Store pointers to heap data in 32-bit words

Store Erlang terms in 32-bit entities on the heap, expanding the
pointers to 64-bit when needed. This works because all terms are stored
on addresses in the 32-bit address range (the 32 most significant bits
of pointers to term data are always 0).

Introduce a new datatype called UWord (along with its companion SWord),
which is an integer having the exact same size as the machine word
(a void *), but might be larger than Eterm/Uint.

Store code as machine words, as the instructions are pointers to
executable code which might reside outside the 32-bit address range.

Continuation pointers are stored on the 32-bit stack and hence must
point to addresses in the low range, which means that loaded beam code
much be placed in the low 32-bit address range (but, as said earlier,
the instructions themselves are full words).

No Erlang term data can be stored on C stacks (enforced by an
earlier commit).

This version gives a prompt, but test cases still fail (and dump core).

The loader (and emulator loop) has instruction packing disabled.

The main issues has been in rewriting loader and actual virtual
machine. Subsystems (like distribution) does not work yet.

73 files changed, 1492 insertions, 1049 deletions

diff --git a/configure.in b/configure.in
index 21f9a640ff..643a7cecfa 100644
--- a/configure.in
+++ b/configure.in
@@ -195,6 +195,10 @@ AC_ARG_ENABLE(threads,
 [  --enable-threads        enable async thread support
   --disable-threads       disable async thread support])
 
+AC_ARG_ENABLE(halfword-emulator,
+[  --enable-halfword-emulator        enable halfword emulator (only for 64bit builds)
+  --disable-halfword-emulator        disable halfword emulator (only for 64bit builds)])
+
 AC_ARG_ENABLE(smp-support,
 [  --enable-smp-support    enable smp support
   --disable-smp-support   disable smp support])
diff --git a/erts/configure.in b/erts/configure.in
index 5fa1245b13..6823133936 100644
--- a/erts/configure.in
+++ b/erts/configure.in
@@ -123,6 +123,14 @@ AC_ARG_ENABLE(threads,
     *)  enable_threads=yes ;;
   esac ], enable_threads=unknown)
 
+AC_ARG_ENABLE(halfword-emulator,
+[  --enable-halfword-emulator        enable halfword emulator (only for 64bit builds)
+  --disable-halfword-emulator        disable halfword emulator (only for 64bit builds)],
+[ case "$enableval" in
+    no) enable_halfword_emualtor=no ;;
+    *)  enable_halfword_emulator=yes ;;
+  esac ], enable_halfword_emulator=unknown)
+
 AC_ARG_ENABLE(smp-support,
 [  --enable-smp-support    enable smp support
   --disable-smp-support   disable smp support],
@@ -749,6 +757,25 @@ esac
 	
 AC_SUBST(LIBCARBON)
 
+dnl Check if we should/can build a halfword emulator
+
+AC_MSG_CHECKING(if we are building a halfword emulator (32bit heap on 64bit machine))
+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])
+		AC_MSG_RESULT([yes])
+	else
+		AC_MSG_ERROR(no; halfword emulator not supported on this architecture)
+	fi
+else
+	AC_MSG_RESULT([no])
+fi
+
+
+
+
+
 dnl some tests below will call this if we haven't already - and autoconf
 dnl can't handle those tests being done conditionally at runtime
 AC_PROG_CPP
diff --git a/erts/emulator/beam/beam_bif_load.c b/erts/emulator/beam/beam_bif_load.c
index b1feec7074..db557c27a5 100644
--- a/erts/emulator/beam/beam_bif_load.c
+++ b/erts/emulator/beam/beam_bif_load.c
@@ -39,10 +39,10 @@ static Eterm check_process_code(Process* rp, Module* modp);
 static void delete_code(Process *c_p, ErtsProcLocks c_p_locks, Module* modp);
 static void delete_export_references(Eterm module);
 static int purge_module(int module);
-static int is_native(Eterm* code);
+static int is_native(UWord* code);
 static int any_heap_ref_ptrs(Eterm* start, Eterm* end, char* mod_start, Uint mod_size);
 static int any_heap_refs(Eterm* start, Eterm* end, char* mod_start, Uint mod_size);
-static void remove_from_address_table(Eterm* code);
+static void remove_from_address_table(UWord* code);
 
 Eterm
 load_module_2(BIF_ALIST_2)
@@ -344,8 +344,8 @@ BIF_RETTYPE finish_after_on_load_2(BIF_ALIST_2)
 	modp->code[MI_ON_LOAD_FUNCTION_PTR] = 0;
 	set_default_trace_pattern(BIF_ARG_1);
     } else if (BIF_ARG_2 == am_false) {
-	Eterm* code;
-	Eterm* end;
+	UWord* code;
+	UWord* end;
 
 	/*
 	 * The on_load function failed. Remove the loaded code.
@@ -354,7 +354,7 @@ BIF_RETTYPE finish_after_on_load_2(BIF_ALIST_2)
 	 */
 	erts_total_code_size -= modp->code_length;
 	code = modp->code;
-	end = (Eterm *)((char *)code + modp->code_length);
+	end = (UWord *)((char *)code + modp->code_length);
 	erts_cleanup_funs_on_purge(code, end);
 	beam_catches_delmod(modp->catches, code, modp->code_length);
 	erts_free(ERTS_ALC_T_CODE, (void *) code);
@@ -397,10 +397,10 @@ set_default_trace_pattern(Eterm module)
 static Eterm
 check_process_code(Process* rp, Module* modp)
 {
-    Eterm* start;
+    UWord* start;
     char* mod_start;
     Uint mod_size;
-    Eterm* end;
+    UWord* end;
     Eterm* sp;
 #ifndef HYBRID /* FIND ME! */
     ErlFunThing* funp;
@@ -418,7 +418,7 @@ check_process_code(Process* rp, Module* modp)
      * Pick up limits for the module.
      */
     start = modp->old_code;
-    end = (Eterm *)((char *)start + modp->old_code_length);
+    end = (UWord *)((char *)start + modp->old_code_length);
     mod_start = (char *) start;
     mod_size = modp->old_code_length;
 
@@ -472,11 +472,11 @@ check_process_code(Process* rp, Module* modp)
 #ifndef HYBRID /* FIND ME! */
  rescan:
     for (funp = MSO(rp).funs; funp; funp = funp->next) {
-	Eterm* fun_code;
+	UWord* fun_code;
 
 	fun_code = funp->fe->address;
 
-	if (INSIDE((Eterm *) funp->fe->address)) {
+	if (INSIDE((UWord *) funp->fe->address)) {
 	    if (done_gc) {
 		return am_true;
 	    } else {
@@ -576,7 +576,7 @@ any_heap_ref_ptrs(Eterm* start, Eterm* end, char* mod_start, Uint mod_size)
 	switch (primary_tag(val)) {
 	case TAG_PRIMARY_BOXED:
 	case TAG_PRIMARY_LIST:
-	    if (in_area(val, mod_start, mod_size)) {
+	    if (in_area(EXPAND_POINTER(val), mod_start, mod_size)) {
 		return 1;
 	    }
 	    break;
@@ -596,7 +596,7 @@ any_heap_refs(Eterm* start, Eterm* end, char* mod_start, Uint mod_size)
 	switch (primary_tag(val)) {
 	case TAG_PRIMARY_BOXED:
 	case TAG_PRIMARY_LIST:
-	    if (in_area(val, mod_start, mod_size)) {
+	    if (in_area(EXPAND_POINTER(val), mod_start, mod_size)) {
 		return 1;
 	    }
 	    break;
@@ -617,8 +617,8 @@ any_heap_refs(Eterm* start, Eterm* end, char* mod_start, Uint mod_size)
 static int
 purge_module(int module)
 {
-    Eterm* code;
-    Eterm* end;
+    UWord* code;
+    UWord* end;
     Module* modp;
 
     /*
@@ -653,7 +653,7 @@ purge_module(int module)
     ASSERT(erts_total_code_size >= modp->old_code_length);
     erts_total_code_size -= modp->old_code_length;
     code = modp->old_code;
-    end = (Eterm *)((char *)code + modp->old_code_length);
+    end = (UWord *)((char *)code + modp->old_code_length);
     erts_cleanup_funs_on_purge(code, end);
     beam_catches_delmod(modp->old_catches, code, modp->old_code_length);
     erts_free(ERTS_ALC_T_CODE, (void *) code);
@@ -665,7 +665,7 @@ purge_module(int module)
 }
 
 static void
-remove_from_address_table(Eterm* code)
+remove_from_address_table(UWord* code)
 {
     int i;
 
@@ -738,11 +738,11 @@ delete_export_references(Eterm module)
 	Export *ep = export_list(i);
         if (ep != NULL && (ep->code[0] == module)) {
 	    if (ep->address == ep->code+3 &&
-		(ep->code[3] == (Eterm) em_apply_bif)) {
+		(ep->code[3] == (UWord) em_apply_bif)) {
 		continue;
 	    }
 	    ep->address = ep->code+3;
-	    ep->code[3] = (Uint) em_call_error_handler;
+	    ep->code[3] = (UWord) em_call_error_handler;
 	    ep->code[4] = 0;
 	    MatchSetUnref(ep->match_prog_set);
 	    ep->match_prog_set = NULL;
@@ -774,7 +774,7 @@ beam_make_current_old(Process *c_p, ErtsProcLocks c_p_locks, Eterm module)
 }
 
 static int
-is_native(Eterm* code)
+is_native(UWord* code)
 {
     return ((Eterm *)code[MI_FUNCTIONS])[1] != 0;
 }
diff --git a/erts/emulator/beam/beam_bp.c b/erts/emulator/beam/beam_bp.c
index 1abf1dc10c..0cdc2116ba 100644
--- a/erts/emulator/beam/beam_bp.c
+++ b/erts/emulator/beam/beam_bp.c
@@ -262,7 +262,7 @@ erts_clear_function_break(Module *modp, Uint *pc) {
  * SMP NOTE: Process p may have become exiting on return!
  */
 Uint 
-erts_trace_break(Process *p, Uint *pc, Eterm *args, 
+erts_trace_break(Process *p, UWord *pc, Eterm *args,
 		 Uint32 *ret_flags, Eterm *tracer_pid) {
     Eterm tpid1, tpid2;
     BpDataTrace *bdt = (BpDataTrace *) pc[-4];
@@ -296,7 +296,7 @@ erts_trace_break(Process *p, Uint *pc, Eterm *args,
  * SMP NOTE: Process p may have become exiting on return!
  */
 Uint32
-erts_bif_mtrace(Process *p, Uint *pc, Eterm *args, int local, 
+erts_bif_mtrace(Process *p, UWord *pc, Eterm *args, int local,
 		Eterm *tracer_pid) {
     BpDataTrace *bdt = (BpDataTrace *) pc[-4];
     
@@ -364,7 +364,7 @@ erts_is_mtrace_break(Uint *pc, Binary **match_spec_ret, Eterm *tracer_pid_ret) {
 }
 
 int
-erts_is_mtrace_bif(Uint *pc, Binary **match_spec_ret, Eterm *tracer_pid_ret) {
+erts_is_mtrace_bif(UWord *pc, Binary **match_spec_ret, Eterm *tracer_pid_ret) {
     BpDataTrace *bdt = (BpDataTrace *) pc[-4];
     
     if (bdt) {
diff --git a/erts/emulator/beam/beam_bp.h b/erts/emulator/beam/beam_bp.h
index 44e6b294d8..44b463d39c 100644
--- a/erts/emulator/beam/beam_bp.h
+++ b/erts/emulator/beam/beam_bp.h
@@ -89,11 +89,11 @@ extern erts_smp_spinlock_t erts_bp_lock;
 do {                                                        \
     BpDataCount *bdc = (BpDataCount *) (pc)[-4];            \
                                                             \
-    ASSERT((pc)[-5] == (Uint) BeamOp(op_i_func_info_IaaI)); \
+    ASSERT((pc)[-5] == (UWord) BeamOp(op_i_func_info_IaaI)); \
     ASSERT(bdc);                                            \
     bdc = (BpDataCount *) bdc->next;                        \
     ASSERT(bdc);                                            \
-    (pc)[-4] = (Uint) bdc;                                  \
+    (pc)[-4] = (UWord) bdc;                                  \
     ErtsSmpBPLock(bdc);                                     \
     if (bdc->count >= 0) bdc->count++;                      \
     ErtsSmpBPUnlock(bdc);                                   \
@@ -104,11 +104,11 @@ do {                                                        \
 do {                                                        \
     BpData *bd = (BpData *) (pc)[-4];                       \
                                                             \
-    ASSERT((pc)[-5] == (Uint) BeamOp(op_i_func_info_IaaI)); \
+    ASSERT((pc)[-5] == (UWord) BeamOp(op_i_func_info_IaaI)); \
     ASSERT(bd);                                             \
     bd = bd->next;                                          \
     ASSERT(bd);                                             \
-    (pc)[-4] = (Uint) bd;                                   \
+    (pc)[-4] = (UWord) bd;                                   \
     *(instr_result) = bd->orig_instr;                       \
 } while (0)
 
@@ -146,16 +146,16 @@ int erts_clear_break(Eterm mfa[3], int specified);
 int erts_clear_module_break(Module *modp);
 int erts_clear_function_break(Module *modp, Uint *pc);
 
-Uint erts_trace_break(Process *p, Uint *pc, Eterm *args, 
+Uint erts_trace_break(Process *p, UWord *pc, Eterm *args,
 		      Uint32 *ret_flags, Eterm *tracer_pid);
-Uint32 erts_bif_mtrace(Process *p, Uint *pc, Eterm *args, 
+Uint32 erts_bif_mtrace(Process *p, UWord *pc, Eterm *args,
 		       int local, Eterm *tracer_pid);
 
-int erts_is_trace_break(Uint *pc, Binary **match_spec_ret, 
+int erts_is_trace_break(Uint *pc, Binary **match_spec_ret,
 			Eterm *tracer_pid_ret);
 int erts_is_mtrace_break(Uint *pc, Binary **match_spec_ret, 
 			 Eterm *tracer_pid_rte);
-int erts_is_mtrace_bif(Uint *pc, Binary **match_spec_ret, 
+int erts_is_mtrace_bif(UWord *pc, Binary **match_spec_ret,
 		       Eterm *tracer_pid_ret);
 int erts_is_native_break(Uint *pc);
 int erts_is_count_break(Uint *pc, Sint *count_ret);
diff --git a/erts/emulator/beam/beam_catches.c b/erts/emulator/beam/beam_catches.c
index d5cef1cad2..af99b0499a 100644
--- a/erts/emulator/beam/beam_catches.c
+++ b/erts/emulator/beam/beam_catches.c
@@ -26,7 +26,7 @@
 /* XXX: should use dynamic reallocation */
 #define TABSIZ (16*1024)
 static struct {
-    Eterm *cp;
+    UWord *cp;
     unsigned cdr;
 } beam_catches[TABSIZ];
 
@@ -39,7 +39,7 @@ void beam_catches_init(void)
     high_mark = 0;
 }
 
-unsigned beam_catches_cons(Eterm *cp, unsigned cdr)
+unsigned beam_catches_cons(UWord *cp, unsigned cdr)
 {
     int i;
 
@@ -65,7 +65,7 @@ unsigned beam_catches_cons(Eterm *cp, unsigned cdr)
     return i;
 }
 
-Eterm *beam_catches_car(unsigned i)
+UWord *beam_catches_car(unsigned i)
 {
     if( i >= TABSIZ ) {
 	fprintf(stderr,
@@ -75,7 +75,7 @@ Eterm *beam_catches_car(unsigned i)
     return beam_catches[i].cp;
 }
 
-void beam_catches_delmod(unsigned head, Eterm *code, unsigned code_bytes)
+void beam_catches_delmod(unsigned head, UWord *code, unsigned code_bytes)
 {
     unsigned i, cdr;
 
diff --git a/erts/emulator/beam/beam_catches.h b/erts/emulator/beam/beam_catches.h
index ccf33d5e86..0769d47263 100644
--- a/erts/emulator/beam/beam_catches.h
+++ b/erts/emulator/beam/beam_catches.h
@@ -23,9 +23,9 @@
 #define BEAM_CATCHES_NIL	(-1)
 
 void beam_catches_init(void);
-unsigned beam_catches_cons(Eterm* cp, unsigned cdr);
-Eterm *beam_catches_car(unsigned i);
-void beam_catches_delmod(unsigned head, Eterm* code, unsigned code_bytes);
+unsigned beam_catches_cons(UWord* cp, unsigned cdr);
+UWord *beam_catches_car(unsigned i);
+void beam_catches_delmod(unsigned head, UWord* code, unsigned code_bytes);
 
 #define catch_pc(x)	beam_catches_car(catch_val((x)))
 
diff --git a/erts/emulator/beam/beam_debug.c b/erts/emulator/beam/beam_debug.c
index 4242a4161e..b381d4e546 100644
--- a/erts/emulator/beam/beam_debug.c
+++ b/erts/emulator/beam/beam_debug.c
@@ -38,7 +38,7 @@
 #include "beam_bp.h"
 #include "erl_binary.h"
 
-#ifdef ARCH_64
+#ifdef ARCH_64 /* XXX:PaN Halfword? */
 # define HEXF "%016bpX"
 #else
 # define HEXF "%08bpX"
diff --git a/erts/emulator/beam/beam_emu.c b/erts/emulator/beam/beam_emu.c
index 66c6eb7819..d542c56e28 100644
--- a/erts/emulator/beam/beam_emu.c
+++ b/erts/emulator/beam/beam_emu.c
@@ -117,6 +117,7 @@ do {                                     \
 #endif
 
 #define GET_BIF_ADDRESS(p) ((BifFunction) (((Export *) p)->code[4]))
+#define TermWords(t) (((t) / (sizeof(UWord)/sizeof(Eterm))) + !!((t) % (sizeof(UWord)/sizeof(Eterm))))
 
 
 /*
@@ -138,8 +139,8 @@ do {                                     \
 #define VALID_INSTR(IP) (0 <= (int)(IP) && ((int)(IP) < (NUMBER_OF_OPCODES*2+10)))
 #else
 #define VALID_INSTR(IP) \
-   ((Sint)LabelAddr(emulator_loop) <= (Sint)(IP) && \
-    (Sint)(IP) < (Sint)LabelAddr(end_emulator_loop))
+   ((SWord)LabelAddr(emulator_loop) <= (SWord)(IP) && \
+    (SWord)(IP) < (SWord)LabelAddr(end_emulator_loop))
 #endif /* NO_JUMP_TABLE */
 
 #define SET_CP(p, ip)           \
@@ -181,11 +182,11 @@ do {                                     \
  
 #define StoreBifResult(Dst, Result)                          \
   do {                                                       \
-    Eterm* stb_next;                                         \
+    UWord* stb_next;                                         \
     Eterm stb_reg;                                           \
     stb_reg = Arg(Dst);                                      \
     I += (Dst) + 2;                                          \
-    stb_next = (Eterm *) *I;                                 \
+    stb_next = (UWord *) *I;                                 \
     CHECK_TERM(Result);                                      \
     switch (beam_reg_tag(stb_reg)) {                         \
     case R_REG_DEF:                                          \
@@ -205,7 +206,7 @@ do {                                     \
       c_p->cp = 0;				\
    } while(0)
 
-#define RESTORE_CP(X)		SET_CP(c_p, cp_val(*(X)))
+#define RESTORE_CP(X)		SET_CP(c_p, (UWord *) cp_val(*(X)))
 
 #define ISCATCHEND(instr) ((Eterm *) *(instr) == OpCode(catch_end_y))
 
@@ -213,13 +214,13 @@ do {                                     \
  * Special Beam instructions.
  */
 
-Eterm beam_apply[2];
-Eterm beam_exit[1];
-Eterm beam_continue_exit[1];
+UWord beam_apply[2];
+UWord beam_exit[1];
+UWord beam_continue_exit[1];
 
-Eterm* em_call_error_handler;
-Eterm* em_apply_bif;
-Eterm* em_call_traced_function;
+UWord* em_call_error_handler;
+UWord* em_apply_bif;
+UWord* em_call_traced_function;
 
 
 /* NOTE These should be the only variables containing trace instructions.
@@ -227,9 +228,9 @@ Eterm* em_call_traced_function;
 **      for the refering variable (one of these), and rouge references
 **      will most likely cause chaos.
 */
-Eterm beam_return_to_trace[1]; /* OpCode(i_return_to_trace) */
-Eterm beam_return_trace[1];    /* OpCode(i_return_trace) */
-Eterm beam_exception_trace[1]; /* UGLY also OpCode(i_return_trace) */
+UWord beam_return_to_trace[1]; /* OpCode(i_return_to_trace) */
+UWord beam_return_trace[1];    /* OpCode(i_return_trace) */
+UWord beam_exception_trace[1]; /* UGLY also OpCode(i_return_trace) */
 
 /*
  * All Beam instructions in numerical order.
@@ -522,8 +523,8 @@ extern int count_instructions;
 
 #define DispatchMacro()				\
   do {						\
-     Eterm* dis_next;				\
-     dis_next = (Eterm *) *I;			\
+     UWord* dis_next;				\
+     dis_next = (UWord *) *I;			\
      CHECK_ARGS(I);				\
      if (FCALLS > 0 || FCALLS > neg_o_reds) {	\
         FCALLS--;				\
@@ -535,8 +536,8 @@ extern int count_instructions;
 
 #define DispatchMacroFun()			\
   do {						\
-     Eterm* dis_next;				\
-     dis_next = (Eterm *) *I;			\
+     UWord* dis_next;				\
+     dis_next = (UWord *) *I;			\
      CHECK_ARGS(I);				\
      if (FCALLS > 0 || FCALLS > neg_o_reds) {	\
         FCALLS--;				\
@@ -590,7 +591,7 @@ extern int count_instructions;
     ASSERT(VALID_INSTR(*I));   \
     Goto(*I)
 
-#define PreFetch(N, Dst) do { Dst = (Eterm *) *(I + N + 1); } while (0)
+#define PreFetch(N, Dst) do { Dst = (UWord *) *(I + N + 1); } while (0)
 #define NextPF(N, Dst)         \
     I += N + 1;                \
     ASSERT(VALID_INSTR(Dst));  \
@@ -644,7 +645,7 @@ extern int count_instructions;
 #define DeallocateReturn(Deallocate)       \
   do {                                     \
     int words_to_pop = (Deallocate);       \
-    SET_I(cp_val(*E));                     \
+    SET_I((UWord *) cp_val(*E));                     \
     E = ADD_BYTE_OFFSET(E, words_to_pop);  \
     CHECK_TERM(r(0));                      \
     Goto(*I);                              \
@@ -657,19 +658,19 @@ extern int count_instructions;
 #define MoveCall(Src, Dest, CallDest, Size)	\
     (Dest) = (Src);				\
     SET_CP(c_p, I+Size+1);			\
-    SET_I((Eterm *) CallDest);			\
+    SET_I((UWord *) CallDest);			\
     Dispatch();
 
 #define MoveCallLast(Src, Dest, CallDest, Deallocate)	\
     (Dest) = (Src);					\
     RESTORE_CP(E);					\
     E = ADD_BYTE_OFFSET(E, (Deallocate));		\
-    SET_I((Eterm *) CallDest);				\
+    SET_I((UWord *) CallDest);				\
     Dispatch();
 
 #define MoveCallOnly(Src, Dest, CallDest)	\
     (Dest) = (Src);				\
-    SET_I((Eterm *) CallDest);			\
+    SET_I((UWord *) CallDest);			\
     Dispatch();
 
 #define GetList(Src, H, T) do {			\
@@ -677,47 +678,48 @@ extern int count_instructions;
    H = CAR(tmp_ptr);				\
    T = CDR(tmp_ptr); } while (0)
 
-#define GetTupleElement(Src, Element, Dest)				  \
-  do {									  \
-    tmp_arg1 = (Eterm) (((unsigned char *) tuple_val(Src)) + (Element));  \
-    (Dest) = (*(Eterm *)tmp_arg1);					  \
+#define GetTupleElement(Src, Element, Dest)				\
+  do {									\
+    tmp_arg1 = (Eterm) COMPRESS_POINTER(((unsigned char *) tuple_val(Src)) + 	\
+				(Element));				\
+    (Dest) = (*(Eterm *) EXPAND_POINTER(tmp_arg1));			\
   } while (0)
 
-#define ExtractNextElement(Dest)				\
-    tmp_arg1 += sizeof(Eterm);					\
-    (Dest) = (* (Eterm *) (((unsigned char *) tmp_arg1)))
+#define ExtractNextElement(Dest)					  \
+    tmp_arg1 += sizeof(Eterm);						  \
+    (Dest) = (* (Eterm *) (((unsigned char *) EXPAND_POINTER(tmp_arg1))))
 
-#define ExtractNextElement2(Dest)		\
-  do {						\
-    Eterm* ene_dstp = &(Dest);			\
-    ene_dstp[0] = ((Eterm *) tmp_arg1)[1];	\
-    ene_dstp[1] = ((Eterm *) tmp_arg1)[2];	\
-    tmp_arg1 += sizeof(Eterm) + sizeof(Eterm);	\
+#define ExtractNextElement2(Dest)				\
+  do {								\
+    Eterm* ene_dstp = &(Dest);					\
+    ene_dstp[0] = ((Eterm *) EXPAND_POINTER(tmp_arg1))[1];	\
+    ene_dstp[1] = ((Eterm *) EXPAND_POINTER(tmp_arg1))[2];	\
+    tmp_arg1 += sizeof(Eterm) + sizeof(Eterm);			\
   } while (0)
 
 #define ExtractNextElement3(Dest)		\
   do {						\
     Eterm* ene_dstp = &(Dest);			\
-    ene_dstp[0] = ((Eterm *) tmp_arg1)[1];	\
-    ene_dstp[1] = ((Eterm *) tmp_arg1)[2];	\
-    ene_dstp[2] = ((Eterm *) tmp_arg1)[3];	\
+    ene_dstp[0] = ((Eterm *) EXPAND_POINTER(tmp_arg1))[1];	\
+    ene_dstp[1] = ((Eterm *) EXPAND_POINTER(tmp_arg1))[2];	\
+    ene_dstp[2] = ((Eterm *) EXPAND_POINTER(tmp_arg1))[3];	\
     tmp_arg1 += 3*sizeof(Eterm);		\
   } while (0)
 
 #define ExtractNextElement4(Dest)		\
   do {						\
     Eterm* ene_dstp = &(Dest);			\
-    ene_dstp[0] = ((Eterm *) tmp_arg1)[1];	\
-    ene_dstp[1] = ((Eterm *) tmp_arg1)[2];	\
-    ene_dstp[2] = ((Eterm *) tmp_arg1)[3];	\
-    ene_dstp[3] = ((Eterm *) tmp_arg1)[4];	\
+    ene_dstp[0] = ((Eterm *) EXPAND_POINTER(tmp_arg1))[1];	\
+    ene_dstp[1] = ((Eterm *) EXPAND_POINTER(tmp_arg1))[2];	\
+    ene_dstp[2] = ((Eterm *) EXPAND_POINTER(tmp_arg1))[3];	\
+    ene_dstp[3] = ((Eterm *) EXPAND_POINTER(tmp_arg1))[4];	\
     tmp_arg1 += 4*sizeof(Eterm);		\
   } while (0)
 
 #define ExtractElement(Element, Dest)		\
   do {						\
      tmp_arg1 += (Element);			\
-     (Dest) = (* (Eterm *) tmp_arg1);		\
+     (Dest) = (* (Eterm *) EXPAND_POINTER(tmp_arg1));		\
   } while (0)
 
 #define PutTuple(Arity, Src, Dest)		\
@@ -759,8 +761,13 @@ extern int count_instructions;
 
 #define IsTuple(X, Action) if (is_not_tuple(X)) Action
 
-#define IsArity(Pointer, Arity, Fail) \
-    if (*(Eterm *)(tmp_arg1 = (Eterm)tuple_val(Pointer)) != (Arity)) { Fail; }
+#define IsArity(Pointer, Arity, Fail)					  \
+    if (*(Eterm *)							  \
+	EXPAND_POINTER(tmp_arg1 = (Eterm) 				  \
+		       COMPRESS_POINTER(tuple_val(Pointer))) != (Arity))  \
+    { 									  \
+        Fail; 								  \
+    }
 
 #define IsFunction(X, Action)			\
   do {						\
@@ -776,11 +783,14 @@ extern int count_instructions;
      }						\
   } while (0)
 
-#define IsTupleOfArity(Src, Arity, Fail) \
-  do { \
-    if (is_not_tuple(Src) || *(Eterm *)(tmp_arg1 = (Eterm) tuple_val(Src)) != Arity) { \
-        Fail; \
-    } \
+#define IsTupleOfArity(Src, Arity, Fail)				      \
+  do {									      \
+    if (is_not_tuple(Src) || 						      \
+	*(Eterm *)							      \
+	EXPAND_POINTER(tmp_arg1 = 					      \
+		       (Eterm) COMPRESS_POINTER(tuple_val(Src))) != Arity) { \
+        Fail;								      \
+    }									      \
   } while (0)
 
 #define IsBoolean(X, Fail) if ((X) != am_true && (X) != am_false) { Fail; }
@@ -791,7 +801,7 @@ extern int count_instructions;
 #define IsBitstring(Src, Fail) \
   if (is_not_binary(Src)) { Fail; }
 
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 #define BsSafeMul(A, B, Fail, Target)		\
    do { Uint64 _res = (A) * (B);		\
       if (_res / B != A) { Fail; }		\
@@ -974,23 +984,23 @@ extern int count_instructions;
 #define IsRef(Src, Fail) if (is_not_ref(Src)) { Fail; }
 
 static BifFunction translate_gc_bif(void* gcf);
-static Eterm* handle_error(Process* c_p, Eterm* pc, Eterm* reg, BifFunction bf);
-static Eterm* next_catch(Process* c_p, Eterm *reg);
+static UWord* handle_error(Process* c_p, UWord* pc, Eterm* reg, BifFunction bf);
+static UWord* next_catch(Process* c_p, Eterm *reg);
 static void terminate_proc(Process* c_p, Eterm Value);
 static Eterm add_stacktrace(Process* c_p, Eterm Value, Eterm exc);
-static void save_stacktrace(Process* c_p, Eterm* pc, Eterm* reg,
+static void save_stacktrace(Process* c_p, UWord* pc, Eterm* reg,
 			     BifFunction bf, Eterm args);
 static struct StackTrace * get_trace_from_exc(Eterm exc);
 static Eterm make_arglist(Process* c_p, Eterm* reg, int a);
-static Eterm call_error_handler(Process* p, Eterm* ip, Eterm* reg);
-static Eterm call_breakpoint_handler(Process* p, Eterm* fi, Eterm* reg);
-static Uint* fixed_apply(Process* p, Eterm* reg, Uint arity);
-static Eterm* apply(Process* p, Eterm module, Eterm function,
+static Eterm call_error_handler(Process* p, UWord* ip, Eterm* reg);
+static Eterm call_breakpoint_handler(Process* p, UWord* fi, Eterm* reg);
+static UWord* fixed_apply(Process* p, Eterm* reg, Uint arity);
+static UWord* apply(Process* p, Eterm module, Eterm function,
 		     Eterm args, Eterm* reg);
 static int hibernate(Process* c_p, Eterm module, Eterm function,
 		     Eterm args, Eterm* reg);
-static Eterm* call_fun(Process* p, int arity, Eterm* reg, Eterm args);
-static Eterm* apply_fun(Process* p, Eterm fun, Eterm args, Eterm* reg);
+static UWord* call_fun(Process* p, int arity, Eterm* reg, Eterm args);
+static UWord* apply_fun(Process* p, Eterm fun, Eterm args, Eterm* reg);
 static Eterm new_fun(Process* p, Eterm* reg, ErlFunEntry* fe, int num_free);
 
 #if defined(_OSE_) || defined(VXWORKS)
@@ -1078,7 +1088,7 @@ void process_main(void)
     /*
      * Pointer to next threaded instruction.
      */
-    register Eterm *I REG_I = NULL;
+    register UWord *I REG_I = NULL;
 
     /* Number of reductions left.  This function
      * returns to the scheduler when FCALLS reaches zero.
@@ -1097,6 +1107,7 @@ void process_main(void)
 #endif
 
 #ifndef ERTS_SMP
+#if !HALFWORD_HEAP
     static Eterm save_reg[ERTS_X_REGS_ALLOCATED];
     /* X registers -- not used directly, but
      * through 'reg', because using it directly
@@ -1104,7 +1115,7 @@ void process_main(void)
      * while using it through reg needs only
      * one.
      */
-
+#endif
     /*
      * Floating point registers.
      */
@@ -1150,7 +1161,12 @@ void process_main(void)
 	goto init_emulator;
     }
 #ifndef ERTS_SMP
+#if !HALFWORD_HEAP
     reg = save_reg;	/* XXX: probably wastes a register on x86 */
+#else
+    /* Registers need to be heap allocated (correct memory range) for tracing to work */
+    reg = erts_alloc(ERTS_ALC_T_BEAM_REGISTER, ERTS_X_REGS_ALLOCATED * sizeof(Eterm));
+#endif
 #endif
     c_p = NULL;
     reds_used = 0;
@@ -1178,7 +1194,7 @@ void process_main(void)
     {
 	int reds;
 	Eterm* argp;
-	Eterm* next;
+	UWord *next;
 	int i;
 
 	argp = c_p->arg_reg;
@@ -1205,7 +1221,7 @@ void process_main(void)
 	    FCALLS = REDS_IN(c_p) = reds;
 	}
 
-	next = (Eterm *) *I;
+	next = (UWord *) *I;
 	r(0) = c_p->arg_reg[0];
 #ifdef HARDDEBUG
 	if (c_p->arity > 0) {
@@ -1297,7 +1313,7 @@ void process_main(void)
  }
  /* FALL THROUGH */
  OpCase(i_call_only_f): {
-     SET_I((Eterm *) Arg(0));
+     SET_I((UWord *) Arg(0));
      Dispatch();
  }
 
@@ -1308,7 +1324,7 @@ void process_main(void)
  OpCase(i_call_last_fP): {
      RESTORE_CP(E);
      E = ADD_BYTE_OFFSET(E, Arg(1));
-     SET_I((Eterm *) Arg(0));
+     SET_I((UWord *) Arg(0));
      Dispatch();
  }
 
@@ -1319,7 +1335,7 @@ void process_main(void)
  /* FALL THROUGH */
  OpCase(i_call_f): {
      SET_CP(c_p, I+2);
-     SET_I((Eterm *) Arg(0));
+     SET_I((UWord *) Arg(0));
      Dispatch();
  }
 
@@ -1355,7 +1371,7 @@ void process_main(void)
     Dispatchx();
 
  OpCase(init_y): {
-     Eterm* next;
+     UWord *next;
 
      PreFetch(1, next);
      make_blank(yb(Arg(0)));
@@ -1363,7 +1379,7 @@ void process_main(void)
  }
 
  OpCase(i_trim_I): {
-     Eterm* next;
+     UWord *next;
      Uint words;
      Uint cp;
 
@@ -1389,7 +1405,7 @@ void process_main(void)
  }
 
  OpCase(test_heap_1_put_list_Iy): {
-     Eterm* next;
+     UWord *next;
 
      PreFetch(2, next);
      TestHeap(Arg(0), 1);
@@ -1420,7 +1436,7 @@ void process_main(void)
      */
 
  OpCase(send): {
-     Eterm* next;
+     UWord *next;
      Eterm result;
 
      PRE_BIF_SWAPOUT(c_p);
@@ -1435,7 +1451,7 @@ void process_main(void)
 	 NextPF(0, next);
      } else if (c_p->freason == TRAP) {
 	 SET_CP(c_p, I+1);
-	 SET_I((Eterm *) c_p->def_arg_reg[3]);
+	 SET_I(*((UWord **) (UWord) ((c_p)->def_arg_reg + 3)));
 	 SWAPIN;
 	 r(0) = c_p->def_arg_reg[0];
 	 x(1) = c_p->def_arg_reg[1];
@@ -1561,7 +1577,7 @@ void process_main(void)
      */
  OpCase(i_loop_rec_fr):
  {
-     Eterm* next;
+     UWord *next;
      ErlMessage* msgp;
 
  loop_rec__:
@@ -1585,7 +1601,7 @@ void process_main(void)
 	     erts_smp_proc_unlock(c_p, ERTS_PROC_LOCKS_MSG_RECEIVE);
 	 else {
 #endif
-	     SET_I((Eterm *) Arg(0));
+	     SET_I((UWord *) Arg(0));
 	     Goto(*I);		/* Jump to a wait or wait_timeout instruction */
 #ifdef ERTS_SMP
 	 }
@@ -1621,7 +1637,7 @@ void process_main(void)
   * Remove a (matched) message from the message queue.
   */
  OpCase(remove_message): {
-     Eterm* next;
+     UWord *next;
      ErlMessage* msgp;
 
      PROCESS_MAIN_CHK_LOCKS(c_p);
@@ -1666,7 +1682,7 @@ void process_main(void)
      * message didn't match), then jump to the loop_rec instruction.
      */
  OpCase(loop_rec_end_f): {
-     SET_I((Eterm *) Arg(0));
+     SET_I((UWord *) Arg(0));
      SAVE_MESSAGE(c_p);
      goto loop_rec__;
  }
@@ -1696,12 +1712,12 @@ void process_main(void)
      }
      GetArg1(1, timeout_value);
      if (timeout_value != make_small(0)) {
-#if !defined(ARCH_64)
+#if !defined(ARCH_64) || HALFWORD_HEAP
 	 Uint time_val;
 #endif
 
 	 if (is_small(timeout_value) && signed_val(timeout_value) > 0 &&
-#if defined(ARCH_64)
+#if defined(ARCH_64) && !HALFWORD_HEAP
 	     ((unsigned_val(timeout_value) >> 32) == 0)
 #else
 	     1
@@ -1712,14 +1728,16 @@ void process_main(void)
 	      * c_p->def_arg_reg[0].  Note that it is safe to use this
 	      * location because there are no living x registers in
 	      * a receive statement.
+	      * Note that for the halfword emulator, the two first elements
+	      * of the array are used.
 	      */
-	     c_p->def_arg_reg[0] = (Eterm) (I+3);
+	     *((UWord **) (UWord) c_p->def_arg_reg) = 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)
+#if !defined(ARCH_64) || HALFWORD_HEAP
 	 } else if (term_to_Uint(timeout_value, &time_val)) {
-	     c_p->def_arg_reg[0] = (Eterm) (I+3);
+	     *((UWord **) (UWord) c_p->def_arg_reg) = I+3;
 	     set_timer(c_p, time_val);
 #endif
 	 } else {		/* Wrong time */
@@ -1748,7 +1766,7 @@ void process_main(void)
 
 	 wait2: {
 	     ASSERT(!ERTS_PROC_IS_EXITING(c_p));
-	     c_p->i = (Eterm *) Arg(0); /* L1 */
+	     c_p->i = (UWord *) Arg(0); /* L1 */
 	     SWAPOUT;
 	     c_p->arity = 0;
 	     c_p->status = P_WAITING;
@@ -1776,7 +1794,7 @@ 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) {
-	 c_p->def_arg_reg[0] = (Eterm) (I+3);
+	 *((UWord **) (UWord) c_p->def_arg_reg) = I+3;
 	 set_timer(c_p, Arg(1));
      }
      goto wait2;
@@ -1791,7 +1809,7 @@ void process_main(void)
  }
 
  OpCase(timeout): {
-     Eterm* next;
+     UWord *next;
 
      PreFetch(0, next);
      if (IS_TRACED_FL(c_p, F_TRACE_RECEIVE)) {
@@ -1811,8 +1829,8 @@ void process_main(void)
  do_binary_search:
  {
      struct Pairs {
-	 Eterm val;
-	 Eterm* addr;
+	 UWord val;
+	 UWord* addr;
      };
      struct Pairs* low;
      struct Pairs* high;
@@ -1852,7 +1870,7 @@ void process_main(void)
 	     Goto(*I);
 	 }
      }
-     SET_I((Eterm *) Arg(1));
+     SET_I((UWord *) Arg(1));
      Goto(*I);
  }
 
@@ -1864,11 +1882,11 @@ void process_main(void)
      if (is_small(index)) {
 	 index = signed_val(index);
 	 if (index < Arg(2)) {
-	     SET_I((Eterm *) (&Arg(3))[index]);
+	     SET_I((UWord *) (&Arg(3))[index]);
 	     Goto(*I);
 	 }
      }
-     SET_I((Eterm *) Arg(1));
+     SET_I((UWord *) Arg(1));
      Goto(*I);
  }
 
@@ -1880,11 +1898,11 @@ void process_main(void)
      if (is_small(index)) {
 	 index = (Uint) (signed_val(index) - Arg(3));
 	 if (index < Arg(2)) {
-	     SET_I((Eterm *) (&Arg(4))[index]);
+	     SET_I((UWord *) (&Arg(4))[index]);
 	     Goto(*I);
 	 }
      }
-     SET_I((Eterm *) Arg(1));
+     SET_I((UWord *) Arg(1));
      Goto(*I);
  }
 
@@ -1921,7 +1939,7 @@ void process_main(void)
 	if (is_value(result)) {
 	    StoreBifResult(3, result);
 	}
-	SET_I((Eterm *) Arg(0));
+	SET_I((UWord *) Arg(0));
 	Goto(*I);
     }
 
@@ -1982,7 +2000,7 @@ void process_main(void)
 	    StoreBifResult(4, result);
 	}
 	if (Arg(0) != 0) {
-	    SET_I((Eterm *) Arg(0));
+	    SET_I((UWord *) Arg(0));
 	    Goto(*I);
 	}
 	reg[0] = arg;
@@ -2010,7 +2028,7 @@ void process_main(void)
 	if (is_value(result)) {
 	    StoreBifResult(2, result);
 	}
-	SET_I((Eterm *) Arg(0));
+	SET_I((UWord *) Arg(0));
 	Goto(*I);
     }
 
@@ -2046,7 +2064,7 @@ void process_main(void)
      */
  OpCase(call_bif0_e):
     {
-	Eterm (*bf)(Process*, Uint*) = GET_BIF_ADDRESS(Arg(0));
+	Eterm (*bf)(Process*, UWord*) = GET_BIF_ADDRESS(Arg(0));
 
 	PRE_BIF_SWAPOUT(c_p);
 	c_p->fcalls = FCALLS - 1;
@@ -2079,9 +2097,9 @@ void process_main(void)
 
  OpCase(call_bif1_e):
     {
-	Eterm (*bf)(Process*, Eterm, Uint*) = GET_BIF_ADDRESS(Arg(0));
+	Eterm (*bf)(Process*, Eterm, UWord*) = GET_BIF_ADDRESS(Arg(0));
 	Eterm result;
-	Eterm* next;
+	UWord *next;
 
 	c_p->fcalls = FCALLS - 1;
 	if (FCALLS <= 0) {
@@ -2114,9 +2132,9 @@ void process_main(void)
 
  OpCase(call_bif2_e):
     {
-	Eterm (*bf)(Process*, Eterm, Eterm, Uint*) = GET_BIF_ADDRESS(Arg(0));
+	Eterm (*bf)(Process*, Eterm, Eterm, UWord*) = GET_BIF_ADDRESS(Arg(0));
 	Eterm result;
-	Eterm* next;
+	UWord *next;
 
 	PRE_BIF_SWAPOUT(c_p);
 	c_p->fcalls = FCALLS - 1;
@@ -2151,9 +2169,9 @@ void process_main(void)
 
  OpCase(call_bif3_e):
     {
-	Eterm (*bf)(Process*, Eterm, Eterm, Eterm, Uint*) = GET_BIF_ADDRESS(Arg(0));
+	Eterm (*bf)(Process*, Eterm, Eterm, Eterm, UWord*) = GET_BIF_ADDRESS(Arg(0));
 	Eterm result;
-	Eterm* next;
+	UWord *next;
 
 	PRE_BIF_SWAPOUT(c_p);
 	c_p->fcalls = FCALLS - 1;
@@ -2174,7 +2192,7 @@ void process_main(void)
 	} else if (c_p->freason == TRAP) {
 	call_bif_trap3:
 	    SET_CP(c_p, I+2);
-	    SET_I((Eterm *)c_p->def_arg_reg[3]);
+	    SET_I(*((UWord **) (UWord) ((c_p)->def_arg_reg + 3)));
 	    SWAPIN;
 	    r(0) = c_p->def_arg_reg[0];
 	    x(1) = c_p->def_arg_reg[1];
@@ -2282,7 +2300,7 @@ void process_main(void)
  lb_Cl_error: {
      if (Arg(0) != 0) {
 	 OpCase(jump_f): {
-	     SET_I((Eterm *) Arg(0));
+	     SET_I((UWord *) Arg(0));
 	     Goto(*I);
 	 }
      }
@@ -2474,7 +2492,7 @@ void process_main(void)
     goto lb_Cl_error;
 
  OpCase(i_apply): {
-     Eterm* next;
+     UWord *next;
      SWAPOUT;
      next = apply(c_p, r(0), x(1), x(2), reg);
      SWAPIN;
@@ -2489,13 +2507,13 @@ void process_main(void)
  }
 
  OpCase(i_apply_last_P): {
-     Eterm* next;
+     UWord *next;
      SWAPOUT;
      next = apply(c_p, r(0), x(1), x(2), reg);
      SWAPIN;
      if (next != NULL) {
 	 r(0) = reg[0];
-	 SET_CP(c_p, (Eterm *) E[0]);
+	 SET_CP(c_p, (UWord *) EXPAND_POINTER(E[0]));
 	 E = ADD_BYTE_OFFSET(E, Arg(0));
 	 SET_I(next);
 	 Dispatch();
@@ -2505,7 +2523,7 @@ void process_main(void)
  }
 
  OpCase(i_apply_only): {
-     Eterm* next;
+     UWord *next;
      SWAPOUT;
      next = apply(c_p, r(0), x(1), x(2), reg);
      SWAPIN;
@@ -2519,7 +2537,7 @@ void process_main(void)
  }
 
  OpCase(apply_I): {
-     Eterm* next;
+     UWord *next;
 
      reg[0] = r(0);
      SWAPOUT;
@@ -2536,7 +2554,7 @@ void process_main(void)
  }
 
  OpCase(apply_last_IP): {
-     Eterm* next;
+     UWord *next;
 
      reg[0] = r(0);
      SWAPOUT;
@@ -2544,7 +2562,7 @@ void process_main(void)
      SWAPIN;
      if (next != NULL) {
 	 r(0) = reg[0];
-	 SET_CP(c_p, (Eterm *) E[0]);
+	 SET_CP(c_p, (UWord *) EXPAND_POINTER(E[0]));
 	 E = ADD_BYTE_OFFSET(E, Arg(1));
 	 SET_I(next);
 	 Dispatch();
@@ -2554,7 +2572,7 @@ void process_main(void)
  }
 
  OpCase(i_apply_fun): {
-     Eterm* next;
+     UWord *next;
 
      SWAPOUT;
      next = apply_fun(c_p, r(0), x(1), reg);
@@ -2569,14 +2587,14 @@ void process_main(void)
  }
 
  OpCase(i_apply_fun_last_P): {
-     Eterm* next;
+     UWord *next;
 
      SWAPOUT;
      next = apply_fun(c_p, r(0), x(1), reg);
      SWAPIN;
      if (next != NULL) {
 	 r(0) = reg[0];
-	 SET_CP(c_p, (Eterm *) E[0]);
+	 SET_CP(c_p, (UWord *) EXPAND_POINTER(E[0]));
 	 E = ADD_BYTE_OFFSET(E, Arg(0));
 	 SET_I(next);
 	 Dispatchfun();
@@ -2585,7 +2603,7 @@ void process_main(void)
  }
 
  OpCase(i_apply_fun_only): {
-     Eterm* next;
+     UWord *next;
 
      SWAPOUT;
      next = apply_fun(c_p, r(0), x(1), reg);
@@ -2599,7 +2617,7 @@ void process_main(void)
  }
 
  OpCase(i_call_fun_I): {
-     Eterm* next;
+     UWord *next;
 
      SWAPOUT;
      reg[0] = r(0);
@@ -2615,7 +2633,7 @@ void process_main(void)
  }
 
  OpCase(i_call_fun_last_IP): {
-     Eterm* next;
+     UWord *next;
 
      SWAPOUT;
      reg[0] = r(0);
@@ -2623,7 +2641,7 @@ void process_main(void)
      SWAPIN;
      if (next != NULL) {
 	r(0) = reg[0];
-	SET_CP(c_p, (Eterm *) E[0]);
+	SET_CP(c_p, (UWord *) EXPAND_POINTER(E[0]));
 	E = ADD_BYTE_OFFSET(E, Arg(1));
 	SET_I(next);
 	Dispatchfun();
@@ -2728,7 +2746,7 @@ void process_main(void)
 	 tmp_arg1 = *tuple_val(tmp_arg1);
 	 goto do_binary_search;
      }
-     SET_I((Eterm *) Arg(1));
+     SET_I((UWord *) Arg(1));
      Goto(*I);
  }     
 
@@ -2754,13 +2772,15 @@ void process_main(void)
  	    given = big_val(tmp_arg1);
 	    given_arity = given[0];
 	    given_size = thing_arityval(given_arity);
-	    bigp = &Arg(2);
+	    bigp = (Eterm *) &Arg(2);
 	    while ((arity = bigp[0]) > given_arity) {
 		bigp += thing_arityval(arity) + 2;
 	    }
 	    while (bigp[0] == given_arity) {
 		if (memcmp(bigp+1, given+1, sizeof(Eterm)*given_size) == 0) {
-		    SET_I((Eterm *) bigp[given_size+1]);
+		    UWord *tmp =
+			((UWord *) (UWord) bigp) + TermWords(given_size + 1);
+		    SET_I((UWord *) *tmp);
 		    Goto(*I);
 		}
 		bigp += thing_arityval(arity) + 2;
@@ -2771,18 +2791,18 @@ void process_main(void)
 	 * Failed.
 	 */
 
-	SET_I((Eterm *) Arg(1));
+	SET_I((UWord *) Arg(1));
 	Goto(*I);
     }
 
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
  OpCase(i_select_float_sfI):
  {
      Uint f;
      int n;
      struct ValLabel {
 	 Uint f;
-	 Eterm* addr;
+	 UWord* addr;
      };
      struct ValLabel* ptr;
 
@@ -2810,7 +2830,7 @@ void process_main(void)
      struct ValLabel {
 	 Uint fpart1;
 	 Uint fpart2;
-	 Eterm* addr;
+	 UWord* addr;
      };
      struct ValLabel* ptr;
 
@@ -2828,7 +2848,7 @@ void process_main(void)
 	 }
 	 ptr++;
      }
-     SET_I((Eterm *) Arg(1));
+     SET_I((UWord *) Arg(1));
      Goto(*I);
  }
 #endif
@@ -2836,7 +2856,7 @@ void process_main(void)
  OpCase(set_tuple_element_sdP): {
      Eterm element;
      Eterm tuple;
-     Eterm* next;
+     UWord *next;
      Eterm* p;
      
      PreFetch(3, next);
@@ -3031,7 +3051,7 @@ void process_main(void)
 	switch (tmp_arg2) {
 	case 3:
 	    {
-		Eterm (*bf)(Process*, Eterm, Eterm, Eterm, Uint*) = vbf;
+		Eterm (*bf)(Process*, Eterm, Eterm, Eterm, UWord*) = vbf;
 		ASSERT(!ERTS_PROC_IS_EXITING(c_p));
 		tmp_arg1 = (*bf)(c_p, r(0), x(1), x(2), I);
 		ASSERT(!ERTS_PROC_IS_EXITING(c_p) || is_non_value(tmp_arg1));
@@ -3040,7 +3060,7 @@ void process_main(void)
 	    break;
 	case 2:
 	    {
-		Eterm (*bf)(Process*, Eterm, Eterm, Uint*) = vbf;
+		Eterm (*bf)(Process*, Eterm, Eterm, UWord*) = vbf;
 		ASSERT(!ERTS_PROC_IS_EXITING(c_p));
 		tmp_arg1 = (*bf)(c_p, r(0), x(1), I);
 		ASSERT(!ERTS_PROC_IS_EXITING(c_p) || is_non_value(tmp_arg1));
@@ -3049,7 +3069,7 @@ void process_main(void)
 	    break;
 	case 1:
 	    {
-		Eterm (*bf)(Process*, Eterm, Uint*) = vbf;
+		Eterm (*bf)(Process*, Eterm, UWord*) = vbf;
 		ASSERT(!ERTS_PROC_IS_EXITING(c_p));
 		tmp_arg1 = (*bf)(c_p, r(0), I);
 		ASSERT(!ERTS_PROC_IS_EXITING(c_p) || is_non_value(tmp_arg1));
@@ -3058,7 +3078,7 @@ void process_main(void)
 	    break;
 	case 0:
 	    {
-		Eterm (*bf)(Process*, Uint*) = vbf;
+		Eterm (*bf)(Process*, UWord*) = vbf;
 		ASSERT(!ERTS_PROC_IS_EXITING(c_p));
 		tmp_arg1 = (*bf)(c_p, I);
 		ASSERT(!ERTS_PROC_IS_EXITING(c_p) || is_non_value(tmp_arg1));
@@ -3082,7 +3102,7 @@ apply_bif_or_nif_epilogue:
 	    SET_I(c_p->cp);
 	    Goto(*I);
 	} else if (c_p->freason == TRAP) {
-	    SET_I((Eterm *)c_p->def_arg_reg[3]);
+	    SET_I(*((UWord **) (UWord) ((c_p)->def_arg_reg + 3)));
 	    r(0) = c_p->def_arg_reg[0];
 	    x(1) = c_p->def_arg_reg[1];
 	    x(2) = c_p->def_arg_reg[2];
@@ -3540,7 +3560,7 @@ apply_bif_or_nif_epilogue:
 
  OpCase(bs_put_string_II):
     {
-	Eterm* next;
+	UWord *next;
 	PreFetch(2, next);
 	erts_new_bs_put_string(ERL_BITS_ARGS_2((byte *) Arg(1), Arg(0)));
 	NextPF(2, next);
@@ -3711,7 +3731,7 @@ apply_bif_or_nif_epilogue:
 
  {
      Eterm header;
-     Eterm* next;
+     UWord *next;
      Uint slots;
 
      OpCase(i_bs_start_match2_rfIId): {
@@ -3777,7 +3797,7 @@ apply_bif_or_nif_epilogue:
  }
 
  OpCase(bs_test_zero_tail2_fr): {
-     Eterm* next;
+     UWord *next;
      ErlBinMatchBuffer *_mb;
      
      PreFetch(1, next);
@@ -3789,7 +3809,7 @@ apply_bif_or_nif_epilogue:
  }
 
  OpCase(bs_test_zero_tail2_fx): {
-     Eterm* next;
+     UWord *next;
      ErlBinMatchBuffer *_mb;
      
      PreFetch(2, next);
@@ -3801,7 +3821,7 @@ apply_bif_or_nif_epilogue:
  }
 
  OpCase(bs_test_tail_imm2_frI): {
-     Eterm* next;
+     UWord *next;
      ErlBinMatchBuffer *_mb;
      PreFetch(2, next);
      _mb = ms_matchbuffer(r(0));
@@ -3811,7 +3831,7 @@ apply_bif_or_nif_epilogue:
      NextPF(2, next);
  }
  OpCase(bs_test_tail_imm2_fxI): {
-     Eterm* next;
+     UWord *next;
      ErlBinMatchBuffer *_mb;
      PreFetch(3, next);
      _mb = ms_matchbuffer(xb(Arg(1)));
@@ -3822,7 +3842,7 @@ apply_bif_or_nif_epilogue:
  }
 
  OpCase(bs_test_unit_frI): {
-     Eterm* next;
+     UWord *next;
      ErlBinMatchBuffer *_mb;
      PreFetch(2, next);
      _mb = ms_matchbuffer(r(0));
@@ -3832,7 +3852,7 @@ apply_bif_or_nif_epilogue:
      NextPF(2, next);
  }
  OpCase(bs_test_unit_fxI): {
-     Eterm* next;
+     UWord *next;
      ErlBinMatchBuffer *_mb;
      PreFetch(3, next);
      _mb = ms_matchbuffer(xb(Arg(1)));
@@ -3843,7 +3863,7 @@ apply_bif_or_nif_epilogue:
  }
 
  OpCase(bs_test_unit8_fr): {
-     Eterm* next;
+     UWord *next;
      ErlBinMatchBuffer *_mb;
      PreFetch(1, next);
      _mb = ms_matchbuffer(r(0));
@@ -3853,7 +3873,7 @@ apply_bif_or_nif_epilogue:
      NextPF(1, next);
  }
  OpCase(bs_test_unit8_fx): {
-     Eterm* next;
+     UWord *next;
      ErlBinMatchBuffer *_mb;
      PreFetch(2, next);
      _mb = ms_matchbuffer(xb(Arg(1)));
@@ -3937,11 +3957,11 @@ apply_bif_or_nif_epilogue:
 	 _integer = get_int32(_mb->base + _mb->offset/8);
      }
      _mb->offset += 32;
-#ifndef ARCH_64
+#if !defined(ARCH_64) || HALFWORD_HEAP
      if (IS_USMALL(0, _integer)) {
 #endif
 	 _result = make_small(_integer);
-#ifndef ARCH_64
+#if !defined(ARCH_64) || HALFWORD_HEAP
      } else {
 	 TestHeap(BIG_UINT_HEAP_SIZE, Arg(1));
 	 _result = uint_to_big((Uint) _integer, HTOP);
@@ -4178,7 +4198,7 @@ apply_bif_or_nif_epilogue:
 
  do_bs_match_string:
      {
-	 Eterm* next;
+	 UWord *next;
 	 byte* bytes;
 	 Uint bits;
 	 ErlBinMatchBuffer* mb;
@@ -4205,7 +4225,7 @@ apply_bif_or_nif_epilogue:
  }
 
  OpCase(i_bs_save2_rI): {
-     Eterm* next;
+     UWord *next;
      ErlBinMatchState *_ms;
      PreFetch(1, next);
      _ms = (ErlBinMatchState*) boxed_val((Eterm) r(0));
@@ -4213,7 +4233,7 @@ apply_bif_or_nif_epilogue:
      NextPF(1, next);
  }
  OpCase(i_bs_save2_xI): {
-     Eterm* next;
+     UWord *next;
      ErlBinMatchState *_ms;
      PreFetch(2, next);
      _ms = (ErlBinMatchState*) boxed_val((Eterm) xb(Arg(0)));
@@ -4222,7 +4242,7 @@ apply_bif_or_nif_epilogue:
  }
 
  OpCase(i_bs_restore2_rI): {
-     Eterm* next;
+     UWord *next;
      ErlBinMatchState *_ms;
      PreFetch(1, next);
      _ms = (ErlBinMatchState*) boxed_val((Eterm) r(0));
@@ -4230,7 +4250,7 @@ apply_bif_or_nif_epilogue:
      NextPF(1, next);
  }
  OpCase(i_bs_restore2_xI): {
-     Eterm* next;
+     UWord *next;
      ErlBinMatchState *_ms;
      PreFetch(2, next);
      _ms = (ErlBinMatchState*) boxed_val((Eterm) xb(Arg(0)));
@@ -4247,7 +4267,7 @@ apply_bif_or_nif_epilogue:
   * deallocate not followed by a return, and that should work.
   */
  OpCase(deallocate_I): {
-     Eterm* next;
+     UWord *next;
 
      PreFetch(1, next);
      D(Arg(0));
@@ -4297,26 +4317,25 @@ apply_bif_or_nif_epilogue:
 	     }
 	     E -= 3;
 	     ASSERT(c_p->htop <= E && E <= c_p->hend);
-	     ASSERT(is_CP((Eterm)(ep->code)));
+	     ASSERT(is_CP((UWord)(ep->code)));
 	     ASSERT(is_internal_pid(c_p->tracer_proc) || 
 		    is_internal_port(c_p->tracer_proc));
-	     E[2] = make_cp(c_p->cp);
+	     E[2] = make_cp(c_p->cp); /* XXX:PaN - code in lower range on halfword */
 	     E[1] = am_true; /* Process tracer */
 	     E[0] = make_cp(ep->code);
-	     c_p->cp = (Eterm*)
-		 make_cp(flags & MATCH_SET_EXCEPTION_TRACE
-			 ? beam_exception_trace : beam_return_trace);
+	     c_p->cp = (flags & MATCH_SET_EXCEPTION_TRACE)
+		 ? beam_exception_trace : beam_return_trace;
 	     erts_smp_proc_lock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
 	     c_p->trace_flags |= F_EXCEPTION_TRACE;
 	     erts_smp_proc_unlock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
 	 }
      }
-     SET_I((Uint *) Arg(0));
+     SET_I((UWord *)Arg(0));
      Dispatch();
  }
 
  OpCase(return_trace): {
-     Uint* code = (Uint *) E[0];
+     UWord* code = (UWord *) (UWord) E[0];
      
      SWAPOUT;		/* Needed for shared heap */
      ERTS_SMP_UNREQ_PROC_MAIN_LOCK(c_p);
@@ -4324,24 +4343,24 @@ apply_bif_or_nif_epilogue:
      ERTS_SMP_REQ_PROC_MAIN_LOCK(c_p);
      SWAPIN;
      c_p->cp = NULL;
-     SET_I((Eterm *) E[2]);
+     SET_I((UWord *) cp_val(E[2]));
      E += 3;
      Goto(*I);
  }
 
  OpCase(i_count_breakpoint): {
-     Uint real_I;
+     UWord real_I;
      
-     ErtsCountBreak((Uint *) I, &real_I);
+     ErtsCountBreak((UWord *) I, &real_I);
      ASSERT(VALID_INSTR(real_I));
      Goto(real_I);
  }
 
  OpCase(i_trace_breakpoint):
      if (! IS_TRACED_FL(c_p, F_TRACE_CALLS)) {
-	 Uint real_I;
+	 UWord real_I;
 	 
-	 ErtsBreakSkip((Uint *) I, &real_I);
+	 ErtsBreakSkip((UWord *) I, &real_I);
 	 Goto(real_I);
      }
  /* Fall through to next case */
@@ -4355,11 +4374,10 @@ apply_bif_or_nif_epilogue:
      SWAPOUT;
      reg[0] = r(0);
 
-     if (*cp_val((Eterm)c_p->cp) 
-	 == (Uint) OpCode(return_trace)) {
+     if (*(c_p->cp) == (UWord) OpCode(return_trace)) {
 	 cpp = (Uint*)&E[2];
-     } else if (*cp_val((Eterm)c_p->cp) 
-		== (Uint) OpCode(i_return_to_trace)) {
+     } else if (*(c_p->cp)
+		== (UWord) OpCode(i_return_to_trace)) {
 	 return_to_trace = !0;
 	 cpp = (Uint*)&E[0];
      } else {
@@ -4370,19 +4388,19 @@ apply_bif_or_nif_epilogue:
 	  * return_trace and/or i_return_to_trace stackframes
 	  * on the stack, they are not intermixed with y registers
 	  */
-	 Eterm *cp_save = c_p->cp;
+	 UWord *cp_save = c_p->cp;
 	 for (;;) {
 	     ASSERT(is_CP(*cpp));
-	     if (*cp_val(*cpp) == (Uint) OpCode(return_trace)) {
+	     if (*cp_val(*cpp) == (UWord) OpCode(return_trace)) {
 		 cpp += 3;
-	     } else if (*cp_val(*cpp) == (Uint) OpCode(i_return_to_trace)) {
+	     } else if (*cp_val(*cpp) == (UWord) OpCode(i_return_to_trace)) {
 		 return_to_trace = !0;
 		 cpp += 1;
 	     } else
 		 break;
 	 }
-	 c_p->cp = (Eterm *) *cpp;
-	 ASSERT(is_CP((Eterm)c_p->cp));
+	 c_p->cp = (UWord *) cp_val(*cpp);
+	 ASSERT(is_CP(*cpp));
 	 ERTS_SMP_UNREQ_PROC_MAIN_LOCK(c_p);
 	 real_I = erts_trace_break(c_p, I, reg, &flags, &tracer_pid);
 	 ERTS_SMP_REQ_PROC_MAIN_LOCK(c_p);
@@ -4418,12 +4436,12 @@ apply_bif_or_nif_epilogue:
 	 E -= 1;
 	 ASSERT(c_p->htop <= E && E <= c_p->hend);
 	 E[0] = make_cp(c_p->cp);
-	 c_p->cp = (Eterm *) make_cp(beam_return_to_trace);
+	 c_p->cp = (UWord *) beam_return_to_trace;
      }
      if (flags & MATCH_SET_RX_TRACE) {
 	 E -= 3;
 	 ASSERT(c_p->htop <= E && E <= c_p->hend);
-	 ASSERT(is_CP((Eterm) (I - 3)));
+	 ASSERT(is_CP((Eterm) (UWord) (I - 3)));
 	 ASSERT(am_true == tracer_pid || 
 		is_internal_pid(tracer_pid) || is_internal_port(tracer_pid));
 	 E[2] = make_cp(c_p->cp);
@@ -4431,9 +4449,9 @@ apply_bif_or_nif_epilogue:
 	 E[0] = make_cp(I - 3); /* We ARE at the beginning of an 
 				   instruction,
 				   the funcinfo is above i. */
-	 c_p->cp = (Eterm*)
-	     make_cp(flags & MATCH_SET_EXCEPTION_TRACE
-		     ? beam_exception_trace : beam_return_trace);
+	 c_p->cp =
+	     (flags & MATCH_SET_EXCEPTION_TRACE)
+		     ? beam_exception_trace : beam_return_trace;
 	 erts_smp_proc_lock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
 	 c_p->trace_flags |= F_EXCEPTION_TRACE;
 	 erts_smp_proc_unlock(c_p, ERTS_PROC_LOCKS_ALL_MINOR);
@@ -4446,10 +4464,10 @@ apply_bif_or_nif_epilogue:
 	 Uint *cpp = (Uint*) E;
 	 for(;;) {
 	     ASSERT(is_CP(*cpp));
-	     if (*cp_val(*cpp) == (Uint) OpCode(return_trace)) {
+	     if (*cp_val(*cpp) == (UWord) OpCode(return_trace)) {
 		 do ++cpp; while(is_not_CP(*cpp));
 		 cpp += 2;
-	     } else if (*cp_val(*cpp) == (Uint) OpCode(i_return_to_trace)) {
+	     } else if (*cp_val(*cpp) == (UWord) OpCode(i_return_to_trace)) {
 		 do ++cpp; while(is_not_CP(*cpp));
 	     } else break;
 	 }
@@ -4460,7 +4478,7 @@ apply_bif_or_nif_epilogue:
 	 SWAPIN;
      }
      c_p->cp = NULL;
-     SET_I((Eterm *) E[0]);
+     SET_I((UWord *) cp_val(E[0]));
      E += 1;
      Goto(*I);
  }
@@ -4471,7 +4489,7 @@ apply_bif_or_nif_epilogue:
 
  OpCase(i_global_cons):
  {
-     Eterm *next;
+     UWord *next;
 #ifdef HYBRID
      Eterm *hp;
 
@@ -4493,7 +4511,7 @@ apply_bif_or_nif_epilogue:
 
  OpCase(i_global_tuple):
  {
-     Eterm *next;
+     UWord *next;
      int len;
 #ifdef HYBRID
      Eterm list;
@@ -4528,7 +4546,7 @@ apply_bif_or_nif_epilogue:
 
  OpCase(i_global_copy):
  {
-     Eterm *next;
+     UWord *next;
      PreFetch(0,next);
 #ifdef HYBRID
      if (!IS_CONST(r(0)))
@@ -4557,7 +4575,7 @@ apply_bif_or_nif_epilogue:
 
  OpCase(fmove_ql): {
      Eterm fr = Arg(1);
-     Eterm* next;
+     UWord *next;
 
      PreFetch(2, next);
      GET_DOUBLE(Arg(0), *(FloatDef*)ADD_BYTE_OFFSET(freg, fr));
@@ -4567,7 +4585,7 @@ apply_bif_or_nif_epilogue:
  OpCase(fmove_dl): {
      Eterm targ1;
      Eterm fr = Arg(1);
-     Eterm* next;
+     UWord *next;
 
      PreFetch(2, next);
      GetR(0, targ1);
@@ -4588,7 +4606,7 @@ apply_bif_or_nif_epilogue:
  OpCase(fconv_dl): {
      Eterm targ1;
      Eterm fr = Arg(1);
-     Eterm* next;
+     UWord *next;
 
      GetR(0, targ1);
      PreFetch(2, next);
@@ -4617,7 +4635,7 @@ apply_bif_or_nif_epilogue:
 	 erl_exit(1, "fclearerror/i_fcheckerror without fpe signals (beam_emu)");
 #else
      OpCase(fclearerror): {
-	 Eterm* next;
+	 UWord *next;
 
 	 PreFetch(0, next);
 	 ERTS_FP_CHECK_INIT(c_p);
@@ -4625,7 +4643,7 @@ apply_bif_or_nif_epilogue:
      }
 
      OpCase(i_fcheckerror): {
-	 Eterm* next;
+	 UWord *next;
 
 	 PreFetch(0, next);
 	 ERTS_FP_ERROR(c_p, freg[0].fd, goto fbadarith);
@@ -4639,7 +4657,7 @@ apply_bif_or_nif_epilogue:
 
 
  OpCase(i_fadd_lll): {
-     Eterm* next;
+     UWord *next;
 
      PreFetch(3, next);
      ERTS_FP_CHECK_INIT(c_p);
@@ -4648,7 +4666,7 @@ apply_bif_or_nif_epilogue:
      NextPF(3, next);
  }
  OpCase(i_fsub_lll): {
-     Eterm* next;
+     UWord *next;
 
      PreFetch(3, next);
      ERTS_FP_CHECK_INIT(c_p);
@@ -4657,7 +4675,7 @@ apply_bif_or_nif_epilogue:
      NextPF(3, next);
  }
  OpCase(i_fmul_lll): {
-     Eterm* next;
+     UWord *next;
 
      PreFetch(3, next);
      ERTS_FP_CHECK_INIT(c_p);
@@ -4666,7 +4684,7 @@ apply_bif_or_nif_epilogue:
      NextPF(3, next);
  }
  OpCase(i_fdiv_lll): {
-     Eterm* next;
+     UWord *next;
 
      PreFetch(3, next);
      ERTS_FP_CHECK_INIT(c_p);
@@ -4675,7 +4693,7 @@ apply_bif_or_nif_epilogue:
      NextPF(3, next);
  }
  OpCase(i_fnegate_ll): {
-     Eterm* next;
+     UWord *next;
 
      PreFetch(2, next);
      ERTS_FP_CHECK_INIT(c_p);
@@ -4742,7 +4760,7 @@ apply_bif_or_nif_epilogue:
      neg_o_reds = -c_p->def_arg_reg[4];
      FCALLS = c_p->fcalls;
      SWAPIN;
-     switch( c_p->def_arg_reg[3] ) {
+     switch( c_p->def_arg_reg[3] ) { /* XXX:PaN - Halfword wont work with hipe yet... */
        case HIPE_MODE_SWITCH_RES_RETURN:
 	 ASSERT(is_value(reg[0]));
 	 MoveReturn(reg[0], r(0));
@@ -4754,7 +4772,7 @@ apply_bif_or_nif_epilogue:
 	 /* This can be used to call any function value, but currently it's
 	    only used to call closures referring to unloaded modules. */
 	 {
-	     Eterm *next;
+	     UWord *next;
 
 	     next = call_fun(c_p, c_p->arity - 1, reg, THE_NON_VALUE);
 	     SWAPIN;
@@ -4903,8 +4921,8 @@ apply_bif_or_nif_epilogue:
 			      bif_table[i].name,
 			      bif_table[i].arity);
 	 bif_export[i] = ep;
-	 ep->code[3] = (Eterm) OpCode(apply_bif);
-	 ep->code[4] = (Eterm) bif_table[i].f;
+	 ep->code[3] = (UWord) OpCode(apply_bif);
+	 ep->code[4] = (UWord) bif_table[i].f;
      }
 
      return;
@@ -5007,8 +5025,8 @@ Eterm error_atom[NUMBER_EXIT_CODES] = {
  * at the point of the original exception.
  */
 
-static Eterm*
-handle_error(Process* c_p, Eterm* pc, Eterm* reg, BifFunction bf)
+static UWord*
+handle_error(Process* c_p, UWord* pc, Eterm* reg, BifFunction bf)
 {
     Eterm* hp;
     Eterm Value = c_p->fvalue;
@@ -5062,7 +5080,7 @@ handle_error(Process* c_p, Eterm* pc, Eterm* reg, BifFunction bf)
     /* Find a handler or die */
     if ((c_p->catches > 0 || IS_TRACED_FL(c_p, F_EXCEPTION_TRACE))
 	&& !(c_p->freason & EXF_PANIC)) {
-	Eterm *new_pc;
+	UWord *new_pc;
         /* The Beam handler code (catch_end or try_end) checks reg[0]
 	   for THE_NON_VALUE to see if the previous code finished
 	   abnormally. If so, reg[1], reg[2] and reg[3] should hold the
@@ -5088,13 +5106,13 @@ handle_error(Process* c_p, Eterm* pc, Eterm* reg, BifFunction bf)
 /*
  * Find the nearest catch handler
  */
-static Eterm*
+static UWord*
 next_catch(Process* c_p, Eterm *reg) {
     int active_catches = c_p->catches > 0;
     int have_return_to_trace = 0;
     Eterm *ptr, *prev, *return_to_trace_ptr = NULL;
-    Uint i_return_trace = beam_return_trace[0];
-    Uint i_return_to_trace = beam_return_to_trace[0];
+    UWord i_return_trace = beam_return_trace[0];
+    UWord i_return_to_trace = beam_return_to_trace[0];
     ptr = prev = c_p->stop;
     ASSERT(is_CP(*ptr));
     ASSERT(ptr <= STACK_START(c_p));
@@ -5103,9 +5121,9 @@ next_catch(Process* c_p, Eterm *reg) {
 			     *cp_val(*ptr) != i_return_to_trace))
 	&& c_p->cp) {
 	/* Can not follow cp here - code may be unloaded */
-	Uint *cpp = cp_val((Eterm) c_p->cp);
+	UWord *cpp = c_p->cp;
 	if (cpp == beam_exception_trace) {
-	    erts_trace_exception(c_p, (Eterm*) ptr[0],
+	    erts_trace_exception(c_p, cp_val(ptr[0]),
 				 reg[1], reg[2], ptr+1);
 	    /* Skip return_trace parameters */
 	    ptr += 2;
@@ -5129,7 +5147,7 @@ next_catch(Process* c_p, Eterm *reg) {
 		    if (is_catch(*ptr) && active_catches) goto found_catch;
 		}
 		if (cp_val(*prev) == beam_exception_trace) {
-		    erts_trace_exception(c_p, (Eterm*) ptr[0],
+		    erts_trace_exception(c_p, cp_val(ptr[0]),
 					 reg[1], reg[2], ptr+1);
 		}
 		/* Skip return_trace parameters */
@@ -5258,7 +5276,7 @@ expand_error_value(Process* c_p, Uint freason, Eterm Value) {
  */
 
 static void
-save_stacktrace(Process* c_p, Eterm* pc, Eterm* reg, BifFunction bf,
+save_stacktrace(Process* c_p, UWord* pc, Eterm* reg, BifFunction bf,
 		Eterm args) {
     struct StackTrace* s;
     int sz;
@@ -5269,7 +5287,7 @@ save_stacktrace(Process* c_p, Eterm* pc, Eterm* reg, BifFunction bf,
     }
 
     /* Create a container for the exception data */
-    sz = (offsetof(struct StackTrace, trace) + sizeof(Eterm)*depth
+    sz = (offsetof(struct StackTrace, trace) + sizeof(UWord *)*depth
           + sizeof(Eterm) - 1) / sizeof(Eterm);
     s = (struct StackTrace *) HAlloc(c_p, 1 + sz);
     /* The following fields are inside the bignum */
@@ -5356,9 +5374,10 @@ save_stacktrace(Process* c_p, Eterm* pc, Eterm* reg, BifFunction bf,
 
     /* Save the actual stack trace */
     if (depth > 0) {
-	Eterm *ptr, *prev = s->depth ? s->trace[s->depth-1] : NULL;
-	Uint i_return_trace = beam_return_trace[0];
-	Uint i_return_to_trace = beam_return_to_trace[0];
+	Eterm *ptr;
+	UWord *prev = s->depth ? s->trace[s->depth-1] : NULL;
+	UWord i_return_trace = beam_return_trace[0];
+	UWord i_return_to_trace = beam_return_to_trace[0];
 	/*
 	 * Traverse the stack backwards and add all unique continuation
 	 * pointers to the buffer, up to the maximum stack trace size.
@@ -5371,7 +5390,7 @@ save_stacktrace(Process* c_p, Eterm* pc, Eterm* reg, BifFunction bf,
 				   *cp_val(*ptr) != i_return_to_trace))
 	    && c_p->cp) {
 	    /* Can not follow cp here - code may be unloaded */
-	    Uint *cpp = cp_val((Eterm) c_p->cp);
+	    UWord *cpp = c_p->cp;
 	    if (cpp == beam_exception_trace || cpp == beam_return_trace) {
 		/* Skip return_trace parameters */
 		ptr += 2;
@@ -5391,7 +5410,7 @@ save_stacktrace(Process* c_p, Eterm* pc, Eterm* reg, BifFunction bf,
 		    /* Skip stack frame variables */
 		    do ++ptr; while (is_not_CP(*ptr));
 		} else {
-		    Eterm *cp = (Eterm *)(*ptr);
+		    UWord *cp = cp_val(*ptr);
 		    if (cp != prev) {
 			/* Record non-duplicates only */
 			prev = cp;
@@ -5463,9 +5482,12 @@ build_stacktrace(Process* c_p, Eterm exc) {
     struct StackTrace* s;
     Eterm  args;
     int    depth;
-    Eterm* current;
+    UWord* current;
+#if HALFWORD_HEAP
+    UWord current_buff[3];
+#endif
     Eterm  Where = NIL;
-    Eterm* next_p = &Where;
+    Eterm *next_p = &Where;
 
     if (! (s = get_trace_from_exc(exc))) {
         return NIL;
@@ -5493,7 +5515,14 @@ build_stacktrace(Process* c_p, Eterm exc) {
      * (e.g. spawn_link(erlang, abs, [1])).
      */
     if (current == NULL) {
+#if HALFWORD_HEAP
+	current = current_buff;
+	current[0] = (UWord) c_p->initial[0];
+	current[1] = (UWord) c_p->initial[1];
+	current[2] = (UWord) c_p->initial[2];
+#else
 	current = c_p->initial;
+#endif
 	args = am_true; /* Just in case */
     } else {
 	args = get_args_from_exc(exc);
@@ -5529,7 +5558,7 @@ build_stacktrace(Process* c_p, Eterm exc) {
 	 * Finally, we go through the saved continuation pointers.
 	 */
 	for (i = 0; i < depth; i++) {
-	    Eterm *fi = find_function_from_pc((Eterm *) s->trace[i]);
+	    UWord *fi = find_function_from_pc((UWord *) s->trace[i]);
 	    if (fi == NULL) continue;
 	    mfa = TUPLE3(hp, fi[0], fi[1], make_small(fi[2]));
 	    hp += 4;
@@ -5546,7 +5575,7 @@ build_stacktrace(Process* c_p, Eterm exc) {
 
 
 static Eterm
-call_error_handler(Process* p, Eterm* fi, Eterm* reg)
+call_error_handler(Process* p, UWord* fi, Eterm* reg)
 {
     Eterm* hp;
     Export* ep;
@@ -5594,7 +5623,7 @@ call_error_handler(Process* p, Eterm* fi, Eterm* reg)
 }
 
 static Eterm
-call_breakpoint_handler(Process* p, Eterm* fi, Eterm* reg)
+call_breakpoint_handler(Process* p, UWord* fi, Eterm* reg)
 {
     Eterm* hp;
     Export* ep;
@@ -5687,7 +5716,7 @@ apply_setup_error_handler(Process* p, Eterm module, Eterm function, Uint arity,
     return ep;
 }
 
-static Uint*
+static UWord*
 apply(Process* p, Eterm module, Eterm function, Eterm args, Eterm* reg)
 {
     int arity;
@@ -5769,7 +5798,7 @@ apply(Process* p, Eterm module, Eterm function, Eterm args, Eterm* reg)
     return ep->address;
 }
 
-static Uint*
+static UWord*
 fixed_apply(Process* p, Eterm* reg, Uint arity)
 {
     Export* ep;
@@ -5875,7 +5904,7 @@ hibernate(Process* c_p, Eterm module, Eterm function, Eterm args, Eterm* reg)
     c_p->stop = STACK_START(c_p);
     c_p->catches = 0;
     c_p->i = beam_apply;
-    c_p->cp = (Eterm *) beam_apply+1;
+    c_p->cp = (UWord *) beam_apply+1;
 
     /*
      * If there are no waiting messages, garbage collect and
@@ -5905,7 +5934,7 @@ hibernate(Process* c_p, Eterm module, Eterm function, Eterm args, Eterm* reg)
     return 1;
 }
 
-static Uint*
+static UWord*
 call_fun(Process* p,		/* Current process. */
 	 int arity,		/* Number of arguments for Fun. */
 	 Eterm* reg,		/* Contents of registers. */
@@ -5925,7 +5954,7 @@ call_fun(Process* p,		/* Current process. */
     if (is_fun_header(hdr)) {
 	ErlFunThing* funp = (ErlFunThing *) fun_val(fun);
 	ErlFunEntry* fe;
-	Eterm* code_ptr;
+	UWord* code_ptr;
 	Eterm* var_ptr;
 	int actual_arity;
 	unsigned num_free;
@@ -6020,7 +6049,7 @@ call_fun(Process* p,		/* Current process. */
 	    }
 	}
     } else if (is_export_header(hdr)) {
-	Export* ep = (Export *) (export_val(fun))[1];
+	Export* ep = *((Export **) (export_val(fun) + 1));
 	int actual_arity = (int) ep->code[2];
 	if (arity == actual_arity) {
 	    return ep->address;
@@ -6088,7 +6117,7 @@ call_fun(Process* p,		/* Current process. */
     }
 }
 
-static Eterm*
+static UWord*
 apply_fun(Process* p, Eterm fun, Eterm args, Eterm* reg)
 {
     int arity;
@@ -6182,7 +6211,7 @@ erts_is_builtin(Eterm Mod, Eterm Name, int arity)
     if ((ep = export_get(&e)) == NULL) {
 	return 0;
     }
-    return ep->address == ep->code+3 && (ep->code[3] == (Uint) em_apply_bif);
+    return ep->address == ep->code+3 && (ep->code[3] == (UWord) em_apply_bif);
 }
 
 
diff --git a/erts/emulator/beam/beam_load.c b/erts/emulator/beam/beam_load.c
index 99fab28dce..b928b04210 100644
--- a/erts/emulator/beam/beam_load.c
+++ b/erts/emulator/beam/beam_load.c
@@ -51,9 +51,9 @@ ErlDrvBinary* erts_gzinflate_buffer(char*, int);
 #define EXPORTED  2
 
 #ifdef NO_JUMP_TABLE
-#  define BeamOpCode(Op) ((Uint)(Op))
+#  define BeamOpCode(Op) ((UWord)(Op))
 #else
-#  define BeamOpCode(Op) ((Eterm)beam_ops[Op])
+#  define BeamOpCode(Op) ((UWord)beam_ops[Op])
 #endif
 
 #if defined(WORDS_BIGENDIAN)
@@ -94,7 +94,7 @@ typedef struct {
 
 typedef struct {
     unsigned type;		/* Type of operand. */
-    Uint val;			/* Value of operand. */
+    UWord val;			/* Value of operand. */
     Uint bigarity;		/* Arity for bignumbers (only). */
 } GenOpArg;
 
@@ -142,7 +142,7 @@ typedef struct {
 typedef struct {
     Eterm function;		/* Tagged atom for function. */
     int arity;			/* Arity. */
-    Eterm* address;		/* Address to function in code. */
+    UWord* address;		/* Address to function in code. */
 } ExportEntry;
 
 #define MakeIffId(a, b, c, d) \
@@ -274,13 +274,13 @@ typedef struct {
     int num_functions;		/* Number of functions in module. */
     int num_labels;		/* Number of labels. */
     int code_buffer_size;	/* Size of code buffer in words.  */
-    Eterm* code;		/* Loaded code. */
+    UWord* code;		/* Loaded code. */
     int ci;			/* Current index into loaded code. */
     Label* labels;
-    Uint put_strings;		/* Linked list of put_string instructions. */
-    Uint new_bs_put_strings;	/* Linked list of i_new_bs_put_string instructions. */
+    UWord put_strings;		/* Linked list of put_string instructions. */
+    UWord new_bs_put_strings;	/* Linked list of i_new_bs_put_string instructions. */
     StringPatch* string_patches; /* Linked list of position into string table to patch. */
-    Uint catches;		/* Linked list of catch_yf instructions. */
+    UWord catches;		/* Linked list of catch_yf instructions. */
     unsigned loaded_size;	/* Final size of code when loaded. */
     byte mod_md5[16];		/* MD5 for module code. */
     int may_load_nif;           /* true if NIFs may later be loaded for this module */  
@@ -341,7 +341,7 @@ typedef struct {
 
 #define GetTagAndValue(Stp, Tag, Val) \
    do { \
-      Uint __w; \
+      UWord __w; \
       GetByte(Stp, __w); \
       Tag = __w & 0x07; \
       if ((__w & 0x08) == 0) { \
@@ -388,7 +388,7 @@ typedef struct {
        goto load_error; \
     } else { \
        int __n = (N); \
-       Uint __result = 0; \
+       UWord __result = 0; \
        Stp->file_left -= (unsigned) __n; \
        while (__n-- > 0) { \
           __result = __result << 8 | *Stp->file_p++; \
@@ -465,7 +465,7 @@ static int bin_load(Process *c_p, ErtsProcLocks c_p_locks,
 static void init_state(LoaderState* stp);
 static int insert_new_code(Process *c_p, ErtsProcLocks c_p_locks,
 			   Eterm group_leader, Eterm module,
-			   Eterm* code, Uint size, Uint catches);
+			   UWord* code, Uint size, UWord catches);
 static int scan_iff_file(LoaderState* stp, Uint* chunk_types,
 			 Uint num_types, Uint num_mandatory);
 static int load_atom_table(LoaderState* stp);
@@ -499,8 +499,8 @@ static void load_printf(int line, LoaderState* context, char *fmt, ...);
 static int transform_engine(LoaderState* st);
 static void id_to_string(Uint id, char* s);
 static void new_genop(LoaderState* stp);
-static int get_int_val(LoaderState* stp, Uint len_code, Uint* result);
-static int get_erlang_integer(LoaderState* stp, Uint len_code, Uint* result);
+static int get_int_val(LoaderState* stp, Uint len_code, UWord* result);
+static int get_erlang_integer(LoaderState* stp, Uint len_code, UWord* result);
 static int new_label(LoaderState* stp);
 static void new_literal_patch(LoaderState* stp, int pos);
 static void new_string_patch(LoaderState* stp, int pos);
@@ -513,7 +513,7 @@ static Eterm compilation_info_for_module(Process* p, Eterm mod);
 static Eterm native_addresses(Process* p, Eterm mod);
 int patch_funentries(Eterm Patchlist);
 int patch(Eterm Addresses, Uint fe);
-static int safe_mul(Uint a, Uint b, Uint* resp);
+static int safe_mul(UWord a, UWord b, UWord* resp);
 
 
 static int must_swap_floats;
@@ -592,6 +592,15 @@ erts_load_module(Process *c_p,
     return result;
 }
 
+#ifdef DEBUG
+extern void check_allocators(void);
+extern void check_allocated_block(Uint type, void *blk);
+#define CHKALLOC() check_allocators()
+#define CHKBLK(TYPE,BLK) if ((BLK) != NULL) check_allocated_block((TYPE),(BLK))
+#else
+#define CHKALLOC() /* nothing */
+#define CHKBLK(TYPE,BLK) /* nothing */
+#endif
 
 static int
 bin_load(Process *c_p, ErtsProcLocks c_p_locks,
@@ -608,6 +617,12 @@ bin_load(Process *c_p, ErtsProcLocks c_p_locks,
      * Scan the IFF file.
      */
 
+#ifdef DEBUG
+    fprintf(stderr,"Loading a module\r\n");
+#endif
+
+    CHKALLOC();
+    CHKBLK(ERTS_ALC_T_CODE,state.code);
     state.file_name = "IFF header for Beam file";
     state.file_p = bytes;
     state.file_left = unloaded_size;
@@ -619,6 +634,7 @@ bin_load(Process *c_p, ErtsProcLocks c_p_locks,
      * Read the header for the code chunk.
      */
 
+    CHKBLK(ERTS_ALC_T_CODE,state.code);
     define_file(&state, "code chunk header", CODE_CHUNK);
     if (!read_code_header(&state)) {
 	goto load_error;
@@ -628,6 +644,7 @@ bin_load(Process *c_p, ErtsProcLocks c_p_locks,
      * Read the atom table.
      */
 
+    CHKBLK(ERTS_ALC_T_CODE,state.code);
     define_file(&state, "atom table", ATOM_CHUNK);
     if (!load_atom_table(&state)) {
 	goto load_error;
@@ -637,6 +654,7 @@ bin_load(Process *c_p, ErtsProcLocks c_p_locks,
      * Read the import table.
      */
 
+    CHKBLK(ERTS_ALC_T_CODE,state.code);
     define_file(&state, "import table", IMP_CHUNK);
     if (!load_import_table(&state)) {
 	goto load_error;
@@ -646,6 +664,7 @@ bin_load(Process *c_p, ErtsProcLocks c_p_locks,
      * Read the lambda (fun) table.
      */
 
+    CHKBLK(ERTS_ALC_T_CODE,state.code);
     if (state.chunks[LAMBDA_CHUNK].size > 0) {
 	define_file(&state, "lambda (fun) table", LAMBDA_CHUNK);
 	if (!read_lambda_table(&state)) {
@@ -657,6 +676,7 @@ bin_load(Process *c_p, ErtsProcLocks c_p_locks,
      * Read the literal table.
      */
 
+    CHKBLK(ERTS_ALC_T_CODE,state.code);
     if (state.chunks[LITERAL_CHUNK].size > 0) {
 	define_file(&state, "literals table (constant pool)", LITERAL_CHUNK);
 	if (!read_literal_table(&state)) {
@@ -668,18 +688,25 @@ bin_load(Process *c_p, ErtsProcLocks c_p_locks,
      * Load the code chunk.
      */
 
+    CHKBLK(ERTS_ALC_T_CODE,state.code);
     state.file_name = "code chunk";
     state.file_p = state.code_start;
     state.file_left = state.code_size;
-    if (!load_code(&state) || !freeze_code(&state)) {
+    if (!load_code(&state)) {
 	goto load_error;
     }
+    CHKBLK(ERTS_ALC_T_CODE,state.code);
+    if (!freeze_code(&state)) {
+	goto load_error;
+    }
+
 
     /*
      * Read and validate the export table.  (This must be done after
      * loading the code, because it contains labels.)
      */
     
+    CHKBLK(ERTS_ALC_T_CODE,state.code);
     define_file(&state, "export table", EXP_CHUNK);
     if (!read_export_table(&state)) {
 	goto load_error;
@@ -690,16 +717,19 @@ bin_load(Process *c_p, ErtsProcLocks c_p_locks,
      * exported and imported functions.  This can't fail.
      */
     
+    CHKBLK(ERTS_ALC_T_CODE,state.code);
     rval = insert_new_code(c_p, c_p_locks, state.group_leader, state.module,
 			   state.code, state.loaded_size, state.catches);
     if (rval < 0) {
 	goto load_error;
     }
+    CHKBLK(ERTS_ALC_T_CODE,state.code);
     final_touch(&state);
 
     /*
      * Loading succeded.
      */
+    CHKBLK(ERTS_ALC_T_CODE,state.code);
     rval = 0;
     state.code = NULL;		/* Prevent code from being freed. */
     *modp = state.module;
@@ -755,6 +785,9 @@ bin_load(Process *c_p, ErtsProcLocks c_p_locks,
 	erts_free(ERTS_ALC_T_LOADER_TMP, (void *) state.genop_blocks);
 	state.genop_blocks = next;
     }
+#ifdef DEBUG
+    erts_fprintf(stderr,"Loaded %T\r\n",*modp);
+#endif
 
     return rval;
 }
@@ -791,7 +824,7 @@ init_state(LoaderState* stp)
 
 static int
 insert_new_code(Process *c_p, ErtsProcLocks c_p_locks,
-		Eterm group_leader, Eterm module, Eterm* code, Uint size, Uint catches)
+		Eterm group_leader, Eterm module, UWord* code, Uint size, UWord catches)
 {
     Module* modp;
     int rval;
@@ -833,7 +866,7 @@ insert_new_code(Process *c_p, ErtsProcLocks c_p_locks,
 	modules[i] = modules[i-1];
     }
     modules[i].start = code;
-    modules[i].end = (Eterm *) (((byte *)code) + size);
+    modules[i].end = (UWord *) (((byte *)code) + size);
     num_loaded_modules++;
     mid_module = &modules[num_loaded_modules/2];
     return 0;
@@ -1083,7 +1116,7 @@ load_import_table(LoaderState* stp)
 	 * the BIF function.
 	 */
 	if ((e = erts_find_export_entry(mod, func, arity)) != NULL) {
-	    if (e->code[3] == (Uint) em_apply_bif) {
+	    if (e->code[3] == (UWord) em_apply_bif) {
 		stp->import[i].bf = (BifFunction) e->code[4];
 		if (func == am_load_nif && mod == am_erlang && arity == 2) {
 		    stp->may_load_nif = 1;
@@ -1151,7 +1184,7 @@ read_export_table(LoaderState* stp)
 	 * redefine).
 	 */
 	if ((e = erts_find_export_entry(stp->module, func, arity)) != NULL) {
-	    if (e->code[3] == (Uint) em_apply_bif) {
+	    if (e->code[3] == (UWord) em_apply_bif) {
 		int j;
 
 		for (j = 0; j < sizeof(allow_redef)/sizeof(allow_redef[0]); j++) {
@@ -1220,7 +1253,7 @@ static int
 read_literal_table(LoaderState* stp)
 {
     int i;
-    Uint uncompressed_sz;
+    UWord uncompressed_sz;
     byte* uncompressed = 0;
 
     GetInt(stp, 4, uncompressed_sz);
@@ -1338,8 +1371,8 @@ read_code_header(LoaderState* stp)
      * Initialize code area.
      */
     stp->code_buffer_size = erts_next_heap_size(2048 + stp->num_functions, 0);
-    stp->code = (Eterm*) erts_alloc(ERTS_ALC_T_CODE,
-				    sizeof(Eterm) * stp->code_buffer_size);
+    stp->code = (UWord *) erts_alloc(ERTS_ALC_T_CODE,
+				    sizeof(UWord) * stp->code_buffer_size);
 
     stp->code[MI_NUM_FUNCTIONS] = stp->num_functions;
     stp->ci = MI_FUNCTIONS + stp->num_functions + 1;
@@ -1365,16 +1398,18 @@ read_code_header(LoaderState* stp)
        LoadError2(Stp, "bad tag %d; expected %d", Actual, Expected); \
     } else {}
 
-#define Need(w) \
-    ASSERT(ci <= code_buffer_size); \
-    if (code_buffer_size < ci+(w)) { \
-        code_buffer_size = erts_next_heap_size(ci+(w), 0); \
-	stp->code = code \
-	    = (Eterm *) erts_realloc(ERTS_ALC_T_CODE, \
-				     (void *) code, \
-				     code_buffer_size * sizeof(Eterm)); \
-    }
+#define CodeNeed(w) do {						\
+    ASSERT(ci <= code_buffer_size);					\
+    if (code_buffer_size < ci+(w)) {					\
+        code_buffer_size = erts_next_heap_size(ci+(w), 0);		\
+	stp->code = code						\
+	    = (UWord *) erts_realloc(ERTS_ALC_T_CODE,			\
+				     (void *) code,			\
+				     code_buffer_size * sizeof(UWord));	\
+    } 									\
+} while (0)
     
+#define TermWords(t) (((t) / (sizeof(UWord)/sizeof(Eterm))) + !!((t) % (sizeof(UWord)/sizeof(Eterm))))
 
 
 static int
@@ -1387,7 +1422,7 @@ load_code(LoaderState* stp)
     char* sign;
     int arg;			/* Number of current argument. */
     int num_specific;		/* Number of specific ops for current. */
-    Eterm* code;
+    UWord* code;
     int code_buffer_size;
     int specific;
     Uint last_label = 0;	/* Number of last label. */
@@ -1446,7 +1481,7 @@ load_code(LoaderState* stp)
       if (((First) & 0x08) == 0) { \
 	 Val = (First) >> 4; \
       } else if (((First) & 0x10) == 0) { \
-         Uint __w; \
+         UWord __w; \
 	 GetByte(Stp, __w); \
 	 Val = (((First) >> 5) << 8) | __w; \
       } else { \
@@ -1455,7 +1490,7 @@ load_code(LoaderState* stp)
    } while (0)
 
 	for (arg = 0; arg < arity; arg++) {
-	    Uint first;
+	    UWord first;
 
 	    GetByte(stp, first);
 	    last_op->a[arg].type = first & 0x07;
@@ -1464,7 +1499,7 @@ load_code(LoaderState* stp)
 		if ((first & 0x08) == 0) {
 		    last_op->a[arg].val = first >> 4;
 		} else if ((first & 0x10) == 0) {
-		    Uint w;
+		    UWord w;
 		    GetByte(stp, w);
 		    ASSERT(first < 0x800);
 		    last_op->a[arg].val = ((first >> 5) << 8) | w;
@@ -1523,7 +1558,7 @@ load_code(LoaderState* stp)
 		break;
 	    case TAG_z:
 		{
-		    Uint ext_tag;
+		    UWord ext_tag;
 		    unsigned tag;
 
 		    GetValue(stp, first, ext_tag);
@@ -1531,14 +1566,14 @@ load_code(LoaderState* stp)
 		    case 0:	/* Floating point number */
 			{
 			    Eterm* hp;
-# ifndef ARCH_64
+#if !defined(ARCH_64) || HALFWORD_HEAP /* XXX:PaN - Should use ARCH_64 variant instead */
 			    Uint high, low;
 # endif
 			    last_op->a[arg].val = new_literal(stp, &hp,
 							      FLOAT_SIZE_OBJECT);
 			    hp[0] = HEADER_FLONUM;
 			    last_op->a[arg].type = TAG_q;
-# ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 			    GetInt(stp, 8, hp[1]);
 # else
 			    GetInt(stp, 4, high);
@@ -1575,10 +1610,10 @@ load_code(LoaderState* stp)
 			break;
 		    case 3: 	/* Allocation list. */
 			{
-			    Uint n;
-			    Uint type;
-			    Uint val;
-			    Uint words = 0;
+			    UWord n;
+			    UWord type;
+			    UWord val;
+			    UWord words = 0;
 			    
 			    stp->new_float_instructions = 1;
 			    GetTagAndValue(stp, tag, n);
@@ -1607,7 +1642,7 @@ load_code(LoaderState* stp)
 			}
 		    case 4:	/* Literal. */
 			{
-			    Uint val;
+			    UWord val;
 
 			    GetTagAndValue(stp, tag, val);
 			    VerifyTag(stp, tag, TAG_u);
@@ -1734,7 +1769,7 @@ load_code(LoaderState* stp)
 	    }
 
 	    stp->specific_op = specific;
-	    Need(opc[stp->specific_op].sz+2); /* Extra margin for packing */
+	    CodeNeed(opc[stp->specific_op].sz+2); /* Extra margin for packing */
 	    code[ci++] = BeamOpCode(stp->specific_op);
 	}
 	
@@ -1772,7 +1807,7 @@ load_code(LoaderState* stp)
 	    case 'c':		/* Tagged constant */
 		switch (tag) {
 		case TAG_i:
-		    code[ci++] = make_small(tmp_op->a[arg].val);
+		    code[ci++] = (UWord) make_small((Uint) tmp_op->a[arg].val);
 		    break;
 		case TAG_a:
 		    code[ci++] = tmp_op->a[arg].val;
@@ -1802,7 +1837,7 @@ load_code(LoaderState* stp)
 		    code[ci++] = make_yreg(tmp_op->a[arg].val);
 		    break;
 		case TAG_i:
-		    code[ci++] = make_small(tmp_op->a[arg].val);
+		    code[ci++] = (UWord) make_small((Uint)tmp_op->a[arg].val);
 		    break;
 		case TAG_a:
 		    code[ci++] = tmp_op->a[arg].val;
@@ -1896,12 +1931,12 @@ load_code(LoaderState* stp)
 		if (stp->import[i].bf == NULL) {
 		    LoadError1(stp, "not a BIF: import table index %d", i);
 		}
-		code[ci++] = (Eterm) stp->import[i].bf;
+		code[ci++] = (UWord) stp->import[i].bf;
 		break;
-	    case 'P':		/* Byte offset into tuple */
+	    case 'P':		/* Byte offset into tuple */ /* XXX:PaN - * sizeof(Eterm or Eterm *) ? */
 		VerifyTag(stp, tag, TAG_u);
 		tmp = tmp_op->a[arg].val;
-		code[ci++] = (Eterm) ((tmp_op->a[arg].val+1) * sizeof(Eterm *));
+		code[ci++] = (UWord) ((tmp_op->a[arg].val+1) * sizeof(Eterm));
 		break;
 	    case 'l':		/* Floating point register. */
 		VerifyTag(stp, tag_to_letter[tag], *sign);
@@ -1925,17 +1960,17 @@ load_code(LoaderState* stp)
 	for ( ; arg < tmp_op->arity; arg++) {
 	    switch (tmp_op->a[arg].type) {
 	    case TAG_i:
-		Need(1);
+		CodeNeed(1);
 		code[ci++] = make_small(tmp_op->a[arg].val);
 		break;
 	    case TAG_u:
 	    case TAG_a:
 	    case TAG_v:
-		Need(1);
+		CodeNeed(1);
 		code[ci++] = tmp_op->a[arg].val;
 		break;
 	    case TAG_f:
-		Need(1);
+		CodeNeed(1);
 		code[ci] = stp->labels[tmp_op->a[arg].val].patches;
 		stp->labels[tmp_op->a[arg].val].patches = ci;
 		ci++;
@@ -1947,24 +1982,41 @@ load_code(LoaderState* stp)
 		    lit = stp->literals[tmp_op->a[arg].val].term;
 		    if (is_big(lit)) {
 			Eterm* bigp;
+			Eterm *tmp;
 			Uint size;
+			Uint term_size;
 
 			bigp = big_val(lit);
-			size = bignum_header_arity(*bigp);
-			Need(size+1);
-			code[ci++] = *bigp++;
-			while (size-- > 0) {
-			    code[ci++] = *bigp++;
+			term_size = bignum_header_arity(*bigp);
+			size = TermWords(term_size + 1);
+			CodeNeed(size);
+			tmp = (Eterm *) (code + ci);
+			*tmp++ = *bigp++;
+			while (term_size-- > 0) {
+			    *tmp++ = *bigp++;
 			}
+			ci +=size;
 		    } else if (is_float(lit)) {
-#ifdef ARCH_64
-			Need(1);
+#if defined(ARCH_64) && !HALFWORD_HEAP
+			CodeNeed(1);
 			code[ci++] = float_val(stp->literals[tmp_op->a[arg].val].term)[1];
+#elif HALFWORD_HEAP
+			Eterm* fptr;
+			Uint size;
+			Eterm *tmp;
+
+			fptr = float_val(stp->literals[tmp_op->a[arg].val].term)+1;
+			size = TermWords(2);
+			CodeNeed(size);
+			tmp = (Eterm *) (code + ci);
+			*tmp++ = *fptr++;
+			*tmp = *fptr;
+			ci += size;
 #else
 			Eterm* fptr;
 
 			fptr = float_val(stp->literals[tmp_op->a[arg].val].term)+1;
-			Need(2);
+			CodeNeed(2);
 			code[ci++] = *fptr++;
 			code[ci++] = *fptr;
 #endif
@@ -1978,7 +2030,7 @@ load_code(LoaderState* stp)
 			   tag_to_letter[tmp_op->a[arg].type]);
 	    }
 	}
-
+#if !HALFWORD_HEAP /* XXX:PaN - just disabled during development */
 	/*
 	 * The packing engine.
 	 */
@@ -1987,7 +2039,7 @@ load_code(LoaderState* stp)
 	    Uint stack[8];	/* Stack. */
 	    Uint* sp = stack;	/* Points to next free position. */
 	    Uint packed = 0;	/* Accumulator for packed operations. */
-	    
+	    ASSERT(0); /* XXX:PaN - just to check that this does not happen when packing is disabled... */
 	    for (prog = opc[stp->specific_op].pack; *prog; prog++) {
 		switch (*prog) {
 		case 'g':	/* Get instruction; push on stack. */
@@ -2015,7 +2067,7 @@ load_code(LoaderState* stp)
 	    }
 	    ASSERT(sp == stack); /* Incorrect program? */
 	}
-
+#endif /* !HALFWORD_HEAP */
 	/*
 	 * Handle a few special cases.
 	 */
@@ -2037,9 +2089,9 @@ load_code(LoaderState* stp)
 			/* Must make room for call_nif op */
 			int pad = MIN_FUNC_SZ - (finfo_ix - last_func_start);
 			ASSERT(pad > 0 && pad < MIN_FUNC_SZ);
-			Need(pad);
-			sys_memmove(&code[finfo_ix+pad], &code[finfo_ix], FINFO_SZ*sizeof(Eterm));
-			sys_memset(&code[finfo_ix], 0, pad*sizeof(Eterm));
+			CodeNeed(pad);
+			sys_memmove(&code[finfo_ix+pad], &code[finfo_ix], FINFO_SZ*sizeof(UWord));
+			sys_memset(&code[finfo_ix], 0, pad*sizeof(UWord));
 			ci += pad;
 			stp->labels[last_label].value += pad;
 		    }
@@ -2050,6 +2102,7 @@ load_code(LoaderState* stp)
 		 */
 		stp->function = code[ci-2];
 		stp->arity = code[ci-1];
+
 		ASSERT(stp->labels[last_label].value == ci - FINFO_SZ);
 		offset = MI_FUNCTIONS + function_number;
 		code[offset] = stp->labels[last_label].patches;
@@ -2161,7 +2214,6 @@ load_code(LoaderState* stp)
 	}
     }
     
-#undef Need
 
  load_error:
     return 0;
@@ -2373,7 +2425,7 @@ gen_get_integer2(LoaderState* stp, GenOpArg Fail, GenOpArg Ms, GenOpArg Live,
 		 GenOpArg Flags, GenOpArg Dst)
 {
     GenOp* op;
-    Uint bits;
+    UWord bits;
 
     NEW_GENOP(stp, op);
 
@@ -2838,14 +2890,14 @@ gen_literal_timeout(LoaderState* stp, GenOpArg Fail, GenOpArg Time)
     op->a[1].type = TAG_u;
     
     if (Time.type == TAG_i && (timeout = Time.val) >= 0 &&
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 	(timeout >> 32) == 0
 #else
 	1
 #endif
 	) {
 	op->a[1].val = timeout;
-#if !defined(ARCH_64)
+#if !defined(ARCH_64) || HALFWORD_HEAP
     } else if (Time.type == TAG_q) {
 	Eterm big;
 
@@ -2856,11 +2908,13 @@ gen_literal_timeout(LoaderState* stp, GenOpArg Fail, GenOpArg Time)
 	if (big_arity(big) > 1 || big_sign(big)) {
 	    goto error;
 	} else {
-	    (void) term_to_Uint(big, &op->a[1].val);
+	    Uint u;
+	    (void) term_to_Uint(big, &u);
+	    op->a[1].val = (UWord) u;
 	}
 #endif
     } else {
-#if !defined(ARCH_64)
+#if !defined(ARCH_64) || HALFWORD_HEAP
     error:
 #endif
 	op->op = genop_i_wait_error_0;
@@ -2883,14 +2937,14 @@ gen_literal_timeout_locked(LoaderState* stp, GenOpArg Fail, GenOpArg Time)
     op->a[1].type = TAG_u;
     
     if (Time.type == TAG_i && (timeout = Time.val) >= 0 &&
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 	(timeout >> 32) == 0
 #else
 	1
 #endif
 	) {
 	op->a[1].val = timeout;
-#ifndef ARCH_64
+#if !defined(ARCH_64) || HALFWORD_HEAP
     } else if (Time.type == TAG_q) {
 	Eterm big;
 
@@ -2901,11 +2955,13 @@ gen_literal_timeout_locked(LoaderState* stp, GenOpArg Fail, GenOpArg Time)
 	if (big_arity(big) > 1 || big_sign(big)) {
 	    goto error;
 	} else {
-	    (void) term_to_Uint(big, &op->a[1].val);
+	    Uint u;
+	    (void) term_to_Uint(big, &u);
+	    op->a[1].val = (UWord) u;
 	}
 #endif
     } else {
-#ifndef ARCH_64
+#if !defined(ARCH_64) || HALFWORD_HEAP
     error:
 #endif
 	op->op = genop_i_wait_error_locked_0;
@@ -3321,7 +3377,7 @@ gen_make_fun2(LoaderState* stp, GenOpArg idx)
     op->op = genop_i_make_fun_2;
     op->arity = 2;
     op->a[0].type = TAG_u;
-    op->a[0].val = (Uint) fe;
+    op->a[0].val = (UWord) fe;
     op->a[1].type = TAG_u;
     op->a[1].val = stp->lambdas[idx.val].num_free;
     op->next = NULL;
@@ -3376,7 +3432,8 @@ gen_guard_bif(LoaderState* stp, GenOpArg Fail, GenOpArg Live, GenOpArg Bif,
 static int
 freeze_code(LoaderState* stp)
 {
-    Eterm* code = stp->code;
+    UWord* code = stp->code;
+    Uint *literal_end = NULL;
     Uint index;
     int i;
     byte* str_table;
@@ -3401,46 +3458,49 @@ freeze_code(LoaderState* stp)
      * Calculate the final size of the code.
      */
 
-    size = (stp->ci + stp->total_literal_size) * sizeof(Eterm) +
+    size = (stp->ci * sizeof(UWord)) + (stp->total_literal_size * sizeof(Eterm)) +
 	strtab_size + attr_size + compile_size;
 
     /*
      * Move the code to its final location.
      */
 
-    code = (Eterm *) erts_realloc(ERTS_ALC_T_CODE, (void *) code, size);
-
+    code = (UWord *) erts_realloc(ERTS_ALC_T_CODE, (void *) code, size);
+    CHKBLK(ERTS_ALC_T_CODE,code);
     /*
      * Place a pointer to the op_int_code_end instruction in the
      * function table in the beginning of the file.
      */
 
-    code[MI_FUNCTIONS+stp->num_functions] = (Eterm) (code + stp->ci - 1);
+    code[MI_FUNCTIONS+stp->num_functions] = (UWord) (code + stp->ci - 1);
+    CHKBLK(ERTS_ALC_T_CODE,code);
 
     /*
      * Store the pointer to the on_load function.
      */
 
     if (stp->on_load) {
-	code[MI_ON_LOAD_FUNCTION_PTR] = (Eterm) (code + stp->on_load);
+	code[MI_ON_LOAD_FUNCTION_PTR] = (UWord) (code + stp->on_load);
     } else {
 	code[MI_ON_LOAD_FUNCTION_PTR] = 0;
     }
+    CHKBLK(ERTS_ALC_T_CODE,code);
 
+    literal_end = (Uint *) (code+stp->ci);
     /*
      * Place the literal heap directly after the code and fix up all
      * put_literal instructions that refer to it.
      */
     {
-	Eterm* ptr;
-	Eterm* low;
-	Eterm* high;
+	Uint* ptr;
+	Uint* low;
+	Uint* high;
 	LiteralPatch* lp;
 
-	low = code+stp->ci;
+	low = (Uint *) (code+stp->ci);
 	high = low + stp->total_literal_size;
-	code[MI_LITERALS_START] = (Eterm) low;
-	code[MI_LITERALS_END] = (Eterm) high;
+	code[MI_LITERALS_START] = (UWord) low;
+	code[MI_LITERALS_END] = (UWord) high;
 	ptr = low;
 	for (i = 0; i < stp->num_literals; i++) {
 	    Uint offset;
@@ -3472,7 +3532,7 @@ freeze_code(LoaderState* stp)
 	}
 	lp = stp->literal_patches;
 	while (lp != 0) {
-	    Uint* op_ptr;
+	    UWord* op_ptr;
 	    Uint literal;
 	    Literal* lit;
 
@@ -3485,38 +3545,48 @@ freeze_code(LoaderState* stp)
 	    op_ptr[0] = literal;
 	    lp = lp->next;
 	}
-	stp->ci += stp->total_literal_size;
+	literal_end += stp->total_literal_size;
     }
     
     /*
      * Place the string table and, optionally, attributes, after the literal heap.
      */
+    CHKBLK(ERTS_ALC_T_CODE,code);
 
-    sys_memcpy(code+stp->ci, stp->chunks[STR_CHUNK].start, strtab_size);
-    str_table = (byte *) (code+stp->ci);
+    sys_memcpy(literal_end, stp->chunks[STR_CHUNK].start, strtab_size);
+    CHKBLK(ERTS_ALC_T_CODE,code);
+    str_table = (byte *) literal_end;
     if (attr_size) {
 	byte* attr = str_table + strtab_size;
 	sys_memcpy(attr, stp->chunks[ATTR_CHUNK].start, stp->chunks[ATTR_CHUNK].size);
-	code[MI_ATTR_PTR] = (Eterm) attr;
-	code[MI_ATTR_SIZE] = (Eterm) stp->chunks[ATTR_CHUNK].size;
+	code[MI_ATTR_PTR] = (UWord) attr;
+	code[MI_ATTR_SIZE] = (UWord) stp->chunks[ATTR_CHUNK].size;
 	decoded_size = erts_decode_ext_size(attr, attr_size, 0);
 	if (decoded_size < 0) {
  	    LoadError0(stp, "bad external term representation of module attributes");
  	}
 	code[MI_ATTR_SIZE_ON_HEAP] = decoded_size;
     }
+    CHKBLK(ERTS_ALC_T_CODE,code);
     if (compile_size) {
 	byte* compile_info = str_table + strtab_size + attr_size;
+    CHKBLK(ERTS_ALC_T_CODE,code);
 	sys_memcpy(compile_info, stp->chunks[COMPILE_CHUNK].start,
 	       stp->chunks[COMPILE_CHUNK].size);
-	code[MI_COMPILE_PTR] = (Eterm) compile_info;
-	code[MI_COMPILE_SIZE] = (Eterm) stp->chunks[COMPILE_CHUNK].size;
+    CHKBLK(ERTS_ALC_T_CODE,code);
+	code[MI_COMPILE_PTR] = (UWord) compile_info;
+    CHKBLK(ERTS_ALC_T_CODE,code);
+	code[MI_COMPILE_SIZE] = (UWord) stp->chunks[COMPILE_CHUNK].size;
+    CHKBLK(ERTS_ALC_T_CODE,code);
 	decoded_size = erts_decode_ext_size(compile_info, compile_size, 0);
+    CHKBLK(ERTS_ALC_T_CODE,code);
 	if (decoded_size < 0) {
  	    LoadError0(stp, "bad external term representation of compilation information");
  	}
+    CHKBLK(ERTS_ALC_T_CODE,code);
 	code[MI_COMPILE_SIZE_ON_HEAP] = decoded_size;
     }
+    CHKBLK(ERTS_ALC_T_CODE,code);
 
 
     /*
@@ -3529,9 +3599,10 @@ freeze_code(LoaderState* stp)
     while (index != 0) {
 	Uint next = code[index];
 	code[index] = BeamOpCode(op_put_string_IId);
-	code[index+2] = (Uint) (str_table + code[index+2] + code[index+1] - 1);
+	code[index+2] = (UWord) (str_table + code[index+2] + code[index+1] - 1);
 	index = next;
     }
+    CHKBLK(ERTS_ALC_T_CODE,code);
 
     /*
      * Go through all i_new_bs_put_strings instructions, restore the pointer to
@@ -3543,23 +3614,25 @@ freeze_code(LoaderState* stp)
     while (index != 0) {
 	Uint next = code[index];
 	code[index] = BeamOpCode(op_bs_put_string_II);
-	code[index+2] = (Uint) (str_table + code[index+2]);
+	code[index+2] = (UWord) (str_table + code[index+2]);
 	index = next;
     }
+    CHKBLK(ERTS_ALC_T_CODE,code);
 
     {
 	StringPatch* sp = stp->string_patches;
 
 	while (sp != 0) {
-	    Uint* op_ptr;
+	    UWord* op_ptr;
 	    byte* strp;
 
 	    op_ptr = code + sp->pos;
 	    strp = str_table + op_ptr[0];
-	    op_ptr[0] = (Eterm) strp;
+	    op_ptr[0] = (UWord) strp;
 	    sp = sp->next;
 	}
     }
+    CHKBLK(ERTS_ALC_T_CODE,code);
 
     /*
      * Resolve all labels.
@@ -3579,10 +3652,11 @@ freeze_code(LoaderState* stp)
 	    ASSERT(this_patch < stp->ci);
 	    next_patch = code[this_patch];
 	    ASSERT(next_patch < stp->ci);
-	    code[this_patch] = (Uint) (code + value);
+	    code[this_patch] = (UWord) (code + value);
 	    this_patch = next_patch;
 	}
     }
+    CHKBLK(ERTS_ALC_T_CODE,code);
 
     /*
      * Fix all catch_yf instructions.
@@ -3590,13 +3664,14 @@ freeze_code(LoaderState* stp)
     index = stp->catches;
     catches = BEAM_CATCHES_NIL;
     while (index != 0) {
-	Uint next = code[index];
+	UWord next = code[index];
 	code[index] = BeamOpCode(op_catch_yf);
-	catches = beam_catches_cons((Uint*)code[index+2], catches);
+	catches = beam_catches_cons((UWord *)code[index+2], catches);
 	code[index+2] = make_catch(catches);
 	index = next;
     }
     stp->catches = catches;
+    CHKBLK(ERTS_ALC_T_CODE,code);
 
     /*
      * Save the updated code pointer and code size.
@@ -3605,6 +3680,7 @@ freeze_code(LoaderState* stp)
     stp->code = code;
     stp->loaded_size = size;
 
+    CHKBLK(ERTS_ALC_T_CODE,code);
     return 1;
 
  load_error:
@@ -3638,7 +3714,7 @@ final_touch(LoaderState* stp)
 	     * callable yet.
 	     */
 	    ep->address = ep->code+3;
-	    ep->code[4] = (Eterm) stp->export[i].address;
+	    ep->code[4] = (UWord) stp->export[i].address;
 	}
     }
 
@@ -3650,14 +3726,14 @@ final_touch(LoaderState* stp)
 	Eterm mod;
 	Eterm func;
 	Uint arity;
-	Uint import;
+	UWord import;
 	Uint current;
 	Uint next;
 
 	mod = stp->import[i].module;
 	func = stp->import[i].function;
 	arity = stp->import[i].arity;
-	import = (Uint) erts_export_put(mod, func, arity);
+	import = (UWord) erts_export_put(mod, func, arity);
 	current = stp->import[i].patches;
 	while (current != 0) {
 	    ASSERT(current < stp->ci);
@@ -3675,7 +3751,7 @@ final_touch(LoaderState* stp)
 	for (i = 0; i < stp->num_lambdas; i++) {
 	    unsigned entry_label = stp->lambdas[i].label;
 	    ErlFunEntry* fe = stp->lambdas[i].fe;
-	    Eterm* code_ptr = (Eterm *) (stp->code + stp->labels[entry_label].value);
+	    UWord* code_ptr = (UWord *) (stp->code + stp->labels[entry_label].value);
 
 	    if (fe->address[0] != 0) {
 		/*
@@ -3807,7 +3883,7 @@ transform_engine(LoaderState* st)
 		if (i >= st->num_imports || st->import[i].bf == NULL)
 		    goto restart;
 		if (bif_number != -1 &&
-		    bif_export[bif_number]->code[4] != (Uint) st->import[i].bf) {
+		    bif_export[bif_number]->code[4] != (UWord) st->import[i].bf) {
 		    goto restart;
 		}
 	    }
@@ -4071,7 +4147,7 @@ load_printf(int line, LoaderState* context, char *fmt,...)
 
 
 static int
-get_int_val(LoaderState* stp, Uint len_code, Uint* result)
+get_int_val(LoaderState* stp, Uint len_code, UWord* result)
 {
     Uint count;
     Uint val;
@@ -4103,7 +4179,7 @@ get_int_val(LoaderState* stp, Uint len_code, Uint* result)
 
 
 static int
-get_erlang_integer(LoaderState* stp, Uint len_code, Uint* result)
+get_erlang_integer(LoaderState* stp, Uint len_code, UWord* result)
 {
     Uint count;
     Sint val;
@@ -4124,11 +4200,12 @@ get_erlang_integer(LoaderState* stp, Uint len_code, Uint* result)
 	count = len_code + 2;
     } else {
 	Uint tag;
+	UWord len_word;
 
 	ASSERT(len_code == 7);
-	GetTagAndValue(stp, tag, len_code);
+	GetTagAndValue(stp, tag, len_word);
 	VerifyTag(stp, TAG_u, tag);
-	count = len_code + 9;
+	count = len_word + 9;
     }
 
     /*
@@ -4376,7 +4453,7 @@ functions_in_module(Process* p, /* Process whose heap to use. */
 		     Eterm mod) /* Tagged atom for module. */
 {
     Module* modp;
-    Eterm* code;
+    UWord* code;
     int i;
     Uint num_functions;
     Eterm* hp;
@@ -4394,9 +4471,9 @@ functions_in_module(Process* p, /* Process whose heap to use. */
     num_functions = code[MI_NUM_FUNCTIONS];
     hp = HAlloc(p, 5*num_functions);
     for (i = num_functions-1; i >= 0 ; i--) {
-	Eterm* func_info = (Eterm *) code[MI_FUNCTIONS+i];
-	Eterm name = func_info[3];
-	int arity = func_info[4];
+	UWord* func_info = (UWord *) code[MI_FUNCTIONS+i];
+	Eterm name = (Eterm) func_info[3];
+	int arity = (int) func_info[4];
 	Eterm tuple;
 
 	ASSERT(is_atom(name));
@@ -4419,7 +4496,7 @@ static Eterm
 native_addresses(Process* p, Eterm mod)
 {
     Module* modp;
-    Eterm* code;
+    UWord* code;
     int i;
     Eterm* hp;
     Uint num_functions;
@@ -4442,9 +4519,9 @@ native_addresses(Process* p, Eterm mod)
     hp = HAlloc(p, need);
     hp_end = hp + need;
     for (i = num_functions-1; i >= 0 ; i--) {
-	Eterm* func_info = (Eterm *) code[MI_FUNCTIONS+i];
-	Eterm name = func_info[3];
-	int arity = func_info[4];
+	UWord* func_info = (UWord *) code[MI_FUNCTIONS+i];
+	Eterm name = (Eterm) func_info[3];
+	int arity = (int) func_info[4];
 	Eterm tuple;
 
 	ASSERT(is_atom(name));
@@ -4486,7 +4563,7 @@ exported_from_module(Process* p, /* Process whose heap to use. */
 	    Eterm tuple;
 	    
 	    if (ep->address == ep->code+3 &&
-		ep->code[3] == (Eterm) em_call_error_handler) {
+		ep->code[3] == (UWord) em_call_error_handler) {
 		/* There is a call to the function, but it does not exist. */ 
 		continue;
 	    }
@@ -4519,7 +4596,7 @@ attributes_for_module(Process* p, /* Process whose heap to use. */
 
 {
     Module* modp;
-    Eterm* code;
+    UWord* code;
     Eterm* hp;
     byte* ext;
     Eterm result = NIL;
@@ -4559,7 +4636,7 @@ compilation_info_for_module(Process* p, /* Process whose heap to use. */
 			    Eterm mod) /* Tagged atom for module. */
 {
     Module* modp;
-    Eterm* code;
+    UWord* code;
     Eterm* hp;
     byte* ext;
     Eterm result = NIL;
@@ -4591,8 +4668,8 @@ compilation_info_for_module(Process* p, /* Process whose heap to use. */
 /*
  * Returns a pointer to {module, function, arity}, or NULL if not found.
  */
-Eterm*
-find_function_from_pc(Eterm* pc)
+UWord *
+find_function_from_pc(UWord* pc)
 {
     Range* low = modules;
     Range* high = low + num_loaded_modules;
@@ -4604,9 +4681,9 @@ find_function_from_pc(Eterm* pc)
 	} else if (pc > mid->end) {
 	    low = mid + 1;
 	} else {
-	    Eterm** low1 = (Eterm **) (mid->start + MI_FUNCTIONS);
-	    Eterm** high1 = low1 + mid->start[MI_NUM_FUNCTIONS];
-	    Eterm** mid1;
+	    UWord** low1 = (UWord **) (mid->start + MI_FUNCTIONS);
+	    UWord** high1 = low1 + mid->start[MI_NUM_FUNCTIONS];
+	    UWord** mid1;
 
 	    while (low1 < high1) {
 		mid1 = low1 + (high1-low1) / 2;
@@ -4719,10 +4796,10 @@ code_module_md5_1(Process* p, Eterm Bin)
 
 #define WORDS_PER_FUNCTION 6
 
-static Eterm*
-make_stub(Eterm* fp, Eterm mod, Eterm func, Uint arity, Uint native, Eterm OpCode)
+static UWord*
+make_stub(UWord* fp, Eterm mod, Eterm func, Uint arity, Uint native, UWord OpCode)
 {
-    fp[0] = (Eterm) BeamOp(op_i_func_info_IaaI);
+    fp[0] = (UWord) BeamOp(op_i_func_info_IaaI);
     fp[1] = native;
     fp[2] = mod;
     fp[3] = func;
@@ -4741,14 +4818,14 @@ static byte*
 stub_copy_info(LoaderState* stp,
 	       int chunk,	/* Chunk: ATTR_CHUNK or COMPILE_CHUNK */
 	       byte* info,	/* Where to store info. */
-	       Eterm* ptr_word,	/* Where to store pointer into info. */
-	       Eterm* size_word) /* Where to store size of info. */
+	       UWord* ptr_word,	/* Where to store pointer into info. */
+	       UWord* size_word) /* Where to store size of info. */
 {
     Sint decoded_size;
     Uint size = stp->chunks[chunk].size;
     if (size != 0) {
 	memcpy(info, stp->chunks[chunk].start, size);
-	*ptr_word = (Eterm) info;
+	*ptr_word = (UWord) info;
 	decoded_size = erts_decode_ext_size(info, size, 0);
 	if (decoded_size < 0) {
  	    return 0;
@@ -4791,7 +4868,7 @@ stub_read_export_table(LoaderState* stp)
 }
 
 static void
-stub_final_touch(LoaderState* stp, Eterm* fp)
+stub_final_touch(LoaderState* stp, UWord* fp)
 {
     int i;
     int n = stp->num_exps;
@@ -4978,12 +5055,12 @@ Eterm
 erts_make_stub_module(Process* p, Eterm Mod, Eterm Beam, Eterm Info)
 {
     LoaderState state;
-    Eterm Funcs;
-    Eterm Patchlist;
+    UWord Funcs;
+    UWord Patchlist;
     Eterm* tp;
-    Eterm* code = NULL;
-    Eterm* ptrs;
-    Eterm* fp;
+    UWord* code = NULL;
+    UWord* ptrs;
+    UWord* fp;
     byte* info;
     Uint ci;
     int n;
@@ -5072,7 +5149,7 @@ erts_make_stub_module(Process* p, Eterm Mod, Eterm Beam, Eterm Info)
      * Allocate memory for the stub module.
      */
 
-    code_size = ((WORDS_PER_FUNCTION+1)*n + MI_FUNCTIONS + 2) * sizeof(Eterm);
+    code_size = ((WORDS_PER_FUNCTION+1)*n + MI_FUNCTIONS + 2) * sizeof(UWord);
     code_size += state.chunks[ATTR_CHUNK].size;
     code_size += state.chunks[COMPILE_CHUNK].size;
     code = erts_alloc_fnf(ERTS_ALC_T_CODE, code_size);
@@ -5141,7 +5218,7 @@ erts_make_stub_module(Process* p, Eterm Mod, Eterm Beam, Eterm Info)
 	 * Set the pointer and make the stub. Put a return instruction
 	 * as the body until we know what kind of trap we should put there.
 	 */
-	ptrs[i] = (Eterm) fp;
+	ptrs[i] = (UWord) fp;
 #ifdef HIPE
 	op = (Eterm) BeamOpCode(op_hipe_trap_call); /* Might be changed later. */
 #else
@@ -5154,8 +5231,8 @@ erts_make_stub_module(Process* p, Eterm Mod, Eterm Beam, Eterm Info)
      * Insert the last pointer and the int_code_end instruction.
      */
 
-    ptrs[i] = (Eterm) fp;
-    *fp++ = (Eterm) BeamOp(op_int_code_end);
+    ptrs[i] = (UWord) fp;
+    *fp++ = (UWord) BeamOp(op_int_code_end);
 
     /*
      * Copy attributes and compilation information.
@@ -5222,9 +5299,9 @@ erts_make_stub_module(Process* p, Eterm Mod, Eterm Beam, Eterm Info)
 
 #undef WORDS_PER_FUNCTION
 
-static int safe_mul(Uint a, Uint b, Uint* resp)
+static int safe_mul(UWord a, UWord b, UWord* resp)
 {
-    Uint res = a * b;
+    UWord res = a * b;
     *resp = res;
 
     if (b == 0) {
diff --git a/erts/emulator/beam/beam_load.h b/erts/emulator/beam/beam_load.h
index c17844a553..a24460d2a5 100644
--- a/erts/emulator/beam/beam_load.h
+++ b/erts/emulator/beam/beam_load.h
@@ -44,13 +44,13 @@ extern void** beam_ops;
 #endif
 
 
-extern Eterm beam_debug_apply[];
-extern Eterm* em_call_error_handler;
-extern Eterm* em_apply_bif;
-extern Eterm* em_call_traced_function;
+extern UWord beam_debug_apply[];
+extern UWord* em_call_error_handler;
+extern UWord* em_apply_bif;
+extern UWord* em_call_traced_function;
 typedef struct {
-    Eterm* start;		/* Pointer to start of module. */
-    Eterm* end;			/* Points one word beyond last function in module. */
+    UWord* start;		/* Pointer to start of module. */
+    UWord* end;			/* Points one word beyond last function in module. */
 } Range;
 
 /*
diff --git a/erts/emulator/beam/bif.c b/erts/emulator/beam/bif.c
index 0b21cf3afd..3aa258b675 100644
--- a/erts/emulator/beam/bif.c
+++ b/erts/emulator/beam/bif.c
@@ -3470,7 +3470,12 @@ BIF_RETTYPE make_fun_3(BIF_ALIST_3)
     }
     hp = HAlloc(BIF_P, 2);
     hp[0] = HEADER_EXPORT;
+#if HALFWORD_HEAP
+    /* Yes, May be misaligned, but X86_64 will fix it... */
+    memcpy(hp+1,erts_export_get_or_make_stub(BIF_ARG_1, BIF_ARG_2, (Uint) arity),sizeof(Export *));
+#else
     hp[1] = (Eterm) erts_export_get_or_make_stub(BIF_ARG_1, BIF_ARG_2, (Uint) arity);
+#endif
     BIF_RET(make_export(hp));
 }
 
@@ -3886,7 +3891,7 @@ BIF_RETTYPE hash_2(BIF_ALIST_2)
     if ((range = signed_val(BIF_ARG_2)) <= 0) {  /* [1..MAX_SMALL] */
 	BIF_ERROR(BIF_P, BADARG);
     }
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
     if (range > ((1L << 27) - 1))
 	BIF_ERROR(BIF_P, BADARG);
 #endif
@@ -3958,7 +3963,7 @@ BIF_RETTYPE phash2_2(BIF_ALIST_2)
     /*
      * Return either a small or a big. Use the heap for bigs if there is room.
      */
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
     BIF_RET(make_small(final_hash));
 #else
     if (IS_USMALL(0, final_hash)) {
@@ -4120,8 +4125,8 @@ void erts_init_bif(void)
 #else
     bif_return_trap_export.code[2] = 1;
 #endif
-    bif_return_trap_export.code[3] = (Eterm) em_apply_bif;
-    bif_return_trap_export.code[4] = (Eterm) &bif_return_trap;
+    bif_return_trap_export.code[3] = (UWord) em_apply_bif;
+    bif_return_trap_export.code[4] = (UWord) &bif_return_trap;
 
     flush_monitor_message_trap = erts_export_put(am_erlang,
 						 am_flush_monitor_message,
diff --git a/erts/emulator/beam/bif.h b/erts/emulator/beam/bif.h
index 05e9b78c28..4342a6cf4e 100644
--- a/erts/emulator/beam/bif.h
+++ b/erts/emulator/beam/bif.h
@@ -125,7 +125,7 @@ do {						\
 #define ERTS_BIF_PREP_TRAP0(Ret, Trap, Proc)		\
 do {							\
     (Proc)->arity = 0;					\
-    (Proc)->def_arg_reg[3] = (Eterm) (Trap->address);	\
+    *((UWord *) (UWord) ((Proc)->def_arg_reg + 3)) = (UWord) ((Trap)->address);	\
     (Proc)->freason = TRAP;				\
     (Ret) = THE_NON_VALUE;				\
 } while (0)
@@ -134,7 +134,8 @@ do {							\
 do {							\
     (Proc)->arity = 1;					\
     (Proc)->def_arg_reg[0] = (Eterm) (A0);		\
-    (Proc)->def_arg_reg[3] = (Eterm) ((Trap)->address);	\
+    *((UWord *) (UWord) ((Proc)->def_arg_reg + 3)) = (UWord) ((Trap)->address);	\
+    (Proc)->def_arg_reg[3] = (UWord) ((Trap)->address);	\
     (Proc)->freason = TRAP;				\
     (Ret) = THE_NON_VALUE;				\
 } while (0)
@@ -144,7 +145,8 @@ do {							\
     (Proc)->arity = 2;					\
     (Proc)->def_arg_reg[0] = (Eterm) (A0);		\
     (Proc)->def_arg_reg[1] = (Eterm) (A1);		\
-    (Proc)->def_arg_reg[3] = (Eterm) ((Trap)->address);	\
+    *((UWord *) (UWord) ((Proc)->def_arg_reg + 3)) = (UWord) ((Trap)->address);	\
+    (Proc)->def_arg_reg[3] = (UWord) ((Trap)->address);	\
     (Proc)->freason = TRAP;				\
     (Ret) = THE_NON_VALUE;				\
 } while (0)
@@ -155,14 +157,15 @@ do {							\
     (Proc)->def_arg_reg[0] = (Eterm) (A0);		\
     (Proc)->def_arg_reg[1] = (Eterm) (A1);		\
     (Proc)->def_arg_reg[2] = (Eterm) (A2);		\
-    (Proc)->def_arg_reg[3] = (Eterm) ((Trap)->address);	\
+    *((UWord *) (UWord) ((Proc)->def_arg_reg + 3)) = (UWord) ((Trap)->address);	\
+    (Proc)->def_arg_reg[3] = (UWord) ((Trap)->address);	\
     (Proc)->freason = TRAP;				\
     (Ret) = THE_NON_VALUE;				\
 } while (0)
 
 #define BIF_TRAP0(p, Trap_) do {			\
       (p)->arity = 0;					\
-      (p)->def_arg_reg[3] = (Eterm) ((Trap_)->address);	\
+      *((UWord *) (UWord) ((p)->def_arg_reg + 3)) = (UWord) ((Trap_)->address);	\
       (p)->freason = TRAP;				\
       return THE_NON_VALUE;				\
  } while(0)
@@ -170,7 +173,7 @@ do {							\
 #define BIF_TRAP1(Trap_, p, A0) do {			\
       (p)->arity = 1;					\
       (p)->def_arg_reg[0] = (A0);			\
-      (p)->def_arg_reg[3] = (Eterm) ((Trap_)->address);	\
+      *((UWord *) (UWord) ((p)->def_arg_reg + 3)) = (UWord) ((Trap_)->address);	\
       (p)->freason = TRAP;				\
       return THE_NON_VALUE;				\
  } while(0)
@@ -179,7 +182,7 @@ do {							\
       (p)->arity = 2;					\
       (p)->def_arg_reg[0] = (A0);			\
       (p)->def_arg_reg[1] = (A1);			\
-      (p)->def_arg_reg[3] = (Eterm) ((Trap_)->address);	\
+      *((UWord *) (UWord) ((p)->def_arg_reg + 3)) = (UWord) ((Trap_)->address);	\
       (p)->freason = TRAP;				\
       return THE_NON_VALUE;				\
  } while(0)
@@ -189,14 +192,14 @@ do {							\
       (p)->def_arg_reg[0] = (A0);			\
       (p)->def_arg_reg[1] = (A1);			\
       (p)->def_arg_reg[2] = (A2);			\
-      (p)->def_arg_reg[3] = (Eterm) ((Trap_)->address);	\
+      *((UWord *) (UWord) ((p)->def_arg_reg + 3)) = (UWord) ((Trap_)->address);	\
       (p)->freason = TRAP;				\
       return THE_NON_VALUE;				\
  } while(0)
 
 #define BIF_TRAP_CODE_PTR_0(p, Code_) do {		\
       (p)->arity = 0;					\
-      (p)->def_arg_reg[3] = (Eterm) (Code_);		\
+      *((UWord *) (UWord) ((p)->def_arg_reg + 3)) = (UWord) (Code_);	\
       (p)->freason = TRAP;				\
       return THE_NON_VALUE;				\
  } while(0)
diff --git a/erts/emulator/beam/big.c b/erts/emulator/beam/big.c
index 03c88da8c6..b8464ded28 100644
--- a/erts/emulator/beam/big.c
+++ b/erts/emulator/beam/big.c
@@ -1459,7 +1459,31 @@ Eterm uint_to_big(Uint x, Eterm *y)
     BIG_DIGIT(y, 0) = x;
     return make_big(y);
 }
+/*
+** convert UWord to bigint
+** (must only be used if x is to big to be stored as a small)
+** Allocation is tricky, the heap need has to be calculated
+** with the macro BIG_UWORD_HEAP_SIZE(x)
+*/
 
+Eterm uword_to_big(UWord x, Eterm *y)
+{
+#if HALFWORD_HEAP
+    Uint upper = x >> 32;
+    Uint lower = x & 0xFFFFFFFFUL;
+    if (upper == 0) {
+	*y = make_pos_bignum_header(1);
+    } else {
+	*y = make_pos_bignum_header(2);
+	BIG_DIGIT(y, 1) = upper;
+    }
+    BIG_DIGIT(y, 0) = lower;
+#else
+    *y = make_pos_bignum_header(1);
+    BIG_DIGIT(y, 0) = x;
+#endif
+    return make_big(y);
+}
 
 /*
 ** convert signed int to bigint
diff --git a/erts/emulator/beam/big.h b/erts/emulator/beam/big.h
index b8e38d482c..d74da7cef7 100644
--- a/erts/emulator/beam/big.h
+++ b/erts/emulator/beam/big.h
@@ -34,7 +34,7 @@
 
 typedef Uint     ErtsDigit;
 
-#if (SIZEOF_VOID_P == 4) && defined(SIZEOF_LONG_LONG) && (SIZEOF_LONG_LONG == 8)
+#if ((SIZEOF_VOID_P == 4) || HALFWORD_HEAP) && defined(SIZEOF_LONG_LONG) && (SIZEOF_LONG_LONG == 8)
 /* Assume 32-bit machine with long long support */
 typedef Uint64   ErtsDoubleDigit;
 typedef Uint16   ErtsHalfDigit;
@@ -58,7 +58,7 @@ typedef Uint32   ErtsHalfDigit;
 
 typedef Uint  dsize_t;	 /* Vector size type */
 
-#define D_EXP (SIZEOF_VOID_P*8)
+#define D_EXP (ERTS_SIZEOF_ETERM*8)
 #define D_MASK     ((ErtsDigit)(-1))      /* D_BASE-1 */
 
 /* macros for bignum objects */
@@ -88,6 +88,12 @@ typedef Uint  dsize_t;	 /* Vector size type */
 
 #define BIG_UINT_HEAP_SIZE (1 + 1)	/* always, since sizeof(Uint) <= sizeof(Eterm) */
 
+#if HALFWORD_HEAP
+#define BIG_UWORD_HEAP_SIZE(UW) (((UW) >> (sizeof(Uint) * 8)) ? 3 : 2)
+#else
+#define BIG_UWORD_HEAP_SIZE(UW) BIG_UINT_HEAP_SIZE
+#endif
+
 #ifdef ARCH_32
 
 #define ERTS_UINT64_BIG_HEAP_SIZE__(X) \
@@ -136,6 +142,7 @@ int big_ucomp (Eterm, Eterm);
 int big_to_double(Eterm x, double* resp);
 Eterm small_to_big(Sint, Eterm*);
 Eterm uint_to_big(Uint, Eterm*);
+Eterm uword_to_big(UWord, Eterm*);
 Eterm erts_make_integer(Uint, Process *);
 
 dsize_t big_bytes(Eterm);
diff --git a/erts/emulator/beam/break.c b/erts/emulator/beam/break.c
index cc69977b79..0b7afcac02 100644
--- a/erts/emulator/beam/break.c
+++ b/erts/emulator/beam/break.c
@@ -381,7 +381,7 @@ loaded(int to, void *to_arg)
     int i;
     int old = 0;
     int cur = 0;
-    Eterm* code;
+    UWord* code;
 
     /*
      * Calculate and print totals.
diff --git a/erts/emulator/beam/dist.c b/erts/emulator/beam/dist.c
index 64ab88fc45..eb8fb02904 100644
--- a/erts/emulator/beam/dist.c
+++ b/erts/emulator/beam/dist.c
@@ -1620,9 +1620,9 @@ dsig_send(ErtsDSigData *dsdp, Eterm ctl, Eterm msg, int force_busy)
 	 */
 
 	data_size >>= (10-4);
-#if defined(ARCH_64)
+#if defined(ARCH_64) && !HALFWORD_HEAP
 	data_size &= 0x003fffffffffffff;
-#elif defined(ARCH_32)
+#elif defined(ARCH_32) || HALFWORD_HEAP
 	data_size &= 0x003fffff;
 #else
 #       error "Ohh come on ... !?!"
@@ -1706,9 +1706,9 @@ dist_port_commandv(Port *prt, ErtsDistOutputBuf *obuf)
 }
 
 
-#if defined(ARCH_64)
+#if defined(ARCH_64) && !HALFWORD_HEAP
 #define ERTS_PORT_REDS_MASK__ 0x003fffffffffffffL
-#elif defined(ARCH_32)
+#elif defined(ARCH_32) || HALFWORD_HEAP
 #define ERTS_PORT_REDS_MASK__ 0x003fffff
 #else
 #  error "Ohh come on ... !?!"
diff --git a/erts/emulator/beam/erl_alloc.c b/erts/emulator/beam/erl_alloc.c
index 4403a19f4a..9c5210fed6 100644
--- a/erts/emulator/beam/erl_alloc.c
+++ b/erts/emulator/beam/erl_alloc.c
@@ -220,7 +220,7 @@ set_default_ll_alloc_opts(struct au_init *ip)
     ip->init.util.ramv		= 0;
     ip->init.util.mmsbc		= 0;
     ip->init.util.mmmbc		= 0;
-    ip->init.util.sbct		= ~((Uint) 0);
+    ip->init.util.sbct		= ~((UWord) 0);
     ip->init.util.name_prefix	= "ll_";
     ip->init.util.alloc_no	= ERTS_ALC_A_LONG_LIVED;
 #ifndef SMALL_MEMORY
@@ -394,7 +394,7 @@ static void init_thr_ix(int static_ixs);
 void
 erts_alloc_init(int *argc, char **argv, ErtsAllocInitOpts *eaiop)
 {
-    Uint extra_block_size = 0;
+    UWord extra_block_size = 0;
     int i;
     erts_alc_hndl_args_init_t init = {
 	0,
@@ -542,7 +542,6 @@ erts_alloc_init(int *argc, char **argv, ErtsAllocInitOpts *eaiop)
 
     sys_alloc_opt(SYS_ALLOC_OPT_TRIM_THRESHOLD, init.trim_threshold);
     sys_alloc_opt(SYS_ALLOC_OPT_TOP_PAD, init.top_pad);
-
     if (erts_allctrs_info[ERTS_FIX_CORE_ALLOCATOR].enabled)
 	erts_fix_core_allocator_ix = ERTS_FIX_CORE_ALLOCATOR;
     else
@@ -719,8 +718,8 @@ start_au_allocator(ErtsAlcType_t alctr_n,
 		     init->init.util.name_prefix);
 	tspec->allctr = (Allctr_t **) states;
 	states = ((char *) states) + sizeof(Allctr_t *) * (tspec->size + 1);
-	states = ((((Uint) states) & ERTS_CACHE_LINE_MASK)
-		  ? (void *) ((((Uint) states) & ~ERTS_CACHE_LINE_MASK)
+	states = ((((UWord) states) & ERTS_CACHE_LINE_MASK)
+		  ? (void *) ((((UWord) states) & ~ERTS_CACHE_LINE_MASK)
 			      + ERTS_CACHE_LINE_SIZE)
 		  : (void *) states);
 	tspec->allctr[0] = init->thr_spec > 0 ? (Allctr_t *) state : (Allctr_t *) NULL;
@@ -2948,9 +2947,9 @@ unsigned long erts_alc_test(unsigned long op,
 #endif
 
 
-#define FENCE_SZ		(3*sizeof(Uint))
+#define FENCE_SZ		(3*sizeof(UWord))
 
-#ifdef ARCH_64
+#if defined(ARCH_64)
 #define FENCE_PATTERN 0xABCDEF97ABCDEF97
 #else
 #define FENCE_PATTERN 0xABCDEF97
@@ -2960,7 +2959,7 @@ unsigned long erts_alc_test(unsigned long op,
 #define TYPE_PATTERN_SHIFT 16
 
 #define FIXED_FENCE_PATTERN_MASK \
-  (~((Uint) (TYPE_PATTERN_MASK << TYPE_PATTERN_SHIFT)))
+  (~((UWord) (TYPE_PATTERN_MASK << TYPE_PATTERN_SHIFT)))
 #define FIXED_FENCE_PATTERN \
   (FENCE_PATTERN & FIXED_FENCE_PATTERN_MASK)
 
@@ -2971,21 +2970,60 @@ unsigned long erts_alc_test(unsigned long op,
   (((P) >> TYPE_PATTERN_SHIFT) & TYPE_PATTERN_MASK)
 
 
+#ifdef  ERTS_ALLOC_UTIL_HARD_DEBUG
+static void *check_memory_fence(void *ptr, Uint *size, ErtsAlcType_t n, int func);
+
+void check_allocated_block( Uint type, void *blk)
+{
+    Uint dummy;
+    check_memory_fence(blk, &dummy, ERTS_ALC_T2N(type), ERTS_ALC_O_FREE);
+}
+
+void check_allocators(void)
+{
+    int i;
+    if (!erts_initialized)
+	return;
+    for (i = ERTS_ALC_A_MIN; i <= ERTS_ALC_A_MAX; ++i) {
+	if (erts_allctrs_info[i].alloc_util) {
+	    ErtsAllocatorFunctions_t *real_af = (ErtsAllocatorFunctions_t *) erts_allctrs[i].extra;
+	    Allctr_t *allctr = real_af->extra;
+	    Carrier_t *ct;
+#ifdef USE_THREADS
+	if (allctr->thread_safe)
+	    erts_mtx_lock(&allctr->mutex);
+#endif
+
+	    if (allctr->check_mbc) {
+		for (ct = allctr->mbc_list.first; ct; ct = ct->next) {
+		    fprintf(stderr,"Checking allocator %d\r\n",i);
+		    allctr->check_mbc(allctr,ct);
+		}
+	    }
+#ifdef USE_THREADS
+	if (allctr->thread_safe)
+	    erts_mtx_unlock(&allctr->mutex);
+#endif
+	}
+    }
+}
+#endif
+
 static void *
 set_memory_fence(void *ptr, Uint sz, ErtsAlcType_t n)
 {
-    Uint *ui_ptr;
-    Uint pattern;
+    UWord *ui_ptr;
+    UWord pattern;
 
     if (!ptr)
 	return NULL;
 
-    ui_ptr = (Uint *) ptr;
+    ui_ptr = (UWord *) ptr;
     pattern = MK_PATTERN(n);
     
     *(ui_ptr++) = sz;
     *(ui_ptr++) = pattern;
-    memcpy((void *) (((char *) ui_ptr)+sz), (void *) &pattern, sizeof(Uint));
+    memcpy((void *) (((char *) ui_ptr)+sz), (void *) &pattern, sizeof(UWord));
 
     return (void *) ui_ptr;
 }
@@ -2995,14 +3033,14 @@ check_memory_fence(void *ptr, Uint *size, ErtsAlcType_t n, int func)
 {
     Uint sz;
     Uint found_type;
-    Uint pre_pattern;
-    Uint post_pattern;
-    Uint *ui_ptr;
+    UWord pre_pattern;
+    UWord post_pattern;
+    UWord *ui_ptr;
 
     if (!ptr)
 	return NULL;
 
-    ui_ptr = (Uint *) ptr;
+    ui_ptr = (UWord *) ptr;
     pre_pattern = *(--ui_ptr);
     *size = sz = *(--ui_ptr);
 
@@ -3015,7 +3053,7 @@ check_memory_fence(void *ptr, Uint *size, ErtsAlcType_t n, int func)
 		     (unsigned long) ptr);
     }
 
-    memcpy((void *) &post_pattern, (void *) (((char *)ptr)+sz), sizeof(Uint));
+    memcpy((void *) &post_pattern, (void *) (((char *)ptr)+sz), sizeof(UWord));
 
     if (post_pattern != MK_PATTERN(n)
 	|| pre_pattern != post_pattern) {
diff --git a/erts/emulator/beam/erl_alloc.h b/erts/emulator/beam/erl_alloc.h
index e7a203002f..3f090abdc2 100644
--- a/erts/emulator/beam/erl_alloc.h
+++ b/erts/emulator/beam/erl_alloc.h
@@ -553,7 +553,7 @@ NAME##_free(TYPE *p)							\
 }
 
 #ifdef DEBUG
-#define ERTS_ALC_DBG_BLK_SZ(PTR) (*(((Uint *) (PTR)) - 2))
+#define ERTS_ALC_DBG_BLK_SZ(PTR) (*(((UWord *) (PTR)) - 2))
 #endif /* #ifdef DEBUG */
 
 #undef ERTS_ALC_INLINE
diff --git a/erts/emulator/beam/erl_alloc.types b/erts/emulator/beam/erl_alloc.types
index f808d3b66e..2d7077227f 100644
--- a/erts/emulator/beam/erl_alloc.types
+++ b/erts/emulator/beam/erl_alloc.types
@@ -138,6 +138,7 @@ type	PEND_SUSPEND	SHORT_LIVED	PROCESSES	pending_suspend
 type	PROC_LIST	SHORT_LIVED	PROCESSES	proc_list
 type	FUN_ENTRY	FIXED_SIZE	CODE		fun_entry
 type	ATOM_TXT	LONG_LIVED	ATOM		atom_text
+type 	BEAM_REGISTER	EHEAP		PROCESSES	beam_register
 type	HEAP		EHEAP		PROCESSES	heap
 type	OLD_HEAP	EHEAP		PROCESSES	old_heap
 type	HEAP_FRAG	EHEAP		PROCESSES	heap_frag
diff --git a/erts/emulator/beam/erl_alloc_util.c b/erts/emulator/beam/erl_alloc_util.c
index 8b71b51912..022a2829f6 100644
--- a/erts/emulator/beam/erl_alloc_util.c
+++ b/erts/emulator/beam/erl_alloc_util.c
@@ -69,14 +69,14 @@ static int initialized = 0;
 
 #if HAVE_ERTS_MSEG
 
-#define INV_MSEG_UNIT_MASK	((Uint) (mseg_unit_size - 1))
+#define INV_MSEG_UNIT_MASK	((UWord) (mseg_unit_size - 1))
 #define MSEG_UNIT_MASK		(~INV_MSEG_UNIT_MASK)
 #define MSEG_UNIT_FLOOR(X)	((X) & MSEG_UNIT_MASK)
 #define MSEG_UNIT_CEILING(X)	MSEG_UNIT_FLOOR((X) + INV_MSEG_UNIT_MASK)
 
 #endif
 
-#define INV_SYS_ALLOC_CARRIER_MASK	((Uint) (sys_alloc_carrier_size - 1))
+#define INV_SYS_ALLOC_CARRIER_MASK	((UWord) (sys_alloc_carrier_size - 1))
 #define SYS_ALLOC_CARRIER_MASK		(~INV_SYS_ALLOC_CARRIER_MASK)
 #define SYS_ALLOC_CARRIER_FLOOR(X)	((X) & SYS_ALLOC_CARRIER_MASK)
 #define SYS_ALLOC_CARRIER_CEILING(X) \
@@ -85,7 +85,7 @@ static int initialized = 0;
 #undef ASSERT
 #define ASSERT ASSERT_EXPR
 
-#define ERTS_ALCU_FLG_FAIL_REALLOC_MOVE ((Uint) 1)
+#define ERTS_ALCU_FLG_FAIL_REALLOC_MOVE ((UWord) 1)
 
 #if 0
 /* Can be useful for debugging */
@@ -114,12 +114,12 @@ static Uint mseg_unit_size;
 
 /* Blocks ... */
 
-#define SBC_BLK_FTR_FLG		(((Uint) 1) << 0)
-#define UNUSED1_BLK_FTR_FLG	(((Uint) 1) << 1)
-#define UNUSED2_BLK_FTR_FLG	(((Uint) 1) << 2)
+#define SBC_BLK_FTR_FLG		(((UWord) 1) << 0)
+#define UNUSED1_BLK_FTR_FLG	(((UWord) 1) << 1)
+#define UNUSED2_BLK_FTR_FLG	(((UWord) 1) << 2)
 
 #define ABLK_HDR_SZ (sizeof(Block_t))
-#define FBLK_FTR_SZ (sizeof(Uint))
+#define FBLK_FTR_SZ (sizeof(UWord))
 
 #define UMEMSZ2BLKSZ(AP, SZ)						\
   (ABLK_HDR_SZ + (SZ) <= (AP)->min_block_size				\
@@ -130,14 +130,14 @@ static Uint mseg_unit_size;
 #define BLK2UMEM(P) ((void *)    (((char *) (P)) + ABLK_HDR_SZ))
 
 #define PREV_BLK_SZ(B) \
-  ((Uint) (*(((Uint *) (B)) - 1) & SZ_MASK))
+  ((UWord) (*(((UWord *) (B)) - 1) & SZ_MASK))
 
 #define SET_BLK_SZ_FTR(B, SZ) \
-  (*((Uint *) (((char *) (B)) + (SZ) - sizeof(Uint))) = (SZ))
+  (*((UWord *) (((char *) (B)) + (SZ) - sizeof(UWord))) = (SZ))
 
-#define THIS_FREE_BLK_HDR_FLG 	(((Uint) 1) << 0)
-#define PREV_FREE_BLK_HDR_FLG 	(((Uint) 1) << 1)
-#define LAST_BLK_HDR_FLG 	(((Uint) 1) << 2)
+#define THIS_FREE_BLK_HDR_FLG 	(((UWord) 1) << 0)
+#define PREV_FREE_BLK_HDR_FLG 	(((UWord) 1) << 1)
+#define LAST_BLK_HDR_FLG 	(((UWord) 1) << 2)
 
 #define SET_BLK_SZ(B, SZ) \
   (ASSERT(((SZ) & FLG_MASK) == 0), \
@@ -156,11 +156,11 @@ static Uint mseg_unit_size;
   (*((Block_t *) (B)) &= ~LAST_BLK_HDR_FLG)
 
 #define SBH_THIS_FREE		THIS_FREE_BLK_HDR_FLG
-#define SBH_THIS_ALLOCED	((Uint) 0)
+#define SBH_THIS_ALLOCED	((UWord) 0)
 #define SBH_PREV_FREE		PREV_FREE_BLK_HDR_FLG
-#define SBH_PREV_ALLOCED	((Uint) 0)
+#define SBH_PREV_ALLOCED	((UWord) 0)
 #define SBH_LAST_BLK		LAST_BLK_HDR_FLG
-#define SBH_NOT_LAST_BLK	((Uint) 0)
+#define SBH_NOT_LAST_BLK	((UWord) 0)
 
 #define SET_BLK_HDR(B, Sz, F) \
   (ASSERT(((Sz) & FLG_MASK) == 0), *((Block_t *) (B)) = ((Sz) | (F)))
@@ -200,7 +200,7 @@ static Uint mseg_unit_size;
   ((FTR) = 0)
 
 #define IS_SBC_BLK(B) \
-  (IS_PREV_BLK_FREE((B)) && (((Uint *) (B))[-1] & SBC_BLK_FTR_FLG))
+  (IS_PREV_BLK_FREE((B)) && (((UWord *) (B))[-1] & SBC_BLK_FTR_FLG))
 #define IS_MBC_BLK(B) \
   (!IS_SBC_BLK((B)))
 
@@ -211,8 +211,8 @@ static Uint mseg_unit_size;
 
 /* Carriers ... */
 
-#define MSEG_CARRIER_HDR_FLAG		(((Uint) 1) << 0)
-#define SBC_CARRIER_HDR_FLAG		(((Uint) 1) << 1)
+#define MSEG_CARRIER_HDR_FLAG		(((UWord) 1) << 0)
+#define SBC_CARRIER_HDR_FLAG		(((UWord) 1) << 1)
 
 #define SCH_SYS_ALLOC			0
 #define SCH_MSEG			MSEG_CARRIER_HDR_FLAG
@@ -407,18 +407,18 @@ do {									\
 
 /* Debug stuff... */
 #ifdef DEBUG
-static Uint carrier_alignment;
+static UWord carrier_alignment;
 #define DEBUG_SAVE_ALIGNMENT(C)						\
 do {									\
-    Uint algnmnt__ = sizeof(Unit_t) - (((Uint) (C)) % sizeof(Unit_t));	\
+    UWord algnmnt__ = sizeof(Unit_t) - (((UWord) (C)) % sizeof(Unit_t));	\
     carrier_alignment = MIN(carrier_alignment, algnmnt__);		\
-    ASSERT(((Uint) (C)) % sizeof(Uint) == 0);				\
+    ASSERT(((UWord) (C)) % sizeof(UWord) == 0);				\
 } while (0)
 #define DEBUG_CHECK_ALIGNMENT(P)					\
 do {									\
-    ASSERT(sizeof(Unit_t) - (((Uint) (P)) % sizeof(Unit_t))		\
+    ASSERT(sizeof(Unit_t) - (((UWord) (P)) % sizeof(Unit_t))		\
 	   >= carrier_alignment);					\
-    ASSERT(((Uint) (P)) % sizeof(Uint) == 0);				\
+    ASSERT(((UWord) (P)) % sizeof(UWord) == 0);				\
 } while (0)
 
 #else
@@ -610,7 +610,7 @@ unlink_carrier(CarrierList_t *cl, Carrier_t *crr)
 }
 
 
-static Block_t *create_carrier(Allctr_t *, Uint, Uint);
+static Block_t *create_carrier(Allctr_t *, Uint, UWord);
 static void destroy_carrier(Allctr_t *, Block_t *);
 
 /* Multi block carrier alloc/realloc/free ... */
@@ -662,14 +662,14 @@ static ERTS_INLINE void
 mbc_alloc_finalize(Allctr_t *allctr,
 		   Block_t *blk,
 		   Uint org_blk_sz,
-		   Uint flags,
+		   UWord flags,
 		   Uint want_blk_sz,
 		   int valid_blk_info)
 {
     Uint blk_sz;
     Uint nxt_blk_sz;
     Block_t *nxt_blk;
-    Uint prev_free_flg = flags & PREV_FREE_BLK_HDR_FLG;
+    UWord prev_free_flg = flags & PREV_FREE_BLK_HDR_FLG;
 
     ASSERT(org_blk_sz >= want_blk_sz);
     ASSERT(blk);
@@ -859,7 +859,7 @@ mbc_free(Allctr_t *allctr, void *p)
 }
 
 static void *
-mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint flgs)
+mbc_realloc(Allctr_t *allctr, void *p, Uint size, UWord flgs)
 {
     void *new_p;
     Uint old_blk_sz;
@@ -1150,7 +1150,7 @@ mbc_realloc(Allctr_t *allctr, void *p, Uint size, Uint flgs)
 	}
 	else {
 	    Uint new_blk_sz;
-	    Uint new_blk_flgs;
+	    UWord new_blk_flgs;
 	    Uint prev_blk_sz;
 	    Uint blk_cpy_sz;
 
@@ -1245,7 +1245,7 @@ do {									\
 
 
 static Block_t *
-create_carrier(Allctr_t *allctr, Uint umem_sz, Uint flags)
+create_carrier(Allctr_t *allctr, Uint umem_sz, UWord flags)
 {
     Block_t *blk;
     Carrier_t *crr;
@@ -1289,8 +1289,8 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, Uint flags)
 	if (crr_sz < allctr->mbc_header_size + blk_sz)
 	    crr_sz = allctr->mbc_header_size + blk_sz;
 #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
-	if (sizeof(Unit_t) == sizeof(Uint))
-	    crr_sz += sizeof(Uint);
+	if (sizeof(Unit_t) == sizeof(UWord))
+	    crr_sz += sizeof(UWord);
 #endif
     }
     crr_sz = MSEG_UNIT_CEILING(crr_sz);
@@ -1330,8 +1330,8 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, Uint flags)
 	    && bcrr_sz < allctr->smallest_mbc_size)
 	    bcrr_sz = allctr->smallest_mbc_size;
 #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
-	if (sizeof(Unit_t) == sizeof(Uint))
-	    bcrr_sz += sizeof(Uint);
+	if (sizeof(Unit_t) == sizeof(UWord))
+	    bcrr_sz += sizeof(UWord);
 #endif
 
     }
@@ -1366,7 +1366,7 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, Uint flags)
 
 	blk = SBC2BLK(allctr, crr);
 
-	SET_SBC_BLK_FTR(((Uint *) blk)[-1]);
+	SET_SBC_BLK_FTR(((UWord *) blk)[-1]);
 	SET_BLK_HDR(blk, blk_sz, SBH_THIS_ALLOCED|SBH_PREV_FREE|SBH_LAST_BLK);
 
 	link_carrier(&allctr->sbc_list, crr);
@@ -1385,13 +1385,13 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, Uint flags)
 	blk = MBC2FBLK(allctr, crr);
 
 #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
-	if (sizeof(Unit_t) == sizeof(Uint))
-	    crr_sz -= sizeof(Uint);
+	if (sizeof(Unit_t) == sizeof(UWord))
+	    crr_sz -= sizeof(UWord);
 #endif
 
 	blk_sz = UNIT_FLOOR(crr_sz - allctr->mbc_header_size);
 
-	SET_MBC_BLK_FTR(((Uint *) blk)[-1]);
+	SET_MBC_BLK_FTR(((UWord *) blk)[-1]);
 	SET_BLK_HDR(blk, blk_sz, SBH_THIS_FREE|SBH_PREV_FREE|SBH_LAST_BLK);
 
 #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
@@ -1406,13 +1406,13 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, Uint flags)
 	link_carrier(&allctr->mbc_list, crr);
 
 #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
-	if (sizeof(Unit_t) == sizeof(Uint))
-	    crr_sz += sizeof(Uint);
+	if (sizeof(Unit_t) == sizeof(UWord))
+	    crr_sz += sizeof(UWord);
 #endif
 	CHECK_1BLK_CARRIER(allctr, 0, is_mseg, crr, crr_sz, blk, blk_sz);
 #ifdef ERTS_ALLOC_UTIL_HARD_DEBUG
-	if (sizeof(Unit_t) == sizeof(Uint))
-	    crr_sz -= sizeof(Uint);
+	if (sizeof(Unit_t) == sizeof(UWord))
+	    crr_sz -= sizeof(UWord);
 #endif
 	if (allctr->creating_mbc)
 	    (*allctr->creating_mbc)(allctr, crr);
@@ -1424,11 +1424,12 @@ create_carrier(Allctr_t *allctr, Uint umem_sz, Uint flags)
 }
 
 static Block_t *
-resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, Uint flags)
+resize_carrier(Allctr_t *allctr, Block_t *old_blk, Uint umem_sz, UWord flags)
 {
     Block_t *new_blk;
     Carrier_t *new_crr, *old_crr;
-    Uint create_flags, old_crr_sz, old_blk_sz, new_blk_sz, new_crr_sz;
+    UWord create_flags;
+    Uint old_crr_sz, old_blk_sz, new_blk_sz, new_crr_sz;
     Uint new_bcrr_sz;
 
     if (flags & CFLG_MBC) {
@@ -2605,16 +2606,16 @@ erts_alcu_alloc_thr_pref(ErtsAlcType_t type, void *extra, Uint size)
     Allctr_t *allctr;
     void *res;
 
-    ASSERT(sizeof(Uint) == sizeof(Allctr_t *));
+    ASSERT(sizeof(UWord) == sizeof(Allctr_t *));
     ASSERT(ix > 0);
     if (ix >= tspec->size)
 	ix = (ix % (tspec->size - 1)) + 1;
     allctr = tspec->allctr[ix];
     erts_mtx_lock(&allctr->mutex);
-    res = do_erts_alcu_alloc(type, allctr, size + sizeof(Uint));
+    res = do_erts_alcu_alloc(type, allctr, size + sizeof(UWord));
     if (res) {
 	*((Allctr_t **) res) = allctr;
-	res = (void *) (((char *) res) + sizeof(Uint));
+	res = (void *) (((char *) res) + sizeof(UWord));
     }
     erts_mtx_unlock(&allctr->mutex);
     DEBUG_CHECK_ALIGNMENT(res);
@@ -2692,7 +2693,7 @@ void
 erts_alcu_free_thr_pref(ErtsAlcType_t type, void *unused, void *p)
 {
     if (p) {
-	void *ptr = (void *) (((char *) p) - sizeof(Uint));
+	void *ptr = (void *) (((char *) p) - sizeof(UWord));
 	Allctr_t *allctr = *((Allctr_t **) ptr);
 	erts_mtx_lock(&allctr->mutex);
 	do_erts_alcu_free(type, allctr, ptr);
@@ -2709,7 +2710,7 @@ do_erts_alcu_realloc(ErtsAlcType_t type,
 		     void *extra,
 		     void *p,
 		     Uint size,
-		     Uint flgs)
+		     UWord flgs)
 {
     Allctr_t *allctr = (Allctr_t *) extra; 
     Block_t *blk;
@@ -2981,7 +2982,7 @@ erts_alcu_realloc_thr_pref(ErtsAlcType_t type, void *extra, void *p, Uint size)
     if (!p)
 	return erts_alcu_alloc_thr_pref(type, extra, size);
 
-    ptr = (void *) (((char *) p) - sizeof(Uint));
+    ptr = (void *) (((char *) p) - sizeof(UWord));
     used_allctr = *((Allctr_t **) ptr);
 
     ix = erts_alc_get_thr_ix();
@@ -2995,32 +2996,32 @@ erts_alcu_realloc_thr_pref(ErtsAlcType_t type, void *extra, void *p, Uint size)
     res = do_erts_alcu_realloc(type,
 			       used_allctr,
 			       ptr,
-			       size + sizeof(Uint),
+			       size + sizeof(UWord),
 			       (pref_allctr != used_allctr
 				? ERTS_ALCU_FLG_FAIL_REALLOC_MOVE
 				: 0));
     erts_mtx_unlock(&used_allctr->mutex);
     if (res) {
 	ASSERT(used_allctr == *((Allctr_t **) res));
-	res = (void *) (((char *) res) + sizeof(Uint));
+	res = (void *) (((char *) res) + sizeof(UWord));
 	DEBUG_CHECK_ALIGNMENT(res);
     }
     else {
 	erts_mtx_lock(&pref_allctr->mutex);
-	res = do_erts_alcu_alloc(type, pref_allctr, size + sizeof(Uint));
+	res = do_erts_alcu_alloc(type, pref_allctr, size + sizeof(UWord));
 	erts_mtx_unlock(&pref_allctr->mutex);
 	if (res) {
 	    Block_t *blk;
 	    size_t cpy_size;
 
 	    *((Allctr_t **) res) = pref_allctr;
-	    res = (void *) (((char *) res) + sizeof(Uint));
+	    res = (void *) (((char *) res) + sizeof(UWord));
 
 	    DEBUG_CHECK_ALIGNMENT(res);
 
 	    erts_mtx_lock(&used_allctr->mutex);
 	    blk = UMEM2BLK(ptr);
-	    cpy_size = BLK_SZ(blk) - ABLK_HDR_SZ - sizeof(Uint);
+	    cpy_size = BLK_SZ(blk) - ABLK_HDR_SZ - sizeof(UWord);
 	    if (cpy_size > size)
 		cpy_size = size;
 	    sys_memcpy(res, p, cpy_size);
@@ -3045,7 +3046,7 @@ erts_alcu_realloc_mv_thr_pref(ErtsAlcType_t type, void *extra,
     if (!p)
 	return erts_alcu_alloc_thr_pref(type, extra, size);
 
-    ptr = (void *) (((char *) p) - sizeof(Uint));
+    ptr = (void *) (((char *) p) - sizeof(UWord));
     used_allctr = *((Allctr_t **) ptr);
 
     ix = erts_alc_get_thr_ix();
@@ -3056,7 +3057,7 @@ erts_alcu_realloc_mv_thr_pref(ErtsAlcType_t type, void *extra,
     ASSERT(used_allctr && pref_allctr);
 
     erts_mtx_lock(&pref_allctr->mutex);
-    res = do_erts_alcu_alloc(type, pref_allctr, size + sizeof(Uint));
+    res = do_erts_alcu_alloc(type, pref_allctr, size + sizeof(UWord));
     if (!res) {
 	erts_mtx_unlock(&pref_allctr->mutex);
 	res = erts_alcu_realloc_thr_pref(type, extra, p, size);
@@ -3067,7 +3068,7 @@ erts_alcu_realloc_mv_thr_pref(ErtsAlcType_t type, void *extra,
 	Allctr_t *allctr;
 
 	*((Allctr_t **) res) = pref_allctr;
-	res = (void *) (((char *) res) + sizeof(Uint));
+	res = (void *) (((char *) res) + sizeof(UWord));
 
 	DEBUG_CHECK_ALIGNMENT(res);
 
@@ -3080,7 +3081,7 @@ erts_alcu_realloc_mv_thr_pref(ErtsAlcType_t type, void *extra,
 	}
 
 	blk = UMEM2BLK(ptr);
-	cpy_size = BLK_SZ(blk) - ABLK_HDR_SZ - sizeof(Uint);
+	cpy_size = BLK_SZ(blk) - ABLK_HDR_SZ - sizeof(UWord);
 	if (cpy_size > size)
 	    cpy_size = size;
 	sys_memcpy(res, p, cpy_size);
@@ -3157,11 +3158,11 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init)
     if (allctr->min_block_size < ABLK_HDR_SZ)
 	goto error;
     allctr->min_block_size		= UNIT_CEILING(allctr->min_block_size
-						       + sizeof(Uint));
+						       + sizeof(UWord));
 
 #if HAVE_ERTS_MSEG
-    if (allctr->mseg_opt.abs_shrink_th > ~((Uint) 0) / 100)
-	allctr->mseg_opt.abs_shrink_th = ~((Uint) 0) / 100;
+    if (allctr->mseg_opt.abs_shrink_th > ~((UWord) 0) / 100)
+	allctr->mseg_opt.abs_shrink_th = ~((UWord) 0) / 100;
 #endif
 
 #ifdef USE_THREADS
@@ -3201,15 +3202,15 @@ erts_alcu_start(Allctr_t *allctr, AllctrInit_t *init)
 	allctr->mbc_header_size = (UNIT_CEILING(allctr->mbc_header_size
 						+ FBLK_FTR_SZ
 						+ ABLK_HDR_SZ
-						+ sizeof(Uint))
+						+ sizeof(UWord))
 				   - ABLK_HDR_SZ
-				   - sizeof(Uint));
+				   - sizeof(UWord));
 	allctr->sbc_header_size = (UNIT_CEILING(sizeof(Carrier_t)
 						+ FBLK_FTR_SZ
 						+ ABLK_HDR_SZ
-						+ sizeof(Uint))
+						+ sizeof(UWord))
 				   - ABLK_HDR_SZ
-				   - sizeof(Uint));
+				   - sizeof(UWord));
     }
     else
 #endif
@@ -3437,7 +3438,7 @@ check_blk_carrier(Allctr_t *allctr, Block_t *iblk)
 	    is_free_blk = (int) IS_FREE_BLK(blk);
 	    if(is_free_blk) {
 		if (IS_NOT_LAST_BLK(blk))
-		    ASSERT(*((Uint *) (((char *) blk)+blk_sz-sizeof(Uint)))
+		    ASSERT(*((UWord *) (((char *) blk)+blk_sz-sizeof(UWord)))
 			   == blk_sz);
 	    }
 
@@ -3445,7 +3446,7 @@ check_blk_carrier(Allctr_t *allctr, Block_t *iblk)
 		(*allctr->check_block)(allctr, blk, (int) is_free_blk);
 
 	    if (IS_LAST_BLK(blk)) {
-		carrier_end = ((char *) NXT_BLK(blk)) + sizeof(Uint);
+		carrier_end = ((char *) NXT_BLK(blk)) + sizeof(UWord);
 		mbc = *((Carrier_t **) NXT_BLK(blk));
 		prev_blk = NULL;
 		blk = MBC2FBLK(allctr, mbc);
@@ -3461,7 +3462,7 @@ check_blk_carrier(Allctr_t *allctr, Block_t *iblk)
 	ASSERT((((char *) mbc)
 		+ allctr->mbc_header_size
 		+ tot_blk_sz
-		+ sizeof(Uint)) == carrier_end);
+		+ sizeof(UWord)) == carrier_end);
 	ASSERT(((char *) mbc) + CARRIER_SZ(mbc) == carrier_end);
 
 	if (allctr->check_mbc)
diff --git a/erts/emulator/beam/erl_alloc_util.h b/erts/emulator/beam/erl_alloc_util.h
index 10b11661e6..e805883ddc 100644
--- a/erts/emulator/beam/erl_alloc_util.h
+++ b/erts/emulator/beam/erl_alloc_util.h
@@ -27,8 +27,8 @@
 typedef struct Allctr_t_ Allctr_t;
 
 typedef struct {
-    Uint ycs;
-    Uint mmc;
+    UWord ycs;
+    UWord mmc;
 } AlcUInit_t;
 
 typedef struct {
@@ -38,22 +38,22 @@ typedef struct {
     int tspec;
     int tpref;
     int ramv;
-    Uint sbct;
-    Uint asbcst;
-    Uint rsbcst;
-    Uint rsbcmt;
-    Uint rmbcmt;
-    Uint mmbcs;
-    Uint mmsbc;
-    Uint mmmbc;
-    Uint lmbcs;
-    Uint smbcs;
-    Uint mbcgs;
+    UWord sbct;
+    UWord asbcst;
+    UWord rsbcst;
+    UWord rsbcmt;
+    UWord rmbcmt;
+    UWord mmbcs;
+    UWord mmsbc;
+    UWord mmmbc;
+    UWord lmbcs;
+    UWord smbcs;
+    UWord mbcgs;
 } AllctrInit_t;
 
 typedef struct {
-    Uint blocks;
-    Uint carriers;
+    UWord blocks;
+    UWord carriers;
 } AllctrSize_t;
 
 #ifndef SMALL_MEMORY
@@ -150,7 +150,7 @@ void    erts_alcu_current_size(Allctr_t *, AllctrSize_t *);
 
 #undef ERTS_ALLOC_UTIL_HARD_DEBUG
 #ifdef DEBUG
-#  if 0
+#  if 1
 #    define ERTS_ALLOC_UTIL_HARD_DEBUG
 #  endif
 #endif
@@ -163,19 +163,19 @@ void    erts_alcu_current_size(Allctr_t *, AllctrSize_t *);
 #define CEILING(X, I)  ((((X) - 1)/(I) + 1)*(I))
 
 #undef  WORD_MASK
-#define INV_WORD_MASK	((Uint) (sizeof(Uint) - 1))
+#define INV_WORD_MASK	((UWord) (sizeof(UWord) - 1))
 #define WORD_MASK	(~INV_WORD_MASK)
 #define WORD_FLOOR(X)	((X) & WORD_MASK)
 #define WORD_CEILING(X)	WORD_FLOOR((X) + INV_WORD_MASK)
 
 #undef  UNIT_MASK
-#define INV_UNIT_MASK	((Uint) (sizeof(Unit_t) - 1))
+#define INV_UNIT_MASK	((UWord) (sizeof(Unit_t) - 1))
 #define UNIT_MASK	(~INV_UNIT_MASK)
 #define UNIT_FLOOR(X)	((X) & UNIT_MASK)
 #define UNIT_CEILING(X)	UNIT_FLOOR((X) + INV_UNIT_MASK)
 
 
-#define SZ_MASK			(~((Uint) 0) << 3)
+#define SZ_MASK			(~((UWord) 0) << 3)
 #define FLG_MASK		(~(SZ_MASK))
 
 
@@ -189,7 +189,7 @@ typedef union {char c[8]; long l; double d;} Unit_t;
 
 typedef struct Carrier_t_ Carrier_t;
 struct Carrier_t_ {
-    Uint chdr;
+    UWord chdr;
     Carrier_t *next;
     Carrier_t *prev;
 };
@@ -199,17 +199,17 @@ typedef struct {
     Carrier_t *last;
 } CarrierList_t;
 
-typedef Uint Block_t;
-typedef Uint FreeBlkFtr_t;
+typedef UWord Block_t;
+typedef UWord FreeBlkFtr_t;
 
 typedef struct {
-    Uint giga_no;
-    Uint no;
+    UWord giga_no;
+    UWord no;
 } CallCounter_t;
 
 typedef struct {
-    Uint		no;
-    Uint		size;
+    UWord		no;
+    UWord		size;
 } StatValues_t;
 
 typedef struct {
diff --git a/erts/emulator/beam/erl_async.c b/erts/emulator/beam/erl_async.c
index b090564649..ef9644c3d1 100644
--- a/erts/emulator/beam/erl_async.c
+++ b/erts/emulator/beam/erl_async.c
@@ -390,7 +390,7 @@ int check_async_ready(void)
 **      async_data     data to pass to invoke function
 **      async_free     function for relase async_data in case of failure
 */
-long driver_async(ErlDrvPort ix, unsigned int* key,
+long driver_async(ErlDrvPort ix, unsigned long* key,
 		  void (*async_invoke)(void*), void* async_data,
 		  void (*async_free)(void*))
 {
diff --git a/erts/emulator/beam/erl_bif_chksum.c b/erts/emulator/beam/erl_bif_chksum.c
index 445ba00ca7..e76b0b9252 100644
--- a/erts/emulator/beam/erl_bif_chksum.c
+++ b/erts/emulator/beam/erl_bif_chksum.c
@@ -49,9 +49,9 @@ void erts_init_bif_chksum(void)
     chksum_md5_2_exp.code[1] = am_atom_put("md5_trap",8);
     chksum_md5_2_exp.code[2] = 2;
     chksum_md5_2_exp.code[3] =
-	(Eterm) em_apply_bif;
+	(UWord) em_apply_bif;
     chksum_md5_2_exp.code[4] = 
-	(Eterm) &md5_2;
+	(UWord) &md5_2;
 }
     
 
diff --git a/erts/emulator/beam/erl_bif_info.c b/erts/emulator/beam/erl_bif_info.c
index 8d3618393a..40f8b7f23a 100644
--- a/erts/emulator/beam/erl_bif_info.c
+++ b/erts/emulator/beam/erl_bif_info.c
@@ -127,7 +127,7 @@ bld_bin_list(Uint **hpp, Uint *szp, ProcBin* pb)
     Eterm tuple;
 
     for (; pb; pb = pb->next) {
-	Eterm val = erts_bld_uint(hpp, szp, (Uint) pb->val);
+	Eterm val = erts_bld_uword(hpp, szp, (UWord) pb->val);
 	Eterm orig_size = erts_bld_uint(hpp, szp, pb->val->orig_size);
 
 	if (szp)
@@ -1010,7 +1010,7 @@ process_info_aux(Process *BIF_P,
 	    hp = HAlloc(BIF_P, 3);
 	    res = am_undefined;
 	} else {
-	    Eterm* current;
+	    UWord* current;
 
 	    if (rp->current[0] == am_erlang &&
 		rp->current[1] == am_process_info &&
@@ -2829,7 +2829,7 @@ fun_info_2(Process* p, Eterm fun, Eterm what)
 	    goto error;
 	}
     } else if (is_export(fun)) {
-	Export* exp = (Export *) (export_val(fun))[1];
+	Export* exp = (Export *) ((UWord) (export_val(fun))[1]);
 	switch (what) {
 	case am_type:
 	    hp = HAlloc(p, 3);
@@ -3022,11 +3022,11 @@ BIF_RETTYPE statistics_1(BIF_ALIST_1)
 	res = erts_run_queues_len(NULL);
 	BIF_RET(make_small(res));
     } else if (BIF_ARG_1 == am_wall_clock) {
-	Uint w1, w2;
+	UWord w1, w2;
 	Eterm b1, b2;
 	wall_clock_elapsed_time_both(&w1, &w2);
-	b1 = erts_make_integer(w1,BIF_P);
-	b2 = erts_make_integer(w2,BIF_P);
+	b1 = erts_make_integer((Uint) w1,BIF_P);
+	b2 = erts_make_integer((Uint) w2,BIF_P);
 	hp = HAlloc(BIF_P,3);
 	res = TUPLE2(hp, b1, b2);
 	BIF_RET(res);
diff --git a/erts/emulator/beam/erl_bif_op.c b/erts/emulator/beam/erl_bif_op.c
index 6da72dcef9..372b858832 100644
--- a/erts/emulator/beam/erl_bif_op.c
+++ b/erts/emulator/beam/erl_bif_op.c
@@ -251,7 +251,7 @@ BIF_RETTYPE is_function_2(BIF_ALIST_2)
 	    BIF_RET(am_true);
 	}
     } else if (is_export(BIF_ARG_1)) {
-	Export* exp = (Export *) (export_val(BIF_ARG_1))[1];
+	Export* exp = (Export *) EXPAND_POINTER((export_val(BIF_ARG_1))[1]);
 
 	if (exp->code[2] == (Uint) arity) {
 	    BIF_RET(am_true);
diff --git a/erts/emulator/beam/erl_bif_re.c b/erts/emulator/beam/erl_bif_re.c
index c027cd5984..4b527ce215 100644
--- a/erts/emulator/beam/erl_bif_re.c
+++ b/erts/emulator/beam/erl_bif_re.c
@@ -76,8 +76,8 @@ void erts_init_bif_re(void)
     re_exec_trap_export.code[0] = am_erlang;
     re_exec_trap_export.code[1] = am_re_run_trap;
     re_exec_trap_export.code[2] = 3;
-    re_exec_trap_export.code[3] = (Eterm) em_apply_bif;
-    re_exec_trap_export.code[4] = (Eterm) &re_exec_trap;
+    re_exec_trap_export.code[3] = (UWord) em_apply_bif;
+    re_exec_trap_export.code[4] = (UWord) &re_exec_trap;
 
     grun_trap_exportp =  erts_export_put(am_re,am_grun,3);
     urun_trap_exportp =  erts_export_put(am_re,am_urun,3);
@@ -103,7 +103,7 @@ Sint erts_re_set_loop_limit(Sint limit)
 
 static int term_to_int(Eterm term, int *sp)
 {
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 
     if (is_small(term)) {
 	Uint x = signed_val(term);
@@ -154,7 +154,7 @@ static int term_to_int(Eterm term, int *sp)
 
 static Eterm make_signed_integer(int x, Process *p)
 {
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
     return make_small(x);
 #else
     Eterm* hp;
diff --git a/erts/emulator/beam/erl_bif_timer.c b/erts/emulator/beam/erl_bif_timer.c
index 1740569b69..23f4879d45 100644
--- a/erts/emulator/beam/erl_bif_timer.c
+++ b/erts/emulator/beam/erl_bif_timer.c
@@ -155,7 +155,7 @@ create_ref(Uint *hp, Uint32 *ref_numbers, Uint32 len)
 	erl_exit(1, "%s:%d: Internal error\n", __FILE__, __LINE__);
     }
 
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
     hp[0] = make_ref_thing_header(len/2 + 1);
     datap = (Uint32 *) &hp[1];
     *(datap++) = len;
@@ -173,7 +173,7 @@ create_ref(Uint *hp, Uint32 *ref_numbers, Uint32 len)
 static int
 eq_non_standard_ref_numbers(Uint32 *rn1, Uint32 len1, Uint32 *rn2, Uint32 len2)
 {
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 #define MAX_REF_HEAP_SZ (1+(ERTS_MAX_REF_NUMBERS/2+1))
 #else
 #define MAX_REF_HEAP_SZ (1+ERTS_MAX_REF_NUMBERS)
@@ -398,7 +398,7 @@ setup_bif_timer(Uint32 xflags,
     
     if (!term_to_Uint(time, &timeout))
 	return THE_NON_VALUE;
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
     if ((timeout >> 32) != 0)
 	return THE_NON_VALUE;
 #endif
diff --git a/erts/emulator/beam/erl_bif_trace.c b/erts/emulator/beam/erl_bif_trace.c
index 6623eea4f3..c46a0074fc 100644
--- a/erts/emulator/beam/erl_bif_trace.c
+++ b/erts/emulator/beam/erl_bif_trace.c
@@ -962,12 +962,12 @@ static int function_is_traced(Eterm mfa[3],
     e.code[2] = mfa[2];
     if ((ep = export_get(&e)) != NULL) {
 	if (ep->address == ep->code+3 &&
-	    ep->code[3] != (Uint) em_call_error_handler) {
-	    if (ep->code[3] == (Uint) em_call_traced_function) {
+	    ep->code[3] != (UWord) em_call_error_handler) {
+	    if (ep->code[3] == (UWord) em_call_traced_function) {
 		*ms = ep->match_prog_set;
 		return FUNC_TRACE_GLOBAL_TRACE;
 	    }
-	    if (ep->code[3] == (Uint) em_apply_bif) {
+	    if (ep->code[3] == (UWord) em_apply_bif) {
 		for (i = 0; i < BIF_SIZE; ++i) {
 		    if (bif_export[i] == ep) {
 			int r = 0;
@@ -1439,9 +1439,9 @@ static int
 setup_func_trace(Export* ep, void* match_prog)
 {
     if (ep->address == ep->code+3) {
-	if (ep->code[3] == (Uint) em_call_error_handler) {
+	if (ep->code[3] == (UWord) em_call_error_handler) {
 	    return 0;
-	} else if (ep->code[3] == (Uint) em_call_traced_function) {
+	} else if (ep->code[3] == (UWord) em_call_traced_function) {
 	    MatchSetUnref(ep->match_prog_set);
 	    ep->match_prog_set = match_prog;
 	    MatchSetRef(ep->match_prog_set);
@@ -1461,8 +1461,8 @@ setup_func_trace(Export* ep, void* match_prog)
 	return 0;
     }
     
-    ep->code[3] = (Uint) em_call_traced_function;
-    ep->code[4] = (Uint) ep->address;
+    ep->code[3] = (UWord) em_call_traced_function;
+    ep->code[4] = (UWord) ep->address;
     ep->address = ep->code+3;
     ep->match_prog_set = match_prog;
     MatchSetRef(ep->match_prog_set);
@@ -1474,7 +1474,7 @@ static void setup_bif_trace(int bif_index) {
     
     ASSERT(ExportIsBuiltIn(ep));
     ASSERT(ep->code[4]);
-    ep->code[4] = (Uint) bif_table[bif_index].traced;
+    ep->code[4] = (UWord) bif_table[bif_index].traced;
 }
 
 static void set_trace_bif(int bif_index, void* match_prog) {
@@ -1501,9 +1501,9 @@ static int
 reset_func_trace(Export* ep)
 {
     if (ep->address == ep->code+3) {
-	if (ep->code[3] == (Uint) em_call_error_handler) {
+	if (ep->code[3] == (UWord) em_call_error_handler) {
 	    return 0;
-	} else if (ep->code[3] == (Uint) em_call_traced_function) {
+	} else if (ep->code[3] == (UWord) em_call_traced_function) {
 	    ep->address = (Uint *) ep->code[4];
 	    MatchSetUnref(ep->match_prog_set);
 	    ep->match_prog_set = NULL;
@@ -1536,8 +1536,8 @@ static void reset_bif_trace(int bif_index) {
     ASSERT(ExportIsBuiltIn(ep));
     ASSERT(ep->code[4]);
     ASSERT(! ep->match_prog_set);
-    ASSERT(! erts_is_mtrace_bif((Uint *)ep->code+3, NULL, NULL));
-    ep->code[4] = (Uint) bif_table[bif_index].f;
+    ASSERT(! erts_is_mtrace_bif((UWord *)ep->code+3, NULL, NULL));
+    ep->code[4] = (UWord) bif_table[bif_index].f;
 }
 
 static void clear_trace_bif(int bif_index) {
diff --git a/erts/emulator/beam/erl_binary.h b/erts/emulator/beam/erl_binary.h
index 1f948a9684..5b0b3bcec2 100644
--- a/erts/emulator/beam/erl_binary.h
+++ b/erts/emulator/beam/erl_binary.h
@@ -164,7 +164,7 @@ byte* erts_get_aligned_binary_bytes_extra(Eterm, byte**, unsigned extra);
 #endif
 
 #define ERTS_CHK_BIN_ALIGNMENT(B) \
-  do { ASSERT(!(B) || (((Uint) &((Binary *)(B))->orig_bytes[0]) & ERTS_BIN_ALIGNMENT_MASK) == ((Uint) 0)) } while(0)
+  do { ASSERT(!(B) || (((UWord) &((Binary *)(B))->orig_bytes[0]) & ERTS_BIN_ALIGNMENT_MASK) == ((UWord) 0)) } while(0)
 
 ERTS_GLB_INLINE byte* erts_get_aligned_binary_bytes(Eterm bin, byte** base_ptr);
 ERTS_GLB_INLINE void erts_free_aligned_binary_bytes(byte* buf);
diff --git a/erts/emulator/beam/erl_bits.c b/erts/emulator/beam/erl_bits.c
index e4f5d50ddf..3f69ddcdb6 100644
--- a/erts/emulator/beam/erl_bits.c
+++ b/erts/emulator/beam/erl_bits.c
@@ -255,7 +255,7 @@ erts_bs_get_integer_2(Process *p, Uint num_bits, unsigned flags, ErlBinMatchBuff
 	 * Simply shift whole bytes into the result.
 	 */
 	switch (BYTE_OFFSET(n)) {
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 	case 7: w = (w << 8) | *bp++;
 	case 6: w = (w << 8) | *bp++;
 	case 5: w = (w << 8) | *bp++;
@@ -360,7 +360,7 @@ erts_bs_get_integer_2(Process *p, Uint num_bits, unsigned flags, ErlBinMatchBuff
     case 3: 
 	v32 = LSB[0] + (LSB[1]<<8) + (LSB[2]<<16); 
 	goto big_small;
-#if !defined(ARCH_64)
+#if !defined(ARCH_64) || HALFWORD_HEAP
     case 4:
 	v32 = (LSB[0] + (LSB[1]<<8) + (LSB[2]<<16) + (LSB[3]<<24));
 	if (!IS_USMALL(sgn, v32)) {
diff --git a/erts/emulator/beam/erl_db.c b/erts/emulator/beam/erl_db.c
index 08c448882f..959f7fa37c 100644
--- a/erts/emulator/beam/erl_db.c
+++ b/erts/emulator/beam/erl_db.c
@@ -87,7 +87,7 @@ static union {
 static struct {
     union {
 	DbTable *tb;     /* Only directly readable if slot is ALIVE */
-	Uint next_free;  /* (index<<2)|1 if slot is FREE */
+	UWord next_free;  /* (index<<2)|1 if slot is FREE */
     }u;
 } *meta_main_tab;
 
@@ -187,7 +187,7 @@ static Eterm ms_delete_all_buff[8]; /* To compare with for deletion
 
 static void fix_table_locked(Process* p, DbTable* tb);
 static void unfix_table_locked(Process* p,  DbTable* tb, db_lock_kind_t* kind);
-static void set_heir(Process* me, DbTable* tb, Eterm heir, Eterm heir_data);
+static void set_heir(Process* me, DbTable* tb, Eterm heir, UWord heir_data);
 static void free_heir_data(DbTable*);
 static void free_fixations_locked(DbTable *tb);
 
@@ -1186,7 +1186,7 @@ BIF_RETTYPE ets_new_2(BIF_ALIST_2)
     Eterm val;
     Eterm ret;
     Eterm heir;
-    Eterm heir_data;
+    UWord heir_data;
     Uint32 status;
     Sint keypos;
     int is_named, is_fine_locked;
@@ -1206,7 +1206,7 @@ BIF_RETTYPE ets_new_2(BIF_ALIST_2)
     is_named = 0;
     is_fine_locked = 0;
     heir = am_none;
-    heir_data = am_undefined;
+    heir_data = (UWord) am_undefined;
 
     list = BIF_ARG_2;
     while(is_list(list)) {
@@ -1566,9 +1566,15 @@ BIF_RETTYPE ets_delete_1(BIF_ALIST_1)
 	 * (it looks like an continuation pointer), but that is will crash the
 	 * emulator if this BIF is call traced.
 	 */
+#if HALFWORD_HEAP
+	Eterm *hp = HAlloc(BIF_P, 3);
+	hp[0] = make_pos_bignum_header(2);
+	*((UWord *) (UWord) (hp+1)) = (UWord) tb;
+#else
 	Eterm *hp = HAlloc(BIF_P, 2);
 	hp[0] = make_pos_bignum_header(1);
 	hp[1] = (Eterm) tb;
+#endif
 	BIF_TRAP1(&ets_delete_continue_exp, BIF_P, make_big(hp));
     }
     else {
@@ -1638,7 +1644,7 @@ BIF_RETTYPE ets_setopts_2(BIF_ALIST_2)
     Eterm* tp;
     Eterm opt;
     Eterm heir = THE_NON_VALUE;
-    Eterm heir_data = THE_NON_VALUE;
+    UWord heir_data = (UWord) THE_NON_VALUE;
     Uint32 protection = 0;
     DeclareTmpHeap(fakelist,2,BIF_P);
     Eterm tail;
@@ -2618,7 +2624,7 @@ void init_db(void)
 {
     DbTable init_tb;
     int i;
-    extern Eterm* em_apply_bif;
+    extern UWord* em_apply_bif;
     Eterm *hp;
     unsigned bits;
     size_t size;
@@ -2747,9 +2753,9 @@ void init_db(void)
     ets_select_delete_continue_exp.code[1] = am_atom_put("delete_trap",11);
     ets_select_delete_continue_exp.code[2] = 1;
     ets_select_delete_continue_exp.code[3] =
-	(Eterm) em_apply_bif;
+	(UWord) em_apply_bif;
     ets_select_delete_continue_exp.code[4] = 
-	(Eterm) &ets_select_delete_1;
+	(UWord) &ets_select_delete_1;
 
     /* Non visual BIF to trap to. */
     memset(&ets_select_count_continue_exp, 0, sizeof(Export));
@@ -2759,9 +2765,9 @@ void init_db(void)
     ets_select_count_continue_exp.code[1] = am_atom_put("count_trap",11);
     ets_select_count_continue_exp.code[2] = 1;
     ets_select_count_continue_exp.code[3] =
-	(Eterm) em_apply_bif;
+	(UWord) em_apply_bif;
     ets_select_count_continue_exp.code[4] = 
-	(Eterm) &ets_select_count_1;
+	(UWord) &ets_select_count_1;
 
     /* Non visual BIF to trap to. */
     memset(&ets_select_continue_exp, 0, sizeof(Export));
@@ -2771,9 +2777,9 @@ void init_db(void)
     ets_select_continue_exp.code[1] = am_atom_put("select_trap",11);
     ets_select_continue_exp.code[2] = 1;
     ets_select_continue_exp.code[3] =
-	(Eterm) em_apply_bif;
+	(UWord) em_apply_bif;
     ets_select_continue_exp.code[4] = 
-	(Eterm) &ets_select_trap_1;
+	(UWord) &ets_select_trap_1;
 
     /* Non visual BIF to trap to. */
     memset(&ets_delete_continue_exp, 0, sizeof(Export));
@@ -2781,8 +2787,8 @@ void init_db(void)
     ets_delete_continue_exp.code[0] = am_ets;
     ets_delete_continue_exp.code[1] = am_atom_put("delete_trap",11);
     ets_delete_continue_exp.code[2] = 1;
-    ets_delete_continue_exp.code[3] = (Eterm) em_apply_bif;
-    ets_delete_continue_exp.code[4] = (Eterm) &ets_delete_trap;
+    ets_delete_continue_exp.code[3] = (UWord) em_apply_bif;
+    ets_delete_continue_exp.code[4] = (UWord) &ets_delete_trap;
 
     hp = ms_delete_all_buff;
     ms_delete_all = CONS(hp, am_true, NIL);
@@ -2878,7 +2884,7 @@ static int give_away_to_heir(Process* p, DbTable* tb)
     ErtsProcLocks to_locks = ERTS_PROC_LOCK_MAIN;
     DeclareTmpHeap(buf,5,p);
     Eterm to_pid;
-    Eterm heir_data;
+    UWord heir_data;
 
     ASSERT(tb->common.owner == p->id);
     ASSERT(is_internal_pid(tb->common.heir));
@@ -3301,7 +3307,7 @@ static void free_fixations_locked(DbTable *tb)
     tb->common.fixations = NULL;
 }
 
-static void set_heir(Process* me, DbTable* tb, Eterm heir, Eterm heir_data)
+static void set_heir(Process* me, DbTable* tb, Eterm heir, UWord heir_data)
 {	
     tb->common.heir = heir;
     if (heir == am_none) {
@@ -3326,7 +3332,7 @@ static void set_heir(Process* me, DbTable* tb, Eterm heir, Eterm heir_data)
 
 	UseTmpHeap(2,me);
 	/* Make a dummy 1-tuple around data to use db_get_term() */
-	heir_data = (Eterm) db_get_term(&tb->common, NULL, 0,
+	heir_data = (UWord) db_get_term(&tb->common, NULL, 0,
 					TUPLE1(tmp,heir_data));
 	UnUseTmpHeap(2,me);
 	ASSERT(!is_immed(heir_data));
@@ -3351,7 +3357,7 @@ static BIF_RETTYPE ets_delete_trap(Process *p, Eterm cont)
 {
     int trap;
     Eterm* ptr = big_val(cont);
-    DbTable *tb = (DbTable *) ptr[1];
+    DbTable *tb = *((DbTable **) (UWord) (ptr + 1));
 
     ASSERT(*ptr == make_pos_bignum_header(1));
 
diff --git a/erts/emulator/beam/erl_db_tree.c b/erts/emulator/beam/erl_db_tree.c
index b421da591b..71c594854b 100644
--- a/erts/emulator/beam/erl_db_tree.c
+++ b/erts/emulator/beam/erl_db_tree.c
@@ -443,9 +443,9 @@ void db_initialize_tree(void)
     ets_select_reverse_exp.code[1] = am_reverse;
     ets_select_reverse_exp.code[2] = 3;
     ets_select_reverse_exp.code[3] =
-	(Eterm) em_apply_bif;
+	(UWord) em_apply_bif;
     ets_select_reverse_exp.code[4] = 
-	(Eterm) &ets_select_reverse;
+	(UWord) &ets_select_reverse;
     return;
 };
 
diff --git a/erts/emulator/beam/erl_db_util.c b/erts/emulator/beam/erl_db_util.c
index de12b9bdbf..c565e1ee89 100644
--- a/erts/emulator/beam/erl_db_util.c
+++ b/erts/emulator/beam/erl_db_util.c
@@ -281,7 +281,7 @@ typedef struct dmc_guard_bif {
 */
 DMC_DECLARE_STACK_TYPE(Eterm);
 
-DMC_DECLARE_STACK_TYPE(Uint);
+DMC_DECLARE_STACK_TYPE(UWord);
 
 DMC_DECLARE_STACK_TYPE(unsigned);
 
@@ -382,7 +382,7 @@ cleanup_match_pseudo_process(ErtsMatchPseudoProcess *mpsp, int keep_heap)
 	else {
 	    int i;
 	    for (i = 0; i < ERTS_DEFAULT_MS_HEAP_SIZE; i++) {
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 		mpsp->default_heap[i] = (Eterm) 0xdeadbeefdeadbeef;
 #else
 		mpsp->default_heap[i] = (Eterm) 0xdeadbeef;
@@ -830,42 +830,42 @@ static Uint my_size_object(Eterm t);
 static Eterm my_copy_struct(Eterm t, Eterm **hp, ErlOffHeap* off_heap);
 
 /* Guard compilation */
-static void do_emit_constant(DMCContext *context, DMC_STACK_TYPE(Uint) *text,
+static void do_emit_constant(DMCContext *context, DMC_STACK_TYPE(UWord) *text,
 			     Eterm t);
 static DMCRet dmc_list(DMCContext *context,
 		       DMCHeap *heap,
-		       DMC_STACK_TYPE(Uint) *text,
+		       DMC_STACK_TYPE(UWord) *text,
 		       Eterm t,
 		       int *constant);
 static DMCRet dmc_tuple(DMCContext *context,
 		       DMCHeap *heap,
-		       DMC_STACK_TYPE(Uint) *text,
+		       DMC_STACK_TYPE(UWord) *text,
 		       Eterm t,
 		       int *constant);
 static DMCRet dmc_variable(DMCContext *context,
 			   DMCHeap *heap,
-			   DMC_STACK_TYPE(Uint) *text,
+			   DMC_STACK_TYPE(UWord) *text,
 			   Eterm t,
 			   int *constant);
 static DMCRet dmc_fun(DMCContext *context,
 		      DMCHeap *heap,
-		      DMC_STACK_TYPE(Uint) *text,
+		      DMC_STACK_TYPE(UWord) *text,
 		      Eterm t,
 		      int *constant);
 static DMCRet dmc_expr(DMCContext *context,
 		       DMCHeap *heap,
-		       DMC_STACK_TYPE(Uint) *text,
+		       DMC_STACK_TYPE(UWord) *text,
 		       Eterm t,
 		       int *constant);
 static DMCRet compile_guard_expr(DMCContext *context,
 				    DMCHeap *heap,
-				    DMC_STACK_TYPE(Uint) *text,
+				    DMC_STACK_TYPE(UWord) *text,
 				    Eterm t);
 /* match expression subroutine */
 static DMCRet dmc_one_term(DMCContext *context, 
 			   DMCHeap *heap,
 			   DMC_STACK_TYPE(Eterm) *stack,
-			   DMC_STACK_TYPE(Uint) *text,
+			   DMC_STACK_TYPE(UWord) *text,
 			   Eterm c);
 
 
@@ -1185,7 +1185,7 @@ Eterm erts_match_set_run(Process *p, Binary *mpsp,
     Eterm ret;
 
     ret = db_prog_match(p, mpsp,
-			(Eterm) args, 
+			(Eterm) COMPRESS_POINTER(args),
 			num_args, return_flags);
 #if defined(HARDDEBUG)
     if (is_non_value(ret)) {
@@ -1245,7 +1245,7 @@ Binary *db_match_compile(Eterm *matchexpr,
 {
     DMCHeap heap;
     DMC_STACK_TYPE(Eterm) stack;
-    DMC_STACK_TYPE(Uint) text;
+    DMC_STACK_TYPE(UWord) text;
     DMCContext context;
     MatchProg *ret = NULL;
     Eterm t;
@@ -1491,8 +1491,8 @@ restart:
     ** A special case is when the match expression is a single binding 
     ** (i.e '$1'), then the field single_variable is set to 1.
     */
-    bp = erts_create_magic_binary(((sizeof(MatchProg) - sizeof(Uint)) +
-				   (DMC_STACK_NUM(text) * sizeof(Uint))),
+    bp = erts_create_magic_binary(((sizeof(MatchProg) - sizeof(UWord)) +
+				   (DMC_STACK_NUM(text) * sizeof(UWord))),
 				  erts_db_match_prog_destructor);
     ret = Binary2MatchProg(bp);
     ret->saved_program_buf = NULL;
@@ -1501,7 +1501,7 @@ restart:
     ret->num_bindings = heap.used;
     ret->single_variable = context.special;
     sys_memcpy(ret->text, DMC_STACK_DATA(text), 
-	       DMC_STACK_NUM(text) * sizeof(Uint));
+	       DMC_STACK_NUM(text) * sizeof(UWord));
     ret->heap_size = ((heap.used * sizeof(Eterm)) +
 		      (max_eheap_need * sizeof(Eterm)) +
 		      (context.stack_need * sizeof(Eterm *)) +
@@ -1601,7 +1601,8 @@ Eterm db_prog_match(Process *c_p, Binary *bprog, Eterm term,
     Eterm **sp;
     Eterm *esp;
     Eterm *hp;
-    Uint *pc = prog->text;
+    UWord *cp;
+    UWord *pc = prog->text;
     Eterm *ehp;
     Eterm ret;
     Uint n = 0; /* To avoid warning. */
@@ -1616,9 +1617,9 @@ Eterm db_prog_match(Process *c_p, Binary *bprog, Eterm term,
     int fail_label;
     int atomic_trace;
 #ifdef DMC_DEBUG
-    unsigned long *heap_fence;
-    unsigned long *eheap_fence;
-    unsigned long *stack_fence;
+    Uint *heap_fence;
+    Uint *eheap_fence;
+    Uint *stack_fence;
     Uint save_op;
 #endif /* DMC_DEBUG */
 
@@ -1654,9 +1655,9 @@ Eterm db_prog_match(Process *c_p, Binary *bprog, Eterm term,
 
 #ifdef DMC_DEBUG
     save_op = 0;
-    heap_fence =  (unsigned long *) mpsp->heap + prog->eheap_offset - 1;
-    eheap_fence = (unsigned long *) mpsp->heap + prog->stack_offset - 1;
-    stack_fence = (unsigned long *) mpsp->heap + prog->heap_size - 1;
+    heap_fence =  (Uint *) mpsp->heap + prog->eheap_offset - 1;
+    eheap_fence = (Uint *) mpsp->heap + prog->stack_offset - 1;
+    stack_fence = (Uint *) mpsp->heap + prog->heap_size - 1;
     *heap_fence = FENCE_PATTERN;
     *eheap_fence = FENCE_PATTERN;
     *stack_fence = FENCE_PATTERN;
@@ -1709,11 +1710,12 @@ restart:
 	    n = *pc++;
 	    if ((int) n != arity)
 		FAIL();
-	    ep = (Eterm *) *ep;
+	    ep = (Eterm *) EXPAND_POINTER(*ep);
 	    break;
-	case matchArrayBind: /* When the array size is unknown. */
+	case matchArrayBind: /* When the array size is unknown. */ /* XXX:PaN - where does
+								      this array come from? */
 	    n = *pc++;
-	    hp[n] = dpm_array_to_list(psp, (Eterm *) term, arity);
+	    hp[n] = dpm_array_to_list(psp, (Eterm *) EXPAND_POINTER(term), arity);
 	    break;
 	case matchTuple: /* *ep is a tuple of arity n */
 	    if (!is_tuple(*ep))
@@ -1884,7 +1886,7 @@ restart:
 	    break;
 	case matchPushArrayAsList:
 	    n = arity; /* Only happens when 'term' is an array */
-	    tp = (Eterm *) term;
+	    tp = (Eterm *) EXPAND_POINTER(term);
 	    *esp++  = make_list(ehp);
 	    while (n--) {
 		*ehp++ = *tp++;
@@ -1897,7 +1899,7 @@ restart:
 	    break;
 	case matchPushArrayAsListU:
 	    /* This instruction is NOT efficient. */
-	    *esp++  = dpm_array_to_list(psp, (Eterm *) term, arity); 
+	    *esp++  = dpm_array_to_list(psp, (Eterm *) EXPAND_POINTER(term), arity);
 	    break;
 	case matchTrue:
 	    if (*--esp != am_true)
@@ -2095,14 +2097,14 @@ restart:
 	    }
 	    break;
  	case matchCaller:
- 	    if (!(c_p->cp) || !(hp = find_function_from_pc(c_p->cp))) {
+	    if (!(c_p->cp) || !(cp = find_function_from_pc(c_p->cp))) {
  		*esp++ = am_undefined;
  	    } else {
  		*esp++ = make_tuple(ehp);
  		ehp[0] = make_arityval(3);
- 		ehp[1] = hp[0];
- 		ehp[2] = hp[1];
- 		ehp[3] = make_small(hp[2]);
+		ehp[1] = cp[0];
+		ehp[2] = cp[1];
+		ehp[3] = make_small((Uint) hp[2]);
  		ehp += 4;
  	    }
  	    break;
@@ -2609,7 +2611,7 @@ static void add_dmc_err(DMCErrInfo *err_info,
 static DMCRet dmc_one_term(DMCContext *context, 
 			   DMCHeap *heap,
 			   DMC_STACK_TYPE(Eterm) *stack,
-			   DMC_STACK_TYPE(Uint) *text,
+			   DMC_STACK_TYPE(UWord) *text,
 			   Eterm c)
 {
     Sint n;
@@ -2756,7 +2758,7 @@ static DMCRet dmc_one_term(DMCContext *context,
 ** Match guard compilation
 */
 
-static void do_emit_constant(DMCContext *context, DMC_STACK_TYPE(Uint) *text,
+static void do_emit_constant(DMCContext *context, DMC_STACK_TYPE(UWord) *text,
 			     Eterm t) 
 {
 	int sz;
@@ -2810,7 +2812,7 @@ add_dmc_err((ContextP)->err_info, String, -1, T, dmcWarning)
 
 static DMCRet dmc_list(DMCContext *context,
 		       DMCHeap *heap,
-		       DMC_STACK_TYPE(Uint) *text,
+		       DMC_STACK_TYPE(UWord) *text,
 		       Eterm t,
 		       int *constant)
 {
@@ -2846,11 +2848,11 @@ static DMCRet dmc_list(DMCContext *context,
 
 static DMCRet dmc_tuple(DMCContext *context,
 		       DMCHeap *heap,
-		       DMC_STACK_TYPE(Uint) *text,
+		       DMC_STACK_TYPE(UWord) *text,
 		       Eterm t,
 		       int *constant)
 {
-    DMC_STACK_TYPE(Uint) instr_save;
+    DMC_STACK_TYPE(UWord) instr_save;
     int all_constant = 1;
     int textpos = DMC_STACK_NUM(*text);
     Eterm *p = tuple_val(t);
@@ -2906,7 +2908,7 @@ static DMCRet dmc_tuple(DMCContext *context,
 
 static DMCRet dmc_whole_expression(DMCContext *context,
 				   DMCHeap *heap,
-				   DMC_STACK_TYPE(Uint) *text,
+				   DMC_STACK_TYPE(UWord) *text,
 				   Eterm t,
 				   int *constant)
 {
@@ -2934,7 +2936,7 @@ static DMCRet dmc_whole_expression(DMCContext *context,
 
 static DMCRet dmc_variable(DMCContext *context,
 			   DMCHeap *heap,
-			   DMC_STACK_TYPE(Uint) *text,
+			   DMC_STACK_TYPE(UWord) *text,
 			   Eterm t,
 			   int *constant)
 {
@@ -2955,7 +2957,7 @@ static DMCRet dmc_variable(DMCContext *context,
 
 static DMCRet dmc_all_bindings(DMCContext *context,
 			       DMCHeap *heap,
-			       DMC_STACK_TYPE(Uint) *text,
+			       DMC_STACK_TYPE(UWord) *text,
 			       Eterm t,
 			       int *constant)
 {
@@ -2982,7 +2984,7 @@ static DMCRet dmc_all_bindings(DMCContext *context,
 
 static DMCRet dmc_const(DMCContext *context,
 		       DMCHeap *heap,
-		       DMC_STACK_TYPE(Uint) *text,
+		       DMC_STACK_TYPE(UWord) *text,
 		       Eterm t,
 		       int *constant)
 {
@@ -2999,7 +3001,7 @@ static DMCRet dmc_const(DMCContext *context,
 
 static DMCRet dmc_and(DMCContext *context,
 		      DMCHeap *heap,
-		      DMC_STACK_TYPE(Uint) *text,
+		      DMC_STACK_TYPE(UWord) *text,
 		      Eterm t,
 		      int *constant)
 {
@@ -3028,7 +3030,7 @@ static DMCRet dmc_and(DMCContext *context,
 
 static DMCRet dmc_or(DMCContext *context,
 		     DMCHeap *heap,
-		     DMC_STACK_TYPE(Uint) *text,
+		     DMC_STACK_TYPE(UWord) *text,
 		     Eterm t,
 		     int *constant)
 {
@@ -3058,7 +3060,7 @@ static DMCRet dmc_or(DMCContext *context,
 
 static DMCRet dmc_andalso(DMCContext *context,
 			  DMCHeap *heap,
-			  DMC_STACK_TYPE(Uint) *text,
+			  DMC_STACK_TYPE(UWord) *text,
 			  Eterm t,
 			  int *constant)
 {
@@ -3107,7 +3109,7 @@ static DMCRet dmc_andalso(DMCContext *context,
 
 static DMCRet dmc_orelse(DMCContext *context,
 			 DMCHeap *heap,
-			 DMC_STACK_TYPE(Uint) *text,
+			 DMC_STACK_TYPE(UWord) *text,
 			 Eterm t,
 			 int *constant)
 {
@@ -3155,7 +3157,7 @@ static DMCRet dmc_orelse(DMCContext *context,
 
 static DMCRet dmc_message(DMCContext *context,
 			  DMCHeap *heap,
-			  DMC_STACK_TYPE(Uint) *text,
+			  DMC_STACK_TYPE(UWord) *text,
 			  Eterm t,
 			  int *constant)
 {
@@ -3197,7 +3199,7 @@ static DMCRet dmc_message(DMCContext *context,
 
 static DMCRet dmc_self(DMCContext *context,
 		     DMCHeap *heap,
-		     DMC_STACK_TYPE(Uint) *text,
+		     DMC_STACK_TYPE(UWord) *text,
 		     Eterm t,
 		     int *constant)
 {
@@ -3217,7 +3219,7 @@ static DMCRet dmc_self(DMCContext *context,
 
 static DMCRet dmc_return_trace(DMCContext *context,
 			       DMCHeap *heap,
-			       DMC_STACK_TYPE(Uint) *text,
+			       DMC_STACK_TYPE(UWord) *text,
 			       Eterm t,
 			       int *constant)
 {
@@ -3247,7 +3249,7 @@ static DMCRet dmc_return_trace(DMCContext *context,
 
 static DMCRet dmc_exception_trace(DMCContext *context,
 			       DMCHeap *heap,
-			       DMC_STACK_TYPE(Uint) *text,
+			       DMC_STACK_TYPE(UWord) *text,
 			       Eterm t,
 			       int *constant)
 {
@@ -3279,7 +3281,7 @@ static DMCRet dmc_exception_trace(DMCContext *context,
 
 static DMCRet dmc_is_seq_trace(DMCContext *context,
 			       DMCHeap *heap,
-			       DMC_STACK_TYPE(Uint) *text,
+			       DMC_STACK_TYPE(UWord) *text,
 			       Eterm t,
 			       int *constant)
 {
@@ -3305,7 +3307,7 @@ static DMCRet dmc_is_seq_trace(DMCContext *context,
 
 static DMCRet dmc_set_seq_token(DMCContext *context,
 				DMCHeap *heap,
-				DMC_STACK_TYPE(Uint) *text,
+				DMC_STACK_TYPE(UWord) *text,
 				Eterm t,
 				int *constant)
 {
@@ -3354,7 +3356,7 @@ static DMCRet dmc_set_seq_token(DMCContext *context,
 
 static DMCRet dmc_get_seq_token(DMCContext *context,
 				DMCHeap *heap,
-				DMC_STACK_TYPE(Uint) *text,
+				DMC_STACK_TYPE(UWord) *text,
 				Eterm t,
 				int *constant)
 {
@@ -3391,7 +3393,7 @@ static DMCRet dmc_get_seq_token(DMCContext *context,
 
 static DMCRet dmc_display(DMCContext *context,
 			  DMCHeap *heap,
-			  DMC_STACK_TYPE(Uint) *text,
+			  DMC_STACK_TYPE(UWord) *text,
 			  Eterm t,
 			  int *constant)
 {
@@ -3431,7 +3433,7 @@ static DMCRet dmc_display(DMCContext *context,
 
 static DMCRet dmc_process_dump(DMCContext *context,
 			       DMCHeap *heap,
-			       DMC_STACK_TYPE(Uint) *text,
+			       DMC_STACK_TYPE(UWord) *text,
 			       Eterm t,
 			       int *constant)
 {
@@ -3461,7 +3463,7 @@ static DMCRet dmc_process_dump(DMCContext *context,
 
 static DMCRet dmc_enable_trace(DMCContext *context,
 			       DMCHeap *heap,
-			       DMC_STACK_TYPE(Uint) *text,
+			       DMC_STACK_TYPE(UWord) *text,
 			       Eterm t,
 			       int *constant)
 {
@@ -3521,7 +3523,7 @@ static DMCRet dmc_enable_trace(DMCContext *context,
 
 static DMCRet dmc_disable_trace(DMCContext *context,
 				DMCHeap *heap,
-				DMC_STACK_TYPE(Uint) *text,
+				DMC_STACK_TYPE(UWord) *text,
 				Eterm t,
 				int *constant)
 {
@@ -3581,7 +3583,7 @@ static DMCRet dmc_disable_trace(DMCContext *context,
 
 static DMCRet dmc_trace(DMCContext *context,
 			DMCHeap *heap,
-			DMC_STACK_TYPE(Uint) *text,
+			DMC_STACK_TYPE(UWord) *text,
 			Eterm t,
 			int *constant)
 {
@@ -3655,7 +3657,7 @@ static DMCRet dmc_trace(DMCContext *context,
 
 static DMCRet dmc_caller(DMCContext *context,
  			 DMCHeap *heap,
- 			 DMC_STACK_TYPE(Uint) *text,
+			 DMC_STACK_TYPE(UWord) *text,
  			 Eterm t,
  			 int *constant)
 {
@@ -3688,7 +3690,7 @@ static DMCRet dmc_caller(DMCContext *context,
   
 static DMCRet dmc_silent(DMCContext *context,
  			 DMCHeap *heap,
- 			 DMC_STACK_TYPE(Uint) *text,
+			 DMC_STACK_TYPE(UWord) *text,
  			 Eterm t,
  			 int *constant)
 {
@@ -3730,7 +3732,7 @@ static DMCRet dmc_silent(DMCContext *context,
 
 static DMCRet dmc_fun(DMCContext *context,
 		       DMCHeap *heap,
-		       DMC_STACK_TYPE(Uint) *text,
+		       DMC_STACK_TYPE(UWord) *text,
 		       Eterm t,
 		       int *constant)
 {
@@ -3847,7 +3849,7 @@ static DMCRet dmc_fun(DMCContext *context,
 	erl_exit(1,"ets:match() internal error, "
 		 "guard with more than 3 arguments.");
     }
-    DMC_PUSH(*text, (Uint) b->biff);
+    DMC_PUSH(*text, (UWord) b->biff);
     context->stack_used -= (((int) a) - 2);
     if (context->stack_used > context->stack_need)
  	context->stack_need = context->stack_used;
@@ -3856,7 +3858,7 @@ static DMCRet dmc_fun(DMCContext *context,
 
 static DMCRet dmc_expr(DMCContext *context,
 		       DMCHeap *heap,
-		       DMC_STACK_TYPE(Uint) *text,
+		       DMC_STACK_TYPE(UWord) *text,
 		       Eterm t,
 		       int *constant)
 {
@@ -3919,7 +3921,7 @@ static DMCRet dmc_expr(DMCContext *context,
     
 static DMCRet compile_guard_expr(DMCContext *context,
 				 DMCHeap *heap,
-				 DMC_STACK_TYPE(Uint) *text,
+				 DMC_STACK_TYPE(UWord) *text,
 				 Eterm l)
 {
     DMCRet ret;
@@ -4233,7 +4235,7 @@ static Eterm match_spec_test(Process *p, Eterm against, Eterm spec, int trace)
     Eterm l;
     Uint32 ret_flags;
     Uint sz;
-    Eterm *save_cp;
+    UWord *save_cp;
 
     if (trace && !(is_list(against) || against == NIL)) {
 	return THE_NON_VALUE;
@@ -4276,7 +4278,7 @@ static Eterm match_spec_test(Process *p, Eterm against, Eterm spec, int trace)
 	    }
 	} else {
 	    n = 0;
-	    arr = (Eterm *) against;
+	    arr = (Eterm *) EXPAND_POINTER(against);
 	}
 	
 	/* We are in the context of a BIF, 
@@ -4327,7 +4329,7 @@ static Eterm seq_trace_fake(Process *p, Eterm arg1)
 static void db_match_dis(Binary *bp)
 {
     MatchProg *prog = Binary2MatchProg(bp);
-    Uint *t = prog->text;
+    UWord *t = prog->text;
     Uint n;
     Eterm p;
     int first;
@@ -4402,7 +4404,7 @@ static void db_match_dis(Binary *bp)
 		    first = 0;
 		else
 		    erts_printf(", ");
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 		erts_printf("0x%016bpx", *t);
 #else
 		erts_printf("0x%08bpx", *t);
@@ -4422,7 +4424,7 @@ static void db_match_dis(Binary *bp)
 		    first = 0;
 		else
 		    erts_printf(", ");
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 		erts_printf("0x%016bpx", *t);
 #else
 		erts_printf("0x%08bpx", *t);
diff --git a/erts/emulator/beam/erl_db_util.h b/erts/emulator/beam/erl_db_util.h
index 4fc7b4f52e..fbc25aaf2b 100644
--- a/erts/emulator/beam/erl_db_util.h
+++ b/erts/emulator/beam/erl_db_util.h
@@ -212,7 +212,7 @@ typedef struct db_table_common {
 #endif
     Eterm owner;              /* Pid of the creator */
     Eterm heir;               /* Pid of the heir */
-    Eterm heir_data;          /* To send in ETS-TRANSFER (is_immed or (DbTerm*) */
+    UWord heir_data;          /* To send in ETS-TRANSFER (is_immed or (DbTerm*) */
     SysTimeval heir_started;  /* To further identify the heir */
     Eterm the_name;           /* an atom */
     Eterm id;                 /* atom | integer */
@@ -304,9 +304,9 @@ typedef struct match_prog {
     Uint eheap_offset;
     Uint stack_offset;
 #ifdef DMC_DEBUG
-    Uint* prog_end;		/* End of program */
+    UWord* prog_end;		/* End of program */
 #endif
-    Uint text[1];		/* Beginning of program */
+    UWord text[1];		/* Beginning of program */
 } MatchProg;
 
 /*
diff --git a/erts/emulator/beam/erl_debug.c b/erts/emulator/beam/erl_debug.c
index 99b05fcf26..9c7a47da0a 100644
--- a/erts/emulator/beam/erl_debug.c
+++ b/erts/emulator/beam/erl_debug.c
@@ -235,9 +235,9 @@ pps(Process* p, Eterm* stop)
     }
 
     while(sp >= stop) {
-	erts_print(to, to_arg, "%0*lx: ", PTR_SIZE, (Eterm) sp);
+	erts_print(to, to_arg, "%0*lx: ", PTR_SIZE, (UWord) sp);
 	if (is_catch(*sp)) {
-	    erts_print(to, to_arg, "catch %d", (Uint)catch_pc(*sp));
+	    erts_print(to, to_arg, "catch %ld", (UWord)catch_pc(*sp));
 	} else {
 	    paranoid_display(to, to_arg, p, *sp);
 	}
diff --git a/erts/emulator/beam/erl_driver.h b/erts/emulator/beam/erl_driver.h
index 489e74d960..316d0ef992 100644
--- a/erts/emulator/beam/erl_driver.h
+++ b/erts/emulator/beam/erl_driver.h
@@ -567,7 +567,7 @@ EXTERN int driver_send_term(ErlDrvPort ix, ErlDrvTermData to,
 
 /* Async IO functions */
 EXTERN long driver_async(ErlDrvPort ix,
-			 unsigned int* key,
+			 unsigned long* key,
 			 void (*async_invoke)(void*), 
 			 void* async_data,
 			 void (*async_free)(void*));
diff --git a/erts/emulator/beam/erl_fun.c b/erts/emulator/beam/erl_fun.c
index 79e844b315..e6884c1179 100644
--- a/erts/emulator/beam/erl_fun.c
+++ b/erts/emulator/beam/erl_fun.c
@@ -50,8 +50,8 @@ static void fun_free(ErlFunEntry* obj);
  * to unloaded_fun[]. The -1 in unloaded_fun[0] will be interpreted
  * as an illegal arity when attempting to call a fun.
  */
-static Eterm unloaded_fun_code[3] = {NIL, -1, 0};
-static Eterm* unloaded_fun = unloaded_fun_code + 2;
+static UWord unloaded_fun_code[3] = {NIL, -1, 0};
+static UWord* unloaded_fun = unloaded_fun_code + 2;
 
 void
 erts_init_fun_table(void)
@@ -207,7 +207,7 @@ erts_cleanup_funs(ErlFunThing* funp)
 #endif
 
 void
-erts_cleanup_funs_on_purge(Eterm* start, Eterm* end)
+erts_cleanup_funs_on_purge(UWord* start, UWord* end)
 {
     int limit;
     HashBucket** bucket;
@@ -222,7 +222,7 @@ erts_cleanup_funs_on_purge(Eterm* start, Eterm* end)
 
 	while (b) {
 	    ErlFunEntry* fe = (ErlFunEntry *) b;
-	    Eterm* addr = fe->address;
+	    UWord* addr = fe->address;
 
 	    if (start <= addr && addr < end) {
 		fe->address = unloaded_fun;
diff --git a/erts/emulator/beam/erl_fun.h b/erts/emulator/beam/erl_fun.h
index fb5e75649b..980cdd414a 100644
--- a/erts/emulator/beam/erl_fun.h
+++ b/erts/emulator/beam/erl_fun.h
@@ -33,10 +33,10 @@ typedef struct erl_fun_entry {
     int index;			/* New style index. */
     int old_uniq;		/* Unique number (old_style) */
     int old_index;		/* Old style index */
-    Eterm* address;		/* Pointer to code for fun */
+    UWord* address;		/* Pointer to code for fun */
 
 #ifdef HIPE
-    Eterm* native_address;	/* Native entry code for fun. */
+    UWord* native_address;	/* Native entry code for fun. */
 #endif
 
     Uint arity;			/* The arity of the fun. */
@@ -58,7 +58,7 @@ typedef struct erl_fun_thing {
 #endif
     ErlFunEntry* fe;		/* Pointer to fun entry. */
 #ifdef HIPE
-    Eterm* native_address;	/* Native code for the fun. */
+    UWord* native_address;	/* Native code for the fun. */
 #endif
     Uint arity;			/* The arity of the fun. */
     Uint num_free;		/* Number of free variables (in env). */
@@ -86,7 +86,7 @@ void erts_erase_fun_entry(ErlFunEntry* fe);
 #ifndef HYBRID /* FIND ME! */
 void erts_cleanup_funs(ErlFunThing* funp);
 #endif
-void erts_cleanup_funs_on_purge(Eterm* start, Eterm* end);
+void erts_cleanup_funs_on_purge(UWord* start, UWord* end);
 void erts_dump_fun_entries(int, void *);
 
 #endif
diff --git a/erts/emulator/beam/erl_gc.c b/erts/emulator/beam/erl_gc.c
index e9bf37a173..9ed566e66e 100644
--- a/erts/emulator/beam/erl_gc.c
+++ b/erts/emulator/beam/erl_gc.c
@@ -126,7 +126,7 @@ static void disallow_heap_frag_ref_in_old_heap(Process* p);
 static void disallow_heap_frag_ref(Process* p, Eterm* n_htop, Eterm* objv, int nobj);
 #endif
 
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 # define MAX_HEAP_SIZES 154
 #else
 # define MAX_HEAP_SIZES 55
@@ -1551,7 +1551,7 @@ disallow_heap_frag_ref_in_old_heap(Process* p)
 	val = *hp++;
 	switch (primary_tag(val)) {
 	case TAG_PRIMARY_BOXED:
-	    ptr = (Eterm *) val;
+	    ptr = (Eterm *) EXPAND_POINTER(val);
 	    if (!in_area(ptr, old_heap, old_heap_size)) {
 		if (in_area(ptr, new_heap, new_heap_size)) {
 		    abort();
@@ -1564,7 +1564,7 @@ disallow_heap_frag_ref_in_old_heap(Process* p)
 	    }
 	    break;
 	case TAG_PRIMARY_LIST:
-	    ptr = (Eterm *) val;
+	    ptr = (Eterm *) EXPAND_POINTER(val);
 	    if (!in_area(ptr, old_heap, old_heap_size)) {
 		if (in_area(ptr, new_heap, new_heap_size)) {
 		    abort();
diff --git a/erts/emulator/beam/erl_goodfit_alloc.c b/erts/emulator/beam/erl_goodfit_alloc.c
index ea2ba4d55c..33d93d344d 100644
--- a/erts/emulator/beam/erl_goodfit_alloc.c
+++ b/erts/emulator/beam/erl_goodfit_alloc.c
@@ -49,30 +49,30 @@
 #define MIN_MBC_FIRST_FREE_SZ		(4*1024)
 
 #define MAX_SUB_MASK_IX \
-  ((((Uint)1) << (NO_OF_BKT_IX_BITS - SUB_MASK_IX_SHIFT)) - 1)
-#define MAX_SUB_BKT_IX ((((Uint)1) << SUB_MASK_IX_SHIFT) - 1)
+  ((((UWord)1) << (NO_OF_BKT_IX_BITS - SUB_MASK_IX_SHIFT)) - 1)
+#define MAX_SUB_BKT_IX ((((UWord)1) << SUB_MASK_IX_SHIFT) - 1)
 #define MAX_BKT_IX (NO_OF_BKTS - 1)
 
-#define MIN_BLK_SZ  UNIT_CEILING(sizeof(GFFreeBlock_t) + sizeof(Uint))
+#define MIN_BLK_SZ  UNIT_CEILING(sizeof(GFFreeBlock_t) + sizeof(UWord))
 
-#define IX2SBIX(IX) ((IX) & (~(~((Uint)0) << SUB_MASK_IX_SHIFT)))
+#define IX2SBIX(IX) ((IX) & (~(~((UWord)0) << SUB_MASK_IX_SHIFT)))
 #define IX2SMIX(IX) ((IX) >> SUB_MASK_IX_SHIFT)
 #define MAKE_BKT_IX(SMIX, SBIX)	\
-  ((((Uint)(SMIX)) << SUB_MASK_IX_SHIFT) | ((Uint)(SBIX)))
+  ((((UWord)(SMIX)) << SUB_MASK_IX_SHIFT) | ((UWord)(SBIX)))
 
 #define SET_BKT_MASK_IX(BM, IX)						\
 do {									\
     int sub_mask_ix__ = IX2SMIX((IX));					\
-    (BM).main |= (((Uint) 1) << sub_mask_ix__);				\
-    (BM).sub[sub_mask_ix__] |= (((Uint)1) << IX2SBIX((IX)));		\
+    (BM).main |= (((UWord) 1) << sub_mask_ix__);				\
+    (BM).sub[sub_mask_ix__] |= (((UWord)1) << IX2SBIX((IX)));		\
 } while (0)
 
 #define UNSET_BKT_MASK_IX(BM, IX)					\
 do {									\
     int sub_mask_ix__ = IX2SMIX((IX));					\
-    (BM).sub[sub_mask_ix__] &= ~(((Uint)1) << IX2SBIX((IX)));		\
+    (BM).sub[sub_mask_ix__] &= ~(((UWord)1) << IX2SBIX((IX)));		\
     if (!(BM).sub[sub_mask_ix__])					\
-	(BM).main &= ~(((Uint)1) << sub_mask_ix__);			\
+	(BM).main &= ~(((UWord)1) << sub_mask_ix__);			\
 } while (0)
 
 /* Buckets ... */
@@ -263,8 +263,8 @@ find_bucket(BucketMask_t *bmask, int min_index)
     while(max != min) {					\
 	mid = ((max - min) >> 1) + min;			\
 	if((BitMask)					\
-	   & (~(~((Uint) 0) << (mid + 1)))		\
-	   & (~((Uint) 0) << min))			\
+	   & (~(~((UWord) 0) << (mid + 1)))		\
+	   & (~((UWord) 0) << min))			\
 	    max = mid;					\
 	else						\
 	    min = mid + 1;				\
@@ -272,21 +272,21 @@ find_bucket(BucketMask_t *bmask, int min_index)
     (MinBit) = min
 
 
-    ASSERT(bmask->main < (((Uint) 1) << (MAX_SUB_MASK_IX+1)));
+    ASSERT(bmask->main < (((UWord) 1) << (MAX_SUB_MASK_IX+1)));
 
     sub_mask_ix = IX2SMIX(min_index);
 
-    if ((bmask->main & (~((Uint) 0) << sub_mask_ix)) == 0)
+    if ((bmask->main & (~((UWord) 0) << sub_mask_ix)) == 0)
 	return -1;
 
     /* There exists a non empty bucket; find it... */
 
-    if (bmask->main & (((Uint) 1) << sub_mask_ix)) {
+    if (bmask->main & (((UWord) 1) << sub_mask_ix)) {
 	sub_bkt_ix = IX2SBIX(min_index);
-	if ((bmask->sub[sub_mask_ix] & (~((Uint) 0) << sub_bkt_ix)) == 0) {
+	if ((bmask->sub[sub_mask_ix] & (~((UWord) 0) << sub_bkt_ix)) == 0) {
 	    sub_mask_ix++;
 	    sub_bkt_ix = 0;
-	    if ((bmask->main & (~((Uint) 0)<< sub_mask_ix)) == 0)
+	    if ((bmask->main & (~((UWord) 0)<< sub_mask_ix)) == 0)
 		return -1;
 	}
 	else
@@ -299,17 +299,17 @@ find_bucket(BucketMask_t *bmask, int min_index)
 
     ASSERT(sub_mask_ix <= MAX_SUB_MASK_IX);
     /* Has to be a bit > sub_mask_ix */
-    ASSERT(bmask->main & (~((Uint) 0) << (sub_mask_ix)));
+    ASSERT(bmask->main & (~((UWord) 0) << (sub_mask_ix)));
     GET_MIN_BIT(sub_mask_ix, bmask->main, sub_mask_ix, MAX_SUB_MASK_IX);
 
  find_sub_bkt_ix:
     ASSERT(sub_mask_ix <= MAX_SUB_MASK_IX);
     ASSERT(sub_bkt_ix <= MAX_SUB_BKT_IX);
 
-    if ((bmask->sub[sub_mask_ix] & (((Uint) 1) << sub_bkt_ix)) == 0) {
+    if ((bmask->sub[sub_mask_ix] & (((UWord) 1) << sub_bkt_ix)) == 0) {
 	ASSERT(sub_mask_ix + 1 <= MAX_SUB_BKT_IX);
 	/* Has to be a bit > sub_bkt_ix */
-	ASSERT(bmask->sub[sub_mask_ix] & (~((Uint) 0) << sub_bkt_ix));
+	ASSERT(bmask->sub[sub_mask_ix] & (~((UWord) 0) << sub_bkt_ix));
 
 	GET_MIN_BIT(sub_bkt_ix,
 		    bmask->sub[sub_mask_ix],
@@ -336,7 +336,7 @@ search_bucket(Allctr_t *allctr, int ix, Uint size)
     Uint min_sz;
     Uint blk_sz;
     Uint cand_sz = 0;
-    Uint max_blk_search;
+    UWord max_blk_search;
     GFFreeBlock_t *blk;
     GFFreeBlock_t *cand = NULL;
     int blk_on_lambc;
@@ -615,9 +615,9 @@ check_block(Allctr_t *allctr, Block_t * blk, int free_block)
 	Uint blk_sz = BLK_SZ(blk);
 	bi = BKT_IX(gfallctr, blk_sz);
 
-	ASSERT(gfallctr->bucket_mask.main & (((Uint) 1) << IX2SMIX(bi)));
+	ASSERT(gfallctr->bucket_mask.main & (((UWord) 1) << IX2SMIX(bi)));
 	ASSERT(gfallctr->bucket_mask.sub[IX2SMIX(bi)]
-	       & (((Uint) 1) << IX2SBIX(bi)));
+	       & (((UWord) 1) << IX2SBIX(bi)));
 		
 	found = 0;
 	for (fblk = gfallctr->buckets[bi]; fblk; fblk = fblk->next)
@@ -648,9 +648,9 @@ check_mbc(Allctr_t *allctr, Carrier_t *mbc)
     int bi;
 
     for(bi = 0; bi < NO_OF_BKTS; bi++) {
-	if ((gfallctr->bucket_mask.main & (((Uint) 1) << IX2SMIX(bi)))
+	if ((gfallctr->bucket_mask.main & (((UWord) 1) << IX2SMIX(bi)))
 	    && (gfallctr->bucket_mask.sub[IX2SMIX(bi)]
-		& (((Uint) 1) << IX2SBIX(bi)))) {
+		& (((UWord) 1) << IX2SBIX(bi)))) {
 	    ASSERT(gfallctr->buckets[bi] != NULL);
 	}
 	else {
diff --git a/erts/emulator/beam/erl_goodfit_alloc.h b/erts/emulator/beam/erl_goodfit_alloc.h
index 3d1b8c01f6..648800ce31 100644
--- a/erts/emulator/beam/erl_goodfit_alloc.h
+++ b/erts/emulator/beam/erl_goodfit_alloc.h
@@ -28,7 +28,7 @@
 typedef struct GFAllctr_t_ GFAllctr_t;
 
 typedef struct {
-    Uint mbsd;
+    UWord mbsd;
 } GFAllctrInit_t;
 
 #define ERTS_DEFAULT_GF_ALLCTR_INIT {                                      \
@@ -49,18 +49,18 @@ Allctr_t *erts_gfalc_start(GFAllctr_t *, GFAllctrInit_t *, AllctrInit_t *);
 #include "erl_alloc_util.h"
 
 #define NO_OF_BKT_IX_BITS (8)
-#ifdef ARCH_64
+#if defined(ARCH_64)
 #  define SUB_MASK_IX_SHIFT (6)
 #else
 #  define SUB_MASK_IX_SHIFT (5)
 #endif
-#define NO_OF_BKTS (((Uint) 1) << NO_OF_BKT_IX_BITS)
-#define NO_OF_SUB_MASKS (NO_OF_BKTS/(((Uint) 1) << SUB_MASK_IX_SHIFT))
+#define NO_OF_BKTS (((UWord) 1) << NO_OF_BKT_IX_BITS)
+#define NO_OF_SUB_MASKS (NO_OF_BKTS/(((UWord) 1) << SUB_MASK_IX_SHIFT))
 
 typedef struct {
-    Uint main;
-    Uint sub[NO_OF_SUB_MASKS];
-} BucketMask_t; 
+    UWord main;
+    UWord sub[NO_OF_SUB_MASKS];
+} BucketMask_t;
 
 typedef struct GFFreeBlock_t_ GFFreeBlock_t;
 struct GFFreeBlock_t_ {
@@ -74,11 +74,11 @@ struct GFAllctr_t_ {
 
     char *		last_aux_mbc_start;
     char *		last_aux_mbc_end;
-    Uint		bkt_max_size_d;
-    Uint		bkt_intrvl_d;
+    UWord		bkt_max_size_d;
+    UWord		bkt_intrvl_d;
     BucketMask_t	bucket_mask;
     GFFreeBlock_t *	buckets[NO_OF_BKTS];
-    Uint 		max_blk_search;
+    UWord 		max_blk_search;
 
 };
 
diff --git a/erts/emulator/beam/erl_instrument.c b/erts/emulator/beam/erl_instrument.c
index 3f022f92b8..62be9d379b 100644
--- a/erts/emulator/beam/erl_instrument.c
+++ b/erts/emulator/beam/erl_instrument.c
@@ -540,18 +540,18 @@ static void dump_memory_map_to_stream(FILE *fp)
 	if (is_internal_pid(bp->pid))
 	    fprintf(fp,
 		    "{%lu, %lu, %lu, {%lu,%lu,%lu}}.\n",
-		    (Uint) bp->type_no,
-		    (Uint) bp->mem,
-		    (Uint) bp->size,
-		    (Uint) pid_channel_no(bp->pid),
-		    (Uint) pid_number(bp->pid),
-		    (Uint) pid_serial(bp->pid));
+		    (UWord) bp->type_no,
+		    (UWord) bp->mem,
+		    (UWord) bp->size,
+		    (UWord) pid_channel_no(bp->pid),
+		    (UWord) pid_number(bp->pid),
+		    (UWord) pid_serial(bp->pid));
 	else
 	    fprintf(fp,
 		    "{%lu, %lu, %lu, undefined}.\n",
-		    (Uint) bp->type_no,
-		    (Uint) bp->mem,
-		    (Uint) bp->size);
+		    (UWord) bp->type_no,
+		    (UWord) bp->mem,
+		    (UWord) bp->size);
     }
 
     if (lock)
@@ -638,7 +638,7 @@ Eterm erts_instr_get_memory_map(Process *proc)
 	    hsz += 4;
 	}
 
-	if ((Uint) bp->mem > MAX_SMALL)
+	if ((UWord) bp->mem > MAX_SMALL)
 	    hsz += BIG_UINT_HEAP_SIZE;
 	if (bp->size > MAX_SMALL)
 	    hsz += BIG_UINT_HEAP_SIZE;
@@ -749,12 +749,12 @@ Eterm erts_instr_get_memory_map(Process *proc)
 #endif
 	type = make_small(bp->type_no);
 
-	if ((Uint) bp->mem > MAX_SMALL) {
-	    ptr = uint_to_big((Uint) bp->mem, hp);
+	if ((UWord) bp->mem > MAX_SMALL) {
+	    ptr = uint_to_big((UWord) bp->mem, hp);
 	    hp += BIG_UINT_HEAP_SIZE;
 	}
 	else
-	    ptr = make_small((Uint) bp->mem);
+	    ptr = make_small((UWord) bp->mem);
 
 	if (bp->size > MAX_SMALL) {
 	    size = uint_to_big(bp->size, hp);
@@ -962,12 +962,12 @@ dump_stat_to_stream(FILE *fp, int begin_max_period)
 
     fprintf(fp,
 	    "{total,[{total,[{sizes,%lu,%lu,%lu},{blocks,%lu,%lu,%lu}]}]}.\n",
-	    stats->tot.size,
-	    stats->tot.max_size,
-	    stats->tot.max_size_ever,
-	    stats->tot.blocks,
-	    stats->tot.max_blocks,
-	    stats->tot.max_blocks_ever);
+	    (UWord) stats->tot.size,
+	    (UWord) stats->tot.max_size,
+	    (UWord) stats->tot.max_size_ever,
+	    (UWord) stats->tot.blocks,
+	    (UWord) stats->tot.max_blocks,
+	    (UWord) stats->tot.max_blocks_ever);
 
     a_max = 0;
     a_min = ~0;
@@ -992,12 +992,12 @@ dump_stat_to_stream(FILE *fp, int begin_max_period)
 		    "%s{%s,[{sizes,%lu,%lu,%lu},{blocks,%lu,%lu,%lu}]}%s",
 		    i == a_min ? "{allocators,\n [" : "  ",
 		    ERTS_ALC_A2AD(i),
-		    stats->a[i].size,
-		    stats->a[i].max_size,
-		    stats->a[i].max_size_ever,
-		    stats->a[i].blocks,
-		    stats->a[i].max_blocks,
-		    stats->a[i].max_blocks_ever,
+		    (UWord) stats->a[i].size,
+		    (UWord) stats->a[i].max_size,
+		    (UWord) stats->a[i].max_size_ever,
+		    (UWord) stats->a[i].blocks,
+		    (UWord) stats->a[i].max_blocks,
+		    (UWord) stats->a[i].max_blocks_ever,
 		    i == a_max ? "]}.\n" : ",\n");
 	}
     }
@@ -1009,12 +1009,12 @@ dump_stat_to_stream(FILE *fp, int begin_max_period)
 		"%s{%s,[{sizes,%lu,%lu,%lu},{blocks,%lu,%lu,%lu}]}%s",
 		i == ERTS_ALC_C_MIN ? "{classes,\n [" : "  ",
 		ERTS_ALC_C2CD(i),
-		stats->c[i].size,
-		stats->c[i].max_size,
-		stats->c[i].max_size_ever,
-		stats->c[i].blocks,
-		stats->c[i].max_blocks,
-		stats->c[i].max_blocks_ever,
+		(UWord) stats->c[i].size,
+		(UWord) stats->c[i].max_size,
+		(UWord) stats->c[i].max_size_ever,
+		(UWord) stats->c[i].blocks,
+		(UWord) stats->c[i].max_blocks,
+		(UWord) stats->c[i].max_blocks_ever,
 		i == ERTS_ALC_C_MAX ? "]}.\n" :  ",\n" );
     }
 
@@ -1025,12 +1025,12 @@ dump_stat_to_stream(FILE *fp, int begin_max_period)
 		"%s{%s,[{sizes,%lu,%lu,%lu},{blocks,%lu,%lu,%lu}]}%s",
 		i == ERTS_ALC_N_MIN ? "{types,\n [" : "  ",
 		ERTS_ALC_N2TD(i),
-		stats->n[i].size,
-		stats->n[i].max_size,
-		stats->n[i].max_size_ever,
-		stats->n[i].blocks,
-		stats->n[i].max_blocks,
-		stats->n[i].max_blocks_ever,
+		(UWord) stats->n[i].size,
+		(UWord) stats->n[i].max_size,
+		(UWord) stats->n[i].max_size_ever,
+		(UWord) stats->n[i].blocks,
+		(UWord) stats->n[i].max_blocks,
+		(UWord) stats->n[i].max_blocks_ever,
 		i == ERTS_ALC_N_MAX ? "]}.\n" :  ",\n" );
     }
 
diff --git a/erts/emulator/beam/erl_message.h b/erts/emulator/beam/erl_message.h
index 5cf7c209bd..459c6363aa 100644
--- a/erts/emulator/beam/erl_message.h
+++ b/erts/emulator/beam/erl_message.h
@@ -49,7 +49,7 @@ typedef struct erl_heap_fragment ErlHeapFragment;
 struct erl_heap_fragment {
     ErlHeapFragment* next;	/* Next heap fragment */
     ErlOffHeap off_heap;	/* Offset heap data. */
-    unsigned size;		/* Size in words of mem */
+    unsigned size;		/* Size in (half)words of mem */
     unsigned used_size;         /* With terms to be moved to heap by GC */
     Eterm mem[1];		/* Data */
 };
diff --git a/erts/emulator/beam/erl_mtrace.c b/erts/emulator/beam/erl_mtrace.c
index 8b8ac2ec80..9f2aa62a9a 100644
--- a/erts/emulator/beam/erl_mtrace.c
+++ b/erts/emulator/beam/erl_mtrace.c
@@ -79,7 +79,7 @@ static erts_mtx_t mtrace_buf_mutex;
 #define UI16_SZ					(2)
 #define UI32_SZ					(4)
 #define UI64_SZ					(8)
-#ifdef ARCH_64
+#ifdef ARCH_64 /* XXX:PaN Halfword? (whole file...) */
 #  define UI_SZ					UI64_SZ
 #else
 #  define UI_SZ					UI32_SZ
@@ -188,7 +188,7 @@ check_alloc_entry(byte *sp, byte *ep,
 		  byte tag,
 		  Uint16 ct_no, int ct_no_n,
 		  Uint16 type, int type_n,
-		  Uint res, int res_n,
+		  UWord res, int res_n,
 		  Uint size, int size_n,
 		  Uint32 ti,int ti_n);
 void
@@ -196,8 +196,8 @@ check_realloc_entry(byte *sp, byte *ep,
 		    byte tag,
 		    Uint16 ct_no, int ct_no_n,
 		    Uint16 type, int type_n,
-		    Uint res, int res_n,
-		    Uint ptr, int ptr_n,
+		    UWord res, int res_n,
+		    UWord ptr, int ptr_n,
 		    Uint size, int size_n,
 		    Uint32 ti,int ti_n);
 void
@@ -205,7 +205,7 @@ check_free_entry(byte *sp, byte *ep,
 		 byte tag,
 		 Uint16 ct_no, int ct_no_n,
 		 Uint16 t_no, int t_no_n,
-		 Uint ptr, int ptr_n,
+		 UWord ptr, int ptr_n,
 		 Uint32 ti,int ti_n);
 void
 check_time_inc_entry(byte *sp, byte *ep,
@@ -785,7 +785,7 @@ write_alloc_entry(byte tag,
 			      tag,
 			      ct_no, ct_no_n,
 			      t_no, t_no_n,
-			      (Uint) res, res_n,
+			      (UWord) res, res_n,
 			      size, size_n,
 			      ti, ti_n);
 #endif
@@ -865,8 +865,8 @@ write_realloc_entry(byte tag,
 				tag,
 				ct_no, ct_no_n,
 				t_no, t_no_n,
-				(Uint) res, res_n,
-				(Uint) ptr, ptr_n,
+				(UWord) res, res_n,
+				(UWord) ptr, ptr_n,
 				size, size_n,
 				ti, ti_n);
 #endif
@@ -934,7 +934,7 @@ write_free_entry(byte tag,
 			     tag,
 			     ct_no, ct_no_n,
 			     t_no, t_no_n,
-			     (Uint) ptr, ptr_n,
+			     (UWord) ptr, ptr_n,
 			     ti, ti_n);
 #endif
 	}
@@ -1135,7 +1135,7 @@ check_alloc_entry(byte *sp, byte *ep,
 		  byte tag,
 		  Uint16 ct_no, int ct_no_n,
 		  Uint16 t_no, int t_no_n,
-		  Uint res, int res_n,
+		  UWord res, int res_n,
 		  Uint size, int size_n,
 		  Uint32 ti,int ti_n)
 {
@@ -1163,8 +1163,8 @@ check_realloc_entry(byte *sp, byte *ep,
 		    byte tag,
 		    Uint16 ct_no, int ct_no_n,
 		    Uint16 t_no, int t_no_n,
-		    Uint res, int res_n,
-		    Uint ptr, int ptr_n,
+		    UWord res, int res_n,
+		    UWord ptr, int ptr_n,
 		    Uint size, int size_n,
 		    Uint32 ti,int ti_n)
 {
@@ -1193,7 +1193,7 @@ check_free_entry(byte *sp, byte *ep,
 		 byte tag,
 		 Uint16 ct_no, int ct_no_n,
 		 Uint16 t_no, int t_no_n,
-		 Uint ptr, int ptr_n,
+		 UWord ptr, int ptr_n,
 		 Uint32 ti,int ti_n)
 {
     byte *p = sp;
diff --git a/erts/emulator/beam/erl_nif.c b/erts/emulator/beam/erl_nif.c
index 585a6c1fdf..7c62bcb016 100644
--- a/erts/emulator/beam/erl_nif.c
+++ b/erts/emulator/beam/erl_nif.c
@@ -496,9 +496,9 @@ int enif_get_atom(ErlNifEnv* env, Eterm atom, char* buf, unsigned len)
 
 int enif_get_int(ErlNifEnv* env, Eterm term, int* ip)
 {
-#if SIZEOF_INT == SIZEOF_VOID_P
+#if SIZEOF_INT ==  ERTS_SIZEOF_ETERM
     return term_to_Sint(term, (Sint*)ip);
-#elif SIZEOF_LONG == SIZEOF_VOID_P
+#elif SIZEOF_LONG ==  ERTS_SIZEOF_ETERM
     Sint i;
     if (!term_to_Sint(term, &i) || i < INT_MIN || i > INT_MAX) {
 	return 0;
@@ -512,9 +512,9 @@ int enif_get_int(ErlNifEnv* env, Eterm term, int* ip)
 
 int enif_get_uint(ErlNifEnv* env, Eterm term, unsigned* ip)
 {
-#if SIZEOF_INT == SIZEOF_VOID_P
+#if SIZEOF_LONG == ERTS_SIZEOF_ETERM
     return term_to_Uint(term, (Uint*)ip);
-#elif SIZEOF_LONG == SIZEOF_VOID_P
+#elif SIZEOF_INT == ERTS_SIZEOF_ETERM
     Uint i;
     if (!term_to_Uint(term, &i) || i > UINT_MAX) {
 	return 0;
@@ -526,8 +526,13 @@ int enif_get_uint(ErlNifEnv* env, Eterm term, unsigned* ip)
 
 int enif_get_long(ErlNifEnv* env, Eterm term, long* ip)
 {
-#if SIZEOF_LONG == SIZEOF_VOID_P
+#if SIZEOF_LONG == ERTS_SIZEOF_ETERM
     return term_to_Sint(term, ip);
+#elif SIZEOF_INT == ERTS_SIZEOF_ETERM
+    Uint u;
+    term_to_Sint(term, u);
+    *ip = (long) u;
+    return 1;
 #else
 #  error Unknown long word size 
 #endif     
@@ -535,8 +540,14 @@ int enif_get_long(ErlNifEnv* env, Eterm term, long* ip)
 
 int enif_get_ulong(ErlNifEnv* env, Eterm term, unsigned long* ip)
 {
-#if SIZEOF_LONG == SIZEOF_VOID_P
+#if SIZEOF_LONG == ERTS_SIZEOF_ETERM
     return term_to_Uint(term, ip);
+#elif SIZEOF_INT == ERTS_SIZEOF_ETERM
+    Uint u;
+    int r;
+    r = term_to_Uint(term, &u);
+    *ip = (unsigned long) u;
+    return r;
 #else
 #  error Unknown long word size 
 #endif     
@@ -565,27 +576,24 @@ int enif_get_list_cell(ErlNifEnv* env, Eterm term, Eterm* head, Eterm* tail)
 
 ERL_NIF_TERM enif_make_int(ErlNifEnv* env, int i)
 {
-#if SIZEOF_INT == SIZEOF_VOID_P
+#if SIZEOF_INT == ERTS_SIZEOF_ETERM
     return IS_SSMALL(i) ? make_small(i) : small_to_big(i,alloc_heap(env,2));
-#elif SIZEOF_LONG == SIZEOF_VOID_P
+#elif SIZEOF_LONG == ERTS_SIZEOF_ETERM
     return make_small(i);
 #endif
 }
 
 ERL_NIF_TERM enif_make_uint(ErlNifEnv* env, unsigned i)
 {
-#if SIZEOF_INT == SIZEOF_VOID_P
+#if SIZEOF_INT == ERTS_SIZEOF_ETERM
     return IS_USMALL(0,i) ? make_small(i) : uint_to_big(i,alloc_heap(env,2));
-#elif SIZEOF_LONG == SIZEOF_VOID_P
+#elif SIZEOF_LONG == ERTS_SIZEOF_ETERM
     return make_small(i);
 #endif
 }
 
 ERL_NIF_TERM enif_make_long(ErlNifEnv* env, long i)
 {
-#if SIZEOF_LONG != SIZEOF_VOID_P
-#  error Unknown long word size 
-#endif     
     return IS_SSMALL(i) ? make_small(i) : small_to_big(i, alloc_heap(env,2));
 }
 
@@ -967,34 +975,34 @@ unsigned enif_sizeof_resource(ErlNifEnv* env, void* obj)
  ***************************************************************************/
 
 
-static Uint** get_func_pp(Eterm* mod_code, Eterm f_atom, unsigned arity)
+static UWord** get_func_pp(UWord* mod_code, Eterm f_atom, unsigned arity)
 {
     int n = (int) mod_code[MI_NUM_FUNCTIONS];
     int j;
     for (j = 0; j < n; ++j) {
-	Uint* code_ptr = (Uint*) mod_code[MI_FUNCTIONS+j];
-	ASSERT(code_ptr[0] == (Uint) BeamOp(op_i_func_info_IaaI));
+	UWord* code_ptr = (UWord*) mod_code[MI_FUNCTIONS+j];
+	ASSERT(code_ptr[0] == (UWord) BeamOp(op_i_func_info_IaaI));
 	if (f_atom == ((Eterm) code_ptr[3])
 	    && arity == ((unsigned) code_ptr[4])) {
 
-	    return (Uint**) &mod_code[MI_FUNCTIONS+j];
+	    return (UWord**) &mod_code[MI_FUNCTIONS+j];
 	}
     }
     return NULL;
 }
 
-/*static void refresh_cached_nif_data(Eterm* mod_code,
+/*static void refresh_cached_nif_data(BeamInstr* mod_code,
 				    struct erl_module_nif* mod_nif)
 {
     int i;
     for (i=0; i < mod_nif->entry->num_of_funcs; i++) {
 	Eterm f_atom;
 	ErlNifFunc* func = &mod_nif->entry->funcs[i];	
-	Uint* code_ptr;
+	BeamInstr* code_ptr;
 	
 	erts_atom_get(func->name, sys_strlen(func->name), &f_atom); 
 	code_ptr = *get_func_pp(mod_code, f_atom, func->arity);
-	code_ptr[5+2] = (Uint) mod_nif->priv_data;
+	code_ptr[5+2] = (UWord) mod_nif->priv_data;
     }
 }*/
 
@@ -1098,7 +1106,7 @@ BIF_RETTYPE load_nif_2(BIF_ALIST_2)
     Module* mod;
     Eterm mod_atom;
     Eterm f_atom;
-    Eterm* caller;
+    UWord* caller;
     ErtsSysDdllError errdesc = ERTS_SYS_DDLL_ERROR_INIT;
     Eterm ret = am_ok;
     int veto;
@@ -1165,7 +1173,7 @@ BIF_RETTYPE load_nif_2(BIF_ALIST_2)
 	/*erts_fprintf(stderr, "Found module %T\r\n", mod_atom);*/
     
 	for (i=0; i < entry->num_of_funcs && ret==am_ok; i++) {
-	    Uint** code_pp;
+	    UWord** code_pp;
 	    ErlNifFunc* f = &entry->funcs[i];
 	    if (!erts_atom_get(f->name, sys_strlen(f->name), &f_atom)
 		|| (code_pp = get_func_pp(mod->code, f_atom, f->arity))==NULL) { 
@@ -1268,19 +1276,19 @@ BIF_RETTYPE load_nif_2(BIF_ALIST_2)
         mod->nif = lib; 
 	for (i=0; i < entry->num_of_funcs; i++)
 	{
-	    Uint* code_ptr;
+	    BeamInstr* code_ptr;
 	    erts_atom_get(entry->funcs[i].name, sys_strlen(entry->funcs[i].name), &f_atom); 
 	    code_ptr = *get_func_pp(mod->code, f_atom, entry->funcs[i].arity); 
 	    
 	    if (code_ptr[1] == 0) {
-		code_ptr[5+0] = (Uint) BeamOp(op_call_nif);		
+		code_ptr[5+0] = (BeamInstr) BeamOp(op_call_nif);
 	    }
 	    else { /* Function traced, patch the original instruction word */
 		BpData* bp = (BpData*) code_ptr[1];
-	        bp->orig_instr = (Uint) BeamOp(op_call_nif); 
+	        bp->orig_instr = (BeamInstr) BeamOp(op_call_nif);
 	    }	    
-	    code_ptr[5+1] = (Uint) entry->funcs[i].fptr;
-	    code_ptr[5+2] = (Uint) lib;
+	    code_ptr[5+1] = (BeamInstr) entry->funcs[i].fptr;
+	    code_ptr[5+2] = (BeamInstr) lib;
 	}
     }
     else {    
diff --git a/erts/emulator/beam/erl_node_container_utils.h b/erts/emulator/beam/erl_node_container_utils.h
index 87dbfc2a04..f620582bdb 100644
--- a/erts/emulator/beam/erl_node_container_utils.h
+++ b/erts/emulator/beam/erl_node_container_utils.h
@@ -251,7 +251,7 @@ extern int erts_use_r9_pids_ports;
  * Refs                                                                    *
 \*                                                                         */
 
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 
 #define internal_ref_no_of_numbers(x)					\
   (internal_ref_data((x))[0])
diff --git a/erts/emulator/beam/erl_node_tables.c b/erts/emulator/beam/erl_node_tables.c
index f3f10201b5..6960844e99 100644
--- a/erts/emulator/beam/erl_node_tables.c
+++ b/erts/emulator/beam/erl_node_tables.c
@@ -1109,11 +1109,21 @@ insert_offheap(ErlOffHeap *oh, int type, Eterm id)
 			break;
 		    }
 		if (insert_bin) {
+#if HALFWORD_HEAP
+		    UWord val = (UWord) pb->val;
+		    DeclareTmpHeapNoproc(id_heap,BIG_UINT_HEAP_SIZE*2); /* extra place allocated */
+#else
 		    DeclareTmpHeapNoproc(id_heap,BIG_UINT_HEAP_SIZE);
+#endif
 		    Uint *hp = &id_heap[0];
 		    InsertedBin *nib;
+#if HALFWORD_HEAP
+		    UseTmpHeapNoproc(BIG_UWORD_HEAP_SIZE(val));
+		    a.id = erts_bld_uword(&hp, NULL, (UWord) val);
+#else
 		    UseTmpHeapNoproc(BIG_UINT_HEAP_SIZE);
 		    a.id = erts_bld_uint(&hp, NULL, (Uint) pb->val);
+#endif
 		    erts_match_prog_foreach_offheap(pb->val,
 						    insert_offheap2,
 						    (void *) &a);
@@ -1121,7 +1131,11 @@ insert_offheap(ErlOffHeap *oh, int type, Eterm id)
 		    nib->bin_val = pb->val;
 		    nib->next = inserted_bins;
 		    inserted_bins = nib;
+#if HALFWORD_HEAP
+		    UnUseTmpHeapNoproc(BIG_UINT_HEAP_SIZE*2);
+#else
 		    UnUseTmpHeapNoproc(BIG_UINT_HEAP_SIZE);
+#endif
 		}
 	    }
 	}
diff --git a/erts/emulator/beam/erl_node_tables.h b/erts/emulator/beam/erl_node_tables.h
index c48dac6219..3f53de9cf2 100644
--- a/erts/emulator/beam/erl_node_tables.h
+++ b/erts/emulator/beam/erl_node_tables.h
@@ -61,7 +61,7 @@
 #define ERTS_DE_QFLGS_ALL			(ERTS_DE_QFLG_BUSY \
 						 | ERTS_DE_QFLG_EXIT)
 
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 #define ERTS_DIST_OUTPUT_BUF_DBG_PATTERN ((Uint) 0xf713f713f713f713UL)
 #else
 #define ERTS_DIST_OUTPUT_BUF_DBG_PATTERN ((Uint) 0xf713f713)
diff --git a/erts/emulator/beam/erl_printf_term.c b/erts/emulator/beam/erl_printf_term.c
index 7fe3f3bca5..9ad0965144 100644
--- a/erts/emulator/beam/erl_printf_term.c
+++ b/erts/emulator/beam/erl_printf_term.c
@@ -246,7 +246,7 @@ print_term(fmtfn_t fn, void* arg, Eterm obj, long *dcount)
 
     if (is_CP(obj)) {
 	PRINT_STRING(res, fn, arg, "<cp/header:");
-	PRINT_POINTER(res, fn, arg, obj);
+	PRINT_POINTER(res, fn, arg, cp_val(obj));
 	PRINT_CHAR(res, fn, arg, '>');
 	return res;
     }
@@ -406,7 +406,7 @@ 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(obj) + 1));
 	    Atom* module = atom_tab(atom_val(ep->code[0]));
 	    Atom* name = atom_tab(atom_val(ep->code[1]));
 
@@ -438,7 +438,7 @@ print_term(fmtfn_t fn, void* arg, Eterm obj, long *dcount)
 	break;
     default:
 	PRINT_STRING(res, fn, arg, "<unknown:");
-	PRINT_POINTER(res, fn, arg, obj);
+	PRINT_POINTER(res, fn, arg, (UWord) obj);
 	PRINT_CHAR(res, fn, arg, '>');
 	break;
     }
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index f5b1308ca4..429bdd74fb 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -91,9 +91,9 @@ do {								\
 #define ERTS_EMPTY_RUNQ(RQ) \
   ((RQ)->len == 0 && (RQ)->misc.start == NULL)
 
-extern Eterm beam_apply[];
-extern Eterm beam_exit[];
-extern Eterm beam_continue_exit[];
+extern UWord beam_apply[];
+extern UWord beam_exit[];
+extern UWord beam_continue_exit[];
 
 static Sint p_last;
 static Sint p_next;
@@ -334,7 +334,7 @@ do {									\
 static void init_processes_bif(void);
 static void save_terminating_process(Process *p);
 static void exec_misc_ops(ErtsRunQueue *);
-static void print_function_from_pc(int to, void *to_arg, Eterm* x);
+static void print_function_from_pc(int to, void *to_arg, UWord* x);
 static int stack_element_dump(int to, void *to_arg, Process* p, Eterm* sp,
 			      int yreg);
 #ifdef ERTS_SMP
@@ -2078,9 +2078,9 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online)
 
     erts_aligned_run_queues = erts_alloc(ERTS_ALC_T_RUNQS,
 					 (sizeof(ErtsAlignedRunQueue)*(n+1)));
-    if ((((Uint) erts_aligned_run_queues) & ERTS_CACHE_LINE_MASK) == 0)
+    if ((((UWord) erts_aligned_run_queues) & ERTS_CACHE_LINE_MASK) == 0)
 	erts_aligned_run_queues = ((ErtsAlignedRunQueue *)
-				   ((((Uint) erts_aligned_run_queues)
+				   ((((UWord) erts_aligned_run_queues)
 				     & ~ERTS_CACHE_LINE_MASK)
 				    + ERTS_CACHE_LINE_SIZE));
 
@@ -2175,9 +2175,9 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online)
     erts_aligned_scheduler_data = erts_alloc(ERTS_ALC_T_SCHDLR_DATA,
 					     (sizeof(ErtsAlignedSchedulerData)
 					      *(n+1)));
-    if ((((Uint) erts_aligned_scheduler_data) & ERTS_CACHE_LINE_MASK) == 0)
+    if ((((UWord) erts_aligned_scheduler_data) & ERTS_CACHE_LINE_MASK) == 0)
 	erts_aligned_scheduler_data = ((ErtsAlignedSchedulerData *)
-				       ((((Uint) erts_aligned_scheduler_data)
+				       ((((UWord) erts_aligned_scheduler_data)
 					 & ~ERTS_CACHE_LINE_MASK)
 					+ ERTS_CACHE_LINE_SIZE));
     for (ix = 0; ix < n; ix++) {
@@ -2186,6 +2186,10 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online)
 	erts_bits_init_state(&esdp->erl_bits_state);
 	esdp->match_pseudo_process = NULL;
 	esdp->free_process = NULL;
+#if HALFWORD_HEAP
+	/* Registers need to be heap allocated (correct memory range) for tracing to work */
+	esdp->save_reg = erts_alloc(ERTS_ALC_T_BEAM_REGISTER, ERTS_X_REGS_ALLOCATED * sizeof(Eterm));
+#endif
 #endif
 #if !HEAP_ON_C_STACK
 	esdp->num_tmp_heap_used = 0;
@@ -4484,7 +4488,7 @@ add_pend_suspend(Process *suspendee,
 					 sizeof(ErtsPendingSuspend));
     psp->next = NULL;
 #ifdef DEBUG
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
     psp->end = (ErtsPendingSuspend *) 0xdeaddeaddeaddead;
 #else
     psp->end = (ErtsPendingSuspend *) 0xdeaddead;
@@ -6765,8 +6769,8 @@ erl_create_process(Process* parent, /* Parent of process (default group leader).
 
     p->current = p->initial+INITIAL_MOD;
 
-    p->i = (Eterm *) beam_apply;
-    p->cp = (Eterm *) beam_apply+1;
+    p->i = (UWord *) beam_apply;
+    p->cp = (UWord *) beam_apply+1;
 
     p->arg_reg = p->def_arg_reg;
     p->max_arg_reg = sizeof(p->def_arg_reg)/sizeof(p->def_arg_reg[0]);
@@ -7351,7 +7355,7 @@ set_proc_exiting(Process *p, Eterm reason, ErlHeapFragment *bp)
     p->freason = EXTAG_EXIT;
     KILL_CATCHES(p);
     cancel_timer(p);
-    p->i = (Eterm *) beam_exit;
+    p->i = (UWord *) beam_exit;
 }
 
 
@@ -8279,7 +8283,7 @@ continue_exit_process(Process *p
 
     ASSERT(p->status == P_EXITING);
 
-    p->i = (Eterm *) beam_continue_exit;
+    p->i = (UWord *) beam_continue_exit;
 
     if (!(curr_locks & ERTS_PROC_LOCK_STATUS)) {
 	erts_smp_proc_lock(p, ERTS_PROC_LOCK_STATUS);
@@ -8299,7 +8303,7 @@ continue_exit_process(Process *p
 static void
 timeout_proc(Process* p)
 {
-    p->i = (Eterm *) p->def_arg_reg[0];
+    p->i = *((UWord **) (UWord) p->def_arg_reg);
     p->flags |= F_TIMO;
     p->flags &= ~F_INSLPQUEUE;
 
@@ -8398,9 +8402,9 @@ erts_program_counter_info(int to, void *to_arg, Process *p)
 }
 
 static void
-print_function_from_pc(int to, void *to_arg, Eterm* x)
+print_function_from_pc(int to, void *to_arg, UWord* x)
 {
-    Eterm* addr = find_function_from_pc(x);
+    UWord* addr = find_function_from_pc(x);
     if (addr == NULL) {
         if (x == beam_exit) {
             erts_print(to, to_arg, "<terminate process>");
@@ -8434,7 +8438,7 @@ stack_element_dump(int to, void *to_arg, Process* p, Eterm* sp, int yreg)
     }
 
     if (is_CP(x)) {
-        erts_print(to, to_arg, "Return addr %p (", (Eterm *) x);
+        erts_print(to, to_arg, "Return addr %p (", (Eterm *) EXPAND_POINTER(x));
         print_function_from_pc(to, to_arg, cp_val(x));
         erts_print(to, to_arg, ")\n");
         yreg = 0;
@@ -9263,8 +9267,8 @@ init_processes_bif(void)
     processes_trap_export.code[0] = am_erlang;
     processes_trap_export.code[1] = am_processes_trap;
     processes_trap_export.code[2] = 2;
-    processes_trap_export.code[3] = (Eterm) em_apply_bif;
-    processes_trap_export.code[4] = (Eterm) &processes_trap;
+    processes_trap_export.code[3] = (UWord) em_apply_bif;
+    processes_trap_export.code[4] = (UWord) &processes_trap;
 
 #if ERTS_PROCESSES_BIF_DEBUGLEVEL >= ERTS_PROCS_DBGLVL_CHK_TERM_PROC_LIST
     erts_get_emu_time(&debug_tv_start);
diff --git a/erts/emulator/beam/erl_process.h b/erts/emulator/beam/erl_process.h
index 4466442fbc..df08f2474c 100644
--- a/erts/emulator/beam/erl_process.h
+++ b/erts/emulator/beam/erl_process.h
@@ -179,7 +179,7 @@ extern int erts_sched_thread_suggested_stack_size;
 
 
 #ifdef DEBUG
-#  ifdef ARCH_64
+#  if defined(ARCH_64) && !HALFWORD_HEAP
 #    define ERTS_DBG_SET_INVALID_RUNQP(RQP, N) \
        (*((char **) &(RQP)) = (char *) (0xdeadbeefdead0003 | ((N) << 4)))
 #  define ERTS_DBG_VERIFY_VALID_RUNQP(RQP) \
@@ -344,7 +344,11 @@ struct ErtsSchedulerData_ {
      * numbered registers as possible in the same cache
      * line).
      */
+#if !HALFWORD_HEAP
     Eterm save_reg[ERTS_X_REGS_ALLOCATED]; /* X registers */
+#else
+    Eterm *save_reg;
+#endif
     FloatDef freg[MAX_REG];	/* Floating point registers. */
     ethr_tid tid;		/* Thread id */
     struct erl_bits_state erl_bits_state; /* erl_bits.c state */
@@ -525,8 +529,8 @@ struct process {
     unsigned max_arg_reg;	/* Maximum number of argument registers available. */
     Eterm def_arg_reg[6];	/* Default array for argument registers. */
 
-    Eterm* cp;			/* Continuation pointer (for threaded code). */
-    Eterm* i;			/* Program counter for threaded code. */
+    UWord* cp;			/* Continuation pointer (for threaded code). */
+    UWord* i;			/* Program counter for threaded code. */
     Sint catches;		/* Number of catches on stack */
     Sint fcalls;		/* 
 				 * Number of reductions left to execute.
@@ -573,11 +577,11 @@ struct process {
     Uint seq_trace_lastcnt;
     Eterm seq_trace_token;	/* Sequential trace token (tuple size 5 see below) */
 
-    Eterm initial[3];		/* Initial module(0), function(1), arity(2) */
-    Eterm* current;		/* Current Erlang function:
+    UWord initial[3];		/* Initial module(0), function(1), arity(2) */
+    UWord* current;		/* Current Erlang function, part of the funcinfo:
 				 * module(0), function(1), arity(2)
 				 * (module and functions are tagged atoms;
-				 * arity an untagged integer).
+				 * arity an untagged integer). UWord * because it references code
 				 */
     
     /*
@@ -1216,8 +1220,8 @@ erts_proc_get_error_handler(Process *p)
     if (!val)
 	return am_error_handler;
     else {
-	ASSERT(is_atom(((Eterm) val)));
-	return (Eterm) val;
+	ASSERT(is_atom(((Eterm) (UWord) val)));
+	return (Eterm) (UWord) val;
     }
 }
 
@@ -1227,13 +1231,13 @@ erts_proc_set_error_handler(Process *p, ErtsProcLocks plocks, Eterm handler)
     void *old_val;
     void *new_val;
     ASSERT(is_atom(handler));
-    new_val = handler == am_error_handler ? NULL : (void *) handler;
+    new_val = (handler == am_error_handler) ? NULL : (void *) (UWord) handler;
     old_val = erts_psd_set(p, plocks, ERTS_PSD_ERROR_HANDLER, new_val);
     if (!old_val)
 	return am_error_handler;
     else {
-	ASSERT(is_atom(((Eterm) old_val)));
-	return (Eterm) old_val;
+	ASSERT(is_atom(((Eterm) (UWord) old_val)));
+	return (Eterm) (UWord) old_val;
     }
 }
 
diff --git a/erts/emulator/beam/erl_process_dump.c b/erts/emulator/beam/erl_process_dump.c
index 1666509c72..0b776367ed 100644
--- a/erts/emulator/beam/erl_process_dump.c
+++ b/erts/emulator/beam/erl_process_dump.c
@@ -45,16 +45,16 @@ static void dump_dist_ext(int to, void *to_arg, ErtsDistExternal *edep);
 static void dump_element_nl(int to, void *to_arg, Eterm x);
 static int stack_element_dump(int to, void *to_arg, Process* p, Eterm* sp,
 			      int yreg);
-static void print_function_from_pc(int to, void *to_arg, Eterm* x);
+static void print_function_from_pc(int to, void *to_arg, UWord* x);
 static void heap_dump(int to, void *to_arg, Eterm x);
 static void dump_binaries(int to, void *to_arg, Binary* root);
 static void dump_externally(int to, void *to_arg, Eterm term);
 
 static Binary* all_binaries;
 
-extern Eterm beam_apply[];
-extern Eterm beam_exit[];
-extern Eterm beam_continue_exit[];
+extern UWord beam_apply[];
+extern UWord beam_exit[];
+extern UWord beam_continue_exit[];
 
 
 void
@@ -223,7 +223,7 @@ stack_element_dump(int to, void *to_arg, Process* p, Eterm* sp, int yreg)
     }
 
     if (is_CP(x)) {
-        erts_print(to, to_arg, "SReturn addr 0x%X (", (Eterm *) x);
+        erts_print(to, to_arg, "SReturn addr 0x%X (", cp_val(x));
         print_function_from_pc(to, to_arg, cp_val(x));
         erts_print(to, to_arg, ")\n");
         yreg = 0;
@@ -239,9 +239,9 @@ stack_element_dump(int to, void *to_arg, Process* p, Eterm* sp, int yreg)
 }
 
 static void
-print_function_from_pc(int to, void *to_arg, Eterm* x)
+print_function_from_pc(int to, void *to_arg, UWord* x)
 {
-    Eterm* addr = find_function_from_pc(x);
+    UWord* addr = find_function_from_pc(x);
     if (addr == NULL) {
         if (x == beam_exit) {
             erts_print(to, to_arg, "<terminate process>");
@@ -273,7 +273,7 @@ heap_dump(int to, void *to_arg, Eterm x)
     if (x == OUR_NIL) {	/* We are done. */
 	return;
     } if (is_CP(x)) {
-	next = (Eterm *) x;
+	next = (Eterm *) EXPAND_POINTER(x);
     } else if (is_list(x)) {
 	ptr = list_val(x);
 	if (ptr[0] != OUR_NIL) {
@@ -286,7 +286,7 @@ heap_dump(int to, void *to_arg, Eterm x)
 		ptr[1] = make_small(0);
 	    }
 	    x = ptr[0];
-	    ptr[0] = (Eterm) next;
+	    ptr[0] = (Eterm) COMPRESS_POINTER(next);
 	    next = ptr + 1;
 	    goto again;
 	}
@@ -316,7 +316,7 @@ heap_dump(int to, void *to_arg, Eterm x)
 		    ptr[0] = OUR_NIL;
 		} else {
 		    x = ptr[arity];
-		    ptr[0] = (Eterm) next;
+		    ptr[0] = (Eterm) COMPRESS_POINTER(next);
 		    next = ptr + arity - 1;
 		    goto again;
 		}
@@ -351,7 +351,7 @@ heap_dump(int to, void *to_arg, Eterm x)
 		    Binary* val = pb->val;
 
 		    if (erts_smp_atomic_xchg(&val->refc, 0) != 0) {
-			val->flags = (Uint) all_binaries;
+			val->flags = (UWord) all_binaries;
 			all_binaries = val;
 		    }
 		    erts_print(to, to_arg, "Yc%X:%X:%X", val,
diff --git a/erts/emulator/beam/erl_term.c b/erts/emulator/beam/erl_term.c
index c0ed3cb0a6..dd163b2b21 100644
--- a/erts/emulator/beam/erl_term.c
+++ b/erts/emulator/beam/erl_term.c
@@ -105,7 +105,7 @@ unsigned tag_val_def(Eterm x)
 	  break;
       }
     }
-    sprintf(msg, "tag_val_def: %#lx", x);
+    sprintf(msg, "tag_val_def: %#lx", (unsigned long) x);
     et_abort(msg, file, line);
 #undef file
 #undef line
@@ -165,8 +165,8 @@ ET_DEFINE_CHECKED(Uint32*,external_ref_data,Eterm,is_external_ref);
 ET_DEFINE_CHECKED(struct erl_node_*,external_ref_node,Eterm,is_external_ref);
 ET_DEFINE_CHECKED(Eterm*,export_val,Eterm,is_export);
 
-ET_DEFINE_CHECKED(Eterm,make_cp,Uint*,_is_taggable_pointer);
-ET_DEFINE_CHECKED(Uint*,cp_val,Eterm,is_CP);
+ET_DEFINE_CHECKED(Eterm,make_cp,UWord *,_is_taggable_pointer);
+ET_DEFINE_CHECKED(UWord *,cp_val,Eterm,is_CP);
 ET_DEFINE_CHECKED(Uint,catch_val,Eterm,is_catch);
 ET_DEFINE_CHECKED(Uint,x_reg_offset,Uint,_is_xreg);
 ET_DEFINE_CHECKED(Uint,y_reg_offset,Uint,_is_yreg);
diff --git a/erts/emulator/beam/erl_term.h b/erts/emulator/beam/erl_term.h
index 016d6e66a7..8e74502022 100644
--- a/erts/emulator/beam/erl_term.h
+++ b/erts/emulator/beam/erl_term.h
@@ -20,10 +20,7 @@
 #ifndef __ERL_TERM_H
 #define __ERL_TERM_H
 
-#ifdef ARCH_64
-#define HALFWORD_HEAP 1
-#define HALFWORD_ASSERT 1
-#endif
+#include "sys.h" /* defines HALFWORD_HEAP */
 
 #if HALFWORD_HEAP
 #  define HEAP_ON_C_STACK 0
@@ -35,8 +32,8 @@
 #  endif
 #  if 1
 #    define CHECK_POINTER_MASK 0xFFFFFFFF00000000UL
-#    define COMPRESS_POINTER(AnUint) (AnUint)
-#    define EXPAND_POINTER(APointer) (APointer)
+#    define COMPRESS_POINTER(APointer) ((Eterm) (UWord) (APointer))
+#    define EXPAND_POINTER(AnEterm) ((UWord) (AnEterm))
 #  else
 #    define CHECK_POINTER_MASK 0x0UL
 #    define COMPRESS_POINTER(AnUint) (AnUint)
@@ -188,14 +185,14 @@ struct erl_node_; /* Declared in erl_node_tables.h */
 
 /* boxed object access methods */
 #if HALFWORD_HEAP
-#define _is_taggable_pointer(x)	 (((Uint)(x) & (CHECK_POINTER_MASK | 0x3)) == 0)
+#define _is_taggable_pointer(x)	 (((UWord)(x) & (CHECK_POINTER_MASK | 0x3)) == 0)
 #define _boxed_precond(x)        (is_boxed(x))
 #else
 #define _is_taggable_pointer(x)	 (((Uint)(x) & 0x3) == 0)
 #define  _boxed_precond(x)       (is_boxed(x))
 #endif
 #define _is_aligned(x)		(((Uint)(x) & 0x3) == 0)
-#define _unchecked_make_boxed(x) COMPRESS_POINTER((Uint)(x) + TAG_PRIMARY_BOXED)
+#define _unchecked_make_boxed(x) ((Uint) COMPRESS_POINTER(x) + TAG_PRIMARY_BOXED)
 _ET_DECLARE_CHECKED(Eterm,make_boxed,Eterm*);
 #define make_boxed(x)		_ET_APPLY(make_boxed,(x))
 #if 1
@@ -206,12 +203,12 @@ _ET_DECLARE_CHECKED(int,is_boxed,Eterm)
 #else
 #define is_boxed(x)		(((x) & _TAG_PRIMARY_MASK) == TAG_PRIMARY_BOXED)
 #endif
-#define _unchecked_boxed_val(x) EXPAND_POINTER((Eterm*)((x) - TAG_PRIMARY_BOXED))
+#define _unchecked_boxed_val(x) ((Eterm*) EXPAND_POINTER(((x) - TAG_PRIMARY_BOXED)))
 _ET_DECLARE_CHECKED(Eterm*,boxed_val,Eterm);
 #define boxed_val(x)		_ET_APPLY(boxed_val,(x))
 
 /* cons cell ("list") access methods */
-#define _unchecked_make_list(x)	COMPRESS_POINTER((Uint)(x) + TAG_PRIMARY_LIST)
+#define _unchecked_make_list(x)	((Uint) COMPRESS_POINTER(x) + TAG_PRIMARY_LIST)
 _ET_DECLARE_CHECKED(Eterm,make_list,Eterm*);
 #define make_list(x)		_ET_APPLY(make_list,(x))
 #if 1
@@ -228,7 +225,7 @@ _ET_DECLARE_CHECKED(int,is_not_list,Eterm)
 #else
 #define _list_precond(x)       (is_list(x))
 #endif
-#define _unchecked_list_val(x)	EXPAND_POINTER((Eterm*)((x) - TAG_PRIMARY_LIST))
+#define _unchecked_list_val(x) ((Eterm*) EXPAND_POINTER((x) - TAG_PRIMARY_LIST))
 _ET_DECLARE_CHECKED(Eterm*,list_val,Eterm);
 #define list_val(x)		_ET_APPLY(list_val,(x))
 
@@ -239,13 +236,13 @@ _ET_DECLARE_CHECKED(Eterm*,list_val,Eterm);
 #define CDR(x)  ((x)[1])
 
 /* generic tagged pointer (boxed or list) access methods */
-#define _unchecked_ptr_val(x)	EXPAND_POINTER((Eterm*)((x) & ~((Uint) 0x3)))
+#define _unchecked_ptr_val(x)	((Eterm*) EXPAND_POINTER((x) & ~((Uint) 0x3)))
 #define ptr_val(x)		_unchecked_ptr_val((x))	/*XXX*/
 #define _unchecked_offset_ptr(x,offs)	((x)+((offs)*sizeof(Eterm)))
 #define offset_ptr(x,offs)	_unchecked_offset_ptr(x,offs)	/*XXX*/
 
 /* fixnum ("small") access methods */
-#if defined(ARCH_64)
+#if defined(ARCH_64) && !HALFWORD_HEAP
 #define SMALL_BITS	(64-4)
 #define SMALL_DIGITS	(17)
 #else
@@ -370,7 +367,11 @@ _ET_DECLARE_CHECKED(Eterm*,fun_val,Eterm)
 _ET_DECLARE_CHECKED(Eterm*,export_val,Eterm)
 #define export_val(x)	_ET_APPLY(export_val,(x))
 #define is_export_header(x)	((x) == HEADER_EXPORT)
+#if HALFWORD_HEAP
+#define HEADER_EXPORT   _make_header(2,_TAG_HEADER_EXPORT)
+#else
 #define HEADER_EXPORT   _make_header(1,_TAG_HEADER_EXPORT)
+#endif
 
 /* bignum access methods */
 #define make_pos_bignum_header(sz)	_make_header((sz),_TAG_HEADER_POS_BIG)
@@ -394,7 +395,7 @@ _ET_DECLARE_CHECKED(Eterm*,big_val,Eterm)
 #define big_val(x)		_ET_APPLY(big_val,(x))
 
 /* flonum ("float") access methods */
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 #define HEADER_FLONUM   _make_header(1,_TAG_HEADER_FLOAT)
 #else
 #define HEADER_FLONUM	_make_header(2,_TAG_HEADER_FLOAT)
@@ -415,12 +416,12 @@ typedef union float_def
     byte   fb[sizeof(ieee754_8)];
     Uint16 fs[sizeof(ieee754_8) / sizeof(Uint16)];
     Uint32 fw[sizeof(ieee754_8) / sizeof(Uint32)];
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
     Uint   fdw;
 #endif
 } FloatDef;
 
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 #define GET_DOUBLE(x, f) (f).fdw = *(float_val(x)+1)
 
 #define PUT_DOUBLE(f, x)  *(x) = HEADER_FLONUM, \
@@ -720,7 +721,7 @@ _ET_DECLARE_CHECKED(struct erl_node_*,internal_port_node,Eterm)
 #define ERTS_MAX_REF_NUMBERS	3
 #define ERTS_REF_NUMBERS	ERTS_MAX_REF_NUMBERS
 
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 #  define ERTS_REF_WORDS	(ERTS_REF_NUMBERS/2 + 1)
 #  define ERTS_REF_32BIT_WORDS  (ERTS_REF_NUMBERS+1)
 #else
@@ -742,7 +743,7 @@ typedef struct {
 #define make_ref_thing_header(DW) \
   _make_header((DW)+REF_THING_HEAD_SIZE-1,_TAG_HEADER_REF)
 
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 
 /*
  * Ref layout on a 64-bit little endian machine:
@@ -985,15 +986,15 @@ _ET_DECLARE_CHECKED(struct erl_node_*,external_ref_node,Eterm)
 #error "fix yer arch, like"
 #endif
 
-#define _unchecked_make_cp(x)	((Eterm)(x))
-_ET_DECLARE_CHECKED(Eterm,make_cp,Uint*)
+#define _unchecked_make_cp(x)	((Eterm) COMPRESS_POINTER(x))
+_ET_DECLARE_CHECKED(Eterm,make_cp,UWord*);
 #define make_cp(x)	_ET_APPLY(make_cp,(x))
 
 #define is_not_CP(x)	((x) & _CPMASK)
 #define is_CP(x)	(!is_not_CP(x))
 
-#define _unchecked_cp_val(x)	((Uint*)(x))
-_ET_DECLARE_CHECKED(Uint*,cp_val,Eterm)
+#define _unchecked_cp_val(x)	((UWord*) EXPAND_POINTER(x))
+_ET_DECLARE_CHECKED(UWord*,cp_val,Eterm);
 #define cp_val(x)	_ET_APPLY(cp_val,(x))
 
 #define make_catch(x)	(((x) << _TAG_IMMED2_SIZE) | _TAG_IMMED2_CATCH)
diff --git a/erts/emulator/beam/erl_trace.c b/erts/emulator/beam/erl_trace.c
index f98d5ee522..eb61c72e90 100644
--- a/erts/emulator/beam/erl_trace.c
+++ b/erts/emulator/beam/erl_trace.c
@@ -1182,7 +1182,7 @@ seq_trace_output_generic(Eterm token, Eterm msg, Uint type,
  * or   {trace, Pid, return_to, {Mod, Func, Arity}}
  */
 void 
-erts_trace_return_to(Process *p, Uint *pc)
+erts_trace_return_to(Process *p, UWord *pc)
 {
 #define LOCAL_HEAP_SIZE (4+5+5)
     Eterm* hp;
@@ -1190,7 +1190,7 @@ erts_trace_return_to(Process *p, Uint *pc)
     Eterm mess;
     DeclareTmpHeapNoproc(local_heap,LOCAL_HEAP_SIZE);
 
-    Eterm *code_ptr = find_function_from_pc(pc);
+    UWord *code_ptr = find_function_from_pc(pc);
 
 
     UseTmpHeapNoproc(LOCAL_HEAP_SIZE);
@@ -1249,7 +1249,7 @@ erts_trace_return_to(Process *p, Uint *pc)
  * or   {trace, Pid, return_from, {Mod, Name, Arity}, Retval}
  */
 void
-erts_trace_return(Process* p, Eterm* fi, Eterm retval, Eterm *tracer_pid)
+erts_trace_return(Process* p, UWord* fi, Eterm retval, Eterm *tracer_pid)
 {
     Eterm* hp;
     Eterm mfa;
@@ -1380,7 +1380,7 @@ erts_trace_return(Process* p, Eterm* fi, Eterm retval, Eterm *tracer_pid)
  * Where Class is atomic but Value is any term.
  */
 void
-erts_trace_exception(Process* p, Eterm mfa[3], Eterm class, Eterm value, 
+erts_trace_exception(Process* p, UWord mfa[3], Eterm class, Eterm value,
 		     Eterm *tracer_pid)
 {
     Eterm* hp;
@@ -1439,7 +1439,7 @@ erts_trace_exception(Process* p, Eterm mfa[3], Eterm class, Eterm value,
 	UseTmpHeapNoproc(LOCAL_HEAP_SIZE);
 
 	hp = local_heap;
-	mfa_tuple = TUPLE3(hp, mfa[0], mfa[1], make_small(mfa[2]));
+	mfa_tuple = TUPLE3(hp, (Eterm) mfa[0], (Eterm) mfa[1], make_small((Eterm)mfa[2]));
 	hp += 4;
 	cv = TUPLE2(hp, class, value);
 	hp += 3;
@@ -1485,7 +1485,7 @@ erts_trace_exception(Process* p, Eterm mfa[3], Eterm class, Eterm value,
 	 * Build the trace tuple and put it into receive queue of the tracer process.
 	 */
 	
-	mfa_tuple = TUPLE3(hp, mfa[0], mfa[1], make_small(mfa[2]));
+	mfa_tuple = TUPLE3(hp, (Eterm) mfa[0], (Eterm) mfa[1], make_small((Eterm) mfa[2]));
 	hp += 4;
 	value = copy_struct(value, value_size, &hp, off_heap);
 	cv = TUPLE2(hp, class, value);
@@ -1522,7 +1522,7 @@ erts_trace_exception(Process* p, Eterm mfa[3], Eterm class, Eterm value,
  * if it is a pid or port we do a meta trace.
  */
 Uint32
-erts_call_trace(Process* p, Eterm mfa[3], Binary *match_spec, 
+erts_call_trace(Process* p, UWord mfa[3], Binary *match_spec,
 		Eterm* args, int local, Eterm *tracer_pid)
 {
     Eterm* hp;
@@ -1589,7 +1589,7 @@ erts_call_trace(Process* p, Eterm mfa[3], Binary *match_spec,
      * such as size_object() and copy_struct(), we must make sure that we
      * temporarily convert any match contexts to sub binaries.
      */
-    arity = mfa[2];
+    arity = (Eterm) mfa[2];
     UseTmpHeap(ERL_SUB_BIN_SIZE,p);
 #ifdef DEBUG
     sub_bin_heap->thing_word = 0;
@@ -1725,7 +1725,7 @@ erts_call_trace(Process* p, Eterm mfa[3], Binary *match_spec,
 		hp += 2;
 	    }
 	}
-	mfa_tuple = TUPLE3(hp, mfa[0], mfa[1], mfa_tuple);
+	mfa_tuple = TUPLE3(hp, (Eterm) mfa[0], (Eterm) mfa[1], mfa_tuple);
 	hp += 4;
 	
 	/*
@@ -1888,7 +1888,7 @@ erts_call_trace(Process* p, Eterm mfa[3], Binary *match_spec,
 		hp += 2;
 	    }
 	}
-	mfa_tuple = TUPLE3(hp, mfa[0], mfa[1], mfa_tuple);
+	mfa_tuple = TUPLE3(hp, (Eterm) mfa[0], (Eterm) mfa[1], mfa_tuple);
 	hp += 4;
 	
 	/*
@@ -2092,7 +2092,7 @@ void save_calls(Process *p, Export *e)
  */
 Eterm
 erts_bif_trace(int bif_index, Process* p, 
-	       Eterm arg1, Eterm arg2, Eterm arg3, Uint *I)
+	       Eterm arg1, Eterm arg2, Eterm arg3, UWord *I)
 {
     Eterm result;
     int meta = !!(erts_bif_trace_flags[bif_index] & BIF_TRACE_AS_META);
@@ -2106,10 +2106,10 @@ erts_bif_trace(int bif_index, Process* p,
 	 * no tracing will occur. Doing the whole else branch will 
 	 * also do nothing, only slower.
 	 */
-	Eterm (*func)(Process*, Eterm, Eterm, Eterm, Uint*) = bif_table[bif_index].f;
+	Eterm (*func)(Process*, Eterm, Eterm, Eterm, UWord*) = bif_table[bif_index].f;
 	result = func(p, arg1, arg2, arg3, I);
     } else {
-	Eterm (*func)(Process*, Eterm, Eterm, Eterm, Uint*);
+	Eterm (*func)(Process*, Eterm, Eterm, Eterm, UWord*);
 	Export* ep = bif_export[bif_index];
 	Uint32 flags = 0, flags_meta = 0;
 	int global = !!(erts_bif_trace_flags[bif_index] & BIF_TRACE_AS_GLOBAL);
@@ -2118,16 +2118,9 @@ erts_bif_trace(int bif_index, Process* p,
 	int applying = (I == &(ep->code[3])); /* Yup, the apply code for a bif
 					       * is actually in the 
 					       * export entry */
-	Eterm *cp = p->cp;
+	UWord *cp = p->cp;
 	
-#ifndef _OSE_
 	Eterm args[3] = {arg1, arg2, arg3};
-#else
-	Eterm args[3];
-	args[0] = arg1;
-	args[1] = arg2;
-	args[2] = arg3;
-#endif
 	
 	/* 
 	 * Make continuation pointer OK, it is not during direct BIF calls,
@@ -2157,12 +2150,11 @@ erts_bif_trace(int bif_index, Process* p,
 	    Eterm *cpp;
 	    /* Maybe advance cp to skip trace stack frames */
 	    for (cpp = p->stop;  ;  cp = cp_val(*cpp++)) {
-		ASSERT(is_CP((Eterm) cp));
-		if (*cp_val((Eterm) cp) == i_return_trace) {
+		if (*cp == i_return_trace) {
 		    /* Skip stack frame variables */
 		    while (is_not_CP(*cpp)) cpp++;
 		    cpp += 2; /* Skip return_trace parameters */
-		} else if (*cp_val((Eterm) cp) == i_return_to_trace) {
+		} else if (*cp == i_return_to_trace) {
 		    /* A return_to trace message is going to be generated
 		     * by normal means, so we do not have to.
 		     */
diff --git a/erts/emulator/beam/erl_unicode.c b/erts/emulator/beam/erl_unicode.c
index ab5811c70f..8f16f7338f 100644
--- a/erts/emulator/beam/erl_unicode.c
+++ b/erts/emulator/beam/erl_unicode.c
@@ -90,9 +90,9 @@ void erts_init_unicode(void)
 	am_atom_put("characters_to_utf8_trap",23);
     characters_to_utf8_trap_exp.code[2] = 3;
     characters_to_utf8_trap_exp.code[3] =
-	(Eterm) em_apply_bif;
+	(UWord) em_apply_bif;
     characters_to_utf8_trap_exp.code[4] = 
-	(Eterm) &characters_to_utf8_trap;
+	(UWord) &characters_to_utf8_trap;
 
     memset(&characters_to_list_trap_1_exp, 0, sizeof(Export));
     characters_to_list_trap_1_exp.address = 
@@ -102,9 +102,9 @@ void erts_init_unicode(void)
 	am_atom_put("characters_to_list_trap_1",25);
     characters_to_list_trap_1_exp.code[2] = 3;
     characters_to_list_trap_1_exp.code[3] =
-	(Eterm) em_apply_bif;
+	(UWord) em_apply_bif;
     characters_to_list_trap_1_exp.code[4] = 
-	(Eterm) &characters_to_list_trap_1;
+	(UWord) &characters_to_list_trap_1;
 
     memset(&characters_to_list_trap_2_exp, 0, sizeof(Export));
     characters_to_list_trap_2_exp.address = 
@@ -114,9 +114,9 @@ void erts_init_unicode(void)
 	am_atom_put("characters_to_list_trap_2",25);
     characters_to_list_trap_2_exp.code[2] = 3;
     characters_to_list_trap_2_exp.code[3] =
-	(Eterm) em_apply_bif;
+	(UWord) em_apply_bif;
     characters_to_list_trap_2_exp.code[4] = 
-	(Eterm) &characters_to_list_trap_2;
+	(UWord) &characters_to_list_trap_2;
 
 
     memset(&characters_to_list_trap_3_exp, 0, sizeof(Export));
@@ -127,9 +127,9 @@ void erts_init_unicode(void)
 	am_atom_put("characters_to_list_trap_3",25);
     characters_to_list_trap_3_exp.code[2] = 3;
     characters_to_list_trap_3_exp.code[3] =
-	(Eterm) em_apply_bif;
+	(UWord) em_apply_bif;
     characters_to_list_trap_3_exp.code[4] = 
-	(Eterm) &characters_to_list_trap_3;
+	(UWord) &characters_to_list_trap_3;
 
     memset(&characters_to_list_trap_4_exp, 0, sizeof(Export));
     characters_to_list_trap_4_exp.address = 
@@ -139,9 +139,9 @@ void erts_init_unicode(void)
 	am_atom_put("characters_to_list_trap_4",25);
     characters_to_list_trap_4_exp.code[2] = 1;
     characters_to_list_trap_4_exp.code[3] =
-	(Eterm) em_apply_bif;
+	(UWord) em_apply_bif;
     characters_to_list_trap_4_exp.code[4] = 
-	(Eterm) &characters_to_list_trap_4;
+	(UWord) &characters_to_list_trap_4;
 
     c_to_b_int_trap_exportp =  erts_export_put(am_unicode,am_characters_to_binary_int,2);
     c_to_l_int_trap_exportp =  erts_export_put(am_unicode,am_characters_to_list_int,2);
diff --git a/erts/emulator/beam/erl_vm.h b/erts/emulator/beam/erl_vm.h
index 1b61bca440..235bd7931d 100644
--- a/erts/emulator/beam/erl_vm.h
+++ b/erts/emulator/beam/erl_vm.h
@@ -139,8 +139,12 @@
 #define HeapWordsLeft(p) (HEAP_LIMIT(p) - HEAP_TOP(p))
 
 #if defined(DEBUG) || defined(CHECK_FOR_HOLES)
+#if HALFWORD_HEAP
+# define ERTS_HOLE_MARKER (0xaf5e78ccU)
+#else
 # define ERTS_HOLE_MARKER (((0xaf5e78ccUL << 24) << 8) | 0xaf5e78ccUL)
 #endif
+#endif
 
 /*
  * Allocate heap memory on the ordinary heap, NEVER in a heap
diff --git a/erts/emulator/beam/error.h b/erts/emulator/beam/error.h
index 4930def4ed..8f95140864 100644
--- a/erts/emulator/beam/error.h
+++ b/erts/emulator/beam/error.h
@@ -187,10 +187,10 @@ extern Eterm exception_tag[NUMBER_EXC_TAGS];
 struct StackTrace {
     Eterm header;	/* bignum header - must be first in struct */
     Eterm freason; /* original exception reason is saved in the struct */
-    Eterm* pc;
-    Eterm* current;
+    UWord* pc;
+    UWord* current;
     int depth;	/* number of saved pointers in trace[] */
-    Eterm *trace[1];  /* varying size - must be last in struct */
+    UWord *trace[1];  /* varying size - must be last in struct */
 };
 
 #endif /* __ERROR_H__ */
diff --git a/erts/emulator/beam/export.c b/erts/emulator/beam/export.c
index 271b40cf0f..c39f610924 100644
--- a/erts/emulator/beam/export.c
+++ b/erts/emulator/beam/export.c
@@ -46,8 +46,8 @@ static erts_smp_rwmtx_t export_table_lock; /* Locks the secondary export table.
 #define export_init_lock()	erts_smp_rwmtx_init(&export_table_lock, \
 						    "export_tab")
 
-extern Eterm* em_call_error_handler;
-extern Uint* em_call_traced_function;
+extern UWord* em_call_error_handler;
+extern UWord* em_call_traced_function;
 
 void
 export_info(int to, void *to_arg)
@@ -93,7 +93,7 @@ export_alloc(Export* tmpl)
     obj->code[2] = tmpl->code[2];
     obj->slot.index = -1;
     obj->address = obj->code+3;
-    obj->code[3] = (Eterm) em_call_error_handler;
+    obj->code[3] = (UWord) em_call_error_handler;
     obj->code[4] = 0;
     obj->match_prog_set = NULL;
     return obj;
@@ -140,7 +140,7 @@ init_export_table(void)
 Export*
 erts_find_export_entry(Eterm m, Eterm f, unsigned int a)
 {
-    HashValue hval = EXPORT_HASH(m, f, a);
+    HashValue hval = EXPORT_HASH((UWord) m, (UWord) f, (UWord) a);
     int ix;
     HashBucket* b;
 
@@ -185,7 +185,7 @@ erts_find_function(Eterm m, Eterm f, unsigned int a)
 
     ep = hash_get(&export_table.htable, (void*) &e);
     if (ep != NULL && ep->address == ep->code+3 &&
-	ep->code[3] != (Uint) em_call_traced_function) {
+	ep->code[3] != (UWord) em_call_traced_function) {
 	ep = NULL;
     }
     return ep;
diff --git a/erts/emulator/beam/export.h b/erts/emulator/beam/export.h
index cd6af6dd85..57b6a7afcf 100644
--- a/erts/emulator/beam/export.h
+++ b/erts/emulator/beam/export.h
@@ -52,7 +52,11 @@ typedef struct export
      *		on_load function that has not been run yet.
      *		Otherwise: 0.
      */
+#if HALFWORD_HEAP
+    UWord code[5];
+#else
     Eterm code[5];
+#endif
 } Export;
 
 
@@ -74,6 +78,6 @@ Export *export_get(Export*);
 #include "beam_load.h" /* For em_* extern declarations */ 
 #define ExportIsBuiltIn(EntryPtr) 			\
 (((EntryPtr)->address == (EntryPtr)->code + 3) && 	\
- ((EntryPtr)->code[3] == (Uint) em_apply_bif))
+ ((EntryPtr)->code[3] == (UWord) em_apply_bif))
 
 #endif
diff --git a/erts/emulator/beam/external.c b/erts/emulator/beam/external.c
index f7b498e414..d6cb6efb79 100644
--- a/erts/emulator/beam/external.c
+++ b/erts/emulator/beam/external.c
@@ -1489,20 +1489,28 @@ dec_pid(ErtsDistExternal *edep, Eterm** hpp, byte* ep, ErlOffHeap* off_heap, Ete
 static byte*
 enc_term(ErtsAtomCacheMap *acmp, Eterm obj, byte* ep, Uint32 dflags)
 {
-    DECLARE_ESTACK(s);
+    DECLARE_WSTACK(s);
     Uint n;
     Uint i;
     Uint j;
     Uint* ptr;
     Eterm val;
     FloatDef f;
+#if HALFWORD_HEAP
+    UWord wobj;
+#endif
+
 
     goto L_jump_start;
 
  outer_loop:
-    while (!ESTACK_ISEMPTY(s)) {
-	obj = ESTACK_POP(s);
-	switch (val = ESTACK_POP(s)) {
+    while (!WSTACK_ISEMPTY(s)) {
+#if HALFWORD_HEAP
+	obj = (Eterm) (wobj = WSTACK_POP(s));
+#else
+	obj = WSTACK_POP(s);
+#endif
+	switch (val = WSTACK_POP(s)) {
 	case ENC_TERM:
 	    break;
 	case ENC_ONE_CONS:
@@ -1513,29 +1521,40 @@ enc_term(ErtsAtomCacheMap *acmp, Eterm obj, byte* ep, Uint32 dflags)
 
 		obj = CAR(cons);
 		tl = CDR(cons);
-		ESTACK_PUSH(s, is_list(tl) ? ENC_ONE_CONS : ENC_TERM);
-		ESTACK_PUSH(s, tl);
+		WSTACK_PUSH(s, is_list(tl) ? ENC_ONE_CONS : ENC_TERM);
+		WSTACK_PUSH(s, tl);
 	    }
 	    break;
 	case ENC_PATCH_FUN_SIZE:
 	    {
+#if HALFWORD_HEAP
+		byte* size_p = (byte *) wobj;
+#else
 		byte* size_p = (byte *) obj;
-
+#endif
 		put_int32(ep - size_p, size_p);
 	    }
 	    goto outer_loop;
 	case ENC_LAST_ARRAY_ELEMENT:
 	    {
+#if HALFWORD_HEAP
+		Eterm* ptr = (Eterm *) wobj;
+#else
 		Eterm* ptr = (Eterm *) obj;
+#endif
 		obj = *ptr;
 	    }
 	    break;
 	default:		/* ENC_LAST_ARRAY_ELEMENT+1 and upwards */
 	    {
+#if HALFWORD_HEAP
+		Eterm* ptr = (Eterm *) wobj;
+#else
 		Eterm* ptr = (Eterm *) obj;
+#endif
 		obj = *ptr++;
-		ESTACK_PUSH(s, val-1);
-		ESTACK_PUSH(s, (Eterm) ptr);
+		WSTACK_PUSH(s, val-1);
+		WSTACK_PUSH(s, (UWord) ptr);
 	    }
 	    break;
 	}
@@ -1665,8 +1684,8 @@ enc_term(ErtsAtomCacheMap *acmp, Eterm obj, byte* ep, Uint32 dflags)
 		ep += 4;
 	    }
 	    if (i > 0) {
-		ESTACK_PUSH(s, ENC_LAST_ARRAY_ELEMENT+i-1);
-		ESTACK_PUSH(s, (Eterm) ptr);
+		WSTACK_PUSH(s, ENC_LAST_ARRAY_ELEMENT+i-1);
+		WSTACK_PUSH(s, (UWord) ptr);
 	    }
 	    break;
 
@@ -1744,7 +1763,7 @@ enc_term(ErtsAtomCacheMap *acmp, Eterm obj, byte* ep, Uint32 dflags)
 	    }
 	case EXPORT_DEF:
 	    {
-		Export* exp = (Export *) (export_val(obj))[1];
+		Export* exp = *((Export **) (export_val(obj) + 1));
 		if ((dflags & DFLAG_EXPORT_PTR_TAG) != 0) {
 		    *ep++ = EXPORT_EXT;
 		    ep = enc_atom(acmp, exp->code[0], ep, dflags);
@@ -1773,8 +1792,8 @@ enc_term(ErtsAtomCacheMap *acmp, Eterm obj, byte* ep, Uint32 dflags)
 		    int ei;
 
 		    *ep++ = NEW_FUN_EXT;
-		    ESTACK_PUSH(s, ENC_PATCH_FUN_SIZE);
-		    ESTACK_PUSH(s, (Eterm) ep); /* Position for patching in size */
+		    WSTACK_PUSH(s, ENC_PATCH_FUN_SIZE);
+		    WSTACK_PUSH(s, (UWord) ep); /* Position for patching in size */
 		    ep += 4;
 		    *ep = funp->arity;
 		    ep += 1;
@@ -1791,8 +1810,8 @@ enc_term(ErtsAtomCacheMap *acmp, Eterm obj, byte* ep, Uint32 dflags)
 
 		fun_env:
 		    for (ei = funp->num_free-1; ei > 0; ei--) {
-			ESTACK_PUSH(s, ENC_TERM);
-			ESTACK_PUSH(s, funp->env[ei]);
+			WSTACK_PUSH(s, ENC_TERM);
+			WSTACK_PUSH(s, (UWord) funp->env[ei]);
 		    }
 		    if (funp->num_free != 0) {
 			obj = funp->env[0];
@@ -1835,7 +1854,7 @@ enc_term(ErtsAtomCacheMap *acmp, Eterm obj, byte* ep, Uint32 dflags)
 	    break;
 	}
     }
-    DESTROY_ESTACK(s);
+    DESTROY_WSTACK(s);
     return ep;
 }
 
@@ -1959,7 +1978,7 @@ dec_term(ErtsDistExternal *edep, Eterm** hpp, byte* ep, ErlOffHeap* off_heap, Et
 
     while (next != NULL) {
 	objp = next;
-	next = (Eterm *) (*objp);
+	next = (Eterm *) EXPAND_POINTER(*objp);
 
 	switch (*ep++) {
 	case INTEGER_EXT:
@@ -1967,7 +1986,7 @@ dec_term(ErtsDistExternal *edep, Eterm** hpp, byte* ep, ErlOffHeap* off_heap, Et
 		Sint sn = get_int32(ep);
 
 		ep += 4;
-#if defined(ARCH_64)
+#if defined(ARCH_64) && !HALFWORD_HEAP
 		*objp = make_small(sn);
 #else
 		if (MY_IS_SSMALL(sn)) {
@@ -2064,7 +2083,7 @@ dec_term_atom_common:
 	    hp += n;
 	    objp = hp - 1;
 	    while (n-- > 0) {
-		objp[0] = (Eterm) next;
+		objp[0] = (Eterm) COMPRESS_POINTER(next);
 		next = objp;
 		objp--;
 	    }
@@ -2082,12 +2101,12 @@ dec_term_atom_common:
 	    *objp = make_list(hp);
 	    hp += 2*n;
 	    objp = hp - 2;
-	    objp[0] = (Eterm) (objp+1);
-	    objp[1] = (Eterm) next;
+	    objp[0] = (Eterm) COMPRESS_POINTER((objp+1));
+	    objp[1] = (Eterm) COMPRESS_POINTER(next);
 	    next = objp;
 	    objp -= 2;
 	    while (--n > 0) {
-		objp[0] = (Eterm) next;
+		objp[0] = (Eterm) COMPRESS_POINTER(next);
 		objp[1] = make_list(objp + 2);
 		next = objp;
 		objp -= 2;
@@ -2242,7 +2261,7 @@ dec_term_atom_common:
 		if(node == erts_this_node) {
 		    RefThing *rtp = (RefThing *) hp;
 		    hp += REF_THING_HEAD_SIZE;
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 		    rtp->header = make_ref_thing_header(ref_words/2 + 1);
 #else
 		    rtp->header = make_ref_thing_header(ref_words);
@@ -2253,7 +2272,7 @@ dec_term_atom_common:
 		    ExternalThing *etp = (ExternalThing *) hp;
 		    hp += EXTERNAL_THING_HEAD_SIZE;
 		    
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 		    etp->header = make_external_ref_header(ref_words/2 + 1);
 #else
 		    etp->header = make_external_ref_header(ref_words);
@@ -2266,7 +2285,7 @@ dec_term_atom_common:
 		}
 
 		ref_num = (Uint32 *) hp;
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 		*(ref_num++) = ref_words /* 32-bit arity */;
 #endif
 		ref_num[0] = r0;
@@ -2274,7 +2293,7 @@ dec_term_atom_common:
 		    ref_num[i] = get_int32(ep);
 		    ep += 4;
 		}
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 		if ((1 + ref_words) % 2)
 		    ref_num[ref_words] = 0;
 		hp += ref_words/2 + 1;
@@ -2401,7 +2420,12 @@ dec_term_atom_common:
                 }
 		*objp = make_export(hp);
 		*hp++ = HEADER_EXPORT;
+#if HALFWORD_HEAP
+		*((UWord *) (UWord) hp) =  (UWord) erts_export_get_or_make_stub(mod, name, arity);
+		hp += 2;
+#else
 		*hp++ = (Eterm) erts_export_get_or_make_stub(mod, name, arity);
+#endif
 		break;
 	    }
 	    break;
@@ -2477,11 +2501,11 @@ dec_term_atom_common:
 
 		/* Environment */
 		for (i = num_free-1; i >= 0; i--) {
-		    funp->env[i] = (Eterm) next;
+		    funp->env[i] = (Eterm) COMPRESS_POINTER(next);
 		    next = funp->env + i;
 		}
 		/* Creator */
-		funp->creator = (Eterm) next;
+		funp->creator = (Eterm) COMPRESS_POINTER(next);
 		next = &(funp->creator);
 		break;
 	    }
@@ -2552,7 +2576,7 @@ dec_term_atom_common:
 
 		/* Environment */
 		for (i = num_free-1; i >= 0; i--) {
-		    funp->env[i] = (Eterm) next;
+		    funp->env[i] = (Eterm) COMPRESS_POINTER(next);
 		    next = funp->env + i;
 		}
 		break;
@@ -2583,26 +2607,35 @@ dec_term_atom_common:
 static Uint
 encode_size_struct2(ErtsAtomCacheMap *acmp, Eterm obj, unsigned dflags)
 {
-    DECLARE_ESTACK(s);
+    DECLARE_WSTACK(s);
     Uint m, i, arity;
     Uint result = 0;
+#if HALFWORD_HEAP
+    UWord wobj;
+#endif
 
     goto L_jump_start;
 
  outer_loop:
-    while (!ESTACK_ISEMPTY(s)) {
-	obj = ESTACK_POP(s);
-
+    while (!WSTACK_ISEMPTY(s)) {
+#if HALFWORD_HEAP
+	obj = (Eterm) (wobj = WSTACK_POP(s));
+#else
+	obj = WSTACK_POP(s);
+#endif
     handle_popped_obj:
-	if (is_CP(obj)) {
+	if (is_CP(obj)) { /* Does not look for CP, looks for "no tag" */
+#if HALFWORD_HEAP
+	    Eterm* ptr = (Eterm *) wobj;
+#else
 	    Eterm* ptr = (Eterm *) obj;
-
+#endif
 	    /*
 	     * Pointer into a tuple.
 	     */
 	    obj = *ptr--;
 	    if (!is_header(obj)) {
-		ESTACK_PUSH(s, (Eterm)ptr);
+		WSTACK_PUSH(s, (UWord)ptr);
 	    } else {
 		/* Reached tuple header */
 		ASSERT(header_is_arityval(obj));
@@ -2614,7 +2647,7 @@ encode_size_struct2(ErtsAtomCacheMap *acmp, Eterm obj, unsigned dflags)
 
 	    tl = CDR(cons);
 	    obj = CAR(cons);
-	    ESTACK_PUSH(s, tl);
+	    WSTACK_PUSH(s, tl);
 	} else if (is_nil(obj)) {
 	    result++;
 	    goto outer_loop;
@@ -2703,7 +2736,11 @@ encode_size_struct2(ErtsAtomCacheMap *acmp, Eterm obj, unsigned dflags)
 		    result += 1 + 4;
 		}
 		ptr += arity;
+#if HALFWORD_HEAP
+		obj = (Eterm) (wobj = (UWord) ptr);
+#else
 		obj = (Eterm) ptr;
+#endif
 		goto handle_popped_obj;
 	    }
 	    break;
@@ -2745,14 +2782,14 @@ encode_size_struct2(ErtsAtomCacheMap *acmp, Eterm obj, unsigned dflags)
 
 		    if (is_not_list(obj)) {
 			/* Push any non-list terms on the stack */
-			ESTACK_PUSH(s, obj);
+			WSTACK_PUSH(s, obj);
 		    } else {
 			/* Lists must be handled specially. */
 			if ((m = is_string(obj)) && (m < MAX_STRING_LEN)) {
 			    result += m + 2 + 1;
 			} else {
 			    result += 5;
-			    ESTACK_PUSH(s, obj);
+			    WSTACK_PUSH(s, obj);
 			}
 		    }
 		}
@@ -2765,8 +2802,12 @@ encode_size_struct2(ErtsAtomCacheMap *acmp, Eterm obj, unsigned dflags)
 
 	case EXPORT_DEF:
 	    {
-		Export* ep = (Export *) (export_val(obj))[1];
+		Export* ep = *((Export **) (export_val(obj) + 1));
+#if HALFWORD_HEAP
+		result += 2;
+#else
 		result += 1;
+#endif
 		result += encode_size_struct2(acmp, ep->code[0], dflags);
 		result += encode_size_struct2(acmp, ep->code[1], dflags);
 		result += encode_size_struct2(acmp, make_small(ep->code[2]), dflags);
@@ -2779,7 +2820,7 @@ encode_size_struct2(ErtsAtomCacheMap *acmp, Eterm obj, unsigned dflags)
 	}
     }
 
-    DESTROY_ESTACK(s);
+    DESTROY_WSTACK(s);
     return result;
 }
 
@@ -2891,7 +2932,7 @@ decoded_size(byte *ep, byte* endp, int no_refc_bins)
 		ep += 2;
 		atom_extra_skip = 1 + 4*id_words;
 		/* In case it is an external ref */
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 		heap_size += EXTERNAL_THING_HEAD_SIZE + id_words/2 + 1;
 #else
 		heap_size += EXTERNAL_THING_HEAD_SIZE + id_words;
diff --git a/erts/emulator/beam/external.h b/erts/emulator/beam/external.h
index eada6d4f27..db86b4d796 100644
--- a/erts/emulator/beam/external.h
+++ b/erts/emulator/beam/external.h
@@ -212,7 +212,7 @@ erts_dist_ext_trailer(ErtsDistExternal *edep)
 {
     void *res = (void *) (edep->ext_endp
 			  + ERTS_WORD_ALIGN_PAD_SZ(edep->ext_endp));
-    ASSERT((((Uint) res) % sizeof(Uint)) == 0);
+    ASSERT((((UWord) res) % sizeof(Uint)) == 0);
     return res;
 }
 
diff --git a/erts/emulator/beam/global.h b/erts/emulator/beam/global.h
index 03e97a0ce4..2afb6775b5 100644
--- a/erts/emulator/beam/global.h
+++ b/erts/emulator/beam/global.h
@@ -403,7 +403,7 @@ extern Eterm erts_ddll_monitor_driver(Process *p,
 
 /* Add fields in ERTS_INTERNAL_BINARY_FIELDS, otherwise the drivers crash */
 #define ERTS_INTERNAL_BINARY_FIELDS				\
-    Uint flags;							\
+    UWord flags;							\
     erts_refc_t refc;						\
     ERTS_BINARY_STRUCT_ALIGNMENT
 
@@ -723,6 +723,60 @@ do {									\
 #define ESTACK_POP(s) (*(--ESTK_CONCAT(s,_sp)))
 
 
+void erl_grow_wstack(UWord** start, UWord** sp, UWord** end);
+#define WSTK_CONCAT(a,b) a##b
+#define WSTK_SUBSCRIPT(s,i) *((UWord *)((byte *)WSTK_CONCAT(s,_start) + (i)))
+#define DEF_WSTACK_SIZE (16)
+
+#define DECLARE_WSTACK(s)						\
+    UWord WSTK_CONCAT(s,_default_stack)[DEF_WSTACK_SIZE];		\
+    UWord* WSTK_CONCAT(s,_start) = WSTK_CONCAT(s,_default_stack);	\
+    UWord* WSTK_CONCAT(s,_sp) = WSTK_CONCAT(s,_start);			\
+    UWord* WSTK_CONCAT(s,_end) = WSTK_CONCAT(s,_start) + DEF_WSTACK_SIZE
+
+#define DESTROY_WSTACK(s)						\
+do {									\
+    if (WSTK_CONCAT(s,_start) != WSTK_CONCAT(s,_default_stack)) {	\
+	erts_free(ERTS_ALC_T_ESTACK, WSTK_CONCAT(s,_start));		\
+    }									\
+} while(0)
+
+#define WSTACK_PUSH(s, x)						\
+do {									\
+    if (WSTK_CONCAT(s,_sp) == WSTK_CONCAT(s,_end)) {			\
+	erl_grow_wstack(&WSTK_CONCAT(s,_start), &WSTK_CONCAT(s,_sp),	\
+	               &WSTK_CONCAT(s,_end));				\
+    }									\
+    *WSTK_CONCAT(s,_sp)++ = (x);					\
+} while(0)
+
+#define WSTACK_PUSH2(s, x, y)						\
+do {									\
+    if (WSTK_CONCAT(s,_sp) > WSTK_CONCAT(s,_end) - 2) {			\
+	erl_grow_wstack(&WSTK_CONCAT(s,_start), &WSTK_CONCAT(s,_sp),	\
+		&WSTK_CONCAT(s,_end));	\
+    }									\
+    *WSTK_CONCAT(s,_sp)++ = (x);					\
+    *WSTK_CONCAT(s,_sp)++ = (y);					\
+} while(0)
+
+#define WSTACK_PUSH3(s, x, y, z)					\
+do {									\
+    if (WSTK_CONCAT(s,_sp) > WSTK_CONCAT(s,_end) - 3) {			\
+	erl_grow_wstack(&WSTK_CONCAT(s,_start), &WSTK_CONCAT(s,_sp),	\
+		&WSTK_CONCAT(s,_end));					\
+    }									\
+    *WSTK_CONCAT(s,_sp)++ = (x);					\
+    *WSTK_CONCAT(s,_sp)++ = (y);					\
+    *WSTK_CONCAT(s,_sp)++ = (z);					\
+} while(0)
+
+#define WSTACK_COUNT(s) (WSTK_CONCAT(s,_sp) - WSTK_CONCAT(s,_start))
+
+#define WSTACK_ISEMPTY(s) (WSTK_CONCAT(s,_sp) == WSTK_CONCAT(s,_start))
+#define WSTACK_POP(s) (*(--WSTK_CONCAT(s,_sp)))
+
+
 /* port status flags */
 
 #define ERTS_PORT_SFLG_CONNECTED	((Uint32) (1 <<  0))
@@ -802,7 +856,7 @@ void erts_system_profile_clear(Process *c_p);
 int erts_load_module(Process *c_p, ErtsProcLocks c_p_locks,
 		     Eterm group_leader, Eterm* mod, byte* code, int size);
 void init_load(void);
-Eterm* find_function_from_pc(Eterm* pc);
+UWord* find_function_from_pc(UWord* pc);
 Eterm erts_module_info_0(Process* p, Eterm module);
 Eterm erts_module_info_1(Process* p, Eterm module, Eterm what);
 Eterm erts_make_stub_module(Process* p, Eterm Mod, Eterm Beam, Eterm Info);
@@ -1171,7 +1225,7 @@ erts_smp_port_unlock(Port *prt)
   ERTS_INVALID_PORT_OPT((PP), (ID), ERTS_PORT_SFLGS_INVALID_TRACER_LOOKUP)
 
 #define ERTS_PORT_SCHED_ID(P, ID) \
-  ((Uint) erts_prtsd_set((P), ERTS_PSD_SCHED_ID, (void *) (ID)))
+  ((Uint) (UWord) erts_prtsd_set((P), ERTS_PSD_SCHED_ID, (void *) (UWord) (ID)))
 
 #ifdef ERTS_SMP
 Port *erts_de2port(DistEntry *, Process *, ErtsProcLocks);
@@ -1463,6 +1517,7 @@ Uint32 make_hash(Eterm);
 
 Eterm erts_bld_atom(Uint **hpp, Uint *szp, char *str);
 Eterm erts_bld_uint(Uint **hpp, Uint *szp, Uint ui);
+Eterm erts_bld_uword(Uint **hpp, Uint *szp, UWord uw);
 Eterm erts_bld_uint64(Uint **hpp, Uint *szp, Uint64 ui64);
 Eterm erts_bld_sint64(Uint **hpp, Uint *szp, Sint64 si64);
 Eterm erts_bld_cons(Uint **hpp, Uint *szp, Eterm car, Eterm cdr);
@@ -1554,12 +1609,12 @@ void erts_queue_error_logger_message(Eterm, Eterm, ErlHeapFragment *);
 void erts_send_sys_msg_proc(Eterm, Eterm, Eterm, ErlHeapFragment *);
 void trace_send(Process*, Eterm, Eterm);
 void trace_receive(Process*, Eterm);
-Uint32 erts_call_trace(Process *p, Eterm mfa[], Binary *match_spec, Eterm* args,
+Uint32 erts_call_trace(Process *p, UWord mfa[], Binary *match_spec, Eterm* args,
 		       int local, Eterm *tracer_pid);
-void erts_trace_return(Process* p, Eterm* fi, Eterm retval, Eterm *tracer_pid);
-void erts_trace_exception(Process* p, Eterm mfa[], Eterm class, Eterm value, 
+void erts_trace_return(Process* p, UWord* fi, Eterm retval, Eterm *tracer_pid);
+void erts_trace_exception(Process* p, UWord mfa[], Eterm class, Eterm value,
 			  Eterm *tracer);
-void erts_trace_return_to(Process *p, Uint *pc);
+void erts_trace_return_to(Process *p, UWord *pc);
 void trace_sched(Process*, Eterm);
 void trace_proc(Process*, Process*, Eterm, Eterm);
 void trace_proc_spawn(Process*, Eterm pid, Eterm mod, Eterm func, Eterm args);
@@ -1589,7 +1644,7 @@ Uint erts_trace_flag2bit(Eterm flag);
 int erts_trace_flags(Eterm List, 
 		 Uint *pMask, Eterm *pTracer, int *pCpuTimestamp);
 Eterm erts_bif_trace(int bif_index, Process* p, 
-		     Eterm arg1, Eterm arg2, Eterm arg3, Uint *I);
+		     Eterm arg1, Eterm arg2, Eterm arg3, UWord *I);
 
 #ifdef ERTS_SMP
 void erts_send_pending_trace_msgs(ErtsSchedulerData *esdp);
@@ -1606,7 +1661,7 @@ void bin_write(int, void*, byte*, int);
 int intlist_to_buf(Eterm, char*, int); /* most callers pass plain char*'s */
 
 struct Sint_buf {
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
     char s[22];
 #else
     char s[12];
diff --git a/erts/emulator/beam/io.c b/erts/emulator/beam/io.c
index ac25eb9d5c..f3062c894c 100644
--- a/erts/emulator/beam/io.c
+++ b/erts/emulator/beam/io.c
@@ -3994,7 +3994,7 @@ drv_cancel_timer(Port *prt)
 	erts_port_task_abort(prt->id, &prt->timeout_task);
 }
 
-int driver_set_timer(ErlDrvPort ix, Uint t)
+int driver_set_timer(ErlDrvPort ix, UWord t)
 {
     Port* prt = erts_drvport2port(ix);
 
@@ -4053,12 +4053,16 @@ driver_read_timer(ErlDrvPort ix, unsigned long* t)
 int 
 driver_get_now(ErlDrvNowData *now_data)
 {
+    Uint mega,secs,micro
     ERTS_SMP_CHK_NO_PROC_LOCKS;
 
     if (now_data == NULL) {
 	return -1;
     }
-    get_now(&(now_data->megasecs),&(now_data->secs),&(now_data->microsecs));
+    get_now(&mega,&secs,&micro);
+    now_data->megasecs = (unsigned long) mega;
+    now_data->secs = (unsigned long) secs;
+    now_data->microsecs = (unsigned long) micro;
     return 0;
 }
 
diff --git a/erts/emulator/beam/module.h b/erts/emulator/beam/module.h
index 87d13b3607..a09d5985db 100644
--- a/erts/emulator/beam/module.h
+++ b/erts/emulator/beam/module.h
@@ -29,8 +29,8 @@ typedef struct erl_module {
     IndexSlot slot;		/* Must be located at top of struct! */
     int module;			/* Atom index for module (not tagged). */
 
-    Eterm* code;
-    Eterm* old_code;
+    UWord* code;
+    UWord* old_code;
     int code_length;		/* Length of loaded code in bytes. */
     int old_code_length;	/* Length of old loaded code in bytes */
     unsigned catches, old_catches;
diff --git a/erts/emulator/beam/sys.h b/erts/emulator/beam/sys.h
index 4b949523fa..b48b147953 100644
--- a/erts/emulator/beam/sys.h
+++ b/erts/emulator/beam/sys.h
@@ -20,6 +20,7 @@
 #ifndef __SYS_H__
 #define __SYS_H__
 
+
 #if defined(VALGRIND) && !defined(NO_FPE_SIGNALS)
 #  define NO_FPE_SIGNALS
 #endif
@@ -233,6 +234,10 @@ EXTERN_FUNCTION(int, real_printf, (const char *fmt, ...));
 ** UInt:  An unsigned integer exactly as large as an Eterm.
 ** SInt:  A signed integer exactly as large as an eterm and therefor large
 **        enough to hold the return value of the signed_val() macro.
+** UWord: An unsigned integer at least as large as a void * and also as large
+**          or larger than an Eterm
+** SWord: A signed integer at least as large as a void * and also as large
+**          or larger than an Eterm
 ** Uint32: An unsigned integer of 32 bits exactly
 ** Sint32: A signed integer of 32 bits exactly
 ** Uint16: An unsigned integer of 16 bits exactly
@@ -253,11 +258,43 @@ EXTERN_FUNCTION(int, real_printf, (const char *fmt, ...));
 #else
 #error Neither 32 nor 64 bit architecture
 #endif
+#ifdef ARCH_64
+#  ifdef HALFWORD_HEAP_EMULATOR
+#    define HALFWORD_HEAP 1
+#    define HALFWORD_ASSERT 1
+#  else
+#    define HALFWORD_HEAP 0
+#    define HALFWORD_ASSERT 1
+#  endif
+#endif
 
 #if SIZEOF_VOID_P != SIZEOF_SIZE_T
 #error sizeof(void*) != sizeof(size_t)
 #endif
 
+#if HALFWORD_HEAP
+
+#if SIZEOF_INT == 4
+typedef unsigned int Eterm;
+typedef unsigned int Uint;
+typedef int          Sint;
+#define ERTS_SIZEOF_ETERM SIZEOF_INT
+#else
+#error Found no appropriate type to use for 'Eterm', 'Uint' and 'Sint'
+#endif
+
+#if SIZEOF_VOID_P == SIZEOF_LONG
+typedef unsigned long UWord;
+typedef long          SWord;
+#elif SIZEOF_VOID_P == SIZEOF_INT
+typedef unsigned int UWord;
+typedef int          SWord;
+#else
+#error Found no appropriate type to use for 'Eterm', 'Uint' and 'Sint'
+#endif
+
+#else /* !HALFWORD_HEAP */
+
 #if SIZEOF_VOID_P == SIZEOF_LONG
 typedef unsigned long Eterm;
 typedef unsigned long Uint;
@@ -272,6 +309,11 @@ typedef int          Sint;
 #error Found no appropriate type to use for 'Eterm', 'Uint' and 'Sint'
 #endif
 
+typedef Uint UWord;
+typedef Sint SWord;
+
+#endif /* HALFWORD_HEAP */
+
 #ifndef HAVE_INT64
 #if SIZEOF_LONG == 8
 #define HAVE_INT64 1
diff --git a/erts/emulator/beam/utils.c b/erts/emulator/beam/utils.c
index f59fbbee33..51c12a0b69 100644
--- a/erts/emulator/beam/utils.c
+++ b/erts/emulator/beam/utils.c
@@ -204,6 +204,25 @@ erl_grow_stack(Eterm** start, Eterm** sp, Eterm** end)
     *end = *start + new_size;
     *sp = *start + sp_offs;
 }
+/*
+ * Helper function for the ESTACK macros defined in global.h.
+ */
+void
+erl_grow_wstack(UWord** start, UWord** sp, UWord** end)
+{
+    Uint old_size = (*end - *start);
+    Uint new_size = old_size * 2;
+    Uint sp_offs = *sp - *start;
+    if (new_size > 2 * DEF_ESTACK_SIZE) {
+	*start = erts_realloc(ERTS_ALC_T_ESTACK, (void *) *start, new_size*sizeof(UWord));
+    } else {
+	UWord* new_ptr = erts_alloc(ERTS_ALC_T_ESTACK, new_size*sizeof(UWord));
+	sys_memcpy(new_ptr, *start, old_size*sizeof(UWord));
+	*start = new_ptr;
+    }
+    *end = *start + new_size;
+    *sp = *start + sp_offs;
+}
 
 /* CTYPE macros */
 
@@ -354,6 +373,31 @@ erts_bld_uint(Uint **hpp, Uint *szp, Uint ui)
     return res;
 }
 
+/*
+ * Erts_bld_uword is more or less similar to erts_bld_uint, but a pointer
+ * can safely be passed.
+ */
+
+Eterm
+erts_bld_uword(Uint **hpp, Uint *szp, UWord uw)
+{
+    Eterm res = THE_NON_VALUE;
+    if (IS_USMALL(0, uw)) {
+	if (hpp)
+	    res = make_small((Uint) uw);
+    }
+    else {
+	if (szp)
+	    *szp += BIG_UWORD_HEAP_SIZE(uw);
+	if (hpp) {
+	    res = uword_to_big(uw, *hpp);
+	    *hpp += BIG_UWORD_HEAP_SIZE(uw);
+	}
+    }
+    return res;
+}
+
+
 Eterm
 erts_bld_uint64(Uint **hpp, Uint *szp, Uint64 ui64)
 {
@@ -711,7 +755,7 @@ hash_binary_bytes(Eterm bin, Uint sz, Uint32 hash)
 
 Uint32 make_hash(Eterm term_arg)
 {
-    DECLARE_ESTACK(stack);
+    DECLARE_WSTACK(stack);
     Eterm term = term_arg;
     Eterm hash = 0;
     unsigned op;
@@ -770,7 +814,7 @@ tail_recur:
 	    Uint y2 = y1 < 0 ? -(Uint)y1 : y1;
 
 	    UINT32_HASH_STEP(y2, FUNNY_NUMBER2);
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 	    if (y2 >> 32)
 		UINT32_HASH_STEP(y2 >> 32, FUNNY_NUMBER2);
 #endif
@@ -787,7 +831,7 @@ tail_recur:
 	}
     case EXPORT_DEF:
 	{
-	    Export* ep = (Export *) (export_val(term))[1];
+	    Export* ep = *((Export **) (export_val(term) + 1));
 
 	    hash = hash * FUNNY_NUMBER11 + ep->code[2];
 	    hash = hash*FUNNY_NUMBER1 + 
@@ -809,7 +853,7 @@ tail_recur:
 	    hash = hash*FUNNY_NUMBER2 + funp->fe->old_uniq;
 	    if (num_free > 0) {
 		if (num_free > 1) {
-		    ESTACK_PUSH3(stack, (Eterm) &funp->env[1], (num_free-1), MAKE_HASH_FUN_OP);
+		    WSTACK_PUSH3(stack, (UWord) &funp->env[1], (num_free-1), MAKE_HASH_FUN_OP);
 		}
 		term = funp->env[0];
 		goto tail_recur;
@@ -837,9 +881,9 @@ tail_recur:
 	}
 
     case MAKE_HASH_CDR_PRE_OP:
-	term = ESTACK_POP(stack);
+	term = (Eterm) WSTACK_POP(stack);
 	if (is_not_list(term)) {
-	    ESTACK_PUSH(stack, MAKE_HASH_CDR_POST_OP);
+	    WSTACK_PUSH(stack, (UWord) MAKE_HASH_CDR_POST_OP);
 	    goto tail_recur;
 	}
 	/* fall through */
@@ -854,13 +898,13 @@ tail_recur:
 		hash = hash*FUNNY_NUMBER2 + unsigned_val(*list);
 		
 		if (is_not_list(CDR(list))) {
-		    ESTACK_PUSH(stack, MAKE_HASH_CDR_POST_OP);
+		    WSTACK_PUSH(stack, MAKE_HASH_CDR_POST_OP);
 		    term = CDR(list);
 		    goto tail_recur;
 		}		
 		list = list_val(CDR(list));
 	    }
-	    ESTACK_PUSH2(stack, CDR(list), MAKE_HASH_CDR_PRE_OP);
+	    WSTACK_PUSH2(stack, CDR(list), MAKE_HASH_CDR_PRE_OP);
 	    term = CAR(list);
 	    goto tail_recur;
 	}
@@ -888,7 +932,7 @@ tail_recur:
 	    }
 	    d = BIG_DIGIT(ptr, k);
 	    k = sizeof(ErtsDigit);
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
 	    if (!(d >> 32))
 		k /= 2;
 #endif
@@ -904,21 +948,21 @@ tail_recur:
 	    Eterm* ptr = tuple_val(term);
 	    Uint arity = arityval(*ptr);
 
-	    ESTACK_PUSH3(stack, arity, (Eterm)(ptr+1), arity);
+	    WSTACK_PUSH3(stack, (UWord) arity, (UWord)(ptr+1), (UWord) arity);
 	    op = MAKE_HASH_TUPLE_OP;	    
 	}/*fall through*/
     case MAKE_HASH_TUPLE_OP:
     case MAKE_HASH_FUN_OP:
 	{
-	    Uint i = ESTACK_POP(stack);
-	    Eterm* ptr = (Eterm*) ESTACK_POP(stack);
+	    Uint i = (Uint) WSTACK_POP(stack);
+	    Eterm* ptr = (Eterm*) WSTACK_POP(stack);
 	    if (i != 0) {
 		term = *ptr;
-		ESTACK_PUSH3(stack, (Eterm)(ptr+1), i-1, op);
+		WSTACK_PUSH3(stack, (UWord)(ptr+1), (UWord) i-1, (UWord) op);
 		goto tail_recur;
 	    }
 	    if (op == MAKE_HASH_TUPLE_OP) {
-		Uint32 arity = ESTACK_POP(stack);
+		Uint32 arity = (Uint32) WSTACK_POP(stack);
 		hash = hash*FUNNY_NUMBER9 + arity;
 	    }
 	    break;
@@ -928,10 +972,10 @@ tail_recur:
 	erl_exit(1, "Invalid tag in make_hash(0x%X,0x%X)\n", term, op);
 	return 0;
       }
-      if (ESTACK_ISEMPTY(stack)) break;
-      op = ESTACK_POP(stack);
+      if (WSTACK_ISEMPTY(stack)) break;
+      op = WSTACK_POP(stack);
     }
-    DESTROY_ESTACK(stack);
+    DESTROY_WSTACK(stack);
     return hash;
 
 #undef UINT32_HASH_STEP
@@ -1123,7 +1167,7 @@ make_hash2(Eterm term)
 	    break;
 	    case EXPORT_SUBTAG:
 	    {
-		Export* ep = (Export *) (export_val(term))[1];
+		Export* ep = *((Export **) (export_val(term) + 1));
 
 		UINT32_HASH_2
 		    (ep->code[2], 
@@ -1333,7 +1377,7 @@ make_hash2(Eterm term)
 Uint32 make_broken_hash(Eterm term)
 {
     Uint32 hash = 0;
-    DECLARE_ESTACK(stack);
+    DECLARE_WSTACK(stack);
     unsigned op;
 tail_recur:
     op = tag_val_def(term); 
@@ -1347,7 +1391,7 @@ tail_recur:
 	    (atom_tab(atom_val(term))->slot.bucket.hvalue);
 	break;
     case SMALL_DEF:
-#ifdef ARCH_64
+#if defined(ARCH_64) && !HALFWORD_HEAP
     {
 	Sint y1 = signed_val(term);
 	Uint y2 = y1 < 0 ? -(Uint)y1 : y1;
@@ -1400,7 +1444,7 @@ tail_recur:
 
     case EXPORT_DEF:
 	{
-	    Export* ep = (Export *) (export_val(term))[1];
+	    Export* ep = *((Export **) (export_val(term) + 1));
 
 	    hash = hash * FUNNY_NUMBER11 + ep->code[2];
 	    hash = hash*FUNNY_NUMBER1 + 
@@ -1422,7 +1466,7 @@ tail_recur:
 	    hash = hash*FUNNY_NUMBER2 + funp->fe->old_uniq;
 	    if (num_free > 0) {
 		if (num_free > 1) {
-		    ESTACK_PUSH3(stack, (Eterm) &funp->env[1], (num_free-1), MAKE_HASH_FUN_OP);
+		    WSTACK_PUSH3(stack, (UWord) &funp->env[1], (num_free-1), MAKE_HASH_FUN_OP);
 		}
 		term = funp->env[0];
 		goto tail_recur;
@@ -1457,16 +1501,17 @@ tail_recur:
 	break;
 
     case MAKE_HASH_CDR_PRE_OP:
-	term = ESTACK_POP(stack);
+	term = (Eterm) WSTACK_POP(stack);
 	if (is_not_list(term)) {
-	    ESTACK_PUSH(stack, MAKE_HASH_CDR_POST_OP);
+	    WSTACK_PUSH(stack, (UWord) MAKE_HASH_CDR_POST_OP);
 	    goto tail_recur;
 	}
 	/*fall through*/
     case LIST_DEF:
 	{
 	    Eterm* list = list_val(term);
-	    ESTACK_PUSH2(stack, CDR(list), MAKE_HASH_CDR_PRE_OP);
+	    WSTACK_PUSH2(stack, (UWord) CDR(list),
+			 (UWord) MAKE_HASH_CDR_PRE_OP);
 	    term = CAR(list);
 	    goto tail_recur;
 	}
@@ -1539,21 +1584,21 @@ tail_recur:
 	    Eterm* ptr = tuple_val(term);
 	    Uint arity = arityval(*ptr);
 
-	    ESTACK_PUSH3(stack, arity, (Eterm)(ptr+1), arity);
+	    WSTACK_PUSH3(stack, (UWord) arity, (UWord) (ptr+1), (UWord) arity);
 	    op = MAKE_HASH_TUPLE_OP;
 	}/*fall through*/ 
     case MAKE_HASH_TUPLE_OP:
     case MAKE_HASH_FUN_OP:
 	{
-	    Uint i = ESTACK_POP(stack);
-	    Eterm* ptr = (Eterm*) ESTACK_POP(stack);
+	    Uint i = (Uint) WSTACK_POP(stack);
+	    Eterm* ptr = (Eterm*) WSTACK_POP(stack);
 	    if (i != 0) {
 		term = *ptr;
-		ESTACK_PUSH3(stack, (Eterm)(ptr+1), i-1, op);
+		WSTACK_PUSH3(stack, (UWord)(ptr+1), (UWord) i-1, (UWord) op);
 		goto tail_recur;
 	    }
 	    if (op == MAKE_HASH_TUPLE_OP) {
-		Uint32 arity = ESTACK_POP(stack);
+		Uint32 arity = (UWord) WSTACK_POP(stack);
 		hash = hash*FUNNY_NUMBER9 + arity;
 	    }
 	    break;
@@ -1563,11 +1608,11 @@ tail_recur:
 	erl_exit(1, "Invalid tag in make_broken_hash\n");
 	return 0;
       }
-      if (ESTACK_ISEMPTY(stack)) break;
-      op = ESTACK_POP(stack);
+      if (WSTACK_ISEMPTY(stack)) break;
+      op = (Uint) WSTACK_POP(stack);
     }
 
-    DESTROY_ESTACK(stack);
+    DESTROY_WSTACK(stack);
     return hash;
     
 #undef MAKE_HASH_TUPLE_OP
@@ -1870,7 +1915,7 @@ erts_destroy_tmp_dsbuf(erts_dsprintf_buf_t *dsbufp)
 
 int eq(Eterm a, Eterm b)
 {
-    DECLARE_ESTACK(stack);
+    DECLARE_WSTACK(stack);
     Sint sz;
     Eterm* aa;
     Eterm* bb;	
@@ -1888,7 +1933,7 @@ tailrecur_ne:
 		Eterm atmp = CAR(aval);
 		Eterm btmp = CAR(bval);
 		if (atmp != btmp) {
-		    ESTACK_PUSH2(stack,CDR(bval),CDR(aval));
+		    WSTACK_PUSH2(stack,(UWord) CDR(bval),(UWord) CDR(aval));
 		    a = atmp;
 		    b = btmp;
 		    goto tailrecur_ne;
@@ -1958,8 +2003,8 @@ tailrecur_ne:
 	    case EXPORT_SUBTAG:
 		{
 		    if (is_export(b)) {
-			Export* a_exp = (Export *) (export_val(a))[1];
-			Export* b_exp = (Export *) (export_val(b))[1];
+			Export* a_exp = *((Export **) (export_val(a) + 1));
+			Export* b_exp = *((Export **) (export_val(b) + 1));
 			if (a_exp == b_exp) goto pop_next;
 		    }
 		    break; /* not equal */
@@ -2131,32 +2176,32 @@ term_array: /* arrays in 'aa' and 'bb', length in 'sz' */
 	    goto not_equal;
 	}
 	if (i > 1) { /* push the rest */
-	    ESTACK_PUSH3(stack, i-1, (Eterm)(bp+1),
-			 ((Eterm)(ap+1)) | TAG_PRIMARY_HEADER);
+	    WSTACK_PUSH3(stack, i-1, (UWord)(bp+1),
+			 ((UWord)(ap+1)) | TAG_PRIMARY_HEADER);
 	    /* We (ab)use TAG_PRIMARY_HEADER to recognize a term_array */
 	}
 	goto tailrecur_ne;
     }
    
 pop_next:
-    if (!ESTACK_ISEMPTY(stack)) {
-	Eterm something  = ESTACK_POP(stack);	
-	if (primary_tag(something) == TAG_PRIMARY_HEADER) { /* a term_array */
+    if (!WSTACK_ISEMPTY(stack)) {
+	UWord something  = WSTACK_POP(stack);
+	if (primary_tag((Eterm) something) == TAG_PRIMARY_HEADER) { /* a term_array */
 	    aa = (Eterm*) something;
-	    bb = (Eterm*) ESTACK_POP(stack);
-	    sz = ESTACK_POP(stack);
+	    bb = (Eterm*) WSTACK_POP(stack);
+	    sz = WSTACK_POP(stack);
 	    goto term_array;
 	}
 	a = something;
-	b = ESTACK_POP(stack);
+	b = WSTACK_POP(stack);
 	goto tailrecur;
     }
 
-    DESTROY_ESTACK(stack);
+    DESTROY_WSTACK(stack);
     return 1;
 
 not_equal:
-    DESTROY_ESTACK(stack);
+    DESTROY_WSTACK(stack);
     return 0;
 }
 
@@ -2211,7 +2256,7 @@ static int cmp_atoms(Eterm a, Eterm b)
 
 Sint cmp(Eterm a, Eterm b)
 {
-    DECLARE_ESTACK(stack);
+    DECLARE_WSTACK(stack);
     Eterm* aa;
     Eterm* bb;
     int i;
@@ -2328,7 +2373,7 @@ tailrecur_ne:
 	    Eterm atmp = CAR(aa);
 	    Eterm btmp = CAR(bb);
 	    if (atmp != btmp) {
-		ESTACK_PUSH2(stack,CDR(bb),CDR(aa));
+		WSTACK_PUSH2(stack,(UWord) CDR(bb),(UWord) CDR(aa));
 		a = atmp;
 		b = btmp;
 		goto tailrecur_ne;
@@ -2393,8 +2438,8 @@ tailrecur_ne:
 		    a_tag = EXPORT_DEF;
 		    goto mixed_types;
 		} else {
-		    Export* a_exp = (Export *) (export_val(a))[1];
-		    Export* b_exp = (Export *) (export_val(b))[1];
+		    Export* a_exp = *((Export **) (export_val(a) + 1));
+		    Export* b_exp = *((Export **) (export_val(b) + 1));
 
 		    if ((j = cmp_atoms(a_exp->code[0], b_exp->code[0])) != 0) {
 			RETURN_NEQ(j);
@@ -2649,7 +2694,7 @@ term_array: /* arrays in 'aa' and 'bb', length in 'i' */
 		}
 	    } else {
 		/* (ab)Use TAG_PRIMARY_HEADER to recognize a term_array */
-		ESTACK_PUSH3(stack, i, (Eterm)bb, (Eterm)aa | TAG_PRIMARY_HEADER);
+		WSTACK_PUSH3(stack, i, (UWord)bb, (UWord)aa | TAG_PRIMARY_HEADER);
 		goto tailrecur_ne;
 	    }
 	}
@@ -2659,20 +2704,20 @@ term_array: /* arrays in 'aa' and 'bb', length in 'i' */
     goto tailrecur;    
    
 pop_next:
-    if (!ESTACK_ISEMPTY(stack)) {
-	Eterm something = ESTACK_POP(stack);
-	if (primary_tag(something) == TAG_PRIMARY_HEADER) { /* a term_array */
+    if (!WSTACK_ISEMPTY(stack)) {
+	UWord something = WSTACK_POP(stack);
+	if (primary_tag((Eterm) something) == TAG_PRIMARY_HEADER) { /* a term_array */
 	    aa = (Eterm*) something;
-	    bb = (Eterm*) ESTACK_POP(stack);
-	    i = ESTACK_POP(stack);
+	    bb = (Eterm*) WSTACK_POP(stack);
+	    i = WSTACK_POP(stack);
 	    goto term_array;
 	}
-	a = something;
-	b = ESTACK_POP(stack);
+	a = (Eterm) something;
+	b = (Eterm) WSTACK_POP(stack);
 	goto tailrecur;
     }
 
-    DESTROY_ESTACK(stack);
+    DESTROY_WSTACK(stack);
     return 0;
 
 not_equal:
diff --git a/erts/emulator/drivers/common/efile_drv.c b/erts/emulator/drivers/common/efile_drv.c
index 95510a16b2..cbd4dcb682 100644
--- a/erts/emulator/drivers/common/efile_drv.c
+++ b/erts/emulator/drivers/common/efile_drv.c
@@ -196,9 +196,9 @@ enum e_timer {timer_idle, timer_again, timer_write};
 struct t_data;
 
 typedef struct {
-    Sint            fd;
+    SWord           fd;
     ErlDrvPort      port;
-    unsigned        key;      /* Async queue key */
+    unsigned long   key;      /* Async queue key */
     unsigned        flags;    /* Original flags from FILE_OPEN. */
     void          (*invoke)(void *);
     struct t_data  *d;
@@ -306,7 +306,7 @@ struct t_data
     int            result_ok;
     Efile_error    errInfo;
     int            flags;
-    Sint           fd;
+    SWord          fd;
     /**/
     Efile_info        info;
     EFILE_DIR_HANDLE  dir_handle; /* Handle to open directory. */
@@ -605,7 +605,7 @@ file_start(ErlDrvPort port, char* command)
     }
     desc->fd = FILE_FD_INVALID;
     desc->port = port;
-    desc->key = (unsigned) (Uint) port;
+    desc->key = (unsigned long) (UWord) port;
     desc->flags = 0;
     desc->invoke = NULL;
     desc->d = NULL;
@@ -630,7 +630,7 @@ static void free_data(void *data)
     EF_FREE(data);
 }
 
-static void do_close(int flags, Sint fd) {
+static void do_close(int flags, SWord fd) {
     if (flags & EFILE_COMPRESSED) {
 	erts_gzclose((gzFile)(fd));
     } else {
@@ -709,7 +709,7 @@ static void reply_Uint_posix_error(file_descriptor *desc, Uint num,
     TRACE_C('N');
 
     response[0] = FILE_RESP_NUMERR;
-#if SIZEOF_VOID_P == 4
+#if SIZEOF_VOID_P == 4 || HALFWORD_HEAP
     put_int32(0, response+1);
 #else
     put_int32(num>>32, response+1);
@@ -767,7 +767,7 @@ static int reply_Uint(file_descriptor *desc, Uint result) {
     TRACE_C('R');
 
     tmp[0] = FILE_RESP_NUMBER;
-#if SIZEOF_VOID_P == 4
+#if SIZEOF_VOID_P == 4 || HALFWORD_HEAP
     put_int32(0, tmp+1);
 #else
     put_int32(result>>32, tmp+1);
@@ -1620,7 +1620,7 @@ static void invoke_open(void *data)
 	    status = efile_may_openfile(&d->errInfo, d->b);
 	    if (status || (d->errInfo.posix_errno != EISDIR)) {
 		mode = (d->flags & EFILE_MODE_READ) ? "rb" : "wb";
-		d->fd = (Sint) erts_gzopen(d->b, mode);
+		d->fd = (SWord) erts_gzopen(d->b, mode);
 		if ((gzFile)d->fd) {
 		    status = 1;
 		} else {
diff --git a/erts/emulator/drivers/unix/ttsl_drv.c b/erts/emulator/drivers/unix/ttsl_drv.c
index 4cd54c073f..d782b044a9 100644
--- a/erts/emulator/drivers/unix/ttsl_drv.c
+++ b/erts/emulator/drivers/unix/ttsl_drv.c
@@ -314,7 +314,7 @@ static ErlDrvData ttysl_start(ErlDrvPort port, char* buf)
     sys_sigset(SIGCONT, cont);
     sys_sigset(SIGWINCH, winch);
 
-    driver_select(port, (ErlDrvEvent)(Uint)ttysl_fd, ERL_DRV_READ|ERL_DRV_USE, 1);
+    driver_select(port, (ErlDrvEvent)(UWord)ttysl_fd, ERL_DRV_READ|ERL_DRV_USE, 1);
     ttysl_port = port;
 
     /* we need to know this when we enter the break handler */
@@ -394,7 +394,7 @@ static void ttysl_stop(ErlDrvData ttysl_data)
 	stop_lbuf();
 	stop_termcap();
 	tty_reset(ttysl_fd);
-	driver_select(ttysl_port, (ErlDrvEvent)(Uint)ttysl_fd, ERL_DRV_READ|ERL_DRV_USE, 0);
+	driver_select(ttysl_port, (ErlDrvEvent)(UWord)ttysl_fd, ERL_DRV_READ|ERL_DRV_USE, 0);
 	sys_sigset(SIGCONT, SIG_DFL);
 	sys_sigset(SIGWINCH, SIG_DFL);
     }
@@ -685,7 +685,7 @@ static void ttysl_from_tty(ErlDrvData ttysl_data, ErlDrvEvent fd)
 	utf8buf_size = 0;
     }
     
-    if ((i = read((int)(Sint)fd, (char *) p, left)) >= 0) {
+    if ((i = read((int)(SWord)fd, (char *) p, left)) >= 0) {
 	if (p != b) {
 	    i += (p - b);
 	}
diff --git a/erts/emulator/sys/common/erl_mseg.h b/erts/emulator/sys/common/erl_mseg.h
index 1c5aa63e90..adbfd30f0f 100644
--- a/erts/emulator/sys/common/erl_mseg.h
+++ b/erts/emulator/sys/common/erl_mseg.h
@@ -58,8 +58,8 @@ typedef struct {
 typedef struct {
     int  cache;
     int  preserv;
-    Uint abs_shrink_th;
-    Uint rel_shrink_th;
+    UWord abs_shrink_th;
+    UWord rel_shrink_th;
 } ErtsMsegOpt_t;
 
 #define ERTS_MSEG_DEFAULT_OPT_INITIALIZER				\
diff --git a/erts/emulator/utils/beam_makeops b/erts/emulator/utils/beam_makeops
index 4a859c3094..1a76e80620 100755
--- a/erts/emulator/utils/beam_makeops
+++ b/erts/emulator/utils/beam_makeops
@@ -812,9 +812,10 @@ sub basic_generator {
     # Pack arguments if requested.
     #
 
-    if ($flags =~ /-pack/ && $hot) {
-	($prefix, $pack_spec, @args) = &do_pack(@args);
-    }
+    # /* XXX:PaN temporarilly disabled during halfword implementation step 1 */
+    #if ($flags =~ /-pack/ && $hot) {
+    #    ($prefix, $pack_spec, @args) = &do_pack(@args);
+    #}
 
     #
     # Calculate the size of the instruction and generate each argument for
@@ -916,7 +917,7 @@ sub basic_generator {
 	} else {
 	    $code = join("\n",
 			 "{ $var_decls",
-			 "Eterm* next;",
+			 "UWord* next;",
 			 "PreFetch($size, next);",
 			 "$macro_code",
 			 "NextPF($size, next);",
diff --git a/erts/emulator/utils/make_tables b/erts/emulator/utils/make_tables
index b5391234cf..8e9e414dbe 100755
--- a/erts/emulator/utils/make_tables
+++ b/erts/emulator/utils/make_tables
@@ -184,7 +184,7 @@ for ($i = 0; $i < @bif; $i++) {
     my $arity = $bif[$i]->[2];
     my $args = join(', ', 'Process*', ('Eterm') x $arity);
     print "Eterm $bif[$i]->[3]($args);\n";
-    print "Eterm wrap_$bif[$i]->[3]($args, Uint *I);\n";
+    print "Eterm wrap_$bif[$i]->[3]($args, UWord *I);\n";
 }
 print "#endif\n";
 
@@ -225,15 +225,15 @@ for ($i = 0; $i < @bif; $i++) {
     for ($arg = 1; $arg <= $arity; $arg++) {
 	print ", Eterm arg$arg";
     }
-    print ", Uint *I)\n";
+    print ", UWord *I)\n";
     print "{\n";
     print "    return erts_bif_trace($i, p";
     for ($arg = 1; $arg <= 3; $arg++) {
 	if ($arg <= $arity) {
 	    print ", arg$arg";
-	} elsif ($arg == ($arity + 1)) { 
-	    # Place I in correct position
-	    print ", (Eterm) I";
+	#} elsif ($arg == ($arity + 1)) {
+	  #  # Place I in correct position
+	   # print ", (Eterm) I";
 	} else {
 	    print ", 0";
 	}