Merge branch 'dev' of super:otp into dev

* 'dev' of super:otp: (166 commits)
  Corrected documentation error and added examples to Users Guide
  In TLS 1.1, failure to properly close a connection no longer requires that a session not be resumed.  This is a change from TLS 1.0 to conform with widespread implementation practice.  Erlang ssl will now in TLS 1.0 conform to the widespread implementation practice instead of the specification to avoid performance issues.
  Add escript to bootstrap/bin
  Remove unused variable warning in inet_res
  Remove unused variable in epmd_port
  Remove compiler warnings in inet_drv
  Add SASL test suite
  Allow same module name in multiple applications if explicitely excluded
  Fix bugs concerning the option report_missing_types
  Fix default encoding in SAX parser.
  re: remove gratuitous "it " in manpage
  Spelling in (backward *compatibility*) comment.
  Improve erl_docgen's support for Dialyzer specs and types
  dialyzer warning on mnesia_tm
  Add documentation text about majority checking
  add mnesia_majority_test suite
  where_to_wlock optimization + change_table_majority/2
  bug in mnesia_tm:needs_majority/2
  optimize sticky_lock maj. check
  check majority for sticky locks
  ...

52 files changed, 1211 insertions, 568 deletions

diff --git a/erts/aclocal.m4 b/erts/aclocal.m4
index a1211bbf0c..f7e5c31910 100644
--- a/erts/aclocal.m4
+++ b/erts/aclocal.m4
@@ -1153,7 +1153,7 @@ case "$THR_LIB_NAME" in
 			AO_nop_full();
 			AO_store(&x, (AO_t) 0);
 			z = AO_load(&x);
-			z = AO_compare_and_swap(&x, (AO_t) 0, (AO_t) 1);
+			z = AO_compare_and_swap_full(&x, (AO_t) 0, (AO_t) 1);
 		    ],
 		    [ethr_have_native_atomics=yes
 		     ethr_have_libatomic_ops=yes])
diff --git a/erts/configure.in b/erts/configure.in
index 31d1d55b8a..fac07f8b6a 100644
--- a/erts/configure.in
+++ b/erts/configure.in
@@ -769,6 +769,7 @@ if test "$enable_halfword_emulator" = "yes"; then
 	if test "$ARCH" = "amd64"; then
 		AC_DEFINE(HALFWORD_HEAP_EMULATOR, [1],
 			  [Define if building a halfword-heap 64bit emulator])
+		ENABLE_ALLOC_TYPE_VARS="$ENABLE_ALLOC_TYPE_VARS halfword"
 		AC_MSG_RESULT([yes])
 	else
 		AC_MSG_ERROR(no; halfword emulator not supported on this architecture)
@@ -3523,6 +3524,8 @@ dnl use "PATH/include" and "PATH/lib".
 AC_SUBST(SSL_INCLUDE)
 AC_SUBST(SSL_ROOT)
 AC_SUBST(SSL_LIBDIR)
+AC_SUBST(SSL_CRYPTO_LIBNAME)
+AC_SUBST(SSL_SSL_LIBNAME)
 AC_SUBST(SSL_CC_RUNTIME_LIBRARY_PATH)
 AC_SUBST(SSL_LD_RUNTIME_LIBRARY_PATH)
 AC_SUBST(SSL_DED_LD_RUNTIME_LIBRARY_PATH)
@@ -3686,19 +3689,41 @@ case "$erl_xcomp_without_sysroot-$with_ssl" in
     CRYPTO_APP=crypto
     SSH_APP=ssh
 
-    AC_MSG_CHECKING(for OpenSSL >= 0.9.7 in standard locations) 
+    SSL_CRYPTO_LIBNAME=crypto
+    SSL_SSL_LIBNAME=ssl
+
+    AC_MSG_CHECKING(for OpenSSL >= 0.9.7 in standard locations)
     for rdir in $extra_dir /cygdrive/c/OpenSSL $std_ssl_locations; do
 	dir="$erl_xcomp_sysroot$rdir"
 	if test -f "$erl_xcomp_isysroot$rdir/include/openssl/opensslv.h"; then
 		is_real_ssl=yes
 		SSL_ROOT="$dir"
 		if test "x$MIXED_CYGWIN" = "xyes" ; then
-			if test -f "$dir/lib/VC/ssleay32.lib" || \
-			   test -f "$dir/lib/VC/openssl.lib"; then 
+			if test -f "$dir/lib/VC/libeay32.lib"; then
+				SSL_RUNTIME_LIBDIR="$rdir/lib/VC"
+				SSL_LIBDIR="$dir/lib/VC"
+				SSL_CRYPTO_LIBNAME=libeay32
+				SSL_SSL_LIBNAME=ssleay32
+			elif test -f "$dir/lib/VC/openssl.lib"; then 
 				SSL_RUNTIME_LIBDIR="$rdir/lib/VC"
 				SSL_LIBDIR="$dir/lib/VC"
-			elif test -f "$dir/lib/ssleay32.lib" || \
-			     test -f "$dir/lib/openssl.lib"; then
+			elif test -f $dir/lib/VC/libeay32MD.lib; then
+		            SSL_CRYPTO_LIBNAME=libeay32MD
+		            SSL_SSL_LIBNAME=ssleay32MD
+			    if test "x$enable_dynamic_ssl" = "xno" && \
+	                       test -f  $dir/lib/VC/static/libeay32MD.lib; then
+				  SSL_RUNTIME_LIBDIR="$rdir/lib/VC/static"
+				  SSL_LIBDIR="$dir/lib/VC/static"
+		            else
+			          SSL_RUNTIME_LIBDIR="$rdir/lib/VC"
+				  SSL_LIBDIR="$dir/lib/VC"
+                            fi 
+			elif test -f "$dir/lib/libeay32.lib"; then
+				SSL_RUNTIME_LIBDIR="$rdir/lib"
+				SSL_LIBDIR="$dir/lib"
+				SSL_CRYPTO_LIBNAME=libeay32
+				SSL_CRYPTO_LIBNAME=ssleay32
+			elif test -f "$dir/lib/openssl.lib"; then
 				SSL_RUNTIME_LIBDIR="$rdir/lib"
 				SSL_LIBDIR="$dir/lib"
 			else
@@ -3858,8 +3883,32 @@ dnl		so it is - be adoptable
        AC_MSG_ERROR(Invalid path to option --with-ssl=PATH)
     fi
     SSL_ROOT="$with_ssl"
+    SSL_CRYPTO_LIBNAME=crypto
+    SSL_SSL_LIBNAME=ssl
     if test "x$MIXED_CYGWIN" = "xyes" && test -d "$with_ssl/lib/VC"; then
-	SSL_LIBDIR="$with_ssl/lib/VC"
+	if test -f "$with_ssl/lib/VC/libeay32.lib"; then
+	    SSL_LIBDIR="$with_ssl/lib/VC"
+	    SSL_CRYPTO_LIBNAME=libeay32
+	    SSL_SSL_LIBNAME=ssleay32
+	elif test -f "$with_ssl/lib/VC/openssl.lib"; then 
+	    SSL_LIBDIR="$with_ssl/lib/VC"
+	elif test -f $with_ssl/lib/VC/libeay32MD.lib; then
+	    SSL_CRYPTO_LIBNAME=libeay32MD
+            SSL_SSL_LIBNAME=ssleay32MD
+	    if test "x$enable_dynamic_ssl" = "xno" && \
+	       test -f  $with_ssl/lib/VC/static/libeay32MD.lib; then
+		SSL_LIBDIR="$with_ssl/lib/VC/static"
+	    else
+		SSL_LIBDIR="$with_ssl/lib/VC"
+            fi 
+	elif test -f "$with_ssl/lib/libeay32.lib"; then
+	    SSL_LIBDIR="$with_ssl/lib"
+	    SSL_CRYPTO_LIBNAME=libeay32
+	    SSL_CRYPTO_LIBNAME=ssleay32
+	else
+	    # This probably wont work, but that's what the user said, so...
+	    SSL_LIBDIR="$with_ssl/lib"
+	fi
     elif test "x$ac_cv_sizeof_void_p" = "x8"; then
 	if test -f "$with_ssl/lib64/libcrypto.a"; then
 		SSL_LIBDIR="$with_ssl/lib64"
diff --git a/erts/doc/src/erlang.xml b/erts/doc/src/erlang.xml
index 19f501391f..f98e15cb52 100644
--- a/erts/doc/src/erlang.xml
+++ b/erts/doc/src/erlang.xml
@@ -2356,6 +2356,14 @@ os_prompt%</pre>
               <seealso marker="tools:instrument">instrument(3)</seealso>
               and/or <seealso marker="erts:erl">erl(1)</seealso>.</p>
           </item>
+          <tag><c>low</c></tag>
+          <item>
+            <p>Only on 64-bit halfword emulator.</p>
+            <p>The total amount of memory allocated in low memory areas
+	      that are restricted to less than 4 Gb even though
+	      the system may have more physical memory.</p>
+	    <p>May be removed in future releases of halfword emulator.</p>
+          </item>
         </taglist>
         <note>
           <p>The <c>system</c> value is not complete. Some allocated
diff --git a/erts/doc/src/notes.xml b/erts/doc/src/notes.xml
index f5607945a8..7383ea381d 100644
--- a/erts/doc/src/notes.xml
+++ b/erts/doc/src/notes.xml
@@ -4,7 +4,7 @@
 <chapter>
   <header>
     <copyright>
-      <year>2004</year><year>2010</year>
+      <year>2004</year><year>2011</year>
       <holder>Ericsson AB. All Rights Reserved.</holder>
     </copyright>
     <legalnotice>
diff --git a/erts/emulator/beam/atom.names b/erts/emulator/beam/atom.names
index 327620772f..de76cb9680 100644
--- a/erts/emulator/beam/atom.names
+++ b/erts/emulator/beam/atom.names
@@ -76,6 +76,7 @@ atom allocator_sizes
 atom alloc_util_allocators
 atom allow_passive_connect
 atom already_loaded
+atom amd64
 atom anchored
 atom and
 atom andalso
diff --git a/erts/emulator/beam/beam_bif_load.c b/erts/emulator/beam/beam_bif_load.c
index 1dbf6f9b92..d76a7d8e9f 100644
--- a/erts/emulator/beam/beam_bif_load.c
+++ b/erts/emulator/beam/beam_bif_load.c
@@ -1,7 +1,7 @@
 /*
  * %CopyrightBegin%
  *
- * Copyright Ericsson AB 1999-2010. All Rights Reserved.
+ * Copyright Ericsson AB 1999-2011. All Rights Reserved.
  *
  * The contents of this file are subject to the Erlang Public License,
  * Version 1.1, (the "License"); you may not use this file except in
@@ -142,7 +142,7 @@ BIF_RETTYPE code_is_module_native_1(BIF_ALIST_1)
     if ((modp = erts_get_module(BIF_ARG_1)) == NULL) {
 	return am_undefined;
     }
-    return (is_native(modp->code) ||
+    return ((modp->code && is_native(modp->code)) ||
 	    (modp->old_code != 0 && is_native(modp->old_code))) ?
 		am_true : am_false;
 }
diff --git a/erts/emulator/beam/beam_emu.c b/erts/emulator/beam/beam_emu.c
index 32ea8588d2..fb90a7d4f7 100644
--- a/erts/emulator/beam/beam_emu.c
+++ b/erts/emulator/beam/beam_emu.c
@@ -1911,13 +1911,15 @@ void process_main(void)
 	      * Note that for the halfword emulator, the two first elements
 	      * of the array are used.
 	      */
-	     *((BeamInstr **) (UWord) c_p->def_arg_reg) = I+3;
+	     BeamInstr** pi = (BeamInstr**) c_p->def_arg_reg;
+	     *pi = I+3;
 	     set_timer(c_p, unsigned_val(timeout_value));
 	 } else if (timeout_value == am_infinity) {
 	     c_p->flags |= F_TIMO;
 #if !defined(ARCH_64) || HALFWORD_HEAP
 	 } else if (term_to_Uint(timeout_value, &time_val)) {
-	     *((BeamInstr **) (UWord) c_p->def_arg_reg) = I+3;
+	     BeamInstr** pi = (BeamInstr**) c_p->def_arg_reg;
+	     *pi = I+3;
 	     set_timer(c_p, time_val);
 #endif
 	 } else {		/* Wrong time */
@@ -1974,7 +1976,8 @@ void process_main(void)
       * we must test the F_INSLPQUEUE flag as well as the F_TIMO flag.
       */
      if ((c_p->flags & (F_INSLPQUEUE | F_TIMO)) == 0) {
-	 *((BeamInstr **) (UWord) c_p->def_arg_reg) = I+3;
+	 BeamInstr** p = (BeamInstr **) c_p->def_arg_reg;
+	 *p = I+3;
 	 set_timer(c_p, Arg(1));
      }
      goto wait2;
@@ -5322,7 +5325,7 @@ void process_main(void)
 	 ep->code[3] = (BeamInstr) OpCode(apply_bif);
 	 ep->code[4] = (BeamInstr) bif_table[i].f;
 	 /* XXX: set func info for bifs */
-	 ((BeamInstr*)ep->code + 3)[-5] = (BeamInstr) BeamOp(op_i_func_info_IaaI);
+	 ep->fake_op_func_info_for_hipe[0] = (BeamInstr) BeamOp(op_i_func_info_IaaI);
      }
 
      return;
diff --git a/erts/emulator/beam/bif.c b/erts/emulator/beam/bif.c
index 8c35644125..68b3350d7f 100644
--- a/erts/emulator/beam/bif.c
+++ b/erts/emulator/beam/bif.c
@@ -2168,20 +2168,146 @@ BIF_RETTYPE tl_1(BIF_ALIST_1)
 /**********************************************************************/
 /* return the size of an I/O list */
 
-BIF_RETTYPE iolist_size_1(BIF_ALIST_1)
+static Eterm
+accumulate(Eterm acc, Uint size)
 {
-    Sint size = io_list_len(BIF_ARG_1);
+    if (is_non_value(acc)) {
+	/*
+	 * There is no pre-existing accumulator. Allocate a
+	 * bignum buffer with one extra word to be used if
+	 * the bignum grows in the future.
+	 */
+	Eterm* hp = (Eterm *) erts_alloc(ERTS_ALC_T_TEMP_TERM,
+					 (BIG_UINT_HEAP_SIZE+1) *
+					 sizeof(Eterm));
+	return uint_to_big(size, hp);
+    } else {
+	Eterm* big;
+	int need_heap;
 
-    if (size == -1) {
-	BIF_ERROR(BIF_P, BADARG);
-    } else if (IS_USMALL(0, (Uint) size)) {
-	BIF_RET(make_small(size));
+	/*
+	 * Add 'size' to 'acc' in place. There is always one
+	 * extra word allocated in case the bignum grows by one word.
+	 */
+	big = big_val(acc);
+	need_heap = BIG_NEED_SIZE(BIG_SIZE(big));
+	acc = big_plus_small(acc, size, big);
+	if (BIG_NEED_SIZE(big_size(acc)) > need_heap) {
+	    /*
+	     * The extra word has been consumed. Grow the
+	     * allocation by one word.
+	     */
+	    big = (Eterm *) erts_realloc(ERTS_ALC_T_TEMP_TERM,
+					 big_val(acc),
+					 (need_heap+1) * sizeof(Eterm));
+	    acc = make_big(big);
+	}
+	return acc;
+    }
+}
+
+static Eterm
+consolidate(Process* p, Eterm acc, Uint size)
+{
+    Eterm* hp;
+
+    if (is_non_value(acc)) {
+	return erts_make_integer(size, p);
     } else {
-	Eterm* hp = HAlloc(BIF_P, BIG_UINT_HEAP_SIZE);
-	BIF_RET(uint_to_big(size, hp));
+	Eterm* big;
+	Uint sz;
+	Eterm res;
+	
+	acc = accumulate(acc, size);
+	big = big_val(acc);
+	sz = BIG_NEED_SIZE(BIG_SIZE(big));
+	hp = HAlloc(p, sz);
+	res = make_big(hp);
+	while (sz--) {
+	    *hp++ = *big++;
+	}
+	erts_free(ERTS_ALC_T_TEMP_TERM, (void *) big_val(acc));
+	return res;
     }
 }
 
+BIF_RETTYPE iolist_size_1(BIF_ALIST_1)
+{
+    Eterm obj, hd;
+    Eterm* objp;
+    Uint size = 0;
+    Uint cur_size;
+    Uint new_size;
+    Eterm acc = THE_NON_VALUE;
+    DECLARE_ESTACK(s);
+
+    obj = BIF_ARG_1;
+    goto L_again;
+
+    while (!ESTACK_ISEMPTY(s)) {
+	obj = ESTACK_POP(s);
+    L_again:
+	if (is_list(obj)) {
+	L_iter_list:
+	    objp = list_val(obj);
+	    hd = CAR(objp);
+	    obj = CDR(objp);
+	    /* Head */
+	    if (is_byte(hd)) {
+		size++;
+		if (size == 0) {
+		    acc = accumulate(acc, (Uint) -1);
+		    size = 1;
+		}
+	    } else if (is_binary(hd) && binary_bitsize(hd) == 0) {
+		cur_size = binary_size(hd);
+		if ((new_size = size + cur_size) >= size) {
+		    size = new_size;
+		} else {
+		    acc = accumulate(acc, size);
+		    size = cur_size;
+		}
+	    } else if (is_list(hd)) {
+		ESTACK_PUSH(s, obj);
+		obj = hd;
+		goto L_iter_list;
+	    } else if (is_not_nil(hd)) {
+		goto L_type_error;
+	    }
+	    /* Tail */
+	    if (is_list(obj)) {
+		goto L_iter_list;
+	    } else if (is_binary(obj) && binary_bitsize(obj) == 0) {
+		cur_size = binary_size(obj);
+		if ((new_size = size + cur_size) >= size) {
+		    size = new_size;
+		} else {
+		    acc = accumulate(acc, size);
+		    size = cur_size;
+		}
+	    } else if (is_not_nil(obj)) {
+		goto L_type_error;
+	    }
+	} else if (is_binary(obj) && binary_bitsize(obj) == 0) {
+	    cur_size = binary_size(obj);
+	    if ((new_size = size + cur_size) >= size) {
+		size = new_size;
+	    } else {
+		acc = accumulate(acc, size);
+		size = cur_size;
+	    }
+	} else if (is_not_nil(obj)) {
+	    goto L_type_error;
+	}
+    }
+
+    DESTROY_ESTACK(s);
+    BIF_RET(consolidate(BIF_P, acc, size));
+
+ L_type_error:
+    DESTROY_ESTACK(s);
+    BIF_ERROR(BIF_P, BADARG);
+}
 
 /**********************************************************************/
 
@@ -3282,6 +3408,7 @@ BIF_RETTYPE ports_0(BIF_ALIST_0)
     Eterm* dead_ports;
     int alive, dead;
     Uint32 next_ss;
+    int i;
 
     /* To get a consistent snapshot... 
      * We add alive ports from start of the buffer
@@ -3293,21 +3420,18 @@ BIF_RETTYPE ports_0(BIF_ALIST_0)
     erts_smp_atomic_set(&erts_dead_ports_ptr,
 			(erts_aint_t) (port_buf + erts_max_ports));
 
-    next_ss = erts_smp_atomic_inctest(&erts_ports_snapshot);
+    next_ss = erts_smp_atomic32_inctest(&erts_ports_snapshot);
 
-    if (erts_smp_atomic_read(&erts_ports_alive) > 0) {
-	erts_aint_t i;
-	for (i = erts_max_ports-1; i >= 0; i--) {
-	    Port* prt = &erts_port[i];
-	    erts_smp_port_state_lock(prt);
-	    if (!(prt->status & ERTS_PORT_SFLGS_DEAD)
-		&& prt->snapshot != next_ss) {
-		ASSERT(prt->snapshot == next_ss - 1);
-		*pp++ = prt->id;		
-		prt->snapshot = next_ss; /* Consumed by this snapshot */
-	    }
-	    erts_smp_port_state_unlock(prt);
+    for (i = erts_max_ports-1; i >= 0; i--) {
+	Port* prt = &erts_port[i];
+	erts_smp_port_state_lock(prt);
+	if (!(prt->status & ERTS_PORT_SFLGS_DEAD)
+	    && prt->snapshot != next_ss) {
+	    ASSERT(prt->snapshot == next_ss - 1);
+	    *pp++ = prt->id;		
+	    prt->snapshot = next_ss; /* Consumed by this snapshot */
 	}
+	erts_smp_port_state_unlock(prt);
     }
 
     dead_ports = (Eterm*)erts_smp_atomic_xchg(&erts_dead_ports_ptr,
diff --git a/erts/emulator/beam/big.c b/erts/emulator/beam/big.c
index f47f5a9c0c..d18de9ae5d 100644
--- a/erts/emulator/beam/big.c
+++ b/erts/emulator/beam/big.c
@@ -1588,7 +1588,7 @@ big_to_double(Wterm x, double* resp)
 /*
  ** Estimate the number of decimal digits (include sign)
  */
-int big_decimal_estimate(Eterm x)
+int big_decimal_estimate(Wterm x)
 {
     Eterm* xp = big_val(x);
     int lg = I_lg(BIG_V(xp), BIG_SIZE(xp));
@@ -1602,7 +1602,7 @@ int big_decimal_estimate(Eterm x)
 ** Convert a bignum into a string of decimal numbers
 */
 
-static void write_big(Eterm x, void (*write_func)(void *, char), void *arg)
+static void write_big(Wterm x, void (*write_func)(void *, char), void *arg)
 {
     Eterm* xp = big_val(x);
     ErtsDigit* dx = BIG_V(xp);
@@ -1681,7 +1681,7 @@ write_string(void *arg, char c)
     *(--(*((char **) arg))) = c;
 }
 
-char *erts_big_to_string(Eterm x, char *buf, Uint buf_sz)
+char *erts_big_to_string(Wterm x, char *buf, Uint buf_sz)
 {
     char *big_str = buf + buf_sz - 1;
     *big_str = '\0';
diff --git a/erts/emulator/beam/big.h b/erts/emulator/beam/big.h
index f28a390aea..2afc37004f 100644
--- a/erts/emulator/beam/big.h
+++ b/erts/emulator/beam/big.h
@@ -114,9 +114,9 @@ typedef Uint  dsize_t;	 /* Vector size type */
 
 #endif
 
-int big_decimal_estimate(Eterm);
+int big_decimal_estimate(Wterm);
 Eterm erts_big_to_list(Eterm, Eterm**);
-char *erts_big_to_string(Eterm x, char *buf, Uint buf_sz);
+char *erts_big_to_string(Wterm x, char *buf, Uint buf_sz);
 
 Eterm small_times(Sint, Sint, Eterm*);
 
diff --git a/erts/emulator/beam/binary.c b/erts/emulator/beam/binary.c
index 9486602633..1fb39c6c67 100644
--- a/erts/emulator/beam/binary.c
+++ b/erts/emulator/beam/binary.c
@@ -32,11 +32,11 @@
 #include "erl_bits.h"
 
 #ifdef DEBUG
-static int list_to_bitstr_buf(Eterm obj, char* buf, int len);
+static int list_to_bitstr_buf(Eterm obj, char* buf, Uint len);
 #else
 static int list_to_bitstr_buf(Eterm obj, char* buf);
 #endif
-static Sint bitstr_list_len(Eterm obj);
+static int bitstr_list_len(Eterm obj, Uint* num_bytes);
 
 void
 erts_init_binary(void)
@@ -355,21 +355,24 @@ BIF_RETTYPE bitstring_to_list_1(BIF_ALIST_1)
 BIF_RETTYPE erts_list_to_binary_bif(Process *p, Eterm arg)
 {
     Eterm bin;
-    int i;
+    Uint size;
     int offset;
     byte* bytes;
+
     if (is_nil(arg)) {
 	BIF_RET(new_binary(p,(byte*)"",0));
     }
     if (is_not_list(arg)) {
 	goto error;
     }
-    if ((i = io_list_len(arg)) < 0) {
-	goto error;
+    switch (erts_iolist_size(arg, &size)) {
+    case ERTS_IOLIST_OVERFLOW: BIF_ERROR(p, SYSTEM_LIMIT);
+    case ERTS_IOLIST_TYPE: goto error;
+    default: ;
     }
-    bin = new_binary(p, (byte *)NULL, i);
+    bin = new_binary(p, (byte *)NULL, size);
     bytes = binary_bytes(bin);
-    offset = io_list_to_buf(arg, (char*) bytes, i);
+    offset = io_list_to_buf(arg, (char*) bytes, size);
     ASSERT(offset == 0);
     BIF_RET(bin);
     
@@ -396,7 +399,8 @@ BIF_RETTYPE iolist_to_binary_1(BIF_ALIST_1)
 BIF_RETTYPE list_to_bitstring_1(BIF_ALIST_1)
 {
     Eterm bin;
-    int i,offset;
+    Uint sz;
+    int offset;
     byte* bytes;
     ErlSubBin* sb1; 
     Eterm* hp;
@@ -405,15 +409,19 @@ BIF_RETTYPE list_to_bitstring_1(BIF_ALIST_1)
 	BIF_RET(new_binary(BIF_P,(byte*)"",0));
     }
     if (is_not_list(BIF_ARG_1)) {
-	goto error;
+    error:
+	BIF_ERROR(BIF_P, BADARG);
     }
-    if ((i = bitstr_list_len(BIF_ARG_1)) < 0) {
+    switch (bitstr_list_len(BIF_ARG_1, &sz)) {
+    case ERTS_IOLIST_TYPE:
 	goto error;
+    case ERTS_IOLIST_OVERFLOW:
+	BIF_ERROR(BIF_P, SYSTEM_LIMIT);
     }
-    bin = new_binary(BIF_P, (byte *)NULL, i);
+    bin = new_binary(BIF_P, (byte *)NULL, sz);
     bytes = binary_bytes(bin);
 #ifdef DEBUG
-    offset = list_to_bitstr_buf(BIF_ARG_1, (char*) bytes, i);
+    offset = list_to_bitstr_buf(BIF_ARG_1, (char*) bytes, sz);
 #else
     offset = list_to_bitstr_buf(BIF_ARG_1, (char*) bytes);
 #endif
@@ -422,20 +430,16 @@ BIF_RETTYPE list_to_bitstring_1(BIF_ALIST_1)
 	hp = HAlloc(BIF_P, ERL_SUB_BIN_SIZE);
 	sb1 = (ErlSubBin *) hp;
 	sb1->thing_word = HEADER_SUB_BIN;
-	sb1->size = i-1;
+	sb1->size = sz-1;
 	sb1->offs = 0;
 	sb1->orig = bin;
 	sb1->bitoffs = 0;
 	sb1->bitsize = offset;
 	sb1->is_writable = 0;
-	hp += ERL_SUB_BIN_SIZE;
 	bin = make_binary(sb1);
     }
     
     BIF_RET(bin);
-    
-    error:
-	BIF_ERROR(BIF_P, BADARG);
 }
 
 BIF_RETTYPE split_binary_2(BIF_ALIST_2)
@@ -499,7 +503,7 @@ BIF_RETTYPE split_binary_2(BIF_ALIST_2)
  */
 static int
 #ifdef DEBUG
-list_to_bitstr_buf(Eterm obj, char* buf, int len)
+list_to_bitstr_buf(Eterm obj, char* buf, Uint len)
 #else
 list_to_bitstr_buf(Eterm obj, char* buf)
 #endif
@@ -602,8 +606,8 @@ list_to_bitstr_buf(Eterm obj, char* buf)
     return offset;
 }
 
-static Sint
-bitstr_list_len(Eterm obj)
+static int
+bitstr_list_len(Eterm obj, Uint* num_bytes)
 {
     Eterm* objp;
     Uint len = 0;
@@ -611,6 +615,26 @@ bitstr_list_len(Eterm obj)
     DECLARE_ESTACK(s);
     goto L_again;
 
+#define SAFE_ADD(Var, Val)			\
+    do {					\
+        Uint valvar = (Val);			\
+	Var += valvar;				\
+	if (Var < valvar) {			\
+	    goto L_overflow_error;		\
+	}					\
+    } while (0)
+
+#define SAFE_ADD_BITSIZE(Var, Bin)					\
+    do {								\
+	if (*binary_val(Bin) == HEADER_SUB_BIN) {			\
+            Uint valvar = ((ErlSubBin *) binary_val(Bin))->bitsize;	\
+	    Var += valvar;						\
+	    if (Var < valvar) {						\
+	         goto L_overflow_error;					\
+	    }								\
+        }								\
+    } while (0)
+
     while (!ESTACK_ISEMPTY(s)) {
 	obj = ESTACK_POP(s);
     L_again:
@@ -621,9 +645,12 @@ bitstr_list_len(Eterm obj)
 	    obj = CAR(objp);
 	    if (is_byte(obj)) {
 		len++;
+		if (len == 0) {
+		    goto L_overflow_error;
+		}
 	    } else if (is_binary(obj)) {
-		len += binary_size(obj);
-		offs += binary_bitsize(obj);
+		SAFE_ADD(len, binary_size(obj));
+		SAFE_ADD_BITSIZE(offs, obj);
 	    } else if (is_list(obj)) {
 		ESTACK_PUSH(s, CDR(objp));
 		goto L_iter_list; /* on head */
@@ -635,24 +662,44 @@ bitstr_list_len(Eterm obj)
 	    if (is_list(obj))
 		goto L_iter_list; /* on tail */
 	    else if (is_binary(obj)) {
-		len += binary_size(obj);
-		offs += binary_bitsize(obj);
+		SAFE_ADD(len, binary_size(obj));
+		SAFE_ADD_BITSIZE(offs, obj);
 	    } else if (is_not_nil(obj)) {
 		goto L_type_error;
 	    }
 	} else if (is_binary(obj)) {
-	    len += binary_size(obj);
-	    offs += binary_bitsize(obj);
+	    SAFE_ADD(len, binary_size(obj));
+	    SAFE_ADD_BITSIZE(offs, obj);
 	} else if (is_not_nil(obj)) {
 	    goto L_type_error;
 	}
     }
+#undef SAFE_ADD
+#undef SAFE_ADD_BITSIZE
 
     DESTROY_ESTACK(s);
-    return (Sint) (len + (offs/8) + ((offs % 8) != 0));
+
+    /*
+     * Make sure that the number of bits in the bitstring will fit
+     * in an Uint to ensure that the binary can be matched using
+     * the binary syntax.
+     */
+    if (len << 3 < len) {
+	goto L_overflow_error;
+    }
+    len += (offs >> 3) + ((offs & 7) != 0);
+    if (len << 3 < len) {
+	goto L_overflow_error;
+    }
+    *num_bytes = len;
+    return ERTS_IOLIST_OK;
 
  L_type_error:
     DESTROY_ESTACK(s);
-    return (Sint) -1;
+    return ERTS_IOLIST_TYPE;
+
+ L_overflow_error:
+    DESTROY_ESTACK(s);
+    return ERTS_IOLIST_OVERFLOW;
 }
 
diff --git a/erts/emulator/beam/erl_alloc.c b/erts/emulator/beam/erl_alloc.c
index 673eac7fea..cda404af5e 100644
--- a/erts/emulator/beam/erl_alloc.c
+++ b/erts/emulator/beam/erl_alloc.c
@@ -90,6 +90,10 @@ typedef union {
 static ErtsAllocatorState_t sl_alloc_state;
 static ErtsAllocatorState_t std_alloc_state;
 static ErtsAllocatorState_t ll_alloc_state;
+#if HALFWORD_HEAP
+static ErtsAllocatorState_t std_alloc_low_state;
+static ErtsAllocatorState_t ll_alloc_low_state;
+#endif
 static ErtsAllocatorState_t temp_alloc_state;
 static ErtsAllocatorState_t eheap_alloc_state;
 static ErtsAllocatorState_t binary_alloc_state;
@@ -166,6 +170,10 @@ typedef struct {
     struct au_init binary_alloc;
     struct au_init ets_alloc;
     struct au_init driver_alloc;
+#if HALFWORD_HEAP
+    struct au_init std_alloc_low;
+    struct au_init ll_alloc_low;
+#endif
 } erts_alc_hndl_args_init_t;
 
 #define ERTS_AU_INIT__ {0, 0, GOODFIT, DEFAULT_ALLCTR_INIT, {1,1,1,1}}
@@ -193,6 +201,10 @@ set_default_sl_alloc_opts(struct au_init *ip)
 #endif
     ip->init.util.ts 		= ERTS_ALC_MTA_SHORT_LIVED;
     ip->init.util.rsbcst	= 80;
+#if HALFWORD_HEAP
+    ip->init.util.low_mem       = 1;
+#endif
+
 }
 
 static void
@@ -256,6 +268,9 @@ set_default_temp_alloc_opts(struct au_init *ip)
     ip->init.util.ts 		= ERTS_ALC_MTA_TEMPORARY;
     ip->init.util.rsbcst	= 90;
     ip->init.util.rmbcmt	= 100;
+#if HALFWORD_HEAP
+    ip->init.util.low_mem       = 1;
+#endif
 }
 
 static void
@@ -275,6 +290,9 @@ set_default_eheap_alloc_opts(struct au_init *ip)
 #endif
     ip->init.util.ts 		= ERTS_ALC_MTA_EHEAP;
     ip->init.util.rsbcst	= 50;
+#if HALFWORD_HEAP
+    ip->init.util.low_mem       = 1;
+#endif
 }
 
 static void
@@ -531,6 +549,20 @@ erts_alloc_init(int *argc, char **argv, ErtsAllocInitOpts *eaiop)
     erts_allctrs[ERTS_ALC_A_SYSTEM].free		= erts_sys_free;
     erts_allctrs_info[ERTS_ALC_A_SYSTEM].enabled	= 1;
 
+#if HALFWORD_HEAP
+    /* Init low memory variants by cloning */
+    init.std_alloc_low = init.std_alloc;
+    init.std_alloc_low.init.util.alloc_no = ERTS_ALC_A_STANDARD_LOW;
+    init.std_alloc_low.init.util.low_mem  = 1;
+
+    init.ll_alloc_low = init.ll_alloc;
+    init.ll_alloc_low.init.util.alloc_no = ERTS_ALC_A_LONG_LIVED_LOW;
+    init.ll_alloc_low.init.util.low_mem = 1;
+
+    set_au_allocator(ERTS_ALC_A_STANDARD_LOW, &init.std_alloc_low);
+    set_au_allocator(ERTS_ALC_A_LONG_LIVED_LOW, &init.ll_alloc_low);
+#endif /* HALFWORD */
+
     set_au_allocator(ERTS_ALC_A_TEMPORARY, &init.temp_alloc);
     set_au_allocator(ERTS_ALC_A_SHORT_LIVED, &init.sl_alloc);
     set_au_allocator(ERTS_ALC_A_STANDARD, &init.std_alloc);
@@ -576,7 +608,14 @@ erts_alloc_init(int *argc, char **argv, ErtsAllocInitOpts *eaiop)
     start_au_allocator(ERTS_ALC_A_LONG_LIVED,
 		       &init.ll_alloc,
 		       &ll_alloc_state);
-
+#if HALFWORD_HEAP
+    start_au_allocator(ERTS_ALC_A_LONG_LIVED_LOW,
+		       &init.ll_alloc_low,
+		       &ll_alloc_low_state);
+    start_au_allocator(ERTS_ALC_A_STANDARD_LOW,
+		       &init.std_alloc_low,
+		       &std_alloc_low_state);
+#endif
     start_au_allocator(ERTS_ALC_A_EHEAP,
 		       &init.eheap_alloc,
 		       &eheap_alloc_state);
@@ -612,11 +651,9 @@ erts_alloc_init(int *argc, char **argv, ErtsAllocInitOpts *eaiop)
     erts_set_fix_size(ERTS_ALC_T_PROC,		sizeof(Process));
     erts_set_fix_size(ERTS_ALC_T_DB_TABLE,	sizeof(DbTable));
     erts_set_fix_size(ERTS_ALC_T_ATOM,		sizeof(Atom));
-    erts_set_fix_size(ERTS_ALC_T_EXPORT,	sizeof(Export));
+
     erts_set_fix_size(ERTS_ALC_T_MODULE,	sizeof(Module));
     erts_set_fix_size(ERTS_ALC_T_REG_PROC,	sizeof(RegProc));
-    erts_set_fix_size(ERTS_ALC_T_MONITOR_SH,	ERTS_MONITOR_SH_SIZE*sizeof(Uint));
-    erts_set_fix_size(ERTS_ALC_T_NLINK_SH,	ERTS_LINK_SH_SIZE*sizeof(Uint));
     erts_set_fix_size(ERTS_ALC_T_FUN_ENTRY,	sizeof(ErlFunEntry));
 #ifdef ERTS_ALC_T_DRV_EV_D_STATE
     erts_set_fix_size(ERTS_ALC_T_DRV_EV_D_STATE,
@@ -626,6 +663,11 @@ erts_alloc_init(int *argc, char **argv, ErtsAllocInitOpts *eaiop)
     erts_set_fix_size(ERTS_ALC_T_DRV_SEL_D_STATE,
 		      sizeof(ErtsDrvSelectDataState));
 #endif
+#if !HALFWORD_HEAP
+    erts_set_fix_size(ERTS_ALC_T_EXPORT,	sizeof(Export));
+    erts_set_fix_size(ERTS_ALC_T_MONITOR_SH,	ERTS_MONITOR_SH_SIZE*sizeof(Uint));
+    erts_set_fix_size(ERTS_ALC_T_NLINK_SH,	ERTS_LINK_SH_SIZE*sizeof(Uint));
+#endif
 #endif
 #endif
 
@@ -638,6 +680,15 @@ set_au_allocator(ErtsAlcType_t alctr_n, struct au_init *init)
     ErtsAllocatorInfo_t *ai = &erts_allctrs_info[alctr_n];
     ErtsAllocatorThrSpec_t *tspec = &erts_allctr_thr_spec[alctr_n];
 
+#if HALFWORD_HEAP
+    /* If halfword heap, silently ignore any disabling of internal
+     * allocators for low memory
+     */
+    if (init->init.util.low_mem) {
+	init->enable = 1;
+    }
+#endif
+
     if (!init->enable) {
 	af->alloc = erts_sys_alloc;
 	af->realloc = erts_sys_realloc;
@@ -1348,14 +1399,6 @@ handle_args(int *argc, char **argv, erts_alc_hndl_args_init_t *init)
 	    argv[j++] = argv[i];
     }
     *argc = j;
-#if HALFWORD_HEAP
-    /* If halfword heap, silently ignore any disabling of internal 
-       allocators */
-    for (i = 0; i < aui_sz; ++i)
-	aui[i]->enable = 1;
-#endif
-
-    
 }
 
 static char *type_no_str(ErtsAlcType_t n)
@@ -1528,10 +1571,10 @@ erts_realloc_n_enomem(ErtsAlcType_t n, void *ptr, Uint size)
     erts_alc_fatal_error(ERTS_ALC_E_NOMEM, ERTS_ALC_O_REALLOC, n, size);
 }
 
-static ERTS_INLINE Uint
+static ERTS_INLINE UWord
 alcu_size(ErtsAlcType_t ai)
 {
-    Uint res = 0;
+    UWord res = 0;
 
     ASSERT(erts_allctrs_info[ai].enabled);
     ASSERT(erts_allctrs_info[ai].alloc_util);
@@ -1563,6 +1606,49 @@ alcu_size(ErtsAlcType_t ai)
     return res;
 }
 
+#if HALFWORD_HEAP
+static ERTS_INLINE int
+alcu_is_low(ErtsAlcType_t ai)
+{
+    int is_low = 0;
+    ASSERT(erts_allctrs_info[ai].enabled);
+    ASSERT(erts_allctrs_info[ai].alloc_util);
+
+    if (!erts_allctrs_info[ai].thr_spec) {
+	Allctr_t *allctr = erts_allctrs_info[ai].extra;
+	is_low = allctr->mseg_opt.low_mem;
+    }
+    else {
+	ErtsAllocatorThrSpec_t *tspec = &erts_allctr_thr_spec[ai];
+	int i;
+# ifdef DEBUG
+	int found_one = 0;
+# endif
+
+	ASSERT(tspec->all_thr_safe);
+	ASSERT(tspec->enabled);
+
+	for (i = tspec->size - 1; i >= 0; i--) {
+	    Allctr_t *allctr = tspec->allctr[i];
+	    if (allctr) {
+# ifdef DEBUG
+		if (!found_one) {
+		    is_low = allctr->mseg_opt.low_mem;
+		    found_one = 1;
+		}
+		else ASSERT(is_low == allctr->mseg_opt.low_mem);
+# else
+		is_low = allctr->mseg_opt.low_mem;
+		break;
+# endif
+	    }
+	}
+	ASSERT(found_one);
+    }
+    return is_low;
+}
+#endif /* HALFWORD */
+
 Eterm
 erts_memory(int *print_to_p, void *print_to_arg, void *proc, Eterm earg)
 {
@@ -1579,22 +1665,28 @@ erts_memory(int *print_to_p, void *print_to_arg, void *proc, Eterm earg)
 	int code;
 	int ets;
 	int maximum;
+#if HALFWORD_HEAP
+	int low;
+#endif
     } want = {0};
     struct {
-	Uint total;
-	Uint processes;
-	Uint processes_used;
-	Uint system;
-	Uint atom;
-	Uint atom_used;
-	Uint binary;
-	Uint code;
-	Uint ets;
-	Uint maximum;
+	UWord total;
+	UWord processes;
+	UWord processes_used;
+	UWord system;
+	UWord atom;
+	UWord atom_used;
+	UWord binary;
+	UWord code;
+	UWord ets;
+	UWord maximum;
+#if HALFWORD_HEAP
+	UWord low;
+#endif
     } size = {0};
-    Eterm atoms[sizeof(size)/sizeof(Uint)];
-    Uint *uintps[sizeof(size)/sizeof(Uint)];
-    Eterm euints[sizeof(size)/sizeof(Uint)];
+    Eterm atoms[sizeof(size)/sizeof(UWord)];
+    UWord *uintps[sizeof(size)/sizeof(UWord)];
+    Eterm euints[sizeof(size)/sizeof(UWord)];
     int want_tot_or_sys;
     int length;
     Eterm res = THE_NON_VALUE;
@@ -1646,7 +1738,11 @@ erts_memory(int *print_to_p, void *print_to_arg, void *proc, Eterm earg)
 	    atoms[length] = am_maximum;
 	    uintps[length++] = &size.maximum;
 	}
-
+#if HALFWORD_HEAP
+	want.low = 1;
+	atoms[length] = am_low;
+	uintps[length++] = &size.low;
+#endif
     }
     else {
 	DeclareTmpHeapNoproc(tmp_heap,2);
@@ -1740,6 +1836,15 @@ erts_memory(int *print_to_p, void *print_to_arg, void *proc, Eterm earg)
 		    return am_badarg;
 		}
 		break;
+#if HALFWORD_HEAP
+	    case am_low:
+		if (!want.low) {
+		    want.low = 1;
+		    atoms[length] = am_low;
+		    uintps[length++] = &size.low;
+		}
+		break;
+#endif
 	    default:
 		UnUseTmpHeapNoproc(2);
 		return am_badarg;
@@ -1769,7 +1874,7 @@ erts_memory(int *print_to_p, void *print_to_arg, void *proc, Eterm earg)
 
     ASSERT(length <= sizeof(atoms)/sizeof(Eterm));
     ASSERT(length <= sizeof(euints)/sizeof(Eterm));
-    ASSERT(length <= sizeof(uintps)/sizeof(Uint));
+    ASSERT(length <= sizeof(uintps)/sizeof(UWord));
 
 
     if (proc) {
@@ -1788,8 +1893,8 @@ erts_memory(int *print_to_p, void *print_to_arg, void *proc, Eterm earg)
 
 	for (ai = ERTS_ALC_A_MIN; ai <= ERTS_ALC_A_MAX; ai++) {
 	    if (erts_allctrs_info[ai].alloc_util) {
-		Uint *save;
-		Uint asz;
+		UWord *save;
+		UWord asz;
 		switch (ai) {
 		case ERTS_ALC_A_TEMPORARY:
 		     /*
@@ -1814,6 +1919,11 @@ erts_memory(int *print_to_p, void *print_to_arg, void *proc, Eterm earg)
 		if (save)
 		    *save = asz;
 		size.total += asz;
+#if HALFWORD_HEAP
+		if (alcu_is_low(ai)) {
+		    size.low += asz;
+		}
+#endif
 	    }
 	}
     }
@@ -1821,7 +1931,7 @@ erts_memory(int *print_to_p, void *print_to_arg, void *proc, Eterm earg)
 
 
     if (want_tot_or_sys || want.processes || want.processes_used) {
-	Uint tmp;
+	UWord tmp;
 
 	if (ERTS_MEM_NEED_ALL_ALCU)
 	    tmp = size.processes;
@@ -1836,6 +1946,9 @@ erts_memory(int *print_to_p, void *print_to_arg, void *proc, Eterm earg)
 
 	size.processes = size.processes_used = tmp;
 
+#if HALFWORD_HEAP
+	/* BUG: We ignore link and monitor memory */
+#else
 	erts_fix_info(ERTS_ALC_T_NLINK_SH, &efi);
 	size.processes += efi.total;
 	size.processes_used += efi.used;
@@ -1843,6 +1956,7 @@ erts_memory(int *print_to_p, void *print_to_arg, void *proc, Eterm earg)
 	erts_fix_info(ERTS_ALC_T_MONITOR_SH, &efi);
 	size.processes += efi.total;
 	size.processes_used += efi.used;
+#endif
 
 	erts_fix_info(ERTS_ALC_T_PROC, &efi);
 	size.processes += efi.total;
@@ -1879,8 +1993,12 @@ erts_memory(int *print_to_p, void *print_to_arg, void *proc, Eterm earg)
 	erts_fix_info(ERTS_ALC_T_MODULE, &efi);
 	size.code += efi.used;
 	size.code += export_table_sz();
+#if HALFWORD_HEAP
+	size.code += export_list_size() * sizeof(Export);
+#else
 	erts_fix_info(ERTS_ALC_T_EXPORT, &efi);
 	size.code += efi.used;
+#endif
 	size.code += erts_fun_table_sz();
 	erts_fix_info(ERTS_ALC_T_FUN_ENTRY, &efi);
 	size.code += efi.used;
@@ -1913,7 +2031,7 @@ erts_memory(int *print_to_p, void *print_to_arg, void *proc, Eterm earg)
 	/* Print result... */
 	erts_print(to, arg, "=memory\n");
 	for (i = 0; i < length; i++)
-	    erts_print(to, arg, "%T: %beu\n", atoms[i], *uintps[i]);
+	    erts_print(to, arg, "%T: %bpu\n", atoms[i], *uintps[i]);
     }
 
     if (proc) {
@@ -1926,9 +2044,9 @@ erts_memory(int *print_to_p, void *print_to_arg, void *proc, Eterm earg)
 	if (only_one_value) {
 	    ASSERT(length == 1);
 	    hsz = 0;
-	    erts_bld_uint(NULL, &hsz, *uintps[0]);
+	    erts_bld_uword(NULL, &hsz, *uintps[0]);
 	    hp = hsz ? HAlloc((Process *) proc, hsz) : NULL;
-	    res = erts_bld_uint(&hp, NULL, *uintps[0]);
+	    res = erts_bld_uword(&hp, NULL, *uintps[0]);
 	}
 	else {
 	    Uint **hpp = NULL;
@@ -1938,7 +2056,7 @@ erts_memory(int *print_to_p, void *print_to_arg, void *proc, Eterm earg)
 	    while (1) {
 		int i;
 		for (i = 0; i < length; i++)
-		    euints[i] = erts_bld_uint(hpp, hszp, *uintps[i]);
+		    euints[i] = erts_bld_uword(hpp, hszp, *uintps[i]);
 		res = erts_bld_2tup_list(hpp, hszp, length, atoms, euints);
 		if (hpp)
 		    break;
diff --git a/erts/emulator/beam/erl_alloc.types b/erts/emulator/beam/erl_alloc.types
index ca71798917..c6cc0e1fac 100644
--- a/erts/emulator/beam/erl_alloc.types
+++ b/erts/emulator/beam/erl_alloc.types
@@ -75,6 +75,11 @@ allocator	EHEAP		true		eheap_alloc
 allocator	ETS		true		ets_alloc
 allocator	FIXED_SIZE	true		fix_alloc
 
++if halfword
+allocator	LONG_LIVED_LOW	true		ll_alloc_low
+allocator	STANDARD_LOW	true		std_alloc_low
++endif
+
 +else # Non smp build
 
 allocator	TEMPORARY	false		temp_alloc
@@ -85,12 +90,18 @@ allocator	EHEAP		false		eheap_alloc
 allocator	ETS		false		ets_alloc
 allocator	FIXED_SIZE	false		fix_alloc
 
++if halfword
+allocator	LONG_LIVED_LOW	false		ll_alloc_low
+allocator	STANDARD_LOW	false		std_alloc_low
++endif
+
 +endif
 
 allocator	BINARY		true		binary_alloc
 allocator	DRIVER		true		driver_alloc
 
 
+
 # --- Class declarations -----------------------------------------------------
 #
 # Syntax: class <CLASS> <DESCRIPTION>
@@ -125,14 +136,9 @@ class	SYSTEM		system_data
 
 type	PROC		FIXED_SIZE	PROCESSES	proc
 type	ATOM		FIXED_SIZE	ATOM		atom_entry
-type	EXPORT		FIXED_SIZE	CODE		export_entry
 type	MODULE		FIXED_SIZE	CODE		module_entry
 type	REG_PROC	FIXED_SIZE	PROCESSES	reg_proc
 type	LINK_LH		STANDARD	PROCESSES	link_lh
-type	MONITOR_SH	FIXED_SIZE	PROCESSES	monitor_sh
-type	MONITOR_LH	STANDARD	PROCESSES	monitor_lh
-type	NLINK_SH	FIXED_SIZE	PROCESSES	nlink_sh
-type	NLINK_LH	STANDARD	PROCESSES	nlink_lh
 type	SUSPEND_MON	STANDARD	PROCESSES	suspend_monitor
 type	PEND_SUSPEND	SHORT_LIVED	PROCESSES	pending_suspend
 type	PROC_LIST	SHORT_LIVED	PROCESSES	proc_list
@@ -175,7 +181,6 @@ type	DRIVER		STANDARD	SYSTEM		driver
 type	NIF		DRIVER		SYSTEM		nif_internal
 type	BINARY		BINARY		BINARIES	binary
 type	NBIF_TABLE	SYSTEM		SYSTEM		nbif_tab
-type	CODE		LONG_LIVED	CODE		code
 type	ARG_REG		STANDARD	PROCESSES	arg_reg
 type	PROC_DICT	STANDARD	PROCESSES	proc_dict
 type	CALLS_BUF	STANDARD	PROCESSES	calls_buf
@@ -193,10 +198,8 @@ type	DB_FIXATION	SHORT_LIVED	ETS		db_fixation
 type	DB_FIX_DEL	SHORT_LIVED	ETS		fixed_del
 type	DB_TABLES	LONG_LIVED	ETS		db_tabs
 type    DB_NTAB_ENT	STANDARD	ETS		db_named_table_entry
-type	DB_HEIR_DATA	STANDARD	ETS		db_heir_data
 type	DB_TMP		TEMPORARY	ETS		db_tmp
 type	DB_MC_STK	TEMPORARY	ETS		db_mc_stack
-type	DB_MS_PSDO_PROC	LONG_LIVED	ETS		db_match_pseudo_proc
 type	DB_MS_RUN_HEAP	SHORT_LIVED	ETS		db_match_spec_run_heap
 type	DB_MS_CMPL_HEAP	TEMPORARY	ETS		db_match_spec_cmpl_heap
 type	DB_SEG		ETS		ETS		db_segment
@@ -213,10 +216,8 @@ type	LOGGER_DSBUF	TEMPORARY	SYSTEM		logger_dsbuf
 type	TMP_DSBUF	TEMPORARY	SYSTEM		tmp_dsbuf
 type	INFO_DSBUF	SYSTEM		SYSTEM		info_dsbuf
 # INFO_DSBUF have to use the SYSTEM allocator; otherwise, a deadlock might occur
-type	SCHDLR_DATA	LONG_LIVED	SYSTEM		scheduler_data
 type	SCHDLR_SLP_INFO	LONG_LIVED	SYSTEM		scheduler_sleep_info
 type	RUNQS		LONG_LIVED	SYSTEM		run_queues
-type	DDLL_PROCESS	STANDARD	SYSTEM		ddll_processes
 type	DDLL_HANDLE	STANDARD	SYSTEM		ddll_handle
 type	DDLL_ERRCODES	LONG_LIVED	SYSTEM		ddll_errcodes
 type	DDLL_TMP_BUF	TEMPORARY	SYSTEM		ddll_tmp_buf
@@ -327,13 +328,45 @@ type	SSB		SHORT_LIVED	PROCESSES	ssb
 +endif
 
 
++if halfword
+
+type	DDLL_PROCESS	STANDARD_LOW	SYSTEM		ddll_processes
+type	MONITOR_LH	STANDARD_LOW	PROCESSES	monitor_lh
+type	NLINK_LH	STANDARD_LOW	PROCESSES	nlink_lh
+type	CODE		LONG_LIVED_LOW	CODE		code
+type	DB_HEIR_DATA	STANDARD_LOW	ETS		db_heir_data
+type	DB_MS_PSDO_PROC	LONG_LIVED_LOW	ETS		db_match_pseudo_proc
+type	SCHDLR_DATA	LONG_LIVED_LOW	SYSTEM		scheduler_data
+type  	LL_TEMP_TERM 	LONG_LIVED_LOW	SYSTEM		ll_temp_term
+
+# no FIXED_SIZE for low memory
+type	EXPORT		STANDARD_LOW	CODE		export_entry
+type	MONITOR_SH	STANDARD_LOW	PROCESSES	monitor_sh
+type	NLINK_SH	STANDARD_LOW	PROCESSES	nlink_sh
+
++else # "fullword"
+
+type	DDLL_PROCESS	STANDARD	SYSTEM		ddll_processes
+type	MONITOR_LH	STANDARD	PROCESSES	monitor_lh
+type	NLINK_LH	STANDARD	PROCESSES	nlink_lh
+type	CODE		LONG_LIVED	CODE		code
+type	DB_HEIR_DATA	STANDARD	ETS		db_heir_data
+type	DB_MS_PSDO_PROC	LONG_LIVED	ETS		db_match_pseudo_proc
+type	SCHDLR_DATA	LONG_LIVED	SYSTEM		scheduler_data
+type  	LL_TEMP_TERM 	LONG_LIVED	SYSTEM		ll_temp_term
+
+type	EXPORT		FIXED_SIZE	CODE		export_entry
+type	MONITOR_SH	FIXED_SIZE	PROCESSES	monitor_sh
+type	NLINK_SH	FIXED_SIZE	PROCESSES	nlink_sh
+
++endif
+
+
 #
 # Types used by system specific code
 #
 
 type  	TEMP_TERM 	TEMPORARY	SYSTEM		temp_term
-type  	LL_TEMP_TERM 	LONG_LIVED	SYSTEM		ll_temp_term
-
 type	DRV_TAB		LONG_LIVED	SYSTEM		drv_tab
 type	DRV_EV_STATE	LONG_LIVED	SYSTEM		driver_event_state
 type	DRV_EV_D_STATE	FIXED_SIZE	SYSTEM		driver_event_data_state
diff --git a/erts/emulator/beam/erl_alloc_util.c b/erts/emulator/beam/erl_alloc_util.c
index 84c72439a3..cc04ef65bf 100644
--- a/erts/emulator/beam/erl_alloc_util.c
+++ b/erts/emulator/beam/erl_alloc_util.c
@@ -1639,6 +1639,9 @@ static struct {
     Eterm e;
     Eterm t;
     Eterm ramv;
+#if HALFWORD_HEAP
+    Eterm low;
+#endif
     Eterm sbct;
 #if HAVE_ERTS_MSEG
     Eterm asbcst;
@@ -1724,6 +1727,9 @@ init_atoms(Allctr_t *allctr)
 	AM_INIT(e);
 	AM_INIT(t);
 	AM_INIT(ramv);
+#if HALFWORD_HEAP
+	AM_INIT(low);
+#endif
 	AM_INIT(sbct);
 #if HAVE_ERTS_MSEG
 	AM_INIT(asbcst);
@@ -2168,6 +2174,9 @@ info_options(Allctr_t *allctr,
 		   "option e: true\n"
 		   "option t: %s\n"
 		   "option ramv: %s\n"
+#if HALFWORD_HEAP
+		   "option low: %s\n"
+#endif
 		   "option sbct: %beu\n"
 #if HAVE_ERTS_MSEG
 		   "option asbcst: %bpu\n"
@@ -2185,6 +2194,9 @@ info_options(Allctr_t *allctr,
 		   "option mbcgs: %beu\n",
 		   topt,
 		   allctr->ramv ? "true" : "false",
+#if HALFWORD_HEAP
+		   allctr->mseg_opt.low_mem ? "true" : "false",
+#endif
 		   allctr->sbc_threshold,
 #if HAVE_ERTS_MSEG
 		   allctr->mseg_opt.abs_shrink_th,
@@ -2243,6 +2255,9 @@ info_options(Allctr_t *allctr,
 	add_2tup(hpp, szp, &res,
 		 am.sbct,
 		 bld_uint(hpp, szp, allctr->sbc_threshold));
+#if HALFWORD_HEAP
+	add_2tup(hpp, szp, &res, am.low, allctr->mseg_opt.low_mem ? am_true : am_false);
+#endif
 	add_2tup(hpp, szp, &res, am.ramv, allctr->ramv ? am_true : am_false);
 	add_2tup(hpp, szp, &res, am.t, (allctr->t
 					? bld_uint(hpp, szp, (Uint) allctr->t)
@@ -3105,13 +3120,12 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init)
 	goto error;
 
 #if HAVE_ERTS_MSEG
-    {
-	ErtsMsegOpt_t mseg_opt = ERTS_MSEG_DEFAULT_OPT_INITIALIZER;
-    
-	sys_memcpy((void *) &allctr->mseg_opt,
-		   (void *) &mseg_opt,
-		   sizeof(ErtsMsegOpt_t));
-    }
+    sys_memcpy((void *) &allctr->mseg_opt,
+	       (void *) &erts_mseg_default_opt,
+	       sizeof(ErtsMsegOpt_t));
+# if HALFWORD_HEAP
+    allctr->mseg_opt.low_mem = init->low_mem;
+# endif
 #endif
 
     allctr->name_prefix			= init->name_prefix;
diff --git a/erts/emulator/beam/erl_alloc_util.h b/erts/emulator/beam/erl_alloc_util.h
index d296081714..ddf84c086c 100644
--- a/erts/emulator/beam/erl_alloc_util.h
+++ b/erts/emulator/beam/erl_alloc_util.h
@@ -38,6 +38,7 @@ typedef struct {
     int tspec;
     int tpref;
     int ramv;
+    int low_mem;   /* HALFWORD only */
     UWord sbct;
     UWord asbcst;
     UWord rsbcst;
@@ -70,6 +71,7 @@ typedef struct {
     0,			/* (bool)   tspec:  thread specific              */\
     0,			/* (bool)   tpref:  thread preferred             */\
     0,			/* (bool)   ramv:   realloc always moves         */\
+    0,			/* (bool)   low_mem: HALFWORD only               */\
     512*1024,		/* (bytes)  sbct:   sbc threshold                */\
     2*1024*2024,	/* (amount) asbcst: abs sbc shrink threshold     */\
     20,			/* (%)      rsbcst: rel sbc shrink threshold     */\
@@ -97,6 +99,7 @@ typedef struct {
     0,			/* (bool)   tspec:  thread specific              */\
     0,			/* (bool)   tpref:  thread preferred             */\
     0,			/* (bool)   ramv:   realloc always moves         */\
+    0,			/* (bool)   low_mem: HALFWORD only               */\
     64*1024,		/* (bytes)  sbct:   sbc threshold                */\
     2*1024*2024,	/* (amount) asbcst: abs sbc shrink threshold     */\
     20,			/* (%)      rsbcst: rel sbc shrink threshold     */\
diff --git a/erts/emulator/beam/erl_bif_ddll.c b/erts/emulator/beam/erl_bif_ddll.c
index c9cdcb87a6..9631fb50db 100644
--- a/erts/emulator/beam/erl_bif_ddll.c
+++ b/erts/emulator/beam/erl_bif_ddll.c
@@ -1,7 +1,7 @@
 /*
  * %CopyrightBegin%
  * 
- * Copyright Ericsson AB 2006-2010. All Rights Reserved.
+ * Copyright Ericsson AB 2006-2011. All Rights Reserved.
  * 
  * The contents of this file are subject to the Erlang Public License,
  * Version 1.1, (the "License"); you may not use this file except in
@@ -146,7 +146,7 @@ BIF_RETTYPE erl_ddll_try_load_3(Process *p, Eterm path_term,
 				Eterm name_term, Eterm options)
 {
     char *path = NULL;
-    int path_len;
+    Uint path_len;
     char *name = NULL;
     DE_Handle *dh;
     erts_driver_t *drv;
@@ -221,9 +221,7 @@ BIF_RETTYPE erl_ddll_try_load_3(Process *p, Eterm path_term,
 	goto error;
     }
 
-    path_len = io_list_len(path_term);
-
-    if (path_len <= 0) {
+    if (erts_iolist_size(path_term, &path_len)) {
 	goto error;
     }
     path = erts_alloc(ERTS_ALC_T_DDLL_TMP_BUF, path_len + 1 /* might need path separator */ + sys_strlen(name) + 1);
@@ -1878,7 +1876,7 @@ static Eterm mkatom(char *str)
 static char *pick_list_or_atom(Eterm name_term)
 { 
     char *name = NULL;
-    int name_len;
+    Uint name_len;
     if (is_atom(name_term)) {
 	Atom *ap = atom_tab(atom_val(name_term));
 	if (ap->len == 0) {
@@ -1890,8 +1888,7 @@ static char *pick_list_or_atom(Eterm name_term)
 	memcpy(name,ap->name,ap->len);
 	name[ap->len] = '\0';
     } else {
-	name_len = io_list_len(name_term);
-	if (name_len <= 0) {
+	if (erts_iolist_size(name_term, &name_len)) {
 	    goto error;
 	}
 	name = erts_alloc(ERTS_ALC_T_DDLL_TMP_BUF, name_len + 1);
diff --git a/erts/emulator/beam/erl_bif_info.c b/erts/emulator/beam/erl_bif_info.c
index e50fc18e64..f264bf44df 100644
--- a/erts/emulator/beam/erl_bif_info.c
+++ b/erts/emulator/beam/erl_bif_info.c
@@ -1732,14 +1732,14 @@ info_1_tuple(Process* BIF_P,	/* Pointer to current process. */
 #    define ERTS_ERROR_CHECKER_PRINTF_XML VALGRIND_PRINTF_XML
 #  endif
 #endif
-	    int buf_size = 8*1024; /* Try with 8KB first */
+	    Uint buf_size = 8*1024; /* Try with 8KB first */
 	    char *buf = erts_alloc(ERTS_ALC_T_TMP, buf_size);
 	    int r = io_list_to_buf(*tp, (char*) buf, buf_size - 1);
 	    if (r < 0) {
 		erts_free(ERTS_ALC_T_TMP, (void *) buf);
-		buf_size = io_list_len(*tp);
-		if (buf_size < 0)
+		if (erts_iolist_size(*tp, &buf_size)) {
 		    goto badarg;
+		}
 		buf_size++;
 		buf = erts_alloc(ERTS_ALC_T_TMP, buf_size);
 		r = io_list_to_buf(*tp, (char*) buf, buf_size - 1);
diff --git a/erts/emulator/beam/erl_bif_port.c b/erts/emulator/beam/erl_bif_port.c
index fbc92b9730..3fd35dd963 100644
--- a/erts/emulator/beam/erl_bif_port.c
+++ b/erts/emulator/beam/erl_bif_port.c
@@ -1,7 +1,7 @@
 /*
  * %CopyrightBegin%
  *
- * Copyright Ericsson AB 2001-2010. All Rights Reserved.
+ * Copyright Ericsson AB 2001-2011. All Rights Reserved.
  *
  * The contents of this file are subject to the Erlang Public License,
  * Version 1.1, (the "License"); you may not use this file except in
@@ -996,6 +996,7 @@ static byte* convert_environment(Process* p, Eterm env)
     Eterm* hp;
     Uint heap_size;
     int n;
+    Uint size;
     byte* bytes;
 
     if ((n = list_length(env)) < 0) {
@@ -1039,15 +1040,15 @@ static byte* convert_environment(Process* p, Eterm env)
     if (is_not_nil(env)) {
 	goto done;
     }
-    if ((n = io_list_len(all)) < 0) {
+    if (erts_iolist_size(all, &size)) {
 	goto done;
     }
 
     /*
      * Put the result in a binary (no risk for a memory leak that way).
      */
-    (void) erts_new_heap_binary(p, NULL, n, &bytes);
-    io_list_to_buf(all, (char*)bytes, n);
+    (void) erts_new_heap_binary(p, NULL, size, &bytes);
+    io_list_to_buf(all, (char*)bytes, size);
 
  done:
     erts_free(ERTS_ALC_T_TMP, temp_heap);
diff --git a/erts/emulator/beam/erl_bif_re.c b/erts/emulator/beam/erl_bif_re.c
index d4a8a3aaa7..26891c4348 100644
--- a/erts/emulator/beam/erl_bif_re.c
+++ b/erts/emulator/beam/erl_bif_re.c
@@ -1,7 +1,7 @@
 /*
  * %CopyrightBegin%
  *
- * Copyright Ericsson AB 2008-2010. All Rights Reserved.
+ * Copyright Ericsson AB 2008-2011. All Rights Reserved.
  *
  * The contents of this file are subject to the Erlang Public License,
  * Version 1.1, (the "License"); you may not use this file except in
@@ -417,7 +417,7 @@ build_compile_result(Process *p, Eterm error_tag, pcre *result, int errcode, con
 BIF_RETTYPE
 re_compile_2(BIF_ALIST_2)
 {
-    int slen;
+    Uint slen;
     char *expr;
     pcre *result;
     int errcode = 0;
@@ -444,7 +444,7 @@ re_compile_2(BIF_ALIST_2)
 	BIF_TRAP2(ucompile_trap_exportp, BIF_P, BIF_ARG_1, BIF_ARG_2);
     }
 
-    if ((slen = io_list_len(BIF_ARG_1)) < 0) {
+    if (erts_iolist_size(BIF_ARG_1, &slen)) {
         BIF_ERROR(BIF_P,BADARG);
     }
     expr = erts_alloc(ERTS_ALC_T_RE_TMP_BUF, slen + 1);
@@ -795,8 +795,8 @@ build_capture(Eterm capture_spec[CAPSPEC_SIZE], const pcre *code)
 			memcpy(tmpb,ap->name,ap->len);
 			tmpb[ap->len] = '\0';
 		    } else {
-			int slen = io_list_len(val);
-			if (slen < 0) {
+			Uint slen;
+			if (erts_iolist_size(val, &slen)) {
 			    goto error;
 			}
 			if ((slen + 1) > tmpbsiz) {
@@ -851,7 +851,7 @@ re_run_3(BIF_ALIST_3)
     const pcre *code_tmp;
     RestartContext restart;
     byte *temp_alloc = NULL;
-    int slength;
+    Uint slength;
     int startoffset = 0;
     int options = 0, comp_options = 0;
     int ovsize;
@@ -875,7 +875,7 @@ re_run_3(BIF_ALIST_3)
     if (is_not_tuple(BIF_ARG_2) || (arityval(*tuple_val(BIF_ARG_2)) != 4)) {
 	if (is_binary(BIF_ARG_2) || is_list(BIF_ARG_2) || is_nil(BIF_ARG_2)) { 
 	    /* Compile from textual RE */
-	    int slen;
+	    Uint slen;
 	    char *expr;
 	    pcre *result;
 	    int errcode = 0;
@@ -889,7 +889,7 @@ re_run_3(BIF_ALIST_3)
 		BIF_TRAP3(urun_trap_exportp, BIF_P, BIF_ARG_1, BIF_ARG_2, BIF_ARG_3);
 	    }
 	    
-	    if ((slen = io_list_len(BIF_ARG_2)) < 0) {
+	    if (erts_iolist_size(BIF_ARG_2, &slen)) {
 		BIF_ERROR(BIF_P,BADARG);
 	    }
 	    
@@ -1027,7 +1027,7 @@ re_run_3(BIF_ALIST_3)
 	restart.flags |= RESTART_FLAG_SUBJECT_IN_BINARY;
     } else {
 handle_iolist:
-	if ((slength = io_list_len(BIF_ARG_1)) < 0) {
+	if (erts_iolist_size(BIF_ARG_1, &slength)) {
 	    erts_free(ERTS_ALC_T_RE_SUBJECT, restart.ovector);
 	    erts_free(ERTS_ALC_T_RE_SUBJECT, restart.code);
 	    if (restart.ret_info != NULL) {
diff --git a/erts/emulator/beam/erl_db_hash.c b/erts/emulator/beam/erl_db_hash.c
index 9ef990cc4f..fdc82c8b88 100644
--- a/erts/emulator/beam/erl_db_hash.c
+++ b/erts/emulator/beam/erl_db_hash.c
@@ -105,7 +105,7 @@
 #define NSEG_2   256 /* Size of second segment table */
 #define NSEG_INC 128 /* Number of segments to grow after that */
 
-#define SEGTAB(tb) ((struct segment**)erts_smp_atomic_read(&(tb)->segtab))
+#define SEGTAB(tb) ((struct segment**)erts_smp_atomic_read_acqb(&(tb)->segtab))
 #define NACTIVE(tb) ((int)erts_smp_atomic_read(&(tb)->nactive))
 #define NITEMS(tb) ((int)erts_smp_atomic_read(&(tb)->common.nitems))
 
@@ -122,8 +122,8 @@
 */
 static ERTS_INLINE Uint hash_to_ix(DbTableHash* tb, HashValue hval)
 {
-    Uint mask = erts_smp_atomic_read(&tb->szm);
-    Uint ix = hval & mask; 
+    Uint mask = erts_smp_atomic_read_acqb(&tb->szm);
+    Uint ix = hval & mask;
     if (ix >= erts_smp_atomic_read(&tb->nactive)) {
 	ix &= mask>>1;
 	ASSERT(ix < erts_smp_atomic_read(&tb->nactive));
@@ -668,6 +668,7 @@ int db_create_hash(Process *p, DbTable *tbl)
     else { /* coarse locking */
 	tb->locks = NULL;
     }
+    ERTS_THR_MEMORY_BARRIER;
 #endif /* ERST_SMP */
     return DB_ERROR_NONE;
 }
@@ -2085,7 +2086,14 @@ static void db_print_hash(int to, void *to_arg, int show, DbTable *tbl)
 	    while(list != 0) {
 		if (list->hvalue == INVALID_HASH)
 		    erts_print(to, to_arg, "*");
-		erts_print(to, to_arg, "%T", make_tuple(list->dbterm.tpl));
+		if (tb->common.compress) {
+		    Eterm key = GETKEY(tb, list->dbterm.tpl);
+		    erts_print(to, to_arg, "key=%R", key, list->dbterm.tpl);
+		}
+		else {
+		    Eterm obj = make_tuple_rel(list->dbterm.tpl,list->dbterm.tpl);
+		    erts_print(to, to_arg, "%R", obj, list->dbterm.tpl);
+		}
 		if (list->next != 0)
 		    erts_print(to, to_arg, ",");
 		list = list->next;
@@ -2342,7 +2350,7 @@ static int alloc_seg(DbTableHash *tb)
 	    struct ext_segment* eseg;
 	    eseg = (struct ext_segment*) SEGTAB(tb)[seg_ix-1];
 	    MY_ASSERT(eseg!=NULL && eseg->s.is_ext_segment);
-	    erts_smp_atomic_set(&tb->segtab, (erts_aint_t) eseg->segtab);
+	    erts_smp_atomic_set_relb(&tb->segtab, (erts_aint_t) eseg->segtab);
 	    tb->nsegs = eseg->nsegs;
 	}
 	ASSERT(seg_ix < tb->nsegs);
@@ -2414,7 +2422,7 @@ static int free_seg(DbTableHash *tb, int free_records)
 	    MY_ASSERT(newtop->s.is_ext_segment);
 	    if (newtop->prev_segtab != NULL) {
 		/* Time to use a smaller segtab */
-		erts_smp_atomic_set(&tb->segtab, (erts_aint_t)newtop->prev_segtab);
+		erts_smp_atomic_set_relb(&tb->segtab, (erts_aint_t)newtop->prev_segtab);
 		tb->nsegs = seg_ix;
 		ASSERT(tb->nsegs == EXTSEG(SEGTAB(tb))->nsegs);
 	    }
@@ -2431,7 +2439,7 @@ static int free_seg(DbTableHash *tb, int free_records)
     if (seg_ix > 0) {
 	if (seg_ix < tb->nsegs) SEGTAB(tb)[seg_ix] = NULL;
     } else {
-	erts_smp_atomic_set(&tb->segtab, (erts_aint_t)NULL);
+	erts_smp_atomic_set_relb(&tb->segtab, (erts_aint_t)NULL);
     }
 #endif
     tb->nslots -= SEGSZ;
@@ -2526,9 +2534,9 @@ static void grow(DbTableHash* tb, int nactive)
     }
     erts_smp_atomic_inc(&tb->nactive);
     if (from_ix == 0) {
-	erts_smp_atomic_set(&tb->szm, szm);
+	erts_smp_atomic_set_relb(&tb->szm, szm);
     }
-    erts_smp_atomic_set(&tb->is_resizing, 0);
+    erts_smp_atomic_set_relb(&tb->is_resizing, 0);
 
     /* Finally, let's split the bucket. We try to do it in a smart way
        to keep link order and avoid unnecessary updates of next-pointers */
@@ -2560,7 +2568,7 @@ static void grow(DbTableHash* tb, int nactive)
     return;
    
 abort:
-    erts_smp_atomic_set(&tb->is_resizing, 0);
+    erts_smp_atomic_set_relb(&tb->is_resizing, 0);
 }
 
 
@@ -2604,7 +2612,7 @@ static void shrink(DbTableHash* tb, int nactive)
 	    
 	    erts_smp_atomic_set(&tb->nactive, src_ix);
 	    if (dst_ix == 0) {
-		erts_smp_atomic_set(&tb->szm, low_szm);
+		erts_smp_atomic_set_relb(&tb->szm, low_szm);
 	    }
 	    WUNLOCK_HASH(lck);
 	    
@@ -2618,7 +2626,7 @@ static void shrink(DbTableHash* tb, int nactive)
 
     }
     /*else already done */
-    erts_smp_atomic_set(&tb->is_resizing, 0);
+    erts_smp_atomic_set_relb(&tb->is_resizing, 0);
 }
 
 
diff --git a/erts/emulator/beam/erl_db_tree.c b/erts/emulator/beam/erl_db_tree.c
index a59c0c258d..9a0ba3a418 100644
--- a/erts/emulator/beam/erl_db_tree.c
+++ b/erts/emulator/beam/erl_db_tree.c
@@ -111,7 +111,7 @@ static void release_stack(DbTableTree* tb, DbTreeStack* stack)
 {
     if (stack == &tb->static_stack) {
 	ASSERT(erts_smp_atomic_read(&tb->is_stack_busy) == 1);
-	erts_smp_atomic_set(&tb->is_stack_busy, 0);
+	erts_smp_atomic_set_relb(&tb->is_stack_busy, 0);
     }
     else {
 	erts_db_free(ERTS_ALC_T_DB_STK, (DbTable *) tb,
@@ -179,7 +179,7 @@ static ERTS_INLINE TreeDbTerm* replace_dbterm(DbTableTree *tb, TreeDbTerm* old,
 static TreeDbTerm *traverse_until(TreeDbTerm *t, int *current, int to);
 static void check_slot_pos(DbTableTree *tb);
 static void check_saved_stack(DbTableTree *tb);
-static int check_table_tree(TreeDbTerm *t);
+static int check_table_tree(DbTableTree* tb, TreeDbTerm *t);
 
 #define TREE_DEBUG
 #endif
@@ -194,8 +194,8 @@ static int check_table_tree(TreeDbTerm *t);
 ** Debugging dump
 */
 
-static void do_dump_tree2(int to, void *to_arg, int show, TreeDbTerm *t,
-			  int offset);
+static void do_dump_tree2(DbTableTree*, int to, void *to_arg, int show,
+			  TreeDbTerm *t, int offset);
 
 #else
 
@@ -1730,6 +1730,7 @@ static int db_select_delete_tree(Process *p, DbTable *tbl,
 ** Other interface routines (not directly coupled to one bif)
 */
 
+
 /* Display tree contents (for dump) */
 static void db_print_tree(int to, void *to_arg, 
 			  int show,
@@ -1740,7 +1741,7 @@ static void db_print_tree(int to, void *to_arg,
     if (show)
 	erts_print(to, to_arg, "\nTree data dump:\n"
 		   "------------------------------------------------\n");
-    do_dump_tree2(to, to_arg, show, tb->root, 0);
+    do_dump_tree2(&tbl->tree, to, to_arg, show, tb->root, 0);
     if (show)
 	erts_print(to, to_arg, "\n"
 		   "------------------------------------------------\n");
@@ -2694,7 +2695,7 @@ static Sint do_cmp_partly_bound(Eterm a, Eterm b, Eterm* b_base, int *done)
 	while (1) {
 	    if ((j = do_cmp_partly_bound(*aa++, *bb++, b_base, done)) != 0 || *done)
 		return j;
-	    if (*aa==*bb)
+	    if (is_same(*aa, NULL, *bb, b_base))
 		return 0;
 	    if (is_not_list(*aa) || is_not_list(*bb))
 		return do_cmp_partly_bound(*aa, *bb, b_base, done);
@@ -2742,7 +2743,7 @@ static Sint cmp_partly_bound(Eterm partly_bound_key, Eterm bound_key, Eterm* bk_
 	erts_fprintf(stderr," > ");
     else
 	erts_fprintf(stderr," == ");
-    erts_fprintf(stderr,"%T\n",bound_key);  // HALFWORD BUG: printing rterm
+    erts_fprintf(stderr,"%R\n", bound_key, bk_base);
 #endif
     return ret;
 }
@@ -3084,19 +3085,28 @@ static int doit_select_delete(DbTableTree *tb, TreeDbTerm *this, void *ptr,
 }
 
 #ifdef TREE_DEBUG
-static void do_dump_tree2(int to, void *to_arg, int show, TreeDbTerm *t,
-			  int offset)
+static void do_dump_tree2(DbTableTree* tb, int to, void *to_arg, int show,
+			  TreeDbTerm *t, int offset)
 {
     if (t == NULL)
-	return 0;
-    do_dump_tree2(to, to_arg, show, t->right, offset + 4);
+	return;
+    do_dump_tree2(tb, to, to_arg, show, t->right, offset + 4);
     if (show) {
-	erts_print(to, to_arg, "%*s%T (addr = %p, bal = %d)\n"
-		   offset, "", make_tuple(t->dbterm.tpl),
+	const char* prefix;
+	Eterm term;
+	if (tb->common.compress) {
+	    prefix = "key=";
+	    term = GETKEY(tb, t->dbterm.tpl);
+	}
+	else {
+	    prefix = "";
+	    term = make_tuple_rel(t->dbterm.tpl,t->dbterm.tpl);
+	}
+	erts_print(to, to_arg, "%*s%s%R (addr = %p, bal = %d)\n",
+		   offset, "", prefix, term, t->dbterm.tpl,
 		   t, t->balance);
     }
-    do_dump_tree2(to, to_arg, show, t->left, offset + 4); 
-    return sum;
+    do_dump_tree2(tb, to, to_arg, show, t->left, offset + 4); 
 }
 
 #endif
@@ -3106,7 +3116,7 @@ static void do_dump_tree2(int to, void *to_arg, int show, TreeDbTerm *t,
 void db_check_table_tree(DbTable *tbl)
 {
     DbTableTree *tb = &tbl->tree;
-    check_table_tree(tb->root);
+    check_table_tree(tb, tb->root);
     check_saved_stack(tb);
     check_slot_pos(tb);
 }
@@ -3137,7 +3147,7 @@ static void check_slot_pos(DbTableTree *tb)
 		   "element position %d is really 0x%08X, when stack says "
 		   "it's 0x%08X\n", tb->stack.slot, t, 
 		   tb->stack.array[tb->stack.pos - 1]);
-	do_dump_tree2(ERTS_PRINT_STDERR, NULL, 1, tb->root, 0);
+	do_dump_tree2(tb, ERTS_PRINT_STDERR, NULL, 1, tb->root, 0);
     }
 }
 	
@@ -3152,14 +3162,14 @@ static void check_saved_stack(DbTableTree *tb)
      if (t != stack->array[0]) {
 	 erts_fprintf(stderr,"tb->stack[0] is 0x%08X, should be 0x%08X\n",
 		      stack->array[0], t);
-	 do_dump_tree2(ERTS_PRINT_STDERR, NULL, 1, tb->root, 0);
+	 do_dump_tree2(tb, ERTS_PRINT_STDERR, NULL, 1, tb->root, 0);
 	 return;
      }
      while (n < stack->pos) {
 	 if (t == NULL) {
 	     erts_fprintf(stderr, "NULL pointer in tree when stack not empty,"
 			" stack depth is %d\n", n);
-	     do_dump_tree2(ERTS_PRINT_STDERR, NULL, 1, tb->root, 0);
+	     do_dump_tree2(tb, ERTS_PRINT_STDERR, NULL, 1, tb->root, 0);
 	     return;
 	 }
 	 n++;
@@ -3173,28 +3183,26 @@ static void check_saved_stack(DbTableTree *tb)
 			    "represent child pointer in tree!"
 			    "(left == 0x%08X, right == 0x%08X\n", 
 			    n, tb->stack[n], t->left, t->right);
-		 do_dump_tree2(ERTS_PRINT_STDERR, NULL, 1, tb->root, 0);
+		 do_dump_tree2(tb, ERTS_PRINT_STDERR, NULL, 1, tb->root, 0);
 		 return;
 	     }
 	 }
      }
 }
 
-static int check_table_tree(TreeDbTerm *t)
+static int check_table_tree(DbTableTree* tb, TreeDbTerm *t)
 {
     int lh, rh;
     if (t == NULL)
 	return 0;
-    lh = check_table_tree(t->left);
-    rh = check_table_tree(t->right);
+    lh = check_table_tree(tb, t->left);
+    rh = check_table_tree(tb, t->right);
     if ((rh - lh) != t->balance) {
 	erts_fprintf(stderr, "Invalid tree balance for this node:\n");
-	erts_fprintf(stderr,"balance = %d, left = 0x%08X, right = 0x%08X\n"
-		     "data = %T",
-		     t->balance, t->left, t->right,
-		     make_tuple(t->dbterm.tpl));
+	erts_fprintf(stderr,"balance = %d, left = 0x%08X, right = 0x%08X\n",
+		     t->balance, t->left, t->right);
 	erts_fprintf(stderr,"\nDump:\n---------------------------------\n");
-	do_dump_tree2(ERTS_PRINT_STDERR, NULL, 1, t, 0);
+	do_dump_tree2(tb, ERTS_PRINT_STDERR, NULL, 1, t, 0);
 	erts_fprintf(stderr,"\n---------------------------------\n");
     }
     return ((rh > lh) ? rh : lh) + 1;
diff --git a/erts/emulator/beam/erl_drv_thread.c b/erts/emulator/beam/erl_drv_thread.c
index 39bbe9633b..dc578f6d2a 100644
--- a/erts/emulator/beam/erl_drv_thread.c
+++ b/erts/emulator/beam/erl_drv_thread.c
@@ -700,6 +700,13 @@ erl_drv_thread_join(ErlDrvTid tid, void **respp)
 extern int erts_darwin_main_thread_pipe[2];
 extern int erts_darwin_main_thread_result_pipe[2];
 
+int erl_drv_stolen_main_thread_join(ErlDrvTid tid, void **respp);
+int erl_drv_steal_main_thread(char *name,
+			      ErlDrvTid *dtid,
+			      void* (*func)(void*),
+			      void* arg,
+			      ErlDrvThreadOpts *opts);
+
 
 int
 erl_drv_stolen_main_thread_join(ErlDrvTid tid, void **respp)
diff --git a/erts/emulator/beam/erl_lock_check.c b/erts/emulator/beam/erl_lock_check.c
index 9e18997890..9180508a49 100644
--- a/erts/emulator/beam/erl_lock_check.c
+++ b/erts/emulator/beam/erl_lock_check.c
@@ -240,7 +240,7 @@ typedef struct erts_lc_locked_lock_t_ erts_lc_locked_lock_t;
 struct erts_lc_locked_lock_t_ {
     erts_lc_locked_lock_t *next;
     erts_lc_locked_lock_t *prev;
-    Eterm extra;
+    UWord extra;
     Sint16 id;
     Uint16 flags;
 };
@@ -441,12 +441,12 @@ new_locked_lock(erts_lc_lock_t *lck, Uint16 op_flags)
 }
 
 static void
-print_lock2(char *prefix, Sint16 id, Eterm extra, Uint16 flags, char *suffix)
+print_lock2(char *prefix, Sint16 id, Wterm extra, Uint16 flags, char *suffix)
 {
     char *lname = (0 <= id && id < ERTS_LOCK_ORDER_SIZE
 		   ? erts_lock_order[id].name
 		   : "unknown");
-    if (is_boxed(extra))
+    if (is_not_immed(extra))
 	erts_fprintf(stderr,
 		     "%s'%s:%p%s'%s%s",
 		     prefix,
@@ -1260,7 +1260,8 @@ erts_lc_init_lock(erts_lc_lock_t *lck, char *name, Uint16 flags)
 {
     lck->id = erts_lc_get_lock_order_id(name);
 
-    lck->extra = make_boxed(&lck->extra);
+    lck->extra = &lck->extra;
+    ASSERT(is_not_immed(lck->extra));
     lck->flags = flags;
     lck->inited = ERTS_LC_INITITALIZED;
 }
@@ -1270,6 +1271,7 @@ erts_lc_init_lock_x(erts_lc_lock_t *lck, char *name, Uint16 flags, Eterm extra)
 {
     lck->id = erts_lc_get_lock_order_id(name);
     lck->extra = extra;
+    ASSERT(is_immed(lck->extra));
     lck->flags = flags;
     lck->inited = ERTS_LC_INITITALIZED;
 }
diff --git a/erts/emulator/beam/erl_lock_check.h b/erts/emulator/beam/erl_lock_check.h
index cdb06d4458..b67f36fa06 100644
--- a/erts/emulator/beam/erl_lock_check.h
+++ b/erts/emulator/beam/erl_lock_check.h
@@ -39,7 +39,7 @@ typedef struct {
     int inited;
     Sint16 id;
     Uint16 flags;
-    Eterm extra;
+    UWord extra;
 } erts_lc_lock_t;
 
 #define ERTS_LC_INITITALIZED 0x7f7f7f7f
diff --git a/erts/emulator/beam/erl_nif.c b/erts/emulator/beam/erl_nif.c
index 8b48444904..68421b4387 100644
--- a/erts/emulator/beam/erl_nif.c
+++ b/erts/emulator/beam/erl_nif.c
@@ -472,7 +472,7 @@ static void tmp_alloc_dtor(struct enif_tmp_obj_t* obj)
 int enif_inspect_iolist_as_binary(ErlNifEnv* env, Eterm term, ErlNifBinary* bin)
 {
     struct enif_tmp_obj_t* tobj;
-    int sz;
+    Uint sz;
     if (is_binary(term)) {
 	return enif_inspect_binary(env,term,bin);
     }
@@ -483,7 +483,7 @@ int enif_inspect_iolist_as_binary(ErlNifEnv* env, Eterm term, ErlNifBinary* bin)
 	bin->ref_bin = NULL;
 	return 1;
     }
-    if ((sz = io_list_len(term)) < 0) {
+    if (erts_iolist_size(term, &sz)) {
 	return 0;
     }
     
diff --git a/erts/emulator/beam/erl_port_task.c b/erts/emulator/beam/erl_port_task.c
index 1b07024ca1..326021643f 100644
--- a/erts/emulator/beam/erl_port_task.c
+++ b/erts/emulator/beam/erl_port_task.c
@@ -658,8 +658,6 @@ erts_port_task_free_port(Port *pp)
 	   when scheduled out... */
 	ErtsPortTask *ptp = port_task_alloc();
 	erts_smp_port_state_lock(pp);
-	ASSERT(erts_smp_atomic_read(&erts_ports_alive) > 0);
-	erts_smp_atomic_dec(&erts_ports_alive);
 	pp->status &= ~ERTS_PORT_SFLG_CLOSING;
 	pp->status |= ERTS_PORT_SFLG_FREE_SCHEDULED;
 	erts_may_save_closed_port(pp);
@@ -681,7 +679,6 @@ erts_port_task_free_port(Port *pp)
 	    port_is_dequeued = 1;
 	}
 	erts_smp_port_state_lock(pp);
-	erts_smp_atomic_dec(&erts_ports_alive);
 	pp->status &= ~ERTS_PORT_SFLG_CLOSING;
 	pp->status |= ERTS_PORT_SFLG_FREE_SCHEDULED;
 	erts_may_save_closed_port(pp);
diff --git a/erts/emulator/beam/erl_printf_term.c b/erts/emulator/beam/erl_printf_term.c
index b71404fd27..34da9cab84 100644
--- a/erts/emulator/beam/erl_printf_term.c
+++ b/erts/emulator/beam/erl_printf_term.c
@@ -114,13 +114,13 @@ do {									\
 /* return 0 if list is not a non-empty flat list of printable characters */
 
 static int
-is_printable_string(Eterm list)
+is_printable_string(Eterm list, Eterm* base)
 {
     int len = 0;
     int c;
 
     while(is_list(list)) {
-	Eterm* consp = list_val(list);
+	Eterm* consp = list_val_rel(list, base);
 	Eterm hd = CAR(consp);
 
 	if (!is_byte(hd))
@@ -226,17 +226,20 @@ static int print_atom_name(fmtfn_t fn, void* arg, Eterm atom, long *dcount)
 #define PRT_LAST_ARRAY_ELEMENT ((Eterm) 7) /* Note! Must be last... */
 
 static int
-print_term(fmtfn_t fn, void* arg, Eterm obj, long *dcount)
+print_term(fmtfn_t fn, void* arg, Eterm obj, long *dcount,
+	   Eterm* obj_base) /* ignored if !HALFWORD_HEAP */
 {
     DECLARE_WSTACK(s);
     int res;
     int i;
     Eterm val;
     Uint32 *ref_num;
+    union {
+	UWord word;
+	Eterm* ptr;
+    }popped;
     Eterm* nobj;
-#if HALFWORD_HEAP
-    UWord wobj;
-#endif
+    Wterm wobj;
 
     res = 0;
 
@@ -258,18 +261,17 @@ print_term(fmtfn_t fn, void* arg, Eterm obj, long *dcount)
 	    PRINT_CHAR(res, fn, arg, '}');
 	    goto L_outer_loop;
 	default:
-#if HALFWORD_HEAP
-	    obj = (Eterm) (wobj = WSTACK_POP(s));
-#else
-	    obj = WSTACK_POP(s);
-#endif
+	    popped.word = WSTACK_POP(s);
+
 	    switch (val) {
 	    case PRT_TERM:
+		obj = (Eterm) popped.word;
 		break;
 	    case PRT_ONE_CONS:
+		obj = (Eterm) popped.word;
 	    L_print_one_cons:
 		{
-		    Eterm* cons = list_val(obj);
+		    Eterm* cons = list_val_rel(obj, obj_base);
 		    Eterm tl;
 		    
 		    obj = CAR(cons);
@@ -288,27 +290,13 @@ print_term(fmtfn_t fn, void* arg, Eterm obj, long *dcount)
 		}
 		break;
 	    case PRT_LAST_ARRAY_ELEMENT:
-		{
-#if HALFWORD_HEAP
-		    Eterm* ptr = (Eterm *) wobj;
-#else
-		    Eterm* ptr = (Eterm *) obj;
-#endif
-		    obj = *ptr;
-		}
+		obj = *popped.ptr;
 		break;
 	    default:		/* PRT_LAST_ARRAY_ELEMENT+1 and upwards */
-		{
-#if HALFWORD_HEAP
-		    Eterm* ptr = (Eterm *) wobj;
-#else
-		    Eterm* ptr = (Eterm *) obj;
-#endif
-		    obj = *ptr++;
-		    WSTACK_PUSH(s, (UWord) ptr);
-		    WSTACK_PUSH(s, val-1);
-		    WSTACK_PUSH(s, PRT_COMMA);
-		}
+		obj = *popped.ptr;
+	        WSTACK_PUSH(s, (UWord) (popped.ptr + 1));
+		WSTACK_PUSH(s, val-1);
+		WSTACK_PUSH(s, PRT_COMMA);
 		break;
 	    }
 	    break;
@@ -325,8 +313,12 @@ print_term(fmtfn_t fn, void* arg, Eterm obj, long *dcount)
 	    PRINT_CHAR(res, fn, arg, '>');
 	    goto L_done;
 	}
-
-	switch (tag_val_def(obj)) {
+#if HALFWORD_HEAP
+	wobj = is_immed(obj) ? (Wterm)obj : rterm2wterm(obj, obj_base);
+#else
+	wobj = (Wterm)obj;
+#endif
+	switch (tag_val_def(wobj)) {
 	case NIL_DEF:
 	    PRINT_STRING(res, fn, arg, "[]");
 	    break;
@@ -348,13 +340,13 @@ print_term(fmtfn_t fn, void* arg, Eterm obj, long *dcount)
 	    int print_res;
 	    char def_buf[64];
 	    char *buf, *big_str;
-	    Uint sz = (Uint) big_decimal_estimate(obj);
+	    Uint sz = (Uint) big_decimal_estimate(wobj);
 	    sz++;
 	    if (sz <= 64)
 		buf = &def_buf[0];
 	    else
 		buf = erts_alloc(ERTS_ALC_T_TMP, sz);
-	    big_str = erts_big_to_string(obj, buf, sz);
+	    big_str = erts_big_to_string(wobj, buf, sz);
 	    print_res = erts_printf_string(fn, arg, big_str);
 	    if (buf != &def_buf[0])
 		erts_free(ERTS_ALC_T_TMP, (void *) buf);
@@ -369,9 +361,9 @@ print_term(fmtfn_t fn, void* arg, Eterm obj, long *dcount)
 	case EXTERNAL_REF_DEF:
 	    PRINT_STRING(res, fn, arg, "#Ref<");
 	    PRINT_ULONG(res, fn, arg, 'u', 0, 1,
-			(unsigned long) ref_channel_no(obj));
-	    ref_num = ref_numbers(obj);
-	    for (i = ref_no_of_numbers(obj)-1; i >= 0; i--) {
+			(unsigned long) ref_channel_no(wobj));
+	    ref_num = ref_numbers(wobj);
+	    for (i = ref_no_of_numbers(wobj)-1; i >= 0; i--) {
 		PRINT_CHAR(res, fn, arg, '.');
 		PRINT_ULONG(res, fn, arg, 'u', 0, 1, (unsigned long) ref_num[i]);
 	    }
@@ -381,30 +373,30 @@ print_term(fmtfn_t fn, void* arg, Eterm obj, long *dcount)
 	case EXTERNAL_PID_DEF:
 	    PRINT_CHAR(res, fn, arg, '<');
 	    PRINT_ULONG(res, fn, arg, 'u', 0, 1,
-			(unsigned long) pid_channel_no(obj));
+			(unsigned long) pid_channel_no(wobj));
 	    PRINT_CHAR(res, fn, arg, '.');
 	    PRINT_ULONG(res, fn, arg, 'u', 0, 1,
-			(unsigned long) pid_number(obj));
+			(unsigned long) pid_number(wobj));
 	    PRINT_CHAR(res, fn, arg, '.');
 	    PRINT_ULONG(res, fn, arg, 'u', 0, 1,
-			(unsigned long) pid_serial(obj));
+			(unsigned long) pid_serial(wobj));
 	    PRINT_CHAR(res, fn, arg, '>');
 	    break;
 	case PORT_DEF:
 	case EXTERNAL_PORT_DEF:
 	    PRINT_STRING(res, fn, arg, "#Port<");
 	    PRINT_ULONG(res, fn, arg, 'u', 0, 1,
-			(unsigned long) port_channel_no(obj));
+			(unsigned long) port_channel_no(wobj));
 	    PRINT_CHAR(res, fn, arg, '.');
 	    PRINT_ULONG(res, fn, arg, 'u', 0, 1,
-			(unsigned long) port_number(obj));
+			(unsigned long) port_number(wobj));
 	    PRINT_CHAR(res, fn, arg, '>');
 	    break;
 	case LIST_DEF:
-	    if (is_printable_string(obj)) {
+	    if (is_printable_string(obj, obj_base)) {
 		int c;
 		PRINT_CHAR(res, fn, arg, '"');
-		nobj = list_val(obj);
+		nobj = list_val_rel(obj, obj_base);
 		while (1) {
 		    if ((*dcount)-- <= 0)
 			goto L_done;
@@ -418,7 +410,7 @@ print_term(fmtfn_t fn, void* arg, Eterm obj, long *dcount)
 		    }
 		    if (is_not_list(*nobj))
 			break;
-		    nobj = list_val(*nobj);
+		    nobj = list_val_rel(*nobj, obj_base);
 		}
 		PRINT_CHAR(res, fn, arg, '"');
 	    } else {
@@ -428,7 +420,7 @@ print_term(fmtfn_t fn, void* arg, Eterm obj, long *dcount)
 	    }
 	    break;
 	case TUPLE_DEF:
-	    nobj = tuple_val(obj);	/* pointer to arity */
+	    nobj = tuple_val(wobj);	/* pointer to arity */
 	    i = arityval(*nobj);	/* arity */
 	    PRINT_CHAR(res, fn, arg, '{');
 	    WSTACK_PUSH(s,PRT_CLOSE_TUPLE);
@@ -440,13 +432,13 @@ print_term(fmtfn_t fn, void* arg, Eterm obj, long *dcount)
 	    break;
 	case FLOAT_DEF: {
 	    FloatDef ff;
-	    GET_DOUBLE(obj, ff);
+	    GET_DOUBLE(wobj, ff);
 	    PRINT_DOUBLE(res, fn, arg, 'e', 6, 0, ff.fd);
 	}
 	    break;
 	case BINARY_DEF:
 	    {
-		ProcBin* pb = (ProcBin *) binary_val(obj);
+		ProcBin* pb = (ProcBin *) binary_val(wobj);
 		if (pb->size == 1)
 		    PRINT_STRING(res, fn, arg, "<<1 byte>>");
 		else {
@@ -458,7 +450,7 @@ print_term(fmtfn_t fn, void* arg, Eterm obj, long *dcount)
 	    break;
 	case EXPORT_DEF:
 	    {
-		Export* ep = *((Export **) (export_val(obj) + 1));
+		Export* ep = *((Export **) (export_val(wobj) + 1));
 		Atom* module = atom_tab(atom_val(ep->code[0]));
 		Atom* name = atom_tab(atom_val(ep->code[1]));
 
@@ -474,7 +466,7 @@ print_term(fmtfn_t fn, void* arg, Eterm obj, long *dcount)
 	    break;
 	case FUN_DEF:
 	    {
-		ErlFunThing *funp = (ErlFunThing *) fun_val(obj);
+		ErlFunThing *funp = (ErlFunThing *) fun_val(wobj);
 		Atom *ap = atom_tab(atom_val(funp->fe->module));
 
 		PRINT_STRING(res, fn, arg, "#Fun<");
@@ -490,7 +482,7 @@ print_term(fmtfn_t fn, void* arg, Eterm obj, long *dcount)
 	    break;
 	default:
 	    PRINT_STRING(res, fn, arg, "<unknown:");
-	    PRINT_POINTER(res, fn, arg, (UWord) obj);
+	    PRINT_POINTER(res, fn, arg, wobj);
 	    PRINT_CHAR(res, fn, arg, '>');
 	    break;
 	}
@@ -503,9 +495,10 @@ print_term(fmtfn_t fn, void* arg, Eterm obj, long *dcount)
 }
 
 int
-erts_printf_term(fmtfn_t fn, void* arg, unsigned long term, long precision)
+erts_printf_term(fmtfn_t fn, void* arg, unsigned long term, long precision,
+		 unsigned long* term_base)
 {
-    int res = print_term(fn, arg, (Uint) term, &precision);
+    int res = print_term(fn, arg, (Eterm)term, &precision, (Eterm*)term_base);
     if (res < 0)
 	return res;
     if (precision <= 0)
diff --git a/erts/emulator/beam/erl_printf_term.h b/erts/emulator/beam/erl_printf_term.h
index 4f76028396..4ba22f12de 100644
--- a/erts/emulator/beam/erl_printf_term.h
+++ b/erts/emulator/beam/erl_printf_term.h
@@ -21,6 +21,6 @@
 #define ERL_PRINTF_TERM_H__
 
 #include "erl_printf_format.h"
-int erts_printf_term(fmtfn_t fn, void* arg, unsigned long term, long precision);
-
+int erts_printf_term(fmtfn_t fn, void* arg, unsigned long term, long precision,
+		     unsigned long* term_base);
 #endif
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index 8a56976905..2704359a8f 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -166,9 +166,8 @@ static struct {
 
 static struct {
     erts_smp_mtx_t update_mtx;
-    erts_smp_atomic32_t active_runqs;
+    erts_smp_atomic32_t no_runqs;
     int last_active_runqs;
-    erts_smp_atomic32_t used_runqs;
     int forced_check_balance;
     erts_smp_atomic32_t checking_balance;
     int halftime;
@@ -965,7 +964,7 @@ sched_spin_wait(ErtsSchedulerSleepInfo *ssi, int spincount)
     erts_aint32_t flgs;
 
     do {
-	flgs = erts_smp_atomic32_read(&ssi->flags);
+	flgs = erts_smp_atomic32_read_acqb(&ssi->flags);
 	if ((flgs & (ERTS_SSI_FLG_SLEEPING|ERTS_SSI_FLG_WAITING))
 	    != (ERTS_SSI_FLG_SLEEPING|ERTS_SSI_FLG_WAITING)) {
 	    break;
@@ -1114,7 +1113,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq)
     else {
 	erts_aint_t dt;
 
-	erts_smp_atomic32_set(&function_calls, 0);
+	erts_smp_atomic32_set_relb(&function_calls, 0);
 	*fcalls = 0;
 
 	sched_waiting_sys(esdp->no, rq);
@@ -1147,7 +1146,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq)
 	    nonblockable_aux_work(esdp, ssi, aux_work);
 #endif
 
-	    flgs = erts_smp_atomic32_read(&ssi->flags);
+	    flgs = erts_smp_atomic32_read_acqb(&ssi->flags);
 	    if (!(flgs & ERTS_SSI_FLG_WAITING)) {
 		ASSERT(!(flgs & ERTS_SSI_FLG_SLEEPING));
 		goto sys_woken;
@@ -1333,6 +1332,76 @@ wake_all_schedulers(void)
     }
 }
 
+#define ERTS_NO_USED_RUNQS_SHIFT 16
+#define ERTS_NO_RUNQS_MASK 0xffff
+
+#if ERTS_MAX_NO_OF_SCHEDULERS > ERTS_NO_RUNQS_MASK
+#  error "Too large amount of schedulers allowed"
+#endif
+
+static ERTS_INLINE void
+init_no_runqs(int active, int used)
+{
+    erts_aint32_t no_runqs = (erts_aint32_t) (active & ERTS_NO_RUNQS_MASK);
+    no_runqs |= (erts_aint32_t) ((used & ERTS_NO_RUNQS_MASK) << ERTS_NO_USED_RUNQS_SHIFT);
+    erts_smp_atomic32_init(&balance_info.no_runqs, no_runqs);
+}
+
+static ERTS_INLINE void
+get_no_runqs(int *active, int *used)
+{
+    erts_aint32_t no_runqs = erts_smp_atomic32_read(&balance_info.no_runqs);
+    if (active)
+	*active = (int) (no_runqs & ERTS_NO_RUNQS_MASK);
+    if (used)
+	*used = (int) ((no_runqs >> ERTS_NO_USED_RUNQS_SHIFT) & ERTS_NO_RUNQS_MASK);
+}
+
+static ERTS_INLINE void
+set_no_used_runqs(int used)
+{
+    erts_aint32_t exp = erts_smp_atomic32_read(&balance_info.no_runqs);
+    while (1) {
+	erts_aint32_t act, new;
+	new = (used << ERTS_NO_USED_RUNQS_SHIFT) | (exp & ERTS_NO_RUNQS_MASK);
+	act = erts_smp_atomic32_cmpxchg(&balance_info.no_runqs, new, exp);
+	if (act == exp)
+	    break;
+	exp = act;
+    }
+}
+
+static ERTS_INLINE void
+set_no_active_runqs(int active)
+{
+    erts_aint32_t exp = erts_smp_atomic32_read(&balance_info.no_runqs);
+    while (1) {
+	erts_aint32_t act, new;
+	new = (exp & (ERTS_NO_RUNQS_MASK << ERTS_NO_USED_RUNQS_SHIFT)) | active;
+	act = erts_smp_atomic32_cmpxchg(&balance_info.no_runqs, new, exp);
+	if (act == exp)
+	    break;
+	exp = act;
+    }
+}
+
+static ERTS_INLINE int
+try_inc_no_active_runqs(int active)
+{
+    erts_aint32_t exp = erts_smp_atomic32_read(&balance_info.no_runqs);
+    if (((exp >> ERTS_NO_USED_RUNQS_SHIFT) & ERTS_NO_RUNQS_MASK) < active)
+	return 0;
+    if ((exp & ERTS_NO_RUNQS_MASK) + 1 == active) {
+	erts_aint32_t new, act;
+	new = (exp & ~ERTS_NO_RUNQS_MASK) | active;
+	act = erts_smp_atomic32_cmpxchg(&balance_info.no_runqs, new, exp);
+	if (act == exp)
+	    return 1;
+    }
+    return 0;
+}
+
+
 static ERTS_INLINE int
 chk_wake_sched(ErtsRunQueue *crq, int ix, int activate)
 {
@@ -1344,9 +1413,7 @@ chk_wake_sched(ErtsRunQueue *crq, int ix, int activate)
     iflgs = erts_smp_atomic32_read(&wrq->info_flags);
     if (!(iflgs & (ERTS_RUNQ_IFLG_SUSPENDED|ERTS_RUNQ_IFLG_NONEMPTY))) {
 	if (activate) {
-	    if (ix == erts_smp_atomic32_cmpxchg(&balance_info.active_runqs,
-						ix+1,
-						ix)) {
+	    if (try_inc_no_active_runqs(ix+1)) {
 		erts_smp_xrunq_lock(crq, wrq);
 		wrq->flags &= ~ERTS_RUNQ_FLG_INACTIVE;
 		erts_smp_xrunq_unlock(crq, wrq);
@@ -1363,8 +1430,9 @@ wake_scheduler_on_empty_runq(ErtsRunQueue *crq)
 {
     int ix = crq->ix;
     int stop_ix = ix;
-    int active_ix = erts_smp_atomic32_read(&balance_info.active_runqs);
-    int balance_ix = erts_smp_atomic32_read(&balance_info.used_runqs);
+    int active_ix, balance_ix;
+
+    get_no_runqs(&active_ix, &balance_ix);
 
     if (active_ix > balance_ix)
 	active_ix = balance_ix;
@@ -1416,7 +1484,7 @@ erts_sched_notify_check_cpu_bind(void)
     int ix;
     if (erts_common_run_queue) {
 	for (ix = 0; ix < erts_no_schedulers; ix++)
-	    erts_smp_atomic32_set(&ERTS_SCHEDULER_IX(ix)->chk_cpu_bind, 1);
+	    erts_smp_atomic32_set_relb(&ERTS_SCHEDULER_IX(ix)->chk_cpu_bind, 1);
 	wake_all_schedulers();
     }
     else {
@@ -1871,8 +1939,7 @@ try_steal_task(ErtsRunQueue *rq)
 
     ERTS_SMP_LC_CHK_RUNQ_LOCK(rq, rq_locked);
 
-    active_rqs = erts_smp_atomic32_read(&balance_info.active_runqs);
-    blnc_rqs = erts_smp_atomic32_read(&balance_info.used_runqs);
+    get_no_runqs(&active_rqs, &blnc_rqs);
 
     if (active_rqs > blnc_rqs)
 	active_rqs = blnc_rqs;
@@ -1883,7 +1950,7 @@ try_steal_task(ErtsRunQueue *rq)
 	if (active_rqs < blnc_rqs) {
 	    int no = blnc_rqs - active_rqs;
 	    int stop_ix = vix = active_rqs + rq->ix % no;
-	    while (erts_smp_atomic32_read(&no_empty_run_queues) < blnc_rqs) {
+	    while (erts_smp_atomic32_read_acqb(&no_empty_run_queues) < blnc_rqs) {
 		res = check_possible_steal_victim(rq, &rq_locked, vix);
 		if (res)
 		    goto done;
@@ -1898,7 +1965,7 @@ try_steal_task(ErtsRunQueue *rq)
 	vix = rq->ix;
 
 	/* ... then try to steal a job from another active queue... */
-	while (erts_smp_atomic32_read(&no_empty_run_queues) < blnc_rqs) {
+	while (erts_smp_atomic32_read_acqb(&no_empty_run_queues) < blnc_rqs) {
 	    vix++;
 	    if (vix >= active_rqs)
 		vix = 0;
@@ -1999,7 +2066,7 @@ check_balance(ErtsRunQueue *c_rq)
 	return;
     }
 
-    blnc_no_rqs = (int) erts_smp_atomic32_read(&balance_info.used_runqs);
+    get_no_runqs(NULL, &blnc_no_rqs);
     if (blnc_no_rqs == 1) {
 	c_rq->check_balance_reds = INT_MAX;
 	erts_smp_atomic32_set(&balance_info.checking_balance, 0);
@@ -2038,7 +2105,8 @@ check_balance(ErtsRunQueue *c_rq)
     forced = balance_info.forced_check_balance;
     balance_info.forced_check_balance = 0;
 
-    blnc_no_rqs = (int) erts_smp_atomic32_read(&balance_info.used_runqs);
+    get_no_runqs(&current_active, &blnc_no_rqs);
+
     if (blnc_no_rqs == 1) {
 	erts_smp_mtx_unlock(&balance_info.update_mtx);
 	erts_smp_runq_lock(c_rq);
@@ -2052,8 +2120,6 @@ check_balance(ErtsRunQueue *c_rq)
     if (balance_info.full_reds_history_index >= ERTS_FULL_REDS_HISTORY_SIZE)
 	balance_info.full_reds_history_index = 0;
 
-    current_active = erts_smp_atomic32_read(&balance_info.active_runqs);
-
     /* Read balance information for all run queues */
     for (qix = 0; qix < blnc_no_rqs; qix++) {
 	ErtsRunQueue *rq = ERTS_RUNQ_IX(qix);
@@ -2387,7 +2453,7 @@ erts_fprintf(stderr, "--------------------------------\n");
     }
 
     balance_info.last_active_runqs = active;
-    erts_smp_atomic32_set(&balance_info.active_runqs, active);
+    set_no_active_runqs(active);
 
     balance_info.halftime = 1;
     erts_smp_atomic32_set(&balance_info.checking_balance, 0);
@@ -2695,9 +2761,8 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online)
     erts_smp_atomic32_init(&schdlr_sspnd.msb.ongoing, 0);
     erts_smp_atomic32_init(&schdlr_sspnd.active, no_schedulers);
     schdlr_sspnd.msb.procs = NULL;
-    erts_smp_atomic32_set(&balance_info.used_runqs,
-			  erts_common_run_queue ? 1 : no_schedulers_online);
-    erts_smp_atomic32_init(&balance_info.active_runqs, no_schedulers);
+    init_no_runqs(no_schedulers,
+		  erts_common_run_queue ? 1 : no_schedulers_online);
     balance_info.last_active_runqs = no_schedulers;
     erts_smp_mtx_init(&balance_info.update_mtx, "migration_info_update");
     balance_info.forced_check_balance = 0;
@@ -2939,7 +3004,7 @@ sched_spin_suspended(ErtsSchedulerSleepInfo *ssi, int spincount)
     erts_aint32_t flgs;
 
     do {
-	flgs = erts_smp_atomic32_read(&ssi->flags);
+	flgs = erts_smp_atomic32_read_acqb(&ssi->flags);
 	if ((flgs & (ERTS_SSI_FLG_SLEEPING
 		     | ERTS_SSI_FLG_WAITING
 		     | ERTS_SSI_FLG_SUSPENDED))
@@ -3068,7 +3133,7 @@ suspend_scheduler(ErtsSchedulerData *esdp)
 		wake = 0;
 	    }
 
-	    flgs = erts_smp_atomic32_read(&ssi->flags);
+	    flgs = erts_smp_atomic32_read_acqb(&ssi->flags);
 	    if (!(flgs & ERTS_SSI_FLG_SUSPENDED))
 		break;
 	    erts_smp_mtx_unlock(&schdlr_sspnd.mtx);
@@ -3292,7 +3357,7 @@ erts_set_schedulers_online(Process *p,
 			ErtsRunQueue *to_rq = ERTS_RUNQ_IX(ix % no);
 			evacuate_run_queue(from_rq, to_rq);
 		    }
-		    erts_smp_atomic32_set(&balance_info.used_runqs, no);
+		    set_no_used_runqs(no);
 		    erts_smp_mtx_unlock(&balance_info.update_mtx);
 		    erts_smp_mtx_lock(&schdlr_sspnd.mtx);
 		}
@@ -3346,7 +3411,7 @@ erts_set_schedulers_online(Process *p,
 		    for (ix = erts_no_run_queues-1; ix >= no; ix--)
 			evacuate_run_queue(ERTS_RUNQ_IX(ix),
 					   ERTS_RUNQ_IX(ix % no));
-		    erts_smp_atomic32_set(&balance_info.used_runqs, no);
+		    set_no_used_runqs(no);
 		    erts_smp_mtx_unlock(&balance_info.update_mtx);
 		    erts_smp_mtx_lock(&schdlr_sspnd.mtx);
 		    for (ix = no; ix < online; ix++) {
@@ -3443,7 +3508,7 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int all)
 		else {
 		    erts_smp_mtx_unlock(&schdlr_sspnd.mtx);
 		    erts_smp_mtx_lock(&balance_info.update_mtx);
-		    erts_smp_atomic32_set(&balance_info.used_runqs, 1);
+		    set_no_used_runqs(1);
 		    for (ix = 0; ix < online; ix++) {
 			ErtsRunQueue *rq = ERTS_RUNQ_IX(ix);
 			erts_smp_runq_lock(rq);
@@ -3580,7 +3645,7 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int all)
 		    evacuate_run_queue(ERTS_RUNQ_IX(ix),
 				       ERTS_RUNQ_IX(ix % online));
 
-		erts_smp_atomic32_set(&balance_info.used_runqs, online);
+		set_no_used_runqs(online);
 		/* Make sure that we balance soon... */
 		balance_info.forced_check_balance = 1;
 		erts_smp_runq_lock(ERTS_RUNQ_IX(0));
@@ -5131,7 +5196,7 @@ Process *schedule(Process *p, int calls)
 	esdp = erts_get_scheduler_data();
 	rq = erts_get_runq_current(esdp);
 	ASSERT(esdp);
-	fcalls = (int) erts_smp_atomic32_read(&function_calls);
+	fcalls = (int) erts_smp_atomic32_read_acqb(&function_calls);
 	actual_reds = reds = 0;
 	erts_smp_runq_lock(rq);
     } else {
@@ -5282,14 +5347,14 @@ Process *schedule(Process *p, int calls)
 			 | ERTS_RUNQ_FLG_CHK_CPU_BIND
 			 | ERTS_RUNQ_FLG_SUSPENDED)) {
 	    if ((rq->flags & ERTS_RUNQ_FLG_SUSPENDED)
-		|| (erts_smp_atomic32_read(&esdp->ssi->flags)
+		|| (erts_smp_atomic32_read_acqb(&esdp->ssi->flags)
 		    & ERTS_SSI_FLG_SUSPENDED)) {
 		ASSERT(erts_smp_atomic32_read(&esdp->ssi->flags)
 		       & ERTS_SSI_FLG_SUSPENDED);
 		suspend_scheduler(esdp);
 	    }
 	    if ((rq->flags & ERTS_RUNQ_FLG_CHK_CPU_BIND)
-		|| erts_smp_atomic32_read(&esdp->chk_cpu_bind)) {
+		|| erts_smp_atomic32_read_acqb(&esdp->chk_cpu_bind)) {
 		erts_sched_check_cpu_bind(esdp);
 	    }
 	}
@@ -5340,7 +5405,7 @@ Process *schedule(Process *p, int calls)
 	    if (rq->flags & (ERTS_RUNQ_FLG_SHARED_RUNQ
 			     | ERTS_RUNQ_FLG_SUSPENDED)) {
 		if ((rq->flags & ERTS_RUNQ_FLG_SUSPENDED)
-		    || (erts_smp_atomic32_read(&esdp->ssi->flags)
+		    || (erts_smp_atomic32_read_acqb(&esdp->ssi->flags)
 			& ERTS_SSI_FLG_SUSPENDED)) {
 		    ASSERT(erts_smp_atomic32_read(&esdp->ssi->flags)
 			   & ERTS_SSI_FLG_SUSPENDED);
@@ -5384,7 +5449,7 @@ Process *schedule(Process *p, int calls)
 	     * Schedule system-level activities.
 	     */
 
-	    erts_smp_atomic32_set(&function_calls, 0);
+	    erts_smp_atomic32_set_relb(&function_calls, 0);
 	    fcalls = 0;
 
 	    ASSERT(!erts_port_task_have_outstanding_io_tasks());
@@ -5426,7 +5491,7 @@ Process *schedule(Process *p, int calls)
 			if (erts_common_run_queue->waiting)
 			    wake_scheduler(erts_common_run_queue, 0, 1);
 		    }
-		    else if (erts_smp_atomic32_read(&no_empty_run_queues) != 0) {
+		    else if (erts_smp_atomic32_read_acqb(&no_empty_run_queues) != 0) {
 			wake_scheduler_on_empty_runq(rq);
 			rq->wakeup_other = 0;
 		    }
@@ -7637,7 +7702,8 @@ continue_exit_process(Process *p
 static void
 timeout_proc(Process* p)
 {
-    p->i = *((BeamInstr **) (UWord) p->def_arg_reg);
+    BeamInstr** pi = (BeamInstr **) p->def_arg_reg;
+    p->i = *pi;
     p->flags |= F_TIMO;
     p->flags &= ~F_INSLPQUEUE;
 
diff --git a/erts/emulator/beam/erl_trace.c b/erts/emulator/beam/erl_trace.c
index c0397ca6c3..8833137112 100644
--- a/erts/emulator/beam/erl_trace.c
+++ b/erts/emulator/beam/erl_trace.c
@@ -1538,8 +1538,7 @@ erts_call_trace(Process* p, BeamInstr mfa[3], Binary *match_spec,
     Eterm tracee;
 #endif
     Eterm transformed_args[MAX_ARG];
-    DeclareTmpHeap(sub_bin_heap_et,ERL_SUB_BIN_SIZE,p);
-    ErlSubBin *sub_bin_heap = (ErlSubBin *) sub_bin_heap_et;
+    DeclareTypedTmpHeap(ErlSubBin,sub_bin_heap,p);
 
     ASSERT(tracer_pid);
     if (*tracer_pid == am_true) {
@@ -1600,21 +1599,20 @@ erts_call_trace(Process* p, BeamInstr mfa[3], Binary *match_spec,
 	if (is_boxed(arg) && header_is_bin_matchstate(*boxed_val(arg))) {
 	    ErlBinMatchState* ms = (ErlBinMatchState *) boxed_val(arg);
 	    ErlBinMatchBuffer* mb = &ms->mb;
-	    ErlSubBin* sb = sub_bin_heap;
 	    Uint bit_size;
 
 	    ASSERT(sub_bin_heap->thing_word == 0); /* At most one of match context */
 
 	    bit_size = mb->size - mb->offset;
-	    sb->thing_word = HEADER_SUB_BIN;
-	    sb->size = BYTE_OFFSET(bit_size);
-	    sb->bitsize = BIT_OFFSET(bit_size);
-	    sb->offs = BYTE_OFFSET(mb->offset);
-	    sb->bitoffs = BIT_OFFSET(mb->offset);
-	    sb->is_writable = 0;
-	    sb->orig = mb->orig;
-
-	    arg = make_binary(sb);
+	    sub_bin_heap->thing_word = HEADER_SUB_BIN;
+	    sub_bin_heap->size = BYTE_OFFSET(bit_size);
+	    sub_bin_heap->bitsize = BIT_OFFSET(bit_size);
+	    sub_bin_heap->offs = BYTE_OFFSET(mb->offset);
+	    sub_bin_heap->bitoffs = BIT_OFFSET(mb->offset);
+	    sub_bin_heap->is_writable = 0;
+	    sub_bin_heap->orig = mb->orig;
+
+	    arg = make_binary(sub_bin_heap);
 	}
 	transformed_args[i] = arg;
     }
diff --git a/erts/emulator/beam/global.h b/erts/emulator/beam/global.h
index 96da894d90..499bdd77ba 100644
--- a/erts/emulator/beam/global.h
+++ b/erts/emulator/beam/global.h
@@ -183,7 +183,7 @@ struct port {
 				    process to get (line oriented I/O)*/
     Uint32 status;		 /* Status and type flags */
     int control_flags;		 /* Flags for port_control()  */
-    Uint32 snapshot;             /* Next snapshot that port should be part of */
+    erts_aint32_t snapshot;      /* Next snapshot that port should be part of */
     struct reg_proc *reg;
     ErlDrvPDL port_data_lock;
 
@@ -527,11 +527,10 @@ union erl_off_heap_ptr {
 
 /* arrays that get malloced at startup */
 extern Port* erts_port;
-extern erts_smp_atomic_t erts_ports_alive;
 
 extern Uint erts_max_ports;
 extern Uint erts_port_tab_index_mask;
-extern erts_smp_atomic_t erts_ports_snapshot;
+extern erts_smp_atomic32_t erts_ports_snapshot;
 extern erts_smp_atomic_t erts_dead_ports_ptr;
 
 ERTS_GLB_INLINE void erts_may_save_closed_port(Port *prt);
@@ -541,12 +540,12 @@ ERTS_GLB_INLINE void erts_may_save_closed_port(Port *prt);
 ERTS_GLB_INLINE void erts_may_save_closed_port(Port *prt)
 {
     ERTS_SMP_LC_ASSERT(erts_smp_lc_spinlock_is_locked(&prt->state_lck));
-    if (prt->snapshot != erts_smp_atomic_read(&erts_ports_snapshot)) {
+    if (prt->snapshot != erts_smp_atomic32_read_acqb(&erts_ports_snapshot)) {
 	/* Dead ports are added from the end of the snapshot buffer */
 	Eterm* tombstone = (Eterm*) erts_smp_atomic_addtest(&erts_dead_ports_ptr,
 							    -(erts_aint_t)sizeof(Eterm));
 	ASSERT(tombstone+1 != NULL);
-	ASSERT(prt->snapshot == (Uint32) erts_smp_atomic_read(&erts_ports_snapshot) - 1);
+	ASSERT(prt->snapshot == erts_smp_atomic32_read(&erts_ports_snapshot) - 1);
 	*tombstone = prt->id;
     }
     /*else no ongoing snapshot or port was already included or created after snapshot */
@@ -1653,10 +1652,14 @@ struct Sint_buf {
 };	
 char* Sint_to_buf(Sint, struct Sint_buf*);
 
+#define ERTS_IOLIST_OK 0
+#define ERTS_IOLIST_OVERFLOW 1
+#define ERTS_IOLIST_TYPE 2
+
 Eterm buf_to_intlist(Eterm**, char*, int, Eterm); /* most callers pass plain char*'s */
 int io_list_to_buf(Eterm, char*, int);
 int io_list_to_buf2(Eterm, char*, int);
-int io_list_len(Eterm);
+int erts_iolist_size(Eterm, Uint *);
 int is_string(Eterm);
 void erl_at_exit(void (*) (void*), void*);
 Eterm collect_memory(Process *);
@@ -1890,6 +1893,8 @@ erts_alloc_message_heap(Uint size,
 #  if defined(DEBUG)
 #    define DeclareTmpHeap(VariableName,Size,Process) \
        Eterm *VariableName = erts_debug_allocate_tmp_heap(Size,Process)
+#    define DeclareTypedTmpHeap(Type,VariableName,Process)		\
+      Type *VariableName = (Type *) erts_debug_allocate_tmp_heap(sizeof(Type)/sizeof(Eterm),Process)
 #    define DeclareTmpHeapNoproc(VariableName,Size) \
        Eterm *VariableName = erts_debug_allocate_tmp_heap(Size,NULL)
 #    define UseTmpHeap(Size,Proc) \
@@ -1911,6 +1916,8 @@ erts_alloc_message_heap(Uint size,
 #  else
 #    define DeclareTmpHeap(VariableName,Size,Process) \
        Eterm *VariableName = (ERTS_PROC_GET_SCHDATA(Process)->tmp_heap)+(ERTS_PROC_GET_SCHDATA(Process)->num_tmp_heap_used)
+#    define DeclareTypedTmpHeap(Type,VariableName,Process)		\
+      Type *VariableName = (Type *) (ERTS_PROC_GET_SCHDATA(Process)->tmp_heap)+(ERTS_PROC_GET_SCHDATA(Process)->num_tmp_heap_used)
 #    define DeclareTmpHeapNoproc(VariableName,Size) \
        Eterm *VariableName = (erts_get_scheduler_data()->tmp_heap)+(erts_get_scheduler_data()->num_tmp_heap_used)
 #    define UseTmpHeap(Size,Proc) \
@@ -1936,6 +1943,8 @@ erts_alloc_message_heap(Uint size,
 #else
 #  define DeclareTmpHeap(VariableName,Size,Process) \
      Eterm VariableName[Size]
+#  define DeclareTypedTmpHeap(Type,VariableName,Process)	\
+     Type VariableName[1]
 #  define DeclareTmpHeapNoproc(VariableName,Size) \
      Eterm VariableName[Size]
 #  define UseTmpHeap(Size,Proc) /* Nothing */
diff --git a/erts/emulator/beam/io.c b/erts/emulator/beam/io.c
index f619c6f88b..df5f8b22a3 100644
--- a/erts/emulator/beam/io.c
+++ b/erts/emulator/beam/io.c
@@ -56,7 +56,6 @@ static erts_smp_tsd_key_t driver_list_last_error_key;  /* Save last DDLL error o
 							  per thread basis (for BC interfaces) */
 
 Port*      erts_port; /* The port table */
-erts_smp_atomic_t erts_ports_alive;
 erts_smp_atomic_t erts_bytes_out;	/* No bytes sent out of the system */
 erts_smp_atomic_t erts_bytes_in;	/* No bytes gotten into the system */
 
@@ -82,6 +81,9 @@ static void driver_monitor_unlock_pdl(Port *p);
 #define DRV_MONITOR_UNLOCK_PDL(Port) /* nothing */
 #endif
 
+#define ERL_SMALL_IO_BIN_LIMIT (4*ERL_ONHEAP_BIN_LIMIT)
+#define SMALL_WRITE_VEC  16
+
 static ERTS_INLINE ErlIOQueue*
 drvport2ioq(ErlDrvPort drvport)
 {
@@ -190,7 +192,7 @@ typedef struct line_buf_context {
 static erts_smp_spinlock_t get_free_port_lck;
 static Uint last_port_num;
 static Uint port_num_mask;
-erts_smp_atomic_t erts_ports_snapshot; /* Identifies the _next_ snapshot (not the ongoing) */
+erts_smp_atomic32_t erts_ports_snapshot; /* Identifies the _next_ snapshot (not the ongoing) */
 
 
 static ERTS_INLINE void
@@ -421,10 +423,9 @@ setup_port(Port* prt, Eterm pid, erts_driver_t *driver,
     new_name = (char*) erts_alloc(ERTS_ALC_T_PORT_NAME, sys_strlen(name)+1);
     sys_strcpy(new_name, name);
     erts_smp_runq_lock(runq);
-    erts_smp_atomic_inc(&erts_ports_alive);
     erts_smp_port_state_lock(prt);    
     prt->status = ERTS_PORT_SFLG_CONNECTED | xstatus;
-    prt->snapshot = (Uint32) erts_smp_atomic_read(&erts_ports_snapshot);    
+    prt->snapshot = erts_smp_atomic32_read(&erts_ports_snapshot);    
     old_name = prt->name;
     prt->name = new_name;
 #ifdef ERTS_SMP
@@ -954,13 +955,14 @@ do {									\
     int _bitoffs;							\
     int _bitsize;							\
     ERTS_GET_REAL_BIN(obj, _real, _offset, _bitoffs, _bitsize);		\
-    ASSERT(_bitsize == 0);						\
+    if (_bitsize != 0) goto L_type_error;				\
     if (thing_subtag(*binary_val(_real)) == REFC_BINARY_SUBTAG &&	\
 	_bitoffs == 0) {						\
 	b_size += _size;						\
+        if (b_size < _size) goto L_overflow_error;			\
 	in_clist = 0;							\
 	v_size++;							\
-        if (_size >= bin_limit) {					\
+        if (_size >= ERL_SMALL_IO_BIN_LIMIT) {				\
             p_in_clist = 0;						\
             p_v_size++;							\
         } else {							\
@@ -972,6 +974,7 @@ do {									\
         }								\
     } else {								\
 	c_size += _size;						\
+        if (c_size < _size) goto L_overflow_error;			\
 	if (!in_clist) {						\
 	    in_clist = 1;						\
 	    v_size++;							\
@@ -986,29 +989,30 @@ do {									\
 
 
 /* 
-** Size of a io list in bytes
-** return -1 if error
-** returns:            - Total size of io list
-**           vsize     - SysIOVec size needed for a writev
-**           csize     - Number of bytes not in binary (in the common binary)
-**           pvsize    - SysIOVec size needed if packing small binaries
-**           pcsize    - Number of bytes in the common binary if packing
-*/
+ * Returns 0 if successful and a non-zero value otherwise.
+ *
+ * Return values through pointers:
+ *    *vsize      - SysIOVec size needed for a writev
+ *    *csize      - Number of bytes not in binary (in the common binary)
+ *    *pvsize     - SysIOVec size needed if packing small binaries
+ *    *pcsize     - Number of bytes in the common binary if packing
+ *    *total_size - Total size of iolist in bytes
+ */
 
 static int 
-io_list_vec_len(Eterm obj, int* vsize, int* csize, 
-		int bin_limit, /* small binaries limit */
-		int * pvsize, int * pcsize)
+io_list_vec_len(Eterm obj, Uint* vsize, Uint* csize,
+		Uint* pvsize, Uint* pcsize, Uint* total_size)
 {
     DECLARE_ESTACK(s);
     Eterm* objp;
-    int v_size = 0;
-    int c_size = 0;
-    int b_size = 0;
-    int in_clist = 0;
-    int p_v_size = 0;
-    int p_c_size = 0;
-    int p_in_clist = 0;
+    Uint v_size = 0;
+    Uint c_size = 0;
+    Uint b_size = 0;
+    Uint in_clist = 0;
+    Uint p_v_size = 0;
+    Uint p_c_size = 0;
+    Uint p_in_clist = 0;
+    Uint total;
 
     goto L_jump_start;  /* avoid a push */
 
@@ -1022,6 +1026,9 @@ io_list_vec_len(Eterm obj, int* vsize, int* csize,
 
 	    if (is_byte(obj)) {
 		c_size++;
+		if (c_size == 0) {
+		    goto L_overflow_error;
+		}
 		if (!in_clist) {
 		    in_clist = 1;
 		    v_size++;
@@ -1061,32 +1068,31 @@ io_list_vec_len(Eterm obj, int* vsize, int* csize,
 	}
     }
 
+    total = c_size + b_size;
+    if (total < c_size) {
+	goto L_overflow_error;
+    }
+    *total_size = total;
+
     DESTROY_ESTACK(s);
-    if (vsize != NULL)
-	*vsize = v_size;
-    if (csize != NULL)
-	*csize = c_size;
-    if (pvsize != NULL)
-	*pvsize = p_v_size;
-    if (pcsize != NULL)
-	*pcsize = p_c_size;
-    return c_size + b_size;
+    *vsize = v_size;
+    *csize = c_size;
+    *pvsize = p_v_size;
+    *pcsize = p_c_size;
+    return 0;
 
  L_type_error:
+ L_overflow_error:
     DESTROY_ESTACK(s);
-    return -1;
+    return 1;
 }
 
-#define ERL_SMALL_IO_BIN_LIMIT (4*ERL_ONHEAP_BIN_LIMIT)
-#define SMALL_WRITE_VEC  16
-
-
 /* write data to a port */
 int erts_write_to_port(Eterm caller_id, Port *p, Eterm list)
 {
     char *buf;
     erts_driver_t *drv = p->drv_ptr;
-    int size;
+    Uint size;
     int fpe_was_unmasked;
     
     ERTS_SMP_LC_ASSERT(erts_lc_is_port_locked(p));
@@ -1094,10 +1100,10 @@ int erts_write_to_port(Eterm caller_id, Port *p, Eterm list)
 
     p->caller = caller_id;
     if (drv->outputv != NULL) {
-	int vsize;
-	int csize;
-	int pvsize;
-	int pcsize;
+	Uint vsize;
+	Uint csize;
+	Uint pvsize;
+	Uint pcsize;
 	int blimit;
 	SysIOVec iv[SMALL_WRITE_VEC];
 	ErlDrvBinary* bv[SMALL_WRITE_VEC];
@@ -1106,9 +1112,8 @@ int erts_write_to_port(Eterm caller_id, Port *p, Eterm list)
 	ErlDrvBinary* cbin;
 	ErlIOVec ev;
 
-	if ((size = io_list_vec_len(list, &vsize, &csize, 
-				    ERL_SMALL_IO_BIN_LIMIT,
-				    &pvsize, &pcsize)) < 0) {
+	if (io_list_vec_len(list, &vsize, &csize,
+			    &pvsize, &pcsize, &size)) {
 	    goto bad_value;
 	}
 	/* To pack or not to pack (small binaries) ...? */
@@ -1183,7 +1188,7 @@ int erts_write_to_port(Eterm caller_id, Port *p, Eterm list)
 	else {
 	    ASSERT(r == -1); /* Overflow */
 	    erts_free(ERTS_ALC_T_TMP, buf);
-	    if ((size = io_list_len(list)) < 0) {
+	    if (erts_iolist_size(list, &size)) {
 		goto bad_value;
 	    }
 
@@ -1274,7 +1279,6 @@ void init_io(void)
 
     erts_smp_atomic_init(&erts_bytes_out, 0);
     erts_smp_atomic_init(&erts_bytes_in, 0);
-    erts_smp_atomic_init(&erts_ports_alive, 0);
 
     for (i = 0; i < erts_max_ports; i++) {
 	erts_port_task_init_sched(&erts_port[i].sched);
@@ -1296,7 +1300,7 @@ void init_io(void)
 	erts_port[i].port_data_lock = NULL;
     }
 
-    erts_smp_atomic_init(&erts_ports_snapshot, (erts_aint_t) 0);
+    erts_smp_atomic32_init(&erts_ports_snapshot, (erts_aint32_t) 0);
     last_port_num = 0;
     erts_smp_spinlock_init(&get_free_port_lck, "get_free_port");
 
@@ -2147,7 +2151,7 @@ erts_port_control(Process* p, Port* prt, Uint command, Eterm iolist)
     byte* to_port = NULL;	/* Buffer to write to port. */
 				/* Initialization is for shutting up
 				   warning about use before set. */
-    int to_len = 0;		/* Length of buffer. */
+    Uint to_len = 0;		/* Length of buffer. */
     int must_free = 0;		/* True if the buffer should be freed. */
     char port_result[ERL_ONHEAP_BIN_LIMIT];  /* Default buffer for result from port. */
     char* port_resp;		/* Pointer to result buffer. */
@@ -2192,7 +2196,7 @@ erts_port_control(Process* p, Port* prt, Uint command, Eterm iolist)
 	} else {
 	    ASSERT(r == -1);	/* Overflow */
 	    erts_free(ERTS_ALC_T_TMP, (void *) to_port);
-	    if ((to_len = io_list_len(iolist)) < 0) { /* Type error */
+	    if (erts_iolist_size(iolist, &to_len)) { /* Type error */
 		return THE_NON_VALUE;
 	    }
 	    must_free = 1;
diff --git a/erts/emulator/beam/safe_hash.c b/erts/emulator/beam/safe_hash.c
index 21d6ce9304..3e9243c77d 100644
--- a/erts/emulator/beam/safe_hash.c
+++ b/erts/emulator/beam/safe_hash.c
@@ -99,7 +99,7 @@ static void rehash(SafeHash* h, int grow_limit)
 	erts_free(h->type, (void *) old_tab);
     }
     /*else already done */
-    erts_smp_atomic_set(&h->is_rehashing, 0);
+    erts_smp_atomic_set_relb(&h->is_rehashing, 0);
 }
 
 
diff --git a/erts/emulator/beam/utils.c b/erts/emulator/beam/utils.c
index 6b4f3b3b36..a17de717bc 100644
--- a/erts/emulator/beam/utils.c
+++ b/erts/emulator/beam/utils.c
@@ -3021,13 +3021,25 @@ int io_list_to_buf(Eterm obj, char* buf, int len)
     return -1;
 }
 
-int io_list_len(Eterm obj)
+/*
+ * Return 0 if successful, and non-zero if unsuccessful.
+ */
+int erts_iolist_size(Eterm obj, Uint* sizep)
 {
     Eterm* objp;
-    Sint len = 0;
+    Uint size = 0;
     DECLARE_ESTACK(s);
     goto L_again;
 
+#define SAFE_ADD(Var, Val)			\
+    do {					\
+        Uint valvar = (Val);			\
+	Var += valvar;				\
+	if (Var < valvar) {			\
+	    goto L_overflow_error;		\
+	}					\
+    } while (0)
+
     while (!ESTACK_ISEMPTY(s)) {
 	obj = ESTACK_POP(s);
     L_again:
@@ -3037,9 +3049,12 @@ int io_list_len(Eterm obj)
 	    /* Head */
 	    obj = CAR(objp);
 	    if (is_byte(obj)) {
-		len++;
+		size++;
+		if (size == 0) {
+		    goto L_overflow_error;
+		}
 	    } else if (is_binary(obj) && binary_bitsize(obj) == 0) {
-		len += binary_size(obj);
+		SAFE_ADD(size, binary_size(obj));
 	    } else if (is_list(obj)) {
 		ESTACK_PUSH(s, CDR(objp));
 		goto L_iter_list; /* on head */
@@ -3051,23 +3066,29 @@ int io_list_len(Eterm obj)
 	    if (is_list(obj))
 		goto L_iter_list; /* on tail */
 	    else if (is_binary(obj) && binary_bitsize(obj) == 0) {
-		len += binary_size(obj);
+		SAFE_ADD(size, binary_size(obj));
 	    } else if (is_not_nil(obj)) {
 		goto L_type_error;
 	    }
 	} else if (is_binary(obj) && binary_bitsize(obj) == 0) { /* Tail was binary */
-	    len += binary_size(obj);
+	    SAFE_ADD(size, binary_size(obj));
 	} else if (is_not_nil(obj)) {
 	    goto L_type_error;
 	}
     }
+#undef SAFE_ADD
 
     DESTROY_ESTACK(s);
-    return len;
+    *sizep = size;
+    return ERTS_IOLIST_OK;
+
+ L_overflow_error:
+    DESTROY_ESTACK(s);
+    return ERTS_IOLIST_OVERFLOW;
 
  L_type_error:
     DESTROY_ESTACK(s);
-    return -1;
+    return ERTS_IOLIST_TYPE;
 }
 
 /* return 0 if item is not a non-empty flat list of bytes */
diff --git a/erts/emulator/drivers/common/inet_drv.c b/erts/emulator/drivers/common/inet_drv.c
index aec456151c..40c4a0df08 100644
--- a/erts/emulator/drivers/common/inet_drv.c
+++ b/erts/emulator/drivers/common/inet_drv.c
@@ -1948,7 +1948,7 @@ static int http_response_inetdrv(void *arg, int major, int minor,
     tcp_descriptor* desc = (tcp_descriptor*) arg;
     int i = 0;
     ErlDrvTermData spec[27];
-    ErlDrvTermData caller;
+    ErlDrvTermData caller = ERL_DRV_NIL;
     
     if (desc->inet.active == INET_PASSIVE) {
         /* {inet_async,S,Ref,{ok,{http_response,Version,Status,Phrase}}} */
@@ -2041,7 +2041,7 @@ http_request_inetdrv(void* arg, const http_atom_t* meth, const char* meth_ptr,
     tcp_descriptor* desc = (tcp_descriptor*) arg;
     int i = 0;
     ErlDrvTermData spec[43];
-    ErlDrvTermData caller;
+    ErlDrvTermData caller = ERL_DRV_NIL;
     
     if (desc->inet.active == INET_PASSIVE) {
         /* {inet_async, S, Ref, {ok,{http_request,Meth,Uri,Version}}} */
@@ -2092,7 +2092,7 @@ http_header_inetdrv(void* arg, const http_atom_t* name, const char* name_ptr,
     tcp_descriptor* desc = (tcp_descriptor*) arg;
     int i = 0;
     ErlDrvTermData spec[26];
-    ErlDrvTermData caller;
+    ErlDrvTermData caller = ERL_DRV_NIL;
     
     if (desc->inet.active == INET_PASSIVE) {
         /* {inet_async,S,Ref,{ok,{http_header,Bit,Name,IValue,Value}} */
@@ -2221,7 +2221,7 @@ int ssl_tls_inetdrv(void* arg, unsigned type, unsigned major, unsigned minor,
     tcp_descriptor* desc = (tcp_descriptor*) arg;
     int i = 0;
     ErlDrvTermData spec[28];
-    ErlDrvTermData caller;
+    ErlDrvTermData caller = ERL_DRV_NIL;
     ErlDrvBinary* bin;
     int ret;
 
@@ -7980,11 +7980,11 @@ static int tcp_inet_ctl(ErlDrvData e, unsigned int cmd, char* buf, int len,
 	timeout = get_int32(buf);
 
 	if (desc->inet.state == TCP_STATE_ACCEPTING) {
-	    unsigned long time_left;
-	    int oid;
-	    ErlDrvTermData ocaller;
-	    int oreq;
-	    unsigned otimeout;
+	    unsigned long time_left = 0;
+	    int oid = 0;
+	    ErlDrvTermData ocaller = ERL_DRV_NIL;
+	    int oreq = 0;
+	    unsigned otimeout = 0;
 	    ErlDrvTermData caller = driver_caller(desc->inet.port);
 	    MultiTimerData *mtd = NULL,*omtd = NULL;
 	    ErlDrvMonitor monitor, omonitor;
diff --git a/erts/emulator/hipe/hipe_bif2.c b/erts/emulator/hipe/hipe_bif2.c
index 19cd32c68f..2660f74a82 100644
--- a/erts/emulator/hipe/hipe_bif2.c
+++ b/erts/emulator/hipe/hipe_bif2.c
@@ -1,9 +1,8 @@
 /*
  * %CopyrightBegin%
-
- *
+ * 
  * Copyright Ericsson AB 2001-2011. All Rights Reserved.
- *
+ * 
  * The contents of this file are subject to the Erlang Public License,
  * Version 1.1, (the "License"); you may not use this file except in
  * compliance with the License. You should have received a copy of the
@@ -37,14 +36,25 @@
 #include "hipe_arch.h"
 #include "hipe_stack.h"
 
-BIF_RETTYPE hipe_bifs_show_estack_1(BIF_ALIST_1)
+static void proc_unlock(Process* c_p, Process* rp)
 {
+    ErtsProcLocks locks = ERTS_PROC_LOCKS_ALL;
+    if (rp == c_p) {
+	locks &= ~ERTS_PROC_LOCK_MAIN;
+    }
+    if (rp && locks) {
+	erts_smp_proc_unlock(rp, locks);
+    }
+}
+
+BIF_RETTYPE hipe_bifs_show_estack_1(BIF_ALIST_1)
+{    
     Process *rp = erts_pid2proc(BIF_P, ERTS_PROC_LOCK_MAIN,
 				BIF_ARG_1, ERTS_PROC_LOCKS_ALL);
     if (!rp)
 	BIF_ERROR(BIF_P, BADARG);
     hipe_print_estack(rp);
-    erts_smp_proc_unlock(rp, ERTS_PROC_LOCKS_ALL);
+    proc_unlock(BIF_P, rp);
     BIF_RET(am_true);
 }
 
@@ -55,7 +65,7 @@ BIF_RETTYPE hipe_bifs_show_heap_1(BIF_ALIST_1)
     if (!rp)
 	BIF_ERROR(BIF_P, BADARG);
     hipe_print_heap(rp);
-    erts_smp_proc_unlock(rp, ERTS_PROC_LOCKS_ALL);
+    proc_unlock(BIF_P, rp);
     BIF_RET(am_true);
 }
 
@@ -66,7 +76,7 @@ BIF_RETTYPE hipe_bifs_show_nstack_1(BIF_ALIST_1)
     if (!rp)
 	BIF_ERROR(BIF_P, BADARG);
     hipe_print_nstack(rp);
-    erts_smp_proc_unlock(rp, ERTS_PROC_LOCKS_ALL);
+    proc_unlock(BIF_P, rp);
     BIF_RET(am_true);
 }
 
@@ -82,7 +92,7 @@ BIF_RETTYPE hipe_bifs_show_pcb_1(BIF_ALIST_1)
     if (!rp)
 	BIF_ERROR(BIF_P, BADARG);
     hipe_print_pcb(rp);
-    erts_smp_proc_unlock(rp, ERTS_PROC_LOCKS_ALL);
+    proc_unlock(BIF_P, rp);
     BIF_RET(am_true);
 }
 
diff --git a/erts/emulator/hipe/hipe_mkliterals.c b/erts/emulator/hipe/hipe_mkliterals.c
index 650861b54b..bced90785d 100644
--- a/erts/emulator/hipe/hipe_mkliterals.c
+++ b/erts/emulator/hipe/hipe_mkliterals.c
@@ -587,9 +587,9 @@ static void print_params(FILE *fp, void (*print_param)(FILE*,const struct rts_pa
 	(*print_param)(fp, &rts_params[i]);
 }
 
-static int do_c(FILE *fp)
+static int do_c(FILE *fp, const char* this_exe)
 {
-    fprintf(fp, "/* File: hipe_literals.h, generated by hipe_mkliterals */\n");
+    fprintf(fp, "/* File: hipe_literals.h, generated by %s */\n", this_exe);
     fprintf(fp, "#ifndef __HIPE_LITERALS_H__\n");
     fprintf(fp, "#define __HIPE_LITERALS_H__\n\n");
     print_literals(fp, c_define_literal);
@@ -603,9 +603,9 @@ static int do_c(FILE *fp)
     return 0;
 }
 
-static int do_e(FILE *fp)
+static int do_e(FILE *fp, const char* this_exe)
 {
-    fprintf(fp, "%%%% File: hipe_literals.hrl, generated by hipe_mkliterals");
+    fprintf(fp, "%%%% File: hipe_literals.hrl, generated by %s", this_exe);
     fprintf(fp, "\n\n");
     print_literals(fp, e_define_literal);
     fprintf(fp, "\n");
@@ -622,9 +622,9 @@ int main(int argc, const char **argv)
     compute_crc();
     if (argc == 2) {
 	if (strcmp(argv[1], "-c") == 0)
-	    return do_c(stdout);
+	    return do_c(stdout, argv[0]);
 	if (strcmp(argv[1], "-e") == 0)
-	    return do_e(stdout);
+	    return do_e(stdout, argv[0]);
     }
     fprintf(stderr, "usage: %s [-c | -e] > output-file\n", argv[0]);
     return 1;
diff --git a/erts/emulator/sys/common/erl_check_io.c b/erts/emulator/sys/common/erl_check_io.c
index 218bd79584..71b374527e 100644
--- a/erts/emulator/sys/common/erl_check_io.c
+++ b/erts/emulator/sys/common/erl_check_io.c
@@ -1137,6 +1137,11 @@ ERTS_CIO_EXPORT(erts_check_io)(int do_wait)
 
  restart:
 
+#ifdef ERTS_BREAK_REQUESTED
+    if (ERTS_BREAK_REQUESTED)
+	erts_do_break_handling();
+#endif
+
     /* Figure out timeout value */
     if (do_wait) {
 	erts_time_remaining(&wait_time);
diff --git a/erts/emulator/sys/common/erl_mseg.c b/erts/emulator/sys/common/erl_mseg.c
index ffa3a6328c..eaef6680dd 100644
--- a/erts/emulator/sys/common/erl_mseg.c
+++ b/erts/emulator/sys/common/erl_mseg.c
@@ -134,7 +134,16 @@ static int mmap_fd;
 #define CAN_PARTLY_DESTROY 0
 #endif
 
-static const ErtsMsegOpt_t default_opt = ERTS_MSEG_DEFAULT_OPT_INITIALIZER;
+const ErtsMsegOpt_t erts_mseg_default_opt = {
+    1,			/* Use cache		     */
+    1,			/* Preserv data		     */
+    0,			/* Absolute shrink threshold */
+    0			/* Relative shrink threshold */
+#if HALFWORD_HEAP
+    ,0                  /* need low memory */
+#endif
+};
+
 
 typedef struct cache_desc_t_ {
     void *seg;
@@ -605,18 +614,10 @@ mseg_clear_cache(MemKind* mk)
     INC_CC(clear_cache);
 }
 
-static ERTS_INLINE MemKind* type2mk(ErtsAlcType_t atype)
+static ERTS_INLINE MemKind* memkind(const ErtsMsegOpt_t *opt)
 {
 #if HALFWORD_HEAP
-    switch (atype) {
-    case ERTS_ALC_A_ETS:
-    case ERTS_ALC_A_BINARY:
-    case ERTS_ALC_A_FIXED_SIZE:
-    case ERTS_ALC_A_DRIVER:
-	return &hi_mem;
-    default:
-	return &low_mem;
-    }
+    return opt->low_mem ? &low_mem : &hi_mem;
 #else
     return &the_mem;
 #endif
@@ -628,7 +629,7 @@ mseg_alloc(ErtsAlcType_t atype, Uint *size_p, const ErtsMsegOpt_t *opt)
     Uint max, min, diff_size, size;
     cache_desc_t *cd, *cand_cd;
     void *seg;
-    MemKind* mk = type2mk(atype);
+    MemKind* mk = memkind(opt);
 
     INC_CC(alloc);
 
@@ -742,7 +743,7 @@ static void
 mseg_dealloc(ErtsAlcType_t atype, void *seg, Uint size,
 	     const ErtsMsegOpt_t *opt)
 {
-    MemKind* mk = type2mk(atype);
+    MemKind* mk = memkind(opt);
     cache_desc_t *cd;
 
     ERTS_MSEG_DEALLOC_STAT(mk,size);
@@ -800,7 +801,7 @@ static void *
 mseg_realloc(ErtsAlcType_t atype, void *seg, Uint old_size, Uint *new_size_p,
 	     const ErtsMsegOpt_t *opt)
 {
-    MemKind* mk = type2mk(atype);
+    MemKind* mk = memkind(opt);
     void *new_seg;
     Uint new_size;
 
@@ -1372,7 +1373,7 @@ erts_mseg_alloc_opt(ErtsAlcType_t atype, Uint *size_p, const ErtsMsegOpt_t *opt)
 void *
 erts_mseg_alloc(ErtsAlcType_t atype, Uint *size_p)
 {
-    return erts_mseg_alloc_opt(atype, size_p, &default_opt);
+    return erts_mseg_alloc_opt(atype, size_p, &erts_mseg_default_opt);
 }
 
 void
@@ -1387,7 +1388,7 @@ erts_mseg_dealloc_opt(ErtsAlcType_t atype, void *seg, Uint size,
 void
 erts_mseg_dealloc(ErtsAlcType_t atype, void *seg, Uint size)
 {
-    erts_mseg_dealloc_opt(atype, seg, size, &default_opt);
+    erts_mseg_dealloc_opt(atype, seg, size, &erts_mseg_default_opt);
 }
 
 void *
@@ -1405,7 +1406,7 @@ void *
 erts_mseg_realloc(ErtsAlcType_t atype, void *seg, Uint old_size,
 		  Uint *new_size_p)
 {
-    return erts_mseg_realloc_opt(atype, seg, old_size, new_size_p, &default_opt);
+    return erts_mseg_realloc_opt(atype, seg, old_size, new_size_p, &erts_mseg_default_opt);
 }
 
 void
diff --git a/erts/emulator/sys/common/erl_mseg.h b/erts/emulator/sys/common/erl_mseg.h
index d8053eb0d9..fbb66ee33b 100644
--- a/erts/emulator/sys/common/erl_mseg.h
+++ b/erts/emulator/sys/common/erl_mseg.h
@@ -60,15 +60,12 @@ typedef struct {
     int  preserv;
     UWord abs_shrink_th;
     UWord rel_shrink_th;
+#if HALFWORD_HEAP
+    int low_mem;
+#endif
 } ErtsMsegOpt_t;
 
-#define ERTS_MSEG_DEFAULT_OPT_INITIALIZER				\
-{									\
-    1,			/* Use cache				*/	\
-    1,			/* Preserv data				*/	\
-    0,			/* Absolute shrink threshold		*/	\
-    0			/* Relative shrink threshold		*/	\
-}
+extern const ErtsMsegOpt_t erts_mseg_default_opt;
 
 void *erts_mseg_alloc(ErtsAlcType_t, Uint *);
 void *erts_mseg_alloc_opt(ErtsAlcType_t, Uint *, const ErtsMsegOpt_t *);
diff --git a/erts/emulator/sys/common/erl_poll.c b/erts/emulator/sys/common/erl_poll.c
index 3ae5b8d747..f5c785d683 100644
--- a/erts/emulator/sys/common/erl_poll.c
+++ b/erts/emulator/sys/common/erl_poll.c
@@ -766,7 +766,7 @@ write_batch_buf(ErtsPollSet ps, ErtsPollBatchBuf *bbp)
 		short filter;
 		int fd = (int) ebuf[i].ident;
 
-		switch ((int) ebuf[i].udata) {
+		switch ((int) (long) ebuf[i].udata) {
 
 		    /*
 		     * Since we use a lazy update approach EV_DELETE will
@@ -805,7 +805,7 @@ write_batch_buf(ErtsPollSet ps, ErtsPollBatchBuf *bbp)
 			if (fd == (int) ebuf[j].ident) {
 			    ebuf[j].udata = (void *) ERTS_POLL_KQ_OP_HANDLED;
 			    if (!(ebuf[j].flags & EV_ERROR)) {
-				switch ((int) ebuf[j].udata) {
+				switch ((int) (long) ebuf[j].udata) {
 				case ERTS_POLL_KQ_OP_ADD2_W:
 				    filter = EVFILT_WRITE;
 				    goto rm_add_fb;
@@ -823,7 +823,8 @@ write_batch_buf(ErtsPollSet ps, ErtsPollBatchBuf *bbp)
 			}
 		    }
 		    /* The other add succeded... */
-		    filter = (((int) ebuf[i].udata == ERTS_POLL_KQ_OP_ADD2_W)
+		    filter = ((((int) (long) ebuf[i].udata)
+			       == ERTS_POLL_KQ_OP_ADD2_W)
 			      ? EVFILT_READ
 			      : EVFILT_WRITE);
 		rm_add_fb:
diff --git a/erts/emulator/sys/win32/erl_poll.c b/erts/emulator/sys/win32/erl_poll.c
index 7662f190ef..074e2e247f 100644
--- a/erts/emulator/sys/win32/erl_poll.c
+++ b/erts/emulator/sys/win32/erl_poll.c
@@ -693,6 +693,7 @@ static void *break_waiter(void *param)
 	    ResetEvent(harr[0]);
 	    erts_mtx_lock(&break_waiter_lock);
 	    erts_atomic32_set(&break_waiter_state,BREAK_WAITER_GOT_BREAK);
+	    ERTS_THR_MEMORY_BARRIER;
 	    SetEvent(break_happened_event);
 	    erts_mtx_unlock(&break_waiter_lock);
 	    break;
@@ -700,6 +701,7 @@ static void *break_waiter(void *param)
 	    ResetEvent(harr[1]);
 	    erts_mtx_lock(&break_waiter_lock);
 	    erts_atomic32_set(&break_waiter_state,BREAK_WAITER_GOT_HALT);
+	    ERTS_THR_MEMORY_BARRIER;
 	    SetEvent(break_happened_event);
 	    erts_mtx_unlock(&break_waiter_lock);
 	    break;
diff --git a/erts/emulator/test/binary_SUITE.erl b/erts/emulator/test/binary_SUITE.erl
index 7e409f053e..4e82381fba 100644
--- a/erts/emulator/test/binary_SUITE.erl
+++ b/erts/emulator/test/binary_SUITE.erl
@@ -23,12 +23,12 @@
 %% Tests binaries and the BIFs:
 %%	list_to_binary/1
 %%      iolist_to_binary/1
-%%      bitstr_to_list/1
+%%      list_to_bitstring/1
 %%	binary_to_list/1
 %%	binary_to_list/3
 %%	binary_to_term/1
 %%  	binary_to_term/2
-%%      bitstr_to_list/1
+%%      bitstring_to_list/1
 %%	term_to_binary/1
 %%      erlang:external_size/1
 %%	size(Binary)
@@ -275,12 +275,33 @@ bad_list_to_binary(Config) when is_list(Config) ->
     ?line test_bad_bin(fun(X, Y) -> X*Y end),
     ?line test_bad_bin([1,fun(X) -> X + 1 end,2|fun() -> 0 end]),
     ?line test_bad_bin([fun(X) -> X + 1 end]),
+
+    %% Test iolists that do not fit in the address space.
+    %% Unfortunately, it would be too slow to test in a 64-bit emulator.
+    case erlang:system_info(wordsize) of
+	4 -> huge_iolists();
+	_ -> ok
+    end.
+
+huge_iolists() ->
+    FourGigs = 1 bsl 32,
+    ?line Sizes = [FourGigs+N || N <- lists:seq(0, 64)] ++
+	[1 bsl N || N <- lists:seq(33, 37)],
+    ?line Base = <<0:(1 bsl 20)/unit:8>>,
+    [begin
+	 L = build_iolist(Sz, Base),
+	 ?line {'EXIT',{system_limit,_}} = (catch list_to_binary([L])),
+	 ?line {'EXIT',{system_limit,_}} = (catch list_to_bitstring([L])),
+	 ?line {'EXIT',{system_limit,_}} = (catch binary:list_to_bin([L])),
+	 ?line {'EXIT',{system_limit,_}} = (catch iolist_to_binary(L))
+	 end || Sz <- Sizes],
     ok.
 
 test_bad_bin(List) ->
     {'EXIT',{badarg,_}} = (catch list_to_binary(List)),
     {'EXIT',{badarg,_}} = (catch iolist_to_binary(List)),
-    {'EXIT',{badarg,_}} = (catch list_to_bitstring(List)).
+    {'EXIT',{badarg,_}} = (catch list_to_bitstring(List)),
+    {'EXIT',{badarg,_}} = (catch iolist_size(List)).
 
 bad_binary_to_list(doc) -> "Tries binary_to_list/1,3 with bad arguments.";
 bad_binary_to_list(Config) when is_list(Config) ->
@@ -516,18 +537,65 @@ external_size_1(Term, Size0, Limit) when Size0 < Limit ->
 external_size_1(_, _, _) -> ok.
 
 t_iolist_size(Config) when is_list(Config) ->
-    %% Build a term whose external size only fits in a big num (on 32-bit CPU).
-    Bin = iolist_to_binary(lists:seq(0, 254)),
-    ?line ok = t_iolist_size_1(Bin, 0, 16#7FFFFFFF),
-    ?line ok = t_iolist_size_1(make_unaligned_sub_binary(Bin), 0, 16#7FFFFFFF).
+    ?line Seed = now(),
+    ?line io:format("Seed: ~p", [Seed]),
+    ?line random:seed(Seed),
+    ?line Base = <<0:(1 bsl 20)/unit:8>>,
+    ?line Powers = [1 bsl N || N <- lists:seq(2, 37)],
+    ?line Sizes0 = [[N - random:uniform(N div 2),
+		     lists:seq(N-2, N+2),
+		     N+N div 2,
+		     N + random:uniform(N div 2)] ||
+		       N <- Powers],
+    %% Test sizes around 1^32 more thoroughly.
+    FourGigs = 1 bsl 32,
+    ?line Sizes1 = [FourGigs+N || N <- lists:seq(-8, 40)] ++ Sizes0,
+    ?line Sizes2 = lists:flatten(Sizes1),
+    ?line Sizes = lists:usort(Sizes2),
+    io:format("~p sizes:", [length(Sizes)]),
+    io:format("~p\n", [Sizes]),
+    ?line [Sz = iolist_size(build_iolist(Sz, Base)) || Sz <- Sizes],
+    ok.
 
-t_iolist_size_1(IOList, Size0, Limit) when Size0 < Limit ->
-    case iolist_size(IOList) of
-	Size when is_integer(Size), Size0 < Size ->
-	    io:format("~p", [Size]),
-	    t_iolist_size_1([IOList|IOList], Size, Limit)
+build_iolist(N, Base) when N < 16 ->
+    case random:uniform(3) of
+	1 ->
+	    <<Bin:N/binary,_/binary>> = Base,
+	    Bin;
+	_ ->
+	    lists:seq(1, N)
+    end;
+build_iolist(N, Base) when N =< byte_size(Base) ->
+    case random:uniform(3) of
+	1 ->
+	    <<Bin:N/binary,_/binary>> = Base,
+	    Bin;
+	2 ->
+	    <<Bin:N/binary,_/binary>> = Base,
+	    [Bin];
+	3 ->
+	    case N rem 2 of
+		0 ->
+		    L = build_iolist(N div 2, Base),
+		    [L,L];
+		1 ->
+		    L = build_iolist(N div 2, Base),
+		    [L,L,45]
+	    end
     end;
-t_iolist_size_1(_, _, _) -> ok.
+build_iolist(N0, Base) ->
+    Small = random:uniform(15),
+    Seq = lists:seq(1, Small),
+    N = N0 - Small,
+    case N rem 2 of
+	0 ->
+	    L = build_iolist(N div 2, Base),
+	    [L,L|Seq];
+	1 ->
+	    L = build_iolist(N div 2, Base),
+	    [47,L,L|Seq]
+    end.
+
 
 bad_binary_to_term_2(doc) -> "OTP-4053.";
 bad_binary_to_term_2(suite) -> [];
@@ -1183,34 +1251,7 @@ deep(Config) when is_list(Config) ->
 
 deep_roundtrip(T) ->
     B = term_to_binary(T),
-    true = deep_eq(T, binary_to_term(B)).
-
-%%
-%% FIXME: =:= runs out of stack.
-%%
-deep_eq([H1|T1], [H2|T2]) ->
-    deep_eq(H1, H2) andalso deep_eq(T1, T2);
-deep_eq(T1, T2) when tuple_size(T1) =:= tuple_size(T2) ->
-    deep_eq_tup(T1, T2, tuple_size(T1));
-deep_eq(T1, T2) when is_function(T1), is_function(T2) ->
-    {uniq,U1} = erlang:fun_info(T1, uniq),
-    {index,I1} = erlang:fun_info(T1, index),
-    {arity,A1} = erlang:fun_info(T1, arity),
-    {env,E1} = erlang:fun_info(T1, env),
-    {uniq,U2} = erlang:fun_info(T2, uniq),
-    {index,I2} = erlang:fun_info(T2, index),
-    {arity,A2} = erlang:fun_info(T2, arity),
-    {env,E2} = erlang:fun_info(T2, env),
-    U1 =:= U2 andalso I1 =:= I2 andalso A1 =:= A2 andalso
-	deep_eq(E1, E2);
-deep_eq(T1, T2) ->
-    T1 =:= T2.
-
-deep_eq_tup(_T1, _T2, 0) ->
-    true;
-deep_eq_tup(T1, T2, N) ->
-    deep_eq(element(N, T1), element(N, T2)) andalso
-	deep_eq_tup(T1, T2, N-1).
+    T = binary_to_term(B).
 
 obsolete_funs(Config) when is_list(Config) ->
     erts_debug:set_internal_state(available_internal_state, true),
diff --git a/erts/emulator/test/code_SUITE.erl b/erts/emulator/test/code_SUITE.erl
index c1a048be75..a062cea117 100644
--- a/erts/emulator/test/code_SUITE.erl
+++ b/erts/emulator/test/code_SUITE.erl
@@ -498,7 +498,9 @@ do_false_dependency(Init, Code) ->
     ?line unlink(Pid), exit(Pid, kill),
     ?line true = erlang:purge_module(cpbugx),
     ?line true = erlang:delete_module(cpbugx),
+    ?line code:is_module_native(cpbugx),  % test is_module_native on deleted code
     ?line true = erlang:purge_module(cpbugx),
+    ?line code:is_module_native(cpbugx),  % test is_module_native on purged code
     ok.
     
 false_dependency_loop(Parent, Init, SendInitAck) ->
diff --git a/erts/emulator/test/driver_SUITE.erl b/erts/emulator/test/driver_SUITE.erl
index 7600a44988..520e3e8c76 100644
--- a/erts/emulator/test/driver_SUITE.erl
+++ b/erts/emulator/test/driver_SUITE.erl
@@ -38,7 +38,7 @@
 	 timer_change/1,
 	 timer_delay/1,
 	 queue_echo/1,
-	 fun_to_port/1,
+	 outputv_errors/1,
 	 driver_unloaded/1,
 	 io_ready_exit/1,
 	 use_fallback_pollset/1,
@@ -129,7 +129,7 @@ end_per_testcase(Case, Config) ->
 suite() -> [{ct_hooks,[ts_install_cth]}].
 
 all() -> 
-    [fun_to_port, outputv_echo, queue_echo, {group, timer},
+    [outputv_errors, outputv_echo, queue_echo, {group, timer},
      driver_unloaded, io_ready_exit, use_fallback_pollset,
      bad_fd_in_pollset, driver_event, fd_change,
      steal_control, otp_6602, 'driver_system_info_ver1.0',
@@ -165,37 +165,89 @@ init_per_group(_GroupName, Config) ->
 end_per_group(_GroupName, Config) ->
     Config.
 
-
-fun_to_port(doc) -> "Test sending a fun to port with an outputv-capable driver.";
-fun_to_port(Config) when is_list(Config) ->
+outputv_errors(doc) -> "Test sending bad types to port with an outputv-capable driver.";
+outputv_errors(Config) when is_list(Config) ->
     ?line Path = ?config(data_dir, Config),
     ?line erl_ddll:start(),
     ?line ok = load_driver(Path, outputv_drv),
 
-    ?line fun_to_port_1(fun() -> 33 end),
-    ?line fun_to_port_1([fun() -> 42 end]),
-    ?line fun_to_port_1([1|fun() -> 42 end]),
-    L = build_io_list(65536),
-    ?line fun_to_port_1([L,fun() -> 42 end]),
-    ?line fun_to_port_1([L|fun() -> 42 end]),
+    outputv_bad_types(fun(T) ->
+			      ?line outputv_errors_1(T),
+			      ?line outputv_errors_1([1|T]),
+			      ?line L = [1,2,3],
+			      ?line outputv_errors_1([L,T]),
+			      ?line outputv_errors_1([L|T])
+		      end),
+    outputv_errors_1(42),
+
+    %% Test iolists that do not fit in the address space.
+    %% Unfortunately, it would be too slow to test in a 64-bit emulator.
+    case erlang:system_info(wordsize) of
+	4 -> outputv_huge_iolists();
+	_ -> ok
+    end.
+
+outputv_bad_types(Test) ->
+    Types = [-1,256,atom,42.0,{a,b,c},make_ref(),fun() -> 42 end,
+	     [1|2],<<1:1>>,<<1:9>>,<<1:15>>],
+    _ = [Test(Type) || Type <- Types],
     ok.
 
-fun_to_port_1(Term) ->
-    Port = open_port({spawn,outputv_drv}, []),
+outputv_huge_iolists() ->
+    FourGigs = 1 bsl 32,
+    ?line Sizes = [FourGigs+N || N <- lists:seq(0, 64)] ++
+	[1 bsl N || N <- lists:seq(33, 37)],
+    ?line Base = <<0:(1 bsl 20)/unit:8>>,
+    [begin
+	 ?line L = build_iolist(Sz, Base),
+	 ?line outputv_errors_1(L)
+     end || Sz <- Sizes],
+    ok.
+
+outputv_errors_1(Term) ->
+    Port = open_port({spawn_driver,outputv_drv}, []),
     {'EXIT',{badarg,_}} = (catch port_command(Port, Term)),
     port_close(Port).
 
-build_io_list(0) -> [];
-build_io_list(1) -> [7];
-build_io_list(N) ->
-    L = build_io_list(N div 2),
+build_iolist(N, Base) when N < 16 ->
+    case random:uniform(3) of
+	1 ->
+	    <<Bin:N/binary,_/binary>> = Base,
+	    Bin;
+	_ ->
+	    lists:seq(1, N)
+    end;
+build_iolist(N, Base) when N =< byte_size(Base) ->
+    case random:uniform(3) of
+	1 ->
+	    <<Bin:N/binary,_/binary>> = Base,
+	    Bin;
+	2 ->
+	    <<Bin:N/binary,_/binary>> = Base,
+	    [Bin];
+	3 ->
+	    case N rem 2 of
+		0 ->
+		    L = build_iolist(N div 2, Base),
+		    [L,L];
+		1 ->
+		    L = build_iolist(N div 2, Base),
+		    [L,L,45]
+	    end
+    end;
+build_iolist(N0, Base) ->
+    Small = random:uniform(15),
+    Seq = lists:seq(1, Small),
+    N = N0 - Small,
     case N rem 2 of
-	0 -> [L|L];
-	1 -> [7,L|L]
+	0 ->
+	    L = build_iolist(N div 2, Base),
+	    [L,L|Seq];
+	1 ->
+	    L = build_iolist(N div 2, Base),
+	    [47,L,L|Seq]
     end.
 
-
-
 outputv_echo(doc) -> ["Test echoing data with a driver that supports outputv."];
 outputv_echo(Config) when is_list(Config) ->
     ?line Dog = test_server:timetrap(test_server:minutes(10)),
diff --git a/erts/etc/common/erlexec.c b/erts/etc/common/erlexec.c
index 60b3af7db7..90d3be9448 100644
--- a/erts/etc/common/erlexec.c
+++ b/erts/etc/common/erlexec.c
@@ -394,6 +394,7 @@ int main(int argc, char **argv)
     int print_args_exit = 0;
     int print_qouted_cmd_exit = 0;
     erts_cpu_info_t *cpuinfo = NULL;
+    char* emu_name;
 
 #ifdef __WIN32__
     this_module_handle = module;
@@ -566,6 +567,7 @@ int main(int argc, char **argv)
 	    usage("+MYm");
     }
     emu = add_extra_suffixes(emu, emu_type);
+    emu_name = strsave(emu);
     erts_snprintf(tmpStr, sizeof(tmpStr), "%s" DIRSEP "%s" BINARY_EXT, bindir, emu);
     emu = strsave(tmpStr);
 
@@ -682,6 +684,9 @@ int main(int argc, char **argv)
 			verbose = 1;
 		    } else if (strcmp(argv[i], "-emu_args_exit") == 0) {
 			print_args_exit = 1;
+		    } else if (strcmp(argv[i], "-emu_name_exit") == 0) {
+			printf("%s\n", emu_name);
+			exit(0);
 		    } else if (strcmp(argv[i], "-emu_qouted_cmd_exit") == 0) {
 			print_qouted_cmd_exit = 1;
 		    } else if (strcmp(argv[i], "-env") == 0) { /* -env VARNAME VARVALUE */
@@ -1970,6 +1975,11 @@ initial_argv_massage(int *argc, char ***argv)
      */
 
     vix = 0;
+
+    av = build_args_from_env("ERL_" OTP_SYSTEM_VERSION "_FLAGS");
+    if (av)
+	avv[vix++].argv = av;
+
     av = build_args_from_env("ERL_AFLAGS");
     if (av)
 	avv[vix++].argv = av;
@@ -1984,10 +1994,6 @@ initial_argv_massage(int *argc, char ***argv)
     if (av)
 	avv[vix++].argv = av;
 
-    av = build_args_from_env("ERL_" OTP_SYSTEM_VERSION "_FLAGS");
-    if (av)
-	avv[vix++].argv = av;
-
     av = build_args_from_env("ERL_ZFLAGS");
     if (av)
 	avv[vix++].argv = av;
diff --git a/erts/etc/unix/cerl.src b/erts/etc/unix/cerl.src
index 0355f2629f..0b2d6512ea 100644
--- a/erts/etc/unix/cerl.src
+++ b/erts/etc/unix/cerl.src
@@ -62,13 +62,21 @@ cxargs_add() {
     done
 }
 
+eeargs=
+eeargs_add() {
+    while [ $# -gt 0 ]; do
+	cargs="$cargs $1"
+	eeargs="$eeargs $1"
+	shift
+    done
+}
+
 core=
 
 GDB=
 GDBBP=
 GDBARGS=
 TYPE=
-EMU_TYPE=
 debug=
 run_valgrind=no
 
@@ -127,34 +135,37 @@ while [ $# -gt 0 ]; do
 	    ;;
 	"-smp")
 	    shift
-	    cargs="$cargs -smp"
-	    EMU_TYPE=.smp
+	    if [ $# -le 0 ]; then
+		eeargs_add -smp
+	    else
+		case $1 in
+		    disable)
+			shift
+			eeargs_add -smpdisable
+			;;
+		    enable)
+			shift
+			eeargs_add -smp
+			;;
+		    *)
+			eeargs_add -smp
+		esac
+	    fi
+	    ;;
+	"-smpdisable")
+	    shift
+	    eeargs_add -smpdisable
 	    ;;
 	"-lcnt")
 	    shift
 	    cargs="$cargs -lcnt"
 	    TYPE=.lcnt
 	    ;;
-	"-frag")
-	    shift
-	    cargs="$cargs -frag"
-	    EMU_TYPE=.frag
-	    ;;
-	"-smp_frag")
-	    shift
-	    cargs="$cargs -smp_frag"
-	    EMU_TYPE=.smp_frag
-	    ;;
 	"-gprof")
 	    shift
 	    cargs="$cargs -gprof"
 	    TYPE=.gprof
 	    ;;
-	"-hybrid")
-	    shift
-	    cargs="$cargs -hybrid"
-	    EMU_TYPE=.hybrid
-	    ;;
 	"-debug")
 	    shift
 	    cargs="$cargs -debug"
@@ -180,11 +191,6 @@ while [ $# -gt 0 ]; do
 #	    shift
 #	    GDB=xxgdb
 #	    ;;
-	"-shared")
-	    shift
-	    cargs="$cargs -shared"
-	    TYPE=.shared
-	    ;;
 	"-purify")
 	    shift
 	    cargs="$cargs -purify"
@@ -222,7 +228,9 @@ PATH=$BINDIR:$ROOTDIR/bin:$PATH
 EXEC=$BINDIR/erlexec
 
 PROGNAME="$PROGNAME $cargs"
-EMU=$EMU$TYPE$EMU_TYPE
+EMU="$EMU$TYPE"
+EMU_NAME=`$EXEC -emu_name_exit $eeargs`
+
 if [ $run_valgrind != yes ]; then
     xargs="$xargs -pz $PRELOADED --"
 fi
@@ -248,9 +256,9 @@ if [ "x$GDB" = "x" ]; then
 	    valgrind_log=
 	else
 	    if [ $valmajor -gt 2 -a $valminor -gt 4 ]; then
-		valgrind_log="$log_file_prefix$VALGRIND_LOG_DIR/$VALGRIND_LOGFILE_PREFIX$VALGRIND_LOGFILE_INFIX$EMU.log.$$"
+		valgrind_log="$log_file_prefix$VALGRIND_LOG_DIR/$VALGRIND_LOGFILE_PREFIX$VALGRIND_LOGFILE_INFIX$EMU_NAME.log.$$"
 	    else
-		valgrind_log="$log_file_prefix$VALGRIND_LOG_DIR/$VALGRIND_LOGFILE_PREFIX$VALGRIND_LOGFILE_INFIX$EMU.log"
+		valgrind_log="$log_file_prefix$VALGRIND_LOG_DIR/$VALGRIND_LOGFILE_PREFIX$VALGRIND_LOGFILE_INFIX$EMU_NAME.log"
 	    fi
 	fi
 	if [ "x$VALGRIND_MISC_FLAGS" = "x" ]; then
@@ -263,9 +271,9 @@ if [ "x$GDB" = "x" ]; then
 	early_beam_args=`echo $beam_args | sed "s|^\(.*-progname\).*$|\1|g"`
 	late_beam_args=`echo $beam_args | sed "s|^$pre_beam_args.*\(-- -home.*\)$|\1|g"`
 	
-	exec valgrind $valgrind_xml $valgrind_log $valgrind_misc_flags $BINDIR/$EMU $emu_xargs $early_beam_args "$PROGNAME" $late_beam_args -pz $PRELOADED
+	exec valgrind $valgrind_xml $valgrind_log $valgrind_misc_flags $BINDIR/$EMU_NAME $emu_xargs $early_beam_args "$PROGNAME" $late_beam_args -pz $PRELOADED
     else
-	exec $EXEC $xargs ${1+"$@"}
+	exec $EXEC $eeargs $xargs ${1+"$@"}
     fi
 else
     if [ "x$EMACS" = "x" ]; then
@@ -300,5 +308,5 @@ else
             (insert-string \"source $ROOTDIR/erts/etc/unix/etp-commands\") \
             (comint-send-input)"
     # Fire up gdb in emacs...
-    exec $EMACS --eval "(progn (gdb \"gdb $GDBARGS$EMU\") $gdbcmd)"
+    exec $EMACS --eval "(progn (gdb \"gdb $GDBARGS$EMU_NAME\") $gdbcmd)"
 fi
diff --git a/erts/include/internal/erl_printf_format.h b/erts/include/internal/erl_printf_format.h
index 45818079ea..400cc7dafd 100644
--- a/erts/include/internal/erl_printf_format.h
+++ b/erts/include/internal/erl_printf_format.h
@@ -40,7 +40,7 @@ extern int erts_printf_ulong(fmtfn_t, void*, char, int, int, unsigned long);
 extern int erts_printf_slong(fmtfn_t, void*, char, int, int, signed long);
 extern int erts_printf_double(fmtfn_t, void *, char, int, int, double);
 
-extern int (*erts_printf_eterm_func)(fmtfn_t, void*, unsigned long, long);
+extern int (*erts_printf_eterm_func)(fmtfn_t, void*, unsigned long, long, unsigned long*);
 
 #endif
 
diff --git a/erts/include/internal/libatomic_ops/ethr_atomic.h b/erts/include/internal/libatomic_ops/ethr_atomic.h
index d56693dbf8..2fc82c99a8 100644
--- a/erts/include/internal/libatomic_ops/ethr_atomic.h
+++ b/erts/include/internal/libatomic_ops/ethr_atomic.h
@@ -146,13 +146,13 @@ ETHR_NATMC_FUNC__(read)(ETHR_ATMC_T__ *var)
 static ETHR_INLINE ETHR_AINT_T__
 ETHR_NATMC_FUNC__(add_return)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ incr)
 {
-#ifdef AO_HAVE_fetch_and_add
-    return ((ETHR_AINT_T__) AO_fetch_and_add(&var->counter, (AO_t) incr)) + incr;
+#ifdef AO_HAVE_fetch_and_add_full
+    return ((ETHR_AINT_T__) AO_fetch_and_add_full(&var->counter, (AO_t) incr)) + incr;
 #else
     while (1) {
 	AO_t exp = AO_load(&var->counter);
 	AO_t new = exp + (AO_t) incr;
-	if (AO_compare_and_swap(&var->counter, exp, new))
+	if (AO_compare_and_swap_full(&var->counter, exp, new))
 	    return (ETHR_AINT_T__) new;
     }
 #endif
@@ -167,8 +167,8 @@ ETHR_NATMC_FUNC__(add)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ incr)
 static ETHR_INLINE ETHR_AINT_T__
 ETHR_NATMC_FUNC__(inc_return)(ETHR_ATMC_T__ *var)
 {
-#ifdef AO_HAVE_fetch_and_add1
-    return ((ETHR_AINT_T__) AO_fetch_and_add1(&var->counter)) + 1;
+#ifdef AO_HAVE_fetch_and_add1_full
+    return ((ETHR_AINT_T__) AO_fetch_and_add1_full(&var->counter)) + 1;
 #else
     return ETHR_NATMC_FUNC__(add_return)(var, 1);
 #endif
@@ -183,8 +183,8 @@ ETHR_NATMC_FUNC__(inc)(ETHR_ATMC_T__ *var)
 static ETHR_INLINE ETHR_AINT_T__
 ETHR_NATMC_FUNC__(dec_return)(ETHR_ATMC_T__ *var)
 {
-#ifdef AO_HAVE_fetch_and_sub1
-    return ((ETHR_AINT_T__) AO_fetch_and_sub1(&var->counter)) - 1;
+#ifdef AO_HAVE_fetch_and_sub1_full
+    return ((ETHR_AINT_T__) AO_fetch_and_sub1_full(&var->counter)) - 1;
 #else
     return ETHR_NATMC_FUNC__(add_return)(var, -1);
 #endif
@@ -202,7 +202,7 @@ ETHR_NATMC_FUNC__(and_retold)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ mask)
     while (1) {
 	AO_t exp = AO_load(&var->counter);
 	AO_t new = exp & ((AO_t) mask);
-	if (AO_compare_and_swap(&var->counter, exp, new))
+	if (AO_compare_and_swap_full(&var->counter, exp, new))
 	    return (ETHR_AINT_T__) exp;
     }
 }
@@ -213,7 +213,7 @@ ETHR_NATMC_FUNC__(or_retold)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ mask)
     while (1) {
 	AO_t exp = AO_load(&var->counter);
 	AO_t new = exp | ((AO_t) mask);
-	if (AO_compare_and_swap(&var->counter, exp, new))
+	if (AO_compare_and_swap_full(&var->counter, exp, new))
 	    return (ETHR_AINT_T__) exp;
     }
 }
@@ -225,7 +225,7 @@ ETHR_NATMC_FUNC__(cmpxchg)(ETHR_ATMC_T__ *var,
 {
     ETHR_AINT_T__ act;
     do {
-	if (AO_compare_and_swap(&var->counter, (AO_t) exp, (AO_t) new))
+	if (AO_compare_and_swap_full(&var->counter, (AO_t) exp, (AO_t) new))
 	    return exp;
 	act = (ETHR_AINT_T__) AO_load(&var->counter);
     } while (act == exp);
@@ -237,7 +237,7 @@ ETHR_NATMC_FUNC__(xchg)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ new)
 {
     while (1) {
 	AO_t exp = AO_load(&var->counter);
-	if (AO_compare_and_swap(&var->counter, exp, (AO_t) new))
+	if (AO_compare_and_swap_full(&var->counter, exp, (AO_t) new))
 	    return (ETHR_AINT_T__) exp;
     }
 }
@@ -265,7 +265,6 @@ ETHR_NATMC_FUNC__(inc_return_acqb)(ETHR_ATMC_T__ *var)
     return ((ETHR_AINT_T__) AO_fetch_and_add1_acquire(&var->counter)) + 1;
 #else
     ETHR_AINT_T__ res = ETHR_NATMC_FUNC__(add_return)(var, 1);
-    ETHR_MEMORY_BARRIER;
     return res;
 #endif
 }
@@ -287,7 +286,6 @@ ETHR_NATMC_FUNC__(dec_return_relb)(ETHR_ATMC_T__ *var)
 #ifdef AO_HAVE_fetch_and_sub1_release
     return ((ETHR_AINT_T__) AO_fetch_and_sub1_release(&var->counter)) - 1;
 #else
-    ETHR_MEMORY_BARRIER;
     return ETHR_NATMC_FUNC__(dec_return)(var);
 #endif
 }
@@ -314,7 +312,6 @@ ETHR_NATMC_FUNC__(cmpxchg_acqb)(ETHR_ATMC_T__ *var,
     return act;
 #else
     ETHR_AINT_T__ act = ETHR_NATMC_FUNC__(cmpxchg)(var, new, exp);
-    ETHR_MEMORY_BARRIER;
     return act;
 #endif
 }
@@ -333,7 +330,6 @@ ETHR_NATMC_FUNC__(cmpxchg_relb)(ETHR_ATMC_T__ *var,
     } while (act == exp);
     return act;
 #else
-    ETHR_MEMORY_BARRIER;
     return ETHR_NATMC_FUNC__(cmpxchg)(var, new, exp);
 #endif
 }
diff --git a/erts/include/internal/sparc32/atomic.h b/erts/include/internal/sparc32/atomic.h
index 00380dbf07..16182f8b01 100644
--- a/erts/include/internal/sparc32/atomic.h
+++ b/erts/include/internal/sparc32/atomic.h
@@ -95,7 +95,7 @@ ETHR_NATMC_FUNC__(add_return)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ incr)
 {
     ETHR_AINT_T__ old, tmp;
 
-    __asm__ __volatile__("membar #LoadLoad|#StoreLoad\n");
+    __asm__ __volatile__("membar #LoadLoad|#StoreLoad\n" : : : "memory");
     do {
 	old = var->counter;
 	tmp = old+incr;
@@ -105,7 +105,7 @@ ETHR_NATMC_FUNC__(add_return)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ incr)
 	    : "r"(old), "r"(&var->counter), "0"(tmp)
 	    : "memory");
     } while (__builtin_expect(old != tmp, 0));
-    __asm__ __volatile__("membar #StoreLoad|#StoreStore");
+    __asm__ __volatile__("membar #StoreLoad|#StoreStore" : : : "memory");
     return old+incr;
 }
 
@@ -144,7 +144,7 @@ ETHR_NATMC_FUNC__(and_retold)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ mask)
 {
     ETHR_AINT_T__ old, tmp;
 
-    __asm__ __volatile__("membar #LoadLoad|#StoreLoad\n");
+    __asm__ __volatile__("membar #LoadLoad|#StoreLoad\n" : : : "memory");
     do {
 	old = var->counter;
 	tmp = old & mask;
@@ -154,7 +154,7 @@ ETHR_NATMC_FUNC__(and_retold)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ mask)
 	    : "r"(old), "r"(&var->counter), "0"(tmp)
 	    : "memory");
     } while (__builtin_expect(old != tmp, 0));
-    __asm__ __volatile__("membar #StoreLoad|#StoreStore");
+    __asm__ __volatile__("membar #StoreLoad|#StoreStore" : : : "memory");
     return old;
 }
 
@@ -163,7 +163,7 @@ ETHR_NATMC_FUNC__(or_retold)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ mask)
 {
     ETHR_AINT_T__ old, tmp;
 
-    __asm__ __volatile__("membar #LoadLoad|#StoreLoad\n");
+    __asm__ __volatile__("membar #LoadLoad|#StoreLoad\n" : : : "memory");
     do {
 	old = var->counter;
 	tmp = old | mask;
@@ -173,7 +173,7 @@ ETHR_NATMC_FUNC__(or_retold)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ mask)
 	    : "r"(old), "r"(&var->counter), "0"(tmp)
 	    : "memory");
     } while (__builtin_expect(old != tmp, 0));
-    __asm__ __volatile__("membar #StoreLoad|#StoreStore");
+    __asm__ __volatile__("membar #StoreLoad|#StoreStore" : : : "memory");
     return old;
 }
 
@@ -182,7 +182,7 @@ ETHR_NATMC_FUNC__(xchg)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ val)
 {
     ETHR_AINT_T__ old, new;
 
-    __asm__ __volatile__("membar #LoadLoad|#StoreLoad");
+    __asm__ __volatile__("membar #LoadLoad|#StoreLoad" : : : "memory");
     do {
 	old = var->counter;
 	new = val;
@@ -192,20 +192,20 @@ ETHR_NATMC_FUNC__(xchg)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ val)
 	    : "r"(old), "r"(&var->counter), "0"(new)
 	    : "memory");
     } while (__builtin_expect(old != new, 0));
-    __asm__ __volatile__("membar #StoreLoad|#StoreStore");
+    __asm__ __volatile__("membar #StoreLoad|#StoreStore" : : : "memory");
     return old;
 }
 
 static ETHR_INLINE ETHR_AINT_T__
 ETHR_NATMC_FUNC__(cmpxchg)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ new, ETHR_AINT_T__ old)
 {
-    __asm__ __volatile__("membar #LoadLoad|#StoreLoad\n");
+    __asm__ __volatile__("membar #LoadLoad|#StoreLoad\n" : : : "memory");
     __asm__ __volatile__(
       ETHR_CAS__ " [%2], %1, %0"
       : "=&r"(new)
       : "r"(old), "r"(&var->counter), "0"(new)
       : "memory");
-    __asm__ __volatile__("membar #StoreLoad|#StoreStore");
+    __asm__ __volatile__("membar #StoreLoad|#StoreStore" : : : "memory");
     return new;
 }
 
@@ -213,13 +213,11 @@ ETHR_NATMC_FUNC__(cmpxchg)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ new, ETHR_AINT_T__
  * Atomic ops with at least specified barriers.
  */
 
-/* TODO: relax acquire barriers */
-
 static ETHR_INLINE ETHR_AINT_T__
 ETHR_NATMC_FUNC__(read_acqb)(ETHR_ATMC_T__ *var)
 {
     ETHR_AINT_T__ res = ETHR_NATMC_FUNC__(read)(var);
-    __asm__ __volatile__("membar #LoadLoad|#LoadStore|#StoreLoad|#StoreStore" : : : "memory");
+    __asm__ __volatile__("membar #LoadLoad|#LoadStore" : : : "memory");
     return res;
 }
 
@@ -234,21 +232,18 @@ static ETHR_INLINE ETHR_AINT_T__
 ETHR_NATMC_FUNC__(inc_return_acqb)(ETHR_ATMC_T__ *var)
 {
     ETHR_AINT_T__ res = ETHR_NATMC_FUNC__(inc_return)(var);
-    __asm__ __volatile__("membar #LoadLoad|#LoadStore" : : : "memory");
     return res;
 }
 
 static ETHR_INLINE void
 ETHR_NATMC_FUNC__(dec_relb)(ETHR_ATMC_T__ *var)
 {
-    __asm__ __volatile__("membar #LoadStore|#StoreStore" : : : "memory");
     ETHR_NATMC_FUNC__(dec)(var);
 }
 
 static ETHR_INLINE ETHR_AINT_T__
 ETHR_NATMC_FUNC__(dec_return_relb)(ETHR_ATMC_T__ *var)
 {
-    __asm__ __volatile__("membar #LoadStore|#StoreStore" : : : "memory");
     return ETHR_NATMC_FUNC__(dec_return)(var);
 }
 
@@ -256,14 +251,12 @@ static ETHR_INLINE ETHR_AINT_T__
 ETHR_NATMC_FUNC__(cmpxchg_acqb)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ new, ETHR_AINT_T__ old)
 {
     ETHR_AINT_T__ res = ETHR_NATMC_FUNC__(cmpxchg)(var, new, old);
-    __asm__ __volatile__("membar #LoadLoad|#LoadStore" : : : "memory");
     return res;
 }
 
 static ETHR_INLINE ETHR_AINT_T__
 ETHR_NATMC_FUNC__(cmpxchg_relb)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ new, ETHR_AINT_T__ old)
 {
-    __asm__ __volatile__("membar #LoadStore|#StoreStore" : : : "memory");
     return ETHR_NATMC_FUNC__(cmpxchg)(var, new, old);
 }
 
diff --git a/erts/include/internal/tile/atomic.h b/erts/include/internal/tile/atomic.h
index 48e4c0c6c8..0c7b597a6b 100644
--- a/erts/include/internal/tile/atomic.h
+++ b/erts/include/internal/tile/atomic.h
@@ -65,25 +65,35 @@ ethr_native_atomic32_read(ethr_native_atomic32_t *var)
 static ETHR_INLINE void
 ethr_native_atomic32_add(ethr_native_atomic32_t *var, ethr_sint32_t incr)
 {
+    ETHR_MEMORY_BARRIER;
     atomic_add(&var->counter, incr);
+    ETHR_MEMORY_BARRIER;
 }      
        
 static ETHR_INLINE void
 ethr_native_atomic32_inc(ethr_native_atomic32_t *var)
 {
+    ETHR_MEMORY_BARRIER;
     atomic_increment(&var->counter);
+    ETHR_MEMORY_BARRIER;
 }
 
 static ETHR_INLINE void
 ethr_native_atomic32_dec(ethr_native_atomic32_t *var)
 {
+    ETHR_MEMORY_BARRIER;
     atomic_decrement(&var->counter);
+    ETHR_MEMORY_BARRIER;
 }
 
 static ETHR_INLINE ethr_sint32_t
 ethr_native_atomic32_add_return(ethr_native_atomic32_t *var, ethr_sint32_t incr)
 {
-    return atomic_exchange_and_add(&var->counter, incr) + incr;
+    ethr_sint32_t res;
+    ETHR_MEMORY_BARRIER;
+    res = atomic_exchange_and_add(&var->counter, incr) + incr;
+    ETHR_MEMORY_BARRIER;
+    return res;
 }
 
 static ETHR_INLINE ethr_sint32_t
@@ -101,18 +111,27 @@ ethr_native_atomic32_dec_return(ethr_native_atomic32_t *var)
 static ETHR_INLINE ethr_sint32_t
 ethr_native_atomic32_and_retold(ethr_native_atomic32_t *var, ethr_sint32_t mask)
 {
-    return atomic_and_val(&var->counter, mask);
+    ethr_sint32_t res;
+    ETHR_MEMORY_BARRIER;
+    res = atomic_and_val(&var->counter, mask);
+    ETHR_MEMORY_BARRIER;
+    return res;
 }
 
 static ETHR_INLINE ethr_sint32_t
 ethr_native_atomic32_or_retold(ethr_native_atomic32_t *var, ethr_sint32_t mask)
 {
-    return atomic_or_val(&var->counter, mask);
+    ethr_sint32_t res;
+    ETHR_MEMORY_BARRIER;
+    res = atomic_or_val(&var->counter, mask);
+    ETHR_MEMORY_BARRIER;
+    return res;
 }
 
 static ETHR_INLINE ethr_sint32_t
 ethr_native_atomic32_xchg(ethr_native_atomic32_t *var, ethr_sint32_t val)
 {   
+    ETHR_MEMORY_BARRIER;
     return atomic_exchange_acq(&var->counter, val);
 } 
 
@@ -121,6 +140,7 @@ ethr_native_atomic32_cmpxchg(ethr_native_atomic32_t *var,
 			     ethr_sint32_t new,
 			     ethr_sint32_t expected)
 {
+    ETHR_MEMORY_BARRIER;
     return atomic_compare_and_exchange_val_acq(&var->counter, new, expected);
 }
 
@@ -139,9 +159,7 @@ ethr_native_atomic32_read_acqb(ethr_native_atomic32_t *var)
 static ETHR_INLINE ethr_sint32_t
 ethr_native_atomic32_inc_return_acqb(ethr_native_atomic32_t *var)
 {
-    ethr_sint32_t res = ethr_native_atomic32_inc_return(var);
-    ETHR_MEMORY_BARRIER;
-    return res;
+    return ethr_native_atomic32_inc_return(var);
 }
 
 static ETHR_INLINE void
@@ -154,14 +172,12 @@ ethr_native_atomic32_set_relb(ethr_native_atomic32_t *var, ethr_sint32_t val)
 static ETHR_INLINE void
 ethr_native_atomic32_dec_relb(ethr_native_atomic32_t *var)
 {
-    ETHR_MEMORY_BARRIER;
     ethr_native_atomic32_dec(var);
 }
 
 static ETHR_INLINE ethr_sint32_t
 ethr_native_atomic32_dec_return_relb(ethr_native_atomic32_t *var)
 {
-    ETHR_MEMORY_BARRIER;
     return ethr_native_atomic32_dec_return(var);
 }
 
@@ -178,7 +194,6 @@ ethr_native_atomic32_cmpxchg_relb(ethr_native_atomic32_t *var,
 				  ethr_sint32_t new,
 				  ethr_sint32_t exp)
 {
-    ETHR_MEMORY_BARRIER;
     return ethr_native_atomic32_cmpxchg(var, new, exp);
 }
 
diff --git a/erts/lib_src/common/erl_printf_format.c b/erts/lib_src/common/erl_printf_format.c
index 968d563325..fba3fd723c 100644
--- a/erts/lib_src/common/erl_printf_format.c
+++ b/erts/lib_src/common/erl_printf_format.c
@@ -25,7 +25,7 @@
  *  width:      [0-9]+ | '*' 
  *  precision:  [0-9]+ | '*' 
  *  length: hh | h | l | ll | L | j | t | b<sz>
- *  conversion: d,i | o,u,x,X | e,E | f,F | g,G | a,A | c | s | T |
+ *  conversion: d,i | o,u,x,X | e,E | f,F | g,G | a,A | c | s | T | R |
  *              p | n | %
  *  sz: 8 | 16 | 32 | 64 | p | e
  */
@@ -101,7 +101,7 @@
 #endif
 
 #define FMTC_d    0x0000
-#define FMTC_i    0x0001
+#define FMTC_R    0x0001
 #define FMTC_o    0x0002
 #define FMTC_u    0x0003
 #define FMTC_x    0x0004
@@ -165,7 +165,7 @@ static char heX[] = "0123456789ABCDEF";
 #define SIGN(X) ((X) > 0 ? 1 : ((X) < 0 ? -1 : 0)) 
 #define USIGN(X) ((X) == 0 ? 0 : 1)
 
-int (*erts_printf_eterm_func)(fmtfn_t, void*, unsigned long, long) = NULL;
+int (*erts_printf_eterm_func)(fmtfn_t, void*, unsigned long, long, unsigned long*) = NULL;
 
 static int
 noop_fn(void *vfp, char* buf, size_t len)
@@ -183,8 +183,8 @@ static int fmt_fld(fmtfn_t fn,void* arg,
     int len;
 
     /* format the prefix */
-    if ((sign || (fmt & (FMTF_sgn|FMTF_blk))) &&
-	(((fmt & FMTC_MASK) == FMTC_d) || ((fmt & FMTC_MASK) == FMTC_i))) {
+    if ((sign || (fmt & (FMTF_sgn|FMTF_blk)))
+	&& (fmt & FMTC_MASK) == FMTC_d) {
 	if (sign < 0)
 	    *pp++ = '-';
 	else if ((fmt & FMTF_sgn))
@@ -245,7 +245,6 @@ static int fmt_long(fmtfn_t fn,void* arg,int sign,unsigned long uval,
 
     switch(fmt & FMTC_MASK) {
     case FMTC_d:
-    case FMTC_i:
     case FMTC_u:
 	break;
     case FMTC_o: 
@@ -298,7 +297,6 @@ static int fmt_long_long(fmtfn_t fn,void* arg,int sign,
 
     switch(fmt & FMTC_MASK) {
     case FMTC_d:
-    case FMTC_i:
     case FMTC_u:
 	break;
     case FMTC_o: 
@@ -622,7 +620,7 @@ int erts_printf_format(fmtfn_t fn, void* arg, char* fmt, va_list ap)
 	    /* specifier */
 	    switch(*ptr) {
 	    case 'd': ptr++; fmt |= FMTC_d; break;
-	    case 'i': ptr++; fmt |= FMTC_i; break;
+	    case 'i': ptr++; fmt |= FMTC_d; break;
 	    case 'o': ptr++; fmt |= FMTC_o; break;
 	    case 'u': ptr++; fmt |= FMTC_u; break;
 	    case 'x': ptr++; fmt |= FMTC_x; break;
@@ -637,6 +635,7 @@ int erts_printf_format(fmtfn_t fn, void* arg, char* fmt, va_list ap)
 	    case 'p': ptr++; fmt |= FMTC_p; break;
 	    case 'n': ptr++; fmt |= FMTC_n; break;
 	    case 'T': ptr++; fmt |= FMTC_T; break;
+	    case 'R': ptr++; fmt |= FMTC_R; break;
 	    case '%':
 		FMT(fn,arg,ptr,1,count);
 		ptr++;
@@ -650,7 +649,6 @@ int erts_printf_format(fmtfn_t fn, void* arg, char* fmt, va_list ap)
 
 	    switch(fmt & FMTC_MASK) {
 	    case FMTC_d:
-	    case FMTC_i:
 		switch(fmt & FMTL_MASK) {
 		case FMTL_hh: {
 		    signed char tval = (signed char) va_arg(ap,int);
@@ -814,9 +812,12 @@ int erts_printf_format(fmtfn_t fn, void* arg, char* fmt, va_list ap)
 		default: *va_arg(ap,int*) = count; break;
 		}
 		break;
-	    case FMTC_T: {
+	    case FMTC_T:    /* Eterm */
+	    case FMTC_R: {  /* Eterm, Eterm* base  (base ignored if !HALFWORD_HEAP) */
 		long prec;
 		unsigned long eterm;
+		unsigned long* eterm_base;
+		
 		if (!erts_printf_eterm_func)
 		    return -EINVAL;
 		if (precision < 0)
@@ -826,14 +827,16 @@ int erts_printf_format(fmtfn_t fn, void* arg, char* fmt, va_list ap)
 		else
 		    prec = (long) precision;
 		eterm = va_arg(ap, unsigned long);
+		eterm_base = ((fmt & FMTC_MASK) == FMTC_R) ?
+		    va_arg(ap, unsigned long*) : NULL;
 		if (width > 0 && !(fmt & FMTF_adj)) {
-		    res = (*erts_printf_eterm_func)(noop_fn, NULL, eterm, prec);
+		    res = (*erts_printf_eterm_func)(noop_fn, NULL, eterm, prec, eterm_base);
 		    if (res < 0)
 			return res;
 		    if (width > res)
 			BLANKS(fn, arg, width - res, count);
 		}
-		res = (*erts_printf_eterm_func)(fn, arg, eterm, prec);
+		res = (*erts_printf_eterm_func)(fn, arg, eterm, prec, eterm_base);
 		if (res < 0)
 		    return res;
 		count += res;
@@ -924,7 +927,7 @@ erts_printf_slong(fmtfn_t fn, void *arg, char conv, int pad, int width,
     unsigned long ul_val;
     switch (conv) {
     case 'd': fmt |= FMTC_d; break;
-    case 'i': fmt |= FMTC_i; break;
+    case 'i': fmt |= FMTC_d; break;
     case 'o': fmt |= FMTC_o; break;
     case 'x': fmt |= FMTC_x; break;
     case 'X': fmt |= FMTC_X; break;