 ErtsBpIndex;
 
+typedef struct {
+    BeamInstr* pc;
+    Module* mod;
+} BpFunction;
+
+typedef struct {
+    Uint matched;		/* Number matched */
+    BpFunction* matching;	/* Matching functions */
+} BpFunctions;
 
 /*
 ** Function interface exported from beam_bp.c
@@ -199,49 +112,66 @@ enum erts_break_op{
 
 void erts_bp_init(void);
 
-int erts_set_trace_break(Eterm mfa[3], int specified, Binary *match_spec,
-			 Eterm tracer_pid);
-int erts_clear_trace_break(Eterm mfa[3], int specified);
-int erts_set_mtrace_break(Eterm mfa[3], int specified, Binary *match_spec,
+void erts_prepare_bp_staging(void);
+void erts_commit_staged_bp(void);
+
+ERTS_GLB_INLINE ErtsBpIndex erts_active_bp_ix(void);
+ERTS_GLB_INLINE ErtsBpIndex erts_staging_bp_ix(void);
+
+void erts_bp_match_functions(BpFunctions* f, Eterm mfa[3], int specified);
+void erts_bp_match_export(BpFunctions* f, Eterm mfa[3], int specified);
+void erts_bp_free_matched_functions(BpFunctions* f);
+
+void erts_install_breakpoints(BpFunctions* f);
+void erts_uninstall_breakpoints(BpFunctions* f);
+void erts_consolidate_bp_data(BpFunctions* f, int local);
+void erts_consolidate_bif_bp_data(void);
+
+void erts_set_trace_break(BpFunctions *f, Binary *match_spec);
+void erts_clear_trace_break(BpFunctions *f);
+
+void erts_set_call_trace_bif(BeamInstr *pc, Binary *match_spec, int local);
+void erts_clear_call_trace_bif(BeamInstr *pc, int local);
+
+void erts_set_mtrace_break(BpFunctions *f, Binary *match_spec,
 			  Eterm tracer_pid);
-int erts_clear_mtrace_break(Eterm mfa[3], int specified);
+void erts_clear_mtrace_break(BpFunctions *f);
 void erts_set_mtrace_bif(BeamInstr *pc, Binary *match_spec,
 			 Eterm tracer_pid);
 void erts_clear_mtrace_bif(BeamInstr *pc);
-int erts_set_debug_break(Eterm mfa[3], int specified);
-int erts_clear_debug_break(Eterm mfa[3], int specified);
-int erts_set_count_break(Eterm mfa[3], int specified, enum erts_break_op);
-int erts_clear_count_break(Eterm mfa[3], int specified);
 
+void erts_set_debug_break(BpFunctions *f);
+void erts_clear_debug_break(BpFunctions *f);
+void erts_set_count_break(BpFunctions *f, enum erts_break_op);
+void erts_clear_count_break(BpFunctions *f);
 
-int erts_clear_break(Eterm mfa[3], int specified);
+
+void erts_clear_all_breaks(BpFunctions* f);
 int erts_clear_module_break(Module *modp);
-int erts_clear_function_break(Module *modp, BeamInstr *pc);
+void erts_clear_export_break(Module *modp, BeamInstr* pc);
 
+BeamInstr erts_generic_breakpoint(Process* c_p, BeamInstr* I, Eterm* reg);
 BeamInstr erts_trace_break(Process *p, BeamInstr *pc, Eterm *args,
 		      Uint32 *ret_flags, Eterm *tracer_pid);
-Uint32 erts_bif_mtrace(Process *p, BeamInstr *pc, Eterm *args,
-		       int local, Eterm *tracer_pid);
 
-int erts_is_trace_break(BeamInstr *pc, Binary **match_spec_ret,
-			Eterm *tracer_pid_ret);
+int erts_is_trace_break(BeamInstr *pc, Binary **match_spec_ret, int local);
 int erts_is_mtrace_break(BeamInstr *pc, Binary **match_spec_ret,
 			 Eterm *tracer_pid_rte);
 int erts_is_mtrace_bif(BeamInstr *pc, Binary **match_spec_ret,
 		       Eterm *tracer_pid_ret);
 int erts_is_native_break(BeamInstr *pc);
-int erts_is_count_break(BeamInstr *pc, Sint *count_ret);
+int erts_is_count_break(BeamInstr *pc, Uint *count_ret);
 int erts_is_time_break(Process *p, BeamInstr *pc, Eterm *call_time);
 
-void erts_trace_time_break(Process *p, BeamInstr *pc, BpDataTime *bdt, Uint type);
+void erts_trace_time_call(Process* c_p, BeamInstr* pc, BpDataTime* bdt);
+void erts_trace_time_return(Process* c_p, BeamInstr* pc);
 void erts_schedule_time_break(Process *p, Uint out);
-int erts_set_time_break(Eterm mfa[3], int specified, enum erts_break_op);
-int erts_clear_time_break(Eterm mfa[3], int specified);
+void erts_set_time_break(BpFunctions *f, enum erts_break_op);
+void erts_clear_time_break(BpFunctions *f);
 
 int erts_is_time_trace_bif(Process *p, BeamInstr *pc, Eterm *call_time);
 void erts_set_time_trace_bif(BeamInstr *pc, enum erts_break_op);
 void erts_clear_time_trace_bif(BeamInstr *pc);
-BpData *erts_get_time_break(Process *p, BeamInstr *pc);
 
 BeamInstr *erts_find_local_func(Eterm mfa[3]);
 
@@ -258,6 +188,19 @@ ERTS_GLB_INLINE Uint erts_bp_sched2ix(void)
     return 0;
 #endif
 }
+
+extern erts_smp_atomic32_t erts_active_bp_index;
+extern erts_smp_atomic32_t erts_staging_bp_index;
+
+ERTS_GLB_INLINE ErtsBpIndex erts_active_bp_ix(void)
+{
+    return erts_smp_atomic32_read_nob(&erts_active_bp_index);
+}
+
+ERTS_GLB_INLINE ErtsBpIndex erts_staging_bp_ix(void)
+{
+    return erts_smp_atomic32_read_nob(&erts_staging_bp_index);
+}
 #endif
 
 #endif /* _BEAM_BP_H */
diff --git a/erts/emulator/beam/beam_debug.c b/erts/emulator/beam/beam_debug.c
index e69cbc3048..a609ed8c71 100644
--- a/erts/emulator/beam/beam_debug.c
+++ b/erts/emulator/beam/beam_debug.c
@@ -84,6 +84,7 @@ erts_debug_breakpoint_2(BIF_ALIST_2)
     int i;
     int specified = 0;
     Eterm res;
+    BpFunctions f;
 
     if (bool != am_true && bool != am_false)
 	goto error;
@@ -121,11 +122,19 @@ erts_debug_breakpoint_2(BIF_ALIST_2)
     erts_smp_proc_unlock(p, ERTS_PROC_LOCK_MAIN);
     erts_smp_thr_progress_block();
 
+    erts_bp_match_functions(&f, mfa, specified);
     if (bool == am_true) {
-	res = make_small(erts_set_debug_break(mfa, specified));
+	erts_set_debug_break(&f);
+	erts_install_breakpoints(&f);
+	erts_commit_staged_bp();
     } else {
-	res = make_small(erts_clear_debug_break(mfa, specified));
+	erts_clear_debug_break(&f);
+	erts_commit_staged_bp();
+	erts_uninstall_breakpoints(&f);
     }
+    erts_consolidate_bp_data(&f, 1);
+    res = make_small(f.matched);
+    erts_bp_free_matched_functions(&f);
 
     erts_smp_thr_progress_unblock();
     erts_smp_proc_lock(p, ERTS_PROC_LOCK_MAIN);
diff --git a/erts/emulator/beam/beam_emu.c b/erts/emulator/beam/beam_emu.c
index 15a937289d..efce070061 100644
--- a/erts/emulator/beam/beam_emu.c
+++ b/erts/emulator/beam/beam_emu.c
@@ -217,7 +217,6 @@ BeamInstr beam_continue_exit[1];
 
 BeamInstr* em_call_error_handler;
 BeamInstr* em_apply_bif;
-BeamInstr* em_call_traced_function;
 
 
 /* NOTE These should be the only variables containing trace instructions.
@@ -4571,64 +4570,6 @@ void process_main(void)
      * Trace and debugging support.
      */
 
-    /*
-     * At this point, I points to the code[3] in the export entry for
-     * a trace-enabled function.
-     *
-     * code[0]: Module
-     * code[1]: Function
-     * code[2]: Arity
-     * code[3]: &&call_traced_function
-     * code[4]: Address of function.
-     */
- OpCase(call_traced_function): {
-     if (IS_TRACED_FL(c_p, F_TRACE_CALLS)) {
-	 unsigned offset = offsetof(Export, code) + 3*sizeof(BeamInstr);
-	 Export* ep = (Export *) (((char *)I)-offset);
-	 Uint32 flags;
-
-	 SWAPOUT;
-	 reg[0] = r(0);
-	 PROCESS_MAIN_CHK_LOCKS(c_p);
-	 ERTS_SMP_UNREQ_PROC_MAIN_LOCK(c_p);
-	 flags = erts_call_trace(c_p, ep->code, ep->match_prog_set, reg,
-				 0, &c_p->tracer_proc);
-	 ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
-	 ERTS_SMP_REQ_PROC_MAIN_LOCK(c_p);
-	 PROCESS_MAIN_CHK_LOCKS(c_p);
-	 ASSERT(!ERTS_PROC_IS_EXITING(c_p));
-	 SWAPIN;
-	 
-	 if (flags & MATCH_SET_RX_TRACE) {
-	     ASSERT(c_p->htop <= E && E <= c_p->hend);
-	     if (E - 3 < HTOP) {
-		 /* SWAPOUT, SWAPIN was done and r(0) was saved above */
-		 PROCESS_MAIN_CHK_LOCKS(c_p);
-		 FCALLS -= erts_garbage_collect(c_p, 3, reg, ep->code[2]);
-		 ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
-		 PROCESS_MAIN_CHK_LOCKS(c_p);
-		 r(0) = reg[0];
-		 SWAPIN;
-	     }
-	     E -= 3;
-	     ASSERT(c_p->htop <= E && E <= c_p->hend);
-	     ASSERT(is_CP((BeamInstr)(ep->code)));
-	     ASSERT(is_internal_pid(c_p->tracer_proc) || 
-		    is_internal_port(c_p->tracer_proc));
-	     E[2] = make_cp(c_p->cp); /* Code in lower range on halfword */
-	     E[1] = am_true; /* Process tracer */
-	     E[0] = make_cp(ep->code);
-	     c_p->cp = (flags & MATCH_SET_EXCEPTION_TRACE)
-		 ? beam_exception_trace : beam_return_trace;
-	     erts_smp_proc_lock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
-	     c_p->trace_flags |= F_EXCEPTION_TRACE;
-	     erts_smp_proc_unlock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
-	 }
-     }
-     SET_I((BeamInstr *)Arg(0));
-     Dispatch();
- }
-
  OpCase(return_trace): {
      BeamInstr* code = (BeamInstr *) (UWord) E[0];
      
@@ -4643,80 +4584,22 @@ void process_main(void)
      Goto(*I);
  }
 
- OpCase(i_count_breakpoint): {
+ OpCase(i_generic_breakpoint): {
      BeamInstr real_I;
-     
-     ErtsCountBreak(c_p, (BeamInstr *) I, &real_I);
-     ASSERT(VALID_INSTR(real_I));
-     Goto(real_I);
- }
-
- /* need to send mfa instead of bdt pointer
-  * the pointer might be deallocated.
-  */
-
- OpCase(i_time_breakpoint): {
-     BeamInstr real_I;
-     BpData **bds = (BpData **) (I)[-4];
-     BpDataTime *bdt = NULL;
-     Uint ix = 0;
-#ifdef ERTS_SMP
-     ix = c_p->scheduler_data->no - 1;
-#else
-     ix = 0;
-#endif
-     bdt = (BpDataTime *)bds[ix];
-
-     ASSERT((I)[-5] == (BeamInstr) BeamOp(op_i_func_info_IaaI));
-     ASSERT(bdt);
-     bdt = (BpDataTime *) bdt->next;
-     ASSERT(bdt);
-     bds[ix] = (BpData *) bdt;
-     real_I = bdt->orig_instr;
+     ASSERT(I[-5] == (BeamInstr) BeamOp(op_i_func_info_IaaI));
+     SWAPOUT;
+     reg[0] = r(0);
+     real_I = erts_generic_breakpoint(c_p, I, reg);
+     r(0) = reg[0];
+     SWAPIN;
      ASSERT(VALID_INSTR(real_I));
-
-     if (IS_TRACED_FL(c_p, F_TRACE_CALLS) && !(bdt->pause)) {
-	 if (	(*(c_p->cp) == (BeamInstr) OpCode(i_return_time_trace)) ||
-		(*(c_p->cp) == (BeamInstr) OpCode(return_trace)) ||
-		(*(c_p->cp) == (BeamInstr) OpCode(i_return_to_trace))) {
-	     /* This _IS_ a tail recursive call */
-	     SWAPOUT;
-	     erts_trace_time_break(c_p, I, bdt, ERTS_BP_CALL_TIME_TAIL_CALL);
-	     SWAPIN;
-	 } else {
-	     SWAPOUT;
-	     erts_trace_time_break(c_p, I, bdt, ERTS_BP_CALL_TIME_CALL);
-
-	     /* r register needs to be copied to the array
-	      * for the garbage collector
-	      */
-	     ASSERT(c_p->htop <= E && E <= c_p->hend);
-	     if (E - 2 < HTOP) {
-		 reg[0] = r(0);
-		 PROCESS_MAIN_CHK_LOCKS(c_p);
-		 FCALLS -= erts_garbage_collect(c_p, 2, reg, I[-1]);
-		 ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
-		 PROCESS_MAIN_CHK_LOCKS(c_p);
-		 r(0) = reg[0];
-	     }
-	     SWAPIN;
-
-	     ASSERT(c_p->htop <= E && E <= c_p->hend);
-
-	     E -= 2;
-	     E[0] = make_cp(I);
-	     E[1] = make_cp(c_p->cp);     /* original return address */
-	     c_p->cp = beam_return_time_trace;
-	 }
-     }
-
      Goto(real_I);
  }
 
  OpCase(i_return_time_trace): {
      BeamInstr *pc = (BeamInstr *) (UWord) E[0];
      SWAPOUT;
-     erts_trace_time_break(c_p, pc, NULL, ERTS_BP_CALL_TIME_RETURN);
+     erts_trace_time_return(c_p, pc);
      SWAPIN;
      c_p->cp = NULL;
      SET_I((BeamInstr *) cp_val(E[1]));
@@ -4724,114 +4607,6 @@ void process_main(void)
      Goto(*I);
  }
 
- OpCase(i_trace_breakpoint):
-     if (! IS_TRACED_FL(c_p, F_TRACE_CALLS)) {
-	 BeamInstr real_I;
-	 
-	 ErtsBreakSkip(c_p, (BeamInstr *) I, &real_I);
-	 Goto(real_I);
-     }
- /* Fall through to next case */
- OpCase(i_mtrace_breakpoint): {
-     BeamInstr real_I;
-     Uint32 flags;
-     Eterm tracer_pid;
-     Uint* cpp;
-     int return_to_trace = 0, need = 0;
-     flags = 0;
-     SWAPOUT;
-     reg[0] = r(0);
-
-     if (*(c_p->cp) == (BeamInstr) OpCode(return_trace)) {
-	 cpp = &E[2];
-     } else if (*(c_p->cp) == (BeamInstr) OpCode(i_return_to_trace)) {
-	 return_to_trace = !0;
-	 cpp = &E[0];
-     } else if (*(c_p->cp) == (BeamInstr) OpCode(i_return_time_trace)) {
-	 return_to_trace = !0;
-	 cpp = &E[0];
-     } else {
-	 cpp = NULL;
-     }
-     if (cpp) {
-	 /* This _IS_ a tail recursive call, if there are
-	  * return_trace and/or i_return_to_trace stackframes
-	  * on the stack, they are not intermixed with y registers
-	  */
-	 BeamInstr *cp_save = c_p->cp;
-	 for (;;) {
-	     ASSERT(is_CP(*cpp));
-	     if (*cp_val(*cpp) == (BeamInstr) OpCode(return_trace)) {
-		 cpp += 3;
-	     } else if (*cp_val(*cpp) == (BeamInstr) OpCode(i_return_to_trace)) {
-		 return_to_trace = !0;
-		 cpp += 1;
-	     } else if (*cp_val(*cpp) == (BeamInstr) OpCode(i_return_time_trace)) {
-		 cpp += 2;
-	     } else
-		 break;
-	 }
-	 c_p->cp = (BeamInstr *) cp_val(*cpp);
-	 ASSERT(is_CP(*cpp));
-	 ERTS_SMP_UNREQ_PROC_MAIN_LOCK(c_p);
-	 real_I = erts_trace_break(c_p, I, reg, &flags, &tracer_pid);
-	 ERTS_SMP_REQ_PROC_MAIN_LOCK(c_p);
-	 SWAPIN;		/* Needed by shared heap. */
-	 c_p->cp = cp_save;
-     } else {
-	 ERTS_SMP_UNREQ_PROC_MAIN_LOCK(c_p);
-	 real_I = erts_trace_break(c_p, I, reg, &flags, &tracer_pid);
-	 ERTS_SMP_REQ_PROC_MAIN_LOCK(c_p);
-	 SWAPIN;		/* Needed by shared heap. */
-     }
-
-     ASSERT(!ERTS_PROC_IS_EXITING(c_p));
-
-     if ((flags & MATCH_SET_RETURN_TO_TRACE) && !return_to_trace) {
-	 need += 1;
-     }
-     if (flags & MATCH_SET_RX_TRACE) {
-	 need += 3;
-     }
-     if (need) {
-	 ASSERT(c_p->htop <= E && E <= c_p->hend);
-	 if (E - need < HTOP) {
-	     /* SWAPOUT was done and r(0) was saved above */
-	     PROCESS_MAIN_CHK_LOCKS(c_p);
-	     FCALLS -= erts_garbage_collect(c_p, need, reg, I[-1]);
-	     ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
-	     PROCESS_MAIN_CHK_LOCKS(c_p);
-	     r(0) = reg[0];
-	     SWAPIN;
-	 }
-     }
-     if ((flags & MATCH_SET_RETURN_TO_TRACE) && !return_to_trace) {
-	 E -= 1;
-	 ASSERT(c_p->htop <= E && E <= c_p->hend);
-	 E[0] = make_cp(c_p->cp);
-	 c_p->cp = (BeamInstr *) beam_return_to_trace;
-     }
-     if (flags & MATCH_SET_RX_TRACE) {
-	 E -= 3;
-	 ASSERT(c_p->htop <= E && E <= c_p->hend);
-	 ASSERT(is_CP((Eterm) (UWord) (I - 3)));
-	 ASSERT(am_true == tracer_pid || 
-		is_internal_pid(tracer_pid) || is_internal_port(tracer_pid));
-	 E[2] = make_cp(c_p->cp);
-	 E[1] = tracer_pid;
-	 E[0] = make_cp(I - 3); /* We ARE at the beginning of an 
-				   instruction,
-				   the funcinfo is above i. */
-	 c_p->cp =
-	     (flags & MATCH_SET_EXCEPTION_TRACE)
-		     ? beam_exception_trace : beam_return_trace;
-	 erts_smp_proc_lock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
-	 c_p->trace_flags |= F_EXCEPTION_TRACE;
-	 erts_smp_proc_unlock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
-     }
-     Goto(real_I);
- }
- 
  OpCase(i_return_to_trace): {
      if (IS_TRACED_FL(c_p, F_TRACE_RETURN_TO)) {
 	 Uint *cpp = (Uint*) E;
@@ -5181,7 +4956,6 @@ void process_main(void)
 #endif /* NO_JUMP_TABLE */
      
      em_call_error_handler = OpCode(call_error_handler);
-     em_call_traced_function = OpCode(call_traced_function);
      em_apply_bif = OpCode(apply_bif);
 
      beam_apply[0]             = (BeamInstr) OpCode(i_apply);
diff --git a/erts/emulator/beam/beam_load.h b/erts/emulator/beam/beam_load.h
index f1506a8684..1048f258a5 100644
--- a/erts/emulator/beam/beam_load.h
+++ b/erts/emulator/beam/beam_load.h
@@ -49,7 +49,6 @@ extern void** beam_ops;
 extern BeamInstr beam_debug_apply[];
 extern BeamInstr* em_call_error_handler;
 extern BeamInstr* em_apply_bif;
-extern BeamInstr* em_call_traced_function;
 
 /*
  * The following variables keep a sorted list of address ranges for
diff --git a/erts/emulator/beam/code_ix.c b/erts/emulator/beam/code_ix.c
index ae4cca1e58..8025058ee0 100644
--- a/erts/emulator/beam/code_ix.c
+++ b/erts/emulator/beam/code_ix.c
@@ -44,6 +44,10 @@ struct code_ix_queue_item {
 static struct code_ix_queue_item* the_code_ix_queue = NULL;
 static erts_smp_mtx_t the_code_ix_queue_lock;
 
+#ifdef ERTS_ENABLE_LOCK_CHECK
+static erts_tsd_key_t has_code_write_permission;
+#endif
+
 void erts_code_ix_init(void)
 {
     /* We start emulator by initializing preloaded modules
@@ -53,6 +57,9 @@ void erts_code_ix_init(void)
     erts_smp_atomic32_init_nob(&the_active_code_index, 0);
     erts_smp_atomic32_init_nob(&the_staging_code_index, 0);
     erts_smp_mtx_init(&the_code_ix_queue_lock, "code_ix_queue");
+#ifdef ERTS_ENABLE_LOCK_CHECK
+    erts_tsd_key_create(&has_code_write_permission);
+#endif
     CIX_TRACE("init");
 }
 
@@ -112,6 +119,9 @@ int erts_try_seize_code_write_permission(Process* c_p)
     success = !the_code_ix_lock;
     if (success) {
 	the_code_ix_lock = 1;
+#ifdef ERTS_ENABLE_LOCK_CHECK
+	erts_tsd_set(has_code_write_permission, (void *) 1);
+#endif
     }
     else { /* Already locked */
 	struct code_ix_queue_item* qitem;
@@ -128,6 +138,7 @@ int erts_try_seize_code_write_permission(Process* c_p)
 
 void erts_release_code_write_permission(void)
 {
+    ERTS_SMP_LC_ASSERT(erts_is_code_ix_locked());
     erts_smp_mtx_lock(&the_code_ix_queue_lock);
     while (the_code_ix_queue != NULL) { /* unleash the entire herd */
 	struct code_ix_queue_item* qitem = the_code_ix_queue;
@@ -141,12 +152,15 @@ void erts_release_code_write_permission(void)
 	erts_free(ERTS_ALC_T_CODE_IX_LOCK_Q, qitem);
     }
     the_code_ix_lock = 0;
+#ifdef ERTS_ENABLE_LOCK_CHECK
+    erts_tsd_set(has_code_write_permission, (void *) 0);
+#endif
     erts_smp_mtx_unlock(&the_code_ix_queue_lock);
 }
 
 #ifdef ERTS_ENABLE_LOCK_CHECK
 int erts_is_code_ix_locked(void)
 {
-    return the_code_ix_lock;
+    return the_code_ix_lock && erts_tsd_get(has_code_write_permission);
 }
 #endif
diff --git a/erts/emulator/beam/erl_bif_trace.c b/erts/emulator/beam/erl_bif_trace.c
index 7f1b02b9b4..e88fb8c9f4 100644
--- a/erts/emulator/beam/erl_bif_trace.c
+++ b/erts/emulator/beam/erl_bif_trace.c
@@ -42,12 +42,24 @@
 #define DECL_AM(S) Eterm AM_ ## S = am_atom_put(#S, sizeof(#S) - 1)
 
 const struct trace_pattern_flags   erts_trace_pattern_flags_off = {0, 0, 0, 0, 0};
+
+/*
+ * The following variables are protected by code write permission.
+ */
 static int                         erts_default_trace_pattern_is_on;
 static Binary                     *erts_default_match_spec;
 static Binary                     *erts_default_meta_match_spec;
 static struct trace_pattern_flags  erts_default_trace_pattern_flags;
 static Eterm                       erts_default_meta_tracer_pid;
 
+static struct {			/* Protected by code write permission */
+    int current;
+    int install;
+    int local;
+    BpFunctions f;		/* Local functions */
+    BpFunctions e;		/* Export entries */
+} finish_bp;
+
 static Eterm
 trace_pattern(Process* p, Eterm MFA, Eterm Pattern, Eterm flaglist);
 static BIF_RETTYPE
@@ -60,12 +72,11 @@ static Eterm trace_info_pid(Process* p, Eterm pid_spec, Eterm key);
 static Eterm trace_info_func(Process* p, Eterm pid_spec, Eterm key);
 static Eterm trace_info_on_load(Process* p, Eterm key);
 
-static int setup_func_trace(Export* ep, void* match_prog, ErtsCodeIndex);
-static int reset_func_trace(Export* ep, ErtsCodeIndex);
-static void reset_bif_trace(int bif_index);
-static void setup_bif_trace(int bif_index);
-static void set_trace_bif(int bif_index, void* match_prog);
-static void clear_trace_bif(int bif_index);
+static void reset_bif_trace(void);
+static void setup_bif_trace(void);
+static void install_exp_breakpoints(BpFunctions* f);
+static void uninstall_exp_breakpoints(BpFunctions* f);
+static void clean_export_entries(BpFunctions* f);
 
 void
 erts_bif_trace_init(void)
@@ -107,12 +118,12 @@ trace_pattern(Process* p, Eterm MFA, Eterm Pattern, Eterm flaglist)
     int is_global;
     Process *meta_tracer_proc = p;
     Eterm meta_tracer_pid = p->id;
+    int is_blocking = 0;
 
     if (!erts_try_seize_code_write_permission(p)) {
 	ERTS_BIF_YIELD3(bif_export[BIF_trace_pattern_3], p, MFA, Pattern, flaglist);
     }
-    erts_smp_proc_unlock(p, ERTS_PROC_LOCK_MAIN);
-    erts_smp_thr_progress_block();
+    finish_bp.current = -1;
 
     UseTmpHeap(3,p);
     /*
@@ -328,16 +339,24 @@ trace_pattern(Process* p, Eterm MFA, Eterm Pattern, Eterm flaglist)
 	    meta_tracer_proc->trace_flags |= F_TRACER;
 	}
 
-	matches = erts_set_trace_pattern(mfa, specified, 
+	matches = erts_set_trace_pattern(p, mfa, specified,
 					 match_prog_set, match_prog_set,
-					 on, flags, meta_tracer_pid);
+					 on, flags, meta_tracer_pid, 0);
     }
 
  error:
     MatchSetUnref(match_prog_set);
     UnUseTmpHeap(3,p);
-    erts_smp_thr_progress_unblock();
-    erts_smp_proc_lock(p, ERTS_PROC_LOCK_MAIN);
+
+#ifdef ERTS_SMP
+    if (finish_bp.current >= 0) {
+	ASSERT(matches >= 0);
+	erts_notify_finish_breakpointing(p);
+	erts_suspend(p, ERTS_PROC_LOCK_MAIN, NULL);
+	ERTS_BIF_YIELD_RETURN(p, make_small(matches));
+    }
+#endif
+
     erts_release_code_write_permission();
 
     if (matches >= 0) {
@@ -355,6 +374,8 @@ erts_get_default_trace_pattern(int *trace_pattern_is_on,
 			       struct trace_pattern_flags *trace_pattern_flags,
 			       Eterm *meta_tracer_pid)
 {
+    ERTS_SMP_LC_ASSERT(erts_is_code_ix_locked() ||
+		       erts_smp_thr_progress_is_blocking());
     if (trace_pattern_is_on)
 	*trace_pattern_is_on = erts_default_trace_pattern_is_on;
     if (match_spec)
@@ -369,6 +390,8 @@ erts_get_default_trace_pattern(int *trace_pattern_is_on,
 
 int erts_is_default_trace_enabled(void)
 {
+    ERTS_SMP_LC_ASSERT(erts_is_code_ix_locked() ||
+		       erts_smp_thr_progress_is_blocking());
     return erts_default_trace_pattern_is_on;
 }
 
@@ -842,6 +865,11 @@ Eterm trace_info_2(BIF_ALIST_2)
     Eterm What = BIF_ARG_1;
     Eterm Key = BIF_ARG_2;
     Eterm res;
+
+    if (!erts_try_seize_code_write_permission(p)) {
+	ERTS_BIF_YIELD2(bif_export[BIF_trace_info_2], p, What, Key);
+    }
+
     if (What == am_on_load) {
 	res = trace_info_on_load(p, Key);
     } else if (is_atom(What) || is_pid(What)) {
@@ -849,8 +877,10 @@ Eterm trace_info_2(BIF_ALIST_2)
     } else if (is_tuple(What)) {
 	res = trace_info_func(p, What, Key);
     } else {
+	erts_release_code_write_permission();
 	BIF_ERROR(p, BADARG);
     }
+    erts_release_code_write_permission();
     BIF_RET(res);
 }
 
@@ -978,64 +1008,54 @@ static int function_is_traced(Process *p,
 			      Binary **ms,              /* out */
 			      Binary **ms_meta,         /* out */
 			      Eterm   *tracer_pid_meta, /* out */
-			      Sint    *count,           /* out */
+			      Uint    *count,           /* out */
 			      Eterm   *call_time)       /* out */
 {
     Export e;
     Export* ep;
-    int i;
-    BeamInstr *code;
+    BeamInstr* pc;
 
     /* First look for an export entry */
     e.code[0] = mfa[0];
     e.code[1] = mfa[1];
     e.code[2] = mfa[2];
     if ((ep = export_get(&e)) != NULL) {
-	if (ep->addressv[erts_active_code_ix()] == ep->code+3 &&
-	    ep->code[3] != (BeamInstr) em_call_error_handler) {
-	    if (ep->code[3] == (BeamInstr) em_call_traced_function) {
-		*ms = ep->match_prog_set;
+	pc = ep->code+3;
+	if (ep->addressv[erts_active_code_ix()] == pc &&
+	    *pc != (BeamInstr) em_call_error_handler) {
+
+	    int r = 0;
+
+	    ASSERT(*pc == (BeamInstr) em_apply_bif ||
+		   *pc == (BeamInstr) BeamOp(op_i_generic_breakpoint));
+
+	    if (erts_is_trace_break(pc, ms, 0)) {
 		return FUNC_TRACE_GLOBAL_TRACE;
 	    }
-	    if (ep->code[3] == (BeamInstr) em_apply_bif) {
-		for (i = 0; i < BIF_SIZE; ++i) {
-		    if (bif_export[i] == ep) {
-			int r = 0;
-			
-			if (erts_bif_trace_flags[i] & BIF_TRACE_AS_GLOBAL) {
-			    *ms = ep->match_prog_set;
-			    return FUNC_TRACE_GLOBAL_TRACE;
-			} else {
-			    if (erts_bif_trace_flags[i] & BIF_TRACE_AS_LOCAL) {
-				r |= FUNC_TRACE_LOCAL_TRACE;
-				*ms = ep->match_prog_set;
-			    }
-			    if (erts_is_mtrace_break(ep->code+3, ms_meta,
-						   tracer_pid_meta)) {
-				r |= FUNC_TRACE_META_TRACE;
-			    }
-			    if (erts_is_time_break(p, ep->code+3, call_time)) {
-				r |= FUNC_TRACE_TIME_TRACE;
-			    }
-			}
-			return r ? r : FUNC_TRACE_UNTRACED;
-		    }
-		}
-		erl_exit(1,"Impossible ghost bif encountered in trace_info.");
+
+	    if (erts_is_trace_break(pc, ms, 1)) {
+		r |= FUNC_TRACE_LOCAL_TRACE;
+	    }
+	    if (erts_is_mtrace_break(pc, ms_meta, tracer_pid_meta)) {
+		r |= FUNC_TRACE_META_TRACE;
 	    }
+	    if (erts_is_time_break(p, pc, call_time)) {
+		r |= FUNC_TRACE_TIME_TRACE;
+	    }
+	    return r ? r : FUNC_TRACE_UNTRACED;
 	}
     }
     
     /* OK, now look for breakpoint tracing */
-    if ((code = erts_find_local_func(mfa)) != NULL) {
+    if ((pc = erts_find_local_func(mfa)) != NULL) {
 	int r = 
-	    (erts_is_trace_break(code, ms, NULL)
+	    (erts_is_trace_break(pc, ms, 1)
 	     ? FUNC_TRACE_LOCAL_TRACE : 0) 
-	    | (erts_is_mtrace_break(code, ms_meta, tracer_pid_meta)
+	    | (erts_is_mtrace_break(pc, ms_meta, tracer_pid_meta)
 	       ? FUNC_TRACE_META_TRACE : 0)
-	    | (erts_is_count_break(code, count)
+	    | (erts_is_count_break(pc, count)
 	       ? FUNC_TRACE_COUNT_TRACE : 0)
-	    | (erts_is_time_break(p, code, call_time)
+	    | (erts_is_time_break(p, pc, call_time)
 	       ? FUNC_TRACE_TIME_TRACE : 0);
 	
 	return r ? r : FUNC_TRACE_UNTRACED;
@@ -1050,7 +1070,7 @@ trace_info_func(Process* p, Eterm func_spec, Eterm key)
     Eterm* hp;
     DeclareTmpHeap(mfa,3,p); /* Not really heap here, but might be when setting pattern */
     Binary *ms = NULL, *ms_meta = NULL;
-    Sint count = 0;
+    Uint count = 0;
     Eterm traced = am_false;
     Eterm match_spec = am_false;
     Eterm retval = am_false;
@@ -1138,9 +1158,7 @@ trace_info_func(Process* p, Eterm func_spec, Eterm key)
 	break;
     case am_call_count:
 	if (r & FUNC_TRACE_COUNT_TRACE) {
-	    retval = count < 0 ? 
-		erts_make_integer(-count-1, p) : 
-		erts_make_integer(count, p);
+	    retval = erts_make_integer(count, p);
 	}
 	break;
     case am_call_time:
@@ -1329,39 +1347,46 @@ trace_info_on_load(Process* p, Eterm key)
 #undef FUNC_TRACE_LOCAL_TRACE
 
 int
-erts_set_trace_pattern(Eterm* mfa, int specified, 
+erts_set_trace_pattern(Process*p, Eterm* mfa, int specified,
 		       Binary* match_prog_set, Binary *meta_match_prog_set,
 		       int on, struct trace_pattern_flags flags,
-		       Eterm meta_tracer_pid)
+		       Eterm meta_tracer_pid, int is_blocking)
 {
     const ErtsCodeIndex code_ix = erts_active_code_ix();
     int matches = 0;
     int i;
+    int n;
+    BpFunction* fp;
 
     /*
      * First work on normal functions (not real BIFs).
      */
-    
-    for (i = 0; i < export_list_size(code_ix); i++) {
-	Export* ep = export_list(i, code_ix);
-	int j;
-	
-	if (ExportIsBuiltIn(ep)) {
-	    continue;
-	}
-	
-	for (j = 0; j < specified && mfa[j] == ep->code[j]; j++) {
-	    /* Empty loop body */
-	}
 
-	if (j == specified) {
-	    if (on) {
-		if (! flags.breakpoint)
-		    matches += setup_func_trace(ep, match_prog_set, code_ix);
-		else
-		    reset_func_trace(ep, code_ix);
-	    } else if (! flags.breakpoint) {
-		matches += reset_func_trace(ep, code_ix);
+    erts_bp_match_export(&finish_bp.e, mfa, specified);
+    fp = finish_bp.e.matching;
+    n = finish_bp.e.matched;
+
+    for (i = 0; i < n; i++) {
+	BeamInstr* pc = fp[i].pc;
+	Export* ep = (Export *)(((char *)(pc-3)) - offsetof(Export, code));
+
+	if (!on || flags.breakpoint) {
+	    erts_clear_call_trace_bif(pc, 0);
+	    if (pc[0] == (BeamInstr) BeamOp(op_i_generic_breakpoint)) {
+		pc[0] = (BeamInstr) BeamOp(op_jump_f);
+	    }
+	} else {
+	    if (ep->addressv[code_ix] != pc) {
+		fp[i].mod->curr.num_traced_exports++;
+#ifdef DEBUG
+		pc[-5] = (BeamInstr) BeamOp(op_i_func_info_IaaI);
+#endif
+		pc[0] = (BeamInstr) BeamOp(op_jump_f);
+		pc[1] = (BeamInstr) ep->addressv[code_ix];
+	    }
+	    erts_set_call_trace_bif(pc, match_prog_set, 0);
+	    if (ep->addressv[code_ix] != pc) {
+		pc[0] = (BeamInstr) BeamOp(op_i_generic_breakpoint);
 	    }
 	}
     }
@@ -1386,26 +1411,15 @@ erts_set_trace_pattern(Eterm* mfa, int specified,
 	    /* Empty loop body */
 	}
 	if (j == specified) {
+	    BeamInstr* pc = (BeamInstr *)bif_export[i]->code + 3;
+
 	    if (! flags.breakpoint) { /* Export entry call trace */
 		if (on) {
-		    if (erts_bif_trace_flags[i] & BIF_TRACE_AS_META) {
-			ASSERT(ExportIsBuiltIn(bif_export[i]));
-			erts_clear_mtrace_bif
-			    ((BeamInstr *)bif_export[i]->code + 3);
-			erts_bif_trace_flags[i] &= ~BIF_TRACE_AS_META;
-		    }
-		    set_trace_bif(i, match_prog_set);
-		    erts_bif_trace_flags[i] &= ~BIF_TRACE_AS_LOCAL;
-		    erts_bif_trace_flags[i] |= BIF_TRACE_AS_GLOBAL;
-		    setup_bif_trace(i);
+		    erts_clear_call_trace_bif(pc, 1);
+		    erts_clear_mtrace_bif(pc);
+		    erts_set_call_trace_bif(pc, match_prog_set, 0);
 		} else { /* off */
-		    if (erts_bif_trace_flags[i] & BIF_TRACE_AS_GLOBAL) {
-			clear_trace_bif(i);
-			erts_bif_trace_flags[i] &= ~BIF_TRACE_AS_GLOBAL;
-		    }
-		    if (! erts_bif_trace_flags[i]) {
-			reset_bif_trace(i);
-		    }
+		    erts_clear_call_trace_bif(pc, 0);
 		}
 		matches++;
 	    } else { /* Breakpoint call trace */
@@ -1413,52 +1427,33 @@ erts_set_trace_pattern(Eterm* mfa, int specified,
 		
 		if (on) {
 		    if (flags.local) {
-			set_trace_bif(i, match_prog_set);
-			erts_bif_trace_flags[i] |= BIF_TRACE_AS_LOCAL;
-			erts_bif_trace_flags[i] &= ~BIF_TRACE_AS_GLOBAL;
+			erts_clear_call_trace_bif(pc, 0);
+			erts_set_call_trace_bif(pc, match_prog_set, 1);
 			m = 1;
 		    }
 		    if (flags.meta) {
-			erts_set_mtrace_bif
-			    ((BeamInstr *)bif_export[i]->code + 3,
-			     meta_match_prog_set, meta_tracer_pid);
-			erts_bif_trace_flags[i] |= BIF_TRACE_AS_META;
-			erts_bif_trace_flags[i] &= ~BIF_TRACE_AS_GLOBAL;
+			erts_set_mtrace_bif(pc, meta_match_prog_set,
+					    meta_tracer_pid);
 			m = 1;
 		    }
 		    if (flags.call_time) {
-			erts_set_time_trace_bif(bif_export[i]->code + 3, on);
+			erts_set_time_trace_bif(pc, on);
 			/* I don't want to remove any other tracers */
-			erts_bif_trace_flags[i] |= BIF_TRACE_AS_CALL_TIME;
 			m = 1;
 		    }
-		    if (erts_bif_trace_flags[i]) {
-			setup_bif_trace(i);
-		    }
 		} else { /* off */
 		    if (flags.local) {
-			if (erts_bif_trace_flags[i] & BIF_TRACE_AS_LOCAL) {
-			    clear_trace_bif(i);
-			    erts_bif_trace_flags[i] &= ~BIF_TRACE_AS_LOCAL;
-			}
+			erts_clear_call_trace_bif(pc, 1);
 			m = 1;
 		    }
 		    if (flags.meta) {
-			if (erts_bif_trace_flags[i] & BIF_TRACE_AS_META) {
-			    erts_clear_mtrace_bif
-				((BeamInstr *)bif_export[i]->code + 3);
-			    erts_bif_trace_flags[i] &= ~BIF_TRACE_AS_META;
-			}
+			erts_clear_mtrace_bif(pc);
 			m = 1;
 		    }
 		    if (flags.call_time) {
-			erts_clear_time_trace_bif(bif_export[i]->code + 3);
-			erts_bif_trace_flags[i] &= ~BIF_TRACE_AS_CALL_TIME;
+			erts_clear_time_trace_bif(pc);
 			m = 1;
 		    }
-		    if (! erts_bif_trace_flags[i]) {
-			reset_bif_trace(i);
-		    }
 		}
 		matches += m;
 	    }
@@ -1468,186 +1463,242 @@ erts_set_trace_pattern(Eterm* mfa, int specified,
     /*
     ** So, now for breakpoint tracing
     */
+    erts_bp_match_functions(&finish_bp.f, mfa, specified);
     if (on) {
 	if (! flags.breakpoint) {
-	    erts_clear_trace_break(mfa, specified);
-	    erts_clear_mtrace_break(mfa, specified);
-	    erts_clear_count_break(mfa, specified);
-	    erts_clear_time_break(mfa, specified);
+	    erts_clear_all_breaks(&finish_bp.f);
 	} else {
-	    int m = 0;
 	    if (flags.local) {
-		m = erts_set_trace_break(mfa, specified, match_prog_set,
-					 am_true);
+		erts_set_trace_break(&finish_bp.f, match_prog_set);
 	    }
 	    if (flags.meta) {
-		m = erts_set_mtrace_break(mfa, specified, meta_match_prog_set,
-					  meta_tracer_pid);
+		erts_set_mtrace_break(&finish_bp.f, meta_match_prog_set,
+				      meta_tracer_pid);
 	    }
 	    if (flags.call_count) {
-		m = erts_set_count_break(mfa, specified, on);
+		erts_set_count_break(&finish_bp.f, on);
 	    }
 	    if (flags.call_time) {
-		m = erts_set_time_break(mfa, specified, on);
+		erts_set_time_break(&finish_bp.f, on);
 	    }
-	    /* All assignments to 'm' above should give the same value,
-	     * so just use the last */
-	    matches += m;
 	}
     } else {
-	int m = 0;
 	if (flags.local) {
-	    m = erts_clear_trace_break(mfa, specified);
+	    erts_clear_trace_break(&finish_bp.f);
 	}
 	if (flags.meta) {
-	    m = erts_clear_mtrace_break(mfa, specified);
+	    erts_clear_mtrace_break(&finish_bp.f);
 	}
 	if (flags.call_count) {
-	    m = erts_clear_count_break(mfa, specified);
+	    erts_clear_count_break(&finish_bp.f);
 	}
 	if (flags.call_time) {
-	    m = erts_clear_time_break(mfa, specified);
+	    erts_clear_time_break(&finish_bp.f);
 	}
-	/* All assignments to 'm' above should give the same value,
-	 * so just use the last */
-	matches += m;
     }
 
+    finish_bp.current = 0;
+    finish_bp.install = on;
+    finish_bp.local = flags.breakpoint;
+
+#ifdef ERTS_SMP
+    if (is_blocking) {
+	ERTS_SMP_LC_ASSERT(erts_smp_thr_progress_is_blocking());
+#endif
+	while (erts_finish_breakpointing()) {
+	    /* Empty loop body */
+	}
+#ifdef ERTS_SMP
+	finish_bp.current = -1;
+    }
+#endif
+
+    if (flags.breakpoint) {
+	matches += finish_bp.f.matched;
+    } else {
+	matches += finish_bp.e.matched;
+    }
     return matches;
 }
 
-/*
- * Setup function tracing for the given exported function.
- *
- * Return Value: 1 if entry refers to a BIF or loaded function,
- * 0 if the entry refers to a function not loaded.
- */
-
-static int
-setup_func_trace(Export* ep, void* match_prog, ErtsCodeIndex code_ix)
+int
+erts_finish_breakpointing(void)
 {
-    Module* modp;
-
-    if (ep->addressv[code_ix] == ep->code+3) {
-	if (ep->code[3] == (BeamInstr) em_call_error_handler) {
-	    return 0;
-	} else if (ep->code[3] == (BeamInstr) em_call_traced_function) {
-	    MatchSetUnref(ep->match_prog_set);
-	    ep->match_prog_set = match_prog;
-	    MatchSetRef(ep->match_prog_set);
-	    return 1;
-	} else {
-	    /*
-	     * We ignore apply/3 and anything else.
-	     */
-	    return 0;
-	}
-    }
-    
+    ERTS_SMP_LC_ASSERT(erts_is_code_ix_locked());
+
     /*
-     * Currently no trace support for native code.
+     * Memory barriers will be issued for all processes *before*
+     * each of the stages below. (Unless the other schedulers
+     * are blocked, in which case memory barriers will be issued
+     * when they are awaken.)
      */
-    if (erts_is_native_break(ep->addressv[code_ix])) {
+
+    switch (finish_bp.current++) {
+    case 0:
+	/*
+	 * At this point, in all functions that are to be breakpointed,
+	 * a pointer to a GenericBp struct has already been added,
+	 *
+	 * Insert the new breakpoints (if any) into the
+	 * code. Different schedulers may see breakpoint instruction
+	 * at different times, but it does not matter since the newly
+	 * added breakpoints are disabled.
+	 */
+	if (finish_bp.install) {
+	    if (finish_bp.local) {
+		erts_install_breakpoints(&finish_bp.f);
+	    } else {
+		install_exp_breakpoints(&finish_bp.e);
+	    }
+	}
+	setup_bif_trace();
+	return 1;
+    case 1:
+	/*
+	 * Switch index for the breakpoint data, activating the staged
+	 * data. (Depending on the changes in the breakpoint data,
+	 * that could either activate breakpoints or disable
+	 * breakpoints.)
+	 */
+	erts_commit_staged_bp();
+	return 1;
+    case 2:
+	/*
+	 * Remove breakpoints instructions for disabled breakpoints
+	 * (if any).
+	 */
+	if (finish_bp.install) {
+	    if (finish_bp.local) {
+		uninstall_exp_breakpoints(&finish_bp.e);
+	    } else {
+		erts_uninstall_breakpoints(&finish_bp.f);
+	    }
+	} else {
+	    if (finish_bp.local) {
+		erts_uninstall_breakpoints(&finish_bp.f);
+	    } else {
+		uninstall_exp_breakpoints(&finish_bp.e);
+	    }
+	}
+	reset_bif_trace();
+	return 1;
+    case 3:
+	/*
+	 * Now all breakpoints have either been inserted or removed.
+	 * For all updated breakpoints, copy the active breakpoint
+	 * data to the staged breakpoint data to make them equal
+	 * (simplifying for the next time breakpoints are to be
+	 * updated).  If any breakpoints have been totally disabled,
+	 * deallocate the GenericBp structs for them.
+	 */
+	erts_consolidate_bif_bp_data();
+	clean_export_entries(&finish_bp.e);
+	erts_consolidate_bp_data(&finish_bp.e, 0);
+	erts_consolidate_bp_data(&finish_bp.f, 1);
+	erts_bp_free_matched_functions(&finish_bp.e);
+	erts_bp_free_matched_functions(&finish_bp.f);
 	return 0;
+    default:
+	ASSERT(0);
     }
+    return 0;
+}
 
-    ep->code[3] = (BeamInstr) em_call_traced_function;
-    ep->code[4] = (BeamInstr) ep->addressv[code_ix];
-    ep->addressv[code_ix] = ep->code+3;
-    ep->match_prog_set = match_prog;
-    MatchSetRef(ep->match_prog_set);
+static void
+install_exp_breakpoints(BpFunctions* f)
+{
+    const ErtsCodeIndex code_ix = erts_active_code_ix();
+    BpFunction* fp = f->matching;
+    Uint ne = f->matched;
+    Uint i;
+    Uint offset = offsetof(Export, code) + 3*sizeof(BeamInstr);
 
-    modp = erts_get_module(ep->code[0], code_ix);
-    ASSERT(modp);
-    modp->curr.num_traced_exports++;
-    return 1;
-}
+    for (i = 0; i < ne; i++) {
+	BeamInstr* pc = fp[i].pc;
+	Export* ep = (Export *) (((char *)pc)-offset);
 
-static void setup_bif_trace(int bif_index) {
-    Export *ep = bif_export[bif_index];
-    
-    ASSERT(ExportIsBuiltIn(ep));
-    ASSERT(ep->code[4]);
-    ep->code[4] = (BeamInstr) bif_table[bif_index].traced;
+	ep->addressv[code_ix] = pc;
+    }
 }
 
-static void set_trace_bif(int bif_index, void* match_prog) {
-    Export *ep = bif_export[bif_index];
-    
-#ifdef HARDDEBUG
-    erts_fprintf(stderr, "set_trace_bif: %T:%T/%bpu\n",
-		 ep->code[0], ep->code[1], ep->code[2]);
-#endif
-    ASSERT(ExportIsBuiltIn(ep));
-    MatchSetUnref(ep->match_prog_set);
-    ep->match_prog_set = match_prog;
-    MatchSetRef(ep->match_prog_set);
-}
+static void
+uninstall_exp_breakpoints(BpFunctions* f)
+{
+    const ErtsCodeIndex code_ix = erts_active_code_ix();
+    BpFunction* fp = f->matching;
+    Uint ne = f->matched;
+    Uint i;
+    Uint offset = offsetof(Export, code) + 3*sizeof(BeamInstr);
 
-/*
- * Reset function tracing for the given exported function.
- *
- * Return Value: 1 if entry refers to a BIF or loaded function,
- * 0 if the entry refers to a function not loaded.
- */
+    for (i = 0; i < ne; i++) {
+	BeamInstr* pc = fp[i].pc;
+	Export* ep = (Export *) (((char *)pc)-offset);
 
-static int
-reset_func_trace(Export* ep, ErtsCodeIndex code_ix)
-{
-    if (ep->addressv[code_ix] == ep->code+3) {
-	if (ep->code[3] == (BeamInstr) em_call_error_handler) {
-	    return 0;
-	} else if (ep->code[3] == (BeamInstr) em_call_traced_function) {
-	    Module* modp = erts_get_module(ep->code[0], code_ix);
-	    ASSERT(modp);
-	    modp->curr.num_traced_exports--;
-
-	    ep->addressv[code_ix] = (Uint *) ep->code[4];
-	    MatchSetUnref(ep->match_prog_set);
-	    ep->match_prog_set = NULL;
-	    return 1;
-	} else {
-	    /*
-	     * We ignore apply/3 and anything else.
-	     */
-	    return 0;
+	if (ep->addressv[code_ix] != pc) {
+	    continue;
 	}
+	ASSERT(*pc == (BeamInstr) BeamOp(op_jump_f));
+	ep->addressv[code_ix] = (BeamInstr *) ep->code[4];
     }
-    
-    /*
-     * Currently no trace support for native code.
-     */
-    if (erts_is_native_break(ep->addressv[code_ix])) {
-	return 0;
-    }
-    
-    /*
-     * Nothing to do, but the export entry matches.
-     */
+}
+
+static void
+clean_export_entries(BpFunctions* f)
+{
+    const ErtsCodeIndex code_ix = erts_active_code_ix();
+    BpFunction* fp = f->matching;
+    Uint ne = f->matched;
+    Uint i;
+    Uint offset = offsetof(Export, code) + 3*sizeof(BeamInstr);
+
+    for (i = 0; i < ne; i++) {
+	BeamInstr* pc = fp[i].pc;
+	Export* ep = (Export *) (((char *)pc)-offset);
 
-    return 1;
+	if (ep->addressv[code_ix] == pc) {
+	    continue;
+	}
+	if (*pc == (BeamInstr) BeamOp(op_jump_f)) {
+	    ep->code[3] = (BeamInstr) 0;
+	    ep->code[4] = (BeamInstr) 0;
+	}
+    }
 }
 
-static void reset_bif_trace(int bif_index) {
-    Export *ep = bif_export[bif_index];
-    
-    ASSERT(ExportIsBuiltIn(ep));
-    ASSERT(ep->code[4]);
-    ASSERT(! ep->match_prog_set);
-    ASSERT(! erts_is_mtrace_break((BeamInstr *)ep->code+3, NULL, NULL));
-    ep->code[4] = (BeamInstr) bif_table[bif_index].f;
+static void
+setup_bif_trace(void)
+{
+    int i;
+
+    for (i = 0; i < BIF_SIZE; ++i) {
+	Export *ep = bif_export[i];
+	GenericBp* g = (GenericBp *) ep->fake_op_func_info_for_hipe[1];
+	if (g) {
+	    if (ExportIsBuiltIn(ep)) {
+		ASSERT(ep->code[4]);
+		ep->code[4] = (BeamInstr) bif_table[i].traced;
+	    }
+	}
+    }
 }
 
-static void clear_trace_bif(int bif_index) {
-    Export *ep = bif_export[bif_index];
-    
-#ifdef HARDDEBUG
-    erts_fprintf(stderr, "clear_trace_bif: %T:%T/%bpu\n",
-		 ep->code[0], ep->code[1], ep->code[2]);
-#endif
-    ASSERT(ExportIsBuiltIn(ep));
-    MatchSetUnref(ep->match_prog_set);
-    ep->match_prog_set = NULL;
+static void
+reset_bif_trace(void)
+{
+    int i;
+    ErtsBpIndex active = erts_active_bp_ix();
+
+    for (i = 0; i < BIF_SIZE; ++i) {
+	Export *ep = bif_export[i];
+	BeamInstr* pc = ep->code+3;
+	GenericBp* g = (GenericBp *) pc[-4];
+	if (g && g->data[active].flags == 0) {
+	    if (ExportIsBuiltIn(ep)) {
+		ASSERT(ep->code[4]);
+		ep->code[4] = (BeamInstr) bif_table[i].f;
+	    }
+	}
+    }
 }
 
 /*
diff --git a/erts/emulator/beam/erl_nif.c b/erts/emulator/beam/erl_nif.c
index da4376fd0a..594c51a5db 100644
--- a/erts/emulator/beam/erl_nif.c
+++ b/erts/emulator/beam/erl_nif.c
@@ -1711,9 +1711,10 @@ BIF_RETTYPE load_nif_2(BIF_ALIST_2)
 		code_ptr[5+0] = (BeamInstr) BeamOp(op_call_nif);
 	    }
 	    else { /* Function traced, patch the original instruction word */
-		BpData** bps = (BpData**) code_ptr[1];
-		BpData*  bp  = (BpData*) bps[erts_bp_sched2ix()];
-	        bp->orig_instr = (BeamInstr) BeamOp(op_call_nif);
+		GenericBp* g = (GenericBp *) code_ptr[1];
+		ASSERT(code_ptr[5+0] ==
+		       (BeamInstr) BeamOp(op_i_generic_breakpoint));
+	        g->orig_instr = (BeamInstr) BeamOp(op_call_nif);
 	    }	    
 	    code_ptr[5+1] = (BeamInstr) entry->funcs[i].fptr;
 	    code_ptr[5+2] = (BeamInstr) lib;
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index 3d978f4994..e07c9ae2b0 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -530,6 +530,7 @@ dbg_chk_aux_work_val(erts_aint32_t value)
 #endif
 #ifdef ERTS_SMP
     valid |= ERTS_SSI_AUX_WORK_CODE_IX_ACTIVATION;
+    valid |= ERTS_SSI_AUX_WORK_FINISH_BP;
 #endif
 #ifdef ERTS_SSI_AUX_WORK_REAP_PORTS
     valid |= ERTS_SSI_AUX_WORK_REAP_PORTS;
@@ -1446,6 +1447,55 @@ handle_code_ix_activation(ErtsAuxWorkData *awdp, erts_aint32_t aux_work)
 }
 #endif /* ERTS_SMP */
 
+#ifdef ERTS_SMP
+void
+erts_notify_finish_breakpointing(Process* p)
+{
+    ErtsAuxWorkData* awdp = &p->scheduler_data->aux_work_data;
+
+    ASSERT(awdp->bp_ix_activation.stager == NULL);
+    awdp->bp_ix_activation.stager = p;
+    awdp->bp_ix_activation.thr_prgr = erts_thr_progress_later(awdp->esdp);
+    erts_thr_progress_wakeup(awdp->esdp, awdp->bp_ix_activation.thr_prgr);
+    erts_smp_proc_inc_refc(p);
+    set_aux_work_flags_wakeup_relb(p->scheduler_data->ssi,
+				   ERTS_SSI_AUX_WORK_FINISH_BP);
+}
+
+static erts_aint32_t
+handle_finish_bp(ErtsAuxWorkData* awdp, erts_aint32_t aux_work)
+{
+    ErtsThrPrgrVal current = haw_thr_prgr_current(awdp);
+
+    if (!erts_thr_progress_has_reached_this(current,
+					    awdp->bp_ix_activation.thr_prgr)) {
+	return aux_work & ~ERTS_SSI_AUX_WORK_FINISH_BP;
+    }
+    if (erts_finish_breakpointing()) { /* Not done */
+	/* Arrange for being called again */
+	awdp->bp_ix_activation.thr_prgr =
+	    erts_thr_progress_later(awdp->esdp);
+	erts_thr_progress_wakeup(awdp->esdp, awdp->bp_ix_activation.thr_prgr);
+    } else {			/* Done */
+	Process* p;
+
+	unset_aux_work_flags(awdp->ssi, ERTS_SSI_AUX_WORK_FINISH_BP);
+	p = awdp->bp_ix_activation.stager;
+#ifdef DEBUG
+	awdp->bp_ix_activation.stager = NULL;
+#endif
+	erts_smp_proc_lock(p, ERTS_PROC_LOCK_STATUS);
+	if (!ERTS_PROC_IS_EXITING(p)) {
+	    erts_resume(p, ERTS_PROC_LOCK_STATUS);
+	}
+	erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS);
+	erts_smp_proc_dec_refc(p);
+	erts_release_code_write_permission();
+    }
+    return aux_work & ~ERTS_SSI_AUX_WORK_FINISH_BP;
+}
+#endif /* ERTS_SMP */
+
 static ERTS_INLINE erts_aint32_t
 handle_fix_alloc(ErtsAuxWorkData *awdp, erts_aint32_t aux_work)
 {
@@ -1817,6 +1867,11 @@ handle_aux_work(ErtsAuxWorkData *awdp, erts_aint32_t orig_aux_work, int waiting)
     HANDLE_AUX_WORK(ERTS_SSI_AUX_WORK_REAP_PORTS,
 		    handle_reap_ports);
 
+#ifdef ERTS_SMP
+    HANDLE_AUX_WORK(ERTS_SSI_AUX_WORK_FINISH_BP,
+		    handle_finish_bp);
+#endif
+
     ERTS_DBG_CHK_AUX_WORK_VAL(aux_work);
 
 #ifdef ERTS_SMP
diff --git a/erts/emulator/beam/erl_process.h b/erts/emulator/beam/erl_process.h
index e9e5a3365f..93e71681da 100644
--- a/erts/emulator/beam/erl_process.h
+++ b/erts/emulator/beam/erl_process.h
@@ -283,6 +283,7 @@ typedef enum {
 #define ERTS_SSI_AUX_WORK_MSEG_CACHE_CHECK	(((erts_aint32_t) 1) << 10)
 #define ERTS_SSI_AUX_WORK_CODE_IX_ACTIVATION	(((erts_aint32_t) 1) << 11)
 #define ERTS_SSI_AUX_WORK_REAP_PORTS		(((erts_aint32_t) 1) << 12)
+#define ERTS_SSI_AUX_WORK_FINISH_BP	        (((erts_aint32_t) 1) << 13)
 
 typedef struct ErtsSchedulerSleepInfo_ ErtsSchedulerSleepInfo;
 
@@ -472,6 +473,12 @@ typedef struct {
 	ErtsThrPrgrVal thr_prgr;
     } code_ix_activation;
 #endif
+#ifdef ERTS_SMP
+    struct {
+	Process* stager;
+	ErtsThrPrgrVal thr_prgr;
+    } bp_ix_activation;
+#endif
 } ErtsAuxWorkData;
 
 struct ErtsSchedulerData_ {
@@ -1188,6 +1195,7 @@ void erts_notify_check_async_ready_queue(void *);
 #endif
 #ifdef ERTS_SMP
 void erts_notify_code_ix_activation(Process* p, ErtsThrPrgrVal later);
+void erts_notify_finish_breakpointing(Process* p);
 #endif
 void erts_schedule_misc_aux_work(int sched_id,
 				 void (*func)(void *),
diff --git a/erts/emulator/beam/erl_trace.c b/erts/emulator/beam/erl_trace.c
index bc988cd61b..d04a91f18c 100644
--- a/erts/emulator/beam/erl_trace.c
+++ b/erts/emulator/beam/erl_trace.c
@@ -2134,187 +2134,6 @@ void save_calls(Process *p, Export *e)
     }
 }
 
-/* 
- * Entry point called by the trace wrap functions in erl_bif_wrap.c
- *
- * The trace wrap functions are themselves called through the export
- * entries instead of the original BIF functions.
- */
-Eterm
-erts_bif_trace(int bif_index, Process* p, Eterm* args, BeamInstr* I)
-{
-    Eterm result;
-    int meta = !!(erts_bif_trace_flags[bif_index] & BIF_TRACE_AS_META);
-
-    ERTS_SMP_CHK_HAVE_ONLY_MAIN_PROC_LOCK(p);
-
-    if (!ARE_TRACE_FLAGS_ON(p, F_TRACE_CALLS) && (! meta)) {
-	/* Warning! This is an Optimization. 
-	 *
-	 * If neither meta trace is active nor process trace flags then 
-	 * no tracing will occur. Doing the whole else branch will 
-	 * also do nothing, only slower.
-	 */
-	Eterm (*func)(Process*, Eterm*, BeamInstr*) = bif_table[bif_index].f;
-	result = func(p, args, I);
-    } else {
-	Eterm (*func)(Process*, Eterm*, BeamInstr*);
-	Export* ep = bif_export[bif_index];
-	Uint32 flags = 0, flags_meta = 0;
-	int global = !!(erts_bif_trace_flags[bif_index] & BIF_TRACE_AS_GLOBAL);
-	int local  = !!(erts_bif_trace_flags[bif_index] & BIF_TRACE_AS_LOCAL);
-	int time   = !!(erts_bif_trace_flags[bif_index] & BIF_TRACE_AS_CALL_TIME);
-	Eterm meta_tracer_pid = NIL;
-	int applying = (I == &(ep->code[3])); /* Yup, the apply code for a bif
-					       * is actually in the 
-					       * export entry */
-	BeamInstr *cp = p->cp;
-	
-	/* 
-	 * Make continuation pointer OK, it is not during direct BIF calls,
-	 * but it is correct during apply of bif.
-	 */
-	if (!applying) { 
-	    p->cp = I;
-	}
-	if (global || local) {
-	    flags = erts_call_trace(p, ep->code, ep->match_prog_set, args, 
-				       local, &p->tracer_proc);
-	}
-	if (meta) {
-	    flags_meta = erts_bif_mtrace(p, ep->code+3, args, local, 
-					 &meta_tracer_pid);
-	}
-	if (time) {
-	    BpDataTime *bdt = NULL;
-	    BeamInstr *pc = (BeamInstr *)ep->code+3;
-
-	    bdt = (BpDataTime *) erts_get_time_break(p, pc);
-	    ASSERT(bdt);
-
-	    if (!bdt->pause) {
-		erts_trace_time_break(p, pc, bdt, ERTS_BP_CALL_TIME_CALL);
-	    }
-	}
-	/* Restore original continuation pointer (if changed). */
-	p->cp = cp;
-	
-	func = bif_table[bif_index].f;
-
-	result = func(p, args, I);
-	
-	if (applying && (flags & MATCH_SET_RETURN_TO_TRACE)) {
-	    BeamInstr i_return_trace      = beam_return_trace[0];
-	    BeamInstr i_return_to_trace   = beam_return_to_trace[0];
-	    BeamInstr i_return_time_trace = beam_return_time_trace[0];
-	    Eterm *cpp;
-	    /* Maybe advance cp to skip trace stack frames */
-	    for (cpp = p->stop;  ;  cp = cp_val(*cpp++)) {
-		if (*cp == i_return_trace) {
-		    /* Skip stack frame variables */
-		    while (is_not_CP(*cpp)) cpp++;
-		    cpp += 2; /* Skip return_trace parameters */
-		} else if (*cp == i_return_time_trace) {
-		    /* Skip stack frame variables */
-		    while (is_not_CP(*cpp)) cpp++;
-		    cpp += 1; /* Skip return_time_trace parameters */
-		} else if (*cp == i_return_to_trace) {
-		    /* A return_to trace message is going to be generated
-		     * by normal means, so we do not have to.
-		     */
-		    cp = NULL;
-		    break;
-		} else break;
-	    }
-	}
-	
-	/* Try to get these in the order 
-	 * they usually appear in normal code... */
-	if (is_non_value(result)) {
-	    Uint reason = p->freason;
-	    if (reason != TRAP) {
-		Eterm class;
-		Eterm value = p->fvalue;
-		DeclareTmpHeapNoproc(nocatch,3);
-		UseTmpHeapNoproc(3);
-		/* Expand error value like in handle_error() */
-		if (reason & EXF_ARGLIST) {
-		    Eterm *tp;
-		    ASSERT(is_tuple(value));
-		    tp = tuple_val(value);
-		    value = tp[1];
-		}
-		if ((reason & EXF_THROWN) && (p->catches <= 0)) {
-		    value = TUPLE2(nocatch, am_nocatch, value);
-		    reason = EXC_ERROR;
-		}
-		/* Note: expand_error_value() could theoretically 
-		 * allocate on the heap, but not for any error
-		 * returned by a BIF, and it would do no harm,
-		 * just be annoying.
-		 */
-		value = expand_error_value(p, reason, value);
-		class = exception_tag[GET_EXC_CLASS(reason)];
-		
-		if (flags_meta & MATCH_SET_EXCEPTION_TRACE) {
-		    erts_trace_exception(p, ep->code, class, value,
-					 &meta_tracer_pid);
-		}
-		if (flags & MATCH_SET_EXCEPTION_TRACE) {
-		    erts_trace_exception(p, ep->code, class, value,
-					 &p->tracer_proc);
-		}
-		if ((flags & MATCH_SET_RETURN_TO_TRACE) && p->catches > 0) {
-		    /* can only happen if(local)*/
-		    Eterm *ptr = p->stop;
-		    ASSERT(is_CP(*ptr));
-		    ASSERT(ptr <= STACK_START(p));
-		    /* Search the nearest stack frame for a catch */
-		    while (++ptr < STACK_START(p)) {
-			if (is_CP(*ptr)) break;
-			if (is_catch(*ptr)) {
-			    if (applying) {
-				/* Apply of BIF, cp is in calling function */
-				if (cp) erts_trace_return_to(p, cp);
-			    } else {
-				/* Direct bif call, I points into 
-				 * calling function */
-				erts_trace_return_to(p, I);
-			    }
-			}
-		    }
-		}
-		UnUseTmpHeapNoproc(3);
-		if ((flags_meta|flags) & MATCH_SET_EXCEPTION_TRACE) {
-		    erts_smp_proc_lock(p, ERTS_PROC_LOCKS_ALL_MINOR);
-		    p->trace_flags |= F_EXCEPTION_TRACE;
-		    erts_smp_proc_unlock(p, ERTS_PROC_LOCKS_ALL_MINOR);
-		}
-	    }
-	} else {
-	    if (flags_meta & MATCH_SET_RX_TRACE) {
-		erts_trace_return(p, ep->code, result, &meta_tracer_pid);
-	    }
-	    /* MATCH_SET_RETURN_TO_TRACE cannot occur if(meta) */
-	    if (flags & MATCH_SET_RX_TRACE) {
-		erts_trace_return(p, ep->code, result, &p->tracer_proc);
-	    }
-	    if (flags & MATCH_SET_RETURN_TO_TRACE) { 
-		/* can only happen if(local)*/
-		if (applying) {
-		    /* Apply of BIF, cp is in calling function */
-		    if (cp) erts_trace_return_to(p, cp);
-		} else {
-		    /* Direct bif call, I points into calling function */
-		    erts_trace_return_to(p, I);
-		}
-	    }
-	}
-    }
-    ERTS_SMP_CHK_HAVE_ONLY_MAIN_PROC_LOCK(p);
-    return result;
-}
-
 /* Sends trace message:
  *    {trace_ts, Pid, What, Msg, Timestamp}
  * or {trace, Pid, What, Msg}
diff --git a/erts/emulator/beam/export.c b/erts/emulator/beam/export.c
index 229641cb32..6b5121f917 100644
--- a/erts/emulator/beam/export.c
+++ b/erts/emulator/beam/export.c
@@ -137,7 +137,6 @@ export_alloc(struct export_entry* tmpl_e)
 	obj->code[2] = tmpl->code[2];
 	obj->code[3] = (BeamInstr) em_call_error_handler;
 	obj->code[4] = 0;
-	obj->match_prog_set = NULL;
 
 	for (ix=0; ix<ERTS_NUM_CODE_IX; ix++) {
 	    obj->addressv[ix] = obj->code+3;
@@ -260,8 +259,9 @@ erts_find_function(Eterm m, Eterm f, unsigned int a, ErtsCodeIndex code_ix)
     struct export_entry* ee;
 
     ee = hash_get(&export_tables[code_ix].htable, init_template(&templ, m, f, a));
-    if (ee == NULL || (ee->ep->addressv[code_ix] == ee->ep->code+3 &&
-		       ee->ep->code[3] != (BeamInstr) em_call_traced_function)) {
+    if (ee == NULL ||
+	(ee->ep->addressv[code_ix] == ee->ep->code+3 &&
+	 ee->ep->code[3] != (BeamInstr) BeamOp(op_i_generic_breakpoint))) {
 	return NULL;
     }
     return ee->ep;
diff --git a/erts/emulator/beam/export.h b/erts/emulator/beam/export.h
index ec9fcb26f2..ee06e69aff 100644
--- a/erts/emulator/beam/export.h
+++ b/erts/emulator/beam/export.h
@@ -37,7 +37,6 @@
 typedef struct export
 {
     void* addressv[ERTS_NUM_CODE_IX];  /* Pointer to code for function. */
-    struct binary* match_prog_set; /* Match program for tracing. */
 
     BeamInstr fake_op_func_info_for_hipe[2]; /* MUST be just before code[] */
     /*
@@ -46,12 +45,12 @@ typedef struct export
      * code[2]: Arity (untagged integer).
      * code[3]: This entry is 0 unless the 'address' field points to it.
      *          Threaded code instruction to load function
-     *		(em_call_error_handler), execute BIF (em_apply_bif,
-     *		em_apply_apply), or call a traced function
-     *		(em_call_traced_function).
-     * code[4]: Function pointer to BIF function (for BIFs only)
+     *		(em_call_error_handler), execute BIF (em_apply_bif),
+     *		or a breakpoint instruction (op_i_generic_breakpoint).
+     * code[4]: Function pointer to BIF function (for BIFs only),
      *		or pointer to threaded code if the module has an
-     *		on_load function that has not been run yet.
+     *		on_load function that has not been run yet, or pointer
+     *          to code for function code[3] is a breakpont instruction.
      *		Otherwise: 0.
      */
     BeamInstr code[5];
diff --git a/erts/emulator/beam/global.h b/erts/emulator/beam/global.h
index 9e387c550f..c9be20322d 100755
--- a/erts/emulator/beam/global.h
+++ b/erts/emulator/beam/global.h
@@ -1682,10 +1682,10 @@ struct trace_pattern_flags {
 };
 extern const struct trace_pattern_flags erts_trace_pattern_flags_off;
 extern int erts_call_time_breakpoint_tracing;
-int erts_set_trace_pattern(Eterm* mfa, int specified, 
+int erts_set_trace_pattern(Process*p, Eterm* mfa, int specified,
 			   Binary* match_prog_set, Binary *meta_match_prog_set,
 			   int on, struct trace_pattern_flags,
-			   Eterm meta_tracer_pid);
+			   Eterm meta_tracer_pid, int is_blocking);
 void
 erts_get_default_trace_pattern(int *trace_pattern_is_on,
 			       Binary **match_spec,
@@ -1694,6 +1694,7 @@ erts_get_default_trace_pattern(int *trace_pattern_is_on,
 			       Eterm *meta_tracer_pid);
 int erts_is_default_trace_enabled(void);
 void erts_bif_trace_init(void);
+int erts_finish_breakpointing(void);
 
 /*
 ** Call_trace uses this API for the parameter matching functions
@@ -1739,14 +1740,6 @@ extern void erts_match_prog_foreach_offheap(Binary *b,
 					   breakpoint functions */
 #define MATCH_SET_EXCEPTION_TRACE (0x4) /* exception trace requested */
 #define MATCH_SET_RX_TRACE (MATCH_SET_RETURN_TRACE|MATCH_SET_EXCEPTION_TRACE)
-/*
- * Flag values when tracing bif
- * Future note: flag field is 8 bits
- */
-#define BIF_TRACE_AS_LOCAL      (0x1)
-#define BIF_TRACE_AS_GLOBAL     (0x2)
-#define BIF_TRACE_AS_META       (0x4)
-#define BIF_TRACE_AS_CALL_TIME  (0x8)
 
 extern erts_driver_t vanilla_driver;
 extern erts_driver_t spawn_driver;
diff --git a/erts/emulator/beam/ops.tab b/erts/emulator/beam/ops.tab
index c58b36231c..6764e88c81 100644
--- a/erts/emulator/beam/ops.tab
+++ b/erts/emulator/beam/ops.tab
@@ -62,11 +62,8 @@ label L
 i_func_info I a a I
 int_code_end
 
-i_trace_breakpoint
-i_mtrace_breakpoint
+i_generic_breakpoint
 i_debug_breakpoint
-i_count_breakpoint
-i_time_breakpoint
 i_return_time_trace
 i_return_to_trace
 i_yield
@@ -522,7 +519,6 @@ apply_bif
 call_nif
 call_error_handler
 error_action_code
-call_traced_function
 return_trace
 
 #
diff --git a/erts/emulator/test/call_trace_SUITE.erl b/erts/emulator/test/call_trace_SUITE.erl
index a642c3a63a..eaecd32f95 100644
--- a/erts/emulator/test/call_trace_SUITE.erl
+++ b/erts/emulator/test/call_trace_SUITE.erl
@@ -78,7 +78,13 @@ init_per_testcase(Func, Config) when is_atom(Func), is_list(Config) ->
 
 end_per_testcase(_Func, Config) ->
     Dog = ?config(watchdog, Config),
-    ?t:timetrap_cancel(Dog).
+    ?t:timetrap_cancel(Dog),
+
+    %% Reloading the module will clear all trace patterns, and
+    %% in a debug-compiled emulator run assertions of the counters
+    %% for the number of traced exported functions in this module.
+
+    c:l(?MODULE).
 
 hipe(Config) when is_list(Config) ->
     ?line 0 = erlang:trace_pattern({?MODULE,worker_foo,1}, true),
@@ -187,7 +193,12 @@ basic() ->
     %% Trace some functions...
 
     ?line trace_func({lists,'_','_'}, []),
+
+    %% Make sure that tracing the same functions more than once
+    %% does not cause any problems.
     ?line 3 = trace_func({?MODULE,foo,'_'}, true),
+    ?line 3 = trace_func({?MODULE,foo,'_'}, true),
+    ?line 1 = trace_func({?MODULE,bar,0}, true),
     ?line 1 = trace_func({?MODULE,bar,0}, true),
     ?line {traced,global} = trace_info({?MODULE,bar,0}, traced),
     ?line 1 = trace_func({erlang,list_to_integer,1}, true),
diff --git a/erts/emulator/test/trace_local_SUITE.erl b/erts/emulator/test/trace_local_SUITE.erl
index 32e2a98e3c..b238d5c630 100644
--- a/erts/emulator/test/trace_local_SUITE.erl
+++ b/erts/emulator/test/trace_local_SUITE.erl
@@ -80,6 +80,7 @@ config(priv_dir,_) ->
 	 exception_meta_nocatch/1, exception_meta_nocatch_apply/1,
 	 exception_meta_nocatch_function/1,
 	 exception_meta_nocatch_apply_function/1,
+	 concurrency/1,
 	 init_per_testcase/2, end_per_testcase/2]).
 init_per_testcase(_Case, Config) ->
     ?line Dog=test_server:timetrap(test_server:minutes(2)),
@@ -89,7 +90,15 @@ end_per_testcase(_Case, Config) ->
     shutdown(),
     Dog=?config(watchdog, Config),
     test_server:timetrap_cancel(Dog),
-    ok.
+
+    %% Reloading the module will clear all trace patterns, and
+    %% in a debug-compiled emulator run assertions of the counters
+    %% for the number of functions with breakpoints.
+
+    c:l(?MODULE).
+
+
+
 suite() -> [{ct_hooks,[ts_install_cth]}].
 
 all() -> 
@@ -106,7 +115,8 @@ all() ->
 	     exception_meta_apply_function, exception_meta_nocatch,
 	     exception_meta_nocatch_apply,
 	     exception_meta_nocatch_function,
-	     exception_meta_nocatch_apply_function]
+	     exception_meta_nocatch_apply_function,
+	     concurrency]
     end.
 
 groups() -> 
@@ -350,7 +360,8 @@ same(A, B) ->
 
 basic_test() ->
     ?line setup([call]),
-    ?line erlang:trace_pattern({?MODULE,'_','_'},[],[local]),
+    NumMatches = erlang:trace_pattern({?MODULE,'_','_'},[],[local]),
+    NumMatches = erlang:trace_pattern({?MODULE,'_','_'},[],[local]),
     ?line erlang:trace_pattern({?MODULE,slave,'_'},false,[local]),
     ?line [1,1,1,1] = apply_slave(?MODULE,exported_wrap,[1]),
     ?line ?CT(?MODULE,exported_wrap,[1]),
@@ -813,6 +824,42 @@ clean_location({crash,{Reason,Stk0}}) ->
     {crash,{Reason,Stk}};
 clean_location(Term) -> Term.
 
+concurrency(_Config) ->
+    N = erlang:system_info(schedulers),
+
+    %% Spawn 2*N processes that spin in a tight infinite loop,
+    %% and one process that will turn on and off local call
+    %% trace on the infinite_loop/0 function. We expect the
+    %% emulator to crash if there is a memory barrier bug or
+    %% if an aligned word-sized write is not atomic.
+
+    Ps0 = [spawn_monitor(fun() -> infinite_loop() end) ||
+	      _ <- lists:seq(1, 2*N)],
+    OnAndOff = fun() -> concurrency_on_and_off() end,
+    Ps1 = [spawn_monitor(OnAndOff)|Ps0],
+    ?t:sleep(1000),
+
+    %% Now spawn off N more processes that turn on off and off
+    %% a local trace pattern.
+    Ps = [spawn_monitor(OnAndOff) || _ <- lists:seq(1, N)] ++ Ps1,
+    ?t:sleep(1000),
+
+    %% Clean up.
+    [exit(Pid, kill) || {Pid,_} <- Ps],
+    [receive
+	 {'DOWN',Ref,process,Pid,killed} -> ok
+     end || {Pid,Ref} <- Ps],
+    erlang:trace_pattern({?MODULE,infinite_loop,0}, false, [local]),
+    ok.
+
+concurrency_on_and_off() ->
+    1 = erlang:trace_pattern({?MODULE,infinite_loop,0}, true, [local]),
+    1 = erlang:trace_pattern({?MODULE,infinite_loop,0}, false, [local]),
+    concurrency_on_and_off().
+
+infinite_loop() ->
+    infinite_loop().
+
 %%% Tracee target functions %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
 %%%
 
diff --git a/erts/emulator/utils/make_tables b/erts/emulator/utils/make_tables
index 91efb4c023..a841f26d6a 100755
--- a/erts/emulator/utils/make_tables
+++ b/erts/emulator/utils/make_tables
@@ -167,7 +167,6 @@ typedef struct bif_entry {
 
 extern BifEntry bif_table[];
 extern Export* bif_export[];
-extern unsigned char erts_bif_trace_flags[];
 
 #define BIF_SIZE $bif_size
 
@@ -197,7 +196,6 @@ includes("export.h", "sys.h", "erl_vm.h", "erl_process.h", "bif.h",
 	 "erl_bif_table.h", "erl_atom_table.h");
 
 print "\nExport* bif_export[BIF_SIZE];\n";
-print "unsigned char erts_bif_trace_flags[BIF_SIZE];\n\n";
 
 print "BifEntry bif_table[] = {\n";
 for ($i = 0; $i < @bif; $i++) {