Merge branch 'master' into sverker/enif_select

Conflicts:
	erts/emulator/beam/erl_binary.h
	erts/emulator/beam/erl_monitors.c
	erts/emulator/beam/erl_nif.c
	erts/emulator/beam/global.h
	erts/emulator/test/nif_SUITE_data/nif_SUITE.c

70 files changed, 2464 insertions, 1273 deletions

diff --git a/erts/emulator/beam/atom.c b/erts/emulator/beam/atom.c
index 2b5ad097a0..2055c29190 100644
--- a/erts/emulator/beam/atom.c
+++ b/erts/emulator/beam/atom.c
@@ -233,10 +233,10 @@ need_convertion:
 }
 
 /*
- * erts_atom_put() may fail. If it fails THE_NON_VALUE is returned!
+ * erts_atom_put_index() may fail. Returns negative indexes for errors.
  */
-Eterm
-erts_atom_put(const byte *name, int len, ErtsAtomEncoding enc, int trunc)
+int
+erts_atom_put_index(const byte *name, int len, ErtsAtomEncoding enc, int trunc)
 {
     byte utf8_copy[MAX_ATOM_SZ_FROM_LATIN1];
     const byte *text = name;
@@ -253,7 +253,7 @@ erts_atom_put(const byte *name, int len, ErtsAtomEncoding enc, int trunc)
 	if (trunc)
 	    tlen = 0;
 	else
-	    return THE_NON_VALUE;
+	    return ATOM_MAX_CHARS_ERROR;
     }
 
     switch (enc) {
@@ -262,7 +262,7 @@ erts_atom_put(const byte *name, int len, ErtsAtomEncoding enc, int trunc)
 	    if (trunc)
 		tlen = MAX_ATOM_CHARACTERS;
 	    else
-		return THE_NON_VALUE;
+		return ATOM_MAX_CHARS_ERROR;
 	}
 #ifdef DEBUG
 	for (aix = 0; aix < len; aix++) {
@@ -276,7 +276,7 @@ erts_atom_put(const byte *name, int len, ErtsAtomEncoding enc, int trunc)
 	    if (trunc)
 		tlen = MAX_ATOM_CHARACTERS;
 	    else
-		return THE_NON_VALUE;
+		return ATOM_MAX_CHARS_ERROR;
 	}
 	no_latin1_chars = tlen;
 	latin1_to_utf8(utf8_copy, &text, &tlen);
@@ -284,7 +284,7 @@ erts_atom_put(const byte *name, int len, ErtsAtomEncoding enc, int trunc)
     case ERTS_ATOM_ENC_UTF8:
 	/* First sanity check; need to verify later */
 	if (tlen > MAX_ATOM_SZ_LIMIT && !trunc)
-	    return THE_NON_VALUE;
+	    return ATOM_MAX_CHARS_ERROR;
 	break;
     }
 
@@ -295,7 +295,7 @@ erts_atom_put(const byte *name, int len, ErtsAtomEncoding enc, int trunc)
     atom_read_unlock();
     if (aix >= 0) {
 	/* Already in table no need to verify it */
-	return make_atom(aix);
+	return aix;
     }
 
     if (enc == ERTS_ATOM_ENC_UTF8) {
@@ -314,13 +314,13 @@ erts_atom_put(const byte *name, int len, ErtsAtomEncoding enc, int trunc)
 	case ERTS_UTF8_OK_MAX_CHARS:
 	    /* Truncated... */
 	    if (!trunc)
-		return THE_NON_VALUE;
+		return ATOM_MAX_CHARS_ERROR;
 	    ASSERT(no_chars == MAX_ATOM_CHARACTERS);
 	    tlen = err_pos - text;
 	    break;
 	default:
 	    /* Bad utf8... */
-	    return THE_NON_VALUE;
+	    return ATOM_BAD_ENCODING_ERROR;
 	}
     }
 
@@ -333,7 +333,20 @@ erts_atom_put(const byte *name, int len, ErtsAtomEncoding enc, int trunc)
     atom_write_lock();
     aix = index_put(&erts_atom_table, (void*) &a);
     atom_write_unlock();
-    return make_atom(aix);
+    return aix;
+}
+
+/*
+ * erts_atom_put() may fail. If it fails THE_NON_VALUE is returned!
+ */
+Eterm
+erts_atom_put(const byte *name, int len, ErtsAtomEncoding enc, int trunc)
+{
+    int aix = erts_atom_put_index(name, len, enc, trunc);
+    if (aix >= 0)
+	return make_atom(aix);
+    else
+	return THE_NON_VALUE;
 }
 
 Eterm
diff --git a/erts/emulator/beam/atom.h b/erts/emulator/beam/atom.h
index abd3b44993..be998a46bd 100644
--- a/erts/emulator/beam/atom.h
+++ b/erts/emulator/beam/atom.h
@@ -29,6 +29,8 @@
 #define MAX_ATOM_SZ_LIMIT (4*MAX_ATOM_CHARACTERS) /* theoretical byte limit */
 #define ATOM_LIMIT (1024*1024)
 #define MIN_ATOM_TABLE_SIZE 8192
+#define ATOM_BAD_ENCODING_ERROR -1
+#define ATOM_MAX_CHARS_ERROR -2
 
 #ifndef ARCH_32
 /* Internal atom cache needs MAX_ATOM_TABLE_SIZE to be less than an
@@ -133,6 +135,7 @@ int atom_table_sz(void);	/* table size in bytes, excluding stored objects */
 
 Eterm am_atom_put(const char*, int); /* ONLY 7-bit ascii! */
 Eterm erts_atom_put(const byte *name, int len, ErtsAtomEncoding enc, int trunc);
+int erts_atom_put_index(const byte *name, int len, ErtsAtomEncoding enc, int trunc);
 void init_atom_table(void);
 void atom_info(fmtfn_t, void *);
 void dump_atoms(fmtfn_t, void *);
diff --git a/erts/emulator/beam/atom.names b/erts/emulator/beam/atom.names
index dd0a42b5ba..df2866b40e 100644
--- a/erts/emulator/beam/atom.names
+++ b/erts/emulator/beam/atom.names
@@ -195,6 +195,7 @@ atom current_stacktrace
 atom data
 atom debug_flags
 atom decimals
+atom default
 atom delay_trap
 atom dexit
 atom depth
@@ -245,6 +246,7 @@ atom Eq='=:='
 atom Eqeq='=='
 atom erl_tracer
 atom erlang
+atom erl_signal_server
 atom ERROR='ERROR'
 atom error_handler
 atom error_logger
@@ -312,6 +314,7 @@ atom global
 atom Gt='>'
 atom grun
 atom group_leader
+atom handle
 atom have_dt_utag
 atom heap_block_size
 atom heap_size
@@ -376,6 +379,7 @@ atom long_schedule
 atom low
 atom Lt='<'
 atom machine
+atom magic_ref
 atom major
 atom match
 atom match_limit
@@ -572,7 +576,7 @@ atom running_procs
 atom runtime
 atom safe
 atom save_calls
-atom scheduler 
+atom scheduler
 atom scheduler_id
 atom scheduler_wall_time
 atom scheduler_wall_time_all
@@ -600,6 +604,19 @@ atom set_tcw
 atom set_tcw_fake
 atom separate
 atom shared
+atom sighup
+atom sigterm
+atom sigusr1
+atom sigusr2
+atom sigill
+atom sigchld
+atom sigabrt
+atom sigalrm
+atom sigstop
+atom sigint
+atom sigsegv
+atom sigtstp
+atom sigquit
 atom silent
 atom size
 atom sl_alloc
diff --git a/erts/emulator/beam/beam_bif_load.c b/erts/emulator/beam/beam_bif_load.c
index c8dde8caf8..4ba8c2a669 100644
--- a/erts/emulator/beam/beam_bif_load.c
+++ b/erts/emulator/beam/beam_bif_load.c
@@ -156,11 +156,13 @@ BIF_RETTYPE code_make_stub_module_3(BIF_ALIST_3)
     Module* modp;
     Eterm res, mod;
 
-    if (!ERTS_TERM_IS_MAGIC_BINARY(BIF_ARG_1) ||
-	is_not_atom(mod = erts_module_for_prepared_code
-		    (((ProcBin*)binary_val(BIF_ARG_1))->val))) {
+    if (!is_internal_magic_ref(BIF_ARG_1))
+	BIF_ERROR(BIF_P, BADARG);
+
+    mod = erts_module_for_prepared_code(erts_magic_ref2bin(BIF_ARG_1));
+
+    if (is_not_atom(mod))
 	BIF_ERROR(BIF_P, BADARG);
-    }
 
     if (!erts_try_seize_code_write_permission(BIF_P)) {
 	ERTS_BIF_YIELD3(bif_export[BIF_code_make_stub_module_3],
@@ -229,8 +231,8 @@ prepare_loading_2(BIF_ALIST_2)
 	res = TUPLE2(hp, am_error, reason);
 	BIF_RET(res);
     }
-    hp = HAlloc(BIF_P, PROC_BIN_SIZE);
-    res = erts_mk_magic_binary_term(&hp, &MSO(BIF_P), magic);
+    hp = HAlloc(BIF_P, ERTS_MAGIC_REF_THING_SIZE);
+    res = erts_mk_magic_ref(&hp, &MSO(BIF_P), magic);
     erts_refc_dec(&magic->refc, 1);
     BIF_RET(res);
 }
@@ -239,15 +241,13 @@ BIF_RETTYPE
 has_prepared_code_on_load_1(BIF_ALIST_1)
 {
     Eterm res;
-    ProcBin* pb;
 
-    if (!ERTS_TERM_IS_MAGIC_BINARY(BIF_ARG_1)) {
+    if (!is_internal_magic_ref(BIF_ARG_1)) {
     error:
 	BIF_ERROR(BIF_P, BADARG);
     }
 
-    pb = (ProcBin*) binary_val(BIF_ARG_1);
-    res = erts_has_code_on_load(pb->val);
+    res = erts_has_code_on_load(erts_magic_ref2bin(BIF_ARG_1));
     if (res == NIL) {
 	goto error;
     }
@@ -333,13 +333,11 @@ finish_loading_1(BIF_ALIST_1)
     for (i = 0; i < n; i++) {
 	Eterm* cons = list_val(BIF_ARG_1);
 	Eterm term = CAR(cons);
-	ProcBin* pb;
 
-	if (!ERTS_TERM_IS_MAGIC_BINARY(term)) {
+	if (!is_internal_magic_ref(term)) {
 	    goto badarg;
 	}
-	pb = (ProcBin*) binary_val(term);
-	p[i].code = pb->val;
+	p[i].code = erts_magic_ref2bin(term);
 	p[i].module = erts_module_for_prepared_code(p[i].code);
 	if (p[i].module == NIL) {
 	    goto badarg;
@@ -1120,11 +1118,6 @@ check_process_code(Process* rp, Module* modp, int *redsp, int fcalls)
     if (erts_check_nif_export_in_area(rp, mod_start, mod_size))
 	return am_true;
 
-    *redsp += 1;
-
-    if (erts_check_nif_export_in_area(rp, mod_start, mod_size))
-	return am_true;
-
     *redsp += (STACK_START(rp) - rp->stop) / 32;
 
     /*
diff --git a/erts/emulator/beam/beam_bp.c b/erts/emulator/beam/beam_bp.c
index 27329a339e..b32c74ce7a 100644
--- a/erts/emulator/beam/beam_bp.c
+++ b/erts/emulator/beam/beam_bp.c
@@ -392,17 +392,17 @@ consolidate_bp_data(Module* modp, ErtsCodeInfo *ci, int local)
     }
     if (flags & ERTS_BPF_META_TRACE) {
 	dst->meta_tracer = src->meta_tracer;
-	erts_refc_inc(&dst->meta_tracer->refc, 1);
+	erts_smp_refc_inc(&dst->meta_tracer->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);
+	erts_smp_refc_inc(&dst->count->refc, 1);
     }
     if (flags & ERTS_BPF_TIME_TRACE) {
 	dst->time = src->time;
-	erts_refc_inc(&dst->time->refc, 1);
+	erts_smp_refc_inc(&dst->time->refc, 1);
 	ASSERT(dst->time->hash);
     }
 }
@@ -1576,7 +1576,7 @@ set_function_break(ErtsCodeInfo *ci, Binary *match_spec, Uint break_flags,
 	MatchSetRef(match_spec);
 	bp->meta_ms = match_spec;
 	bmt = Alloc(sizeof(BpMetaTracer));
-	erts_refc_init(&bmt->refc, 1);
+	erts_smp_refc_init(&bmt->refc, 1);
         erts_tracer_update(&meta_tracer, tracer); /* copy tracer */
 	erts_smp_atomic_init_nob(&bmt->tracer, (erts_aint_t)meta_tracer);
 	bp->meta_tracer = bmt;
@@ -1585,7 +1585,7 @@ set_function_break(ErtsCodeInfo *ci, Binary *match_spec, Uint break_flags,
 
 	ASSERT((bp->flags & ERTS_BPF_COUNT) == 0);
 	bcp = Alloc(sizeof(BpCount));
-	erts_refc_init(&bcp->refc, 1);
+	erts_smp_refc_init(&bcp->refc, 1);
 	erts_smp_atomic_init_nob(&bcp->acount, 0);
 	bp->count = bcp;
     } else if (break_flags & ERTS_BPF_TIME_TRACE) {
@@ -1594,7 +1594,7 @@ set_function_break(ErtsCodeInfo *ci, Binary *match_spec, Uint break_flags,
 
 	ASSERT((bp->flags & ERTS_BPF_TIME_TRACE) == 0);
 	bdt = Alloc(sizeof(BpDataTime));
-	erts_refc_init(&bdt->refc, 1);
+	erts_smp_refc_init(&bdt->refc, 1);
 #ifdef ERTS_DIRTY_SCHEDULERS
 	bdt->n = erts_no_schedulers + 1;
 #else
@@ -1664,7 +1664,7 @@ clear_function_break(ErtsCodeInfo *ci, Uint break_flags)
 static void
 bp_meta_unref(BpMetaTracer* bmt)
 {
-    if (erts_refc_dectest(&bmt->refc, 0) <= 0) {
+    if (erts_smp_refc_dectest(&bmt->refc, 0) <= 0) {
         ErtsTracer trc = erts_smp_atomic_read_nob(&bmt->tracer);
         ERTS_TRACER_CLEAR(&trc);
 	Free(bmt);
@@ -1674,7 +1674,7 @@ bp_meta_unref(BpMetaTracer* bmt)
 static void
 bp_count_unref(BpCount* bcp)
 {
-    if (erts_refc_dectest(&bcp->refc, 0) <= 0) {
+    if (erts_smp_refc_dectest(&bcp->refc, 0) <= 0) {
 	Free(bcp);
     }
 }
@@ -1682,7 +1682,7 @@ bp_count_unref(BpCount* bcp)
 static void
 bp_time_unref(BpDataTime* bdt)
 {
-    if (erts_refc_dectest(&bdt->refc, 0) <= 0) {
+    if (erts_smp_refc_dectest(&bdt->refc, 0) <= 0) {
 	Uint i = 0;
 	Uint j = 0;
 	Process *h_p = NULL;
diff --git a/erts/emulator/beam/beam_bp.h b/erts/emulator/beam/beam_bp.h
index cccd395e0a..3dba3cc1b5 100644
--- a/erts/emulator/beam/beam_bp.h
+++ b/erts/emulator/beam/beam_bp.h
@@ -41,7 +41,7 @@ typedef struct {
 typedef struct bp_data_time {     /* Call time */
     Uint n;
     bp_time_hash_t *hash;
-    erts_refc_t refc;
+    erts_smp_refc_t refc;
 } BpDataTime;
 
 typedef struct {
@@ -51,12 +51,12 @@ typedef struct {
 
 typedef struct {
     erts_smp_atomic_t acount;
-    erts_refc_t refc;
+    erts_smp_refc_t refc;
 } BpCount;
 
 typedef struct {
     erts_smp_atomic_t tracer;
-    erts_refc_t refc;
+    erts_smp_refc_t refc;
 } BpMetaTracer;
 
 typedef struct generic_bp_data {
diff --git a/erts/emulator/beam/beam_emu.c b/erts/emulator/beam/beam_emu.c
index 85d92321b8..49f932a1c2 100644
--- a/erts/emulator/beam/beam_emu.c
+++ b/erts/emulator/beam/beam_emu.c
@@ -126,7 +126,7 @@ do {                                     \
 
 /*
  * We reuse some of fields in the save area in the process structure.
- * This is safe to do, since this space is only activly used when
+ * This is safe to do, since this space is only actively used when
  * the process is switched out.
  */
 #define REDS_IN(p)  ((p)->def_arg_reg[5])
@@ -220,7 +220,7 @@ BeamInstr* em_call_bif_e;
 
 /* NOTE These should be the only variables containing trace instructions.
 **      Sometimes tests are form the instruction value, and sometimes
-**      for the refering variable (one of these), and rouge references
+**      for the referring variable (one of these), and rouge references
 **      will most likely cause chaos.
 */
 BeamInstr beam_return_to_trace[1];   /* OpCode(i_return_to_trace) */
@@ -2993,7 +2993,7 @@ do {						\
  }
 
  /*
-  * An error occured in an arithmetic operation or test that could
+  * An error occurred in an arithmetic operation or test that could
   * appear either in a head or in a body.
   * In a head, execution should continue at failure address in Arg(0).
   * In a body, Arg(0) == 0 and an exception should be raised.
@@ -6268,7 +6268,7 @@ apply_bif_error_adjustment(Process *p, Export *ep,
 	 * stackframe correct. Without the following adjustment,
 	 * 'p->cp' will point into the function that called
 	 * current function when handling the error. We add a
-	 * dummy stackframe in order to achive this.
+	 * dummy stackframe in order to achieve this.
 	 *
 	 * Note that these BIFs unconditionally will cause
 	 * an exception to be raised. That is, our modifications
@@ -6612,7 +6612,7 @@ erts_hibernate(Process* c_p, Eterm module, Eterm function, Eterm args, Eterm* re
 #ifndef ERTS_SMP
 	if (ERTS_PROC_IS_EXITING(c_p)) {
 	    /*
-	     * See comment in the begining of the function...
+	     * See comment in the beginning of the function...
 	     *
 	     * This second test is needed since gc might be traced.
 	     */
@@ -6846,7 +6846,7 @@ new_fun(Process* p, Eterm* reg, ErlFunEntry* fe, int num_free)
     p->htop = hp + needed;
     funp = (ErlFunThing *) hp;
     hp = funp->env;
-    erts_refc_inc(&fe->refc, 2);
+    erts_smp_refc_inc(&fe->refc, 2);
     funp->thing_word = HEADER_FUN;
     funp->next = MSO(p).first;
     MSO(p).first = (struct erl_off_heap_header*) funp;
diff --git a/erts/emulator/beam/beam_load.c b/erts/emulator/beam/beam_load.c
index 119c004f1f..48206a75a8 100644
--- a/erts/emulator/beam/beam_load.c
+++ b/erts/emulator/beam/beam_load.c
@@ -157,13 +157,15 @@ typedef struct {
 #define STR_CHUNK 2
 #define IMP_CHUNK 3
 #define EXP_CHUNK 4
-#define NUM_MANDATORY 5
+#define MIN_MANDATORY 1
+#define MAX_MANDATORY 5
 
 #define LAMBDA_CHUNK 5
 #define LITERAL_CHUNK 6
 #define ATTR_CHUNK 7
 #define COMPILE_CHUNK 8
 #define LINE_CHUNK 9
+#define UTF8_ATOM_CHUNK 10
 
 #define NUM_CHUNK_TYPES (sizeof(chunk_types)/sizeof(chunk_types[0]))
 
@@ -173,9 +175,13 @@ typedef struct {
 
 static Uint chunk_types[] = {
     /*
-     * Mandatory chunk types -- these MUST be present.
+     * Atom chunk types -- Atom or AtU8 MUST be present.
      */
     MakeIffId('A', 't', 'o', 'm'), /* 0 */
+
+    /*
+     * Mandatory chunk types -- these MUST be present.
+     */
     MakeIffId('C', 'o', 'd', 'e'), /* 1 */
     MakeIffId('S', 't', 'r', 'T'), /* 2 */
     MakeIffId('I', 'm', 'p', 'T'), /* 3 */
@@ -189,6 +195,7 @@ static Uint chunk_types[] = {
     MakeIffId('A', 't', 't', 'r'), /* 7 */
     MakeIffId('C', 'I', 'n', 'f'), /* 8 */
     MakeIffId('L', 'i', 'n', 'e'), /* 9 */
+    MakeIffId('A', 't', 'U', '8'), /* 10 */
 };
 
 /*
@@ -490,9 +497,9 @@ static Eterm stub_insert_new_code(Process *c_p, ErtsProcLocks c_p_locks,
 #endif
 static int init_iff_file(LoaderState* stp, byte* code, Uint size);
 static int scan_iff_file(LoaderState* stp, Uint* chunk_types,
-			 Uint num_types, Uint num_mandatory);
+			 Uint num_types);
 static int verify_chunks(LoaderState* stp);
-static int load_atom_table(LoaderState* stp);
+static int load_atom_table(LoaderState* stp, ErtsAtomEncoding enc);
 static int load_import_table(LoaderState* stp);
 static int read_export_table(LoaderState* stp);
 static int is_bif(Eterm mod, Eterm func, unsigned arity);
@@ -629,7 +636,7 @@ erts_prepare_loading(Binary* magic, Process *c_p, Eterm group_leader,
     CHKALLOC();
     CHKBLK(ERTS_ALC_T_CODE,stp->code);
     if (!init_iff_file(stp, code, unloaded_size) ||
-	!scan_iff_file(stp, chunk_types, NUM_CHUNK_TYPES, NUM_MANDATORY) ||
+	!scan_iff_file(stp, chunk_types, NUM_CHUNK_TYPES) ||
 	!verify_chunks(stp)) {
 	goto load_error;
     }
@@ -674,9 +681,16 @@ erts_prepare_loading(Binary* magic, Process *c_p, Eterm group_leader,
      */
 
     CHKBLK(ERTS_ALC_T_CODE,stp->code);
-    define_file(stp, "atom table", ATOM_CHUNK);
-    if (!load_atom_table(stp)) {
-	goto load_error;
+    if (stp->chunks[UTF8_ATOM_CHUNK].size > 0) {
+        define_file(stp, "utf8 atom table", UTF8_ATOM_CHUNK);
+        if (!load_atom_table(stp, ERTS_ATOM_ENC_UTF8)) {
+            goto load_error;
+        }
+    } else {
+        define_file(stp, "atom table", ATOM_CHUNK);
+        if (!load_atom_table(stp, ERTS_ATOM_ENC_LATIN1)) {
+            goto load_error;
+        }
     }
 
     /*
@@ -1213,7 +1227,7 @@ init_iff_file(LoaderState* stp, byte* code, Uint size)
  * Scan the IFF file. The header should have been verified by init_iff_file().
  */
 static int
-scan_iff_file(LoaderState* stp, Uint* chunk_types, Uint num_types, Uint num_mandatory)
+scan_iff_file(LoaderState* stp, Uint* chunk_types, Uint num_types)
 {
     Uint count;
     Uint id;
@@ -1292,7 +1306,16 @@ verify_chunks(LoaderState* stp)
     MD5_CTX context;
 
     MD5Init(&context);
-    for (i = 0; i < NUM_MANDATORY; i++) {
+
+    if (stp->chunks[UTF8_ATOM_CHUNK].start != NULL) {
+	MD5Update(&context, stp->chunks[UTF8_ATOM_CHUNK].start, stp->chunks[UTF8_ATOM_CHUNK].size);
+    } else if (stp->chunks[ATOM_CHUNK].start != NULL) {
+	MD5Update(&context, stp->chunks[ATOM_CHUNK].start, stp->chunks[ATOM_CHUNK].size);
+    } else {
+        LoadError0(stp, "mandatory chunk of type 'Atom' or 'AtU8' not found\n");
+    }
+
+    for (i = MIN_MANDATORY; i < MAX_MANDATORY; i++) {
 	if (stp->chunks[i].start != NULL) {
 	    MD5Update(&context, stp->chunks[i].start, stp->chunks[i].size);
 	} else {
@@ -1353,7 +1376,7 @@ verify_chunks(LoaderState* stp)
 }
 
 static int
-load_atom_table(LoaderState* stp)
+load_atom_table(LoaderState* stp, ErtsAtomEncoding enc)
 {
     unsigned int i;
 
@@ -1372,7 +1395,7 @@ load_atom_table(LoaderState* stp)
 
 	GetByte(stp, n);
 	GetString(stp, atom, n);
-	stp->atom[i] = erts_atom_put(atom, n, ERTS_ATOM_ENC_LATIN1, 1);
+	stp->atom[i] = erts_atom_put(atom, n, enc, 1);
     }
 
     /*
@@ -4840,7 +4863,7 @@ final_touch(LoaderState* stp, struct erl_module_instance* inst_p)
 		/*
 		 * We are hiding a pointer into older code.
 		 */
-		erts_refc_dec(&fe->refc, 1);
+		erts_smp_refc_dec(&fe->refc, 1);
 	    }
 	    fe->address = code_ptr;
 #ifdef HIPE
@@ -5939,7 +5962,7 @@ code_get_chunk_2(BIF_ALIST_2)
 	goto error;
     }
     if (!init_iff_file(stp, start, binary_size(Bin)) ||
-	!scan_iff_file(stp, &chunk, 1, 1) ||
+	!scan_iff_file(stp, &chunk, 1) ||
 	stp->chunks[0].start == NULL) {
 	res = am_undefined;
 	goto done;
@@ -5988,7 +6011,7 @@ code_module_md5_1(BIF_ALIST_1)
     }
     stp->module = THE_NON_VALUE; /* Suppress diagnostiscs */
     if (!init_iff_file(stp, bytes, binary_size(Bin)) ||
-	!scan_iff_file(stp, chunk_types, NUM_CHUNK_TYPES, NUM_MANDATORY) ||
+	!scan_iff_file(stp, chunk_types, NUM_CHUNK_TYPES) ||
 	!verify_chunks(stp)) {
 	res = am_undefined;
 	goto done;
@@ -6252,7 +6275,7 @@ patch_funentries(Eterm Patchlist)
      *
      * Reproduced on a debug emulator with stdlib_test/qlc_SUITE:join_merge
      *
-     * erts_refc_dec(&fe->refc, 1);
+     * erts_smp_refc_dec(&fe->refc, 1);
      */
 
     if (!patch(Addresses, (Uint) fe))
@@ -6300,8 +6323,8 @@ erts_make_stub_module(Process* p, Eterm hipe_magic_bin, Eterm Beam, Eterm Info)
     stp = ERTS_MAGIC_BIN_DATA(magic);
     hipe_code = erts_alloc(ERTS_ALC_T_HIPE, sizeof(*hipe_code));
 
-    if (!ERTS_TERM_IS_MAGIC_BINARY(hipe_magic_bin) ||
-	!(hipe_magic = ((ProcBin*)binary_val(hipe_magic_bin))->val,
+    if (!is_internal_magic_ref(hipe_magic_bin) ||
+	!(hipe_magic = erts_magic_ref2bin(hipe_magic_bin),
 	  hipe_stp = hipe_get_loader_state(hipe_magic)) ||
 	hipe_stp->module == NIL || hipe_stp->text_segment == 0) {
 	goto error;
@@ -6337,7 +6360,7 @@ erts_make_stub_module(Process* p, Eterm hipe_magic_bin, Eterm Beam, Eterm Info)
     if (!init_iff_file(stp, bytes, size)) {
 	goto error;
     }
-    if (!scan_iff_file(stp, chunk_types, NUM_CHUNK_TYPES, NUM_MANDATORY) ||
+    if (!scan_iff_file(stp, chunk_types, NUM_CHUNK_TYPES) ||
 	!verify_chunks(stp)) {
 	goto error;
     }
@@ -6345,9 +6368,16 @@ erts_make_stub_module(Process* p, Eterm hipe_magic_bin, Eterm Beam, Eterm Info)
     if (!read_code_header(stp)) {
 	goto error;
     }
-    define_file(stp, "atom table", ATOM_CHUNK);
-    if (!load_atom_table(stp)) {
-	goto error;
+    if (stp->chunks[UTF8_ATOM_CHUNK].size > 0) {
+        define_file(stp, "utf8 atom table", UTF8_ATOM_CHUNK);
+        if (!load_atom_table(stp, ERTS_ATOM_ENC_UTF8)) {
+            goto error;
+        }
+    } else {
+        define_file(stp, "atom table", ATOM_CHUNK);
+        if (!load_atom_table(stp, ERTS_ATOM_ENC_LATIN1)) {
+            goto error;
+        }
     }
     define_file(stp, "export table", EXP_CHUNK);
     if (!stub_read_export_table(stp)) {
@@ -6539,8 +6569,8 @@ int erts_commit_hipe_patch_load(Eterm hipe_magic_bin)
     HipeModule *hipe_code;
     Module* modp;
 
-    if (!ERTS_TERM_IS_MAGIC_BINARY(hipe_magic_bin) ||
-	!(hipe_magic = ((ProcBin*)binary_val(hipe_magic_bin))->val,
+    if (!is_internal_magic_ref(hipe_magic_bin) ||
+	!(hipe_magic = erts_magic_ref2bin(hipe_magic_bin),
 	  hipe_stp = hipe_get_loader_state(hipe_magic)) ||
 	hipe_stp->module == NIL || hipe_stp->text_segment == 0) {
 	return 0;
diff --git a/erts/emulator/beam/bif.c b/erts/emulator/beam/bif.c
index 51057c3ebb..614bedab12 100644
--- a/erts/emulator/beam/bif.c
+++ b/erts/emulator/beam/bif.c
@@ -199,7 +199,7 @@ BIF_RETTYPE link_1(BIF_ALIST_1)
 		goto res_no_proc;
 	    case ERTS_PORT_OP_SCHEDULED:
 		if (refp) {
-		    ASSERT(is_internal_ref(ref));
+		    ASSERT(is_internal_ordinary_ref(ref));
 		    BIF_TRAP3(await_port_send_result_trap, BIF_P, ref, am_true, am_true);
 		}
 	    default:
@@ -467,7 +467,7 @@ demonitor_local_port(Process *origin, Eterm ref, Eterm target)
 }
 
 /* Can return atom true, false, yield, internal_error, badarg or
- * THE_NON_VALUE if error occured or trap has been set up
+ * THE_NON_VALUE if error occurred or trap has been set up
  */
 static
 BIF_RETTYPE demonitor(Process *c_p, Eterm ref, Eterm *multip)
@@ -597,7 +597,7 @@ BIF_RETTYPE demonitor_2(BIF_ALIST_2)
 
     switch (demonitor(BIF_P, BIF_ARG_1, &multi)) {
     case THE_NON_VALUE:
-        /* If other error occured or trap has been set up - pass through */
+        /* If other error occurred or trap has been set up - pass through */
         BIF_RET(THE_NON_VALUE);
     case am_false:
 	if (info)
@@ -782,7 +782,7 @@ local_name_monitor(Process *self, Eterm type, Eterm target_name)
         case ERTS_PORT_OP_DONE:
             return ret;
         case ERTS_PORT_OP_SCHEDULED: { /* Scheduled a signal */
-            ASSERT(is_internal_ref(ret));
+            ASSERT(is_internal_ordinary_ref(ret));
             BIF_TRAP3(await_port_send_result_trap, self,
                       ret, am_true, ret);
             /* bif_trap returns */
@@ -1180,7 +1180,7 @@ BIF_RETTYPE unlink_1(BIF_ALIST_1)
 		res = erts_port_unlink(BIF_P, prt, BIF_P->common.id, refp);
 
 		if (refp && res == ERTS_PORT_OP_SCHEDULED) {
-		    ASSERT(is_internal_ref(ref));
+		    ASSERT(is_internal_ordinary_ref(ref));
 		    BIF_TRAP3(await_port_send_result_trap, BIF_P, ref, am_true, am_true);
 		}
 	    }
@@ -1586,7 +1586,7 @@ BIF_RETTYPE exit_2(BIF_ALIST_2)
 	     ERTS_BIF_CHK_EXITED(BIF_P);
 
 	     if (refp && res == ERTS_PORT_OP_SCHEDULED) {
-		 ASSERT(is_internal_ref(ref));
+		 ASSERT(is_internal_ordinary_ref(ref));
 		 BIF_TRAP3(await_port_send_result_trap, BIF_P, ref, am_true, am_true);
 	     }
 
@@ -2221,7 +2221,7 @@ do_send(Process *p, Eterm to, Eterm msg, Eterm *refp, ErtsSendContext* ctx)
 		/* Fall through */
 	    case ERTS_PORT_OP_SCHEDULED:
 		if (is_not_nil(*refp)) {
-		    ASSERT(is_internal_ref(*refp));
+		    ASSERT(is_internal_ordinary_ref(*refp));
 		    ret_val = SEND_AWAIT_RESULT;
 		}
 		break;
@@ -2409,7 +2409,7 @@ BIF_RETTYPE send_3(BIF_ALIST_3)
 	ERTS_BIF_PREP_YIELD_RETURN(retval, p, am_ok);
         break;
     case SEND_AWAIT_RESULT:
-	ASSERT(is_internal_ref(ref));
+	ASSERT(is_internal_ordinary_ref(ref));
 	ERTS_BIF_PREP_TRAP3(retval, await_port_send_result_trap, p, ref, am_nosuspend, am_ok);
 	break;
     case SEND_BADARG:
@@ -2446,7 +2446,7 @@ BIF_RETTYPE send_2(BIF_ALIST_2)
 
 static BIF_RETTYPE dsend_continue_trap_1(BIF_ALIST_1)
 {
-    Binary* bin = ((ProcBin*) binary_val(BIF_ARG_1))->val;
+    Binary* bin = erts_magic_ref2bin(BIF_ARG_1);
     ErtsSendContext* ctx = (ErtsSendContext*) ERTS_MAGIC_BIN_DATA(bin);
     Sint initial_reds = (Sint) (ERTS_BIF_REDS_LEFT(BIF_P) * TERM_TO_BINARY_LOOP_FACTOR);
     int result;
@@ -2526,7 +2526,7 @@ Eterm erl_send(Process *p, Eterm to, Eterm msg)
 	ERTS_BIF_PREP_YIELD_RETURN(retval, p, msg);
         break;
     case SEND_AWAIT_RESULT:
-	ASSERT(is_internal_ref(ref));
+	ASSERT(is_internal_ordinary_ref(ref));
 	ERTS_BIF_PREP_TRAP3(retval,
 			    await_port_send_result_trap, p, ref, msg, msg);
 	break;
@@ -3022,8 +3022,8 @@ BIF_RETTYPE atom_to_list_1(BIF_ALIST_1)
 BIF_RETTYPE list_to_atom_1(BIF_ALIST_1)
 {
     Eterm res;
-    char *buf = (char *) erts_alloc(ERTS_ALC_T_TMP, MAX_ATOM_CHARACTERS);
-    Sint i = intlist_to_buf(BIF_ARG_1, buf, MAX_ATOM_CHARACTERS);
+    byte *buf = (byte *) erts_alloc(ERTS_ALC_T_TMP, MAX_ATOM_SZ_LIMIT);
+    Sint i = erts_unicode_list_to_buf(BIF_ARG_1, buf, MAX_ATOM_CHARACTERS);
 
     if (i < 0) {
 	erts_free(ERTS_ALC_T_TMP, (void *) buf);
@@ -3033,7 +3033,7 @@ BIF_RETTYPE list_to_atom_1(BIF_ALIST_1)
 	}
 	BIF_ERROR(BIF_P, BADARG);
     }
-    res = erts_atom_put((byte *) buf, i, ERTS_ATOM_ENC_LATIN1, 1);
+    res = erts_atom_put(buf, i, ERTS_ATOM_ENC_UTF8, 1);
     ASSERT(is_atom(res));
     erts_free(ERTS_ALC_T_TMP, (void *) buf);
     BIF_RET(res);
@@ -3043,17 +3043,17 @@ BIF_RETTYPE list_to_atom_1(BIF_ALIST_1)
  
 BIF_RETTYPE list_to_existing_atom_1(BIF_ALIST_1)
 {
-    Sint i;
-    char *buf = (char *) erts_alloc(ERTS_ALC_T_TMP, MAX_ATOM_CHARACTERS);
+    byte *buf = (byte *) erts_alloc(ERTS_ALC_T_TMP, MAX_ATOM_SZ_LIMIT);
+    Sint i = erts_unicode_list_to_buf(BIF_ARG_1, buf, MAX_ATOM_CHARACTERS);
 
-    if ((i = intlist_to_buf(BIF_ARG_1, buf, MAX_ATOM_CHARACTERS)) < 0) {
+    if (i < 0) {
     error:
 	erts_free(ERTS_ALC_T_TMP, (void *) buf);
 	BIF_ERROR(BIF_P, BADARG);
     } else {
 	Eterm a;
 	
-	if (erts_atom_get(buf, i, &a, ERTS_ATOM_ENC_LATIN1)) {
+	if (erts_atom_get((char *) buf, i, &a, ERTS_ATOM_ENC_UTF8)) {
 	    erts_free(ERTS_ALC_T_TMP, (void *) buf);
 	    BIF_RET(a);
 	} else {
@@ -4109,6 +4109,7 @@ BIF_RETTYPE ref_to_list_1(BIF_ALIST_1)
 {
     if (is_not_ref(BIF_ARG_1))
 	BIF_ERROR(BIF_P, BADARG);
+    erts_magic_ref_save_bin(BIF_ARG_1);
     BIF_RET(term2list_dsprintf(BIF_P, BIF_ARG_1));
 }
 
@@ -4723,25 +4724,6 @@ BIF_RETTYPE system_flag_2(BIF_ALIST_2)
 
 /**********************************************************************/
 
-BIF_RETTYPE hash_2(BIF_ALIST_2)
-{
-    Uint32 hash;
-    Sint range;
-
-    if (is_not_small(BIF_ARG_2)) {
-	BIF_ERROR(BIF_P, BADARG);
-    }
-    if ((range = signed_val(BIF_ARG_2)) <= 0) {  /* [1..MAX_SMALL] */
-	BIF_ERROR(BIF_P, BADARG);
-    }
-#if defined(ARCH_64)
-    if (range > ((1L << 27) - 1))
-	BIF_ERROR(BIF_P, BADARG);
-#endif
-    hash = make_broken_hash(BIF_ARG_1);
-    BIF_RET(make_small(1 + (hash % range)));   /* [1..range] */
-}
-
 BIF_RETTYPE phash_2(BIF_ALIST_2)
 {
     Uint32 hash;
@@ -5574,12 +5556,10 @@ BIF_RETTYPE dt_restore_tag_1(BIF_ALIST_1)
 	    SEQ_TRACE_TOKEN(BIF_P) = am_have_dt_utag;
 	}
     }
-#else    
+#else
     if (BIF_ARG_1 != am_true) {
 	BIF_ERROR(BIF_P,BADARG);
     }
 #endif
     BIF_RET(am_true);
 }
-
-
diff --git a/erts/emulator/beam/bif.tab b/erts/emulator/beam/bif.tab
index 47fdcfa7a4..a6510ace81 100644
--- a/erts/emulator/beam/bif.tab
+++ b/erts/emulator/beam/bif.tab
@@ -670,9 +670,4 @@ gcbif erlang:ceil/1
 bif math:floor/1
 bif math:ceil/1
 bif math:fmod/2
-
-#
-# Obsolete
-#
-
-bif erlang:hash/2
+bif os:set_signal/2
diff --git a/erts/emulator/beam/binary.c b/erts/emulator/beam/binary.c
index 9894992b70..4dd8316dad 100644
--- a/erts/emulator/beam/binary.c
+++ b/erts/emulator/beam/binary.c
@@ -47,13 +47,8 @@ void
 erts_init_binary(void)
 {
     /* Verify Binary alignment... */
-    if ((((UWord) &((Binary *) 0)->orig_bytes[0]) % ((UWord) 8)) != 0) {
-	/* I assume that any compiler should be able to optimize this
-	   away. If not, this test is not very expensive... */
-	erts_exit(ERTS_ABORT_EXIT,
-		 "Internal error: Address of orig_bytes[0] of a Binary"
-		 " is *not* 8-byte aligned\n");
-    }
+    ERTS_CT_ASSERT((offsetof(Binary,orig_bytes) % 8) == 0);
+    ERTS_CT_ASSERT((offsetof(ErtsMagicBinary,u.aligned.data) % 8) == 0);
 
     erts_init_trap_export(&binary_to_list_continue_export,
 			  am_erts_internal, am_binary_to_list_continue, 1,
@@ -107,7 +102,7 @@ new_binary(Process *p, byte *buf, Uint len)
     pb->flags = 0;
 
     /*
-     * Miscellanous updates. Return the tagged binary.
+     * Miscellaneous updates. Return the tagged binary.
      */
     OH_OVERHEAD(&(MSO(p)), pb->size / sizeof(Eterm));
     return make_binary(pb);
@@ -144,7 +139,7 @@ Eterm erts_new_mso_binary(Process *p, byte *buf, Uint len)
     pb->flags = 0;
 
     /*
-     * Miscellanous updates. Return the tagged binary.
+     * Miscellaneous updates. Return the tagged binary.
      */
     OH_OVERHEAD(&(MSO(p)), pb->size / sizeof(Eterm));
     return make_binary(pb);
@@ -446,18 +441,17 @@ binary_to_list(Process *c_p, Eterm *hp, Eterm tail, byte *bytes,
 	sp->size = size;
 	sp->bitoffs = bitoffs;
 
-	hp = HAlloc(c_p, PROC_BIN_SIZE);
-	mb = erts_mk_magic_binary_term(&hp, &MSO(c_p), mbp);
+	hp = HAlloc(c_p, ERTS_MAGIC_REF_THING_SIZE);
+	mb = erts_mk_magic_ref(&hp, &MSO(c_p), mbp);
 	return binary_to_list_chunk(c_p, mb, sp, reds_left, 0);
     }
 }
 
 static BIF_RETTYPE binary_to_list_continue(BIF_ALIST_1)
 {
-    Binary *mbp = ((ProcBin *) binary_val(BIF_ARG_1))->val;
+    Binary *mbp = erts_magic_ref2bin(BIF_ARG_1);
 
     ASSERT(ERTS_MAGIC_BIN_DESTRUCTOR(mbp) == b2l_state_destructor);
-
     ASSERT(BIF_P->flags & F_DISABLE_GC);
 
     return binary_to_list_chunk(BIF_P,
@@ -794,8 +788,8 @@ list_to_binary_chunk(Eterm mb_eterm,
 	    ERTS_L2B_STATE_MOVE(new_sp, sp);
 	    sp = new_sp;
 
-	    hp = HAlloc(c_p, PROC_BIN_SIZE);
-	    mb_eterm = erts_mk_magic_binary_term(&hp, &MSO(c_p), mbp);
+	    hp = HAlloc(c_p, ERTS_MAGIC_REF_THING_SIZE);
+	    mb_eterm = erts_mk_magic_ref(&hp, &MSO(c_p), mbp);
 
 	    ASSERT(is_value(mb_eterm));
 
@@ -841,9 +835,9 @@ list_to_binary_chunk(Eterm mb_eterm,
 
 static BIF_RETTYPE list_to_binary_continue(BIF_ALIST_1)
 {
-    Binary *mbp = ((ProcBin *) binary_val(BIF_ARG_1))->val;
-    ASSERT(ERTS_MAGIC_BIN_DESTRUCTOR(mbp) == l2b_state_destructor);
+    Binary *mbp = erts_magic_ref2bin(BIF_ARG_1);
 
+    ASSERT(ERTS_MAGIC_BIN_DESTRUCTOR(mbp) == l2b_state_destructor);
     ASSERT(BIF_P->flags & F_DISABLE_GC);
 
     return list_to_binary_chunk(BIF_ARG_1,
diff --git a/erts/emulator/beam/break.c b/erts/emulator/beam/break.c
index 5509c50a8e..787f1e4b7e 100644
--- a/erts/emulator/beam/break.c
+++ b/erts/emulator/beam/break.c
@@ -389,6 +389,16 @@ info(fmtfn_t to, void *to_arg)
 
 }
 
+static int code_size(struct erl_module_instance* modi)
+{
+    ErtsLiteralArea* lit = modi->code_hdr->literal_area;
+    int size = modi->code_length;
+    if (lit) {
+        size += (lit->end - lit->start) * sizeof(Eterm);
+    }
+    return size;
+}
+
 void
 loaded(fmtfn_t to, void *to_arg)
 {
@@ -409,9 +419,9 @@ loaded(fmtfn_t to, void *to_arg)
 	if ((modp = module_code(i, code_ix)) != NULL &&
 	    ((modp->curr.code_length != 0) ||
 	     (modp->old.code_length != 0))) {
-	    cur += modp->curr.code_length;
+	    cur += code_size(&modp->curr);
 	    if (modp->old.code_length != 0) {
-		old += modp->old.code_length;
+		old += code_size(&modp->old);
 	    }
 	}
     }
@@ -432,12 +442,12 @@ loaded(fmtfn_t to, void *to_arg)
 	    ((modp->curr.code_length != 0) ||
 	     (modp->old.code_length != 0))) {
 		erts_print(to, to_arg, "%T", make_atom(modp->module));
-		cur += modp->curr.code_length;
-		erts_print(to, to_arg, " %d", modp->curr.code_length );
+		cur += code_size(&modp->curr);
+		erts_print(to, to_arg, " %d", code_size(&modp->curr));
 		if (modp->old.code_length != 0) {
 		    erts_print(to, to_arg, " (%d old)",
-			       modp->old.code_length );
-		    old += modp->old.code_length;
+			       code_size(&modp->old));
+		    old += code_size(&modp->old);
 		}
 		erts_print(to, to_arg, "\n");
 	    }
@@ -452,7 +462,7 @@ loaded(fmtfn_t to, void *to_arg)
 		erts_print(to, to_arg, "%T", make_atom(modp->module));
 		erts_print(to, to_arg, "\n");
 		erts_print(to, to_arg, "Current size: %d\n",
-			   modp->curr.code_length);
+			   code_size(&modp->curr));
 		code = modp->curr.code_hdr;
 		if (code != NULL && code->attr_ptr) {
 		    erts_print(to, to_arg, "Current attributes: ");
@@ -466,7 +476,7 @@ loaded(fmtfn_t to, void *to_arg)
 		}
 
 		if (modp->old.code_length != 0) {
-		    erts_print(to, to_arg, "Old size: %d\n", modp->old.code_length);
+		    erts_print(to, to_arg, "Old size: %d\n", code_size(&modp->old));
 		    code = modp->old.code_hdr;
 		    if (code->attr_ptr) {
 			erts_print(to, to_arg, "Old attributes: ");
@@ -657,7 +667,7 @@ bin_check(void)
 		erts_printf("%p orig_size: %bpd, norefs = %bpd\n",
 			    bp->val, 
 			    bp->val->orig_size, 
-			    erts_smp_atomic_read_nob(&bp->val->refc));
+			    erts_refc_read(&bp->val->refc, 1));
 	    }
 	}
 	if (printed) {
@@ -727,6 +737,8 @@ erl_crash_dump_v(char *file, int line, char* fmt, va_list args)
      * We have to be very very careful when doing this as the schedulers
      * could be anywhere.
      */
+    sys_init_suspend_handler();
+
     for (i = 0; i < erts_no_schedulers; i++) {
         erts_tid_t tid = ERTS_SCHEDULER_IX(i)->tid;
         if (!erts_equal_tids(tid,erts_thr_self()))
diff --git a/erts/emulator/beam/copy.c b/erts/emulator/beam/copy.c
index c4dcd6a3cc..85db23393c 100644
--- a/erts/emulator/beam/copy.c
+++ b/erts/emulator/beam/copy.c
@@ -850,7 +850,7 @@ Eterm copy_struct_x(Eterm obj, Uint sz, Eterm** hpp, ErlOffHeap* off_heap, Uint
 		    funp = (ErlFunThing *) tp;
 		    funp->next = off_heap->first;
 		    off_heap->first = (struct erl_off_heap_header*) funp;
-		    erts_refc_inc(&funp->fe->refc, 2);
+		    erts_smp_refc_inc(&funp->fe->refc, 2);
 		    *argp = make_fun(tp);
 		}
 		break;
@@ -858,20 +858,24 @@ Eterm copy_struct_x(Eterm obj, Uint sz, Eterm** hpp, ErlOffHeap* off_heap, Uint
 	    case EXTERNAL_PORT_SUBTAG:
 	    case EXTERNAL_REF_SUBTAG:
 		{
-		  ExternalThing *etp = (ExternalThing *) htop;
-
+		  ExternalThing *etp = (ExternalThing *) objp;
+		  erts_smp_refc_inc(&etp->node->refc, 2);
+		}
+	    L_off_heap_node_container_common:
+		{
+		  struct erl_off_heap_header *ohhp;
+		  ohhp = (struct erl_off_heap_header *) htop;
 		  i  = thing_arityval(hdr) + 1;
+		  *argp = make_boxed(htop);
 		  tp = htop;
 
 		  while (i--)  {
 		    *htop++ = *objp++;
 		  }
 
-		  etp->next = off_heap->first;
-		  off_heap->first = (struct erl_off_heap_header*)etp;
-		  erts_refc_inc(&etp->node->refc, 2);
+		  ohhp->next = off_heap->first;
+		  off_heap->first = ohhp;
 
-		  *argp = make_external(tp);
 		}
 		break;
 	    case MAP_SUBTAG:
@@ -899,6 +903,13 @@ Eterm copy_struct_x(Eterm obj, Uint sz, Eterm** hpp, ErlOffHeap* off_heap, Uint
 	    case BIN_MATCHSTATE_SUBTAG:
 		erts_exit(ERTS_ABORT_EXIT,
 			 "copy_struct: matchstate term not allowed");
+	    case REF_SUBTAG:
+		if (is_magic_ref_thing(objp)) {
+		    ErtsMRefThing *mreft = (ErtsMRefThing *) objp;
+		    erts_refc_inc(&mreft->mb->refc, 2);
+		    goto L_off_heap_node_container_common;
+		}
+		/* Fall through... */
 	    default:
 		i = thing_arityval(hdr)+1;
 		hbot -= i;
@@ -1525,7 +1536,7 @@ Uint copy_shared_perform(Eterm obj, Uint size, erts_shcopy_t *info,
 		}
 		funp->next = off_heap->first;
 		off_heap->first = (struct erl_off_heap_header*) funp;
-		erts_refc_inc(&funp->fe->refc, 2);
+		erts_smp_refc_inc(&funp->fe->refc, 2);
 		goto cleanup_next;
 	    }
 	    case MAP_SUBTAG:
@@ -1649,20 +1660,33 @@ Uint copy_shared_perform(Eterm obj, Uint size, erts_shcopy_t *info,
 	    }
 	    case EXTERNAL_PID_SUBTAG:
 	    case EXTERNAL_PORT_SUBTAG:
-	    case EXTERNAL_REF_SUBTAG: {
-		ExternalThing *etp = (ExternalThing *) hp;
+	    case EXTERNAL_REF_SUBTAG:
+	    {
+		ExternalThing *etp = (ExternalThing *) ptr;
+		erts_smp_refc_inc(&etp->node->refc, 2);
+	    }
+	  off_heap_node_container_common:
+	    {
+		struct erl_off_heap_header *ohhp;
+		ohhp = (struct erl_off_heap_header *) hp;
 		sz = thing_arityval(hdr);
-		*resp = make_external(hp);
+		*resp = make_boxed(hp);
 		*hp++ = hdr;
 		ptr++;
 		while (sz-- > 0) {
 		    *hp++ = *ptr++;
 		}
-		etp->next = off_heap->first;
-		off_heap->first = (struct erl_off_heap_header*) etp;
-		erts_refc_inc(&etp->node->refc, 2);
+		ohhp->next = off_heap->first;
+		off_heap->first = ohhp;
 		goto cleanup_next;
 	    }
+	    case REF_SUBTAG:
+		if (is_magic_ref_thing(ptr)) {
+		    ErtsMRefThing *mreft = (ErtsMRefThing *) ptr;
+		    erts_refc_inc(&mreft->mb->refc, 2);
+		    goto off_heap_node_container_common;
+		}
+		/* Fall through... */
 	    default:
 		sz = thing_arityval(hdr);
 		*resp = make_boxed(hp);
@@ -1836,7 +1860,7 @@ Eterm copy_shallow(Eterm* ptr, Uint sz, Eterm** hpp, ErlOffHeap* off_heap)
 	    case FUN_SUBTAG:
 		{
 		    ErlFunThing* funp = (ErlFunThing *) (tp-1);
-		    erts_refc_inc(&funp->fe->refc, 2);
+		    erts_smp_refc_inc(&funp->fe->refc, 2);
 		}
 		goto off_heap_common;
 	    case EXTERNAL_PID_SUBTAG:
@@ -1844,7 +1868,7 @@ Eterm copy_shallow(Eterm* ptr, Uint sz, Eterm** hpp, ErlOffHeap* off_heap)
 	    case EXTERNAL_REF_SUBTAG:
 		{
 		    ExternalThing* etp = (ExternalThing *) (tp-1);
-		    erts_refc_inc(&etp->node->refc, 2);
+		    erts_smp_refc_inc(&etp->node->refc, 2);
 		}
 	    off_heap_common:
 		{
@@ -1859,6 +1883,15 @@ Eterm copy_shallow(Eterm* ptr, Uint sz, Eterm** hpp, ErlOffHeap* off_heap)
 		    off_heap->first = ohh;
 		}
 		break;
+	    case REF_SUBTAG: {
+		ErtsRefThing *rtp = (ErtsRefThing *) (tp - 1);
+		if (is_magic_ref_thing(rtp)) {
+		    ErtsMRefThing *mreft = (ErtsMRefThing *) rtp;
+		    erts_refc_inc(&mreft->mb->refc, 2);
+		    goto off_heap_common;
+		}
+		/* Fall through... */
+	    }
 	    default:
 		{
 		    int tari = header_arity(val);
@@ -1959,6 +1992,9 @@ move_one_frag(Eterm** hpp, ErlHeapFragment* frag, ErlOffHeap* off_heap, int lite
 	    ASSERT(ptr + header_arity(val) < end);
 	    MOVE_BOXED(ptr, val, hp, &dummy_ref);
 	    switch (val & _HEADER_SUBTAG_MASK) {
+	    case REF_SUBTAG:
+		if (is_ordinary_ref_thing(hdr))
+		    break;
 	    case REFC_BINARY_SUBTAG:
 	    case FUN_SUBTAG:
 	    case EXTERNAL_PID_SUBTAG:
diff --git a/erts/emulator/beam/dist.c b/erts/emulator/beam/dist.c
index 016abaf717..d1c2da9074 100644
--- a/erts/emulator/beam/dist.c
+++ b/erts/emulator/beam/dist.c
@@ -264,7 +264,7 @@ static void doit_monitor_net_exits(ErtsMonitor *mon, void *vnecp)
 	goto done;
 
     if (mon->type == MON_ORIGIN) {
-	/* local pid is beeing monitored */
+	/* local pid is being monitored */
 	rmon = erts_remove_monitor(&ERTS_P_MONITORS(rp), mon->ref);
 	/* ASSERT(rmon != NULL); nope, can happen during process exit */
 	if (rmon != NULL) {
@@ -744,7 +744,7 @@ Eterm erts_dsend_export_trap_context(Process* p, ErtsSendContext* ctx)
     Binary* ctx_bin = erts_create_magic_binary(sizeof(struct exported_ctx),
 					       erts_dsend_context_dtor);
     struct exported_ctx* dst = ERTS_MAGIC_BIN_DATA(ctx_bin);
-    Eterm* hp = HAlloc(p, PROC_BIN_SIZE);
+    Eterm* hp = HAlloc(p, ERTS_MAGIC_REF_THING_SIZE);
 
     sys_memcpy(&dst->ctx, ctx, sizeof(ErtsSendContext));
     ASSERT(ctx->dss.ctl == make_tuple(ctx->ctl_heap));
@@ -753,7 +753,7 @@ Eterm erts_dsend_export_trap_context(Process* p, ErtsSendContext* ctx)
 	sys_memcpy(&dst->acm, ctx->dss.acmp, sizeof(ErtsAtomCacheMap));
 	dst->ctx.dss.acmp = &dst->acm;
     }
-    return erts_mk_magic_binary_term(&hp, &MSO(p), ctx_bin);
+    return erts_mk_magic_ref(&hp, &MSO(p), ctx_bin);
 }
 
 
@@ -798,7 +798,7 @@ erts_dsig_send_unlink(ErtsDSigData *dsdp, Eterm local, Eterm remote)
 }
 
 
-/* A local process that's beeing monitored by a remote one exits. We send:
+/* A local process that's being monitored by a remote one exits. We send:
    {DOP_MONITOR_P_EXIT, Local pid or name, Remote pid, ref, reason},
    which is rather sad as only the ref is needed, no pid's... */
 int
@@ -2085,7 +2085,7 @@ erts_dist_command(Port *prt, int reds_limit)
 
     ERTS_SMP_LC_ASSERT(erts_lc_is_port_locked(prt));
 
-    erts_refc_inc(&dep->refc, 1); /* Otherwise dist_entry might be
+    erts_smp_refc_inc(&dep->refc, 1); /* Otherwise dist_entry might be
 				     removed if port command fails */
 
     erts_smp_atomic_set_mb(&dep->dist_cmd_scheduled, 0);
@@ -2518,7 +2518,7 @@ info_dist_entry(fmtfn_t to, void *arg, DistEntry *dep, int visible, int connecte
 
   erts_print(to, arg, "Name: %T", dep->sysname);
 #ifdef DEBUG
-  erts_print(to, arg, " (refc=%d)", erts_refc_read(&dep->refc, 0));
+  erts_print(to, arg, " (refc=%d)", erts_smp_refc_read(&dep->refc, 0));
 #endif
   erts_print(to, arg, "\n");
   if (!connected && is_nil(dep->cid)) {
diff --git a/erts/emulator/beam/dist.h b/erts/emulator/beam/dist.h
index 8799e54057..8f6be1061a 100644
--- a/erts/emulator/beam/dist.h
+++ b/erts/emulator/beam/dist.h
@@ -115,8 +115,8 @@ extern int erts_is_alive;
  * erts_dsig_prepare() prepares a send of a distributed signal.
  * One of the values defined below are returned. If the returned
  * value is another than ERTS_DSIG_PREP_CONNECTED, the
- * distributed signal cannot be sent before apropriate actions
- * have been taken. Apropriate actions would typically be setting
+ * distributed signal cannot be sent before appropriate actions
+ * have been taken. Appropriate actions would typically be setting
  * up the connection.
  */
 
diff --git a/erts/emulator/beam/dtrace-wrapper.h b/erts/emulator/beam/dtrace-wrapper.h
index 6f70d5961e..f6fc28b801 100644
--- a/erts/emulator/beam/dtrace-wrapper.h
+++ b/erts/emulator/beam/dtrace-wrapper.h
@@ -74,7 +74,7 @@
 
 #if defined(_SDT_PROBE) && !defined(STAP_PROBE11)
 /* SLF: This is Ubuntu 11-style SystemTap hackery */
-/* work arround for missing STAP macro */
+/* workaround for missing STAP macro */
 #define STAP_PROBE11(provider,name,arg1,arg2,arg3,arg4,arg5,arg6,arg7,arg8,arg9,arg10,arg11) \
   _SDT_PROBE(provider, name, 11, \
              (arg1,arg2,arg3,arg4,arg5,arg6,arg7,arg8,arg9,arg10,arg11))
diff --git a/erts/emulator/beam/erl_alloc.c b/erts/emulator/beam/erl_alloc.c
index 43c22e2e7c..a374593c5d 100644
--- a/erts/emulator/beam/erl_alloc.c
+++ b/erts/emulator/beam/erl_alloc.c
@@ -48,6 +48,7 @@
 #if defined(ERTS_ALC_T_DRV_SEL_D_STATE) || defined(ERTS_ALC_T_DRV_EV_D_STATE)
 #include "erl_check_io.h"
 #endif
+#include "erl_bif_unique.h"
 
 #define GET_ERL_GF_ALLOC_IMPL
 #include "erl_goodfit_alloc.h"
@@ -152,8 +153,7 @@ typedef struct {
     int internal;
     Uint req_sched;
     Process *proc;
-    Eterm ref;
-    Eterm ref_heap[REF_THING_SIZE];
+    ErtsIRefStorage iref;
     int allocs[ERTS_ALC_INFO_A_END - ERTS_ALC_A_MIN + 1];
 } ErtsAllocInfoReq;
 
@@ -684,6 +684,10 @@ erts_alloc_init(int *argc, char **argv, ErtsAllocInitOpts *eaiop)
 #endif
     fix_type_sizes[ERTS_ALC_FIX_TYPE_IX(ERTS_ALC_T_NIF_EXP_TRACE)]
 	= sizeof(NifExportTrace);
+    fix_type_sizes[ERTS_ALC_FIX_TYPE_IX(ERTS_ALC_T_MREF_NSCHED_ENT)]
+	= sizeof(ErtsNSchedMagicRefTableEntry);
+    fix_type_sizes[ERTS_ALC_FIX_TYPE_IX(ERTS_ALC_T_MINDIRECTION)]
+	= ERTS_MAGIC_BIN_UNALIGNED_SIZE(sizeof(ErtsMagicIndirectionWord));
 
 #ifdef HARD_DEBUG
     hdbg_init();
@@ -3141,9 +3145,10 @@ reply_alloc_info(void *vair)
     while (1) {
 
 	if (hpp)
-	    ref_copy = STORE_NC(hpp, ohp, air->ref);
+	    ref_copy = erts_iref_storage_make_ref(&air->iref,
+                                                  hpp, ohp, 0);
 	else
-	    *szp += REF_THING_SIZE;
+	    *szp += erts_iref_storage_heap_size(&air->iref);
 
 	ai_list = NIL;
 	for (i = 0; air->allocs[i] != ERTS_ALC_A_INVALID; i++);
@@ -3372,8 +3377,10 @@ reply_alloc_info(void *vair)
     erts_smp_proc_unlock(rp, rp_locks);
     erts_proc_dec_refc(rp);
 
-    if (erts_smp_atomic32_dec_read_nob(&air->refc) == 0)
+    if (erts_smp_atomic32_dec_read_nob(&air->refc) == 0) {
+        erts_iref_storage_clean(&air->iref);
 	aireq_free(air);
+    }
 }
 
 int
@@ -3386,7 +3393,6 @@ erts_request_alloc_info(struct process *c_p,
     ErtsAllocInfoReq *air = aireq_alloc();
     Eterm req_ai[ERTS_ALC_INFO_A_END] = {0};
     Eterm alist;
-    Eterm *hp;
     int airix = 0, ai;
 
     air->req_sched = erts_get_scheduler_id();
@@ -3400,8 +3406,7 @@ erts_request_alloc_info(struct process *c_p,
     if (is_not_internal_ref(ref))
 	return 0;
 
-    hp = &air->ref_heap[0];
-    air->ref = STORE_NC(&hp, NULL, ref);
+    erts_iref_storage_save(&air->iref, ref);
 
     if (is_not_list(allocs))
 	return 0;
diff --git a/erts/emulator/beam/erl_alloc.types b/erts/emulator/beam/erl_alloc.types
index 1f7e4400e1..3d5de72ee7 100644
--- a/erts/emulator/beam/erl_alloc.types
+++ b/erts/emulator/beam/erl_alloc.types
@@ -65,7 +65,7 @@
 
 # --- Allocator declarations -------------------------------------------------
 #
-# If, and only if, the same thread performes *all* allocations,
+# If, and only if, the same thread performs *all* allocations,
 # reallocations and deallocations of all memory types that are handled
 # by a specific allocator (<ALLOCATOR> in type declaration), set 
 # <MULTI_THREAD> for this specific allocator to false; otherwise, set
@@ -280,6 +280,11 @@ type    TRACE_MSG_QUEUE SHORT_LIVED     SYSTEM          trace_message_queue
 type    SCHED_ASYNC_JOB SHORT_LIVED     SYSTEM          async_calls
 type    DIRTY_START     STANDARD	PROCESSES       dirty_start
 type    DIRTY_SL	SHORT_LIVED	SYSTEM		dirty_short_lived
+type    MREF_NSCHED_ENT	FIXED_SIZE	SYSTEM		nsched_magic_ref_entry
+type    MREF_ENT	STANDARD	SYSTEM		magic_ref_entry
+type	MREF_TAB_BKTS	STANDARD	SYSTEM		magic_ref_table_buckets
+type	MREF_TAB	LONG_LIVED	SYSTEM		magic_ref_table
+type	MINDIRECTION	FIXED_SIZE	SYSTEM		magic_indirection
 
 +if threads_no_smp
 # Need thread safe allocs, but std_alloc and fix_alloc are not;
diff --git a/erts/emulator/beam/erl_alloc_util.c b/erts/emulator/beam/erl_alloc_util.c
index d346f8c8b6..a6bac06478 100644
--- a/erts/emulator/beam/erl_alloc_util.c
+++ b/erts/emulator/beam/erl_alloc_util.c
@@ -5186,7 +5186,7 @@ erts_alcu_sz_info(Allctr_t *allctr,
 
     ERTS_ALCU_DBG_CHK_THR_ACCESS(allctr);
 
-    /* Update sbc values not continously updated */
+    /* Update sbc values not continuously updated */
     allctr->sbcs.blocks.curr.no
 	= allctr->sbcs.curr.norm.mseg.no + allctr->sbcs.curr.norm.sys_alloc.no;
     allctr->sbcs.blocks.max.no = allctr->sbcs.max.no;
@@ -5272,7 +5272,7 @@ erts_alcu_info(Allctr_t *allctr,
 
     ERTS_ALCU_DBG_CHK_THR_ACCESS(allctr);
 
-    /* Update sbc values not continously updated */
+    /* Update sbc values not continuously updated */
     allctr->sbcs.blocks.curr.no
 	= allctr->sbcs.curr.norm.mseg.no + allctr->sbcs.curr.norm.sys_alloc.no;
     allctr->sbcs.blocks.max.no = allctr->sbcs.max.no;
diff --git a/erts/emulator/beam/erl_bif_binary.c b/erts/emulator/beam/erl_bif_binary.c
index 24debedce2..62a752d854 100644
--- a/erts/emulator/beam/erl_bif_binary.c
+++ b/erts/emulator/beam/erl_bif_binary.c
@@ -380,7 +380,7 @@ static void ac_compute_failure_functions(ACTrie *act, ACNode **qbuff)
 		qbuff[qt++] = child;
 		/* Search for correct failure function, follow the parent's
 		   failure function until you find a similar transition
-		   funtion to this child's */
+		   function to this child's */
 		r =  parent->h;
 		while (r != NULL && r->g[i] == NULL) {
 		    r = r->h;
@@ -1010,8 +1010,8 @@ BIF_RETTYPE binary_compile_pattern_1(BIF_ALIST_1)
     if (do_binary_match_compile(BIF_ARG_1,&tag,&bin)) {
 	BIF_ERROR(BIF_P,BADARG);
     }
-    hp = HAlloc(BIF_P, PROC_BIN_SIZE+3);
-    ret = erts_mk_magic_binary_term(&hp, &MSO(BIF_P), bin);
+    hp = HAlloc(BIF_P, ERTS_MAGIC_REF_THING_SIZE+3);
+    ret = erts_mk_magic_ref(&hp, &MSO(BIF_P), bin);
     ret = TUPLE2(hp, tag, ret);
     BIF_RET(ret);
 }
@@ -1426,11 +1426,11 @@ binary_match(Process *p, Eterm arg1, Eterm arg2, Eterm arg3, Uint flags)
 	    goto badarg;
 	}
 	if (((tp[1] != am_bm) && (tp[1] != am_ac)) ||
-	    !ERTS_TERM_IS_MAGIC_BINARY(tp[2])) {
+	    !is_internal_magic_ref(tp[2])) {
 	    goto badarg;
 	}
 	bfs.type = tp[1];
-	bin = ((ProcBin *) binary_val(tp[2]))->val;
+	bin = erts_magic_ref2bin(tp[2]);
 	if (bfs.type == am_bm &&
 	    ERTS_MAGIC_BIN_DESTRUCTOR(bin) != cleanup_my_data_bm) {
 	    goto badarg;
@@ -1448,8 +1448,8 @@ binary_match(Process *p, Eterm arg1, Eterm arg2, Eterm arg3, Uint flags)
     bfs.global_result = &do_match_global_result;
     runres = do_binary_find(p, arg1, &bfs, bin, NIL, &result);
     if (runres == DO_BIN_MATCH_RESTART && bin_term == NIL) {
-	Eterm *hp = HAlloc(p, PROC_BIN_SIZE);
-	bin_term = erts_mk_magic_binary_term(&hp, &MSO(p), bin);
+	Eterm *hp = HAlloc(p, ERTS_MAGIC_REF_THING_SIZE);
+	bin_term = erts_mk_magic_ref(&hp, &MSO(p), bin);
     } else if (bin_term == NIL) {
 	erts_bin_free(bin);
     }
@@ -1512,11 +1512,11 @@ binary_split(Process *p, Eterm arg1, Eterm arg2, Eterm arg3)
 	    goto badarg;
 	}
 	if (((tp[1] != am_bm) && (tp[1] != am_ac)) ||
-	    !ERTS_TERM_IS_MAGIC_BINARY(tp[2])) {
+	    !is_internal_magic_ref(tp[2])) {
 	    goto badarg;
 	}
 	bfs.type = tp[1];
-	bin = ((ProcBin *) binary_val(tp[2]))->val;
+	bin = erts_magic_ref2bin(tp[2]);
 	if (bfs.type == am_bm &&
 	    ERTS_MAGIC_BIN_DESTRUCTOR(bin) != cleanup_my_data_bm) {
 	    goto badarg;
@@ -1534,8 +1534,8 @@ binary_split(Process *p, Eterm arg1, Eterm arg2, Eterm arg3)
     bfs.global_result = &do_split_global_result;
     runres = do_binary_find(p, arg1, &bfs, bin, NIL, &result);
     if (runres == DO_BIN_MATCH_RESTART && bin_term == NIL) {
-	Eterm *hp = HAlloc(p, PROC_BIN_SIZE);
-	bin_term = erts_mk_magic_binary_term(&hp, &MSO(p), bin);
+	Eterm *hp = HAlloc(p, ERTS_MAGIC_REF_THING_SIZE);
+	bin_term = erts_mk_magic_ref(&hp, &MSO(p), bin);
     } else if (bin_term == NIL) {
 	erts_bin_free(bin);
     }
@@ -1767,7 +1767,8 @@ static BIF_RETTYPE binary_find_trap(BIF_ALIST_3)
 {
     int runres;
     Eterm result;
-    Binary *bin = ((ProcBin *) binary_val(BIF_ARG_3))->val;
+    Binary *bin = erts_magic_ref2bin(BIF_ARG_3);
+
     runres = do_binary_find(BIF_P, BIF_ARG_1, NULL, bin, BIF_ARG_2, &result);
     if (runres == DO_BIN_MATCH_OK) {
 	BIF_RET(result);
@@ -2186,8 +2187,8 @@ static BIF_RETTYPE do_longest_common(Process *p, Eterm list, int direction)
 		cd[i].type = CL_TYPE_HEAP;
 	    }
 	}
-	hp = HAlloc(p, PROC_BIN_SIZE);
-	bin_term = erts_mk_magic_binary_term(&hp, &MSO(p), mb);
+	hp = HAlloc(p, ERTS_MAGIC_REF_THING_SIZE);
+	bin_term = erts_mk_magic_ref(&hp, &MSO(p), mb);
 	BUMP_ALL_REDS(p);
 	BIF_TRAP3(trapper, p, bin_term, epos,list);
     }
@@ -2214,8 +2215,7 @@ static BIF_RETTYPE do_longest_common_trap(Process *p, Eterm bin_term, Eterm curr
 #else
     term_to_Uint(current_pos, &pos);
 #endif
-    ASSERT(ERTS_TERM_IS_MAGIC_BINARY(bin_term));
-    bin = ((ProcBin *) binary_val(bin_term))->val;
+    bin = erts_magic_ref2bin(bin_term);
     cd = (CommonData *) ERTS_MAGIC_BIN_DATA(bin);
     if (direction == DIRECTION_PREFIX) {
 	trapper = &binary_longest_prefix_trap_export;
@@ -2681,8 +2681,8 @@ static BIF_RETTYPE do_binary_copy(Process *p, Eterm bin, Eterm en)
 	cbs->source_size = size;
 	cbs->result_pos = pos;
 	cbs->times_left = n-i;
-	hp = HAlloc(p,PROC_BIN_SIZE);
-	trap_term = erts_mk_magic_binary_term(&hp, &MSO(p), mb);
+	hp = HAlloc(p, ERTS_MAGIC_REF_THING_SIZE);
+	trap_term = erts_mk_magic_ref(&hp, &MSO(p), mb);
 	BUMP_ALL_REDS(p);
 	BIF_TRAP2(&binary_copy_trap_export, p, bin, trap_term);
     } else {
@@ -2717,7 +2717,7 @@ BIF_RETTYPE binary_copy_trap(BIF_ALIST_2)
     Uint reds = get_reds(BIF_P, BINARY_COPY_LOOP_FACTOR);
     byte *t;
     Uint pos;
-    Binary *mb = ((ProcBin *) binary_val(BIF_ARG_2))->val;
+    Binary *mb = erts_magic_ref2bin(BIF_ARG_2);
     CopyBinState *cbs = (CopyBinState *) ERTS_MAGIC_BIN_DATA(mb);
     Uint opos;
 
diff --git a/erts/emulator/beam/erl_bif_ddll.c b/erts/emulator/beam/erl_bif_ddll.c
index 587859e413..2ce872d2bc 100644
--- a/erts/emulator/beam/erl_bif_ddll.c
+++ b/erts/emulator/beam/erl_bif_ddll.c
@@ -1109,25 +1109,25 @@ void erts_ddll_decrement_port_count(DE_Handle *dh)
 static void first_ddll_reference(DE_Handle *dh) 
 {
     assert_drv_list_rwlocked();
-    erts_refc_init(&(dh->refc),1);
+    erts_smp_refc_init(&(dh->refc),1);
 }
 
 void erts_ddll_reference_driver(DE_Handle *dh)
 {
     assert_drv_list_locked();
-    if (erts_refc_inctest(&(dh->refc),1) == 1) {
-	erts_refc_inc(&(dh->refc),2); /* add a reference for the scheduled operation */
+    if (erts_smp_refc_inctest(&(dh->refc),1) == 1) {
+	erts_smp_refc_inc(&(dh->refc),2); /* add a reference for the scheduled operation */
     }
 }
 
 void erts_ddll_reference_referenced_driver(DE_Handle *dh)
 {
-    erts_refc_inc(&(dh->refc),2);
+    erts_smp_refc_inc(&(dh->refc),2);
 }
 
 void erts_ddll_dereference_driver(DE_Handle *dh)
 {
-    if (erts_refc_dectest(&(dh->refc),0) == 0) {
+    if (erts_smp_refc_dectest(&(dh->refc),0) == 0) {
 	/* No lock here, but if the driver is referenced again,
 	   the scheduled deletion is added as a reference too, see above */
 	erts_schedule_misc_op(ddll_no_more_references, (void *) dh);
@@ -1150,11 +1150,11 @@ static void restore_process_references(DE_Handle *dh)
 {
     DE_ProcEntry *p;
     assert_drv_list_rwlocked();
-    ASSERT(erts_refc_read(&(dh->refc),0) == 0);
+    ASSERT(erts_smp_refc_read(&(dh->refc),0) == 0);
     for(p  = dh->procs;p != NULL; p = p->next) {
 	if (p->awaiting_status == ERL_DE_PROC_LOADED) {
 	    ASSERT(p->flags & ERL_DE_FL_DEREFERENCED);
-	    erts_refc_inc(&(dh->refc),1);
+	    erts_smp_refc_inc(&(dh->refc),1);
 	    p->flags &= ~ERL_DE_FL_DEREFERENCED;
 	}
     }
@@ -1176,9 +1176,9 @@ static void ddll_no_more_references(void *vdh)
 
     lock_drv_list();
 
-    x = erts_refc_read(&(dh->refc),0);
+    x = erts_smp_refc_read(&(dh->refc),0);
     if (x > 0) {
-	x = erts_refc_dectest(&(dh->refc),0); /* delete the reference added for me */
+	x = erts_smp_refc_dectest(&(dh->refc),0); /* delete the reference added for me */
     }
 
 
@@ -1476,8 +1476,10 @@ static void add_proc_loaded_deref(DE_Handle *dh, Process *proc)
 
 static Eterm copy_ref(Eterm ref, Eterm *hp)
 {
-    RefThing *ptr = ref_thing_ptr(ref);
-    memcpy(hp, ptr, sizeof(RefThing));
+    ErtsORefThing *ptr;
+    ASSERT(is_internal_ordinary_ref(ref));
+    ptr = ordinary_ref_thing_ptr(ref);
+    memcpy(hp, ptr, sizeof(ErtsORefThing));
     return (make_internal_ref(hp));
 }
 
@@ -1643,7 +1645,7 @@ static int load_driver_entry(DE_Handle **dhp, char *path, char *name)
     dh->handle = NULL;
     dh->procs = NULL;
     erts_smp_atomic32_init_nob(&dh->port_count, 0);
-    erts_refc_init(&(dh->refc), (erts_aint_t) 0);
+    erts_smp_refc_init(&(dh->refc), (erts_aint_t) 0);
     dh->status = -1;
     dh->reload_full_path = NULL;
     dh->reload_driver_name = NULL;
@@ -1681,7 +1683,7 @@ static int reload_driver_entry(DE_Handle *dh)
     dh->reload_full_path = NULL;
     dh->reload_driver_name = NULL;
 
-    ASSERT(erts_refc_read(&(dh->refc),0) == 0);
+    ASSERT(erts_smp_refc_read(&(dh->refc),0) == 0);
     ASSERT(dh->full_path != NULL);
     erts_free(ERTS_ALC_T_DDLL_HANDLE, (void *) dh->full_path);
     dh->full_path = NULL;
@@ -1720,10 +1722,10 @@ static void notify_proc(Process *proc, Eterm ref, Eterm driver_name, Eterm type,
 	Eterm e;
 	mp = erts_alloc_message_heap(proc, &rp_locks,
 				     (6 /* tuple */ + 3 /* Error tuple */ + 
-				      REF_THING_SIZE + need),
+				      ERTS_REF_THING_SIZE + need),
 				     &hp, &ohp);
 	r = copy_ref(ref,hp);
-	hp += REF_THING_SIZE;
+	hp += ERTS_REF_THING_SIZE;
 	e = build_load_error_hp(hp, errcode);
 	hp += need;
 	mess = TUPLE2(hp,tag,e);
@@ -1731,10 +1733,10 @@ static void notify_proc(Process *proc, Eterm ref, Eterm driver_name, Eterm type,
 	mess = TUPLE5(hp,type,r,am_driver,driver_name,mess);
     } else {	
 	mp = erts_alloc_message_heap(proc, &rp_locks,
-				     6 /* tuple */ + REF_THING_SIZE,
+				     6 /* tuple */ + ERTS_REF_THING_SIZE,
 				     &hp, &ohp);
 	r = copy_ref(ref,hp);
-	hp += REF_THING_SIZE;
+	hp += ERTS_REF_THING_SIZE;
 	mess = TUPLE5(hp,type,r,am_driver,driver_name,tag);
     }
     erts_queue_message(proc, rp_locks, mp, mess, am_system);
diff --git a/erts/emulator/beam/erl_bif_guard.c b/erts/emulator/beam/erl_bif_guard.c
index 458315f293..cdac4c1d11 100644
--- a/erts/emulator/beam/erl_bif_guard.c
+++ b/erts/emulator/beam/erl_bif_guard.c
@@ -190,7 +190,7 @@ BIF_RETTYPE round_1(BIF_ALIST_1)
     GET_DOUBLE(BIF_ARG_1, f);
 
     /* round it and return the resultant integer */
-    res = double_to_integer(BIF_P, (f.fd > 0.0) ? f.fd + 0.5 : f.fd - 0.5);
+    res = double_to_integer(BIF_P, round(f.fd));
     BIF_RET(res);
 }
 
@@ -630,8 +630,7 @@ Eterm erts_gc_round_1(Process* p, Eterm* reg, Uint live)
     }
     GET_DOUBLE(arg, f);
 
-    return gc_double_to_integer(p, (f.fd > 0.0) ? f.fd + 0.5 : f.fd - 0.5,
-				reg, live);
+    return gc_double_to_integer(p, round(f.fd), reg, live);
 }
 
 Eterm erts_gc_trunc_1(Process* p, Eterm* reg, Uint live)
diff --git a/erts/emulator/beam/erl_bif_info.c b/erts/emulator/beam/erl_bif_info.c
index 9f32e0e7bd..06a73ffea5 100644
--- a/erts/emulator/beam/erl_bif_info.c
+++ b/erts/emulator/beam/erl_bif_info.c
@@ -177,7 +177,33 @@ bld_bin_list(Uint **hpp, Uint *szp, ErlOffHeap* oh)
 	    if (szp)
 		*szp += 4+2;
 	    if (hpp) {
-		Uint refc = (Uint) erts_smp_atomic_read_nob(&pb->val->refc);
+		Uint refc = (Uint) erts_refc_read(&pb->val->refc, 1);
+		tuple = TUPLE3(*hpp, val, orig_size, make_small(refc));
+		res = CONS(*hpp + 4, tuple, res);
+		*hpp += 4+2;
+	    }
+	}
+    }
+    return res;
+}
+
+static Eterm
+bld_magic_ref_bin_list(Uint **hpp, Uint *szp, ErlOffHeap* oh)
+{
+    struct erl_off_heap_header* ohh;
+    Eterm res = NIL;
+    Eterm tuple;
+
+    for (ohh = oh->first; ohh; ohh = ohh->next) {
+	if (is_ref_thing_header((*((Eterm *) ohh)))) {
+            ErtsMRefThing *mrtp = (ErtsMRefThing *) ohh;
+	    Eterm val = erts_bld_uword(hpp, szp, (UWord) mrtp->mb);
+	    Eterm orig_size = erts_bld_uint(hpp, szp, mrtp->mb->orig_size);
+    
+	    if (szp)
+		*szp += 4+2;
+	    if (hpp) {
+		Uint refc = (Uint) erts_refc_read(&mrtp->mb->refc, 1);
 		tuple = TUPLE3(*hpp, val, orig_size, make_small(refc));
 		res = CONS(*hpp + 4, tuple, res);
 		*hpp += 4+2;
@@ -627,7 +653,8 @@ static Eterm pi_args[] = {
     am_current_location,
     am_current_stacktrace,
     am_message_queue_data,
-    am_garbage_collection_info
+    am_garbage_collection_info,
+    am_magic_ref
 };
 
 #define ERTS_PI_ARGS ((int) (sizeof(pi_args)/sizeof(Eterm)))
@@ -678,6 +705,7 @@ pi_arg2ix(Eterm arg)
     case am_current_stacktrace:			return 31;
     case am_message_queue_data:			return 32;
     case am_garbage_collection_info:		return 33;
+    case am_magic_ref:                          return 34;
     default:					return -1;
     }
 }
@@ -1618,6 +1646,14 @@ process_info_aux(Process *BIF_P,
 	hp = HAlloc(BIF_P, 3);
 	break;
 
+    case am_magic_ref: {
+	Uint sz = 3;
+	(void) bld_magic_ref_bin_list(NULL, &sz, &MSO(rp));
+	hp = HAlloc(BIF_P, sz);
+	res = bld_magic_ref_bin_list(&hp, NULL, &MSO(rp));
+	break;
+    }
+
     default:
 	return THE_NON_VALUE; /* will produce badarg */
 
@@ -3551,6 +3587,11 @@ BIF_RETTYPE error_logger_warning_map_0(BIF_ALIST_0)
 
 static erts_smp_atomic_t available_internal_state;
 
+static int empty_magic_ref_destructor(Binary *bin)
+{
+    return 1;
+}
+
 BIF_RETTYPE erts_debug_get_internal_state_1(BIF_ALIST_1)
 {
     /*
@@ -3916,6 +3957,34 @@ BIF_RETTYPE erts_debug_get_internal_state_1(BIF_ALIST_1)
 		}
 		return make_atom(ix);
 	    }
+	    else if (ERTS_IS_ATOM_STR("magic_ref", tp[1])) {
+                Binary *bin;
+                UWord bin_addr, refc;
+                Eterm bin_addr_term, refc_term, test_type;
+                Uint sz;
+                Eterm *hp;
+                if (!is_internal_magic_ref(tp[2])) {
+                    if (is_internal_ordinary_ref(tp[2])) {
+                        ErtsORefThing *rtp;
+                        rtp = (ErtsORefThing *) internal_ref_val(tp[2]);
+                        if (erts_is_ref_numbers_magic(rtp->num))
+                            BIF_RET(am_true);
+                    }
+                    BIF_RET(am_false);
+                }
+                bin = erts_magic_ref2bin(tp[2]);
+                refc = erts_refc_read(&bin->refc, 1);
+                bin_addr = (UWord) bin;
+                sz = 4;
+                erts_bld_uword(NULL, &sz, bin_addr);
+                erts_bld_uword(NULL, &sz, refc);
+                hp = HAlloc(BIF_P, sz);
+                bin_addr_term = erts_bld_uword(&hp, NULL, bin_addr);
+                refc_term = erts_bld_uword(&hp, NULL, refc);
+                test_type = (ERTS_MAGIC_BIN_DESTRUCTOR(bin) == empty_magic_ref_destructor
+                             ? am_true : am_false);
+                BIF_RET(TUPLE3(hp, bin_addr_term, refc_term, test_type));
+	    }
 
 	    break;
 	}
@@ -4004,7 +4073,6 @@ static void broken_halt_test(Eterm bif_arg_2)
     erts_exit(ERTS_DUMP_EXIT, "%T", bif_arg_2);
 }
 
-
 BIF_RETTYPE erts_debug_set_internal_state_2(BIF_ALIST_2)
 {
     /*
@@ -4334,6 +4402,21 @@ BIF_RETTYPE erts_debug_set_internal_state_2(BIF_ALIST_2)
             /* Used by ets_SUITE (stdlib) */
             BIF_RET(erts_ets_restore_meta_state(BIF_P, BIF_ARG_2));
         }
+        else if (ERTS_IS_ATOM_STR("make", BIF_ARG_1)) {
+            if (ERTS_IS_ATOM_STR("magic_ref", BIF_ARG_2)) {
+                Binary *bin = erts_create_magic_binary(0, empty_magic_ref_destructor);
+                UWord bin_addr = (UWord) bin;
+                Eterm bin_addr_term, magic_ref, res;
+                Eterm *hp;
+                Uint sz = ERTS_MAGIC_REF_THING_SIZE + 3;
+                erts_bld_uword(NULL, &sz, bin_addr);
+                hp = HAlloc(BIF_P, sz);
+                bin_addr_term = erts_bld_uword(&hp, NULL, bin_addr);
+                magic_ref = erts_mk_magic_ref(&hp, &BIF_P->off_heap, bin);
+                res = TUPLE2(hp, magic_ref, bin_addr_term);
+                BIF_RET(res);
+            }
+        }
     }
 
     BIF_ERROR(BIF_P, BADARG);
diff --git a/erts/emulator/beam/erl_bif_os.c b/erts/emulator/beam/erl_bif_os.c
index 46777d3aa5..edc3c82b23 100644
--- a/erts/emulator/beam/erl_bif_os.c
+++ b/erts/emulator/beam/erl_bif_os.c
@@ -203,3 +203,17 @@ BIF_RETTYPE os_unsetenv_1(BIF_ALIST_1)
     }
     BIF_RET(am_true);
 }
+
+BIF_RETTYPE os_set_signal_2(BIF_ALIST_2) {
+    if (is_atom(BIF_ARG_1) && ((BIF_ARG_2 == am_ignore) ||
+                               (BIF_ARG_2 == am_default) ||
+                               (BIF_ARG_2 == am_handle))) {
+        if (!erts_set_signal(BIF_ARG_1, BIF_ARG_2))
+            goto error;
+
+        BIF_RET(am_ok);
+    }
+
+error:
+    BIF_ERROR(BIF_P, BADARG);
+}
diff --git a/erts/emulator/beam/erl_bif_port.c b/erts/emulator/beam/erl_bif_port.c
index 90e78a9b0b..09d2c5376b 100644
--- a/erts/emulator/beam/erl_bif_port.c
+++ b/erts/emulator/beam/erl_bif_port.c
@@ -214,7 +214,7 @@ BIF_RETTYPE erts_internal_port_command_3(BIF_ALIST_3)
 	ASSERT(!(flags & ERTS_PORT_SIG_FLG_FORCE));
 	/* Fall through... */
     case ERTS_PORT_OP_SCHEDULED:
-	ASSERT(is_internal_ref(ref));
+	ASSERT(is_internal_ordinary_ref(ref));
 	ERTS_BIF_PREP_RET(res, ref);
 	break;
     case ERTS_PORT_OP_DONE:
@@ -260,7 +260,7 @@ BIF_RETTYPE erts_internal_port_call_3(BIF_ALIST_3)
 	retval = am_badarg;
 	break;
     case ERTS_PORT_OP_SCHEDULED:
-	ASSERT(is_internal_ref(retval));
+	ASSERT(is_internal_ordinary_ref(retval));
 	break;
     case ERTS_PORT_OP_DONE:
 	ASSERT(is_not_internal_ref(retval));
@@ -310,7 +310,7 @@ BIF_RETTYPE erts_internal_port_control_3(BIF_ALIST_3)
 	retval = am_badarg;
 	break;
     case ERTS_PORT_OP_SCHEDULED:
-	ASSERT(is_internal_ref(retval));
+	ASSERT(is_internal_ordinary_ref(retval));
 	break;
     case ERTS_PORT_OP_DONE:
 	ASSERT(is_not_internal_ref(retval));
@@ -356,7 +356,7 @@ BIF_RETTYPE erts_internal_port_close_1(BIF_ALIST_1)
     case ERTS_PORT_OP_DROPPED:
 	BIF_RET(am_badarg);
     case ERTS_PORT_OP_SCHEDULED:
-	ASSERT(is_internal_ref(ref));
+	ASSERT(is_internal_ordinary_ref(ref));
 	BIF_RET(ref);
     case ERTS_PORT_OP_DONE:
 	BIF_RET(am_true);
@@ -389,7 +389,7 @@ BIF_RETTYPE erts_internal_port_connect_2(BIF_ALIST_2)
     case ERTS_PORT_OP_DROPPED:
 	BIF_RET(am_badarg);
     case ERTS_PORT_OP_SCHEDULED:
-	ASSERT(is_internal_ref(ref));
+	ASSERT(is_internal_ordinary_ref(ref));
 	BIF_RET(ref);
 	break;
     case ERTS_PORT_OP_DONE:
@@ -428,7 +428,7 @@ BIF_RETTYPE erts_internal_port_info_1(BIF_ALIST_1)
     case ERTS_PORT_OP_DROPPED:
 	BIF_RET(am_undefined);
     case ERTS_PORT_OP_SCHEDULED:
-	ASSERT(is_internal_ref(retval));
+	ASSERT(is_internal_ordinary_ref(retval));
 	BIF_RET(retval);
     case ERTS_PORT_OP_DONE:
 	ASSERT(is_not_internal_ref(retval));
@@ -467,7 +467,7 @@ BIF_RETTYPE erts_internal_port_info_2(BIF_ALIST_2)
     case ERTS_PORT_OP_DROPPED:
 	BIF_RET(am_undefined);
     case ERTS_PORT_OP_SCHEDULED:
-	ASSERT(is_internal_ref(retval));
+	ASSERT(is_internal_ordinary_ref(retval));
 	BIF_RET(retval);
     case ERTS_PORT_OP_DONE:
 	ASSERT(is_not_internal_ref(retval));
diff --git a/erts/emulator/beam/erl_bif_re.c b/erts/emulator/beam/erl_bif_re.c
index 391fee5753..a66b05c6ff 100644
--- a/erts/emulator/beam/erl_bif_re.c
+++ b/erts/emulator/beam/erl_bif_re.c
@@ -1320,17 +1320,17 @@ handle_iolist:
 	Binary *mbp = erts_create_magic_binary(sizeof(RestartContext),
 					       cleanup_restart_context_bin);
 	RestartContext *restartp = ERTS_MAGIC_BIN_DATA(mbp);
-	Eterm magic_bin;
+	Eterm magic_ref;
 	Eterm *hp;
 	memcpy(restartp,&restart,sizeof(RestartContext));
 	BUMP_ALL_REDS(p);
-	hp = HAlloc(p, PROC_BIN_SIZE);
-	magic_bin = erts_mk_magic_binary_term(&hp, &MSO(p), mbp);
+	hp = HAlloc(p, ERTS_MAGIC_REF_THING_SIZE);
+	magic_ref = erts_mk_magic_ref(&hp, &MSO(p), mbp);
 	BIF_TRAP3(&re_exec_trap_export, 
 		  p,
 		  arg1,
 		  arg2 /* To avoid GC of precompiled code, XXX: not utilized yet */,
-		  magic_bin);
+		  magic_ref);
     }
 
     res = build_exec_return(p, rc, &restart, arg1);
@@ -1367,9 +1367,7 @@ static BIF_RETTYPE re_exec_trap(BIF_ALIST_3)
     Uint loop_limit_tmp;
     Eterm res;
 
-    ASSERT(ERTS_TERM_IS_MAGIC_BINARY(BIF_ARG_3));
-
-    mbp = ((ProcBin *) binary_val(BIF_ARG_3))->val;
+    mbp = erts_magic_ref2bin(BIF_ARG_3);
 
     ASSERT(ERTS_MAGIC_BIN_DESTRUCTOR(mbp)
 	   == cleanup_restart_context_bin);
diff --git a/erts/emulator/beam/erl_bif_trace.c b/erts/emulator/beam/erl_bif_trace.c
index a480754bdb..f471390501 100644
--- a/erts/emulator/beam/erl_bif_trace.c
+++ b/erts/emulator/beam/erl_bif_trace.c
@@ -2363,7 +2363,7 @@ BIF_RETTYPE system_profile_2(BIF_ALIST_2)
 typedef struct {
     Process *proc;
     Eterm ref;
-    Eterm ref_heap[REF_THING_SIZE];
+    Eterm ref_heap[ERTS_REF_THING_SIZE];
     Eterm target;
     erts_smp_atomic32_t refc;
 } ErtsTraceDeliveredAll;
diff --git a/erts/emulator/beam/erl_bif_unique.c b/erts/emulator/beam/erl_bif_unique.c
index 7c70217d8d..3fd4c87094 100644
--- a/erts/emulator/beam/erl_bif_unique.c
+++ b/erts/emulator/beam/erl_bif_unique.c
@@ -22,12 +22,15 @@
 #  include "config.h"
 #endif
 
+#define ERL_BIF_UNIQUE_C__
 #include "sys.h"
 #include "erl_vm.h"
 #include "erl_alloc.h"
 #include "export.h"
 #include "bif.h"
 #include "erl_bif_unique.h"
+#include "hash.h"
+#include "erl_binary.h"
 
 /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *\
  * Reference                                                         *
@@ -58,9 +61,20 @@ static union {
 static Uint32 max_thr_id;
 #endif
 
+static void init_magic_ref_tables(void);
+
+static Uint64 ref_init_value;
+
 static void
 init_reference(void)
 {
+    SysTimeval tv;
+    sys_gettimeofday(&tv);
+    ref_init_value = 0;
+    ref_init_value |= (Uint64) tv.tv_sec;
+    ref_init_value |= ((Uint64) tv.tv_usec) << 32;
+    ref_init_value *= (Uint64) 268438039;
+    ref_init_value += (Uint64) tv.tv_usec;
 #ifdef DEBUG
     max_thr_id = (Uint32) erts_no_schedulers;
 #ifdef ERTS_DIRTY_SCHEDULERS
@@ -68,11 +82,13 @@ init_reference(void)
     max_thr_id += (Uint32) erts_no_dirty_io_schedulers;
 #endif
 #endif
-    erts_atomic64_init_nob(&global_reference.count, 0);
+    erts_atomic64_init_nob(&global_reference.count,
+			   (erts_aint64_t) ref_init_value);
+    init_magic_ref_tables();
 }
 
 static ERTS_INLINE void
-global_make_ref_in_array(Uint32 thr_id, Uint32 ref[ERTS_MAX_REF_NUMBERS])
+global_make_ref_in_array(Uint32 thr_id, Uint32 ref[ERTS_REF_NUMBERS])
 {
     Uint64 value;
 
@@ -82,7 +98,7 @@ global_make_ref_in_array(Uint32 thr_id, Uint32 ref[ERTS_MAX_REF_NUMBERS])
 }
 
 static ERTS_INLINE void
-make_ref_in_array(Uint32 ref[ERTS_MAX_REF_NUMBERS])
+make_ref_in_array(Uint32 ref[ERTS_REF_NUMBERS])
 {
     ErtsSchedulerData *esdp = erts_get_scheduler_data();
     if (esdp)
@@ -92,15 +108,23 @@ make_ref_in_array(Uint32 ref[ERTS_MAX_REF_NUMBERS])
 }
 
 void
-erts_make_ref_in_array(Uint32 ref[ERTS_MAX_REF_NUMBERS])
+erts_make_ref_in_array(Uint32 ref[ERTS_REF_NUMBERS])
+{
+    make_ref_in_array(ref);
+}
+
+void
+erts_make_magic_ref_in_array(Uint32 ref[ERTS_REF_NUMBERS])
 {
     make_ref_in_array(ref);
+    ASSERT(!(ref[1] & ERTS_REF1_MAGIC_MARKER_BIT__));
+    ref[1] |= ERTS_REF1_MAGIC_MARKER_BIT__;
 }
 
-Eterm erts_make_ref_in_buffer(Eterm buffer[REF_THING_SIZE])
+Eterm erts_make_ref_in_buffer(Eterm buffer[ERTS_REF_THING_SIZE])
 {
     Eterm* hp = buffer;
-    Uint32 ref[ERTS_MAX_REF_NUMBERS];
+    Uint32 ref[ERTS_REF_NUMBERS];
 
     make_ref_in_array(ref);
     write_ref_thing(hp, ref[0], ref[1], ref[2]);
@@ -110,11 +134,11 @@ Eterm erts_make_ref_in_buffer(Eterm buffer[REF_THING_SIZE])
 Eterm erts_make_ref(Process *c_p)
 {
     Eterm* hp;
-    Uint32 ref[ERTS_MAX_REF_NUMBERS];
+    Uint32 ref[ERTS_REF_NUMBERS];
 
     ERTS_SMP_LC_ASSERT(ERTS_PROC_LOCK_MAIN & erts_proc_lc_my_proc_locks(c_p));
 
-    hp = HAlloc(c_p, REF_THING_SIZE);
+    hp = HAlloc(c_p, ERTS_REF_THING_SIZE);
 
     make_ref_in_array(ref);
     write_ref_thing(hp, ref[0], ref[1], ref[2]);
@@ -122,6 +146,272 @@ Eterm erts_make_ref(Process *c_p)
     return make_internal_ref(hp);
 }
 
+/*
+ * Magic reference tables
+ */
+
+typedef struct {
+    HashBucket hash;
+    ErtsMagicBinary *mb;
+    Uint64 value;
+    Uint32 thr_id;
+} ErtsMagicRefTableEntry;
+
+typedef struct {
+    erts_rwmtx_t rwmtx;
+    Hash hash;
+    char name[32];
+} ErtsMagicRefTable;
+
+typedef struct {
+    union {
+	ErtsMagicRefTable table;
+	char align__[ERTS_ALC_CACHE_LINE_ALIGN_SIZE(sizeof(ErtsMagicRefTable))];
+    } u;
+} ErtsAlignedMagicRefTable;
+
+ErtsAlignedMagicRefTable *magic_ref_table;
+
+ErtsMagicBinary *
+erts_magic_ref_lookup_bin__(Uint32 refn[ERTS_REF_NUMBERS])
+{
+    ErtsMagicRefTableEntry tmpl;
+    ErtsMagicRefTableEntry *tep;
+    ErtsMagicBinary *mb;
+    ErtsMagicRefTable *tblp;
+
+    ASSERT(erts_is_ref_numbers_magic(refn));
+
+    tmpl.value = erts_get_ref_numbers_value(refn);
+    tmpl.thr_id = erts_get_ref_numbers_thr_id(refn);
+    if (tmpl.thr_id > erts_no_schedulers)
+	tblp = &magic_ref_table[0].u.table;
+    else
+	tblp = &magic_ref_table[tmpl.thr_id].u.table;
+
+    erts_rwmtx_rlock(&tblp->rwmtx);
+
+    tep = (ErtsMagicRefTableEntry *) hash_get(&tblp->hash, &tmpl);
+    if (!tep)
+	mb = NULL;
+    else {
+	erts_aint_t refc;
+	mb = tep->mb;
+        refc = erts_refc_inc_unless(&mb->refc, 0, 0);
+        if (refc == 0)
+            mb = NULL;
+    }
+
+    erts_rwmtx_runlock(&tblp->rwmtx);
+
+    return mb;
+}
+
+void
+erts_magic_ref_save_bin__(Eterm ref)
+{
+    ErtsMagicRefTableEntry tmpl;
+    ErtsMagicRefTableEntry *tep;
+    ErtsMRefThing *mrtp;
+    ErtsMagicRefTable *tblp;
+    Uint32 *refn;
+
+    ASSERT(is_internal_magic_ref(ref));
+
+    mrtp = (ErtsMRefThing *) internal_ref_val(ref);
+    refn = mrtp->mb->refn;
+
+    tmpl.value = erts_get_ref_numbers_value(refn);
+    tmpl.thr_id = erts_get_ref_numbers_thr_id(refn);
+
+    if (tmpl.thr_id > erts_no_schedulers)
+	tblp = &magic_ref_table[0].u.table;
+    else
+	tblp = &magic_ref_table[tmpl.thr_id].u.table;
+
+    erts_rwmtx_rlock(&tblp->rwmtx);
+
+    tep = (ErtsMagicRefTableEntry *) hash_get(&tblp->hash, &tmpl);
+
+    erts_rwmtx_runlock(&tblp->rwmtx);
+
+    if (!tep) {
+	ErtsMagicRefTableEntry *used_tep;
+	
+	ASSERT(tmpl.value == erts_get_ref_numbers_value(refn));
+	ASSERT(tmpl.thr_id == erts_get_ref_numbers_thr_id(refn));
+
+	if (tblp != &magic_ref_table[0].u.table) {
+	    tep = erts_alloc(ERTS_ALC_T_MREF_NSCHED_ENT,
+			     sizeof(ErtsNSchedMagicRefTableEntry));
+	}
+	else {
+	    tep = erts_alloc(ERTS_ALC_T_MREF_ENT,
+			     sizeof(ErtsMagicRefTableEntry));
+	    tep->thr_id = tmpl.thr_id;
+	}
+
+	tep->value = tmpl.value;
+	tep->mb = mrtp->mb;
+
+	erts_rwmtx_rwlock(&tblp->rwmtx);
+
+	used_tep = hash_put(&tblp->hash, tep);
+
+	erts_rwmtx_rwunlock(&tblp->rwmtx);
+
+	if (used_tep != tep) {
+	    if (tblp != &magic_ref_table[0].u.table)
+		erts_free(ERTS_ALC_T_MREF_NSCHED_ENT, (void *) tep);
+	    else
+		erts_free(ERTS_ALC_T_MREF_ENT, (void *) tep);
+	}
+    }
+}
+
+void
+erts_magic_ref_remove_bin(Uint32 refn[ERTS_REF_NUMBERS])
+{
+    ErtsMagicRefTableEntry tmpl;
+    ErtsMagicRefTableEntry *tep;
+    ErtsMagicRefTable *tblp;
+
+    tmpl.value = erts_get_ref_numbers_value(refn);
+    tmpl.thr_id = erts_get_ref_numbers_thr_id(refn);
+
+    if (tmpl.thr_id > erts_no_schedulers)
+	tblp = &magic_ref_table[0].u.table;
+    else
+	tblp = &magic_ref_table[tmpl.thr_id].u.table;
+
+    erts_rwmtx_rlock(&tblp->rwmtx);
+
+    tep = (ErtsMagicRefTableEntry *) hash_get(&tblp->hash, &tmpl);
+
+    erts_rwmtx_runlock(&tblp->rwmtx);
+
+    if (tep) {
+
+	ASSERT(tmpl.value == erts_get_ref_numbers_value(refn));
+	ASSERT(tmpl.thr_id == erts_get_ref_numbers_thr_id(refn));
+
+	erts_rwmtx_rwlock(&tblp->rwmtx);
+
+	tep = hash_remove(&tblp->hash, &tmpl);
+	ASSERT(tep);
+
+	erts_rwmtx_rwunlock(&tblp->rwmtx);
+
+	if (tblp != &magic_ref_table[0].u.table)
+	    erts_free(ERTS_ALC_T_MREF_NSCHED_ENT, (void *) tep);
+	else
+	    erts_free(ERTS_ALC_T_MREF_ENT, (void *) tep);
+    }
+}
+
+static int nsched_mreft_cmp(void *ve1, void *ve2)
+{
+    ErtsNSchedMagicRefTableEntry *e1 = ve1;
+    ErtsNSchedMagicRefTableEntry *e2 = ve2;
+    return e1->value != e2->value;
+}
+
+static int non_nsched_mreft_cmp(void *ve1, void *ve2)
+{
+    ErtsMagicRefTableEntry *e1 = ve1;
+    ErtsMagicRefTableEntry *e2 = ve2;
+    return e1->value != e2->value || e1->thr_id != e2->thr_id;
+}
+
+static HashValue nsched_mreft_hash(void *ve)
+{
+    ErtsNSchedMagicRefTableEntry *e = ve;
+    return (HashValue) e->value;
+}
+
+static HashValue non_nsched_mreft_hash(void *ve)
+{
+    ErtsMagicRefTableEntry *e = ve;
+    HashValue h;
+    h = (HashValue) e->thr_id;
+    h *= 268440163;
+    h += (HashValue) e->value;
+    return h;
+}
+
+static void *mreft_alloc(void *ve)
+{
+    /*
+     * We allocate the element before
+     * hash_put() and pass it as
+     * template which we get as
+     * input...
+     */
+    return ve;
+}
+
+static void mreft_free(void *ve)
+{
+    /*
+     * We free the element ourselves
+     * after hash_remove()...
+     */
+}
+
+static void *mreft_meta_alloc(int i, size_t size)
+{
+    return erts_alloc(ERTS_ALC_T_MREF_TAB_BKTS, size);
+}
+
+static void mreft_meta_free(int i, void *ptr)
+{
+    erts_free(ERTS_ALC_T_MREF_TAB_BKTS, ptr);
+}
+
+static void
+init_magic_ref_tables(void)
+{
+    HashFunctions hash_funcs;
+    int i;
+    ErtsMagicRefTable *tblp;
+
+    magic_ref_table = erts_alloc_permanent_cache_aligned(ERTS_ALC_T_MREF_TAB,
+							 (sizeof(ErtsAlignedMagicRefTable)
+							  * (erts_no_schedulers + 1)));
+
+    hash_funcs.hash = non_nsched_mreft_hash;
+    hash_funcs.cmp = non_nsched_mreft_cmp;
+
+    hash_funcs.alloc = mreft_alloc;
+    hash_funcs.free = mreft_free;
+    hash_funcs.meta_alloc = mreft_meta_alloc;
+    hash_funcs.meta_free = mreft_meta_free;
+    hash_funcs.meta_print = erts_print;
+
+    tblp = &magic_ref_table[0].u.table;
+    erts_snprintf(&tblp->name[0], sizeof(tblp->name),
+		  "magic_ref_table_0");
+    hash_init(0, &tblp->hash, &tblp->name[0], 1, hash_funcs);
+    erts_rwmtx_init(&tblp->rwmtx, "magic_ref_table");
+
+    hash_funcs.hash = nsched_mreft_hash;
+    hash_funcs.cmp = nsched_mreft_cmp;
+
+    for (i = 1; i <= erts_no_schedulers; i++) {
+	ErtsMagicRefTable *tblp = &magic_ref_table[i].u.table;
+	erts_snprintf(&tblp->name[0], sizeof(tblp->name),
+		      "magic_ref_table_%d", i);
+	hash_init(0, &tblp->hash, &tblp->name[0], 1, hash_funcs);
+	erts_rwmtx_init(&tblp->rwmtx, "magic_ref_table");
+    }
+}
+
+void erts_ref_bin_free(ErtsMagicBinary *mb)
+{
+    erts_bin_free((Binary *) mb);
+}
+
+
 /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *\
  * Unique Integer                                                    *
 \*                                                                   */
@@ -498,7 +788,7 @@ void
 erts_sched_bif_unique_init(ErtsSchedulerData *esdp)
 {
     esdp->unique = (Uint64) 0;
-    esdp->ref = (Uint64) 0;
+    esdp->ref = (Uint64) ref_init_value;
 }
 
 /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *\
@@ -513,7 +803,7 @@ BIF_RETTYPE make_ref_0(BIF_ALIST_0)
 
     ERTS_SMP_LC_ASSERT(ERTS_PROC_LOCK_MAIN & erts_proc_lc_my_proc_locks(BIF_P));
 
-    hp = HAlloc(BIF_P, REF_THING_SIZE);
+    hp = HAlloc(BIF_P, ERTS_REF_THING_SIZE);
 
     res = erts_sched_make_ref_in_buffer(erts_proc_sched_data(BIF_P), hp);
 
diff --git a/erts/emulator/beam/erl_bif_unique.h b/erts/emulator/beam/erl_bif_unique.h
index c6481864d0..27c2a15a5e 100644
--- a/erts/emulator/beam/erl_bif_unique.h
+++ b/erts/emulator/beam/erl_bif_unique.h
@@ -21,16 +21,25 @@
 #ifndef ERTS_BIF_UNIQUE_H__
 #define ERTS_BIF_UNIQUE_H__
 
+#include "erl_term.h"
 #include "erl_process.h"
 #include "big.h"
+#define ERTS_BINARY_TYPES_ONLY__
+#include "erl_binary.h"
+#undef ERTS_BINARY_TYPES_ONLY__
 
 void erts_bif_unique_init(void);
 void erts_sched_bif_unique_init(ErtsSchedulerData *esdp);
 
 /* reference */
 Eterm erts_make_ref(Process *);
-Eterm erts_make_ref_in_buffer(Eterm buffer[REF_THING_SIZE]);
-void erts_make_ref_in_array(Uint32 ref[ERTS_MAX_REF_NUMBERS]);
+Eterm erts_make_ref_in_buffer(Eterm buffer[ERTS_REF_THING_SIZE]);
+void erts_make_ref_in_array(Uint32 ref[ERTS_REF_NUMBERS]);
+void erts_make_magic_ref_in_array(Uint32 ref[ERTS_REF_NUMBERS]);
+void erts_magic_ref_remove_bin(Uint32 refn[ERTS_REF_NUMBERS]);
+void erts_magic_ref_save_bin__(Eterm ref);
+ErtsMagicBinary *erts_magic_ref_lookup_bin__(Uint32 refn[ERTS_REF_NUMBERS]);
+
 
 /* strict monotonic counter */
 
@@ -67,19 +76,35 @@ Eterm erts_debug_make_unique_integer(Process *c_p,
 				     Eterm etval1);
 
 
-ERTS_GLB_INLINE void erts_set_ref_numbers(Uint32 ref[ERTS_MAX_REF_NUMBERS],
+ERTS_GLB_INLINE void erts_set_ref_numbers(Uint32 ref[ERTS_REF_NUMBERS],
 					  Uint32 thr_id, Uint64 value);
-ERTS_GLB_INLINE Uint32 erts_get_ref_numbers_thr_id(Uint32 ref[ERTS_MAX_REF_NUMBERS]);
-ERTS_GLB_INLINE Uint64 erts_get_ref_numbers_value(Uint32 ref[ERTS_MAX_REF_NUMBERS]);
+ERTS_GLB_INLINE Uint32 erts_get_ref_numbers_thr_id(Uint32 ref[ERTS_REF_NUMBERS]);
+ERTS_GLB_INLINE int erts_is_ref_numbers_magic(Uint32 ref[ERTS_REF_NUMBERS]);
+ERTS_GLB_INLINE Uint64 erts_get_ref_numbers_value(Uint32 ref[ERTS_REF_NUMBERS]);
 ERTS_GLB_INLINE void erts_sched_make_ref_in_array(ErtsSchedulerData *esdp,
-						  Uint32 ref[ERTS_MAX_REF_NUMBERS]);
+						  Uint32 ref[ERTS_REF_NUMBERS]);
+ERTS_GLB_INLINE void erts_sched_make_magic_ref_in_array(ErtsSchedulerData *esdp,
+							Uint32 ref[ERTS_REF_NUMBERS]);
 ERTS_GLB_INLINE Eterm erts_sched_make_ref_in_buffer(ErtsSchedulerData *esdp,
-						    Eterm buffer[REF_THING_SIZE]);
+						    Eterm buffer[ERTS_REF_THING_SIZE]);
+ERTS_GLB_INLINE Eterm erts_mk_magic_ref(Eterm **hpp, ErlOffHeap *ohp, Binary *mbp);
+ERTS_GLB_INLINE Binary *erts_magic_ref2bin(Eterm mref);
+ERTS_GLB_INLINE void erts_magic_ref_save_bin(Eterm ref);
+ERTS_GLB_INLINE ErtsMagicBinary *erts_magic_ref_lookup_bin(Uint32 ref[ERTS_REF_NUMBERS]);
+
+#define ERTS_REF1_MAGIC_MARKER_BIT_NO__ \
+    (_REF_NUM_SIZE-1)
+#define ERTS_REF1_MAGIC_MARKER_BIT__ \
+    (((Uint32) 1) << ERTS_REF1_MAGIC_MARKER_BIT_NO__)
+#define ERTS_REF1_THR_ID_MASK__ \
+    (ERTS_REF1_MAGIC_MARKER_BIT__-1)
+#define ERTS_REF1_NUM_MASK__ \
+    (~(ERTS_REF1_THR_ID_MASK__|ERTS_REF1_MAGIC_MARKER_BIT__))
 
 #if ERTS_GLB_INLINE_INCL_FUNC_DEF
 
 ERTS_GLB_INLINE void
-erts_set_ref_numbers(Uint32 ref[ERTS_MAX_REF_NUMBERS], Uint32 thr_id, Uint64 value)
+erts_set_ref_numbers(Uint32 ref[ERTS_REF_NUMBERS], Uint32 thr_id, Uint64 value)
 {
     /*
      * We cannot use thread id in the first 18-bit word since
@@ -87,30 +112,39 @@ erts_set_ref_numbers(Uint32 ref[ERTS_MAX_REF_NUMBERS], Uint32 thr_id, Uint64 val
      * we did, we would get really poor hash values. Instead
      * we have to shuffle the bits a bit.
      */
-    ASSERT(thr_id == (thr_id & ((Uint32) 0x3ffff)));
-    ref[0] = (Uint32) (value & ((Uint64) 0x3ffff));
-    ref[1] = (((Uint32) (value & ((Uint64) 0xfffc0000)))
-	      | (thr_id & ((Uint32) 0x3ffff)));
+    ASSERT(thr_id == (thr_id & ((Uint32) ERTS_REF1_THR_ID_MASK__)));
+    ref[0] = (Uint32) (value & ((Uint64) REF_MASK));
+    ref[1] = (((Uint32) (value & ((Uint64) ERTS_REF1_NUM_MASK__)))
+	      | (thr_id & ((Uint32) ERTS_REF1_THR_ID_MASK__)));
     ref[2] = (Uint32) ((value >> 32) & ((Uint64) 0xffffffff));
 }
 
 ERTS_GLB_INLINE Uint32
-erts_get_ref_numbers_thr_id(Uint32 ref[ERTS_MAX_REF_NUMBERS])
+erts_get_ref_numbers_thr_id(Uint32 ref[ERTS_REF_NUMBERS])
 {
-    return ref[1] & ((Uint32) 0x3ffff);
+    return ref[1] & ((Uint32) ERTS_REF1_THR_ID_MASK__);
+}
+
+ERTS_GLB_INLINE int
+erts_is_ref_numbers_magic(Uint32 ref[ERTS_REF_NUMBERS])
+{
+    return !!(ref[1] & ERTS_REF1_MAGIC_MARKER_BIT__);
 }
 
 ERTS_GLB_INLINE Uint64
-erts_get_ref_numbers_value(Uint32 ref[ERTS_MAX_REF_NUMBERS])
+erts_get_ref_numbers_value(Uint32 ref[ERTS_REF_NUMBERS])
 {
+    ERTS_CT_ASSERT((ERTS_REF1_NUM_MASK__ | REF_MASK) == 0xffffffff);
+    ERTS_CT_ASSERT((ERTS_REF1_NUM_MASK__ & REF_MASK) == 0);
+
     return (((((Uint64) ref[2]) & ((Uint64) 0xffffffff)) << 32)
-	    | (((Uint64) ref[1]) & ((Uint64) 0xfffc0000))
-	    | (((Uint64) ref[0]) & ((Uint64) 0x3ffff)));
+	    | (((Uint64) ref[1]) & ((Uint64) ERTS_REF1_NUM_MASK__))
+	    | (((Uint64) ref[0]) & ((Uint64) REF_MASK)));
 }
 
 ERTS_GLB_INLINE void
 erts_sched_make_ref_in_array(ErtsSchedulerData *esdp,
-			     Uint32 ref[ERTS_MAX_REF_NUMBERS])
+			     Uint32 ref[ERTS_REF_NUMBERS])
 {
     Uint64 value;
 
@@ -119,18 +153,288 @@ erts_sched_make_ref_in_array(ErtsSchedulerData *esdp,
     erts_set_ref_numbers(ref, (Uint32) esdp->thr_id, value);
 }
 
+ERTS_GLB_INLINE void
+erts_sched_make_magic_ref_in_array(ErtsSchedulerData *esdp,
+				   Uint32 ref[ERTS_REF_NUMBERS])
+{
+    erts_sched_make_ref_in_array(esdp, ref);
+    ASSERT(!(ref[1] & ERTS_REF1_MAGIC_MARKER_BIT__));
+    ref[1] |= ERTS_REF1_MAGIC_MARKER_BIT__;
+}
+
 ERTS_GLB_INLINE Eterm
 erts_sched_make_ref_in_buffer(ErtsSchedulerData *esdp,
-			      Eterm buffer[REF_THING_SIZE])
+			      Eterm buffer[ERTS_REF_THING_SIZE])
 {
     Eterm* hp = buffer;
-    Uint32 ref[ERTS_MAX_REF_NUMBERS];
+    Uint32 ref[ERTS_REF_NUMBERS];
 
     erts_sched_make_ref_in_array(esdp, ref);
     write_ref_thing(hp, ref[0], ref[1], ref[2]);
     return make_internal_ref(hp);
 }
 
+ERTS_GLB_INLINE Eterm
+erts_mk_magic_ref(Eterm **hpp, ErlOffHeap *ohp, Binary *bp)
+{
+    Eterm *hp = *hpp;
+    ASSERT(bp->flags & BIN_FLAG_MAGIC);
+    write_magic_ref_thing(hp, ohp, (ErtsMagicBinary *) bp);
+    *hpp += ERTS_MAGIC_REF_THING_SIZE;
+    erts_refc_inc(&bp->refc, 1);
+    OH_OVERHEAD(ohp, bp->orig_size / sizeof(Eterm));
+    return make_internal_ref(hp);
+}
+
+ERTS_GLB_INLINE Binary *
+erts_magic_ref2bin(Eterm mref)
+{
+    ErtsMRefThing *mrtp;
+    ASSERT(is_internal_magic_ref(mref));
+    mrtp = (ErtsMRefThing *) internal_ref_val(mref);
+    return (Binary *) mrtp->mb;
+}
+
+/*
+ * Save the magic binary of a ref when the
+ * ref is exposed to the outside world...
+ */
+ERTS_GLB_INLINE void
+erts_magic_ref_save_bin(Eterm ref)
+{
+    if (is_internal_magic_ref(ref))
+	erts_magic_ref_save_bin__(ref);
+}
+
+/*
+ * Look up the magic binary of a magic ref
+ * when the ref comes from the outside world...
+ */
+ERTS_GLB_INLINE ErtsMagicBinary *
+erts_magic_ref_lookup_bin(Uint32 ref[ERTS_REF_NUMBERS])
+{
+    if (!erts_is_ref_numbers_magic(ref))
+	return NULL;
+    return erts_magic_ref_lookup_bin__(ref);
+}
+
+
 #endif /* ERTS_GLB_INLINE_INCL_FUNC_DEF */
 
+
+/*
+ * Storage of internal refs in misc structures...
+ */
+
+#include "erl_message.h"
+
+#if ERTS_REF_NUMBERS != 3
+#  error fix this...
+#endif
+
+ERTS_GLB_INLINE int erts_internal_ref_number_cmp(Uint32 num1[ERTS_REF_NUMBERS],
+						 Uint32 num2[ERTS_REF_NUMBERS]);
+
+#if ERTS_GLB_INLINE_INCL_FUNC_DEF
+
+ERTS_GLB_INLINE int
+erts_internal_ref_number_cmp(Uint32 num1[ERTS_REF_NUMBERS],
+			     Uint32 num2[ERTS_REF_NUMBERS])
+{
+    if (num1[2] != num2[2])
+	return (int) ((Sint64) num1[2] - (Sint64) num2[2]);
+    if (num1[1] != num2[1])
+	return (int) ((Sint64) num1[1] - (Sint64) num2[1]);
+    if (num1[0] != num2[0])
+	return (int) ((Sint64) num1[0] - (Sint64) num2[0]);
+    return 0;
+}
+
+#endif /* #if ERTS_GLB_INLINE_INCL_FUNC_DEF */
+
+/* Iref storage for all internal references... */ 
+typedef struct {
+    Uint32 is_magic;
+    union {
+	ErtsMagicBinary *mb;
+	Uint32 num[ERTS_REF_NUMBERS];
+    } u;
+} ErtsIRefStorage;
+
+void erts_ref_bin_free(ErtsMagicBinary *mb);
+
+ERTS_GLB_INLINE void erts_iref_storage_save(ErtsIRefStorage *iref, Eterm ref);
+ERTS_GLB_INLINE void erts_iref_storage_clean(ErtsIRefStorage *iref);
+ERTS_GLB_INLINE Uint erts_iref_storage_heap_size(ErtsIRefStorage *iref);
+ERTS_GLB_INLINE Eterm erts_iref_storage_make_ref(ErtsIRefStorage *iref,
+						 Eterm **hpp, ErlOffHeap *ohp,
+						 int clean_storage);
+ERTS_GLB_INLINE int erts_iref_storage_cmp(ErtsIRefStorage *iref1,
+					  ErtsIRefStorage *iref2);
+
+
+#if ERTS_GLB_INLINE_INCL_FUNC_DEF
+
+ERTS_GLB_INLINE void
+erts_iref_storage_save(ErtsIRefStorage *iref, Eterm ref)
+{
+    Eterm *hp;
+
+    ERTS_CT_ASSERT(ERTS_REF_NUMBERS == 3);
+    ASSERT(is_internal_ref(ref));
+
+    hp = boxed_val(ref);
+
+    if (is_ordinary_ref_thing(hp)) {
+	ErtsORefThing *rtp = (ErtsORefThing *) hp;
+	iref->is_magic = 0;
+	iref->u.num[0] = rtp->num[0];
+	iref->u.num[1] = rtp->num[1];
+	iref->u.num[2] = rtp->num[2];
+    }
+    else {
+	ErtsMRefThing *mrtp = (ErtsMRefThing *) hp;
+	ASSERT(is_magic_ref_thing(hp));
+	iref->is_magic = 1;
+	iref->u.mb = mrtp->mb;
+	erts_refc_inc(&mrtp->mb->refc, 1);
+    }
+}
+
+ERTS_GLB_INLINE void
+erts_iref_storage_clean(ErtsIRefStorage *iref)
+{
+    if (iref->is_magic && erts_refc_dectest(&iref->u.mb->refc, 0) == 0)
+	erts_ref_bin_free(iref->u.mb);
+#ifdef DEBUG
+    memset((void *) iref, 0xf, sizeof(ErtsIRefStorage));
+#endif
+}
+
+ERTS_GLB_INLINE Uint
+erts_iref_storage_heap_size(ErtsIRefStorage *iref)
+{
+    return iref->is_magic ? ERTS_MAGIC_REF_THING_SIZE : ERTS_REF_THING_SIZE;
+}
+
+ERTS_GLB_INLINE Eterm
+erts_iref_storage_make_ref(ErtsIRefStorage *iref,
+			   Eterm **hpp, ErlOffHeap *ohp,
+			   int clean_storage)
+{
+    Eterm *hp = *hpp;
+    if (!iref->is_magic) {
+	write_ref_thing(hp, iref->u.num[0], iref->u.num[1],
+			iref->u.num[2]);
+	*hpp += ERTS_REF_THING_SIZE;
+    }
+    else {
+	write_magic_ref_thing(hp, ohp, iref->u.mb);
+        OH_OVERHEAD(ohp, iref->u.mb->orig_size / sizeof(Eterm));
+	*hpp += ERTS_MAGIC_REF_THING_SIZE;
+	/*
+	 * If we clean storage, the term inherits the
+	 * refc increment of the cleaned storage...
+	 */
+	if (!clean_storage)
+	    erts_refc_inc(&iref->u.mb->refc, 1);
+    }
+
+#ifdef DEBUG
+    if (clean_storage)
+	memset((void *) iref, 0xf, sizeof(ErtsIRefStorage));
+#endif
+
+    return make_internal_ref(hp);
+}
+
+ERTS_GLB_INLINE int
+erts_iref_storage_cmp(ErtsIRefStorage *iref1,
+		      ErtsIRefStorage *iref2)
+{
+    Uint32 *num1 = iref1->is_magic ? iref1->u.mb->refn : iref1->u.num;
+    Uint32 *num2 = iref2->is_magic ? iref2->u.mb->refn : iref2->u.num;
+    return erts_internal_ref_number_cmp(num1, num2);
+}
+
+#endif /* #if ERTS_GLB_INLINE_INCL_FUNC_DEF */
+
+
+/* OIref storage for ordinary internal references only... */ 
+typedef struct {
+    Uint32 num[ERTS_REF_NUMBERS];
+} ErtsOIRefStorage;
+
+ERTS_GLB_INLINE void erts_oiref_storage_save(ErtsOIRefStorage *oiref,
+					     Eterm ref);
+ERTS_GLB_INLINE Eterm erts_oiref_storage_make_ref(ErtsOIRefStorage *oiref,
+						  Eterm **hpp);
+ERTS_GLB_INLINE int erts_oiref_storage_cmp(ErtsOIRefStorage *oiref1,
+					   ErtsOIRefStorage *oiref2);
+
+#if ERTS_GLB_INLINE_INCL_FUNC_DEF
+
+ERTS_GLB_INLINE void
+erts_oiref_storage_save(ErtsOIRefStorage *oiref, Eterm ref)
+{
+    ErtsORefThing *rtp;
+    ERTS_CT_ASSERT(ERTS_REF_NUMBERS == 3);
+    ASSERT(is_internal_ordinary_ref(ref));
+
+    rtp = (ErtsORefThing *) internal_ref_val(ref);
+
+    oiref->num[0] = rtp->num[0];
+    oiref->num[1] = rtp->num[1];
+    oiref->num[2] = rtp->num[2];
+}
+
+ERTS_GLB_INLINE Eterm
+erts_oiref_storage_make_ref(ErtsOIRefStorage *oiref, Eterm **hpp)
+{
+    Eterm *hp = *hpp;
+    ERTS_CT_ASSERT(ERTS_REF_NUMBERS == 3);
+    write_ref_thing(hp, oiref->num[0], oiref->num[1], oiref->num[2]);
+    *hpp += ERTS_REF_THING_SIZE;
+    return make_internal_ref(hp);
+}
+
+ERTS_GLB_INLINE int
+erts_oiref_storage_cmp(ErtsOIRefStorage *oiref1,
+		       ErtsOIRefStorage *oiref2)
+{
+    return erts_internal_ref_number_cmp(oiref1->num, oiref2->num);
+}
+
+#endif /* #if ERTS_GLB_INLINE_INCL_FUNC_DEF */
+
+
+ERTS_GLB_INLINE Eterm
+erts_proc_store_ref(Process *c_p, Uint32 ref[ERTS_MAX_REF_NUMBERS]);
+
+#if ERTS_GLB_INLINE_INCL_FUNC_DEF
+
+ERTS_GLB_INLINE Eterm
+erts_proc_store_ref(Process *c_p, Uint32 ref[ERTS_MAX_REF_NUMBERS])
+{
+    Eterm *hp = HAlloc(c_p, ERTS_REF_THING_SIZE);
+    write_ref_thing(hp, ref[0], ref[1], ref[2]);
+    return make_internal_ref(hp);
+}
+
+#endif
+
 #endif /*  ERTS_BIF_UNIQUE_H__ */
+
+#if (defined(ERTS_ALLOC_C__) || defined(ERL_BIF_UNIQUE_C__)) \
+    && !defined(ERTS_BIF_UNIQUE_H__FIX_ALLOC_TYPES__)
+#define ERTS_BIF_UNIQUE_H__FIX_ALLOC_TYPES__
+
+#include "hash.h"
+
+typedef struct {
+    HashBucket hash;
+    ErtsMagicBinary *mb;
+    Uint64 value;
+} ErtsNSchedMagicRefTableEntry;
+
+#endif
diff --git a/erts/emulator/beam/erl_binary.h b/erts/emulator/beam/erl_binary.h
index 52efa06816..db259be2a7 100644
--- a/erts/emulator/beam/erl_binary.h
+++ b/erts/emulator/beam/erl_binary.h
@@ -18,11 +18,146 @@
  * %CopyrightEnd%
  */
 
-#ifndef __ERL_BINARY_H
-#define __ERL_BINARY_H
+#ifndef ERL_BINARY_H__TYPES__
+#define ERL_BINARY_H__TYPES__
+
+/*
+** Just like the driver binary but with initial flags
+** Note that the two structures Binary and ErlDrvBinary HAVE to
+** be equal except for extra fields in the beginning of the struct.
+** ErlDrvBinary is defined in erl_driver.h.
+** When driver_alloc_binary is called, a Binary is allocated, but 
+** the pointer returned is to the address of the first element that
+** also occurs in the ErlDrvBinary struct (driver.*binary takes care if this).
+** The driver need never know about additions to the internal Binary of the
+** emulator. One should however NEVER be sloppy when mixing ErlDrvBinary
+** and Binary, the macros below can convert one type to the other, as they both
+** in reality are equal.
+*/
+
+#ifdef ARCH_32
+ /* *DO NOT USE* only for alignment. */
+#define ERTS_BINARY_STRUCT_ALIGNMENT Uint32 align__;
+#else
+#define ERTS_BINARY_STRUCT_ALIGNMENT
+#endif
+
+/* Add fields in ERTS_INTERNAL_BINARY_FIELDS, otherwise the drivers crash */
+#define ERTS_INTERNAL_BINARY_FIELDS				\
+    UWord flags;						\
+    erts_refc_t refc;						\
+    ERTS_BINARY_STRUCT_ALIGNMENT
+
+typedef struct binary {
+    ERTS_INTERNAL_BINARY_FIELDS
+    SWord orig_size;
+    char orig_bytes[1]; /* to be continued */
+} Binary;
+
+#define ERTS_SIZEOF_Binary(Sz) \
+    (offsetof(Binary,orig_bytes) + (Sz))
+
+#if ERTS_REF_NUMBERS != 3
+#error "Update ErtsMagicBinary"
+#endif
+
+typedef struct magic_binary ErtsMagicBinary;
+struct magic_binary {
+    ERTS_INTERNAL_BINARY_FIELDS
+    SWord orig_size;
+    int (*destructor)(Binary *);
+    Uint32 refn[ERTS_REF_NUMBERS];
+    ErtsAlcType_t alloc_type;
+    union {
+        struct {
+            ERTS_BINARY_STRUCT_ALIGNMENT
+            char data[1];
+        } aligned;
+        struct {
+            char data[1];
+        } unaligned;
+    } u;
+};
+
+#ifdef ARCH_32
+#define ERTS_MAGIC_BIN_BYTES_TO_ALIGN 4
+#else
+#define ERTS_MAGIC_BIN_BYTES_TO_ALIGN 0
+#endif
+
+typedef union {
+    Binary binary;
+    ErtsMagicBinary magic_binary;
+    struct {
+	ERTS_INTERNAL_BINARY_FIELDS
+	ErlDrvBinary binary;
+    } driver;
+} ErtsBinary;
+
+/*
+ * 'Binary' alignment:
+ *   Address of orig_bytes[0] of a Binary should always be 8-byte aligned.
+ * It is assumed that the flags, refc, and orig_size fields are 4 bytes on
+ * 32-bits architectures and 8 bytes on 64-bits architectures.
+ */
+
+#define ERTS_MAGIC_BIN_REFN(BP) \
+  ((ErtsBinary *) (BP))->magic_binary.refn
+#define ERTS_MAGIC_BIN_ATYPE(BP) \
+  ((ErtsBinary *) (BP))->magic_binary.alloc_type
+#define ERTS_MAGIC_DATA_OFFSET \
+  (offsetof(ErtsMagicBinary,u.aligned.data) - offsetof(Binary,orig_bytes))
+#define ERTS_MAGIC_BIN_DESTRUCTOR(BP) \
+  ((ErtsBinary *) (BP))->magic_binary.destructor
+#define ERTS_MAGIC_BIN_DATA(BP) \
+  ((void *) ((ErtsBinary *) (BP))->magic_binary.u.aligned.data)
+#define ERTS_MAGIC_BIN_DATA_SIZE(BP) \
+  ((BP)->orig_size - ERTS_MAGIC_DATA_OFFSET)
+#define ERTS_MAGIC_DATA_OFFSET \
+  (offsetof(ErtsMagicBinary,u.aligned.data) - offsetof(Binary,orig_bytes))
+#define ERTS_MAGIC_BIN_ORIG_SIZE(Sz) \
+  (ERTS_MAGIC_DATA_OFFSET + (Sz))
+#define ERTS_MAGIC_BIN_SIZE(Sz) \
+  (offsetof(ErtsMagicBinary,u.aligned.data) + (Sz))
+#define ERTS_MAGIC_BIN_FROM_DATA(DATA) \
+  ((ErtsBinary*)((char*)(DATA) - offsetof(ErtsMagicBinary,u.aligned.data)))
+
+/* On 32-bit arch these macro variants will save memory
+   by not forcing 8-byte alignment for the magic payload.
+*/
+#define ERTS_MAGIC_BIN_UNALIGNED_DATA(BP) \
+  ((void *) ((ErtsBinary *) (BP))->magic_binary.u.unaligned.data)
+#define ERTS_MAGIC_UNALIGNED_DATA_OFFSET \
+  (offsetof(ErtsMagicBinary,u.unaligned.data) - offsetof(Binary,orig_bytes))
+#define ERTS_MAGIC_BIN_UNALIGNED_DATA_SIZE(BP) \
+  ((BP)->orig_size - ERTS_MAGIC_UNALIGNED_DATA_OFFSET)
+#define ERTS_MAGIC_BIN_UNALIGNED_ORIG_SIZE(Sz) \
+  (ERTS_MAGIC_UNALIGNED_DATA_OFFSET + (Sz))
+#define ERTS_MAGIC_BIN_UNALIGNED_SIZE(Sz) \
+  (offsetof(ErtsMagicBinary,u.unaligned.data) + (Sz))
+#define ERTS_MAGIC_BIN_FROM_UNALIGNED_DATA(DATA) \
+  ((ErtsBinary*)((char*)(DATA) - offsetof(ErtsMagicBinary,u.unaligned.data)))
+
+
+#define Binary2ErlDrvBinary(B) (&((ErtsBinary *) (B))->driver.binary)
+#define ErlDrvBinary2Binary(D) ((Binary *) \
+				(((char *) (D)) \
+				 - offsetof(ErtsBinary, driver.binary)))
+
+/* A "magic" binary flag */
+#define BIN_FLAG_MAGIC      1
+#define BIN_FLAG_USR1       2 /* Reserved for use by different modules too mark */
+#define BIN_FLAG_USR2       4 /*  certain binaries as special (used by ets) */
+#define BIN_FLAG_DRV        8
+
+#endif /* ERL_BINARY_H__TYPES__ */
+
+#if !defined(ERL_BINARY_H__) && !defined(ERTS_BINARY_TYPES_ONLY__)
+#define ERL_BINARY_H__
 
 #include "erl_threads.h"
 #include "bif.h"
+#include "erl_bif_unique.h"
 
 /*
  * Maximum number of bytes to place in a heap binary.
@@ -165,6 +300,17 @@ BIF_RETTYPE erts_gc_binary_part(Process *p, Eterm *reg, Eterm live, int range_is
 BIF_RETTYPE erts_binary_part(Process *p, Eterm binary, Eterm epos, Eterm elen);
 
 
+typedef union {
+    /*
+     * These two are almost always of
+     * the same size, but when fallback
+     * atomics are used they might
+     * differ in size.
+     */
+    erts_smp_atomic_t smp_atomic_word;
+    erts_atomic_t atomic_word;
+} ErtsMagicIndirectionWord;
+
 #if defined(__i386__) || !defined(__GNUC__)
 /*
  * Doubles aren't required to be 8-byte aligned on intel x86.
@@ -190,9 +336,13 @@ ERTS_GLB_INLINE Binary *erts_bin_realloc(Binary *bp, Uint size);
 ERTS_GLB_INLINE void erts_bin_free(Binary *bp);
 ERTS_GLB_INLINE Binary *erts_create_magic_binary_x(Uint size,
                                                   int (*destructor)(Binary *),
+                                                   ErtsAlcType_t alloc_type,
                                                   int unaligned);
 ERTS_GLB_INLINE Binary *erts_create_magic_binary(Uint size,
 						 int (*destructor)(Binary *));
+ERTS_GLB_INLINE Binary *erts_create_magic_indirection(int (*destructor)(Binary *));
+ERTS_GLB_INLINE erts_smp_atomic_t *erts_smp_binary_to_magic_indirection(Binary *bp);
+ERTS_GLB_INLINE erts_atomic_t *erts_binary_to_magic_indirection(Binary *bp);
 
 #if ERTS_GLB_INLINE_INCL_FUNC_DEF
 
@@ -321,12 +471,14 @@ ERTS_GLB_INLINE void
 erts_bin_free(Binary *bp)
 {
     if (bp->flags & BIN_FLAG_MAGIC) {
-	if (!ERTS_MAGIC_BIN_DESTRUCTOR(bp)(bp)) {
+        if (!ERTS_MAGIC_BIN_DESTRUCTOR(bp)(bp)) {
             /* Destructor took control of the deallocation */
             return;
         }
+	erts_magic_ref_remove_bin(ERTS_MAGIC_BIN_REFN(bp));
+        erts_free(ERTS_MAGIC_BIN_ATYPE(bp), (void *) bp);
     }
-    if (bp->flags & BIN_FLAG_DRV)
+    else if (bp->flags & BIN_FLAG_DRV)
 	erts_free(ERTS_ALC_T_DRV_BINARY, (void *) bp);
     else
 	erts_free(ERTS_ALC_T_BINARY, (void *) bp);
@@ -334,29 +486,63 @@ erts_bin_free(Binary *bp)
 
 ERTS_GLB_INLINE Binary *
 erts_create_magic_binary_x(Uint size, int (*destructor)(Binary *),
+                           ErtsAlcType_t alloc_type,
                            int unaligned)
 {
     Uint bsize = unaligned ? ERTS_MAGIC_BIN_UNALIGNED_SIZE(size)
                            : ERTS_MAGIC_BIN_SIZE(size);
-    Binary* bptr = erts_alloc_fnf(ERTS_ALC_T_BINARY, bsize);
+    Binary* bptr = erts_alloc_fnf(alloc_type, bsize);
     ASSERT(bsize > size);
     if (!bptr)
-	erts_alloc_n_enomem(ERTS_ALC_T2N(ERTS_ALC_T_BINARY), bsize);
+	erts_alloc_n_enomem(ERTS_ALC_T2N(alloc_type), bsize);
     ERTS_CHK_BIN_ALIGNMENT(bptr);
     bptr->flags = BIN_FLAG_MAGIC;
     bptr->orig_size = unaligned ? ERTS_MAGIC_BIN_UNALIGNED_ORIG_SIZE(size)
                                 : ERTS_MAGIC_BIN_ORIG_SIZE(size);
     erts_refc_init(&bptr->refc, 0);
     ERTS_MAGIC_BIN_DESTRUCTOR(bptr) = destructor;
+    ERTS_MAGIC_BIN_ATYPE(bptr) = alloc_type;
+    erts_make_magic_ref_in_array(ERTS_MAGIC_BIN_REFN(bptr));
     return bptr;
 }
 
 ERTS_GLB_INLINE Binary *
 erts_create_magic_binary(Uint size, int (*destructor)(Binary *))
 {
-    return erts_create_magic_binary_x(size, destructor, 0);
+    return erts_create_magic_binary_x(size, destructor,
+                                      ERTS_ALC_T_BINARY, 0);
+}
+
+ERTS_GLB_INLINE Binary *
+erts_create_magic_indirection(int (*destructor)(Binary *))
+{
+    return erts_create_magic_binary_x(sizeof(ErtsMagicIndirectionWord),
+                                      destructor,
+                                      ERTS_ALC_T_MINDIRECTION,
+                                      1); /* Not 64-bit aligned,
+                                             but word aligned */
+}
+
+ERTS_GLB_INLINE erts_smp_atomic_t *
+erts_smp_binary_to_magic_indirection(Binary *bp)
+{
+    ErtsMagicIndirectionWord *mip;
+    ASSERT(bp->flags & BIN_FLAG_MAGIC);
+    ASSERT(ERTS_MAGIC_BIN_ATYPE(bp) == ERTS_ALC_T_MINDIRECTION);
+    mip = ERTS_MAGIC_BIN_UNALIGNED_DATA(bp);
+    return &mip->smp_atomic_word;
+}
+
+ERTS_GLB_INLINE erts_atomic_t *
+erts_binary_to_magic_indirection(Binary *bp)
+{
+    ErtsMagicIndirectionWord *mip;
+    ASSERT(bp->flags & BIN_FLAG_MAGIC);
+    ASSERT(ERTS_MAGIC_BIN_ATYPE(bp) == ERTS_ALC_T_MINDIRECTION);
+    mip = ERTS_MAGIC_BIN_UNALIGNED_DATA(bp);
+    return &mip->atomic_word;
 }
 
 #endif /* #if ERTS_GLB_INLINE_INCL_FUNC_DEF */
 
-#endif /* !__ERL_BINARY_H */
+#endif /* !ERL_BINARY_H__ */
diff --git a/erts/emulator/beam/erl_db.c b/erts/emulator/beam/erl_db.c
index fd092015bd..9f077dd407 100644
--- a/erts/emulator/beam/erl_db.c
+++ b/erts/emulator/beam/erl_db.c
@@ -266,7 +266,7 @@ static void schedule_free_dbtable(DbTable* tb)
      *  	     Caller is *not* allowed to access the specialized part
      *  	     (hash or tree) of *tb after this function has returned.
      */
-    ASSERT(erts_refc_read(&tb->common.ref, 0) == 0);
+    ASSERT(erts_smp_refc_read(&tb->common.ref, 0) == 0);
     erts_schedule_thr_prgr_later_cleanup_op(free_dbtable,
 					    (void *) tb,
 					    &tb->release.data,
@@ -600,11 +600,11 @@ done:
 */
 static ERTS_INLINE void local_fix_table(DbTable* tb)
 {
-    erts_refc_inc(&tb->common.ref, 1);
+    erts_smp_refc_inc(&tb->common.ref, 1);
 }	    
 static ERTS_INLINE void local_unfix_table(DbTable* tb)
 {	
-    if (erts_refc_dectest(&tb->common.ref, 0) == 0) {
+    if (erts_smp_refc_dectest(&tb->common.ref, 0) == 0) {
 	ASSERT(IS_HASH_TABLE(tb->common.status));
 	db_unfix_table_hash(&(tb->hash));
     }
@@ -1487,7 +1487,7 @@ BIF_RETTYPE ets_new_2(BIF_ALIST_2)
     tb->common.type = status & ERTS_ETS_TABLE_TYPES;
     /* Note, 'type' is *read only* from now on... */
 #endif
-    erts_refc_init(&tb->common.ref, 0);
+    erts_smp_refc_init(&tb->common.ref, 0);
     db_init_lock(tb, status & (DB_FINE_LOCKED|DB_FREQ_READ),
 		 "db_tab", "db_tab_fix");
     tb->common.keypos = keypos;
@@ -2779,7 +2779,7 @@ BIF_RETTYPE ets_info_2(BIF_ALIST_2)
 
 BIF_RETTYPE ets_is_compiled_ms_1(BIF_ALIST_1)
 {
-    if (erts_db_is_compiled_ms(BIF_ARG_1)) {
+    if (erts_db_get_match_prog_binary(BIF_ARG_1)) {
 	BIF_RET(am_true);
     } else {
 	BIF_RET(am_false);
@@ -2794,9 +2794,9 @@ BIF_RETTYPE ets_match_spec_compile_1(BIF_ALIST_1)
 	BIF_ERROR(BIF_P, BADARG);
     }
 
-    hp = HAlloc(BIF_P, PROC_BIN_SIZE);
+    hp = HAlloc(BIF_P, ERTS_MAGIC_REF_THING_SIZE);
 
-    BIF_RET(erts_mk_magic_binary_term(&hp, &MSO(BIF_P), mp));
+    BIF_RET(erts_db_make_match_prog_ref(BIF_P, mp, &hp));
 }
 
 BIF_RETTYPE ets_match_spec_run_r_3(BIF_ALIST_3)
@@ -2805,24 +2805,18 @@ BIF_RETTYPE ets_match_spec_run_r_3(BIF_ALIST_3)
     int i = 0;
     Eterm *hp;
     Eterm lst;
-    ProcBin *bp;
     Binary *mp;
     Eterm res;
     Uint32 dummy;
 
-    if (!(is_list(BIF_ARG_1) || BIF_ARG_1 == NIL) || !is_binary(BIF_ARG_2)) {
+    if (!(is_list(BIF_ARG_1) || BIF_ARG_1 == NIL)) {
     error:
 	BIF_ERROR(BIF_P, BADARG);
     }
     
-    bp = (ProcBin*) binary_val(BIF_ARG_2);
-    if (thing_subtag(bp->thing_word) != REFC_BINARY_SUBTAG) {
+    mp = erts_db_get_match_prog_binary(BIF_ARG_2);
+    if (!mp)
 	goto error;
-    }
-    mp = bp->val;
-    if (!IsMatchProgBinary(mp)) {
-	goto error;
-    }
 
     if (BIF_ARG_1 == NIL) {
 	BIF_RET(BIF_ARG_3);
@@ -2990,7 +2984,7 @@ void init_db(ErtsDbSpinCount db_spin_count)
     meta_pid_to_tab->common.meth   = &db_hash;
     meta_pid_to_tab->common.compress = 0;
 
-    erts_refc_init(&meta_pid_to_tab->common.ref, 0);
+    erts_smp_refc_init(&meta_pid_to_tab->common.ref, 0);
     /* Neither rwlock or fixlock used
     db_init_lock(meta_pid_to_tab, "meta_pid_to_tab", "meta_pid_to_tab_FIX");*/
 
@@ -3021,7 +3015,7 @@ void init_db(ErtsDbSpinCount db_spin_count)
     meta_pid_to_fixed_tab->common.meth   = &db_hash;
     meta_pid_to_fixed_tab->common.compress = 0;
 
-    erts_refc_init(&meta_pid_to_fixed_tab->common.ref, 0);
+    erts_smp_refc_init(&meta_pid_to_fixed_tab->common.ref, 0);
     /* Neither rwlock or fixlock used
     db_init_lock(meta_pid_to_fixed_tab, "meta_pid_to_fixed_tab", "meta_pid_to_fixed_tab_FIX");*/
 
@@ -3382,7 +3376,7 @@ erts_db_process_exiting(Process *c_p, ErtsProcLocks c_p_locks)
 			    if ((*pp)->pid == pid) {
 				DbFixation* fix = *pp;
 				erts_aint_t diff = -((erts_aint_t) fix->counter);
-				erts_refc_add(&tb->common.ref,diff,0);
+				erts_smp_refc_add(&tb->common.ref,diff,0);
 				*pp = fix->next;
 				erts_db_free(ERTS_ALC_T_DB_FIXATION,
 					     tb, fix, sizeof(DbFixation));
@@ -3458,7 +3452,7 @@ static void fix_table_locked(Process* p, DbTable* tb)
 #ifdef ERTS_SMP
     erts_smp_mtx_lock(&tb->common.fixlock);
 #endif
-    erts_refc_inc(&tb->common.ref,1);
+    erts_smp_refc_inc(&tb->common.ref,1);
     fix = tb->common.fixations;
     if (fix == NULL) {
 	tb->common.time.monotonic
@@ -3514,7 +3508,7 @@ static void unfix_table_locked(Process* p,  DbTable* tb,
     for (pp = &tb->common.fixations; *pp != NULL; pp = &(*pp)->next) {
 	if ((*pp)->pid == p->common.id) {
 	    DbFixation* fix = *pp;
-	    erts_refc_dec(&tb->common.ref,0);
+	    erts_smp_refc_dec(&tb->common.ref,0);
 	    --(fix->counter);
 	    ASSERT(fix->counter >= 0);
 	    if (fix->counter > 0) {
@@ -3563,7 +3557,7 @@ static void free_fixations_locked(DbTable *tb)
     fix = tb->common.fixations;
     while (fix != NULL) {
 	erts_aint_t diff = -((erts_aint_t) fix->counter);
-	erts_refc_add(&tb->common.ref,diff,0);
+	erts_smp_refc_add(&tb->common.ref,diff,0);
 	next_fix = fix->next;
 	db_meta_lock(meta_pid_to_fixed_tab, LCK_WRITE_REC);
 	db_erase_bag_exact2(meta_pid_to_fixed_tab,
diff --git a/erts/emulator/beam/erl_db_hash.c b/erts/emulator/beam/erl_db_hash.c
index b2a231467b..bb29d56aa7 100644
--- a/erts/emulator/beam/erl_db_hash.c
+++ b/erts/emulator/beam/erl_db_hash.c
@@ -1287,15 +1287,13 @@ static int db_select_continue_hash(Process *p,
     if (arityval(*tptr) != 6)
 	RET_TO_BIF(NIL,DB_ERROR_BADPARAM);
     
-    if (!is_small(tptr[2]) || !is_small(tptr[3]) || !is_binary(tptr[4]) || 
+    if (!is_small(tptr[2]) || !is_small(tptr[3]) || 
 	!(is_list(tptr[5]) || tptr[5] == NIL) || !is_small(tptr[6]))
 	RET_TO_BIF(NIL,DB_ERROR_BADPARAM);
     if ((chunk_size = signed_val(tptr[3])) < 0)
 	RET_TO_BIF(NIL,DB_ERROR_BADPARAM);
-    if (!(thing_subtag(*binary_val(tptr[4])) == REFC_BINARY_SUBTAG))
-	RET_TO_BIF(NIL,DB_ERROR_BADPARAM);
-    mp = ((ProcBin *) binary_val(tptr[4]))->val;
-    if (!IsMatchProgBinary(mp))
+    mp = erts_db_get_match_prog_binary(tptr[4]);
+    if (!mp)
 	RET_TO_BIF(NIL,DB_ERROR_BADPARAM);
     all_objects = mp->flags & BIN_FLAG_ALL_OBJECTS;
     match_list = tptr[5];
@@ -1554,8 +1552,8 @@ done:
 					 been in 'user space' */ 
 	}
 	if (rest != NIL || slot_ix >= 0) { /* Need more calls */
-	    hp = HAlloc(p,3+7+PROC_BIN_SIZE);
-	    mpb =db_make_mp_binary(p,(mpi.mp),&hp);
+	    hp = HAlloc(p,3+7+ERTS_MAGIC_REF_THING_SIZE);
+	    mpb = erts_db_make_match_prog_ref(p,(mpi.mp),&hp);
 	    if (mpi.all_objects)
 		(mpi.mp)->flags |= BIN_FLAG_ALL_OBJECTS;
 	    continuation = TUPLE6(hp, tb->common.id,make_small(slot_ix), 
@@ -1580,8 +1578,8 @@ trap:
     BUMP_ALL_REDS(p);
     if (mpi.all_objects)
 	(mpi.mp)->flags |= BIN_FLAG_ALL_OBJECTS;
-    hp = HAlloc(p,7+PROC_BIN_SIZE);
-    mpb =db_make_mp_binary(p,(mpi.mp),&hp);
+    hp = HAlloc(p,7+ERTS_MAGIC_REF_THING_SIZE);
+    mpb = erts_db_make_match_prog_ref(p,(mpi.mp),&hp);
     continuation = TUPLE6(hp, tb->common.id, make_small(slot_ix), 
 			  make_small(chunk_size), 
 			  mpb, match_list, 
@@ -1693,15 +1691,15 @@ done:
 trap:
     BUMP_ALL_REDS(p);
     if (IS_USMALL(0, got)) {
-	hp = HAlloc(p,  PROC_BIN_SIZE + 5);
+	hp = HAlloc(p,  ERTS_MAGIC_REF_THING_SIZE + 5);
 	egot = make_small(got);
     }
     else {
-	hp = HAlloc(p, BIG_UINT_HEAP_SIZE + PROC_BIN_SIZE + 5);
+	hp = HAlloc(p, BIG_UINT_HEAP_SIZE + ERTS_MAGIC_REF_THING_SIZE + 5);
 	egot = uint_to_big(got, hp);
 	hp += BIG_UINT_HEAP_SIZE;
     }
-    mpb = db_make_mp_binary(p,mpi.mp,&hp);
+    mpb = erts_db_make_match_prog_ref(p,mpi.mp,&hp);
     continuation = TUPLE4(hp, tb->common.id, make_small(slot_ix), 
 			  mpb, 
 			  egot);
@@ -1838,15 +1836,15 @@ done:
 trap:
     BUMP_ALL_REDS(p);
     if (IS_USMALL(0, got)) {
-	hp = HAlloc(p,  PROC_BIN_SIZE + 5);
+	hp = HAlloc(p,  ERTS_MAGIC_REF_THING_SIZE + 5);
 	egot = make_small(got);
     }
     else {
-	hp = HAlloc(p, BIG_UINT_HEAP_SIZE + PROC_BIN_SIZE + 5);
+	hp = HAlloc(p, BIG_UINT_HEAP_SIZE + ERTS_MAGIC_REF_THING_SIZE + 5);
 	egot = uint_to_big(got, hp);
 	hp += BIG_UINT_HEAP_SIZE;
     }
-    mpb = db_make_mp_binary(p,mpi.mp,&hp);
+    mpb = erts_db_make_match_prog_ref(p,mpi.mp,&hp);
     continuation = TUPLE4(hp, tb->common.id, make_small(slot_ix), 
 			  mpb, 
 			  egot);
@@ -1887,7 +1885,7 @@ static int db_select_delete_continue_hash(Process *p,
     
     tptr = tuple_val(continuation);
     slot_ix = unsigned_val(tptr[2]);
-    mp = ((ProcBin *) binary_val(tptr[3]))->val;
+    mp = erts_db_get_match_prog_binary_unchecked(tptr[3]);
     if (is_big(tptr[4])) {
 	got = big_to_uint32(tptr[4]);
     } else {
@@ -1999,7 +1997,7 @@ static int db_select_count_continue_hash(Process *p,
     
     tptr = tuple_val(continuation);
     slot_ix = unsigned_val(tptr[2]);
-    mp = ((ProcBin *) binary_val(tptr[3]))->val;
+    mp = erts_db_get_match_prog_binary_unchecked(tptr[3]);
     if (is_big(tptr[4])) {
 	got = big_to_uint32(tptr[4]);
     } else {
diff --git a/erts/emulator/beam/erl_db_tree.c b/erts/emulator/beam/erl_db_tree.c
index dd9403e132..c4ecd2ba37 100644
--- a/erts/emulator/beam/erl_db_tree.c
+++ b/erts/emulator/beam/erl_db_tree.c
@@ -935,17 +935,15 @@ static int db_select_continue_tree(Process *p,
     if (arityval(*tptr) != 8)
 	RET_TO_BIF(NIL,DB_ERROR_BADPARAM);
     
-    if (!is_small(tptr[4]) || !is_binary(tptr[5]) || 
+    if (!is_small(tptr[4]) ||
 	!(is_list(tptr[6]) || tptr[6] == NIL) || !is_small(tptr[7]) ||
 	!is_small(tptr[8]))
 	RET_TO_BIF(NIL,DB_ERROR_BADPARAM);
     
     lastkey = tptr[2];
     end_condition = tptr[3];
-    if (!(thing_subtag(*binary_val(tptr[5])) == REFC_BINARY_SUBTAG))
-	RET_TO_BIF(NIL,DB_ERROR_BADPARAM);
-    mp = ((ProcBin *) binary_val(tptr[5]))->val;
-    if (!IsMatchProgBinary(mp))
+    mp = erts_db_get_match_prog_binary(tptr[5]);
+    if (!mp)
 	RET_TO_BIF(NIL,DB_ERROR_BADPARAM);
     chunk_size = signed_val(tptr[4]);
 
@@ -1145,11 +1143,11 @@ static int db_select_tree(Process *p, DbTable *tbl,
 
     key = GETKEY(tb, sc.lastobj);
     sz = size_object(key);
-    hp = HAlloc(p, 9 + sz + PROC_BIN_SIZE);
+    hp = HAlloc(p, 9 + sz + ERTS_MAGIC_REF_THING_SIZE);
     key = copy_struct(key, sz, &hp, &MSO(p));
     if (mpi.all_objects)
 	(mpi.mp)->flags |= BIN_FLAG_ALL_OBJECTS;
-    mpb=db_make_mp_binary(p,mpi.mp,&hp);
+    mpb= erts_db_make_match_prog_ref(p,mpi.mp,&hp);
 	    
     continuation = TUPLE8
 	(hp,
@@ -1208,10 +1206,8 @@ static int db_select_count_continue_tree(Process *p,
     
     lastkey = tptr[2];
     end_condition = tptr[3];
-    if (!(thing_subtag(*binary_val(tptr[4])) == REFC_BINARY_SUBTAG))
-	RET_TO_BIF(NIL,DB_ERROR_BADPARAM);
-    mp = ((ProcBin *) binary_val(tptr[4]))->val;
-    if (!IsMatchProgBinary(mp))
+    mp = erts_db_get_match_prog_binary(tptr[4]);
+    if (!mp)
 	RET_TO_BIF(NIL,DB_ERROR_BADPARAM);
 
     sc.p = p;
@@ -1338,18 +1334,18 @@ static int db_select_count_tree(Process *p, DbTable *tbl,
     key = GETKEY(tb, sc.lastobj);
     sz = size_object(key);
     if (IS_USMALL(0, sc.got)) {
-	hp = HAlloc(p, sz + PROC_BIN_SIZE + 6);
+	hp = HAlloc(p, sz + ERTS_MAGIC_REF_THING_SIZE + 6);
 	egot = make_small(sc.got);
     }
     else {
-	hp = HAlloc(p, BIG_UINT_HEAP_SIZE + sz + PROC_BIN_SIZE + 6);
+	hp = HAlloc(p, BIG_UINT_HEAP_SIZE + sz + ERTS_MAGIC_REF_THING_SIZE + 6);
 	egot = uint_to_big(sc.got, hp);
 	hp += BIG_UINT_HEAP_SIZE;
     }
     key = copy_struct(key, sz, &hp, &MSO(p));
     if (mpi.all_objects)
 	(mpi.mp)->flags |= BIN_FLAG_ALL_OBJECTS;
-    mpb = db_make_mp_binary(p,mpi.mp,&hp);
+    mpb = erts_db_make_match_prog_ref(p,mpi.mp,&hp);
 	    
     continuation = TUPLE5
 	(hp,
@@ -1470,11 +1466,11 @@ static int db_select_chunk_tree(Process *p, DbTable *tbl,
 
 	key = GETKEY(tb, sc.lastobj);
 	sz = size_object(key);
-	hp = HAlloc(p, 9 + sz + PROC_BIN_SIZE);
+	hp = HAlloc(p, 9 + sz + ERTS_MAGIC_REF_THING_SIZE);
 	key = copy_struct(key, sz, &hp, &MSO(p));
 	if (mpi.all_objects)
 	    (mpi.mp)->flags |= BIN_FLAG_ALL_OBJECTS;
-	mpb = db_make_mp_binary(p,mpi.mp,&hp);
+	mpb = erts_db_make_match_prog_ref(p,mpi.mp,&hp);
 	
 	continuation = TUPLE8
 	    (hp,
@@ -1495,12 +1491,12 @@ static int db_select_chunk_tree(Process *p, DbTable *tbl,
 
     key = GETKEY(tb, sc.lastobj);
     sz = size_object(key);
-    hp = HAlloc(p, 9 + sz + PROC_BIN_SIZE);
+    hp = HAlloc(p, 9 + sz + ERTS_MAGIC_REF_THING_SIZE);
     key = copy_struct(key, sz, &hp, &MSO(p));
 
     if (mpi.all_objects)
 	(mpi.mp)->flags |= BIN_FLAG_ALL_OBJECTS;
-    mpb = db_make_mp_binary(p,mpi.mp,&hp);    
+    mpb = erts_db_make_match_prog_ref(p,mpi.mp,&hp);
     continuation = TUPLE8
 	(hp,
 	 tb->common.id,
@@ -1558,7 +1554,7 @@ static int db_select_delete_continue_tree(Process *p,
     sc.erase_lastterm = 0; /* Before first RET_TO_BIF */
     sc.lastterm = NULL;
 
-    mp = ((ProcBin *) binary_val(tptr[4]))->val;
+    mp = erts_db_get_match_prog_binary_unchecked(tptr[4]);
     sc.p = p;
     sc.tb = tb;
     if (is_big(tptr[5])) {
@@ -1682,16 +1678,16 @@ static int db_select_delete_tree(Process *p, DbTable *tbl,
     key = GETKEY(tb, (sc.lastterm)->dbterm.tpl);
     sz = size_object(key);
     if (IS_USMALL(0, sc.accum)) {
-	hp = HAlloc(p, sz + PROC_BIN_SIZE + 6);
+	hp = HAlloc(p, sz + ERTS_MAGIC_REF_THING_SIZE + 6);
 	eaccsum = make_small(sc.accum);
     }
     else {
-	hp = HAlloc(p, BIG_UINT_HEAP_SIZE + sz + PROC_BIN_SIZE + 6);
+	hp = HAlloc(p, BIG_UINT_HEAP_SIZE + sz + ERTS_MAGIC_REF_THING_SIZE + 6);
 	eaccsum = uint_to_big(sc.accum, hp);
 	hp += BIG_UINT_HEAP_SIZE;
     }
     key = copy_struct(key, sz, &hp, &MSO(p));
-    mpb = db_make_mp_binary(p,mpi.mp,&hp);
+    mpb = erts_db_make_match_prog_ref(p,mpi.mp,&hp);
     
     continuation = TUPLE5
 	(hp,
diff --git a/erts/emulator/beam/erl_db_util.c b/erts/emulator/beam/erl_db_util.c
index 991d63a18b..6f30b1d3dd 100644
--- a/erts/emulator/beam/erl_db_util.c
+++ b/erts/emulator/beam/erl_db_util.c
@@ -2546,14 +2546,6 @@ success:
 }
 
 
-/*
- * Convert a match program to a "magic" binary to return up to erlang
- */
-Eterm db_make_mp_binary(Process *p, Binary *mp, Eterm **hpp)
-{
-    return erts_mk_magic_binary_term(hpp, &MSO(p), mp);
-}
-
 DMCErrInfo *db_new_dmc_err_info(void) 
 {
     DMCErrInfo *ret = erts_alloc(ERTS_ALC_T_DB_DMC_ERR_INFO,
@@ -3108,10 +3100,15 @@ void db_cleanup_offheap_comp(DbTerm* obj)
 	    break;
 	case FUN_SUBTAG:
 	    ASSERT(u.pb != &tmp);
-	    if (erts_refc_dectest(&u.fun->fe->refc, 0) == 0) {
+	    if (erts_smp_refc_dectest(&u.fun->fe->refc, 0) == 0) {
 		erts_erase_fun_entry(u.fun->fe);
 	    }
 	    break;
+	case REF_SUBTAG:
+	    ASSERT(is_magic_ref_thing(u.hdr));
+	    if (erts_refc_dectest(&u.mref->mb->refc, 0) == 0)
+		erts_bin_free((Binary *)u.mref->mb);		    
+	    break;
 	default:
 	    ASSERT(is_external_header(u.hdr->thing_word));
 	    ASSERT(u.pb != &tmp);
@@ -3267,13 +3264,6 @@ int db_has_variable(Eterm node) {
     return 0;
 }
 
-int erts_db_is_compiled_ms(Eterm term)
-{
-    return (is_binary(term)
-	    && (thing_subtag(*binary_val(term)) == REFC_BINARY_SUBTAG)
-	    && IsMatchProgBinary((((ProcBin *) binary_val(term))->val)));
-}
-
 /* 
 ** Local (static) utilities.
 */
@@ -5290,24 +5280,35 @@ void db_match_dis(Binary *bp)
 	case matchEqRef:
 	    ++t;
 	    {
-		RefThing *rt = (RefThing *) t;
+		Uint32 *num;
 		int ri;
-		n = thing_arityval(rt->header);
-		erts_printf("EqRef\t(%d) {", (int) n);
+
+		if (is_ordinary_ref_thing(t)) {
+		    ErtsORefThing *rt = (ErtsORefThing *) t;
+		    num = rt->num;
+		    t += TermWords(ERTS_REF_THING_SIZE);
+		}
+		else {
+		    ErtsMRefThing *mrt = (ErtsMRefThing *) t;
+		    ASSERT(is_magic_ref_thing(t));
+		    num = mrt->mb->refn;
+		    t += TermWords(ERTS_MAGIC_REF_THING_SIZE);
+		}
+
+		erts_printf("EqRef\t(%d) {", (int) ERTS_REF_NUMBERS);
 		first = 1;
-		for (ri = 0; ri < n; ++ri) {
+		for (ri = 0; ri < ERTS_REF_NUMBERS; ++ri) {
 		    if (first)
 			first = 0;
 		    else
 			erts_printf(", ");
 #if defined(ARCH_64)
-		    erts_printf("0x%016bex", rt->data.ui[ri]);
+		    erts_printf("0x%016bex", num[ri]);
 #else
-		    erts_printf("0x%08bex", rt->data.ui[ri]);
+		    erts_printf("0x%08bex", num[ri]);
 #endif
 		}
 	    }
-	    t += TermWords(REF_THING_SIZE);
 	    erts_printf("}\n");
 	    break;
 	case matchEqBig:
diff --git a/erts/emulator/beam/erl_db_util.h b/erts/emulator/beam/erl_db_util.h
index 3e29fdcbcf..471fefe3cb 100644
--- a/erts/emulator/beam/erl_db_util.h
+++ b/erts/emulator/beam/erl_db_util.h
@@ -23,6 +23,7 @@
 
 #include "global.h"
 #include "erl_message.h"
+#include "erl_bif_unique.h"
 
 /*#define HARDDEBUG 1*/
 
@@ -211,7 +212,7 @@ typedef struct db_fixation {
  */
 
 typedef struct db_table_common {
-    erts_refc_t ref;          /* fixation counter */
+    erts_smp_refc_t ref;          /* fixation counter */
 #ifdef ERTS_SMP
     erts_smp_rwmtx_t rwlock;  /* rw lock on table */
     erts_smp_mtx_t fixlock;   /* Protects fixations,megasec,sec,microsec */
@@ -260,7 +261,7 @@ typedef struct db_table_common {
 				  (DB_BAG | DB_SET | DB_DUPLICATE_BAG)))
 #define IS_TREE_TABLE(Status) (!!((Status) & \
 				  DB_ORDERED_SET))
-#define NFIXED(T) (erts_refc_read(&(T)->common.ref,0))
+#define NFIXED(T) (erts_smp_refc_read(&(T)->common.ref,0))
 #define IS_FIXED(T) (NFIXED(T) != 0) 
 
 /*
@@ -456,10 +457,44 @@ Eterm db_format_dmc_err_info(Process *p, DMCErrInfo *ei);
 void db_free_dmc_err_info(DMCErrInfo *ei);
 /* Completely free's an error info structure, including all recorded 
    errors */
-Eterm db_make_mp_binary(Process *p, Binary *mp, Eterm **hpp);
-/* Convert a match program to a erlang "magic" binary to be returned to userspace,
-   increments the reference counter. */
-int erts_db_is_compiled_ms(Eterm term);
+
+ERTS_GLB_INLINE Eterm erts_db_make_match_prog_ref(Process *p, Binary *mp, Eterm **hpp);
+ERTS_GLB_INLINE Binary *erts_db_get_match_prog_binary(Eterm term);
+ERTS_GLB_INLINE Binary *erts_db_get_match_prog_binary_unchecked(Eterm term);
+
+#if ERTS_GLB_INLINE_INCL_FUNC_DEF
+
+/*
+ * Convert a match program to a "magic" ref to return up to erlang
+ */
+ERTS_GLB_INLINE Eterm erts_db_make_match_prog_ref(Process *p, Binary *mp, Eterm **hpp)
+{
+    return erts_mk_magic_ref(hpp, &MSO(p), mp);
+}
+
+ERTS_GLB_INLINE Binary *
+erts_db_get_match_prog_binary_unchecked(Eterm term)
+{
+    Binary *bp = erts_magic_ref2bin(term);
+    ASSERT(bp->flags & BIN_FLAG_MAGIC);
+    ASSERT((ERTS_MAGIC_BIN_DESTRUCTOR(bp) == erts_db_match_prog_destructor));
+    return bp;
+}
+
+ERTS_GLB_INLINE Binary *
+erts_db_get_match_prog_binary(Eterm term)
+{
+    Binary *bp;
+    if (!is_internal_magic_ref(term))
+	return NULL;
+    bp = erts_magic_ref2bin(term);
+    ASSERT(bp->flags & BIN_FLAG_MAGIC);
+    if (ERTS_MAGIC_BIN_DESTRUCTOR(bp) != erts_db_match_prog_destructor)
+	return NULL;
+    return bp;
+}
+
+#endif
 
 /*
 ** Convenience when compiling into Binary structures
diff --git a/erts/emulator/beam/erl_debug.c b/erts/emulator/beam/erl_debug.c
index 3526bb684d..517dad9cbb 100644
--- a/erts/emulator/beam/erl_debug.c
+++ b/erts/emulator/beam/erl_debug.c
@@ -130,7 +130,7 @@ pdisplay1(fmtfn_t to, void *to_arg, Process* p, Eterm obj)
 	Uint32 *ref_num;
 	erts_print(to, to_arg, "#Ref<%lu", ref_channel_no(obj));
 	ref_num = ref_numbers(obj);
-	for (i = ref_no_of_numbers(obj)-1; i >= 0; i--)
+	for (i = ref_no_numbers(obj)-1; i >= 0; i--)
 	    erts_print(to, to_arg, ",%lu", ref_num[i]);
 	erts_print(to, to_arg, ">");
 	break;
diff --git a/erts/emulator/beam/erl_fun.c b/erts/emulator/beam/erl_fun.c
index 41097203d4..cedc0517e1 100644
--- a/erts/emulator/beam/erl_fun.c
+++ b/erts/emulator/beam/erl_fun.c
@@ -113,9 +113,9 @@ erts_put_fun_entry(Eterm mod, int uniq, int index)
     fe = (ErlFunEntry *) hash_put(&erts_fun_table, (void*) &template);
     sys_memset(fe->uniq, 0, sizeof(fe->uniq));
     fe->index = 0;
-    refc = erts_refc_inctest(&fe->refc, 0);
+    refc = erts_smp_refc_inctest(&fe->refc, 0);
     if (refc < 2) /* New or pending delete */
-	erts_refc_inc(&fe->refc, 1);
+	erts_smp_refc_inc(&fe->refc, 1);
     erts_fun_write_unlock();
     return fe;
 }
@@ -137,9 +137,9 @@ erts_put_fun_entry2(Eterm mod, int old_uniq, int old_index,
     sys_memcpy(fe->uniq, uniq, sizeof(fe->uniq));
     fe->index = index;
     fe->arity = arity;
-    refc = erts_refc_inctest(&fe->refc, 0);
+    refc = erts_smp_refc_inctest(&fe->refc, 0);
     if (refc < 2) /* New or pending delete */
-	erts_refc_inc(&fe->refc, 1);
+	erts_smp_refc_inc(&fe->refc, 1);
     erts_fun_write_unlock();
     return fe;
 }
@@ -164,9 +164,9 @@ erts_get_fun_entry(Eterm mod, int uniq, int index)
     erts_fun_read_lock();
     ret = (ErlFunEntry *) hash_get(&erts_fun_table, (void*) &template);
     if (ret) {
-	erts_aint_t refc = erts_refc_inctest(&ret->refc, 1);
+	erts_aint_t refc = erts_smp_refc_inctest(&ret->refc, 1);
 	if (refc < 2) /* Pending delete */
-	    erts_refc_inc(&ret->refc, 1);
+	    erts_smp_refc_inc(&ret->refc, 1);
     }
     erts_fun_read_unlock();
     return ret;
@@ -187,7 +187,7 @@ erts_erase_fun_entry(ErlFunEntry* fe)
      * We have to check refc again since someone might have looked up
      * the fun entry and incremented refc after last check.
      */
-    if (erts_refc_dectest(&fe->refc, -1) <= 0)
+    if (erts_smp_refc_dectest(&fe->refc, -1) <= 0)
 #endif
     {
 	if (fe->address != unloaded_fun)
@@ -274,7 +274,7 @@ erts_fun_purge_complete(ErlFunEntry **funs, Uint no)
 #ifdef HIPE
         fe->pend_purge_native_address = NULL;
 #endif
-	if (erts_refc_dectest(&fe->refc, 0) == 0)
+	if (erts_smp_refc_dectest(&fe->refc, 0) == 0)
 	    erts_erase_fun_entry(fe);
     }
     ERTS_SMP_WRITE_MEMORY_BARRIER;
@@ -306,7 +306,7 @@ erts_dump_fun_entries(fmtfn_t to, void *to_arg)
 #ifdef HIPE
 	    erts_print(to, to_arg, "Native_address: %p\n", fe->native_address);
 #endif
-	    erts_print(to, to_arg, "Refc: %ld\n", erts_refc_read(&fe->refc, 1));
+	    erts_print(to, to_arg, "Refc: %ld\n", erts_smp_refc_read(&fe->refc, 1));
 	    b = b->next;
 	}
     }
@@ -337,7 +337,7 @@ fun_alloc(ErlFunEntry* template)
     obj->old_uniq = template->old_uniq;
     obj->old_index = template->old_index;
     obj->module = template->module;
-    erts_refc_init(&obj->refc, -1);
+    erts_smp_refc_init(&obj->refc, -1);
     obj->address = unloaded_fun;
     obj->pend_purge_address = NULL;
 #ifdef HIPE
diff --git a/erts/emulator/beam/erl_fun.h b/erts/emulator/beam/erl_fun.h
index 83899f24a6..e3073eb874 100644
--- a/erts/emulator/beam/erl_fun.h
+++ b/erts/emulator/beam/erl_fun.h
@@ -42,7 +42,7 @@ typedef struct erl_fun_entry {
 
     Uint arity;			/* The arity of the fun. */
     Eterm module;		/* Tagged atom for module. */
-    erts_refc_t refc;		/* Reference count: One for code + one for each
+    erts_smp_refc_t refc;		/* Reference count: One for code + one for each
 				   fun object in each process. */
     BeamInstr *pend_purge_address; /* address stored during a pending purge */
 #ifdef HIPE
diff --git a/erts/emulator/beam/erl_gc.c b/erts/emulator/beam/erl_gc.c
index 6815d76776..8cc7fc0142 100644
--- a/erts/emulator/beam/erl_gc.c
+++ b/erts/emulator/beam/erl_gc.c
@@ -178,7 +178,7 @@ Uint erts_test_long_gc_sleep; /* Only used for testing... */
 typedef struct {
     Process *proc;
     Eterm ref;
-    Eterm ref_heap[REF_THING_SIZE];
+    Eterm ref_heap[ERTS_REF_THING_SIZE];
     Uint req_sched;
     erts_smp_atomic32_t refc;
 } ErtsGCInfoReq;
@@ -2354,6 +2354,9 @@ copy_one_frag(Eterm** hpp, ErlOffHeap* off_heap,
 	    switch (val & _HEADER_SUBTAG_MASK) {
 	    case ARITYVAL_SUBTAG:
 		break;
+	    case REF_SUBTAG:
+		if (is_ordinary_ref_thing(fhp - 1))
+		    goto the_default;
 	    case REFC_BINARY_SUBTAG:
 	    case FUN_SUBTAG:
 	    case EXTERNAL_PID_SUBTAG:
@@ -2363,6 +2366,7 @@ copy_one_frag(Eterm** hpp, ErlOffHeap* off_heap,
 		cpy_sz = thing_arityval(val);
 		goto cpy_words;
 	    default:
+	    the_default:
 		cpy_sz = header_arity(val);
 
 	    cpy_words:
@@ -2796,6 +2800,16 @@ link_live_proc_bin(struct shrink_cand_data *shrink,
     *prevppp = &pbp->next;
 }
 
+#ifdef ERTS_MAGIC_REF_THING_HEADER
+/*
+ * ERTS_MAGIC_REF_THING_HEADER only defined when there
+ * is a size difference between magic and ordinary references...
+ */
+# define ERTS_USED_MAGIC_REF_THING_HEADER__ ERTS_MAGIC_REF_THING_HEADER
+#else
+# define ERTS_USED_MAGIC_REF_THING_HEADER__ ERTS_REF_THING_HEADER
+#endif
+
 
 static void
 sweep_off_heap(Process *p, int fullsweep)
@@ -2820,7 +2834,7 @@ sweep_off_heap(Process *p, int fullsweep)
     prev = &MSO(p).first;
     ptr = MSO(p).first;
 
-    /* Firts part of the list will reside on the (old) new-heap.
+    /* First part of the list will reside on the (old) new-heap.
      * Keep if moved, otherwise deref.
      */
     while (ptr) {
@@ -2828,7 +2842,8 @@ sweep_off_heap(Process *p, int fullsweep)
 	    ASSERT(!ErtsInArea(ptr, oheap, oheap_sz));
 	    *prev = ptr = (struct erl_off_heap_header*) boxed_val(ptr->thing_word);
 	    ASSERT(!IS_MOVED_BOXED(ptr->thing_word));
-	    if (ptr->thing_word == HEADER_PROC_BIN) {
+	    switch (ptr->thing_word) {
+	    case HEADER_PROC_BIN: {
 		int to_new_heap = !ErtsInArea(ptr, oheap, oheap_sz);
 		ASSERT(to_new_heap == !seen_mature || (!to_new_heap && (seen_mature=1)));
 		if (to_new_heap) {
@@ -2837,13 +2852,28 @@ sweep_off_heap(Process *p, int fullsweep)
 		    BIN_OLD_VHEAP(p) += ptr->size / sizeof(Eterm); /* for binary gc (words)*/
 		}		
 		link_live_proc_bin(&shrink, &prev, &ptr, to_new_heap);
-	    }
-	    else {
+                break;
+            }
+            case ERTS_USED_MAGIC_REF_THING_HEADER__: {
+                Uint size;
+		int to_new_heap = !ErtsInArea(ptr, oheap, oheap_sz);
+                ASSERT(is_magic_ref_thing(ptr));
+		ASSERT(to_new_heap == !seen_mature || (!to_new_heap && (seen_mature=1)));
+                size = (Uint) ((ErtsMRefThing *) ptr)->mb->orig_size;
+		if (to_new_heap)
+		    bin_vheap += size / sizeof(Eterm);
+                else
+		    BIN_OLD_VHEAP(p) += size / sizeof(Eterm); /* for binary gc (words)*/
+                /* fall through... */
+            }
+            default:
 		prev = &ptr->next;
 		ptr = ptr->next;
 	    }
 	}
-	else if (!ErtsInArea(ptr, oheap, oheap_sz)) {
+	else if (ErtsInArea(ptr, oheap, oheap_sz))
+            break; /* and let old-heap loop continue */
+        else {
 	    /* garbage */
 	    switch (thing_subtag(ptr->thing_word)) {
 	    case REFC_BINARY_SUBTAG:
@@ -2857,18 +2887,26 @@ sweep_off_heap(Process *p, int fullsweep)
 	    case FUN_SUBTAG:
 		{
 		    ErlFunEntry* fe = ((ErlFunThing*)ptr)->fe;
-		    if (erts_refc_dectest(&fe->refc, 0) == 0) {
+		    if (erts_smp_refc_dectest(&fe->refc, 0) == 0) {
 			erts_erase_fun_entry(fe);
 		    }
 		    break;
 		}
+	    case REF_SUBTAG:
+		{
+		    ErtsMagicBinary *bptr;
+		    ASSERT(is_magic_ref_thing(ptr));
+		    bptr = ((ErtsMRefThing *) ptr)->mb;
+		    if (erts_refc_dectest(&bptr->refc, 0) == 0)
+			erts_bin_free((Binary *) bptr);		    
+		    break;
+		}
 	    default:
 		ASSERT(is_external_header(ptr->thing_word));
 		erts_deref_node_entry(((ExternalThing*)ptr)->node);
 	    }
 	    *prev = ptr = ptr->next;
 	}
-	else break; /* and let old-heap loop continue */
     }
 
     /* The rest of the list resides on old-heap, and we just did a
@@ -2877,15 +2915,24 @@ sweep_off_heap(Process *p, int fullsweep)
     while (ptr) {
 	ASSERT(ErtsInArea(ptr, oheap, oheap_sz));
 	ASSERT(!IS_MOVED_BOXED(ptr->thing_word));       
-	if (ptr->thing_word == HEADER_PROC_BIN) {
+        switch (ptr->thing_word) {
+        case HEADER_PROC_BIN:
 	    BIN_OLD_VHEAP(p) += ptr->size / sizeof(Eterm); /* for binary gc (words)*/
 	    link_live_proc_bin(&shrink, &prev, &ptr, 0);
-	}
-	else {
-	    ASSERT(is_fun_header(ptr->thing_word) ||
-		   is_external_header(ptr->thing_word));
-	    prev = &ptr->next;
-	    ptr = ptr->next;
+            break;
+        case ERTS_USED_MAGIC_REF_THING_HEADER__:
+            ASSERT(is_magic_ref_thing(ptr));
+            BIN_OLD_VHEAP(p) +=
+                (((Uint) ((ErtsMRefThing *) ptr)->mb->orig_size)
+                 / sizeof(Eterm)); /* for binary gc (words)*/
+            /* fall through... */
+        default:
+            ASSERT(is_fun_header(ptr->thing_word) ||
+                   is_external_header(ptr->thing_word)
+                   || is_magic_ref_thing(ptr));
+            prev = &ptr->next;
+            ptr = ptr->next;
+            break;
 	}
     }
 
@@ -2978,6 +3025,9 @@ offset_heap(Eterm* hp, Uint sz, Sint offs, char* area, Uint area_size)
 	      }
 	      tari = thing_arityval(val);
 	      switch (thing_subtag(val)) {
+	      case REF_SUBTAG:
+		  if (is_ordinary_ref_thing(hp))
+		      break;
 	      case REFC_BINARY_SUBTAG:
 	      case FUN_SUBTAG:
 	      case EXTERNAL_PID_SUBTAG:
@@ -3175,7 +3225,7 @@ reply_gc_info(void *vgcirp)
 	if (hpp)
 	    ref_copy = STORE_NC(hpp, ohp, gcirp->ref);
 	else
-	    *szp += REF_THING_SIZE;
+	    *szp += ERTS_REF_THING_SIZE;
 
 	msg = erts_bld_tuple(hpp, szp, 3,
 			     make_small(esdp->no),
@@ -3439,8 +3489,8 @@ erts_max_heap_size(Eterm arg, Uint *max_heap_size, Uint *max_heap_flags)
 
 #if defined(DEBUG) || defined(ERTS_OFFHEAP_DEBUG)
 
-static int
-within2(Eterm *ptr, Process *p, Eterm *real_htop)
+int
+erts_dbg_within_proc(Eterm *ptr, Process *p, Eterm *real_htop)
 {
     ErlHeapFragment* bp;
     ErtsMessage* mp;
@@ -3486,12 +3536,6 @@ within2(Eterm *ptr, Process *p, Eterm *real_htop)
     return 0;
 }
 
-int
-within(Eterm *ptr, Process *p)
-{
-    return within2(ptr, p, NULL);
-}
-
 #endif
 
 #ifdef ERTS_OFFHEAP_DEBUG
@@ -3525,12 +3569,16 @@ erts_check_off_heap2(Process *p, Eterm *htop)
 	    refc = erts_refc_read(&u.pb->val->refc, 1);		
 	    break;
 	case FUN_SUBTAG:
-	    refc = erts_refc_read(&u.fun->fe->refc, 1);
+	    refc = erts_smp_refc_read(&u.fun->fe->refc, 1);
 	    break;
 	case EXTERNAL_PID_SUBTAG:
 	case EXTERNAL_PORT_SUBTAG:
 	case EXTERNAL_REF_SUBTAG:
-	    refc = erts_refc_read(&u.ext->node->refc, 1);
+	    refc = erts_smp_refc_read(&u.ext->node->refc, 1);
+	    break;
+	case REF_SUBTAG:
+	    ASSERT(is_magic_ref_thing(u.hdr));
+	    refc = erts_refc_read(&u.mref->mb->refc, 1);
 	    break;
 	default:
 	    ASSERT(!"erts_check_off_heap2: Invalid thing_word");
@@ -3546,7 +3594,7 @@ erts_check_off_heap2(Process *p, Eterm *htop)
 	else if (oheap <= u.ep && u.ep < ohtop)
 	    old = 1;
 	else {
-	    ERTS_CHK_OFFHEAP_ASSERT(within2(u.ep, p, htop));
+	    ERTS_CHK_OFFHEAP_ASSERT(erts_dbg_within_proc(u.ep, p, htop));
 	}
     }
 
diff --git a/erts/emulator/beam/erl_gc.h b/erts/emulator/beam/erl_gc.h
index f4cbe732ce..1cce426d21 100644
--- a/erts/emulator/beam/erl_gc.h
+++ b/erts/emulator/beam/erl_gc.h
@@ -67,10 +67,6 @@ do {                                                                    \
     while (nelts--) *HTOP++ = *PTR++;                                   \
 } while(0)
 
-#if defined(DEBUG) || defined(ERTS_OFFHEAP_DEBUG)
-int within(Eterm *ptr, Process *p);
-#endif
-
 #define ErtsInYoungGen(TPtr, Ptr, OldHeap, OldHeapSz)			\
     (!erts_is_literal((TPtr), (Ptr))					\
      & !ErtsInArea((Ptr), (OldHeap), (OldHeapSz)))
@@ -157,5 +153,8 @@ void erts_free_heap_frags(struct process* p);
 Eterm erts_max_heap_size_map(Sint, Uint, Eterm **, Uint *);
 int erts_max_heap_size(Eterm, Uint *, Uint *);
 void erts_deallocate_young_generation(Process *c_p);
+#if defined(DEBUG) || defined(ERTS_OFFHEAP_DEBUG)
+int erts_dbg_within_proc(Eterm *ptr, Process *p, Eterm* real_htop);
+#endif
 
 #endif /* __ERL_GC_H__ */
diff --git a/erts/emulator/beam/erl_hl_timer.c b/erts/emulator/beam/erl_hl_timer.c
index 647fa26811..26be8c7edf 100644
--- a/erts/emulator/beam/erl_hl_timer.c
+++ b/erts/emulator/beam/erl_hl_timer.c
@@ -1771,7 +1771,7 @@ setup_bif_timer(Process *c_p, ErtsMonotonicTime timeout_pos,
 
     esdp = erts_proc_sched_data(c_p);
 
-    hp = HAlloc(c_p, REF_THING_SIZE);
+    hp = HAlloc(c_p, ERTS_REF_THING_SIZE);
     ref = erts_sched_make_ref_in_buffer(esdp, hp);
 
     ASSERT(erts_get_ref_numbers_thr_id(
@@ -1939,7 +1939,7 @@ access_sched_local_btm(Process *c_p, Eterm pid,
 	Eterm *hp_end;
 #endif
 
-	hsz = REF_THING_SIZE;
+	hsz = ERTS_REF_THING_SIZE;
 	if (async) {
 	    refn = trefn; /* timer ref */
 	    hsz += 4; /* 3-tuple */
@@ -1973,7 +1973,7 @@ access_sched_local_btm(Process *c_p, Eterm pid,
 			refn[1],
 			refn[2]);
 	ref = make_internal_ref(hp);
-	hp += REF_THING_SIZE;
+	hp += ERTS_REF_THING_SIZE;
 
 	msg = (async
 	       ? TUPLE3(hp, (cancel
@@ -2087,7 +2087,7 @@ try_access_sched_remote_btm(ErtsSchedulerData *esdp,
 	ErtsProcLocks proc_locks = ERTS_PROC_LOCK_MAIN;
 	ErlOffHeap *ohp;
 
-	hsz = 4 + REF_THING_SIZE;
+	hsz = 4 + ERTS_REF_THING_SIZE;
 	if (time_left > (Sint64) MAX_SMALL)
 	    hsz += ERTS_SINT64_HEAP_SIZE(time_left);
 
@@ -2103,7 +2103,7 @@ try_access_sched_remote_btm(ErtsSchedulerData *esdp,
 			trefn[1],
 			trefn[2]);
 	tref = make_internal_ref(hp);
-	hp += REF_THING_SIZE;
+	hp += ERTS_REF_THING_SIZE;
 
 	if (time_left < 0)
 	    res = am_false;
@@ -2189,7 +2189,7 @@ access_bif_timer(Process *c_p, Eterm tref, int cancel, int async, int info)
 	    Eterm *hp, rref;
 	    Uint32 *rrefn;
 
-	    hp = HAlloc(c_p, REF_THING_SIZE);
+	    hp = HAlloc(c_p, ERTS_REF_THING_SIZE);
 	    rref = erts_sched_make_ref_in_buffer(esdp, hp);
 	    rrefn = internal_ref_numbers(rref);
 
diff --git a/erts/emulator/beam/erl_init.c b/erts/emulator/beam/erl_init.c
index c5904b375e..b2c307e826 100644
--- a/erts/emulator/beam/erl_init.c
+++ b/erts/emulator/beam/erl_init.c
@@ -111,12 +111,19 @@ const int etp_lock_check = 1;
 #else
 const int etp_lock_check = 0;
 #endif
-#ifdef WORDS_BIGENDIAN
-const int etp_big_endian = 1;
+const int etp_endianness = ERTS_ENDIANNESS;
+const Eterm etp_ref_header = ERTS_REF_THING_HEADER;
+#ifdef ERTS_MAGIC_REF_THING_HEADER
+const Eterm etp_magic_ref_header = ERTS_MAGIC_REF_THING_HEADER;
 #else
-const int etp_big_endian = 0;
+const Eterm etp_magic_ref_header = ERTS_REF_THING_HEADER;
 #endif
 const Eterm etp_the_non_value = THE_NON_VALUE;
+#ifdef ERTS_HOLE_MARKER
+const Eterm etp_hole_marker = ERTS_HOLE_MARKER;
+#else
+const Eterm etp_hole_marker = 0;
+#endif
 
 /*
  * Note about VxWorks: All variables must be initialized by executable code,
@@ -767,6 +774,8 @@ early_init(int *argc, char **argv) /*
     H_MAX_SIZE = H_DEFAULT_MAX_SIZE;
     H_MAX_FLAGS = MAX_HEAP_SIZE_KILL|MAX_HEAP_SIZE_LOG;
 
+    erts_term_init();
+
     erts_initialized = 0;
 
     erts_use_sender_punish = 1;
@@ -1138,6 +1147,8 @@ early_init(int *argc, char **argv) /*
     no_schedulers_online = schdlrs_onln;
 
     erts_no_schedulers = (Uint) no_schedulers;
+#else
+    erts_no_schedulers = 1;
 #endif
 #ifdef ERTS_DIRTY_SCHEDULERS
     erts_no_dirty_cpu_schedulers = no_dirty_cpu_schedulers = dirty_cpu_scheds;
@@ -2220,7 +2231,6 @@ erl_start(int argc, char **argv)
 	init_break_handler();
     if (replace_intr)
 	erts_replace_intr();
-    sys_init_suspend_handler();
 #endif
 
     boot_argc = argc - i;  /* Number of arguments to init */
diff --git a/erts/emulator/beam/erl_lock_check.c b/erts/emulator/beam/erl_lock_check.c
index 0798faf53a..6ff9aea5ab 100644
--- a/erts/emulator/beam/erl_lock_check.c
+++ b/erts/emulator/beam/erl_lock_check.c
@@ -159,6 +159,7 @@ static erts_lc_lock_order_t erts_lock_order[] = {
     {	"tracer_mtx", 				NULL			},
     {   "port_table",                           NULL                    },
 #endif
+    {	"magic_ref_table",			"address"		},
     {	"mtrace_op",				NULL			},
     {	"instr_x",				NULL			},
     {	"instr",				NULL			},
diff --git a/erts/emulator/beam/erl_lock_count.h b/erts/emulator/beam/erl_lock_count.h
index 4f838f7faa..71cd73ee27 100644
--- a/erts/emulator/beam/erl_lock_count.h
+++ b/erts/emulator/beam/erl_lock_count.h
@@ -129,7 +129,7 @@ typedef struct {
 
 typedef struct erts_lcnt_lock_stats_s {
     /* "tries" and "colls" needs to be atomic since
-     * trylock busy does not aquire a lock and there
+     * trylock busy does not acquire a lock and there
      * is no post action to rectify the situation
      */
 
diff --git a/erts/emulator/beam/erl_map.c b/erts/emulator/beam/erl_map.c
index 3015158b24..9062e44b10 100644
--- a/erts/emulator/beam/erl_map.c
+++ b/erts/emulator/beam/erl_map.c
@@ -1198,7 +1198,7 @@ static int hashmap_merge_ctx_destructor(Binary* ctx_bin)
 }
 
 BIF_RETTYPE maps_merge_trap_1(BIF_ALIST_1) {
-    Binary* ctx_bin = ((ProcBin *) binary_val(BIF_ARG_1))->val;
+    Binary* ctx_bin = erts_magic_ref2bin(BIF_ARG_1);
 
     ASSERT(ERTS_MAGIC_BIN_DESTRUCTOR(ctx_bin) == hashmap_merge_ctx_destructor);
 
@@ -1454,9 +1454,9 @@ trap:  /* Yield */
                                                  hashmap_merge_ctx_destructor);
         ctx = ERTS_MAGIC_BIN_DATA(ctx_b);
         sys_memcpy(ctx, &local_ctx, sizeof(HashmapMergeContext));
-        hp = HAlloc(p, PROC_BIN_SIZE);
+        hp = HAlloc(p, ERTS_MAGIC_REF_THING_SIZE);
         ASSERT(ctx->trap_bin == THE_NON_VALUE);
-        ctx->trap_bin = erts_mk_magic_binary_term(&hp, &MSO(p), ctx_b);
+        ctx->trap_bin = erts_mk_magic_ref(&hp, &MSO(p), ctx_b);
 
         erts_set_gc_state(p, 0);
     }
diff --git a/erts/emulator/beam/erl_message.c b/erts/emulator/beam/erl_message.c
index 792b69bb37..f181c1e3cb 100644
--- a/erts/emulator/beam/erl_message.c
+++ b/erts/emulator/beam/erl_message.c
@@ -172,10 +172,15 @@ erts_cleanup_offheap(ErlOffHeap *offheap)
 	    }
 	    break;
 	case FUN_SUBTAG:
-	    if (erts_refc_dectest(&u.fun->fe->refc, 0) == 0) {
+	    if (erts_smp_refc_dectest(&u.fun->fe->refc, 0) == 0) {
 		erts_erase_fun_entry(u.fun->fe);		    
 	    }
 	    break;
+	case REF_SUBTAG:
+	    ASSERT(is_magic_ref_thing(u.hdr));
+	    if (erts_refc_dectest(&u.mref->mb->refc, 0) == 0)
+		erts_bin_free((Binary *)u.mref->mb);		    
+	    break;
 	default:
 	    ASSERT(is_external_header(u.hdr->thing_word));
 	    erts_deref_node_entry(u.ext->node);
@@ -285,9 +290,11 @@ erts_queue_dist_message(Process *rcvr,
     if (!(rcvr_locks & ERTS_PROC_LOCK_MSGQ)) {
 	if (erts_smp_proc_trylock(rcvr, ERTS_PROC_LOCK_MSGQ) == EBUSY) {
 	    ErtsProcLocks need_locks = ERTS_PROC_LOCK_MSGQ;
-	    if (rcvr_locks & ERTS_PROC_LOCK_STATUS) {
-		erts_smp_proc_unlock(rcvr, ERTS_PROC_LOCK_STATUS);
-		need_locks |= ERTS_PROC_LOCK_STATUS;
+            ErtsProcLocks unlocks =
+                rcvr_locks & ERTS_PROC_LOCKS_HIGHER_THAN(ERTS_PROC_LOCK_MSGQ);
+	    if (unlocks) {
+		erts_smp_proc_unlock(rcvr, unlocks);
+		need_locks |= unlocks;
 	    }
 	    erts_smp_proc_lock(rcvr, need_locks);
 	}
@@ -406,7 +413,7 @@ queue_messages(Process* receiver,
 	    if (state & (ERTS_PSFLG_EXITING|ERTS_PSFLG_PENDING_EXIT))
 		goto exiting;
 
-            need_locks = receiver_locks & (ERTS_PROC_LOCK_STATUS|ERTS_PROC_LOCK_TRACE);
+            need_locks = receiver_locks & ERTS_PROC_LOCKS_HIGHER_THAN(ERTS_PROC_LOCK_MSGQ);
 	    if (need_locks) {
 		erts_smp_proc_unlock(receiver, need_locks);
 	    }
diff --git a/erts/emulator/beam/erl_monitors.c b/erts/emulator/beam/erl_monitors.c
index e5be66be79..6dee1d5ef3 100644
--- a/erts/emulator/beam/erl_monitors.c
+++ b/erts/emulator/beam/erl_monitors.c
@@ -24,7 +24,7 @@
  * key in the monitor case and the pid of the linked process as key in the 
  * link case. Lookups the order of the references is somewhat special. Local 
  * references are strictly smaller than remote references and are sorted 
- * by inlined comparision functionality. Remote references are handled by the
+ * by inlined comparison functionality. Remote references are handled by the
  * usual cmp function.
  * Each Monitor is tagged with different tags depending on which end of the 
  * monitor it is.
@@ -45,6 +45,7 @@
 #include "bif.h"
 #include "big.h"
 #include "erl_monitors.h"
+#include "erl_bif_unique.h"
 
 #define STACK_NEED 50
 #define MAX_MONITORS 0xFFFFFFFFUL
@@ -79,7 +80,24 @@ static ERTS_INLINE int cmp_mon_ref(Eterm ref1, Eterm ref2)
     b2 = boxed_val(ref2);
     if (is_ref_thing_header(*b1)) {
 	if (is_ref_thing_header(*b2)) {
-	    return memcmp(b1+1,b2+1,ERTS_REF_WORDS*sizeof(Uint));
+	    Uint32 *num1, *num2;
+	    if (is_ordinary_ref_thing(b1)) {
+		ErtsORefThing *rtp = (ErtsORefThing *) b1;
+		num1 = rtp->num;
+	    }
+	    else {
+		ErtsMRefThing *mrtp = (ErtsMRefThing *) b1;
+		num1 = mrtp->mb->refn;
+	    }
+	    if (is_ordinary_ref_thing(b2)) {
+		ErtsORefThing *rtp = (ErtsORefThing *) b2;
+		num2 = rtp->num;
+	    }
+	    else {
+		ErtsMRefThing *mrtp = (ErtsMRefThing *) b2;
+		num2 = mrtp->mb->refn;
+	    }
+	    return erts_internal_ref_number_cmp(num1, num2);
 	}
 	return -1;
     }
@@ -97,14 +115,15 @@ do {								\
 	Uint i__;						\
 	Uint len__;						\
 	ASSERT((Hp));						\
-	ASSERT(is_internal_ref((From)) || is_external((From)));	\
+	ASSERT(is_internal_ordinary_ref((From))                 \
+               || is_external((From)));                         \
 	(To) = make_boxed((Hp));				\
 	len__ = thing_arityval(*boxed_val((From))) + 1;		\
 	for(i__ = 0; i__ < len__; i__++)			\
 	    (*((Hp)++)) = boxed_val((From))[i__];		\
 	if (is_external((To))) {				\
 	    external_thing_ptr((To))->next = NULL;		\
-	    erts_refc_inc(&(external_thing_ptr((To))->node->refc), 2);\
+	    erts_smp_refc_inc(&(external_thing_ptr((To))->node->refc), 2);\
 	}							\
     }								\
 } while (0)
@@ -135,7 +154,7 @@ static ErtsMonitor *create_monitor(Uint type, Eterm ref, UWord entity, Eterm nam
      n->type = (Uint16) type;
      n->balance = 0;            /* Always the same initial value */
      n->name = name; /* atom() or [] */
-     CP_LINK_VAL(n->ref, hp, ref); /*XXX Unneccesary check, never immediate*/
+     CP_LINK_VAL(n->ref, hp, ref); /*XXX Unnecessary check, never immediate*/
      if (type == MON_NIF_TARGET)
          n->u.resource = (ErtsResource*)entity;
      else
@@ -342,6 +361,8 @@ void erts_add_monitor(ErtsMonitor **root, Uint type, Eterm ref, UWord entity,
     int state = 0;
     ErtsMonitor **this = root;
     Sint c;
+
+    ASSERT(is_internal_ordinary_ref(ref) || is_external_ref(ref));
   
     dstack[0] = DIR_END;
     for (;;) {
diff --git a/erts/emulator/beam/erl_monitors.h b/erts/emulator/beam/erl_monitors.h
index f90ac930e5..f659829e6c 100644
--- a/erts/emulator/beam/erl_monitors.h
+++ b/erts/emulator/beam/erl_monitors.h
@@ -92,7 +92,7 @@
 
 /* Size of a monitor without heap, in words (fixalloc) */
 #define ERTS_MONITOR_SIZE ((sizeof(ErtsMonitor) - sizeof(Uint))/sizeof(Uint))
-#define ERTS_MONITOR_SH_SIZE (ERTS_MONITOR_SIZE + REF_THING_SIZE)
+#define ERTS_MONITOR_SH_SIZE (ERTS_MONITOR_SIZE + ERTS_REF_THING_SIZE)
 #define ERTS_LINK_SIZE ((sizeof(ErtsLink) - sizeof(Uint))/sizeof(Uint))
 #define ERTS_LINK_SH_SIZE ERTS_LINK_SIZE /* Size of fix-alloced links */
 
diff --git a/erts/emulator/beam/erl_msacc.c b/erts/emulator/beam/erl_msacc.c
index 66bb55e6c8..1c3160efaf 100644
--- a/erts/emulator/beam/erl_msacc.c
+++ b/erts/emulator/beam/erl_msacc.c
@@ -66,9 +66,9 @@ static Uint msacc_unmanaged_count = 0;
 #endif
 
 #if ERTS_MSACC_STATE_COUNT < MAP_SMALL_MAP_LIMIT
-#define DEFAULT_MSACC_MSG_SIZE (3 + 1 + ERTS_MSACC_STATE_COUNT * 2 + 3 + REF_THING_SIZE)
+#define DEFAULT_MSACC_MSG_SIZE (3 + 1 + ERTS_MSACC_STATE_COUNT * 2 + 3 + ERTS_REF_THING_SIZE)
 #else
-#define DEFAULT_MSACC_MSG_SIZE (3 + ERTS_MSACC_STATE_COUNT * 3 + 3 + REF_THING_SIZE)
+#define DEFAULT_MSACC_MSG_SIZE (3 + ERTS_MSACC_STATE_COUNT * 3 + 3 + ERTS_REF_THING_SIZE)
 #endif
 
 /* we have to split initiation as atoms are not inited in early init */
@@ -212,7 +212,7 @@ typedef struct {
     int action;
     Process *proc;
     Eterm ref;
-    Eterm ref_heap[REF_THING_SIZE];
+    Eterm ref_heap[ERTS_REF_THING_SIZE];
     Uint req_sched;
     erts_smp_atomic32_t refc;
 } ErtsMSAccReq;
@@ -245,7 +245,7 @@ static void send_reply(ErtsMsAcc *msacc, ErtsMSAccReq *msaccrp) {
 
         if (msacc->unmanaged) erts_mtx_lock(&msacc->mtx);
 
-        hp = erts_produce_heap(&factory, REF_THING_SIZE + 3 /* tuple */, 0);
+        hp = erts_produce_heap(&factory, ERTS_REF_THING_SIZE + 3 /* tuple */, 0);
         ref_copy = STORE_NC(&hp, &msgp->hfrag.off_heap, msaccrp->ref);
         msg = erts_msacc_gather_stats(msacc, &factory);
         msg = TUPLE2(hp, ref_copy, msg);
@@ -255,7 +255,7 @@ static void send_reply(ErtsMsAcc *msacc, ErtsMSAccReq *msaccrp) {
         erts_factory_close(&factory);
     } else {
         ErlOffHeap *ohp = NULL;
-        msgp = erts_alloc_message_heap(rp, &rp_locks, REF_THING_SIZE, &hp, &ohp);
+        msgp = erts_alloc_message_heap(rp, &rp_locks, ERTS_REF_THING_SIZE, &hp, &ohp);
         msg = STORE_NC(&hp, &msgp->hfrag.off_heap, msaccrp->ref);
     }
 
diff --git a/erts/emulator/beam/erl_nif.c b/erts/emulator/beam/erl_nif.c
index cd6fbf276b..7e7b78ec14 100644
--- a/erts/emulator/beam/erl_nif.c
+++ b/erts/emulator/beam/erl_nif.c
@@ -95,6 +95,14 @@ static void add_readonly_check(ErlNifEnv*, unsigned char* ptr, unsigned sz);
 #  define ADD_READONLY_CHECK(ENV,PTR,SIZE) ((void)0)
 #endif
 
+#ifdef ERTS_NIF_ASSERT_IN_ENV
+#  define ASSERT_IN_ENV(ENV, TERM, NR, TYPE) dbg_assert_in_env(ENV, TERM, NR, TYPE, __func__)
+static void dbg_assert_in_env(ErlNifEnv*, Eterm term, int nr, const char* type, const char* func);
+#  include "erl_gc.h"
+#else
+#  define ASSERT_IN_ENV(ENV, TERM, NR, TYPE)
+#endif
+
 #ifdef DEBUG
 static int is_offheap(const ErlOffHeap* off_heap);
 #endif
@@ -202,6 +210,9 @@ void erts_pre_nif(ErlNifEnv* env, Process* p, struct erl_module_nif* mod_nif,
 
     ASSERT(p->common.id != ERTS_INVALID_PID);
 
+#ifdef ERTS_NIF_ASSERT_IN_ENV
+    env->dbg_disable_assert_in_env = 0;
+#endif
 #if defined(DEBUG) && defined(ERTS_DIRTY_SCHEDULERS)
     {
 	ErtsSchedulerData *esdp = erts_get_scheduler_data();
@@ -487,11 +498,15 @@ setup_nif_env(struct enif_msg_environment_t* msg_env,
     HEAP_END(&msg_env->phony_proc) = phony_heap;
     MBUF(&msg_env->phony_proc) = NULL;
     msg_env->phony_proc.common.id = ERTS_INVALID_PID;
+    msg_env->env.tracee = tracee;
+
 #ifdef FORCE_HEAP_FRAGS
     msg_env->phony_proc.space_verified = 0;
     msg_env->phony_proc.space_verified_from = NULL;
 #endif
-    msg_env->env.tracee = tracee;
+#ifdef ERTS_NIF_ASSERT_IN_ENV
+    msg_env->env.dbg_disable_assert_in_env = 0;
+#endif
 }
 
 ErlNifEnv* enif_alloc_env(void)
@@ -708,7 +723,7 @@ int enif_send(ErlNifEnv* env, const ErlNifPid* to_pid,
         Uint sz;
         INITIALIZE_LITERAL_PURGE_AREA(litarea);
         sz = size_object_litopt(msg, &litarea);
-	if (env && !env->tracee) {
+	if (c_p && !env->tracee) {
 	    full_flush_env(env);
 	    mp = erts_alloc_message_heap(rp, &rp_locks, sz, &hp, &ohp);
 	    full_cache_env(env);
@@ -1592,6 +1607,9 @@ int enif_make_existing_atom_len(ErlNifEnv* env, const char* name, size_t len,
 
 ERL_NIF_TERM enif_make_tuple(ErlNifEnv* env, unsigned cnt, ...)
 {
+#ifdef ERTS_NIF_ASSERT_IN_ENV
+    int nr = 0;
+#endif
     Eterm* hp = alloc_heap(env,cnt+1);
     Eterm ret = make_tuple(hp);
     va_list ap;
@@ -1599,7 +1617,9 @@ ERL_NIF_TERM enif_make_tuple(ErlNifEnv* env, unsigned cnt, ...)
     *hp++ = make_arityval(cnt);
     va_start(ap,cnt);
     while (cnt--) {
-	*hp++ = va_arg(ap,Eterm);	   
+        Eterm elem = va_arg(ap,Eterm);
+        ASSERT_IN_ENV(env, elem, ++nr, "tuple");
+	*hp++ = elem;
     }
     va_end(ap);
     return ret;
@@ -1607,12 +1627,16 @@ ERL_NIF_TERM enif_make_tuple(ErlNifEnv* env, unsigned cnt, ...)
 
 ERL_NIF_TERM enif_make_tuple_from_array(ErlNifEnv* env, const ERL_NIF_TERM arr[], unsigned cnt)
 {
+#ifdef ERTS_NIF_ASSERT_IN_ENV
+    int nr = 0;
+#endif
     Eterm* hp = alloc_heap(env,cnt+1);
     Eterm ret = make_tuple(hp);
     const Eterm* src = arr;
 
     *hp++ = make_arityval(cnt);
     while (cnt--) {
+        ASSERT_IN_ENV(env, *src, ++nr, "tuple");
 	*hp++ = *src++;	   
     }
     return ret;
@@ -1623,6 +1647,8 @@ ERL_NIF_TERM enif_make_list_cell(ErlNifEnv* env, Eterm car, Eterm cdr)
     Eterm* hp = alloc_heap(env,2);
     Eterm ret = make_list(hp);
 
+    ASSERT_IN_ENV(env, car, 0, "head of list cell");
+    ASSERT_IN_ENV(env, cdr, 0, "tail of list cell");
     CAR(hp) = car;
     CDR(hp) = cdr;
     return ret;
@@ -1634,6 +1660,9 @@ ERL_NIF_TERM enif_make_list(ErlNifEnv* env, unsigned cnt, ...)
 	return NIL;
     }
     else {
+#ifdef ERTS_NIF_ASSERT_IN_ENV
+        int nr = 0;
+#endif
 	Eterm* hp = alloc_heap(env,cnt*2);
 	Eterm ret = make_list(hp);
 	Eterm* last = &ret;
@@ -1641,8 +1670,10 @@ ERL_NIF_TERM enif_make_list(ErlNifEnv* env, unsigned cnt, ...)
 
 	va_start(ap,cnt);
 	while (cnt--) {
+            Eterm term = va_arg(ap,Eterm);
 	    *last = make_list(hp);
-	    *hp = va_arg(ap,Eterm);
+            ASSERT_IN_ENV(env, term, ++nr, "list");
+	    *hp = term;
 	    last = ++hp;
 	    ++hp;
 	}
@@ -1654,14 +1685,19 @@ ERL_NIF_TERM enif_make_list(ErlNifEnv* env, unsigned cnt, ...)
 
 ERL_NIF_TERM enif_make_list_from_array(ErlNifEnv* env, const ERL_NIF_TERM arr[], unsigned cnt)
 {
+#ifdef ERTS_NIF_ASSERT_IN_ENV
+    int nr = 0;
+#endif
     Eterm* hp = alloc_heap(env,cnt*2);
     Eterm ret = make_list(hp);
     Eterm* last = &ret;
     const Eterm* src = arr;
 
     while (cnt--) {
+        Eterm term = *src++;
 	*last = make_list(hp);
-	*hp = *src++;
+        ASSERT_IN_ENV(env, term, ++nr, "list");
+	*hp = term;
 	last = ++hp;
 	++hp;
     }
@@ -1685,7 +1721,7 @@ ERL_NIF_TERM enif_make_string_len(ErlNifEnv* env, const char* string,
 
 ERL_NIF_TERM enif_make_ref(ErlNifEnv* env)
 {
-    Eterm* hp = alloc_heap(env, REF_THING_SIZE);
+    Eterm* hp = alloc_heap(env, ERTS_REF_THING_SIZE);
     return erts_make_ref_in_buffer(hp);
 }
 
@@ -1694,13 +1730,9 @@ void enif_system_info(ErlNifSysInfo *sip, size_t si_size)
     driver_system_info(sip, si_size);
 }
 
-int enif_make_reverse_list(ErlNifEnv* env, ERL_NIF_TERM term, ERL_NIF_TERM *list) {
-    Eterm *listptr, ret = NIL, *hp;
-
-    if (is_nil(term)) {
-	*list = term;
-        return 1;
-    }
+int enif_make_reverse_list(ErlNifEnv* env, ERL_NIF_TERM term, ERL_NIF_TERM *list)
+{
+    Eterm *listptr, ret, *hp;
 
     ret = NIL;
 
@@ -2349,6 +2381,7 @@ void* enif_alloc_resource(ErlNifResourceType* type, size_t data_sz)
         magic_sz += data_sz;
     }
     bin = erts_create_magic_binary_x(magic_sz, NIF_RESOURCE_DTOR,
+                                     ERTS_ALC_T_BINARY,
                                      1); /* unaligned */
     resource = ERTS_MAGIC_BIN_UNALIGNED_DATA(bin);
 
@@ -2404,42 +2437,63 @@ Eterm erts_bld_resource_ref(Eterm** hpp, ErlOffHeap* oh, ErtsResource* resource)
 {
     ErtsBinary* bin = ERTS_MAGIC_BIN_FROM_UNALIGNED_DATA(resource);
     ASSERT(!resource->dbg_is_dying);
-    return erts_mk_magic_binary_term(hpp, oh, &bin->binary);
+    return erts_mk_magic_ref(hpp, oh, &bin->binary);
 }
 
 ERL_NIF_TERM enif_make_resource(ErlNifEnv* env, void* obj)
 {
     ErtsResource* resource = DATA_TO_RESOURCE(obj);
     ErtsBinary* bin = ERTS_MAGIC_BIN_FROM_UNALIGNED_DATA(resource);
-    Eterm* hp = alloc_heap(env,PROC_BIN_SIZE);
-    ASSERT(!resource->dbg_is_dying);
-    return erts_mk_magic_binary_term(&hp, &MSO(env->proc), &bin->binary);
+    Eterm* hp = alloc_heap(env, ERTS_MAGIC_REF_THING_SIZE);
+    return erts_mk_magic_ref(&hp, &MSO(env->proc), &bin->binary);
 }
 
 ERL_NIF_TERM enif_make_resource_binary(ErlNifEnv* env, void* obj,
 				       const void* data, size_t size)
 {
-    Eterm bin = enif_make_resource(env, obj);
-    ProcBin* pb = (ProcBin*) binary_val(bin);
-    pb->bytes = (byte*) data;
+    ErtsResource* resource = DATA_TO_RESOURCE(obj);
+    ErtsBinary* bin = ERTS_MAGIC_BIN_FROM_UNALIGNED_DATA(resource);
+    ErlOffHeap *ohp = &MSO(env->proc);
+    Eterm* hp = alloc_heap(env,PROC_BIN_SIZE);
+    ProcBin* pb = (ProcBin *) hp;
+
+    pb->thing_word = HEADER_PROC_BIN;
     pb->size = size;
-    return bin;
+    pb->next = ohp->first;
+    ohp->first = (struct erl_off_heap_header*) pb;
+    pb->val = &bin->binary;
+    pb->bytes = (byte*) data;
+    pb->flags = 0;
+
+    OH_OVERHEAD(ohp, size / sizeof(Eterm));
+    erts_refc_inc(&bin->binary.refc, 1);
+
+    return make_binary(hp);
 }
 
 int enif_get_resource(ErlNifEnv* env, ERL_NIF_TERM term, ErlNifResourceType* type,
 		      void** objp)
 {
-    ProcBin* pb;
     Binary* mbin;
     ErtsResource* resource;
-    if (!ERTS_TERM_IS_MAGIC_BINARY(term)) {
-	return 0;
+    if (is_internal_magic_ref(term))
+	mbin = erts_magic_ref2bin(term);
+    else {
+        Eterm *hp;
+        if (!is_binary(term))
+            return 0;
+        hp = binary_val(term);
+        if (thing_subtag(*hp) != REFC_BINARY_SUBTAG)
+            return 0;
+        /*
+        if (((ProcBin *) hp)->size != 0) {	
+            return 0; / * Or should we allow "resource binaries" as handles? * /
+        }
+        */
+        mbin = ((ProcBin *) hp)->val;
+        if (!(mbin->flags & BIN_FLAG_MAGIC))
+            return 0;
     }
-    pb = (ProcBin*) binary_val(term);
-    /*if (pb->size != 0) {	
-	return 0; / * Or should we allow "resource binaries" as handles? * /
-    }*/
-    mbin = pb->val;
     resource = (ErtsResource*) ERTS_MAGIC_BIN_UNALIGNED_DATA(mbin);
     if (ERTS_MAGIC_BIN_DESTRUCTOR(mbin) != NIF_RESOURCE_DTOR
 	|| resource->type != type) {	
@@ -2746,10 +2800,13 @@ enif_schedule_nif(ErlNifEnv* env, const char* fun_name, int flags,
     if (flags == 0)
 	result = schedule(env, execute_nif, fp, proc->current->module,
 			  fun_name_atom, argc, argv);
+    else if (!(flags & ~(ERL_NIF_DIRTY_JOB_IO_BOUND|ERL_NIF_DIRTY_JOB_CPU_BOUND))) {
 #ifdef ERTS_DIRTY_SCHEDULERS
-    else if (!(flags & ~(ERL_NIF_DIRTY_JOB_IO_BOUND|ERL_NIF_DIRTY_JOB_CPU_BOUND)))
 	result = schedule_dirty_nif(env, flags, fp, fun_name_atom, argc, argv);
+#else
+        result = enif_raise_exception(env, am_notsup);
 #endif
+    }
     else
 	result = enif_make_badarg(env);
 
@@ -2824,6 +2881,10 @@ int enif_make_map_put(ErlNifEnv* env,
     if (!is_map(map_in)) {
 	return 0;
     }
+    ASSERT_IN_ENV(env, map_in, 0, "old map");
+    ASSERT_IN_ENV(env, key, 0, "key");
+    ASSERT_IN_ENV(env, value, 0, "value");
+
     flush_env(env);
     *map_out = erts_maps_put(env->proc, key, value, map_in);
     cache_env(env);
@@ -2858,6 +2919,10 @@ int enif_make_map_update(ErlNifEnv* env,
 	return 0;
     }
 
+    ASSERT_IN_ENV(env, map_in, 0, "old map");
+    ASSERT_IN_ENV(env, key, 0, "key");
+    ASSERT_IN_ENV(env, value, 0, "value");
+
     flush_env(env);
     res = erts_maps_update(env->proc, key, value, map_in, map_out);
     cache_env(env);
@@ -3065,7 +3130,7 @@ int enif_monitor_process(ErlNifEnv* env, void* obj, const ErlNifPid* target_pid,
     int scheduler;
     ErtsResource* rsrc = DATA_TO_RESOURCE(obj);
     Process *rp;
-    Eterm tmp[REF_THING_SIZE];
+    Eterm tmp[ERTS_REF_THING_SIZE];
     Eterm ref;
     int retval;
 
@@ -3132,7 +3197,7 @@ int enif_demonitor_process(ErlNifEnv* env, void* obj, const ErlNifMonitor* monit
     Process *rp;
     ErtsMonitor *mon;
     ErtsMonitor *rmon = NULL;
-    Eterm ref_heap[REF_THING_SIZE];
+    Eterm ref_heap[ERTS_REF_THING_SIZE];
     Eterm ref;
     int is_exiting;
 
@@ -3141,7 +3206,7 @@ int enif_demonitor_process(ErlNifEnv* env, void* obj, const ErlNifMonitor* monit
 
     execution_state(env, NULL, &scheduler);
 
-    memcpy(ref_heap, monitor, sizeof(Eterm)*REF_THING_SIZE);
+    memcpy(ref_heap, monitor, sizeof(Eterm)*ERTS_REF_THING_SIZE);
     ref = make_internal_ref(ref_heap);
     erts_smp_mtx_lock(&rsrc->monitors->lock);
     mon = erts_remove_monitor(&rsrc->monitors->root, ref);
@@ -3205,7 +3270,8 @@ int enif_demonitor_process(ErlNifEnv* env, void* obj, const ErlNifMonitor* monit
 int enif_compare_monitors(const ErlNifMonitor *monitor1,
                           const ErlNifMonitor *monitor2)
 {
-    return memcmp(monitor1,monitor2,sizeof(ErlNifMonitor));
+    return sys_memcmp((void *) monitor1, (void *) monitor2,
+                      ERTS_REF_THING_SIZE*sizeof(Eterm));
 }
 
 /***************************************************************************
@@ -3733,6 +3799,9 @@ Eterm erts_nif_call_function(Process *p, Process *tracee,
         clear_offheap(&MSO(p));
 
         erts_pre_nif(&env, p, mod, tracee);
+#ifdef ERTS_NIF_ASSERT_IN_ENV
+        env.dbg_disable_assert_in_env = 1;
+#endif
         nif_result = (*fun->fptr)(&env, argc, argv);
         if (env.exception_thrown)
             nif_result = THE_NON_VALUE;
@@ -3755,6 +3824,9 @@ Eterm erts_nif_call_function(Process *p, Process *tracee,
            so we create a phony one. */
         struct enif_msg_environment_t msg_env;
         pre_nif_noproc(&msg_env, mod, tracee);
+#ifdef ERTS_NIF_ASSERT_IN_ENV
+        msg_env.env.dbg_disable_assert_in_env = 1;
+#endif
         nif_result = (*fun->fptr)(&msg_env.env, argc, argv);
         if (msg_env.env.exception_thrown)
             nif_result = THE_NON_VALUE;
@@ -3819,6 +3891,55 @@ static unsigned calc_checksum(unsigned char* ptr, unsigned size)
 
 #endif /* READONLY_CHECK */
 
+#ifdef ERTS_NIF_ASSERT_IN_ENV
+static void dbg_assert_in_env(ErlNifEnv* env, Eterm term,
+                              int nr, const char* type, const char* func)
+{
+    Uint saved_used_size;
+    Eterm* real_htop;
+
+    if (is_immed(term)
+        || (is_non_value(term) && env->exception_thrown)
+        || erts_is_literal(term, ptr_val(term)))
+        return;
+
+    if (env->dbg_disable_assert_in_env) {
+        /*
+         * Trace nifs may cheat as built terms are discarded after return.
+         * ToDo: Check if 'term' is part of argv[].
+         */
+        return;
+    }
+
+    if (env->heap_frag) {
+        ASSERT(env->heap_frag == MBUF(env->proc));
+        ASSERT(env->hp >= env->heap_frag->mem);
+        ASSERT(env->hp <= env->heap_frag->mem + env->heap_frag->alloc_size);
+        saved_used_size = env->heap_frag->used_size;
+        env->heap_frag->used_size = env->hp - env->heap_frag->mem;
+        real_htop = NULL;
+    }
+    else {
+        real_htop = env->hp;
+    }
+    if (!erts_dbg_within_proc(ptr_val(term), env->proc, real_htop)) {
+        fprintf(stderr, "\r\nFAILED ASSERTION in %s:\r\n", func);
+        if (nr) {
+            fprintf(stderr, "Term #%d of the %s is not from same ErlNifEnv.",
+                    nr, type);
+        }
+        else {
+            fprintf(stderr, "The %s is not from the same ErlNifEnv.", type);
+        }
+        fprintf(stderr, "\r\nABORTING\r\n");
+        abort();
+    }
+    if (env->heap_frag) {
+        env->heap_frag->used_size = saved_used_size;
+    }
+}
+#endif
+
 #ifdef HAVE_USE_DTRACE
 
 #define MESSAGE_BUFSIZ 1024
diff --git a/erts/emulator/beam/erl_node_container_utils.h b/erts/emulator/beam/erl_node_container_utils.h
index 0c76c6fe7d..e27f1d121a 100644
--- a/erts/emulator/beam/erl_node_container_utils.h
+++ b/erts/emulator/beam/erl_node_container_utils.h
@@ -255,19 +255,16 @@ extern ErtsPTab erts_port;
  * Refs                                                                    *
 \*                                                                         */
 
+#define internal_ref_no_numbers(x)	ERTS_REF_NUMBERS
+#define internal_ref_numbers(x)		(is_internal_ordinary_ref((x)) \
+					 ? internal_ordinary_ref_numbers((x)) \
+					 : (ASSERT(is_internal_magic_ref((x))), \
+					    internal_magic_ref_numbers((x))))
 #if defined(ARCH_64)
 
-#define internal_ref_no_of_numbers(x)					\
-  (internal_ref_data((x))[0])
-#define internal_thing_ref_no_of_numbers(thing)			        \
-  (internal_thing_ref_data(thing)[0])
-#define internal_ref_numbers(x)						\
-  (&internal_ref_data((x))[1])
-#define internal_thing_ref_numbers(thing)				\
-  (&internal_thing_ref_data(thing)[1])
-#define external_ref_no_of_numbers(x)					\
+#define external_ref_no_numbers(x)					\
   (external_ref_data((x))[0])
-#define external_thing_ref_no_of_numbers(thing)			        \
+#define external_thing_ref_no_numbers(thing)			        \
   (external_thing_ref_data(thing)[0])
 #define external_ref_numbers(x)						\
   (&external_ref_data((x))[1])
@@ -277,12 +274,8 @@ extern ErtsPTab erts_port;
 
 #else
 
-#define internal_ref_no_of_numbers(x)	(internal_ref_data_words((x)))
-#define internal_thing_ref_no_of_numbers(t) (internal_thing_ref_data_words(t))
-#define internal_ref_numbers(x)		(internal_ref_data((x)))
-#define internal_thing_ref_numbers(t)   (internal_thing_ref_data(t))
-#define external_ref_no_of_numbers(x)	(external_ref_data_words((x)))
-#define external_thing_ref_no_of_numbers(t) (external_thing_ref_data_words((t)))
+#define external_ref_no_numbers(x)	(external_ref_data_words((x)))
+#define external_thing_ref_no_numbers(t) (external_thing_ref_data_words((t)))
 #define external_ref_numbers(x)		(external_ref_data((x)))
 #define external_thing_ref_numbers(t)   (external_thing_ref_data((t)))
 
@@ -299,15 +292,9 @@ extern ErtsPTab erts_port;
 #define internal_ref_channel_no(x)	(internal_channel_no((x)))
 #define external_ref_channel_no(x)	(external_channel_no((x)))
 
-#define ref_data_words(x)		(is_internal_ref((x))		\
-					 ? internal_ref_data_words((x))	\
-					 : external_ref_data_words((x)))
-#define ref_data(x)			(is_internal_ref((x))		\
-					 ? internal_ref_data((x))	\
-					 : external_ref_data((x)))
-#define ref_no_of_numbers(x)		(is_internal_ref((x))		\
-					 ? internal_ref_no_of_numbers((x))\
-					 : external_ref_no_of_numbers((x)))
+#define ref_no_numbers(x)		(is_internal_ref((x))		\
+					 ? internal_ref_no_numbers((x))\
+					 : external_ref_no_numbers((x)))
 #define ref_numbers(x)			(is_internal_ref((x))		\
 					 ? internal_ref_numbers((x))	\
 					 : external_ref_numbers((x)))
diff --git a/erts/emulator/beam/erl_node_tables.c b/erts/emulator/beam/erl_node_tables.c
index f463f7fdf4..e2072fe30f 100644
--- a/erts/emulator/beam/erl_node_tables.c
+++ b/erts/emulator/beam/erl_node_tables.c
@@ -98,7 +98,7 @@ dist_table_alloc(void *dep_tmpl)
     dist_entries++;
 
     dep->prev				= NULL;
-    erts_refc_init(&dep->refc, -1);
+    erts_smp_refc_init(&dep->refc, -1);
     erts_smp_rwmtx_init_opt_x(&dep->rwmtx, &rwmtx_opt, "dist_entry", chnl_nr);
     dep->sysname			= sysname;
     dep->cid				= NIL;
@@ -208,7 +208,7 @@ erts_channel_no_to_dist_entry(Uint cno)
  * to the node name is used as channel no.
  */
     if(cno == ERST_INTERNAL_CHANNEL_NO) {
-	erts_refc_inc(&erts_this_dist_entry->refc, 2);
+	erts_smp_refc_inc(&erts_this_dist_entry->refc, 2);
 	return erts_this_dist_entry;
     }
 
@@ -231,16 +231,16 @@ erts_sysname_to_connected_dist_entry(Eterm sysname)
     de.sysname = sysname;
   
     if(erts_this_dist_entry->sysname == sysname) {
-	erts_refc_inc(&erts_this_dist_entry->refc, 2);
+	erts_smp_refc_inc(&erts_this_dist_entry->refc, 2);
 	return erts_this_dist_entry;
     }
 
     erts_smp_rwmtx_rlock(&erts_dist_table_rwmtx);
     res_dep = (DistEntry *) hash_get(&erts_dist_table, (void *) &de);
     if (res_dep) {
-	erts_aint_t refc = erts_refc_inctest(&res_dep->refc, 1);
+	erts_aint_t refc = erts_smp_refc_inctest(&res_dep->refc, 1);
 	if (refc < 2) /* Pending delete */
-	    erts_refc_inc(&res_dep->refc, 1);
+	    erts_smp_refc_inc(&res_dep->refc, 1);
     }
     erts_smp_rwmtx_runlock(&erts_dist_table_rwmtx);
     if (res_dep) {
@@ -267,9 +267,9 @@ DistEntry *erts_find_or_insert_dist_entry(Eterm sysname)
     de.sysname = sysname;
     erts_smp_rwmtx_rwlock(&erts_dist_table_rwmtx);
     res = hash_put(&erts_dist_table, (void *) &de);
-    refc = erts_refc_inctest(&res->refc, 0);
+    refc = erts_smp_refc_inctest(&res->refc, 0);
     if (refc < 2) /* New or pending delete */
-	erts_refc_inc(&res->refc, 1);
+	erts_smp_refc_inc(&res->refc, 1);
     erts_smp_rwmtx_rwunlock(&erts_dist_table_rwmtx);
     return res;
 }
@@ -282,9 +282,9 @@ DistEntry *erts_find_dist_entry(Eterm sysname)
     erts_smp_rwmtx_rlock(&erts_dist_table_rwmtx);
     res = hash_get(&erts_dist_table, (void *) &de);
     if (res) {
-	erts_aint_t refc = erts_refc_inctest(&res->refc, 1);
+	erts_aint_t refc = erts_smp_refc_inctest(&res->refc, 1);
 	if (refc < 2) /* Pending delete */
-	    erts_refc_inc(&res->refc, 1);
+	    erts_smp_refc_inc(&res->refc, 1);
     }
     erts_smp_rwmtx_runlock(&erts_dist_table_rwmtx);
     return res;
@@ -311,7 +311,7 @@ static void try_delete_dist_entry(void *vdep)
      *
      * If refc > 0, the entry is in use. Keep the entry.
      */
-    refc = erts_refc_dectest(&dep->refc, -1);
+    refc = erts_smp_refc_dectest(&dep->refc, -1);
     if (refc == -1)
 	(void) hash_erase(&erts_dist_table, (void *) dep);
     erts_smp_rwmtx_rwunlock(&erts_dist_table_rwmtx);
@@ -518,7 +518,7 @@ node_table_alloc(void *venp_tmpl)
 
     node_entries++;
 
-    erts_refc_init(&enp->refc, -1);
+    erts_smp_refc_init(&enp->refc, -1);
     enp->creation = ((ErlNode *) venp_tmpl)->creation;
     enp->sysname = ((ErlNode *) venp_tmpl)->sysname;
     enp->dist_entry = erts_find_or_insert_dist_entry(((ErlNode *) venp_tmpl)->sysname);
@@ -585,9 +585,9 @@ ErlNode *erts_find_or_insert_node(Eterm sysname, Uint32 creation)
     erts_smp_rwmtx_rlock(&erts_node_table_rwmtx);
     res = hash_get(&erts_node_table, (void *) &ne);
     if (res && res != erts_this_node) {
-	erts_aint_t refc = erts_refc_inctest(&res->refc, 0);
+	erts_aint_t refc = erts_smp_refc_inctest(&res->refc, 0);
 	if (refc < 2) /* New or pending delete */
-	    erts_refc_inc(&res->refc, 1);
+	    erts_smp_refc_inc(&res->refc, 1);
     }
     erts_smp_rwmtx_runlock(&erts_node_table_rwmtx);
     if (res)
@@ -597,9 +597,9 @@ ErlNode *erts_find_or_insert_node(Eterm sysname, Uint32 creation)
     res = hash_put(&erts_node_table, (void *) &ne);
     ASSERT(res);
     if (res != erts_this_node) {
-	erts_aint_t refc = erts_refc_inctest(&res->refc, 0);
+	erts_aint_t refc = erts_smp_refc_inctest(&res->refc, 0);
 	if (refc < 2) /* New or pending delete */
-	    erts_refc_inc(&res->refc, 1);
+	    erts_smp_refc_inc(&res->refc, 1);
     }
     erts_smp_rwmtx_rwunlock(&erts_node_table_rwmtx);
     return res;
@@ -626,7 +626,7 @@ static void try_delete_node(void *venp)
      *
      * If refc > 0, the entry is in use. Keep the entry.
      */
-    refc = erts_refc_dectest(&enp->refc, -1);
+    refc = erts_smp_refc_dectest(&enp->refc, -1);
     if (refc == -1)
 	(void) hash_erase(&erts_node_table, (void *) enp);
     erts_smp_rwmtx_rwunlock(&erts_node_table_rwmtx);
@@ -672,7 +672,7 @@ static void print_node(void *venp, void *vpndp)
 	erts_print(pndp->to, pndp->to_arg, " %d", enp->creation);
 #ifdef DEBUG
 	erts_print(pndp->to, pndp->to_arg, " (refc=%ld)",
-		   erts_refc_read(&enp->refc, 0));
+		   erts_smp_refc_read(&enp->refc, 0));
 #endif
 	pndp->no_sysname++;
     }
@@ -715,19 +715,19 @@ void
 erts_set_this_node(Eterm sysname, Uint creation)
 {
     ERTS_SMP_LC_ASSERT(erts_thr_progress_is_blocking());
-    ASSERT(erts_refc_read(&erts_this_dist_entry->refc, 2));
+    ASSERT(erts_smp_refc_read(&erts_this_dist_entry->refc, 2));
 
-    if (erts_refc_dectest(&erts_this_node->refc, 0) == 0)
+    if (erts_smp_refc_dectest(&erts_this_node->refc, 0) == 0)
         try_delete_node(erts_this_node);
 
-    if (erts_refc_dectest(&erts_this_dist_entry->refc, 0) == 0)
+    if (erts_smp_refc_dectest(&erts_this_dist_entry->refc, 0) == 0)
         try_delete_dist_entry(erts_this_dist_entry);
 
     erts_this_node = NULL; /* to make sure refc is bumped for this node */
     erts_this_node = erts_find_or_insert_node(sysname, creation);
     erts_this_dist_entry = erts_this_node->dist_entry;
 
-    erts_refc_inc(&erts_this_dist_entry->refc, 2);
+    erts_smp_refc_inc(&erts_this_dist_entry->refc, 2);
 
     erts_this_node_sysname = erts_this_node_sysname_BUFFER;
     erts_snprintf(erts_this_node_sysname, sizeof(erts_this_node_sysname_BUFFER),
@@ -789,13 +789,13 @@ void erts_init_node_tables(int dd_sec)
     node_tmpl.creation = 0;
     erts_this_node = hash_put(&erts_node_table, &node_tmpl);
      /* +1 for erts_this_node */
-    erts_refc_init(&erts_this_node->refc, 1);
+    erts_smp_refc_init(&erts_this_node->refc, 1);
 
     ASSERT(erts_this_node->dist_entry != NULL);
     erts_this_dist_entry = erts_this_node->dist_entry;
     /* +1 for erts_this_dist_entry */
     /* +1 for erts_this_node->dist_entry */
-    erts_refc_init(&erts_this_dist_entry->refc, 2);
+    erts_smp_refc_init(&erts_this_dist_entry->refc, 2);
 
 
     erts_this_node_sysname = erts_this_node_sysname_BUFFER;
@@ -1126,12 +1126,12 @@ insert_offheap(ErlOffHeap *oh, int type, Eterm id)
 
     for (u.hdr = oh->first; u.hdr; u.hdr = u.hdr->next) {
 	switch (thing_subtag(u.hdr->thing_word)) {
-	case REFC_BINARY_SUBTAG:
-	    if(IsMatchProgBinary(u.pb->val)) {
+	case REF_SUBTAG:
+	    if(IsMatchProgBinary(u.mref->mb)) {
 		InsertedBin *ib;
 		int insert_bin = 1;
 		for (ib = inserted_bins; ib; ib = ib->next)
-		    if(ib->bin_val == u.pb->val) {
+		    if(ib->bin_val == (Binary *) u.mref->mb) {
 			insert_bin = 0;
 			break;
 		    }
@@ -1140,18 +1140,19 @@ insert_offheap(ErlOffHeap *oh, int type, Eterm id)
 		    Uint *hp = &id_heap[0];
 		    InsertedBin *nib;
 		    UseTmpHeapNoproc(BIG_UINT_HEAP_SIZE);
-		    a.id = erts_bld_uint(&hp, NULL, (Uint) u.pb->val);
-		    erts_match_prog_foreach_offheap(u.pb->val,
+		    a.id = erts_bld_uint(&hp, NULL, (Uint) u.mref->mb);
+		    erts_match_prog_foreach_offheap((Binary *) u.mref->mb,
 						    insert_offheap2,
 						    (void *) &a);
 		    nib = erts_alloc(ERTS_ALC_T_NC_TMP, sizeof(InsertedBin));
-		    nib->bin_val = u.pb->val;
+		    nib->bin_val = (Binary *) u.mref->mb;
 		    nib->next = inserted_bins;
 		    inserted_bins = nib;
 		    UnUseTmpHeapNoproc(BIG_UINT_HEAP_SIZE);
 		}
 	    }		
 	    break;
+	case REFC_BINARY_SUBTAG:
 	case FUN_SUBTAG:
 	    break; /* No need to */
 	default:
@@ -1623,7 +1624,7 @@ reference_table_term(Uint **hpp, Uint *szp)
 
 	tup = MK_2TUP(referred_nodes[i].node->sysname,
 		      MK_UINT(referred_nodes[i].node->creation));
-	tup = MK_3TUP(tup, MK_UINT(erts_refc_read(&referred_nodes[i].node->refc, 0)), nril);
+	tup = MK_3TUP(tup, MK_UINT(erts_smp_refc_read(&referred_nodes[i].node->refc, 0)), nril);
 	nl = MK_CONS(tup, nl);
     }
 
@@ -1684,7 +1685,7 @@ reference_table_term(Uint **hpp, Uint *szp)
 
 	/* DistList = [{Dist, Refc, ReferenceIdList}] */
 	tup = MK_3TUP(referred_dists[i].dist->sysname,
-		      MK_UINT(erts_refc_read(&referred_dists[i].dist->refc, 0)),
+		      MK_UINT(erts_smp_refc_read(&referred_dists[i].dist->refc, 0)),
 		      dril);
 	dl = MK_CONS(tup, dl);
     }
diff --git a/erts/emulator/beam/erl_node_tables.h b/erts/emulator/beam/erl_node_tables.h
index 47a6724c21..35051173d0 100644
--- a/erts/emulator/beam/erl_node_tables.h
+++ b/erts/emulator/beam/erl_node_tables.h
@@ -107,7 +107,7 @@ typedef struct dist_entry_ {
     HashBucket hash_bucket;     /* Hash bucket */
     struct dist_entry_ *next;	/* Next entry in dist_table (not sorted) */
     struct dist_entry_ *prev;	/* Previous entry in dist_table (not sorted) */
-    erts_refc_t refc;		/* Reference count */
+    erts_smp_refc_t refc;		/* Reference count */
 
     erts_smp_rwmtx_t rwmtx;     /* Protects all fields below until lck_mtx. */
     Eterm sysname;		/* name@host atom for efficiency */
@@ -149,7 +149,7 @@ typedef struct dist_entry_ {
 
 typedef struct erl_node_ {
   HashBucket hash_bucket;	/* Hash bucket */
-  erts_refc_t refc;		/* Reference count */
+  erts_smp_refc_t refc;		/* Reference count */
   Eterm	sysname;		/* name@host atom for efficiency */
   Uint32 creation;		/* Creation */
   DistEntry *dist_entry;	/* Corresponding dist entry */
@@ -210,7 +210,7 @@ ERTS_GLB_INLINE void
 erts_deref_dist_entry(DistEntry *dep)
 {
     ASSERT(dep);
-    if (erts_refc_dectest(&dep->refc, 0) == 0)
+    if (erts_smp_refc_dectest(&dep->refc, 0) == 0)
 	erts_schedule_delete_dist_entry(dep);
 }
 
@@ -218,7 +218,7 @@ ERTS_GLB_INLINE void
 erts_deref_node_entry(ErlNode *np)
 {
     ASSERT(np);
-    if (erts_refc_dectest(&np->refc, 0) == 0)
+    if (erts_smp_refc_dectest(&np->refc, 0) == 0)
 	erts_schedule_delete_node(np);
 }
 
diff --git a/erts/emulator/beam/erl_printf_term.c b/erts/emulator/beam/erl_printf_term.c
index b43a1b0190..6b64bbd2f1 100644
--- a/erts/emulator/beam/erl_printf_term.c
+++ b/erts/emulator/beam/erl_printf_term.c
@@ -382,12 +382,15 @@ print_term(fmtfn_t fn, void* arg, Eterm obj, long *dcount) {
 	    break;
 	}
 	case REF_DEF:
+            if (!ERTS_IS_CRASH_DUMPING)
+                erts_magic_ref_save_bin(obj);
+            /* fall through... */
 	case EXTERNAL_REF_DEF:
 	    PRINT_STRING(res, fn, arg, "#Ref<");
 	    PRINT_UWORD(res, fn, arg, 'u', 0, 1,
 			(ErlPfUWord) ref_channel_no(wobj));
 	    ref_num = ref_numbers(wobj);
-	    for (i = ref_no_of_numbers(wobj)-1; i >= 0; i--) {
+	    for (i = ref_no_numbers(wobj)-1; i >= 0; i--) {
 		PRINT_CHAR(res, fn, arg, '.');
 		PRINT_UWORD(res, fn, arg, 'u', 0, 1, (ErlPfUWord) ref_num[i]);
 	    }
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index 9db77585f2..52d9f9ddf7 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -127,11 +127,13 @@ runq_got_work_to_execute_flags(Uint32 flags)
     return !ERTS_IS_RUNQ_EMPTY_FLGS(flags);
 }
 
+#ifdef ERTS_SMP
 static ERTS_INLINE int
 runq_got_work_to_execute(ErtsRunQueue *rq)
 {
     return runq_got_work_to_execute_flags(ERTS_RUNQ_FLGS_GET_NOB(rq));
 }
+#endif
 
 #undef RUNQ_READ_RQ
 #undef RUNQ_SET_RQ
@@ -1171,7 +1173,7 @@ typedef struct {
     int enable;
     Process *proc;
     Eterm ref;
-    Eterm ref_heap[REF_THING_SIZE];
+    Eterm ref_heap[ERTS_REF_THING_SIZE];
     Uint req_sched;
     erts_smp_atomic32_t refc;
 #ifdef ERTS_DIRTY_SCHEDULERS
@@ -1183,7 +1185,7 @@ typedef struct {
 typedef struct {
     Process *proc;
     Eterm ref;
-    Eterm ref_heap[REF_THING_SIZE];
+    Eterm ref_heap[ERTS_REF_THING_SIZE];
     Uint req_sched;
     erts_smp_atomic32_t refc;
 } ErtsSystemCheckReq;
@@ -1295,7 +1297,7 @@ reply_sched_wall_time(void *vswtrp)
             if (hpp)
                 ref_copy = STORE_NC(hpp, ohp, swtrp->ref);
             else
-                *szp += REF_THING_SIZE;
+                *szp += ERTS_REF_THING_SIZE;
 
             ASSERT(!swtrp->set);
 
@@ -1340,7 +1342,7 @@ reply_sched_wall_time(void *vswtrp)
             if (hpp)
                 ref_copy = STORE_NC(hpp, ohp, swtrp->ref);
             else
-                *szp += REF_THING_SIZE;
+                *szp += ERTS_REF_THING_SIZE;
 
             if (swtrp->set)
                 msg = ref_copy;
@@ -1442,7 +1444,7 @@ reply_system_check(void *vscrp)
     ASSERT(!ERTS_SCHEDULER_IS_DIRTY(esdp));
 #endif
 
-    sz = REF_THING_SIZE;
+    sz = ERTS_REF_THING_SIZE;
     mp = erts_alloc_message_heap(rp, &rp_locks, sz, &hp, &ohp);
     hpp = &hp;
     msg = STORE_NC(hpp, ohp, scrp->ref);
@@ -5048,7 +5050,7 @@ check_balance(ErtsRunQueue *c_rq)
 	    sched_util_balancing = 1;
 	    /*
 	     * In order to avoid renaming a large amount of fields
-	     * we write utilization values instead of lenght values
+	     * we write utilization values instead of length values
 	     * in the 'max_len' and 'migration_limit' fields...
 	     */
 	    for (qix = 0; qix < blnc_no_rqs; qix++) {
@@ -6379,7 +6381,6 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online
 	erts_tsd_set(sched_data_key, (void *) esdp);
 #endif
     }
-    erts_no_schedulers = 1;
     erts_no_dirty_cpu_schedulers = 0;
     erts_no_dirty_io_schedulers = 0;
 #endif
@@ -10174,7 +10175,8 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
 
             esdp->current_process = NULL;
 #ifdef ERTS_SMP
-            p->scheduler_data = NULL;
+            if (is_normal_sched)
+                p->scheduler_data = NULL;
 #endif
 
             erts_smp_proc_unlock(p, (ERTS_PROC_LOCK_MAIN
@@ -10631,8 +10633,8 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
 
 	state = erts_smp_atomic32_read_nob(&p->state);
 
-	ASSERT(!p->scheduler_data);
 #ifndef ERTS_DIRTY_SCHEDULERS
+	ASSERT(!p->scheduler_data);
 	p->scheduler_data = esdp;
 #else /* ERTS_DIRTY_SCHEDULERS */
 	if (is_normal_sched) {
@@ -10643,6 +10645,7 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
 		erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS);
 		goto sched_out_proc;
 	    }
+	    ASSERT(!p->scheduler_data);
 	    p->scheduler_data = esdp;
 	}
 	else {
diff --git a/erts/emulator/beam/erl_process_dump.c b/erts/emulator/beam/erl_process_dump.c
index 7c23b5c76a..77c7c5e73c 100644
--- a/erts/emulator/beam/erl_process_dump.c
+++ b/erts/emulator/beam/erl_process_dump.c
@@ -447,7 +447,7 @@ heap_dump(fmtfn_t to, void *to_arg, Eterm x)
 			ProcBin* pb = (ProcBin *) binary_val(x);
 			Binary* val = pb->val;
 
-			if (erts_smp_atomic_xchg_nob(&val->refc, 0) != 0) {
+			if (erts_atomic_xchg_nob(&val->refc, 0) != 0) {
 			    val->flags = (UWord) all_binaries;
 			    all_binaries = val;
 			}
diff --git a/erts/emulator/beam/erl_process_lock.h b/erts/emulator/beam/erl_process_lock.h
index 773529384f..46a72fcb0c 100644
--- a/erts/emulator/beam/erl_process_lock.h
+++ b/erts/emulator/beam/erl_process_lock.h
@@ -219,6 +219,10 @@ typedef struct erts_proc_lock_t_ {
 #define ERTS_PROC_LOCKS_ALL_MINOR	(ERTS_PROC_LOCKS_ALL \
                                          & ~ERTS_PROC_LOCK_MAIN)
 
+/* All locks we first must unlock to lock L */
+#define ERTS_PROC_LOCKS_HIGHER_THAN(L) \
+  (ERTS_PROC_LOCKS_ALL & (~(L) & ~((L)-1)))
+
 
 #define ERTS_PIX_LOCKS_BITS		10
 #define ERTS_NO_OF_PIX_LOCKS		(1 << ERTS_PIX_LOCKS_BITS)
diff --git a/erts/emulator/beam/erl_ptab.c b/erts/emulator/beam/erl_ptab.c
index 578ae14b9f..abf2fe44f3 100644
--- a/erts/emulator/beam/erl_ptab.c
+++ b/erts/emulator/beam/erl_ptab.c
@@ -474,7 +474,7 @@ erts_ptab_init_table(ErtsPTab *ptab,
 	 * we don't want to shrink the size to ERTS_PTAB_MAX_SIZE/2.
 	 *
 	 * In order to fix this, we insert a pointer from the table
-	 * to the invalid_element, wich will be interpreted as a
+	 * to the invalid_element, which will be interpreted as a
 	 * slot currently being modified. This way we will be able to
 	 * have ERTS_PTAB_MAX_SIZE-1 valid elements in the table while
 	 * still having a table size of the power of 2.
@@ -771,18 +771,18 @@ erts_ptab_list(Process *c_p, ErtsPTab *ptab)
     }
     else {
 	Eterm *hp;
-	Eterm magic_bin;
+	Eterm magic_ref;
 	ERTS_PTAB_LIST_DBG_CHK_RESLIST(res_acc);
-	hp = HAlloc(c_p, PROC_BIN_SIZE);
-	ERTS_PTAB_LIST_DBG_SAVE_HEAP_ALLOC(ptlbdp, hp, PROC_BIN_SIZE);
-	magic_bin = erts_mk_magic_binary_term(&hp, &MSO(c_p), mbp);
+	hp = HAlloc(c_p, ERTS_MAGIC_REF_THING_SIZE);
+	ERTS_PTAB_LIST_DBG_SAVE_HEAP_ALLOC(ptlbdp, hp, ERTS_MAGIC_REF_THING_SIZE);
+	magic_ref = erts_mk_magic_ref(&hp, &MSO(c_p), mbp);
 	ERTS_PTAB_LIST_DBG_VERIFY_HEAP_ALLOC_USED(ptlbdp, hp);
 	ERTS_PTAB_LIST_DBG_TRACE(c_p->common.id, trap);
 	ERTS_BIF_PREP_YIELD2(ret_val,
 			     &ptab_list_continue_export,
 			     c_p,
 			     res_acc,
-			     magic_bin);
+			     magic_ref);
     }
     return ret_val;
 }
@@ -1289,9 +1289,7 @@ static BIF_RETTYPE ptab_list_continue(BIF_ALIST_2)
 
     res_acc = BIF_ARG_1;
 
-    ERTS_PTAB_LIST_ASSERT(ERTS_TERM_IS_MAGIC_BINARY(BIF_ARG_2));
-
-    mbp = ((ProcBin *) binary_val(BIF_ARG_2))->val;
+    mbp = erts_magic_ref2bin(BIF_ARG_2);
 
     ERTS_PTAB_LIST_ASSERT(ERTS_MAGIC_BIN_DESTRUCTOR(mbp)
 			  == cleanup_ptab_list_bif_data);
diff --git a/erts/emulator/beam/erl_term.c b/erts/emulator/beam/erl_term.c
index 7d857ad326..6b25728af7 100644
--- a/erts/emulator/beam/erl_term.c
+++ b/erts/emulator/beam/erl_term.c
@@ -59,6 +59,42 @@ erts_set_literal_tag(Eterm *term, Eterm *hp_start, Eterm hsz)
 #endif
 }
 
+void
+erts_term_init(void)
+{
+#ifdef ERTS_ORDINARY_REF_MARKER
+    /* Ordinary and magic references of same size... */
+
+    ErtsRefThing ref_thing;
+
+    ERTS_CT_ASSERT(ERTS_ORDINARY_REF_MARKER == ~((Uint32)0));
+    ref_thing.m.header = ERTS_REF_THING_HEADER;
+    ref_thing.m.mb = (ErtsMagicBinary *) ~((UWord) 3);
+    ref_thing.m.next = (struct erl_off_heap_header *) ~((UWord) 3);
+    if (ref_thing.o.marker == ERTS_ORDINARY_REF_MARKER)
+        ERTS_INTERNAL_ERROR("Cannot differentiate between magic and ordinary references");
+
+    ERTS_CT_ASSERT(offsetof(ErtsORefThing,marker) != 0);
+    ERTS_CT_ASSERT(sizeof(ErtsORefThing) == sizeof(ErtsMRefThing));
+#  ifdef ERTS_MAGIC_REF_THING_HEADER
+#    error Magic ref thing header should not have been defined...
+#  endif
+
+#else
+    /* Ordinary and magic references of different sizes... */
+
+#  ifndef ERTS_MAGIC_REF_THING_HEADER
+#    error Magic ref thing header should have been defined...
+#  endif
+    ERTS_CT_ASSERT(sizeof(ErtsORefThing) != sizeof(ErtsMRefThing));
+
+#endif
+
+    ERTS_CT_ASSERT(ERTS_REF_THING_SIZE*sizeof(Eterm) == sizeof(ErtsORefThing));
+    ERTS_CT_ASSERT(ERTS_MAGIC_REF_THING_SIZE*sizeof(Eterm) == sizeof(ErtsMRefThing));
+
+}
+
 /*
  * XXX: define NUMBER_CODE() here when new representation is used
  */
@@ -95,8 +131,8 @@ ET_DEFINE_CHECKED(Eterm*,tuple_val,Wterm,is_tuple);
 ET_DEFINE_CHECKED(struct erl_node_*,internal_pid_node,Eterm,is_internal_pid);
 ET_DEFINE_CHECKED(struct erl_node_*,internal_port_node,Eterm,is_internal_port);
 ET_DEFINE_CHECKED(Eterm*,internal_ref_val,Wterm,is_internal_ref);
-ET_DEFINE_CHECKED(Uint,internal_ref_data_words,Wterm,is_internal_ref);
-ET_DEFINE_CHECKED(Uint32*,internal_ref_data,Wterm,is_internal_ref);
+ET_DEFINE_CHECKED(Uint32*,internal_magic_ref_numbers,Wterm,is_internal_magic_ref);
+ET_DEFINE_CHECKED(Uint32*,internal_ordinary_ref_numbers,Wterm,is_internal_ordinary_ref);
 ET_DEFINE_CHECKED(struct erl_node_*,internal_ref_node,Eterm,is_internal_ref);
 ET_DEFINE_CHECKED(Eterm*,external_val,Wterm,is_external);
 ET_DEFINE_CHECKED(Uint,external_data_words,Wterm,is_external);
diff --git a/erts/emulator/beam/erl_term.h b/erts/emulator/beam/erl_term.h
index c3234ee349..a602a8f7c6 100644
--- a/erts/emulator/beam/erl_term.h
+++ b/erts/emulator/beam/erl_term.h
@@ -23,6 +23,8 @@
 
 #include "erl_mmap.h"
 
+void erts_term_init(void);
+
 typedef UWord Wterm;  /* Full word terms */
 
 struct erl_node_; /* Declared in erl_node_tables.h */
@@ -718,73 +720,224 @@ _ET_DECLARE_CHECKED(struct erl_node_*,internal_port_node,Eterm)
 #define ERTS_MAX_REF_NUMBERS	3
 #define ERTS_REF_NUMBERS	ERTS_MAX_REF_NUMBERS
 
-#if defined(ARCH_64)
-#  define ERTS_REF_WORDS	(ERTS_REF_NUMBERS/2 + 1)
-#  define ERTS_REF_32BIT_WORDS  (ERTS_REF_NUMBERS+1)
-#else
-#  define ERTS_REF_WORDS	ERTS_REF_NUMBERS
-#  define ERTS_REF_32BIT_WORDS  ERTS_REF_NUMBERS
+#ifndef ERTS_ENDIANNESS
+# error ERTS_ENDIANNESS not defined...
 #endif
 
-typedef struct {
-    Eterm      header;
-    union {
-	Uint32 ui32[ERTS_REF_32BIT_WORDS];
-	Uint   ui[ERTS_REF_WORDS];
-    } data;
-} RefThing;
-
-#define REF_THING_SIZE (sizeof(RefThing)/sizeof(Uint))
-#define REF_THING_HEAD_SIZE (sizeof(Eterm)/sizeof(Uint))
+#if ERTS_REF_NUMBERS != 3
+#  error "A new reference layout for 64-bit needs to be implemented..."
+#endif
 
-#define make_ref_thing_header(DW) \
-  _make_header((DW)+REF_THING_HEAD_SIZE-1,_TAG_HEADER_REF)
+struct magic_binary;
 
 #if defined(ARCH_64)
 
+# define ERTS_ORDINARY_REF_MARKER (~((Uint32) 0))
+
+typedef struct {
+    Eterm header;
+#if ERTS_ENDIANNESS <= 0
+    Uint32 marker;
+#endif
+    Uint32 num[ERTS_REF_NUMBERS];
+#if ERTS_ENDIANNESS > 0
+    Uint32 marker;
+#endif
+} ErtsORefThing;
+
+typedef struct {
+    Eterm header;
+    struct magic_binary *mb;
+    struct erl_off_heap_header* next;
+#if !ERTS_ENDIANNESS
+    Uint32 num[ERTS_REF_NUMBERS];
+    Uint32 marker;
+#endif
+} ErtsMRefThing;
+
 /*
- * Ref layout on a 64-bit little endian machine:
+ * Ordinary ref layout on a 64-bit little endian machine:
  *
  * 63             31             0
  * +--------------+--------------+
  * |         Thing word          |
  * +--------------+--------------+
- * |  Data 0      | 32-bit arity |
+ * |  Data 0      | 0xffffffff   |
  * +--------------+--------------+
  * |  Data 2      | Data 1       |
  * +--------------+--------------+
  *
- * Data is stored as an Uint32 array with 32-bit arity as first number.
+ * Ordinary ref layout on a 64-bit big endian machine:
+ *
+ * 63             31             0
+ * +--------------+--------------+
+ * |         Thing word          |
+ * +--------------+--------------+
+ * |  Data 0      | Data 1       |
+ * +--------------+--------------+
+ * |  Data 2      | 0xffffffff   |
+ * +--------------+--------------+
+ *
+ * Magic Ref layout on a 64-bit machine:
+ *
+ * 63             31             0
+ * +--------------+--------------+
+ * |         Thing word          |
+ * +--------------+--------------+
+ * |  Magic Binary Pointer       |
+ * +--------------+--------------+
+ * |  Next Off Heap Pointer      |
+ * +--------------+--------------+
+ *
+ * Both pointers in the magic ref are 64-bit aligned. That is,
+ * least significant bits are zero. The marker 32-bit word is
+ * placed over the least significant bits of one of the pointers.
+ * That is, we can distinguish between magic and ordinary ref
+ * by looking at the marker field.
+ * 
  */
 
 #define write_ref_thing(Hp, R0, R1, R2)					\
 do {									\
-  ((RefThing *) (Hp))->header  = make_ref_thing_header(ERTS_REF_WORDS);	\
-  ((RefThing *) (Hp))->data.ui32[0] = ERTS_REF_NUMBERS;			\
-  ((RefThing *) (Hp))->data.ui32[1] = (R0);				\
-  ((RefThing *) (Hp))->data.ui32[2] = (R1);				\
-  ((RefThing *) (Hp))->data.ui32[3] = (R2);				\
+  ((ErtsORefThing *) (Hp))->header = ERTS_REF_THING_HEADER;		\
+  ((ErtsORefThing *) (Hp))->marker = ERTS_ORDINARY_REF_MARKER;		\
+  ((ErtsORefThing *) (Hp))->num[0] = (R0);				\
+  ((ErtsORefThing *) (Hp))->num[1] = (R1);				\
+  ((ErtsORefThing *) (Hp))->num[2] = (R2);				\
 } while (0)
 
-#else
+#if ERTS_ENDIANNESS
+/* Known big or little endian */
+
+#define write_magic_ref_thing(Hp, Ohp, Binp)				\
+do {									\
+  ((ErtsMRefThing *) (Hp))->header = ERTS_REF_THING_HEADER;		\
+  ((ErtsMRefThing *) (Hp))->mb = (Binp);				\
+  ((ErtsMRefThing *) (Hp))->next = (Ohp)->first;			\
+  (Ohp)->first = (struct erl_off_heap_header*) (Hp);			\
+  ASSERT(erts_is_ref_numbers_magic((Binp)->refn));			\
+} while (0)
+
+#else /* !ERTS_ENDIANNESS */
+
+#define write_magic_ref_thing(Hp, Ohp, Binp)				\
+do {									\
+  ((ErtsMRefThing *) (Hp))->header = ERTS_MAGIC_REF_THING_HEADER;	\
+  ((ErtsMRefThing *) (Hp))->mb = (Binp);				\
+  ((ErtsMRefThing *) (Hp))->next = (Ohp)->first;			\
+  (Ohp)->first = (struct erl_off_heap_header*) (Hp);			\
+  ((ErtsMRefThing *) (Hp))->marker = 0;					\
+  ((ErtsMRefThing *) (Hp))->num[0] = (Binp)->refn[0];			\
+  ((ErtsMRefThing *) (Hp))->num[1] = (Binp)->refn[1];			\
+  ((ErtsMRefThing *) (Hp))->num[2] = (Binp)->refn[2];			\
+  ASSERT(erts_is_ref_numbers_magic((Binp)->refn));			\
+} while (0)
+
+#endif /* !ERTS_ENDIANNESS */
+
+#else /* ARCH_32 */
+
+typedef struct {
+    Eterm header;
+    Uint32 num[ERTS_REF_NUMBERS];
+} ErtsORefThing;
+
+typedef struct {
+    Eterm header;
+    struct magic_binary *mb;
+    struct erl_off_heap_header* next;
+} ErtsMRefThing;
+
 
 #define write_ref_thing(Hp, R0, R1, R2)					\
 do {									\
-  ((RefThing *) (Hp))->header  = make_ref_thing_header(ERTS_REF_WORDS);	\
-  ((RefThing *) (Hp))->data.ui32[0] = (R0);				\
-  ((RefThing *) (Hp))->data.ui32[1] = (R1);				\
-  ((RefThing *) (Hp))->data.ui32[2] = (R2);				\
+  ((ErtsORefThing *) (Hp))->header = ERTS_REF_THING_HEADER;		\
+  ((ErtsORefThing *) (Hp))->num[0] = (R0);				\
+  ((ErtsORefThing *) (Hp))->num[1] = (R1);				\
+  ((ErtsORefThing *) (Hp))->num[2] = (R2);				\
 } while (0)
 
+#define write_magic_ref_thing(Hp, Ohp, Binp)				\
+do {									\
+  ((ErtsMRefThing *) (Hp))->header = ERTS_MAGIC_REF_THING_HEADER;	\
+  ((ErtsMRefThing *) (Hp))->mb = (Binp);				\
+  ((ErtsMRefThing *) (Hp))->next = (Ohp)->first;			\
+  (Ohp)->first = (struct erl_off_heap_header*) (Hp);			\
+  ASSERT(erts_is_ref_numbers_magic(&(Binp)->refn));			\
+} while (0)
+
+#endif /* ARCH_32 */
+
+typedef union {
+    ErtsMRefThing m;
+    ErtsORefThing o;
+} ErtsRefThing;
+
+#define ERTS_REF_THING_SIZE (sizeof(ErtsORefThing)/sizeof(Uint))
+#define ERTS_MAGIC_REF_THING_SIZE (sizeof(ErtsMRefThing)/sizeof(Uint))
+#define ERTS_MAX_INTERNAL_REF_SIZE (sizeof(ErtsRefThing)/sizeof(Uint))
+
+#define make_ref_thing_header(Words) \
+    _make_header((Words)-1,_TAG_HEADER_REF)
+
+#define ERTS_REF_THING_HEADER _make_header(ERTS_REF_THING_SIZE-1,_TAG_HEADER_REF)
+
+#if defined(ARCH_64) && ERTS_ENDIANNESS /* All internal refs of same size... */
+
+#  undef ERTS_MAGIC_REF_THING_HEADER
+
+#  define is_ref_thing_header(x) ((x) == ERTS_REF_THING_HEADER)
+
+#define is_ordinary_ref_thing(x)					\
+    (ASSERT(is_ref_thing_header(*((Eterm *)(x)))),			\
+     ((ErtsRefThing *) (x))->o.marker == ERTS_ORDINARY_REF_MARKER)
+
+#define is_magic_ref_thing(x)						\
+    (!is_ordinary_ref_thing((x)))
+
+#define is_internal_magic_ref(x)					\
+    ((_unchecked_is_boxed((x)) && *boxed_val((x)) == ERTS_REF_THING_HEADER) \
+     && is_magic_ref_thing(boxed_val((x))))
+
+#define is_internal_ordinary_ref(x)					\
+    ((_unchecked_is_boxed((x)) && *boxed_val((x)) == ERTS_REF_THING_HEADER) \
+     && is_ordinary_ref_thing(boxed_val((x))))
+
+#else /* Ordinary and magic references of different sizes... */
+
+#  define ERTS_MAGIC_REF_THING_HEADER					\
+    _make_header(ERTS_MAGIC_REF_THING_SIZE-1,_TAG_HEADER_REF)
+
+#  define is_ref_thing_header(x)					\
+    (((x) & _TAG_HEADER_MASK) == _TAG_HEADER_REF)
+
+#define is_ordinary_ref_thing(x)					\
+    (ASSERT(is_ref_thing_header(*((Eterm *)(x)))),			\
+     *((Eterm *)(x)) == ERTS_REF_THING_HEADER)
+
+#define is_magic_ref_thing(x)						\
+    (ASSERT(is_ref_thing_header(*((Eterm *)(x)))),			\
+     *((Eterm *)(x)) == ERTS_MAGIC_REF_THING_HEADER)
+
+#define is_internal_magic_ref(x)					\
+    (_unchecked_is_boxed((x)) && *boxed_val((x)) == ERTS_MAGIC_REF_THING_HEADER)
+
+#define is_internal_ordinary_ref(x)					\
+    (_unchecked_is_boxed((x)) && *boxed_val((x)) == ERTS_REF_THING_HEADER)
+
 #endif
 
-#define is_ref_thing_header(x)	(((x) & _TAG_HEADER_MASK) == _TAG_HEADER_REF)
 #define make_internal_ref(x)	make_boxed((Eterm*)(x))
 
-#define _unchecked_ref_thing_ptr(x) \
-  ((RefThing*) _unchecked_internal_ref_val(x))
-#define ref_thing_ptr(x) \
-  ((RefThing*) internal_ref_val(x))
+#define _unchecked_ordinary_ref_thing_ptr(x) \
+  ((ErtsORefThing*) _unchecked_internal_ref_val(x))
+#define ordinary_ref_thing_ptr(x) \
+  ((ErtsORefThing*) internal_ref_val(x))
+
+#define _unchecked_magic_ref_thing_ptr(x) \
+  ((ErtsMRefThing*) _unchecked_internal_ref_val(x))
+#define magic_ref_thing_ptr(x) \
+  ((ErtsMRefThing*) internal_ref_val(x))
 
 #define is_internal_ref(x) \
     (_unchecked_is_boxed((x)) && is_ref_thing_header(*boxed_val((x))))
@@ -796,16 +949,21 @@ do {									\
 _ET_DECLARE_CHECKED(Eterm*,internal_ref_val,Wterm)
 #define internal_ref_val(x) _ET_APPLY(internal_ref_val,(x))
 
-#define internal_thing_ref_data_words(t) (thing_arityval(*(Eterm*)(t)))
-#define _unchecked_internal_ref_data_words(x) \
- (_unchecked_thing_arityval(*_unchecked_internal_ref_val(x)))
-_ET_DECLARE_CHECKED(Uint,internal_ref_data_words,Wterm)
-#define internal_ref_data_words(x) _ET_APPLY(internal_ref_data_words,(x))
+#define internal_ordinary_thing_ref_numbers(ort) (((ErtsORefThing *)(ort))->num)
+#define _unchecked_internal_ordinary_ref_numbers(x) (internal_ordinary_thing_ref_numbers(_unchecked_ordinary_ref_thing_ptr(x)))
+_ET_DECLARE_CHECKED(Uint32*,internal_ordinary_ref_numbers,Wterm)
+#define internal_ordinary_ref_numbers(x) _ET_APPLY(internal_ordinary_ref_numbers,(x))
+
+#if defined(ARCH_64) && !ERTS_ENDIANNESS
+#define internal_magic_thing_ref_numbers(mrt) (((ErtsMRefThing *)(mrt))->num)
+#else
+#define internal_magic_thing_ref_numbers(mrt) (((ErtsMRefThing *)(mrt))->mb->refn)
+#endif
+
+#define _unchecked_internal_magic_ref_numbers(x) (internal_magic_thing_ref_numbers(_unchecked_magic_ref_thing_ptr(x)))
+_ET_DECLARE_CHECKED(Uint32*,internal_magic_ref_numbers,Wterm)
+#define internal_magic_ref_numbers(x) _ET_APPLY(internal_magic_ref_numbers,(x))
 
-#define internal_thing_ref_data(thing) ((thing)->data.ui32)
-#define _unchecked_internal_ref_data(x) (internal_thing_ref_data(_unchecked_ref_thing_ptr(x)))
-_ET_DECLARE_CHECKED(Uint32*,internal_ref_data,Wterm)
-#define internal_ref_data(x) _ET_APPLY(internal_ref_data,(x))
 
 #define _unchecked_internal_ref_node(x) erts_this_node
 _ET_DECLARE_CHECKED(struct erl_node_*,internal_ref_node,Eterm)
diff --git a/erts/emulator/beam/erl_time_sup.c b/erts/emulator/beam/erl_time_sup.c
index c01cc26246..cf9d3adc86 100644
--- a/erts/emulator/beam/erl_time_sup.c
+++ b/erts/emulator/beam/erl_time_sup.c
@@ -1502,7 +1502,7 @@ static time_t gregday(int year, int month, int day)
     pyear = gyear - 1;
     ndays = (pyear/4) - (pyear/100) + (pyear/400) + pyear*365 + 366;
   }
-  /* number of days in all months preceeding month */
+  /* number of days in all months preceding month */
   for (m = 1; m < month; m++)
     ndays += mdays[m];
   /* Extra day if leap year and March or later */
@@ -1831,7 +1831,10 @@ erts_demonitor_time_offset(Eterm ref)
     ErtsMonitor *mon;
     ASSERT(is_internal_ref(ref));
     erts_smp_mtx_lock(&erts_get_time_mtx);
-    mon = erts_remove_monitor(&time_offset_monitors, ref);
+    if (is_internal_ordinary_ref(ref))
+        mon = erts_remove_monitor(&time_offset_monitors, ref);
+    else
+        mon = NULL;
     if (!mon)
 	res = 0;
     else {
@@ -1848,7 +1851,7 @@ erts_demonitor_time_offset(Eterm ref)
 typedef struct {
     Eterm pid;
     Eterm ref;
-    Eterm heap[REF_THING_SIZE];
+    Eterm heap[ERTS_REF_THING_SIZE];
 } ErtsTimeOffsetMonitorInfo;
 
 typedef struct {
@@ -1869,11 +1872,11 @@ save_time_offset_monitor(ErtsMonitor *mon, void *vcntxt)
     cntxt->to_mon_info[mix].pid = mon->u.pid;
     to_hp = &cntxt->to_mon_info[mix].heap[0];
 
-    ASSERT(is_internal_ref(mon->ref));
+    ASSERT(is_internal_ordinary_ref(mon->ref));
     from_hp = internal_ref_val(mon->ref);
-    ASSERT(thing_arityval(*from_hp) + 1 == REF_THING_SIZE);
+    ASSERT(thing_arityval(*from_hp) + 1 == ERTS_REF_THING_SIZE);
 
-    for (hix = 0; hix < REF_THING_SIZE; hix++)
+    for (hix = 0; hix < ERTS_REF_THING_SIZE; hix++)
 	to_hp[hix] = from_hp[hix];
 
     cntxt->to_mon_info[mix].ref
@@ -1933,7 +1936,7 @@ send_time_offset_changed_notifications(void *new_offsetp)
 	hp = (Eterm *) (tmp + no_monitors*sizeof(ErtsTimeOffsetMonitorInfo));
 
 	hsz = 6; /* 5-tuple */
-	hsz += REF_THING_SIZE;
+	hsz += ERTS_REF_THING_SIZE;
 	hsz += ERTS_SINT64_HEAP_SIZE(new_offset);
 
 	if (IS_SSMALL(new_offset))
diff --git a/erts/emulator/beam/erl_unicode.c b/erts/emulator/beam/erl_unicode.c
index 3167ca5eef..01629db9ad 100644
--- a/erts/emulator/beam/erl_unicode.c
+++ b/erts/emulator/beam/erl_unicode.c
@@ -130,13 +130,9 @@ static int cleanup_restart_context_bin(Binary *bp)
     return 1;
 }
 
-static RestartContext *get_rc_from_bin(Eterm bin)
+static RestartContext *get_rc_from_bin(Eterm mref)
 {
-    Binary *mbp;
-    ASSERT(ERTS_TERM_IS_MAGIC_BINARY(bin));
-
-    mbp = ((ProcBin *) binary_val(bin))->val;
-
+    Binary *mbp = erts_magic_ref2bin(mref);
     ASSERT(ERTS_MAGIC_BIN_DESTRUCTOR(mbp)
 	   == cleanup_restart_context_bin);
     return (RestartContext *) ERTS_MAGIC_BIN_DATA(mbp);    
@@ -149,8 +145,8 @@ static Eterm make_magic_bin_for_restart(Process *p, RestartContext *rc)
     RestartContext *restartp = ERTS_MAGIC_BIN_DATA(mbp);
     Eterm *hp;
     memcpy(restartp,rc,sizeof(RestartContext));
-    hp = HAlloc(p, PROC_BIN_SIZE);
-    return erts_mk_magic_binary_term(&hp, &MSO(p), mbp);
+    hp = HAlloc(p, ERTS_MAGIC_REF_THING_SIZE);
+    return erts_mk_magic_ref(&hp, &MSO(p), mbp);
 }
 
 	
@@ -1891,74 +1887,57 @@ binary_to_atom(Process* proc, Eterm bin, Eterm enc, int must_exist)
     byte* bytes;
     byte *temp_alloc = NULL;
     Uint bin_size;
+    Eterm a;
 
     if ((bytes = erts_get_aligned_binary_bytes(bin, &temp_alloc)) == 0) {
 	BIF_ERROR(proc, BADARG);
     }
     bin_size = binary_size(bin);
+
     if (enc == am_latin1) {
-	Eterm a;
-	if (bin_size > MAX_ATOM_CHARACTERS) {
-	system_limit:
-	    erts_free_aligned_binary_bytes(temp_alloc);
-	    BIF_ERROR(proc, SYSTEM_LIMIT);
-	}
 	if (!must_exist) {
-	    a = erts_atom_put((byte *) bytes,
-			      bin_size,
-			      ERTS_ATOM_ENC_LATIN1,
-			      0);
-    	    erts_free_aligned_binary_bytes(temp_alloc);
-	    if (is_non_value(a))
-		goto badarg;
-	    BIF_RET(a);
-	} else if (erts_atom_get((char *)bytes, bin_size, &a, ERTS_ATOM_ENC_LATIN1)) {
-	    erts_free_aligned_binary_bytes(temp_alloc);
-	    BIF_RET(a);
-	} else {
+	    int lix = erts_atom_put_index((byte *) bytes,
+					  bin_size,
+					  ERTS_ATOM_ENC_LATIN1,
+					  0);
+	    if (lix == ATOM_BAD_ENCODING_ERROR) {
+	    badarg:
+		erts_free_aligned_binary_bytes(temp_alloc);
+		BIF_ERROR(proc, BADARG);
+	    } else if (lix == ATOM_MAX_CHARS_ERROR) {
+	    system_limit:
+		erts_free_aligned_binary_bytes(temp_alloc);
+		BIF_ERROR(proc, SYSTEM_LIMIT);
+	    }
+
+	    a = make_atom(lix);
+	} else if (!erts_atom_get((char *)bytes, bin_size, &a, ERTS_ATOM_ENC_LATIN1)) {
 	    goto badarg;
 	}
-    } else if (enc == am_utf8 || enc == am_unicode) {
-	Eterm res;
-	Uint num_chars = 0;
-	const byte* p = bytes;
-	Uint left = bin_size;
 
-	while (left) {
-	    if (++num_chars > MAX_ATOM_CHARACTERS) {
+    } else if (enc == am_utf8 || enc == am_unicode) {
+	if (!must_exist) {
+	    int uix = erts_atom_put_index((byte *) bytes,
+					  bin_size,
+					  ERTS_ATOM_ENC_UTF8,
+					  0);
+	    if (uix == ATOM_BAD_ENCODING_ERROR) {
+		goto badarg;
+	    } else if (uix == ATOM_MAX_CHARS_ERROR) {
 		goto system_limit;
 	    }
-	    if ((p[0] & 0x80) == 0) {
-		++p;
-		--left;
-	    }
-	    else if (left >= 2
-		     && (p[0] & 0xFE) == 0xC2 /* only allow latin1 subset */
-		     && (p[1] & 0xC0) == 0x80) {
-		p += 2;
-		left -= 2;
-	    }
-	    else goto badarg;
-	}
 
-	if (!must_exist) {
-	    res = erts_atom_put((byte *) bytes,
-				bin_size,
-				ERTS_ATOM_ENC_UTF8,
-				0);
+	    a = make_atom(uix);
 	}
-	else if (!erts_atom_get((char*)bytes, bin_size, &res, ERTS_ATOM_ENC_UTF8)) {
+	else if (!erts_atom_get((char*)bytes, bin_size, &a, ERTS_ATOM_ENC_UTF8)) {
 	    goto badarg;
 	}
-	erts_free_aligned_binary_bytes(temp_alloc);
-	if (is_non_value(res))
-	    goto badarg;
-	BIF_RET(res);
     } else {
-    badarg:
-	erts_free_aligned_binary_bytes(temp_alloc);
-	BIF_ERROR(proc, BADARG);
+	goto badarg;
     }
+
+    erts_free_aligned_binary_bytes(temp_alloc);
+    BIF_RET(a);
 }
 
 BIF_RETTYPE binary_to_atom_2(BIF_ALIST_2)
diff --git a/erts/emulator/beam/erl_utils.h b/erts/emulator/beam/erl_utils.h
index 81800752f0..47289a0af1 100644
--- a/erts/emulator/beam/erl_utils.h
+++ b/erts/emulator/beam/erl_utils.h
@@ -117,7 +117,6 @@ int erts_fit_in_bits_int32(Sint32);
 int erts_fit_in_bits_uint(Uint);
 Sint erts_list_length(Eterm);
 int erts_is_builtin(Eterm, Eterm, int);
-Uint32 make_broken_hash(Eterm);
 Uint32 block_hash(byte *, unsigned, Uint32);
 Uint32 make_hash2(Eterm);
 Uint32 make_hash(Eterm);
diff --git a/erts/emulator/beam/erl_vm.h b/erts/emulator/beam/erl_vm.h
index d88cafc5bf..5366ee3644 100644
--- a/erts/emulator/beam/erl_vm.h
+++ b/erts/emulator/beam/erl_vm.h
@@ -110,8 +110,21 @@
 #define HeapWordsLeft(p) (HEAP_LIMIT(p) - HEAP_TOP(p))
 
 #if defined(DEBUG) || defined(CHECK_FOR_HOLES)
-# define ERTS_HOLE_MARKER (((0xdeadbeef << 24) << 8) | 0xdeadbeef)
-#endif /* egil: 32-bit ? */
+
+/*
+ * ERTS_HOLE_MARKER must *not* be mistaken for a valid term
+ * on the heap...
+ */
+#  ifdef ARCH_64
+#    define ERTS_HOLE_MARKER \
+    make_catch(UWORD_CONSTANT(0xdeadbeaf00000000) >> _TAG_IMMED2_SIZE)
+/* Will (at the time of writing) appear as 0xdeadbeaf0000001b */
+#  else
+#    define ERTS_HOLE_MARKER \
+    make_catch(UWORD_CONSTANT(0xdead0000) >> _TAG_IMMED2_SIZE)
+/* Will (at the time of writing) appear as 0xdead001b */
+#  endif
+#endif
 
 /*
  * Allocate heap memory on the ordinary heap, NEVER in a heap
diff --git a/erts/emulator/beam/external.c b/erts/emulator/beam/external.c
index cb3486fde2..205a7711ec 100644
--- a/erts/emulator/beam/external.c
+++ b/erts/emulator/beam/external.c
@@ -616,7 +616,7 @@ erts_make_dist_ext_copy(ErtsDistExternal *edep, Uint xsize)
     sys_memcpy((void *) ep, (void *) edep, dist_ext_sz);
     ep += dist_ext_sz;
     if (new_edep->dep)
-	erts_refc_inc(&new_edep->dep->refc, 1);
+	erts_smp_refc_inc(&new_edep->dep->refc, 1);
     new_edep->extp = ep;
     new_edep->ext_endp = ep + ext_sz;
     new_edep->heap_size = -1;
@@ -1079,7 +1079,7 @@ static BIF_RETTYPE term_to_binary_trap_1(BIF_ALIST_1)
     Eterm *tp = tuple_val(BIF_ARG_1);
     Eterm Term = tp[1];
     Eterm bt = tp[2];
-    Binary *bin = ((ProcBin *) binary_val(bt))->val;
+    Binary *bin = erts_magic_ref2bin(bt);
     Eterm res = erts_term_to_binary_int(BIF_P, Term, 0, 0,bin);
     if (is_tuple(res)) {
 	ASSERT(BIF_P->flags & F_DISABLE_GC);
@@ -1222,6 +1222,7 @@ typedef struct B2TContext_t {
     } u;
 } B2TContext;
 
+static B2TContext* b2t_export_context(Process*, B2TContext* src);
 
 static uLongf binary2term_uncomp_size(byte* data, Sint size)
 {
@@ -1254,7 +1255,7 @@ static uLongf binary2term_uncomp_size(byte* data, Sint size)
 
 static ERTS_INLINE int
 binary2term_prepare(ErtsBinary2TermState *state, byte *data, Sint data_size,
-		    B2TContext* ctx)
+		    B2TContext** ctxp, Process* p)
 {
     byte *bytes = data;
     Sint size = data_size;
@@ -1268,8 +1269,8 @@ binary2term_prepare(ErtsBinary2TermState *state, byte *data, Sint data_size,
     size--;
     if (size < 5 || *bytes != COMPRESSED) {
 	state->extp = bytes;
-        if (ctx)
-	    ctx->state = B2TSizeInit;
+        if (ctxp)
+	    (*ctxp)->state = B2TSizeInit;
     }
     else  {
 	uLongf dest_len = (Uint32) get_int32(bytes+1);
@@ -1286,16 +1287,26 @@ binary2term_prepare(ErtsBinary2TermState *state, byte *data, Sint data_size,
                 return -1;
 	    }
 	    state->extp = erts_alloc(ERTS_ALC_T_EXT_TERM_DATA, dest_len);
-            ctx->reds -= dest_len;
+            if (ctxp)
+                (*ctxp)->reds -= dest_len;
 	}
 	state->exttmp = 1;
-        if (ctx) {
+        if (ctxp) {
+            /*
+             * Start decompression by exporting trap context
+             * so we don't have to deal with deep-copying z_stream.
+             */
+            B2TContext* ctx = b2t_export_context(p, *ctxp);
+            ASSERT(state = &(*ctxp)->b2ts);
+            state = &ctx->b2ts;
+
 	    if (erl_zlib_inflate_start(&ctx->u.uc.stream, bytes, size) != Z_OK)
 		return -1;
 
 	    ctx->u.uc.dbytes = state->extp;
 	    ctx->u.uc.dleft = dest_len;
 	    ctx->state = B2TUncompressChunk;
+            *ctxp = ctx;
         }
 	else {
 	    uLongf dlen = dest_len;
@@ -1339,7 +1350,7 @@ erts_binary2term_prepare(ErtsBinary2TermState *state, byte *data, Sint data_size
 {
     Sint res;
 
-    if (binary2term_prepare(state, data, data_size, NULL) < 0 ||
+    if (binary2term_prepare(state, data, data_size, NULL, NULL) < 0 ||
         (res=decoded_size(state->extp, state->extp + state->extsize, 0, NULL)) < 0) {
 
         if (state->exttmp)
@@ -1398,7 +1409,7 @@ static int b2t_context_destructor(Binary *context_bin)
 
 static BIF_RETTYPE binary_to_term_trap_1(BIF_ALIST_1)
 {
-    Binary *context_bin = ((ProcBin *) binary_val(BIF_ARG_1))->val;
+    Binary *context_bin = erts_magic_ref2bin(BIF_ARG_1);
     ASSERT(ERTS_MAGIC_BIN_DESTRUCTOR(context_bin) == b2t_context_destructor);
 
     return binary_to_term_int(BIF_P, 0, THE_NON_VALUE, context_bin, NULL,
@@ -1432,8 +1443,8 @@ static B2TContext* b2t_export_context(Process* p, B2TContext* src)
     if (ctx->state >= B2TDecode && ctx->u.dc.next == &src->u.dc.res) {
         ctx->u.dc.next = &ctx->u.dc.res;
     }
-    hp = HAlloc(p, PROC_BIN_SIZE);
-    ctx->trap_bin = erts_mk_magic_binary_term(&hp, &MSO(p), context_b);
+    hp = HAlloc(p, ERTS_MAGIC_REF_THING_SIZE);
+    ctx->trap_bin = erts_mk_magic_ref(&hp, &MSO(p), context_b);
     return ctx;
 }
 
@@ -1486,7 +1497,7 @@ static BIF_RETTYPE binary_to_term_int(Process* p, Uint32 flags, Eterm bin, Binar
             if (ctx->aligned_alloc) {
                 ctx->reds -= bin_size / 8;
             }
-            if (binary2term_prepare(&ctx->b2ts, bytes, bin_size, ctx) < 0) {
+            if (binary2term_prepare(&ctx->b2ts, bytes, bin_size, &ctx, p) < 0) {
 		ctx->state = B2TBadArg;
 	    }
             break;
@@ -1862,8 +1873,8 @@ static Eterm erts_term_to_binary_int(Process* p, Eterm Term, int level, Uint fla
 
 #define RETURN_STATE()							\
     do {								\
-	hp = HAlloc(p, PROC_BIN_SIZE+3);				\
-	c_term = erts_mk_magic_binary_term(&hp, &MSO(p), context_b);	\
+	hp = HAlloc(p, ERTS_MAGIC_REF_THING_SIZE+3);                    \
+	c_term = erts_mk_magic_ref(&hp, &MSO(p), context_b);            \
 	res = TUPLE2(hp, Term, c_term);					\
 	BUMP_ALL_REDS(p);                                               \
 	return res;							\
@@ -2570,7 +2581,9 @@ enc_term_int(TTBEncodeContext* ctx, ErtsAtomCacheMap *acmp, Eterm obj, byte* ep,
 
 	    ASSERT(dflags & DFLAG_EXTENDED_REFERENCES);
 
-	    i = ref_no_of_numbers(obj);
+	    erts_magic_ref_save_bin(obj);
+
+	    i = ref_no_numbers(obj);
 	    put_int16(i, ep);
 	    ep += 2;
 	    ep = enc_atom(acmp, sysname, ep, dflags);
@@ -3426,26 +3439,35 @@ dec_term_atom_common:
 		r0 = get_int32(ep); /* allow full word */
 		ep += 4;
 
-	    ref_ext_common:
+	    ref_ext_common: {
+		ErtsORefThing *rtp;
+
 		if (ref_words > ERTS_MAX_REF_NUMBERS)
 		    goto error;
 
 		node = dec_get_node(sysname, cre);
 		if(node == erts_this_node) {
-		    RefThing *rtp = (RefThing *) hp;
-		    ref_num = (Uint32 *) (hp + REF_THING_HEAD_SIZE);
+	
+		    rtp = (ErtsORefThing *) hp;
+		    ref_num = &rtp->num[0];
+		    if (ref_words != ERTS_REF_NUMBERS) {
+                        int i;
+                        if (ref_words > ERTS_REF_NUMBERS)
+                            goto error; /* Not a ref that we created... */
+                        for (i = ref_words; i < ERTS_REF_NUMBERS; i++)
+                            ref_num[i] = 0;
+		    }
 
-#if defined(ARCH_64)
-		    hp += REF_THING_HEAD_SIZE + ref_words/2 + 1;
-		    rtp->header = make_ref_thing_header(ref_words/2 + 1);
-#else
-		    hp += REF_THING_HEAD_SIZE + ref_words;
-		    rtp->header = make_ref_thing_header(ref_words);
+#ifdef ERTS_ORDINARY_REF_MARKER
+		    rtp->marker = ERTS_ORDINARY_REF_MARKER;
 #endif
+		    hp += ERTS_REF_THING_SIZE;
+		    rtp->header = ERTS_REF_THING_HEADER;
 		    *objp = make_internal_ref(rtp);
 		}
 		else {
 		    ExternalThing *etp = (ExternalThing *) hp;
+		    rtp = NULL;
 #if defined(ARCH_64)
 		    hp += EXTERNAL_THING_HEAD_SIZE + ref_words/2 + 1;
 #else
@@ -3463,12 +3485,13 @@ dec_term_atom_common:
 		    factory->off_heap->first = (struct erl_off_heap_header*)etp;
 		    *objp = make_external_ref(etp);
 		    ref_num = &(etp->data.ui32[0]);
-		}
-
 #if defined(ARCH_64)
-		*(ref_num++) = ref_words /* 32-bit arity */;
+		    *(ref_num++) = ref_words /* 32-bit arity */;
 #endif
+		}
+
 		ref_num[0] = r0;
+
 		for(i = 1; i < ref_words; i++) {
 		    ref_num[i] = get_int32(ep);
 		    ep += 4;
@@ -3477,8 +3500,26 @@ dec_term_atom_common:
 		if ((1 + ref_words) % 2)
 		    ref_num[ref_words] = 0;
 #endif
+		if (node == erts_this_node) {
+		    /* Check if it was a magic reference... */
+		    ErtsMagicBinary *mb = erts_magic_ref_lookup_bin(ref_num);
+		    if (mb) {
+			/*
+			 * Was a magic ref; adjust it...
+			 *
+			 * Refc on binary was increased by lookup above...
+			 */
+			ASSERT(rtp);
+			hp = (Eterm *) rtp;
+			write_magic_ref_thing(hp, factory->off_heap, mb);
+                        OH_OVERHEAD(factory->off_heap,
+                                    mb->orig_size / sizeof(Eterm));
+			hp += ERTS_MAGIC_REF_THING_SIZE;
+		    }
+		}
 		break;
 	    }
+	    }
 	case BINARY_EXT:
 	    {
 		n = get_int32(ep);
@@ -4100,7 +4141,7 @@ encode_size_struct_int(TTBSizeContext* ctx, ErtsAtomCacheMap *acmp, Eterm obj,
             /*fall through*/
 	case REF_DEF:
 	    ASSERT(dflags & DFLAG_EXTENDED_REFERENCES);
-	    i = ref_no_of_numbers(obj);
+	    i = ref_no_numbers(obj);
 	    result += (1 + 2 + encode_size_struct2(acmp, ref_node_name(obj), dflags) +
 		       1 + 4*i);
 	    break;
diff --git a/erts/emulator/beam/global.h b/erts/emulator/beam/global.h
index 511e357d14..d4a1225bdd 100644
--- a/erts/emulator/beam/global.h
+++ b/erts/emulator/beam/global.h
@@ -43,9 +43,15 @@
 #include "erl_port.h"
 #include "erl_gc.h"
 #include "erl_nif.h"
+#define ERTS_BINARY_TYPES_ONLY__
+#include "erl_binary.h"
+#undef ERTS_BINARY_TYPES_ONLY__
 
 struct enif_func_t;
 
+#ifdef  DEBUG
+#  define ERTS_NIF_ASSERT_IN_ENV
+#endif
 struct enif_environment_t /* ErlNifEnv */
 {
     struct erl_module_nif* mod_nif;
@@ -58,6 +64,10 @@ struct enif_environment_t /* ErlNifEnv */
     int exception_thrown; /* boolean */
     Process *tracee;
     int exiting; /* boolean (dirty nifs might return in exiting state) */
+
+#ifdef ERTS_NIF_ASSERT_IN_ENV
+    int dbg_disable_assert_in_env;
+#endif
 };
 struct enif_resource_type_t
 {
@@ -98,7 +108,7 @@ typedef struct ErtsResource_
 }ErtsResource;
 
 #define DATA_TO_RESOURCE(PTR) ErtsContainerStruct(PTR, ErtsResource, data)
-#define erts_resource_ref_size(P) PROC_BIN_SIZE
+#define erts_resource_ref_size(P) ERTS_MAGIC_REF_THING_SIZE
 
 extern Eterm erts_bld_resource_ref(Eterm** hp, ErlOffHeap*, ErtsResource*);
 
@@ -165,7 +175,7 @@ typedef struct de_proc_entry {
 			                PROC_AWAIT_LOAD == Wants to be notified when we
 			                reloaded the driver (old was locked) */
     Uint    flags;                   /* ERL_FL_DE_DEREFERENCED when reload in progress */
-    Eterm   heap[REF_THING_SIZE];    /* "ref heap" */
+    Eterm   heap[ERTS_REF_THING_SIZE];    /* "ref heap" */
     struct  de_proc_entry *next;
 } DE_ProcEntry;
 
@@ -173,7 +183,7 @@ typedef struct {
     void         *handle;             /* Handle for DLL or SO (for dyn. drivers). */
     DE_ProcEntry *procs;              /* List of pids that have loaded this driver,
 				         or that wait for it to change state */
-    erts_refc_t  refc;                /* Number of ports/processes having
+    erts_smp_refc_t  refc;                /* Number of ports/processes having
 					 references to the driver */
     erts_smp_atomic32_t port_count;   /* Number of ports using the driver */
     Uint         flags;               /* ERL_DE_FL_KILL_PORTS */
@@ -253,116 +263,6 @@ extern Eterm erts_ddll_monitor_driver(Process *p,
 				      ErtsProcLocks plocks);
 
 /*
-** Just like the driver binary but with initial flags
-** Note that the two structures Binary and ErlDrvBinary HAVE to
-** be equal except for extra fields in the beginning of the struct.
-** ErlDrvBinary is defined in erl_driver.h.
-** When driver_alloc_binary is called, a Binary is allocated, but 
-** the pointer returned is to the address of the first element that
-** also occurs in the ErlDrvBinary struct (driver.*binary takes care if this).
-** The driver need never know about additions to the internal Binary of the
-** emulator. One should however NEVER be sloppy when mixing ErlDrvBinary
-** and Binary, the macros below can convert one type to the other, as they both
-** in reality are equal.
-*/
-
-#ifdef ARCH_32
- /* *DO NOT USE* only for alignment. */
-#define ERTS_BINARY_STRUCT_ALIGNMENT Uint32 align__;
-#else
-#define ERTS_BINARY_STRUCT_ALIGNMENT
-#endif
-
-/* Add fields in ERTS_INTERNAL_BINARY_FIELDS, otherwise the drivers crash */
-#define ERTS_INTERNAL_BINARY_FIELDS				\
-    UWord flags;							\
-    erts_refc_t refc;						\
-    ERTS_BINARY_STRUCT_ALIGNMENT
-
-typedef struct binary {
-    ERTS_INTERNAL_BINARY_FIELDS
-    SWord orig_size;
-    char orig_bytes[1]; /* to be continued */
-} Binary;
-
-#define ERTS_SIZEOF_Binary(Sz) \
-    (offsetof(Binary,orig_bytes) + (Sz))
-
-typedef struct {
-    ERTS_INTERNAL_BINARY_FIELDS
-    SWord orig_size;
-    int (*destructor)(Binary *);
-    union {
-        struct {
-            ERTS_BINARY_STRUCT_ALIGNMENT
-            char data[1];
-        } aligned;
-        struct {
-            char data[1];
-        } unaligned;
-    } u;
-} ErtsMagicBinary;
-
-#define ERTS_MAGIC_BIN_BYTES_TO_ALIGN \
-  (offsetof(ErtsMagicBinary,u.aligned.data) - \
-   offsetof(ErtsMagicBinary,u.unaligned.data))
-
-typedef union {
-    Binary binary;
-    ErtsMagicBinary magic_binary;
-    struct {
-	ERTS_INTERNAL_BINARY_FIELDS
-	ErlDrvBinary binary;
-    } driver;
-} ErtsBinary;
-
-/*
- * 'Binary' alignment:
- *   Address of orig_bytes[0] of a Binary should always be 8-byte aligned.
- * It is assumed that the flags, refc, and orig_size fields are 4 bytes on
- * 32-bits architectures and 8 bytes on 64-bits architectures.
- */
-
-#define ERTS_MAGIC_BIN_DESTRUCTOR(BP) \
-  ((ErtsBinary *) (BP))->magic_binary.destructor
-#define ERTS_MAGIC_BIN_DATA(BP) \
-  ((void *) ((ErtsBinary *) (BP))->magic_binary.u.aligned.data)
-#define ERTS_MAGIC_DATA_OFFSET \
-  (offsetof(ErtsMagicBinary,u.aligned.data) - offsetof(Binary,orig_bytes))
-#define ERTS_MAGIC_BIN_ORIG_SIZE(Sz) \
-  (ERTS_MAGIC_DATA_OFFSET + (Sz))
-#define ERTS_MAGIC_BIN_SIZE(Sz) \
-  (offsetof(ErtsMagicBinary,u.aligned.data) + (Sz))
-
-/* On 32-bit arch these macro variants will save memory
-   by not forcing 8-byte alignment for the magic payload.
-*/
-#define ERTS_MAGIC_BIN_UNALIGNED_DATA(BP) \
-  ((void *) ((ErtsBinary *) (BP))->magic_binary.u.unaligned.data)
-#define ERTS_MAGIC_UNALIGNED_DATA_OFFSET \
-  (offsetof(ErtsMagicBinary,u.unaligned.data) - offsetof(Binary,orig_bytes))
-#define ERTS_MAGIC_BIN_UNALIGNED_DATA_SIZE(BP) \
-  ((BP)->orig_size - ERTS_MAGIC_UNALIGNED_DATA_OFFSET)
-#define ERTS_MAGIC_BIN_UNALIGNED_ORIG_SIZE(Sz) \
-  (ERTS_MAGIC_UNALIGNED_DATA_OFFSET + (Sz))
-#define ERTS_MAGIC_BIN_UNALIGNED_SIZE(Sz) \
-  (offsetof(ErtsMagicBinary,u.unaligned.data) + (Sz))
-#define ERTS_MAGIC_BIN_FROM_UNALIGNED_DATA(DATA) \
-  ((ErtsBinary*)((char*)(DATA) - offsetof(ErtsMagicBinary,u.unaligned.data)))
-
-
-#define Binary2ErlDrvBinary(B) (&((ErtsBinary *) (B))->driver.binary)
-#define ErlDrvBinary2Binary(D) ((Binary *) \
-				(((char *) (D)) \
-				 - offsetof(ErtsBinary, driver.binary)))
-
-/* A "magic" binary flag */
-#define BIN_FLAG_MAGIC      1
-#define BIN_FLAG_USR1       2 /* Reserved for use by different modules too mark */
-#define BIN_FLAG_USR2       4 /*  certain binaries as special (used by ets) */
-#define BIN_FLAG_DRV        8
-
-/*
  * This structure represents one type of a binary in a process.
  */
 
@@ -383,46 +283,12 @@ typedef struct proc_bin {
  */
 #define PROC_BIN_SIZE (sizeof(ProcBin)/sizeof(Eterm))
 
-ERTS_GLB_INLINE Eterm erts_mk_magic_binary_term(Eterm **hpp,
-						ErlOffHeap *ohp,
-						Binary *mbp);
-
-#if ERTS_GLB_INLINE_INCL_FUNC_DEF
-
-ERTS_GLB_INLINE Eterm
-erts_mk_magic_binary_term(Eterm **hpp, ErlOffHeap *ohp, Binary *mbp)
-{
-    ProcBin *pb = (ProcBin *) *hpp;
-    *hpp += PROC_BIN_SIZE;
-
-    ASSERT(mbp->flags & BIN_FLAG_MAGIC);
-
-    pb->thing_word = HEADER_PROC_BIN;
-    pb->size = 0;
-    pb->next = ohp->first;
-    ohp->first = (struct erl_off_heap_header*) pb;
-    pb->val = mbp;
-    pb->bytes = (byte *) mbp->orig_bytes;
-    pb->flags = 0;
-
-    erts_refc_inc(&mbp->refc, 1);
-
-    return make_binary(pb);    
-}
-
-#endif
-
-#define ERTS_TERM_IS_MAGIC_BINARY(T) \
-  (is_binary((T)) \
-   && (thing_subtag(*binary_val((T))) == REFC_BINARY_SUBTAG) \
-   && (((ProcBin *) binary_val((T)))->val->flags & BIN_FLAG_MAGIC))
-
-
 union erl_off_heap_ptr {
     struct erl_off_heap_header* hdr;
     ProcBin *pb;
     struct erl_fun_thing* fun;
     struct external_thing_* ext;
+    ErtsMRefThing *mref;
     Eterm* ep;
     void* voidp;
 };
@@ -1015,21 +881,6 @@ void erts_bif_info_init(void);
 
 /* bif.c */
 
-ERTS_GLB_INLINE Eterm
-erts_proc_store_ref(Process *c_p, Uint32 ref[ERTS_MAX_REF_NUMBERS]);
-
-#if ERTS_GLB_INLINE_INCL_FUNC_DEF
-
-ERTS_GLB_INLINE Eterm
-erts_proc_store_ref(Process *c_p, Uint32 ref[ERTS_MAX_REF_NUMBERS])
-{
-    Eterm *hp = HAlloc(c_p, REF_THING_SIZE);
-    write_ref_thing(hp, ref[0], ref[1], ref[2]);
-    return make_internal_ref(hp);
-}
-
-#endif
-
 void erts_queue_monitor_message(Process *,
 				ErtsProcLocks*,
 				Eterm,
@@ -1114,6 +965,9 @@ void print_process_info(fmtfn_t, void *, Process*);
 void info(fmtfn_t, void *);
 void loaded(fmtfn_t, void *);
 
+/* sighandler sys.c */
+int erts_set_signal(Eterm signal, Eterm type);
+
 /* erl_arith.c */
 double erts_get_positive_zero_float(void);
 
@@ -1418,6 +1272,7 @@ int erts_utf8_to_latin1(byte* dest, const byte* source, int slen);
 
 void bin_write(fmtfn_t, void*, byte*, size_t);
 Sint intlist_to_buf(Eterm, char*, Sint); /* most callers pass plain char*'s */
+Sint erts_unicode_list_to_buf(Eterm list, byte *buf, Sint len);
 
 struct Sint_buf {
 #if defined(ARCH_64)
diff --git a/erts/emulator/beam/io.c b/erts/emulator/beam/io.c
index 7fa0280396..9b525cc100 100644
--- a/erts/emulator/beam/io.c
+++ b/erts/emulator/beam/io.c
@@ -1445,7 +1445,7 @@ finalize_force_imm_drv_call(ErtsTryImmDrvCallState *sp)
     erts_unblock_fpe(sp->fpe_was_unmasked);
 }
 
-#define ERTS_QUEUE_PORT_SCHED_OP_REPLY_SIZE (REF_THING_SIZE + 3)
+#define ERTS_QUEUE_PORT_SCHED_OP_REPLY_SIZE (ERTS_REF_THING_SIZE + 3)
 
 static ERTS_INLINE void
 queue_port_sched_op_reply(Process *rp,
@@ -1460,7 +1460,7 @@ queue_port_sched_op_reply(Process *rp,
 
     ref= make_internal_ref(hp);
     write_ref_thing(hp, ref_num[0], ref_num[1], ref_num[2]);
-    hp += REF_THING_SIZE;
+    hp += ERTS_REF_THING_SIZE;
 
     msg = TUPLE2(hp, ref, msg);
 
@@ -3099,7 +3099,7 @@ port_monitor(Port *prt, erts_aint32_t state, Eterm origin,
 
     ASSERT(is_pid(origin));
     ASSERT(is_atom(name) || is_port(name) || name == NIL);
-    ASSERT(is_internal_ref(ref));
+    ASSERT(is_internal_ordinary_ref(ref));
 
     if (!(state & ERTS_PORT_SFLGS_INVALID_LOOKUP)) {
         ErtsProcLocks p_locks = ERTS_PROC_LOCK_LINK;
@@ -3124,7 +3124,7 @@ static int
 port_sig_monitor(Port *prt, erts_aint32_t state, int op,
                  ErtsProc2PortSigData *sigdp)
 {
-    Eterm hp[REF_THING_SIZE];
+    Eterm hp[ERTS_REF_THING_SIZE];
     Eterm ref = make_internal_ref(&hp);
     write_ref_thing(hp, sigdp->ref[0], sigdp->ref[1], sigdp->ref[2]);
 
@@ -3245,7 +3245,7 @@ static int
 port_sig_demonitor(Port *prt, erts_aint32_t state, int op,
                    ErtsProc2PortSigData *sigdp)
 {
-    Eterm hp[REF_THING_SIZE];
+    Eterm hp[ERTS_REF_THING_SIZE];
     Eterm ref = make_internal_ref(&hp);
     write_ref_thing(hp, sigdp->u.demonitor.ref[0],
                     sigdp->u.demonitor.ref[1],
@@ -3302,10 +3302,10 @@ ErtsPortOpResult erts_port_demonitor(Process *origin, ErtsDemonitorMode mode,
     sigdp->u.demonitor.origin = origin->common.id;
     sigdp->u.demonitor.name = target->common.id;
     {
-        RefThing *reft = ref_thing_ptr(ref);
+	Uint32 *nums = internal_ref_numbers(ref);
         /* Start from 1 skip ref arity */
         sys_memcpy(sigdp->u.demonitor.ref,
-                   internal_thing_ref_numbers(reft),
+		   nums,
                    sizeof(sigdp->u.demonitor.ref));
     }
 
@@ -3431,7 +3431,7 @@ void erts_init_io(int port_tab_size,
 			 NULL,
 			 (ErtsPTabElementCommon *) &erts_invalid_port.common,
 			 port_tab_size,
-			 common_element_size, /* Doesn't need to be excact */
+			 common_element_size, /* Doesn't need to be exact */
 			 "port_table",
 			 legacy_port_tab,
 			 1);
@@ -5413,7 +5413,7 @@ reply_io_bytes(void *vreq)
 	    rp_locks = ERTS_PROC_LOCK_MAIN;
 	}
 
-	hsz = 5 /* 4-tuple */ + REF_THING_SIZE;
+	hsz = 5 /* 4-tuple */ + ERTS_REF_THING_SIZE;
 
 	erts_bld_uint64(NULL, &hsz, in);
 	erts_bld_uint64(NULL, &hsz, out);
@@ -5422,7 +5422,7 @@ reply_io_bytes(void *vreq)
 
 	ref = make_internal_ref(hp);
 	write_ref_thing(hp, req->refn[0], req->refn[1], req->refn[2]);
-	hp += REF_THING_SIZE;
+	hp += ERTS_REF_THING_SIZE;
 
 	ein = erts_bld_uint64(&hp, NULL, in);
 	eout = erts_bld_uint64(&hp, NULL, out);
@@ -5451,7 +5451,7 @@ erts_request_io_bytes(Process *c_p)
     ErtsIOBytesReq *req = erts_alloc(ERTS_ALC_T_IOB_REQ,
 				     sizeof(ErtsIOBytesReq));
 
-    hp = HAlloc(c_p, REF_THING_SIZE);
+    hp = HAlloc(c_p, ERTS_REF_THING_SIZE);
     ref = erts_sched_make_ref_in_buffer(esdp, hp);
     refn = internal_ref_numbers(ref);
 
@@ -6888,12 +6888,6 @@ ErlDrvSizeT driver_vec_to_buf(ErlIOVec *vec, char *buf, ErlDrvSizeT len)
     return (orig_len - len);
 }
 
-
-/*
- * - driver_alloc_binary() is thread safe (efile driver depend on it).
- * - driver_realloc_binary(), and driver_free_binary() are *not* thread safe.
- */
-
 /*
  * reference count on driver binaries...
  */
@@ -6936,26 +6930,15 @@ driver_alloc_binary(ErlDrvSizeT size)
     return Binary2ErlDrvBinary(bin);
 }
 
-/* Reallocate space hold by binary */
+/* Reallocate space held by binary */
 
 ErlDrvBinary* driver_realloc_binary(ErlDrvBinary* bin, ErlDrvSizeT size)
 {
     Binary* oldbin;
     Binary* newbin;
 
-    if (!bin) {
-	erts_dsprintf_buf_t *dsbufp = erts_create_logger_dsbuf();
-	erts_dsprintf(dsbufp,
-		      "Bad use of driver_realloc_binary(%p, %lu): "
-		      "called with ",
-		      bin, (unsigned long)size);
-	if (!bin) {
-	    erts_dsprintf(dsbufp, "NULL pointer as first argument");
-	}
-	erts_send_warning_to_logger_nogl(dsbufp);
-	if (!bin)
-	    return driver_alloc_binary(size);
-    }
+    if (!bin)
+	return driver_alloc_binary(size);
 
     oldbin = ErlDrvBinary2Binary(bin);
     newbin = (Binary *) erts_bin_realloc_fnf(oldbin, size);
@@ -6969,14 +6952,8 @@ ErlDrvBinary* driver_realloc_binary(ErlDrvBinary* bin, ErlDrvSizeT size)
 void driver_free_binary(ErlDrvBinary* dbin)
 {
     Binary *bin;
-    if (!dbin) {
-	erts_dsprintf_buf_t *dsbufp = erts_create_logger_dsbuf();
-	erts_dsprintf(dsbufp,
-		      "Bad use of driver_free_binary(%p): called with "
-		      "NULL pointer as argument", dbin);
-	erts_send_warning_to_logger_nogl(dsbufp);
+    if (!dbin)
 	return;
-    }
 
     bin = ErlDrvBinary2Binary(dbin);
     if (erts_refc_dectest(&bin->refc, 0) == 0)
@@ -7632,20 +7609,27 @@ erl_drv_convert_time_unit(ErlDrvTime val,
 
 void erts_ref_to_driver_monitor(Eterm ref, ErlDrvMonitor *mon)
 {
-    RefThing *refp;
-    ASSERT(is_internal_ref(ref));
-    ERTS_CT_ASSERT(sizeof(RefThing) <= sizeof(ErlDrvMonitor));
-    refp = ref_thing_ptr(ref);
-    memset(mon,0,sizeof(ErlDrvMonitor));
-    memcpy(mon,refp,sizeof(RefThing));
+    ERTS_CT_ASSERT(ERTS_REF_THING_SIZE*sizeof(Uint) <= sizeof(ErlDrvMonitor));
+    ASSERT(is_internal_ordinary_ref(ref));
+    sys_memcpy((void *) mon, (void *) internal_ref_val(ref),
+               ERTS_REF_THING_SIZE*sizeof(Uint));
 }
 
+static Eterm driver_monitor_to_ref(Eterm *hp, const ErlDrvMonitor *mon)
+{
+    Eterm ref;
+    ERTS_CT_ASSERT(ERTS_REF_THING_SIZE*sizeof(Uint) <= sizeof(ErlDrvMonitor));
+    sys_memcpy((void *) hp, (void *) mon, ERTS_REF_THING_SIZE*sizeof(Uint));
+    ref = make_internal_ref(hp);
+    ASSERT(is_internal_ordinary_ref(ref));
+    return ref;
+}
 
 static int do_driver_monitor_process(Port *prt,
-				     Eterm *buf,
 				     ErlDrvTermData process,
 				     ErlDrvMonitor *monitor)
 {
+    Eterm buf[ERTS_REF_THING_SIZE];
     Process *rp;
     Eterm ref;
 
@@ -7688,26 +7672,21 @@ int driver_monitor_process(ErlDrvPort drvport,
     /* Now (in SMP) we should have either the port lock (if we have a scheduler) or the port data lock
        (if we're a driver thread) */
     ERTS_SMP_LC_ASSERT((sched != NULL || prt->port_data_lock));
-    {
-	DeclareTmpHeapNoproc(buf,REF_THING_SIZE);
-	UseTmpHeapNoproc(REF_THING_SIZE);
-	ret = do_driver_monitor_process(prt,buf,process,monitor);
-	UnUseTmpHeapNoproc(REF_THING_SIZE);
-    }
+    ret = do_driver_monitor_process(prt,process,monitor);
     DRV_MONITOR_UNLOCK_PDL(prt);
     return ret;
 }
 
-static int do_driver_demonitor_process(Port *prt, Eterm *buf,
-				       const ErlDrvMonitor *monitor)
+static int do_driver_demonitor_process(Port *prt, const ErlDrvMonitor *monitor)
 {
+    Eterm heap[ERTS_REF_THING_SIZE];
     Process *rp;
     Eterm ref;
     ErtsMonitor *mon;
     Eterm to;
 
-    memcpy(buf,monitor,sizeof(Eterm)*REF_THING_SIZE);
-    ref = make_internal_ref(buf);
+    ref = driver_monitor_to_ref(heap, monitor);
+
     mon = erts_lookup_monitor(ERTS_P_MONITORS(prt), ref);
     if (mon == NULL) {
 	return 1;
@@ -7751,25 +7730,20 @@ int driver_demonitor_process(ErlDrvPort drvport,
     /* Now we should have either the port lock (if we have a scheduler) or the port data lock
        (if we're a driver thread) */
     ERTS_SMP_LC_ASSERT((sched != NULL || prt->port_data_lock));
-    {
-	DeclareTmpHeapNoproc(buf,REF_THING_SIZE);
-	UseTmpHeapNoproc(REF_THING_SIZE);
-	ret = do_driver_demonitor_process(prt,buf,monitor);
-	UnUseTmpHeapNoproc(REF_THING_SIZE);
-    }
+    ret = do_driver_demonitor_process(prt,monitor);
     DRV_MONITOR_UNLOCK_PDL(prt);
     return ret;
 }
 
-static ErlDrvTermData do_driver_get_monitored_process(Port *prt, Eterm *buf,
-					    const ErlDrvMonitor *monitor)
+static ErlDrvTermData do_driver_get_monitored_process(Port *prt,const ErlDrvMonitor *monitor)
 {
     Eterm ref;
     ErtsMonitor *mon;
     Eterm to;
+    Eterm heap[ERTS_REF_THING_SIZE];
+
+    ref = driver_monitor_to_ref(heap, monitor);
 
-    memcpy(buf,monitor,sizeof(Eterm)*REF_THING_SIZE);
-    ref = make_internal_ref(buf);
     mon = erts_lookup_monitor(ERTS_P_MONITORS(prt), ref);
     if (mon == NULL) {
 	return driver_term_nil;
@@ -7797,21 +7771,16 @@ ErlDrvTermData driver_get_monitored_process(ErlDrvPort drvport,
     /* Now we should have either the port lock (if we have a scheduler) or the port data lock
        (if we're a driver thread) */
     ERTS_SMP_LC_ASSERT((sched != NULL || prt->port_data_lock));
-    {
-	DeclareTmpHeapNoproc(buf,REF_THING_SIZE);
-	UseTmpHeapNoproc(REF_THING_SIZE);
-	ret = do_driver_get_monitored_process(prt,buf,monitor);
-	UnUseTmpHeapNoproc(REF_THING_SIZE);
-    }
+    ret = do_driver_get_monitored_process(prt,monitor);
     DRV_MONITOR_UNLOCK_PDL(prt);
     return ret;
 }
 
-
 int driver_compare_monitors(const ErlDrvMonitor *monitor1,
 			    const ErlDrvMonitor *monitor2)
 {
-    return memcmp(monitor1,monitor2,sizeof(ErlDrvMonitor));
+    return sys_memcmp((void *) monitor1, (void *) monitor2,
+                      ERTS_REF_THING_SIZE*sizeof(Eterm));
 }
 
 void erts_fire_port_monitor(Port *prt, Eterm ref)
diff --git a/erts/emulator/beam/register.c b/erts/emulator/beam/register.c
index acda51c9fc..7f60710124 100644
--- a/erts/emulator/beam/register.c
+++ b/erts/emulator/beam/register.c
@@ -272,13 +272,13 @@ erts_whereis_name_to_id(Process *c_p, Eterm name)
     int ix;
     HashBucket* b;
 #ifdef ERTS_SMP
-    ErtsProcLocks c_p_locks = c_p ? ERTS_PROC_LOCK_MAIN : 0;
-
-#ifdef ERTS_ENABLE_LOCK_CHECK
-    if (c_p) ERTS_SMP_CHK_HAVE_ONLY_MAIN_PROC_LOCK(c_p);
-#endif
-
+    ErtsProcLocks c_p_locks = 0;
+    if (c_p) {
+        c_p_locks = ERTS_PROC_LOCK_MAIN;
+        ERTS_SMP_CHK_HAVE_ONLY_MAIN_PROC_LOCK(c_p);
+    }
     reg_safe_read_lock(c_p, &c_p_locks);
+
     if (c_p && !c_p_locks)
         erts_smp_proc_lock(c_p, ERTS_PROC_LOCK_MAIN);
 #endif
diff --git a/erts/emulator/beam/sys.h b/erts/emulator/beam/sys.h
index 4b3ac594a0..c6ea8049c3 100644
--- a/erts/emulator/beam/sys.h
+++ b/erts/emulator/beam/sys.h
@@ -487,20 +487,6 @@ extern volatile int erts_break_requested;
 void erts_do_break_handling(void);
 #endif
 
-#ifdef ERTS_WANT_GOT_SIGUSR1
-#  ifndef UNIX
-#    define ERTS_GOT_SIGUSR1 0
-#  else
-#    ifdef ERTS_SMP
-extern erts_smp_atomic32_t erts_got_sigusr1;
-#      define ERTS_GOT_SIGUSR1 ((int) erts_smp_atomic32_read_mb(&erts_got_sigusr1))
-#    else
-extern volatile int erts_got_sigusr1;
-#      define ERTS_GOT_SIGUSR1 erts_got_sigusr1
-#    endif
-#  endif
-#endif
-
 #ifdef ERTS_SMP
 extern erts_smp_atomic32_t erts_writing_erl_crash_dump;
 extern erts_tsd_key_t erts_is_crash_dumping_key;
@@ -865,9 +851,12 @@ int erts_sys_unsetenv(char *key);
 char *erts_read_env(char *key);
 void erts_free_read_env(void *value);
 
+#if defined(ERTS_SMP)
 #if defined(ERTS_THR_HAVE_SIG_FUNCS) && !defined(ETHR_UNUSABLE_SIGUSRX)
 extern void sys_thr_resume(erts_tid_t tid);
 extern void sys_thr_suspend(erts_tid_t tid);
+#define ERTS_SYS_SUSPEND_SIGNAL SIGUSR2
+#endif
 #endif
 
 /* utils.c */
@@ -900,10 +889,13 @@ void sys_alloc_stat(SysAllocStat *);
 #define ERTS_REFC_DEBUG
 #endif
 
-typedef erts_smp_atomic_t erts_refc_t;
+typedef erts_atomic_t erts_refc_t;
 
 ERTS_GLB_INLINE void erts_refc_init(erts_refc_t *refcp, erts_aint_t val);
 ERTS_GLB_INLINE void erts_refc_inc(erts_refc_t *refcp, erts_aint_t min_val);
+ERTS_GLB_INLINE erts_aint_t erts_refc_inc_unless(erts_refc_t *refcp,
+                                                 erts_aint_t unless_val,
+                                                 erts_aint_t min_val);
 ERTS_GLB_INLINE erts_aint_t erts_refc_inctest(erts_refc_t *refcp,
 					      erts_aint_t min_val);
 ERTS_GLB_INLINE void erts_refc_dec(erts_refc_t *refcp, erts_aint_t min_val);
@@ -919,27 +911,51 @@ ERTS_GLB_INLINE erts_aint_t erts_refc_read(erts_refc_t *refcp,
 ERTS_GLB_INLINE void
 erts_refc_init(erts_refc_t *refcp, erts_aint_t val)
 {
-    erts_smp_atomic_init_nob((erts_smp_atomic_t *) refcp, val);
+    erts_atomic_init_nob((erts_atomic_t *) refcp, val);
 }
 
 ERTS_GLB_INLINE void
 erts_refc_inc(erts_refc_t *refcp, erts_aint_t min_val)
 {
 #ifdef ERTS_REFC_DEBUG
-    erts_aint_t val = erts_smp_atomic_inc_read_nob((erts_smp_atomic_t *) refcp);
+    erts_aint_t val = erts_atomic_inc_read_nob((erts_atomic_t *) refcp);
     if (val < min_val)
 	erts_exit(ERTS_ABORT_EXIT,
 		 "erts_refc_inc(): Bad refc found (refc=%ld < %ld)!\n",
 		 val, min_val);
 #else
-    erts_smp_atomic_inc_nob((erts_smp_atomic_t *) refcp);
+    erts_atomic_inc_nob((erts_atomic_t *) refcp);
 #endif
 }
 
 ERTS_GLB_INLINE erts_aint_t
+erts_refc_inc_unless(erts_refc_t *refcp,
+                     erts_aint_t unless_val,
+                     erts_aint_t min_val)
+{
+    erts_aint_t val = erts_atomic_read_nob((erts_atomic_t *) refcp);
+    while (1) {
+        erts_aint_t exp, new;
+#ifdef ERTS_REFC_DEBUG
+        if (val < 0)
+            erts_exit(ERTS_ABORT_EXIT,
+                      "erts_refc_inc_unless(): Bad refc found (refc=%ld < %ld)!\n",
+                      val, min_val);
+#endif
+        if (val == unless_val)
+            return val;
+        new = val + 1;
+        exp = val;
+        val = erts_atomic_cmpxchg_nob((erts_atomic_t *) refcp, new, exp);
+        if (val == exp)
+            return new;
+    }
+}
+
+ERTS_GLB_INLINE erts_aint_t
 erts_refc_inctest(erts_refc_t *refcp, erts_aint_t min_val)
 {
-    erts_aint_t val = erts_smp_atomic_inc_read_nob((erts_smp_atomic_t *) refcp);
+    erts_aint_t val = erts_atomic_inc_read_nob((erts_atomic_t *) refcp);
 #ifdef ERTS_REFC_DEBUG
     if (val < min_val)
 	erts_exit(ERTS_ABORT_EXIT,
@@ -953,20 +969,20 @@ ERTS_GLB_INLINE void
 erts_refc_dec(erts_refc_t *refcp, erts_aint_t min_val)
 {
 #ifdef ERTS_REFC_DEBUG
-    erts_aint_t val = erts_smp_atomic_dec_read_nob((erts_smp_atomic_t *) refcp);
+    erts_aint_t val = erts_atomic_dec_read_nob((erts_atomic_t *) refcp);
     if (val < min_val)
 	erts_exit(ERTS_ABORT_EXIT,
 		 "erts_refc_dec(): Bad refc found (refc=%ld < %ld)!\n",
 		 val, min_val);
 #else
-    erts_smp_atomic_dec_nob((erts_smp_atomic_t *) refcp);
+    erts_atomic_dec_nob((erts_atomic_t *) refcp);
 #endif
 }
 
 ERTS_GLB_INLINE erts_aint_t
 erts_refc_dectest(erts_refc_t *refcp, erts_aint_t min_val)
 {
-    erts_aint_t val = erts_smp_atomic_dec_read_nob((erts_smp_atomic_t *) refcp);
+    erts_aint_t val = erts_atomic_dec_read_nob((erts_atomic_t *) refcp);
 #ifdef ERTS_REFC_DEBUG
     if (val < min_val)
 	erts_exit(ERTS_ABORT_EXIT,
@@ -980,20 +996,20 @@ ERTS_GLB_INLINE void
 erts_refc_add(erts_refc_t *refcp, erts_aint_t diff, erts_aint_t min_val)
 {
 #ifdef ERTS_REFC_DEBUG
-    erts_aint_t val = erts_smp_atomic_add_read_nob((erts_smp_atomic_t *) refcp, diff);
+    erts_aint_t val = erts_atomic_add_read_nob((erts_atomic_t *) refcp, diff);
     if (val < min_val)
 	erts_exit(ERTS_ABORT_EXIT,
 		 "erts_refc_add(%ld): Bad refc found (refc=%ld < %ld)!\n",
 		 diff, val, min_val);
 #else
-    erts_smp_atomic_add_nob((erts_smp_atomic_t *) refcp, diff);
+    erts_atomic_add_nob((erts_atomic_t *) refcp, diff);
 #endif
 }
 
 ERTS_GLB_INLINE erts_aint_t
 erts_refc_read(erts_refc_t *refcp, erts_aint_t min_val)
 {
-    erts_aint_t val = erts_smp_atomic_read_nob((erts_smp_atomic_t *) refcp);
+    erts_aint_t val = erts_atomic_read_nob((erts_atomic_t *) refcp);
 #ifdef ERTS_REFC_DEBUG
     if (val < min_val)
 	erts_exit(ERTS_ABORT_EXIT,
@@ -1005,6 +1021,140 @@ erts_refc_read(erts_refc_t *refcp, erts_aint_t min_val)
 
 #endif /* #if ERTS_GLB_INLINE_INCL_FUNC_DEF */
 
+typedef erts_smp_atomic_t erts_smp_refc_t;
+
+ERTS_GLB_INLINE void erts_smp_refc_init(erts_smp_refc_t *refcp, erts_aint_t val);
+ERTS_GLB_INLINE void erts_smp_refc_inc(erts_smp_refc_t *refcp, erts_aint_t min_val);
+ERTS_GLB_INLINE erts_aint_t erts_smp_refc_inc_unless(erts_smp_refc_t *refcp,
+                                                     erts_aint_t unless_val,
+                                                     erts_aint_t min_val);
+ERTS_GLB_INLINE erts_aint_t erts_smp_refc_inctest(erts_smp_refc_t *refcp,
+                                                  erts_aint_t min_val);
+ERTS_GLB_INLINE void erts_smp_refc_dec(erts_smp_refc_t *refcp, erts_aint_t min_val);
+ERTS_GLB_INLINE erts_aint_t erts_smp_refc_dectest(erts_smp_refc_t *refcp,
+                                                  erts_aint_t min_val);
+ERTS_GLB_INLINE void erts_smp_refc_add(erts_smp_refc_t *refcp, erts_aint_t diff,
+                                       erts_aint_t min_val);
+ERTS_GLB_INLINE erts_aint_t erts_smp_refc_read(erts_smp_refc_t *refcp,
+                                               erts_aint_t min_val);
+
+#if ERTS_GLB_INLINE_INCL_FUNC_DEF
+
+ERTS_GLB_INLINE void
+erts_smp_refc_init(erts_smp_refc_t *refcp, erts_aint_t val)
+{
+    erts_smp_atomic_init_nob((erts_smp_atomic_t *) refcp, val);
+}
+
+ERTS_GLB_INLINE void
+erts_smp_refc_inc(erts_smp_refc_t *refcp, erts_aint_t min_val)
+{
+#ifdef ERTS_REFC_DEBUG
+    erts_aint_t val = erts_smp_atomic_inc_read_nob((erts_smp_atomic_t *) refcp);
+    if (val < min_val)
+	erts_exit(ERTS_ABORT_EXIT,
+		 "erts_smp_refc_inc(): Bad refc found (refc=%ld < %ld)!\n",
+		 val, min_val);
+#else
+    erts_smp_atomic_inc_nob((erts_smp_atomic_t *) refcp);
+#endif
+}
+
+ERTS_GLB_INLINE erts_aint_t
+erts_smp_refc_inc_unless(erts_smp_refc_t *refcp,
+                         erts_aint_t unless_val,
+                         erts_aint_t min_val)
+{
+    erts_aint_t val = erts_smp_atomic_read_nob((erts_smp_atomic_t *) refcp);
+    while (1) {
+        erts_aint_t exp, new;
+#ifdef ERTS_REFC_DEBUG
+        if (val < 0)
+            erts_exit(ERTS_ABORT_EXIT,
+                      "erts_smp_refc_inc_unless(): Bad refc found (refc=%ld < %ld)!\n",
+                      val, min_val);
+#endif
+        if (val == unless_val)
+            return val;
+        new = val + 1;
+        exp = val;
+        val = erts_smp_atomic_cmpxchg_nob((erts_smp_atomic_t *) refcp, new, exp);
+        if (val == exp)
+            return new;
+    }
+}
+
+
+ERTS_GLB_INLINE erts_aint_t
+erts_smp_refc_inctest(erts_smp_refc_t *refcp, erts_aint_t min_val)
+{
+    erts_aint_t val = erts_smp_atomic_inc_read_nob((erts_smp_atomic_t *) refcp);
+#ifdef ERTS_REFC_DEBUG
+    if (val < min_val)
+	erts_exit(ERTS_ABORT_EXIT,
+		 "erts_smp_refc_inctest(): Bad refc found (refc=%ld < %ld)!\n",
+		 val, min_val);
+#endif
+    return val;
+}
+
+ERTS_GLB_INLINE void
+erts_smp_refc_dec(erts_smp_refc_t *refcp, erts_aint_t min_val)
+{
+#ifdef ERTS_REFC_DEBUG
+    erts_aint_t val = erts_smp_atomic_dec_read_nob((erts_smp_atomic_t *) refcp);
+    if (val < min_val)
+	erts_exit(ERTS_ABORT_EXIT,
+		 "erts_smp_refc_dec(): Bad refc found (refc=%ld < %ld)!\n",
+		 val, min_val);
+#else
+    erts_smp_atomic_dec_nob((erts_smp_atomic_t *) refcp);
+#endif
+}
+
+ERTS_GLB_INLINE erts_aint_t
+erts_smp_refc_dectest(erts_smp_refc_t *refcp, erts_aint_t min_val)
+{
+    erts_aint_t val = erts_smp_atomic_dec_read_nob((erts_smp_atomic_t *) refcp);
+#ifdef ERTS_REFC_DEBUG
+    if (val < min_val)
+	erts_exit(ERTS_ABORT_EXIT,
+		 "erts_smp_refc_dectest(): Bad refc found (refc=%ld < %ld)!\n",
+		 val, min_val);
+#endif
+    return val;
+}
+
+ERTS_GLB_INLINE void
+erts_smp_refc_add(erts_smp_refc_t *refcp, erts_aint_t diff, erts_aint_t min_val)
+{
+#ifdef ERTS_REFC_DEBUG
+    erts_aint_t val = erts_smp_atomic_add_read_nob((erts_smp_atomic_t *) refcp, diff);
+    if (val < min_val)
+	erts_exit(ERTS_ABORT_EXIT,
+		 "erts_smp_refc_add(%ld): Bad refc found (refc=%ld < %ld)!\n",
+		 diff, val, min_val);
+#else
+    erts_smp_atomic_add_nob((erts_smp_atomic_t *) refcp, diff);
+#endif
+}
+
+ERTS_GLB_INLINE erts_aint_t
+erts_smp_refc_read(erts_smp_refc_t *refcp, erts_aint_t min_val)
+{
+    erts_aint_t val = erts_smp_atomic_read_nob((erts_smp_atomic_t *) refcp);
+#ifdef ERTS_REFC_DEBUG
+    if (val < min_val)
+	erts_exit(ERTS_ABORT_EXIT,
+		 "erts_smp_refc_read(): Bad refc found (refc=%ld < %ld)!\n",
+		 val, min_val);
+#endif
+    return val;
+}
+
+#endif /* #if ERTS_GLB_INLINE_INCL_FUNC_DEF */
+
+
 #ifdef ERTS_ENABLE_KERNEL_POLL
 extern int erts_use_kernel_poll;
 #endif
diff --git a/erts/emulator/beam/utils.c b/erts/emulator/beam/utils.c
index 3fa48da1ec..092a5320ba 100644
--- a/erts/emulator/beam/utils.c
+++ b/erts/emulator/beam/utils.c
@@ -700,12 +700,7 @@ erts_bld_atom_2uint_3tup_list(Uint **hpp, Uint *szp, Sint length,
 /* make a hash index from an erlang term */
 
 /*
-** There are three hash functions.
-** make_broken_hash: the one used for backward compatibility
-** is called from the bif erlang:hash/2. Should never be used
-** as it a) hashes only a part of binaries, b) hashes bignums really poorly,
-** c) hashes bignums differently on different endian processors and d) hashes 
-** small integers with different weights on different bytes.
+** There are two hash functions.
 **
 ** make_hash: A hash function that will give the same values for the same
 ** terms regardless of the internal representation. Small integers are 
@@ -1045,6 +1040,10 @@ tail_recur:
     DESTROY_WSTACK(stack);
     return hash;
 
+#undef MAKE_HASH_TUPLE_OP
+#undef MAKE_HASH_TERM_ARRAY_OP
+#undef MAKE_HASH_CDR_PRE_OP
+#undef MAKE_HASH_CDR_POST_OP
 #undef UINT32_HASH_STEP
 #undef UINT32_HASH_RET
 }
@@ -1839,7 +1838,7 @@ make_internal_hash(Eterm term)
 	    break;
 	    case REF_SUBTAG:
 		UINT32_HASH(internal_ref_numbers(term)[0], HCONST_7);
-                ASSERT(internal_ref_no_of_numbers(term) == 3);
+                ASSERT(internal_ref_no_numbers(term) == 3);
                 UINT32_HASH_2(internal_ref_numbers(term)[1],
                               internal_ref_numbers(term)[2], HCONST_8);
                 goto pop_next;
@@ -1848,7 +1847,7 @@ make_internal_hash(Eterm term)
             {
                 ExternalThing* thing = external_thing_ptr(term);
 
-                ASSERT(external_thing_ref_no_of_numbers(thing) == 3);
+                ASSERT(external_thing_ref_no_numbers(thing) == 3);
                 /* See limitation #2 */
             #ifdef ARCH_64
                 POINTER_HASH(thing->node, HCONST_7);
@@ -1954,259 +1953,6 @@ make_internal_hash(Eterm term)
 #undef HCONST
 #undef MIX
 
-
-Uint32 make_broken_hash(Eterm term)
-{
-    Uint32 hash = 0;
-    DECLARE_WSTACK(stack);
-    unsigned op;
-tail_recur:
-    op = tag_val_def(term);
-    for (;;) {
-    switch (op) {
-    case NIL_DEF:
-	hash = hash*FUNNY_NUMBER3 + 1;
-	break;
-    case ATOM_DEF:
-	hash = hash*FUNNY_NUMBER1 +
-	    (atom_tab(atom_val(term))->slot.bucket.hvalue);
-	break;
-    case SMALL_DEF:
-#if defined(ARCH_64)
-    {
-	Sint y1 = signed_val(term);
-	Uint y2 = y1 < 0 ? -(Uint)y1 : y1;
-	Uint32 y3 = (Uint32) (y2 >> 32);
-	int arity = 1;
-
-#if defined(WORDS_BIGENDIAN)
-	if (!IS_SSMALL28(y1))
-	{   /* like a bignum */
-	    Uint32 y4 = (Uint32) y2;
-	    hash = hash*FUNNY_NUMBER2 + ((y4 << 16) | (y4 >> 16));
-	    if (y3) {
-		hash = hash*FUNNY_NUMBER2 + ((y3 << 16) | (y3 >> 16));
-		arity++;
-	    }
-	    hash = hash * (y1 < 0 ? FUNNY_NUMBER3 : FUNNY_NUMBER2) + arity;
-	} else {
-	    hash = hash*FUNNY_NUMBER2 + (((Uint) y1) & 0xfffffff);
-	}
-#else
-	if  (!IS_SSMALL28(y1))
-	{   /* like a bignum */
-	    hash = hash*FUNNY_NUMBER2 + ((Uint32) y2);
-	    if (y3)
-	    {
-		hash = hash*FUNNY_NUMBER2 + y3;
-		arity++;
-	    }
-	    hash = hash * (y1 < 0 ? FUNNY_NUMBER3 : FUNNY_NUMBER2) + arity;
-	} else {
-	    hash = hash*FUNNY_NUMBER2 + (((Uint) y1) & 0xfffffff);
-	}
-#endif
-    }
-#else
-	hash = hash*FUNNY_NUMBER2 + unsigned_val(term);
-#endif
-	break;
-
-    case BINARY_DEF:
-	{
-	    size_t sz = binary_size(term);
-	    size_t i = (sz < 15) ? sz : 15;
-
-	    hash = hash_binary_bytes(term, i, hash);
-	    hash = hash*FUNNY_NUMBER4 + sz;
-	    break;
-	}
-
-    case EXPORT_DEF:
-	{
-	    Export* ep = *((Export **) (export_val(term) + 1));
-
-	    hash = hash * FUNNY_NUMBER11 + ep->info.mfa.arity;
-	    hash = hash*FUNNY_NUMBER1 +
-		(atom_tab(atom_val(ep->info.mfa.module))->slot.bucket.hvalue);
-	    hash = hash*FUNNY_NUMBER1 +
-		(atom_tab(atom_val(ep->info.mfa.function))->slot.bucket.hvalue);
-	    break;
-	}
-
-    case FUN_DEF:
-	{
-	    ErlFunThing* funp = (ErlFunThing *) fun_val(term);
-	    Uint num_free = funp->num_free;
-
-	    hash = hash * FUNNY_NUMBER10 + num_free;
-	    hash = hash*FUNNY_NUMBER1 +
-		(atom_tab(atom_val(funp->fe->module))->slot.bucket.hvalue);
-	    hash = hash*FUNNY_NUMBER2 + funp->fe->old_index;
-	    hash = hash*FUNNY_NUMBER2 + funp->fe->old_uniq;
-	    if (num_free > 0) {
-		if (num_free > 1) {
-		    WSTACK_PUSH3(stack, (UWord) &funp->env[1], (num_free-1), MAKE_HASH_TERM_ARRAY_OP);
-		}
-		term = funp->env[0];
-		goto tail_recur;
-	    }
-	    break;
-	}
-
-    case PID_DEF:
-	hash = hash*FUNNY_NUMBER5 + internal_pid_number(term);
-	break;
-    case EXTERNAL_PID_DEF:
-	hash = hash*FUNNY_NUMBER5 + external_pid_number(term);
-        break;
-    case PORT_DEF:
-	hash = hash*FUNNY_NUMBER9 + internal_port_number(term);
-	break;
-    case EXTERNAL_PORT_DEF:
-	hash = hash*FUNNY_NUMBER9 + external_port_number(term);
-	break;
-    case REF_DEF:
-	hash = hash*FUNNY_NUMBER9 + internal_ref_numbers(term)[0];
-	break;
-    case EXTERNAL_REF_DEF:
-	hash = hash*FUNNY_NUMBER9 + external_ref_numbers(term)[0];
-	break;
-    case FLOAT_DEF:
-	{
-            FloatDef ff;
-            GET_DOUBLE(term, ff);
-            if (ff.fd == 0.0f) {
-                /* ensure positive 0.0 */
-                ff.fd = erts_get_positive_zero_float();
-            }
-            hash = hash*FUNNY_NUMBER6 + (ff.fw[0] ^ ff.fw[1]);
-	}
-	break;
-    case MAKE_HASH_CDR_PRE_OP:
-	term = (Eterm) WSTACK_POP(stack);
-	if (is_not_list(term)) {
-	    WSTACK_PUSH(stack, (UWord) MAKE_HASH_CDR_POST_OP);
-	    goto tail_recur;
-	}
-	/*fall through*/
-    case LIST_DEF:
-	{
-	    Eterm* list = list_val(term);
-	    WSTACK_PUSH2(stack, (UWord) CDR(list),
-			 (UWord) MAKE_HASH_CDR_PRE_OP);
-	    term = CAR(list);
-	    goto tail_recur;
-	}
-
-    case MAKE_HASH_CDR_POST_OP:
-	hash *= FUNNY_NUMBER8;
-	break;
-
-     case BIG_DEF:
-	{
-	    Eterm* ptr  = big_val(term);
-	    int is_neg = BIG_SIGN(ptr);
-	    Uint arity = BIG_ARITY(ptr);
-	    Uint i = arity;
-	    ptr++;
-#if D_EXP == 16
-	    /* hash over 32 bit LE */
-
-	    while(i--) {
-		hash = hash*FUNNY_NUMBER2 + *ptr++;
-	    }
-#elif D_EXP == 32
-
-#if defined(WORDS_BIGENDIAN)
-	    while(i--) {
-		Uint d = *ptr++;
-		hash = hash*FUNNY_NUMBER2 + ((d << 16) | (d >> 16));
-	    }
-#else
-	    while(i--) {
-		hash = hash*FUNNY_NUMBER2 + *ptr++;
-	    }
-#endif
-
-#elif D_EXP == 64
-	    {
-	      Uint32 h = 0, l;
-#if defined(WORDS_BIGENDIAN)
-	      while(i--) {
-		  Uint d = *ptr++;
-		  l = d & 0xffffffff;
-		  h = d >> 32;
-		  hash = hash*FUNNY_NUMBER2 + ((l << 16) | (l >> 16));
-		  if (h || i)
-		      hash = hash*FUNNY_NUMBER2 + ((h << 16) | (h >> 16));
-	      }
-#else
-	      while(i--) {
-		  Uint d = *ptr++;
-		  l = d & 0xffffffff;
-		  h = d >> 32;
-		  hash = hash*FUNNY_NUMBER2 + l;
-		  if (h || i)
-		      hash = hash*FUNNY_NUMBER2 + h;
-	      }
-#endif
-	      /* adjust arity to match 32 bit mode */
-	      arity = (arity << 1) - (h == 0);
-	    }
-
-#else
-#error "unsupported D_EXP size"
-#endif
-	    hash = hash * (is_neg ? FUNNY_NUMBER3 : FUNNY_NUMBER2) + arity;
-	}
-	break;
-
-    case MAP_DEF:
-        hash = hash*FUNNY_NUMBER13 + FUNNY_NUMBER14 + make_hash2(term);
-        break;
-    case TUPLE_DEF:
-	{
-	    Eterm* ptr = tuple_val(term);
-	    Uint arity = arityval(*ptr);
-
-	    WSTACK_PUSH3(stack, (UWord) arity, (UWord) (ptr+1), (UWord) arity);
-	    op = MAKE_HASH_TUPLE_OP;
-	}/*fall through*/
-    case MAKE_HASH_TUPLE_OP:
-    case MAKE_HASH_TERM_ARRAY_OP:
-	{
-	    Uint i = (Uint) WSTACK_POP(stack);
-	    Eterm* ptr = (Eterm*) WSTACK_POP(stack);
-	    if (i != 0) {
-		term = *ptr;
-		WSTACK_PUSH3(stack, (UWord)(ptr+1), (UWord) i-1, (UWord) op);
-		goto tail_recur;
-	    }
-	    if (op == MAKE_HASH_TUPLE_OP) {
-		Uint32 arity = (UWord) WSTACK_POP(stack);
-		hash = hash*FUNNY_NUMBER9 + arity;
-	    }
-	    break;
-	}
-
-    default:
-	erts_exit(ERTS_ERROR_EXIT, "Invalid tag in make_broken_hash\n");
-	return 0;
-      }
-      if (WSTACK_ISEMPTY(stack)) break;
-      op = (Uint) WSTACK_POP(stack);
-    }
-
-    DESTROY_WSTACK(stack);
-    return hash;
-
-#undef MAKE_HASH_TUPLE_OP
-#undef MAKE_HASH_TERM_ARRAY_OP
-#undef MAKE_HASH_CDR_PRE_OP
-#undef MAKE_HASH_CDR_POST_OP
-}
-
 static Eterm
 do_allocate_logger_message(Eterm gleader, Eterm **hp, ErlOffHeap **ohp,
 			   ErlHeapFragment **bp, Process **p, Uint sz)
@@ -2740,22 +2486,20 @@ tailrecur_ne:
 
 		anum = external_thing_ref_numbers(athing);
 		bnum = external_thing_ref_numbers(bthing);
-		alen = external_thing_ref_no_of_numbers(athing);
-		blen = external_thing_ref_no_of_numbers(bthing);
+		alen = external_thing_ref_no_numbers(athing);
+		blen = external_thing_ref_no_numbers(bthing);
 
 		goto ref_common;
+
 	    case REF_SUBTAG:
-		    if (!is_internal_ref(b))
-			goto not_equal;
 
-		    {
-			RefThing* athing = ref_thing_ptr(a);
-			RefThing* bthing = ref_thing_ptr(b);
-			alen = internal_thing_ref_no_of_numbers(athing);
-			blen = internal_thing_ref_no_of_numbers(bthing);
-			anum = internal_thing_ref_numbers(athing);
-			bnum = internal_thing_ref_numbers(bthing);
-		    }
+		if (!is_internal_ref(b))
+		    goto not_equal;
+
+		alen = internal_ref_no_numbers(a);
+		anum = internal_ref_numbers(a);
+		blen = internal_ref_no_numbers(b);
+		bnum = internal_ref_numbers(b);
 
 	    ref_common:
 		    ASSERT(alen > 0 && blen > 0);
@@ -3387,25 +3131,21 @@ tailrecur_ne:
 		 */
 
 		if (is_internal_ref(b)) {
-		    RefThing* bthing = ref_thing_ptr(b);
 		    bnode = erts_this_node;
-		    bnum = internal_thing_ref_numbers(bthing);
-		    blen = internal_thing_ref_no_of_numbers(bthing);
+		    blen = internal_ref_no_numbers(b);
+		    bnum = internal_ref_numbers(b);
 		} else if(is_external_ref(b)) {
 		    ExternalThing* bthing = external_thing_ptr(b);
 		    bnode = bthing->node;
 		    bnum = external_thing_ref_numbers(bthing);
-		    blen = external_thing_ref_no_of_numbers(bthing);
+		    blen = external_thing_ref_no_numbers(bthing);
 		} else {
 		    a_tag = REF_DEF;
 		    goto mixed_types;
 		}
-		{
-		    RefThing* athing = ref_thing_ptr(a);
-		    anode = erts_this_node;
-		    anum = internal_thing_ref_numbers(athing);
-		    alen = internal_thing_ref_no_of_numbers(athing);
-		}
+		anode = erts_this_node;
+		alen = internal_ref_no_numbers(a);
+		anum = internal_ref_numbers(a);
 
 	    ref_common:
 		CMP_NODES(anode, bnode);
@@ -3435,15 +3175,14 @@ tailrecur_ne:
 		goto pop_next;
 	    case (_TAG_HEADER_EXTERNAL_REF >> _TAG_PRIMARY_SIZE):
 		if (is_internal_ref(b)) {
-		    RefThing* bthing = ref_thing_ptr(b);
 		    bnode = erts_this_node;
-		    bnum = internal_thing_ref_numbers(bthing);
-		    blen = internal_thing_ref_no_of_numbers(bthing);
+		    blen = internal_ref_no_numbers(b);
+		    bnum = internal_ref_numbers(b);
 		} else if (is_external_ref(b)) {
 		    ExternalThing* bthing = external_thing_ptr(b);
 		    bnode = bthing->node;
 		    bnum = external_thing_ref_numbers(bthing);
-		    blen = external_thing_ref_no_of_numbers(bthing);
+		    blen = external_thing_ref_no_numbers(bthing);
 		} else {
 		    a_tag = EXTERNAL_REF_DEF;
 		    goto mixed_types;
@@ -3452,7 +3191,7 @@ tailrecur_ne:
 		    ExternalThing* athing = external_thing_ptr(a);
 		    anode = athing->node;
 		    anum = external_thing_ref_numbers(athing);
-		    alen = external_thing_ref_no_of_numbers(athing);
+		    alen = external_thing_ref_no_numbers(athing);
 		}
 		goto ref_common;
 	    default:
@@ -3829,40 +3568,41 @@ not_equal:
 Eterm
 store_external_or_ref_(Uint **hpp, ErlOffHeap* oh, Eterm ns)
 {
+    struct erl_off_heap_header *ohhp;
     Uint i;
     Uint size;
-    Uint *from_hp;
-    Uint *to_hp = *hpp;
+    Eterm *from_hp;
+    Eterm *to_hp = *hpp;
 
     ASSERT(is_external(ns) || is_internal_ref(ns));
 
-    if(is_external(ns)) {
-	from_hp = external_val(ns);
-	size = thing_arityval(*from_hp) + 1;
-	*hpp += size;
-
-	for(i = 0; i < size; i++)
-	    to_hp[i] = from_hp[i];
-
-	erts_refc_inc(&((ExternalThing *) to_hp)->node->refc, 2);
-
-	((struct erl_off_heap_header*) to_hp)->next = oh->first;
-	oh->first = (struct erl_off_heap_header*) to_hp;
-
-	return make_external(to_hp);
-    }
-
-    /* Internal ref */
-    from_hp = internal_ref_val(ns);
-
+    from_hp = boxed_val(ns);
     size = thing_arityval(*from_hp) + 1;
-
     *hpp += size;
 
     for(i = 0; i < size; i++)
 	to_hp[i] = from_hp[i];
 
-    return make_internal_ref(to_hp);
+    if (is_external_header(*from_hp)) {
+	ExternalThing *etp = (ExternalThing *) from_hp;
+	ASSERT(is_external(ns));
+	erts_smp_refc_inc(&etp->node->refc, 2);
+    }
+    else if (is_ordinary_ref_thing(from_hp))
+	return make_internal_ref(to_hp);
+    else {
+	ErtsMRefThing *mreft = (ErtsMRefThing *) from_hp;
+        ErtsMagicBinary *mb = mreft->mb;
+	ASSERT(is_magic_ref_thing(from_hp));
+	erts_refc_inc(&mb->refc, 2);
+        OH_OVERHEAD(oh, mb->orig_size / sizeof(Eterm));
+    }
+
+    ohhp = (struct erl_off_heap_header*) to_hp;
+    ohhp->next = oh->first;
+    oh->first = ohhp;
+
+    return make_boxed(to_hp);
 }
 
 Eterm
@@ -3925,6 +3665,68 @@ intlist_to_buf(Eterm list, char *buf, Sint len)
     return -2;			/* not enough space */
 }
 
+/* Fill buf with the contents of the unicode list.
+ * Return the number of bytes in the buffer,
+ * or -1 for type error,
+ * or -2 for not enough buffer space (buffer contains truncated result).
+ */
+Sint
+erts_unicode_list_to_buf(Eterm list, byte *buf, Sint len)
+{
+    Eterm* listptr;
+    Sint sz = 0;
+
+    if (is_nil(list)) {
+	return 0;
+    }
+    if (is_not_list(list)) {
+	return -1;
+    }
+    listptr = list_val(list);
+
+    while (len-- > 0) {
+	Sint val;
+
+	if (is_not_small(CAR(listptr))) {
+	    return -1;
+	}
+	val = signed_val(CAR(listptr));
+	if (0 <= val && val < 0x80) {
+	    buf[sz] = val;
+	    sz++;
+	} else if (val < 0x800) {
+	    buf[sz+0] = 0xC0 | (val >> 6);
+	    buf[sz+1] = 0x80 | (val & 0x3F);
+	    sz += 2;
+	} else if (val < 0x10000UL) {
+	    if (0xD800 <= val && val <= 0xDFFF) {
+		return -1;
+	    }
+	    buf[sz+0] = 0xE0 | (val >> 12);
+	    buf[sz+1] = 0x80 | ((val >> 6) & 0x3F);
+	    buf[sz+2] = 0x80 | (val & 0x3F);
+	    sz += 3;
+	} else if (val < 0x110000) {
+	    buf[sz+0] = 0xF0 | (val >> 18);
+	    buf[sz+1] = 0x80 | ((val >> 12) & 0x3F);
+	    buf[sz+2] = 0x80 | ((val >> 6) & 0x3F);
+	    buf[sz+3] = 0x80 | (val & 0x3F);
+	    sz += 4;
+	} else {
+	    return -1;
+	}
+	list = CDR(listptr);
+	if (is_nil(list)) {
+	    return sz;
+	}
+	if (is_not_list(list)) {
+	    return -1;
+	}
+	listptr = list_val(list);
+    }
+    return -2;			/* not enough space */
+}
+
 /*
 ** Convert an integer to a byte list
 ** return pointer to converted stuff (need not to be at start of buf!)