Merge branch 'np/sharing-preserved-copy/OTP-12590'

* np/sharing-preserved-copy/OTP-12590: (28 commits)
  Refactor have seq_trace token test
  Use sharing preserving copy error messages and exceptions
  Use sharing preserving copy in enif_make_copy
  Refactor sharing preserved copy flags
  Fix rebase of SHCOPY seq_tokens
  Fix erts_debug:copy_shared/1 prototype
  Update preloaded module erlang.beam
  Add erlang:copy_literals/2 spec
  Add copy_literals testcase
  Do not use GCC extensions in copy
  Use copy literal range check in message passing and purging
  Add BIF for setting internal copy literal range
  Copy literals in copy sharing
  Refactor copy sharing
  Add support for HAMT maps in preserved copy
  Fix Map preserved sharing copy implementation
  Fix Halfword removal
  Fix internal stacks
  Add support for maps in preserved copy
  Add --enable-sharing-preserving configure flag
  ...

29 files changed, 1895 insertions, 213 deletions

diff --git a/configure.in b/configure.in
index e8c8680e19..51728a847e 100644
--- a/configure.in
+++ b/configure.in
@@ -278,6 +278,10 @@ AC_ARG_ENABLE(builtin-zlib,
 AS_HELP_STRING([--enable-builtin-zlib],
                [force use of our own built-in zlib]))
 
+AC_ARG_ENABLE(sharing-preserving,
+AS_HELP_STRING([--enable-sharing-preserving],
+               [enable copying of terms without destroying sharing]))
+
 dnl This functionality has been lost along the way... :(
 dnl It could perhaps be nice to reintroduce some day; therefore,
 dnl it is not removed just commented out.
diff --git a/erts/configure.in b/erts/configure.in
index 2419925c33..9ad1588b6c 100644
--- a/erts/configure.in
+++ b/erts/configure.in
@@ -788,6 +788,18 @@ esac
 	
 AC_SUBST(LIBCARBON)
 
+dnl Check if we should/can build a sharing-preserving emulator
+
+AC_MSG_CHECKING(if we are building a sharing-preserving emulator)
+if test "$enable_sharing_preserving" = "yes"; then
+        AC_DEFINE(SHCOPY, [1],
+		  [Define if building a sharing-preserving emulator])
+	AC_MSG_RESULT([yes])
+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 c3ebf71a01..6b6c066211 100644
--- a/erts/emulator/beam/beam_bif_load.c
+++ b/erts/emulator/beam/beam_bif_load.c
@@ -154,7 +154,7 @@ static struct /* Protected by code_write_permission */
 {
     Process* stager;
     ErtsThrPrgrLaterOp lop;
-}commiter_state;
+} committer_state;
 #endif
 
 static Eterm
@@ -367,9 +367,9 @@ staging_epilogue(Process* c_p, int commit, Eterm res, int is_blocking,
 	 * schedulers to read active code_ix in a safe way while executing
 	 * without any memory barriers at all. 
 	 */
-	ASSERT(commiter_state.stager == NULL);
-	commiter_state.stager = c_p;
-	erts_schedule_thr_prgr_later_op(smp_code_ix_commiter, NULL, &commiter_state.lop);
+	ASSERT(committer_state.stager == NULL);
+	committer_state.stager = c_p;
+	erts_schedule_thr_prgr_later_op(smp_code_ix_commiter, NULL, &committer_state.lop);
 	erts_proc_inc_refc(c_p);
 	erts_suspend(c_p, ERTS_PROC_LOCK_MAIN, NULL);
 	/*
@@ -385,11 +385,11 @@ staging_epilogue(Process* c_p, int commit, Eterm res, int is_blocking,
 #ifdef ERTS_SMP
 static void smp_code_ix_commiter(void* null)
 {
-    Process* p = commiter_state.stager;
+    Process* p = committer_state.stager;
 
     erts_commit_staging_code_ix();
 #ifdef DEBUG
-    commiter_state.stager = NULL;
+    committer_state.stager = NULL;
 #endif
     erts_release_code_write_permission();
     erts_smp_proc_lock(p, ERTS_PROC_LOCK_STATUS);
@@ -1007,6 +1007,93 @@ any_heap_refs(Eterm* start, Eterm* end, char* mod_start, Uint mod_size)
     return 0;
 }
 
+#undef in_area
+
+#ifdef ERTS_SMP
+static void copy_literals_commit(void*);
+#endif
+
+copy_literals_t erts_clrange = {NULL, 0};
+
+/* copy literals
+ *
+ * copy_literals.ptr = LitPtr
+ * copy_literals.sz  = LitSz
+ * ------ THR PROG COMMIT -----
+ *
+ *  - check process code
+ *  - check process code
+ *  ...
+ * copy_literals.ptr = NULL
+ * copy_literals.sz  = 0
+ * ------ THR PROG COMMIT -----
+ *  ...
+ */
+
+
+BIF_RETTYPE copy_literals_2(BIF_ALIST_2)
+{
+    Module* modp;
+    ErtsCodeIndex code_ix;
+    Eterm res = am_true;
+
+    if (is_not_atom(BIF_ARG_1) || (am_true != BIF_ARG_2 && am_false != BIF_ARG_2)) {
+        BIF_ERROR(BIF_P, BADARG);
+    }
+
+    if (!erts_try_seize_code_write_permission(BIF_P)) {
+	ERTS_BIF_YIELD2(bif_export[BIF_copy_literals_2], BIF_P, BIF_ARG_1, BIF_ARG_2);
+    }
+
+    code_ix = erts_active_code_ix();
+
+    if ((modp = erts_get_module(BIF_ARG_1, code_ix)) == NULL || !modp->old.code_hdr) {
+        res = am_false;
+        goto done;
+    }
+
+    if (BIF_ARG_2 == am_true) {
+        if (erts_clrange.ptr != NULL) {
+            res = am_aborted;
+            goto done;
+       }
+        erts_clrange.ptr = (Eterm*) modp->old.code_hdr->literals_start;
+        erts_clrange.sz  = (Eterm*) modp->old.code_hdr->literals_end - erts_clrange.ptr;
+    } else if (BIF_ARG_2 == am_false) {
+        erts_clrange.ptr = NULL;
+        erts_clrange.sz  = 0;
+    }
+
+#ifdef ERTS_SMP
+    ASSERT(committer_state.stager == NULL);
+    committer_state.stager = BIF_P;
+    erts_schedule_thr_prgr_later_op(copy_literals_commit, NULL, &committer_state.lop);
+    erts_proc_inc_refc(BIF_P);
+    erts_suspend(BIF_P, ERTS_PROC_LOCK_MAIN, NULL);
+    ERTS_BIF_YIELD_RETURN(BIF_P, am_true);
+#endif
+done:
+    erts_release_code_write_permission();
+    BIF_RET(res);
+}
+
+#ifdef ERTS_SMP
+static void copy_literals_commit(void* null) {
+    Process* p = committer_state.stager;
+#ifdef DEBUG
+    committer_state.stager = NULL;
+#endif
+    erts_release_code_write_permission();
+    erts_smp_proc_lock(p, ERTS_PROC_LOCK_STATUS);
+    if (!ERTS_PROC_IS_EXITING(p)) {
+	erts_resume(p, ERTS_PROC_LOCK_STATUS);
+    }
+    erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS);
+    erts_proc_dec_refc(p);
+}
+#endif /* ERTS_SMP */
+
+
 BIF_RETTYPE purge_module_1(BIF_ALIST_1)
 {
     ErtsCodeIndex code_ix;
diff --git a/erts/emulator/beam/beam_debug.c b/erts/emulator/beam/beam_debug.c
index e989310789..e37bd4d78c 100644
--- a/erts/emulator/beam/beam_debug.c
+++ b/erts/emulator/beam/beam_debug.c
@@ -73,6 +73,40 @@ erts_debug_flat_size_1(BIF_ALIST_1)
     }
 }
 
+BIF_RETTYPE
+erts_debug_size_shared_1(BIF_ALIST_1)
+{
+    Process* p = BIF_P;
+    Eterm term = BIF_ARG_1;
+    Uint size = size_shared(term);
+
+    if (IS_USMALL(0, size)) {
+	BIF_RET(make_small(size));
+    } else {
+	Eterm* hp = HAlloc(p, BIG_UINT_HEAP_SIZE);
+	BIF_RET(uint_to_big(size, hp));
+    }
+}
+
+BIF_RETTYPE
+erts_debug_copy_shared_1(BIF_ALIST_1)
+{
+    Process* p = BIF_P;
+    Eterm term = BIF_ARG_1;
+    Uint size;
+    Eterm* hp;
+    Eterm copy;
+    erts_shcopy_t info;
+    INITIALIZE_SHCOPY(info);
+
+    size = copy_shared_calculate(term, &info);
+    if (size > 0) {
+      hp = HAlloc(p, size);
+    }
+    copy = copy_shared_perform(term, size, &info, &hp, &p->off_heap);
+    DESTROY_SHCOPY(info);
+    BIF_RET(copy);
+}
 
 BIF_RETTYPE
 erts_debug_breakpoint_2(BIF_ALIST_2)
diff --git a/erts/emulator/beam/beam_emu.c b/erts/emulator/beam/beam_emu.c
index 208a16dfd0..1a4133bceb 100644
--- a/erts/emulator/beam/beam_emu.c
+++ b/erts/emulator/beam/beam_emu.c
@@ -1914,20 +1914,7 @@ void process_main(void)
 	 if (DT_UTAG(c_p) != NIL) {
 	     if (DT_UTAG_FLAGS(c_p) & DT_UTAG_PERMANENT) {
 		 SEQ_TRACE_TOKEN(c_p) = am_have_dt_utag;
-#ifdef DTRACE_TAG_HARDDEBUG
-		 if (DT_UTAG_FLAGS(c_p) & DT_UTAG_SPREADING) 
-		     erts_fprintf(stderr,
-				  "Dtrace -> (%T) stop spreading "
-				  "tag %T with message %T\r\n",
-				  c_p->common.id,DT_UTAG(c_p),ERL_MESSAGE_TERM(msgp));
-#endif
 	     } else {
-#ifdef DTRACE_TAG_HARDDEBUG
-		 erts_fprintf(stderr,
-			      "Dtrace -> (%T) kill tag %T with "
-			      "message %T\r\n",
-			      c_p->common.id,DT_UTAG(c_p),ERL_MESSAGE_TERM(msgp));
-#endif
 		 DT_UTAG(c_p) = NIL;
 		 SEQ_TRACE_TOKEN(c_p) = NIL;
 	     }
@@ -1947,12 +1934,6 @@ void process_main(void)
 		 DT_UTAG(c_p) = ERL_MESSAGE_DT_UTAG(msgp);
 	     }
 	     DT_UTAG_FLAGS(c_p) |= DT_UTAG_SPREADING;
-#ifdef DTRACE_TAG_HARDDEBUG
-	     erts_fprintf(stderr,
-			  "Dtrace -> (%T) receive tag (%T) "
-			  "with message %T\r\n",
-			  c_p->common.id, DT_UTAG(c_p), ERL_MESSAGE_TERM(msgp));
-#endif
 	 } else {
 #endif
 	     ASSERT(is_tuple(SEQ_TRACE_TOKEN(c_p)));
@@ -1982,7 +1963,7 @@ void process_main(void)
 
          dtrace_proc_str(c_p, receiver_name);
          token2 = SEQ_TRACE_TOKEN(c_p);
-         if (token2 != NIL && token2 != am_have_dt_utag) {
+         if (have_seqtrace(token2)) {
              tok_label = signed_val(SEQ_TRACE_T_LABEL(token2));
              tok_lastcnt = signed_val(SEQ_TRACE_T_LASTCNT(token2));
              tok_serial = signed_val(SEQ_TRACE_T_SERIAL(token2));
diff --git a/erts/emulator/beam/bif.c b/erts/emulator/beam/bif.c
index f0340540cb..14ab113b32 100644
--- a/erts/emulator/beam/bif.c
+++ b/erts/emulator/beam/bif.c
@@ -2038,11 +2038,7 @@ do_send(Process *p, Eterm to, Eterm msg, Eterm *refp, ErtsSendContext* ctx)
 	if (ERTS_PROC_GET_SAVED_CALLS_BUF(p))
 	    save_calls(p, &exp_send);
 	
-	if (SEQ_TRACE_TOKEN(p) != NIL
-#ifdef USE_VM_PROBES
-	    && SEQ_TRACE_TOKEN(p) != am_have_dt_utag
-#endif
-	    ) {
+        if (have_seqtrace(SEQ_TRACE_TOKEN(p))) {
 	    seq_trace_update_send(p);
 	    seq_trace_output(SEQ_TRACE_TOKEN(p), msg, 
 			     SEQ_TRACE_SEND, portid, p);
diff --git a/erts/emulator/beam/bif.tab b/erts/emulator/beam/bif.tab
index 65f8d6f1f5..c49a3ff313 100644
--- a/erts/emulator/beam/bif.tab
+++ b/erts/emulator/beam/bif.tab
@@ -639,14 +639,17 @@ bif ets:update_counter/4
 bif erts_debug:map_info/1
 
 #
-# Obsolete
+# New in 19.0
 #
 
-bif erlang:hash/2
+bif erlang:copy_literals/2
+bif binary:split/2
+bif binary:split/3
+bif erts_debug:size_shared/1
+bif erts_debug:copy_shared/1
 
 #
-# New in 19.0
+# Obsolete
 #
 
-bif binary:split/2
-bif binary:split/3
+bif erlang:hash/2
diff --git a/erts/emulator/beam/copy.c b/erts/emulator/beam/copy.c
index f27c526413..67a96f6442 100644
--- a/erts/emulator/beam/copy.c
+++ b/erts/emulator/beam/copy.c
@@ -75,8 +75,14 @@ Uint size_object(Eterm obj)
     Uint sum = 0;
     Eterm* ptr;
     int arity;
+#ifdef DEBUG
+    Eterm mypid = erts_get_current_pid();
+#endif
 
     DECLARE_ESTACK(s);
+
+    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] size_object %p\n", mypid, obj));
+
     for (;;) {
 	switch (primary_tag(obj)) {
 	case TAG_PRIMARY_LIST:
@@ -211,6 +217,7 @@ Uint size_object(Eterm obj)
 	pop_next:
 	    if (ESTACK_ISEMPTY(s)) {
 		DESTROY_ESTACK(s);
+		VERBOSE(DEBUG_SHCOPY, ("[pid=%T] size was: %u\n", mypid, sum));
 		return sum;
 	    }
 	    obj = ESTACK_POP(s);
@@ -222,9 +229,377 @@ Uint size_object(Eterm obj)
 }
 
 /*
+ *  Machinery for sharing preserving information
+ *  Using a WSTACK but not very transparently; consider refactoring
+ */
+
+#define DECLARE_BITSTORE(s)						\
+    DECLARE_WSTACK(s);							\
+    int WSTK_CONCAT(s,_bitoffs) = 0;					\
+    int WSTK_CONCAT(s,_offset) = 0;					\
+    UWord WSTK_CONCAT(s,_buffer) = 0
+
+#define DESTROY_BITSTORE(s) DESTROY_WSTACK(s)
+#define BITSTORE_PUT(s,i)						\
+do {									\
+    WSTK_CONCAT(s,_buffer) |= i << WSTK_CONCAT(s,_bitoffs);		\
+    WSTK_CONCAT(s,_bitoffs) += 2;					\
+    if (WSTK_CONCAT(s,_bitoffs) >= 8*sizeof(UWord)) {			\
+	WSTACK_PUSH(s, WSTK_CONCAT(s,_buffer));				\
+	WSTK_CONCAT(s,_bitoffs) = 0;					\
+	WSTK_CONCAT(s,_buffer) = 0;					\
+    }									\
+} while(0)
+#define BITSTORE_CLOSE(s)						\
+do {									\
+    if (WSTK_CONCAT(s,_bitoffs) > 0) {					\
+	WSTACK_PUSH(s, WSTK_CONCAT(s,_buffer));				\
+	WSTK_CONCAT(s,_bitoffs) = 0;					\
+    }									\
+} while(0)
+
+#define BITSTORE_FETCH(s,dst)                                           \
+do {                                                                    \
+    UWord result;                                                       \
+    if (WSTK_CONCAT(s,_bitoffs) <= 0) {                                 \
+        WSTK_CONCAT(s,_buffer) = s.wstart[WSTK_CONCAT(s,_offset)];      \
+        WSTK_CONCAT(s,_offset)++;                                       \
+        WSTK_CONCAT(s,_bitoffs) = 8*sizeof(UWord);                      \
+    }                                                                   \
+    WSTK_CONCAT(s,_bitoffs) -= 2;                                       \
+    result = WSTK_CONCAT(s,_buffer) & 3;                                \
+    WSTK_CONCAT(s,_buffer) >>= 2;                                       \
+    (dst) = result;                                                     \
+} while(0)
+
+#define BOXED_VISITED_MASK	 ((Eterm) 3)
+#define BOXED_VISITED		 ((Eterm) 1)
+#define BOXED_SHARED_UNPROCESSED ((Eterm) 2)
+#define BOXED_SHARED_PROCESSED	 ((Eterm) 3)
+
+#define COUNT_OFF_HEAP (0)
+
+#define IN_LITERAL_PURGE_AREA(info, ptr)                 \
+    ((info)->range_ptr && (                              \
+        (info)->range_ptr <= (ptr) &&                    \
+        (ptr) < ((info)->range_ptr + (info)->range_sz)))
+/*
+ *  Return the real size of an object and find sharing information
+ *  This currently returns the same as erts_debug:size/1.
+ *  It is argued whether the size of subterms in constant pools
+ *  should be counted or not.
+ */
+
+Uint size_shared(Eterm obj)
+{
+    Eterm saved_obj = obj;
+    Uint sum = 0;
+    Eterm* ptr;
+
+    DECLARE_EQUEUE(s);
+    DECLARE_BITSTORE(b);
+
+    for (;;) {
+	switch (primary_tag(obj)) {
+	case TAG_PRIMARY_LIST: {
+	    Eterm head, tail;
+	    ptr = list_val(obj);
+	    /* we're not counting anything that's outside our heap */
+	    if (!COUNT_OFF_HEAP && erts_is_literal(obj,ptr)) {
+		goto pop_next;
+	    }
+	    head = CAR(ptr);
+	    tail = CDR(ptr);
+	    /* if it's visited, don't count it */
+	    if (primary_tag(tail) == TAG_PRIMARY_HEADER ||
+		primary_tag(head) == TAG_PRIMARY_HEADER) {
+		goto pop_next;
+	    }
+	    /* else make it visited now */
+	    switch (primary_tag(tail)) {
+	    case TAG_PRIMARY_LIST:
+		ptr[1] = (tail - TAG_PRIMARY_LIST) | TAG_PRIMARY_HEADER;
+		break;
+	    case TAG_PRIMARY_IMMED1:
+		CAR(ptr) = (head - primary_tag(head)) | TAG_PRIMARY_HEADER;
+		CDR(ptr) = (tail - TAG_PRIMARY_IMMED1) | primary_tag(head);
+		break;
+	    case TAG_PRIMARY_BOXED:
+		BITSTORE_PUT(b, primary_tag(head));
+		CAR(ptr) = (head - primary_tag(head)) | TAG_PRIMARY_HEADER;
+		CDR(ptr) = (tail - TAG_PRIMARY_BOXED) | TAG_PRIMARY_HEADER;
+		break;
+	    }
+	    /* and count it */
+	    sum += 2;
+	    if (!IS_CONST(head)) {
+		EQUEUE_PUT(s, head);
+	    }
+	    obj = tail;
+	    break;
+	}
+	case TAG_PRIMARY_BOXED: {
+	    Eterm hdr;
+	    ptr = boxed_val(obj);
+	    /* we're not counting anything that's outside our heap */
+	    if (!COUNT_OFF_HEAP && erts_is_literal(obj,ptr)) {
+		goto pop_next;
+	    }
+	    hdr = *ptr;
+	    /* if it's visited, don't count it */
+	    if (primary_tag(hdr) != TAG_PRIMARY_HEADER) {
+		goto pop_next;
+	    }
+	    /* else make it visited now */
+	    *ptr = (hdr - primary_tag(hdr)) + BOXED_VISITED;
+	    /* and count it */
+	    ASSERT(is_header(hdr));
+	    switch (hdr & _TAG_HEADER_MASK) {
+	    case ARITYVAL_SUBTAG: {
+		int arity = header_arity(hdr);
+		sum += arity + 1;
+		if (arity == 0) { /* Empty tuple -- unusual. */
+		    goto pop_next;
+		}
+		while (arity-- > 0) {
+		    obj = *++ptr;
+		    if (!IS_CONST(obj)) {
+			EQUEUE_PUT(s, obj);
+		    }
+		}
+		goto pop_next;
+	    }
+	    case FUN_SUBTAG: {
+		ErlFunThing* funp = (ErlFunThing *) ptr;
+		unsigned eterms = 1 /* creator */ + funp->num_free;
+		unsigned sz = thing_arityval(hdr);
+		sum += 1 /* header */ + sz + eterms;
+		ptr += 1 /* header */ + sz;
+		while (eterms-- > 0) {
+		    obj = *ptr++;
+		    if (!IS_CONST(obj)) {
+			EQUEUE_PUT(s, obj);
+		    }
+		}
+		goto pop_next;
+	    }
+	    case SUB_BINARY_SUBTAG: {
+		ErlSubBin* sb = (ErlSubBin *) ptr;
+		Uint extra_bytes;
+		Eterm hdr;
+		ASSERT((sb->thing_word & ~BOXED_VISITED_MASK) == HEADER_SUB_BIN);
+		if (sb->bitsize + sb->bitoffs > 8) {
+		    sum += ERL_SUB_BIN_SIZE;
+		    extra_bytes = 2;
+		} else if (sb->bitsize + sb->bitoffs > 0) {
+		    sum += ERL_SUB_BIN_SIZE;
+		    extra_bytes = 1;
+		} else {
+		    extra_bytes = 0;
+		}
+		ptr = binary_val(sb->orig);
+		hdr = (*ptr) & ~BOXED_VISITED_MASK;
+		if (thing_subtag(hdr) == REFC_BINARY_SUBTAG) {
+		    sum += PROC_BIN_SIZE;
+		} else {
+		    ASSERT(thing_subtag(hdr) == HEAP_BINARY_SUBTAG);
+		    sum += heap_bin_size(binary_size(obj) + extra_bytes);
+		}
+		goto pop_next;
+	    }
+            case MAP_SUBTAG:
+                switch (MAP_HEADER_TYPE(hdr)) {
+                    case MAP_HEADER_TAG_FLATMAP_HEAD : {
+                        flatmap_t *mp  = (flatmap_t*)flatmap_val(obj);
+                        Uint n = flatmap_get_size(mp) + 1;
+                        ptr  = (Eterm *)mp;
+                        sum += n + 2;
+                        ptr += 2; /* hdr + size words */
+                        while (n--) {
+                            obj = *ptr++;
+                            if (!IS_CONST(obj)) {
+                                EQUEUE_PUT(s, obj);
+                            }
+                        }
+                        goto pop_next;
+                    }
+                    case MAP_HEADER_TAG_HAMT_HEAD_BITMAP :
+                    case MAP_HEADER_TAG_HAMT_HEAD_ARRAY :
+                    case MAP_HEADER_TAG_HAMT_NODE_BITMAP : {
+                        Uint n = hashmap_bitcount(MAP_HEADER_VAL(hdr));
+                        sum += 1 + n + header_arity(hdr);
+                        ptr += 1 + header_arity(hdr);
+                        while (n--) {
+                            obj = *ptr++;
+                            if (!IS_CONST(obj)) {
+                                EQUEUE_PUT(s, obj);
+                            }
+                        }
+                        goto pop_next;
+                    }
+                    default:
+                        erl_exit(ERTS_ABORT_EXIT, "size_shared: bad hashmap type %d\n", MAP_HEADER_TYPE(hdr));
+                }
+	    case BIN_MATCHSTATE_SUBTAG:
+		erl_exit(ERTS_ABORT_EXIT,
+			 "size_shared: matchstate term not allowed");
+	    default:
+		sum += thing_arityval(hdr) + 1;
+		goto pop_next;
+	    }
+	    break;
+	}
+	case TAG_PRIMARY_IMMED1:
+	pop_next:
+	    if (EQUEUE_ISEMPTY(s)) {
+		goto cleanup;
+	    }
+	    obj = EQUEUE_GET(s);
+	    break;
+	default:
+	    erl_exit(ERTS_ABORT_EXIT, "size_shared: bad tag for %#x\n", obj);
+	}
+    }
+
+cleanup:
+    obj = saved_obj;
+    BITSTORE_CLOSE(b);
+    for (;;) {
+	switch (primary_tag(obj)) {
+	case TAG_PRIMARY_LIST: {
+	    Eterm head, tail;
+	    ptr = list_val(obj);
+	    if (!COUNT_OFF_HEAP && erts_is_literal(obj,ptr)) {
+		goto cleanup_next;
+	    }
+	    head = CAR(ptr);
+	    tail = CDR(ptr);
+	    /* if not already clean, clean it up */
+	    if (primary_tag(tail) == TAG_PRIMARY_HEADER) {
+		if (primary_tag(head) == TAG_PRIMARY_HEADER) {
+		    Eterm saved;
+                    BITSTORE_FETCH(b, saved);
+		    CAR(ptr) = head = (head - TAG_PRIMARY_HEADER) | saved;
+		    CDR(ptr) = tail = (tail - TAG_PRIMARY_HEADER) | TAG_PRIMARY_BOXED;
+		} else {
+		    CDR(ptr) = tail = (tail - TAG_PRIMARY_HEADER) | TAG_PRIMARY_LIST;
+		}
+	    } else if (primary_tag(head) == TAG_PRIMARY_HEADER) {
+		CAR(ptr) = head = (head - TAG_PRIMARY_HEADER) | primary_tag(tail);
+		CDR(ptr) = tail = (tail - primary_tag(tail)) | TAG_PRIMARY_IMMED1;
+	    } else {
+		goto cleanup_next;
+	    }
+	    /* and its children too */
+	    if (!IS_CONST(head)) {
+		EQUEUE_PUT_UNCHECKED(s, head);
+	    }
+	    obj = tail;
+	    break;
+	}
+	case TAG_PRIMARY_BOXED: {
+	    Eterm hdr;
+	    ptr = boxed_val(obj);
+	    if (!COUNT_OFF_HEAP && erts_is_literal(obj,ptr)) {
+		goto cleanup_next;
+	    }
+	    hdr = *ptr;
+	    /* if not already clean, clean it up */
+	    if (primary_tag(hdr) == TAG_PRIMARY_HEADER) {
+		goto cleanup_next;
+	    }
+	    else {
+		ASSERT(primary_tag(hdr) == BOXED_VISITED);
+		*ptr = hdr = (hdr - BOXED_VISITED) + TAG_PRIMARY_HEADER;
+	    }
+	    /* and its children too */
+	    switch (hdr & _TAG_HEADER_MASK) {
+	    case ARITYVAL_SUBTAG: {
+		int arity = header_arity(hdr);
+		if (arity == 0) { /* Empty tuple -- unusual. */
+		    goto cleanup_next;
+		}
+		while (arity-- > 0) {
+		    obj = *++ptr;
+		    if (!IS_CONST(obj)) {
+			EQUEUE_PUT_UNCHECKED(s, obj);
+		    }
+		}
+		goto cleanup_next;
+	    }
+	    case FUN_SUBTAG: {
+		ErlFunThing* funp = (ErlFunThing *) ptr;
+		unsigned eterms = 1 /* creator */ + funp->num_free;
+		unsigned sz = thing_arityval(hdr);
+		ptr += 1 /* header */ + sz;
+		while (eterms-- > 0) {
+		    obj = *ptr++;
+		    if (!IS_CONST(obj)) {
+			EQUEUE_PUT_UNCHECKED(s, obj);
+		    }
+		}
+		goto cleanup_next;
+	    }
+            case MAP_SUBTAG:
+                switch (MAP_HEADER_TYPE(hdr)) {
+                    case MAP_HEADER_TAG_FLATMAP_HEAD : {
+                        flatmap_t *mp = (flatmap_t *) ptr;
+                        Uint n = flatmap_get_size(mp) + 1;
+                        ptr += 2; /* hdr + size words */
+                        while (n--) {
+                            obj = *ptr++;
+                            if (!IS_CONST(obj)) {
+                                EQUEUE_PUT_UNCHECKED(s, obj);
+                            }
+                        }
+                        goto cleanup_next;
+                    }
+                    case MAP_HEADER_TAG_HAMT_HEAD_BITMAP :
+                    case MAP_HEADER_TAG_HAMT_HEAD_ARRAY :
+                    case MAP_HEADER_TAG_HAMT_NODE_BITMAP : {
+                        Uint n = hashmap_bitcount(MAP_HEADER_VAL(hdr));
+                        sum += 1 + n + header_arity(hdr);
+                        ptr += 1 + header_arity(hdr);
+                        while (n--) {
+                            obj = *ptr++;
+                            if (!IS_CONST(obj)) {
+                                EQUEUE_PUT_UNCHECKED(s, obj);
+                            }
+                        }
+                        goto cleanup_next;
+                    }
+                    default:
+                        erl_exit(ERTS_ABORT_EXIT, "size_shared: bad hashmap type %d\n", MAP_HEADER_TYPE(hdr));
+                }
+	    default:
+		goto cleanup_next;
+	    }
+	    break;
+	}
+	case TAG_PRIMARY_IMMED1:
+	cleanup_next:
+	    if (EQUEUE_ISEMPTY(s)) {
+		goto all_clean;
+	    }
+	    obj = EQUEUE_GET(s);
+	    break;
+	default:
+	    erl_exit(ERTS_ABORT_EXIT, "size_shared: bad tag for %#x\n", obj);
+	}
+    }
+
+ all_clean:
+    /* Return the result */
+    DESTROY_EQUEUE(s);
+    DESTROY_BITSTORE(b);
+    return sum;
+}
+
+
+/*
  *  Copy a structure to a heap.
  */
-Eterm copy_struct(Eterm obj, Uint sz, Eterm** hpp, ErlOffHeap* off_heap)
+Eterm copy_struct_x(Eterm obj, Uint sz, Eterm** hpp, ErlOffHeap* off_heap, Uint *bsz)
 {
     char* hstart;
     Uint hsize;
@@ -239,19 +614,23 @@ Eterm copy_struct(Eterm obj, Uint sz, Eterm** hpp, ErlOffHeap* off_heap)
     Eterm* argp;
     Eterm* const_tuple;
     Eterm hdr;
+    Eterm *hend;
     int i;
 #ifdef DEBUG
     Eterm org_obj = obj;
     Uint org_sz = sz;
+    Eterm mypid = erts_get_current_pid();
 #endif
 
     if (IS_CONST(obj))
 	return obj;
 
+    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] copy_struct %p\n", mypid, obj));
+
     DTRACE1(copy_struct, (int32_t)sz);
 
     hp = htop = *hpp;
-    hbot   = htop + sz;
+    hbot = hend = htop + sz;
     hstart = (char *)htop;
     hsize = (char*) hbot - hstart;
     const_tuple = 0;
@@ -516,22 +895,870 @@ Eterm copy_struct(Eterm obj, Uint sz, Eterm** hpp, ErlOffHeap* off_heap)
 	}
     }
 
+    if (bsz) {
+        *hpp = htop;
+        *bsz = hend - hbot;
+    } else {
 #ifdef DEBUG
-    if (htop != hbot)
-	erl_exit(ERTS_ABORT_EXIT,
-		 "Internal error in copy_struct() when copying %T:"
-		 " htop=%p != hbot=%p (sz=%beu)\n",
-		 org_obj, htop, hbot, org_sz); 
+        if (htop != hbot)
+            erl_exit(ERTS_ABORT_EXIT,
+                    "Internal error in copy_struct() when copying %T:"
+                    " htop=%p != hbot=%p (sz=%beu)\n",
+                    org_obj, htop, hbot, org_sz);
 #else
-    if (htop > hbot) {
-	erl_exit(ERTS_ABORT_EXIT,
-		 "Internal error in copy_struct(): htop, hbot overrun\n");
-    }
+        if (htop > hbot) {
+            erl_exit(ERTS_ABORT_EXIT,
+                    "Internal error in copy_struct(): htop, hbot overrun\n");
+        }
 #endif
-    *hpp = (Eterm *) (hstart+hsize);
+        *hpp = (Eterm *) (hstart+hsize);
+    }
+    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] result is at %p\n", mypid, res));
     return res;
 }
 
+
+/*
+ *  Machinery for the table used by the sharing preserving copier
+ *  Using an ESTACK but not very transparently; consider refactoring
+ */
+
+#define DECLARE_SHTABLE(s)					\
+    DECLARE_ESTACK(s);						\
+    Uint ESTK_CONCAT(s,_offset) = 0
+#define DESTROY_SHTABLE(s) DESTROY_ESTACK(s)
+#define SHTABLE_INCR 4
+#define SHTABLE_NEXT(s)	ESTK_CONCAT(s,_offset)
+#define SHTABLE_PUSH(s,x,y,b)					\
+do {								\
+    if (s.sp > s.end - SHTABLE_INCR) {				\
+        erl_grow_estack(&(s), SHTABLE_INCR);	       		\
+    }								\
+    *s.sp++ = (x);						\
+    *s.sp++ = (y);						\
+    *s.sp++ = (Eterm) NULL;					\
+    *s.sp++ = (Eterm) (b);		                	\
+    ESTK_CONCAT(s,_offset) += SHTABLE_INCR;			\
+} while(0)
+#define SHTABLE_X(s,e) (s.start[e])
+#define SHTABLE_Y(s,e) (s.start[(e)+1])
+#define SHTABLE_FWD(s,e) ((Eterm *) (s.start[(e)+2]))
+#define SHTABLE_FWD_UPD(s,e,p) (s.start[(e)+2] = (Eterm) (p))
+#define SHTABLE_REV(s,e) ((Eterm *) (s.start[(e)+3]))
+
+#define LIST_SHARED_UNPROCESSED ((Eterm) 0)
+#define LIST_SHARED_PROCESSED	((Eterm) 1)
+
+#define HEAP_ELEM_TO_BE_FILLED	_unchecked_make_list(NULL)
+
+
+/*
+ *  Specialized macros for using/reusing the persistent state
+ */
+
+#define DECLARE_EQUEUE_INIT_INFO(q, info)		\
+    UWord* EQUE_DEF_QUEUE(q) = info->queue_default;	\
+    ErtsEQueue q = {					\
+        EQUE_DEF_QUEUE(q), /* start */			\
+        EQUE_DEF_QUEUE(q), /* front */			\
+        EQUE_DEF_QUEUE(q), /* back */			\
+        1,                 /* possibly_empty */		\
+        EQUE_DEF_QUEUE(q) + DEF_EQUEUE_SIZE, /* end */	\
+        ERTS_ALC_T_ESTACK  /* alloc_type */		\
+    }
+
+#define DECLARE_EQUEUE_FROM_INFO(q, info)		\
+    /* no EQUE_DEF_QUEUE(q), read-only */		\
+    ErtsEQueue q = {					\
+        info->queue_start,      /* start */		\
+        info->queue_start,      /* front */		\
+        info->queue_start,      /* back */		\
+        1,                      /* possibly_empty */	\
+        info->queue_end,        /* end */		\
+        info->queue_alloc_type  /* alloc_type */	\
+    }
+
+#define DECLARE_BITSTORE_INIT_INFO(s, info)		\
+    UWord* WSTK_DEF_STACK(s) = info->bitstore_default;	\
+    ErtsWStack s = {					\
+        WSTK_DEF_STACK(s),  /* wstart */		\
+        WSTK_DEF_STACK(s),  /* wsp */			\
+        WSTK_DEF_STACK(s) + DEF_WSTACK_SIZE, /* wend */	\
+        WSTK_DEF_STACK(s),  /* wdflt */ 		\
+        ERTS_ALC_T_ESTACK /* alloc_type */		\
+    };							\
+    int WSTK_CONCAT(s,_bitoffs) = 0;			\
+    /* no WSTK_CONCAT(s,_offset), write-only */		\
+    UWord WSTK_CONCAT(s,_buffer) = 0
+
+#define DECLARE_BITSTORE_FROM_INFO(s, info)		\
+    /* no WSTK_DEF_STACK(s), read-only */		\
+    ErtsWStack s = {					\
+        info->bitstore_start,  /* wstart */		\
+        NULL,                  /* wsp,  read-only */	\
+        NULL,                  /* wend, read-only */	\
+        NULL,                  /* wdef, read-only */	\
+        info->bitstore_alloc_type /* alloc_type */	\
+    };							\
+    int WSTK_CONCAT(s,_bitoffs) = 0;			\
+    int WSTK_CONCAT(s,_offset) = 0;			\
+    UWord WSTK_CONCAT(s,_buffer) = 0
+
+#define DECLARE_SHTABLE_INIT_INFO(s, info)		\
+    Eterm* ESTK_DEF_STACK(s) = info->shtable_default;	\
+    ErtsEStack s = {					\
+        ESTK_DEF_STACK(s),  /* start */			\
+        ESTK_DEF_STACK(s),  /* sp */			\
+        ESTK_DEF_STACK(s) + DEF_ESTACK_SIZE, /* end */	\
+        ESTK_DEF_STACK(s),  /* default */		\
+        ERTS_ALC_T_ESTACK /* alloc_type */		\
+    };							\
+    Uint ESTK_CONCAT(s,_offset) = 0
+
+#define DECLARE_SHTABLE_FROM_INFO(s, info)		\
+    /* no ESTK_DEF_STACK(s), read-only */		\
+    ErtsEStack s = {					\
+        info->shtable_start,     /* start */		\
+        NULL,                    /* sp,  read-only */	\
+        NULL,                    /* end, read-only */	\
+        NULL,                    /* def, read-only */	\
+        info->shtable_alloc_type /* alloc_type */	\
+    };							\
+    /* no ESTK_CONCAT(s,_offset), read-only */
+
+/*
+ *  Copy object "obj" preserving sharing.
+ *  First half: count size and calculate sharing.
+ */
+Uint copy_shared_calculate(Eterm obj, erts_shcopy_t *info)
+{
+    Uint sum;
+    Uint e;
+    unsigned sz;
+    Eterm* ptr;
+#ifdef DEBUG
+    Eterm mypid = erts_get_current_pid();
+#endif
+
+    DECLARE_EQUEUE_INIT_INFO(s, info);
+    DECLARE_BITSTORE_INIT_INFO(b, info);
+    DECLARE_SHTABLE_INIT_INFO(t, info);
+
+    /* step #0:
+       -------------------------------------------------------
+       get rid of the easy cases first:
+       - copying constants
+       - if not a proper process, do flat copy
+    */
+
+    if (IS_CONST(obj))
+	return 0;
+
+    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] copy_shared_calculate %p\n", mypid, obj));
+    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] message is %T\n", mypid, obj));
+
+    /* step #1:
+       -------------------------------------------------------
+       traverse the term and calculate the size;
+       when traversing, transform as you do in size_shared
+       but when you find shared objects:
+
+       a. add entry in the table, indexed by i
+       b. mark them:
+	  b1. boxed terms, set header to (i | 11)
+	      store (old header, NONV, NULL, backptr) in the entry
+	  b2. cons cells, set CDR to NONV, set CAR to i
+	      store (old CAR, old CDR, NULL, backptr) in the entry
+    */
+
+    sum = 0;
+
+    for (;;) {
+	switch (primary_tag(obj)) {
+	case TAG_PRIMARY_LIST: {
+	    Eterm head, tail;
+	    ptr = list_val(obj);
+	    /* off heap list pointers are copied verbatim */
+	    if (erts_is_literal(obj,ptr)) {
+		VERBOSE(DEBUG_SHCOPY, ("[pid=%T] bypassed copying %p is %T\n", mypid, ptr, obj));
+                if (IN_LITERAL_PURGE_AREA(info,ptr))
+                    info->literal_size += size_object(obj);
+		goto pop_next;
+	    }
+	    head = CAR(ptr);
+	    tail = CDR(ptr);
+	    /* if it's visited, don't count it;
+	       if not already shared, make it shared and store it in the table */
+	    if (primary_tag(tail) == TAG_PRIMARY_HEADER ||
+		primary_tag(head) == TAG_PRIMARY_HEADER) {
+		if (tail != THE_NON_VALUE) {
+		    e = SHTABLE_NEXT(t);
+		    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] tabling L %p\n", mypid, ptr));
+		    SHTABLE_PUSH(t, head, tail, ptr);
+		    CAR(ptr) = (e << _TAG_PRIMARY_SIZE) | LIST_SHARED_UNPROCESSED;
+		    CDR(ptr) = THE_NON_VALUE;
+		}
+		goto pop_next;
+	    }
+	    /* else make it visited now */
+	    switch (primary_tag(tail)) {
+	    case TAG_PRIMARY_LIST:
+		VERBOSE(DEBUG_SHCOPY, ("[pid=%T] mangling L/L %p\n", mypid, ptr));
+		CDR(ptr) = (tail - TAG_PRIMARY_LIST) | TAG_PRIMARY_HEADER;
+		break;
+	    case TAG_PRIMARY_IMMED1:
+		VERBOSE(DEBUG_SHCOPY, ("[pid=%T] mangling L/I %p\n", mypid, ptr));
+		CAR(ptr) = (head - primary_tag(head)) | TAG_PRIMARY_HEADER;
+		CDR(ptr) = (tail - TAG_PRIMARY_IMMED1) | primary_tag(head);
+		break;
+	    case TAG_PRIMARY_BOXED:
+		BITSTORE_PUT(b, primary_tag(head));
+		VERBOSE(DEBUG_SHCOPY, ("[pid=%T] mangling L/B %p\n", mypid, ptr));
+		CAR(ptr) = (head - primary_tag(head)) | TAG_PRIMARY_HEADER;
+		CDR(ptr) = (tail - TAG_PRIMARY_BOXED) | TAG_PRIMARY_HEADER;
+		break;
+	    }
+	    /* and count it */
+	    sum += 2;
+	    if (!IS_CONST(head)) {
+		EQUEUE_PUT(s, head);
+	    }
+	    obj = tail;
+	    break;
+	}
+	case TAG_PRIMARY_BOXED: {
+	    Eterm hdr;
+	    ptr = boxed_val(obj);
+	    /* off heap pointers to boxes are copied verbatim */
+	    if (erts_is_literal(obj,ptr)) {
+		VERBOSE(DEBUG_SHCOPY, ("[pid=%T] bypassed copying %p is %T\n", mypid, ptr, obj));
+                if (IN_LITERAL_PURGE_AREA(info,ptr))
+                    info->literal_size += size_object(obj);
+		goto pop_next;
+	    }
+	    hdr = *ptr;
+	    /* if it's visited, don't count it;
+	       if not already shared, make it shared and store it in the table */
+	    if (primary_tag(hdr) != TAG_PRIMARY_HEADER) {
+		if (primary_tag(hdr) == BOXED_VISITED) {
+		    e = SHTABLE_NEXT(t);
+		    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] tabling B %p\n", mypid, ptr));
+		    SHTABLE_PUSH(t, hdr, THE_NON_VALUE, ptr);
+		    *ptr = (e << _TAG_PRIMARY_SIZE) | BOXED_SHARED_UNPROCESSED;
+		}
+		goto pop_next;
+	    }
+	    /* else make it visited now */
+	    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] mangling B %p\n", mypid, ptr));
+	    *ptr = (hdr - primary_tag(hdr)) + BOXED_VISITED;
+	    /* and count it */
+	    ASSERT(is_header(hdr));
+	    switch (hdr & _TAG_HEADER_MASK) {
+	    case ARITYVAL_SUBTAG: {
+		int arity = header_arity(hdr);
+		sum += arity + 1;
+		if (arity == 0) { /* Empty tuple -- unusual. */
+		    goto pop_next;
+		}
+		while (arity-- > 0) {
+		    obj = *++ptr;
+		    if (!IS_CONST(obj)) {
+			EQUEUE_PUT(s, obj);
+		    }
+		}
+		goto pop_next;
+	    }
+	    case FUN_SUBTAG: {
+		ErlFunThing* funp = (ErlFunThing *) ptr;
+		unsigned eterms = 1 /* creator */ + funp->num_free;
+		sz = thing_arityval(hdr);
+		sum += 1 /* header */ + sz + eterms;
+		ptr += 1 /* header */ + sz;
+		while (eterms-- > 0) {
+		    obj = *ptr++;
+		    if (!IS_CONST(obj)) {
+			EQUEUE_PUT(s, obj);
+		    }
+		}
+		goto pop_next;
+	    }
+	    case SUB_BINARY_SUBTAG: {
+		ErlSubBin* sb = (ErlSubBin *) ptr;
+		Eterm real_bin = sb->orig;
+		Uint bit_offset = sb->bitoffs;
+		Uint bit_size = sb->bitsize;
+		size_t size = sb->size;
+		Uint extra_bytes;
+		Eterm hdr;
+		if (bit_size + bit_offset > 8) {
+		    sum += ERL_SUB_BIN_SIZE;
+		    extra_bytes = 2;
+		} else if (bit_size + bit_offset > 0) {
+		    sum += ERL_SUB_BIN_SIZE;
+		    extra_bytes = 1;
+		} else {
+		    extra_bytes = 0;
+		}
+		ASSERT(is_boxed(real_bin) &&
+		       (((*boxed_val(real_bin)) &
+			 (_TAG_HEADER_MASK - _BINARY_XXX_MASK - BOXED_VISITED_MASK))
+			== _TAG_HEADER_REFC_BIN));
+		hdr = *_unchecked_binary_val(real_bin) & ~BOXED_VISITED_MASK;
+		if (thing_subtag(hdr) == HEAP_BINARY_SUBTAG) {
+		    sum += heap_bin_size(size+extra_bytes);
+		} else {
+		    ASSERT(thing_subtag(hdr) == REFC_BINARY_SUBTAG);
+		    sum += PROC_BIN_SIZE;
+		}
+		goto pop_next;
+	    }
+            case MAP_SUBTAG:
+                switch (MAP_HEADER_TYPE(hdr)) {
+                    case MAP_HEADER_TAG_FLATMAP_HEAD : {
+                        flatmap_t *mp = (flatmap_t *) ptr;
+                        Uint n = flatmap_get_size(mp) + 1;
+                        sum += n + 2;
+                        ptr += 2; /* hdr + size words */
+                        while (n--) {
+                            obj = *ptr++;
+                            if (!IS_CONST(obj)) {
+                                EQUEUE_PUT(s, obj);
+                            }
+                        }
+                        goto pop_next;
+                    }
+                    case MAP_HEADER_TAG_HAMT_HEAD_BITMAP :
+                    case MAP_HEADER_TAG_HAMT_HEAD_ARRAY :
+                    case MAP_HEADER_TAG_HAMT_NODE_BITMAP : {
+                        Uint n = hashmap_bitcount(MAP_HEADER_VAL(hdr));
+                        sum += 1 + n + header_arity(hdr);
+                        ptr += 1 + header_arity(hdr);
+
+                        if (n == 0) {
+                            goto pop_next;
+                        }
+                        while(n--) {
+                            obj = *ptr++;
+                            if (!IS_CONST(obj)) {
+                                EQUEUE_PUT(s, obj);
+                            }
+                        }
+                        goto pop_next;
+                    }
+                    default:
+                        erl_exit(ERTS_ABORT_EXIT, "copy_shared_calculate: bad hashmap type %d\n", MAP_HEADER_TYPE(hdr));
+                }
+            case BIN_MATCHSTATE_SUBTAG:
+		erl_exit(ERTS_ABORT_EXIT,
+			 "size_shared: matchstate term not allowed");
+	    default:
+		sum += thing_arityval(hdr) + 1;
+		goto pop_next;
+	    }
+	    break;
+	}
+	case TAG_PRIMARY_IMMED1:
+	pop_next:
+	    if (EQUEUE_ISEMPTY(s)) {
+                /* add sentinel to the table */
+                SHTABLE_PUSH(t, THE_NON_VALUE, THE_NON_VALUE, NULL);
+                /* store persistent info */
+                BITSTORE_CLOSE(b);
+                info->queue_start = s.start;
+                info->queue_end = s.end;
+                info->queue_alloc_type = s.alloc_type;
+                info->bitstore_start = b.wstart;
+                info->bitstore_alloc_type = b.alloc_type;
+                info->shtable_start = t.start;
+                info->shtable_alloc_type = t.alloc_type;
+                /* single point of return: the size of the object */
+                VERBOSE(DEBUG_SHCOPY, ("[pid=%T] size was: %u\n", mypid, sum));
+                return sum + info->literal_size;
+	    }
+	    obj = EQUEUE_GET(s);
+	    break;
+	default:
+	    erl_exit(ERTS_ABORT_EXIT, "[pid=%T] size_shared: bad tag for %#x\n", obj);
+	}
+    }
+}
+
+/*
+ *  Copy object "obj" preserving sharing.
+ *  Second half: copy and restore the object.
+ */
+Uint copy_shared_perform(Eterm obj, Uint size, erts_shcopy_t *info,
+                         Eterm** hpp, ErlOffHeap* off_heap) {
+    Uint e;
+    unsigned sz;
+    Eterm* ptr;
+    Eterm* hp;
+    Eterm* hscan;
+    Eterm result;
+    Eterm* resp;
+    Eterm *hbot, *hend;
+    unsigned remaining;
+#ifdef DEBUG
+    Eterm mypid = erts_get_current_pid();
+    Eterm saved_obj = obj;
+#endif
+
+    DECLARE_EQUEUE_FROM_INFO(s, info);
+    DECLARE_BITSTORE_FROM_INFO(b, info);
+    DECLARE_SHTABLE_FROM_INFO(t, info);
+
+    /* step #0:
+       -------------------------------------------------------
+       get rid of the easy cases first:
+       - copying constants
+       - if not a proper process, do flat copy
+    */
+
+    if (IS_CONST(obj))
+	return obj;
+
+    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] copy_shared_perform %p\n", mypid, obj));
+
+    /* step #2: was performed before this function was called
+       -------------------------------------------------------
+       allocate new space
+    */
+
+    hscan = hp = *hpp;
+    hbot  = hend = hp + size;
+
+    /* step #3:
+       -------------------------------------------------------
+       traverse the term a second time and when traversing:
+       a. if the object is marked as shared
+	  a1. if the entry contains a forwarding ptr, use that
+	  a2. otherwise, copy it to the new space and store the
+	      forwarding ptr to the entry
+      b. otherwise, reverse-transform as you do in size_shared
+	 and copy to the new space
+    */
+
+    resp = &result;
+    remaining = 0;
+    for (;;) {
+	switch (primary_tag(obj)) {
+	case TAG_PRIMARY_LIST: {
+	    Eterm head, tail;
+	    ptr = list_val(obj);
+	    /* off heap list pointers are copied verbatim */
+	    if (erts_is_literal(obj,ptr)) {
+                if (!IN_LITERAL_PURGE_AREA(info,ptr)) {
+                    *resp = obj;
+                } else {
+                    Uint bsz = 0;
+                    *resp = copy_struct_x(obj, hbot - hp, &hp, off_heap, &bsz);
+                    hbot -= bsz;
+                }
+		goto cleanup_next;
+	    }
+	    head = CAR(ptr);
+	    tail = CDR(ptr);
+	    /* if it is shared */
+	    if (tail == THE_NON_VALUE) {
+		e = head >> _TAG_PRIMARY_SIZE;
+		/* if it has been processed, just use the forwarding pointer */
+		if (primary_tag(head) == LIST_SHARED_PROCESSED) {
+		    *resp = make_list(SHTABLE_FWD(t, e));
+		    goto cleanup_next;
+		}
+		/* else, let's process it now,
+		   copy it and keep the forwarding pointer */
+		else {
+		    CAR(ptr) = (head - primary_tag(head)) + LIST_SHARED_PROCESSED;
+		    head = SHTABLE_X(t, e);
+		    tail = SHTABLE_Y(t, e);
+		    ptr = &(SHTABLE_X(t, e));
+		    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] tabled L %p is %p\n", mypid, ptr, SHTABLE_REV(t, e)));
+		    SHTABLE_FWD_UPD(t, e, hp);
+		}
+	    }
+	    /* if not already clean, clean it up and copy it */
+	    if (primary_tag(tail) == TAG_PRIMARY_HEADER) {
+		if (primary_tag(head) == TAG_PRIMARY_HEADER) {
+		    Eterm saved;
+                    BITSTORE_FETCH(b, saved);
+		    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] unmangling L/B %p\n", mypid, ptr));
+		    CAR(ptr) = head = (head - TAG_PRIMARY_HEADER) + saved;
+		    CDR(ptr) = tail = (tail - TAG_PRIMARY_HEADER) + TAG_PRIMARY_BOXED;
+		} else {
+		    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] unmangling L/L %p\n", mypid, ptr));
+		    CDR(ptr) = tail = (tail - TAG_PRIMARY_HEADER) + TAG_PRIMARY_LIST;
+		}
+	    } else if (primary_tag(head) == TAG_PRIMARY_HEADER) {
+		VERBOSE(DEBUG_SHCOPY, ("[pid=%T] unmangling L/I %p\n", mypid, ptr));
+		CAR(ptr) = head = (head - TAG_PRIMARY_HEADER) | primary_tag(tail);
+		CDR(ptr) = tail = (tail - primary_tag(tail)) | TAG_PRIMARY_IMMED1;
+	    } else {
+		ASSERT(0 && "cannot come here");
+		goto cleanup_next;
+	    }
+	    /* and its children too */
+	    if (IS_CONST(head)) {
+		CAR(hp) = head;
+	    } else {
+		EQUEUE_PUT_UNCHECKED(s, head);
+		CAR(hp) = HEAP_ELEM_TO_BE_FILLED;
+	    }
+	    *resp = make_list(hp);
+	    resp = &(CDR(hp));
+	    hp += 2;
+	    obj = tail;
+	    break;
+	}
+	case TAG_PRIMARY_BOXED: {
+	    Eterm hdr;
+	    ptr = boxed_val(obj);
+	    /* off heap pointers to boxes are copied verbatim */
+	    if (erts_is_literal(obj,ptr)) {
+                if (!IN_LITERAL_PURGE_AREA(info,ptr)) {
+                    *resp = obj;
+                } else {
+                    Uint bsz = 0;
+                    *resp = copy_struct_x(obj, hbot - hp, &hp, off_heap, &bsz);
+                    hbot -= bsz;
+                }
+		goto cleanup_next;
+	    }
+	    hdr = *ptr;
+	    /* clean it up, unless it's already clean or shared and processed */
+	    switch (primary_tag(hdr)) {
+	    case TAG_PRIMARY_HEADER:
+		ASSERT(0 && "cannot come here");
+	    /* if it is shared and has been processed,
+	       just use the forwarding pointer */
+	    case BOXED_SHARED_PROCESSED:
+		e = hdr >> _TAG_PRIMARY_SIZE;
+		*resp = make_boxed(SHTABLE_FWD(t, e));
+		goto cleanup_next;
+	    /* if it is shared but has not been processed yet, let's process
+	       it now: copy it and keep the forwarding pointer */
+	    case BOXED_SHARED_UNPROCESSED:
+		e = hdr >> _TAG_PRIMARY_SIZE;
+		*ptr = (hdr - primary_tag(hdr)) + BOXED_SHARED_PROCESSED;
+		hdr = SHTABLE_X(t, e);
+		ASSERT(primary_tag(hdr) == BOXED_VISITED);
+		VERBOSE(DEBUG_SHCOPY, ("[pid=%T] tabled B %p is %p\n", mypid, ptr, SHTABLE_REV(t, e)));
+		VERBOSE(DEBUG_SHCOPY, ("[pid=%T] unmangling B %p\n", mypid, ptr));
+		SHTABLE_X(t, e) = hdr = (hdr - BOXED_VISITED) + TAG_PRIMARY_HEADER;
+		SHTABLE_FWD_UPD(t, e, hp);
+		break;
+	    case BOXED_VISITED:
+		VERBOSE(DEBUG_SHCOPY, ("[pid=%T] unmangling B %p\n", mypid, ptr));
+		*ptr = hdr = (hdr - BOXED_VISITED) + TAG_PRIMARY_HEADER;
+		break;
+	    }
+	    /* and its children too */
+	    switch (hdr & _TAG_HEADER_MASK) {
+	    case ARITYVAL_SUBTAG: {
+		int arity = header_arity(hdr);
+		*resp = make_boxed(hp);
+		*hp++ = hdr;
+		while (arity-- > 0) {
+		    obj = *++ptr;
+		    if (IS_CONST(obj)) {
+			*hp++ = obj;
+		    } else {
+			EQUEUE_PUT_UNCHECKED(s, obj);
+			*hp++ = HEAP_ELEM_TO_BE_FILLED;
+		    }
+		}
+		goto cleanup_next;
+	    }
+	    case FUN_SUBTAG: {
+		ErlFunThing* funp = (ErlFunThing *) ptr;
+		unsigned eterms = 1 /* creator */ + funp->num_free;
+		sz = thing_arityval(hdr);
+		funp = (ErlFunThing *) hp;
+		*resp = make_fun(hp);
+		*hp++ = hdr;
+		ptr++;
+		while (sz-- > 0) {
+		    *hp++ = *ptr++;
+		}
+		while (eterms-- > 0) {
+		    obj = *ptr++;
+		    if (IS_CONST(obj)) {
+			*hp++ = obj;
+		    } else {
+			EQUEUE_PUT_UNCHECKED(s, obj);
+			*hp++ = HEAP_ELEM_TO_BE_FILLED;
+		    }
+		}
+		funp->next = off_heap->first;
+		off_heap->first = (struct erl_off_heap_header*) funp;
+		erts_refc_inc(&funp->fe->refc, 2);
+		goto cleanup_next;
+	    }
+	    case MAP_SUBTAG:
+                *resp  = make_flatmap(hp);
+                *hp++  = hdr;
+                switch (MAP_HEADER_TYPE(hdr)) {
+                    case MAP_HEADER_TAG_FLATMAP_HEAD : {
+                        flatmap_t *mp = (flatmap_t *) ptr;
+                        Uint n = flatmap_get_size(mp) + 1;
+                        *hp++  = *++ptr; /* keys */
+                        while (n--) {
+                            obj = *++ptr;
+                            if (IS_CONST(obj)) {
+                                *hp++ = obj;
+                            } else {
+                                EQUEUE_PUT_UNCHECKED(s, obj);
+                                *hp++ = HEAP_ELEM_TO_BE_FILLED;
+                            }
+                        }
+                        goto cleanup_next;
+                    }
+                    case MAP_HEADER_TAG_HAMT_HEAD_BITMAP :
+                    case MAP_HEADER_TAG_HAMT_HEAD_ARRAY :
+			*hp++ = *++ptr; /* total map size */
+                    case MAP_HEADER_TAG_HAMT_NODE_BITMAP : {
+                         Uint n = hashmap_bitcount(MAP_HEADER_VAL(hdr));
+                         while (n--)  {
+                             obj = *++ptr;
+                             if (IS_CONST(obj)) {
+                                 *hp++ = obj;
+                             } else {
+                                 EQUEUE_PUT_UNCHECKED(s, obj);
+                                 *hp++ = HEAP_ELEM_TO_BE_FILLED;
+                             }
+                         }
+                        goto cleanup_next;
+                    }
+                    default:
+                        erl_exit(ERTS_ABORT_EXIT, "copy_shared_perform: bad hashmap type %d\n", MAP_HEADER_TYPE(hdr));
+                }
+	    case REFC_BINARY_SUBTAG: {
+		ProcBin* pb = (ProcBin *) ptr;
+		sz = thing_arityval(hdr);
+		if (pb->flags) {
+		    erts_emasculate_writable_binary(pb);
+		}
+		pb = (ProcBin *) hp;
+		*resp = make_binary(hp);
+		*hp++ = hdr;
+		ptr++;
+		while (sz-- > 0) {
+		    *hp++ = *ptr++;
+		}
+		erts_refc_inc(&pb->val->refc, 2);
+		pb->next = off_heap->first;
+		pb->flags = 0;
+		off_heap->first = (struct erl_off_heap_header*) pb;
+		OH_OVERHEAD(off_heap, pb->size / sizeof(Eterm));
+		goto cleanup_next;
+	    }
+	    case SUB_BINARY_SUBTAG: {
+		ErlSubBin* sb = (ErlSubBin *) ptr;
+		Eterm real_bin = sb->orig;
+		Uint bit_offset = sb->bitoffs;
+		Uint bit_size = sb->bitsize;
+		Uint offset = sb->offs;
+		size_t size = sb->size;
+		Uint extra_bytes;
+		Uint real_size;
+		if ((bit_size + bit_offset) > 8) {
+		    extra_bytes = 2;
+		} else if ((bit_size + bit_offset) > 0) {
+		    extra_bytes = 1;
+		} else {
+		    extra_bytes = 0;
+		}
+		real_size = size+extra_bytes;
+		ASSERT(is_boxed(real_bin) &&
+		       (((*boxed_val(real_bin)) &
+			 (_TAG_HEADER_MASK - _BINARY_XXX_MASK - BOXED_VISITED_MASK))
+			== _TAG_HEADER_REFC_BIN));
+		ptr = _unchecked_binary_val(real_bin);
+		*resp = make_binary(hp);
+		if (extra_bytes != 0) {
+		    ErlSubBin* res = (ErlSubBin *) hp;
+		    hp += ERL_SUB_BIN_SIZE;
+		    res->thing_word = HEADER_SUB_BIN;
+		    res->size = size;
+		    res->bitsize = bit_size;
+		    res->bitoffs = bit_offset;
+		    res->offs = 0;
+		    res->is_writable = 0;
+		    res->orig = make_binary(hp);
+		}
+		if (thing_subtag(*ptr & ~BOXED_VISITED_MASK) == HEAP_BINARY_SUBTAG) {
+		    ErlHeapBin* from = (ErlHeapBin *) ptr;
+		    ErlHeapBin* to = (ErlHeapBin *) hp;
+		    hp += heap_bin_size(real_size);
+		    to->thing_word = header_heap_bin(real_size);
+		    to->size = real_size;
+		    sys_memcpy(to->data, ((byte *)from->data)+offset, real_size);
+		} else {
+		    ProcBin* from = (ProcBin *) ptr;
+		    ProcBin* to = (ProcBin *) hp;
+		    ASSERT(thing_subtag(*ptr & ~BOXED_VISITED_MASK) == REFC_BINARY_SUBTAG);
+		    if (from->flags) {
+			erts_emasculate_writable_binary(from);
+		    }
+		    hp += PROC_BIN_SIZE;
+		    to->thing_word = HEADER_PROC_BIN;
+		    to->size = real_size;
+		    to->val = from->val;
+		    erts_refc_inc(&to->val->refc, 2);
+		    to->bytes = from->bytes + offset;
+		    to->next = off_heap->first;
+		    to->flags = 0;
+		    off_heap->first = (struct erl_off_heap_header*) to;
+		    OH_OVERHEAD(off_heap, to->size / sizeof(Eterm));
+		}
+		goto cleanup_next;
+	    }
+	    case EXTERNAL_PID_SUBTAG:
+	    case EXTERNAL_PORT_SUBTAG:
+	    case EXTERNAL_REF_SUBTAG: {
+		ExternalThing *etp = (ExternalThing *) hp;
+		sz = thing_arityval(hdr);
+		*resp = make_external(hp);
+		*hp++ = hdr;
+		ptr++;
+		while (sz-- > 0) {
+		    *hp++ = *ptr++;
+		}
+		etp->next = off_heap->first;
+		off_heap->first = (struct erl_off_heap_header*) etp;
+		erts_refc_inc(&etp->node->refc, 2);
+		goto cleanup_next;
+	    }
+	    default:
+		sz = thing_arityval(hdr);
+		*resp = make_boxed(hp);
+		*hp++ = hdr;
+		ptr++;
+		while (sz-- > 0) {
+		    *hp++ = *ptr++;
+		}
+		goto cleanup_next;
+	    }
+	    break;
+	}
+	case TAG_PRIMARY_IMMED1:
+	    *resp = obj;
+	cleanup_next:
+	    if (EQUEUE_ISEMPTY(s)) {
+		goto all_clean;
+	    }
+	    obj = EQUEUE_GET(s);
+	    for (;;) {
+		ASSERT(hscan < hp);
+		if (remaining == 0) {
+		    if (*hscan == HEAP_ELEM_TO_BE_FILLED) {
+			resp = hscan;
+			hscan += 2;
+			break; /* scanning loop */
+		    } else if (primary_tag(*hscan) == TAG_PRIMARY_HEADER) {
+			switch (*hscan & _TAG_HEADER_MASK) {
+			case ARITYVAL_SUBTAG:
+			    remaining = header_arity(*hscan);
+			    hscan++;
+			    break;
+			case FUN_SUBTAG: {
+			    ErlFunThing* funp = (ErlFunThing *) hscan;
+			    hscan += 1 + thing_arityval(*hscan);
+			    remaining = 1 + funp->num_free;
+			    break;
+			}
+			case MAP_SUBTAG:
+                            switch (MAP_HEADER_TYPE(*hscan)) {
+                                case MAP_HEADER_TAG_FLATMAP_HEAD : {
+                                    flatmap_t *mp = (flatmap_t *) hscan;
+                                    remaining = flatmap_get_size(mp) + 1;
+                                    hscan += 2;
+                                    break;
+                                }
+                                case MAP_HEADER_TAG_HAMT_HEAD_BITMAP :
+                                case MAP_HEADER_TAG_HAMT_HEAD_ARRAY :
+                                case MAP_HEADER_TAG_HAMT_NODE_BITMAP :
+                                    remaining = hashmap_bitcount(MAP_HEADER_VAL(*hscan));
+                                    hscan += MAP_HEADER_ARITY(*hscan) + 1;
+                                    break;
+                                default:
+                                    erl_exit(ERTS_ABORT_EXIT,
+                                            "copy_shared_perform: bad hashmap type %d\n",
+                                            MAP_HEADER_TYPE(*hscan));
+                            }
+                            break;
+			case SUB_BINARY_SUBTAG:
+			    ASSERT(((ErlSubBin *) hscan)->bitoffs +
+				   ((ErlSubBin *) hscan)->bitsize > 0);
+			    hscan += ERL_SUB_BIN_SIZE;
+			    break;
+			default:
+			    hscan += 1 + thing_arityval(*hscan);
+			    break;
+			}
+		    } else {
+			hscan++;
+		    }
+		} else if (*hscan == HEAP_ELEM_TO_BE_FILLED) {
+		    resp = hscan++;
+		    remaining--;
+		    break; /* scanning loop */
+		} else {
+		    hscan++;
+		    remaining--;
+		}
+	    }
+	    ASSERT(resp < hp);
+	    break;
+	default:
+	    erl_exit(ERTS_ABORT_EXIT, "size_shared: bad tag for %#x\n", obj);
+	}
+    }
+
+    /* step #4:
+       -------------------------------------------------------
+       traverse the table and reverse-transform all stored entries
+    */
+
+all_clean:
+    for (e = 0; ; e += SHTABLE_INCR) {
+	ptr = SHTABLE_REV(t, e);
+	if (ptr == NULL)
+	    break;
+	VERBOSE(DEBUG_SHCOPY, ("[copy] restoring shared: %x\n", ptr));
+	/* entry was a list */
+	if (SHTABLE_Y(t, e) != THE_NON_VALUE) {
+	    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] untabling L %p\n", mypid, ptr));
+	    CAR(ptr) = SHTABLE_X(t, e);
+	    CDR(ptr) = SHTABLE_Y(t, e);
+	}
+	/* entry was boxed */
+	else {
+	    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] untabling B %p\n", mypid, ptr));
+	    *ptr = SHTABLE_X(t, e);
+	    ASSERT(primary_tag(*ptr) == TAG_PRIMARY_HEADER);
+	}
+    }
+
+#ifdef DEBUG
+    if (eq(saved_obj, result) == 0) {
+	erts_fprintf(stderr, "original = %T\n", saved_obj);
+	erts_fprintf(stderr, "copy = %T\n", result);
+	erl_exit(ERTS_ABORT_EXIT, "copy (shared) not equal to source\n");
+    }
+#endif
+
+    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] original was %T\n", mypid, saved_obj));
+    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] copy is %T\n", mypid, result));
+    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] result is at %p\n", mypid, result));
+
+    ASSERT(hbot == hp);
+    ASSERT(size == ((hp - *hpp) + (hend - hbot)));
+    *hpp = hend;
+    return result;
+}
+
+
 /*
  * Copy a term that is guaranteed to be contained in a single
  * heap block. The heap block is copied word by word, and any
diff --git a/erts/emulator/beam/dist.c b/erts/emulator/beam/dist.c
index 68745fc448..f480915256 100644
--- a/erts/emulator/beam/dist.c
+++ b/erts/emulator/beam/dist.c
@@ -881,11 +881,7 @@ erts_dsig_send_msg(Eterm remote, Eterm message, ErtsSendContext* ctx)
     DTRACE_CHARBUF(receiver_name, 64);
 #endif
 
-    if (SEQ_TRACE_TOKEN(sender) != NIL
-#ifdef USE_VM_PROBES
-	&& SEQ_TRACE_TOKEN(sender) != am_have_dt_utag 
-#endif
-	) {
+    if (have_seqtrace(SEQ_TRACE_TOKEN(sender))) {
 	seq_trace_update_send(sender);
 	token = SEQ_TRACE_TOKEN(sender);
 	seq_trace_output(token, message, SEQ_TRACE_SEND, remote, sender);
@@ -900,7 +896,7 @@ erts_dsig_send_msg(Eterm remote, Eterm message, ErtsSendContext* ctx)
         erts_snprintf(receiver_name, sizeof(DTRACE_CHARBUF_NAME(receiver_name)),
                       "%T", remote);
         msize = size_object(message);
-        if (token != NIL && token != am_have_dt_utag) {
+        if (have_seqtrace(token)) {
             tok_label = signed_val(SEQ_TRACE_T_LABEL(token));
             tok_lastcnt = signed_val(SEQ_TRACE_T_LASTCNT(token));
             tok_serial = signed_val(SEQ_TRACE_T_SERIAL(token));
@@ -942,11 +938,7 @@ erts_dsig_send_reg_msg(Eterm remote_name, Eterm message,
     DTRACE_CHARBUF(receiver_name, 128);
 #endif
 
-    if (SEQ_TRACE_TOKEN(sender) != NIL
-#ifdef USE_VM_PROBES
-	&& SEQ_TRACE_TOKEN(sender) != am_have_dt_utag 
-#endif
-	) {
+    if (have_seqtrace(SEQ_TRACE_TOKEN(sender))) {
 	seq_trace_update_send(sender);
 	token = SEQ_TRACE_TOKEN(sender);
 	seq_trace_output(token, message, SEQ_TRACE_SEND, remote_name, sender);
@@ -961,7 +953,7 @@ erts_dsig_send_reg_msg(Eterm remote_name, Eterm message,
         erts_snprintf(receiver_name, sizeof(DTRACE_CHARBUF_NAME(receiver_name)),
                       "{%T,%s}", remote_name, node_name);
         msize = size_object(message);
-        if (token != NIL && token != am_have_dt_utag) {
+        if (have_seqtrace(token)) {
             tok_label = signed_val(SEQ_TRACE_T_LABEL(token));
             tok_lastcnt = signed_val(SEQ_TRACE_T_LASTCNT(token));
             tok_serial = signed_val(SEQ_TRACE_T_SERIAL(token));
@@ -1006,11 +998,7 @@ erts_dsig_send_exit_tt(ErtsDSigData *dsdp, Eterm local, Eterm remote,
 #endif
 
     UseTmpHeapNoproc(6);
-    if (token != NIL 
-#ifdef USE_VM_PROBES
-	&& token != am_have_dt_utag
-#endif
-	) {	
+    if (have_seqtrace(token)) {
 	seq_trace_update_send(dsdp->proc);
 	seq_trace_output_exit(token, reason, SEQ_TRACE_SEND, remote, local);
 	ctl = TUPLE5(&ctl_heap[0],
@@ -1029,7 +1017,7 @@ erts_dsig_send_exit_tt(ErtsDSigData *dsdp, Eterm local, Eterm remote,
                       "{%T,%s}", remote, node_name);
         erts_snprintf(reason_str, sizeof(DTRACE_CHARBUF_NAME(reason_str)),
                       "%T", reason);
-        if (token != NIL && token != am_have_dt_utag) {
+        if (have_seqtrace(token)) {
             tok_label = signed_val(SEQ_TRACE_T_LABEL(token));
             tok_lastcnt = signed_val(SEQ_TRACE_T_LASTCNT(token));
             tok_serial = signed_val(SEQ_TRACE_T_SERIAL(token));
diff --git a/erts/emulator/beam/erl_bif_info.c b/erts/emulator/beam/erl_bif_info.c
index 1eb106a551..82c2aa4b9e 100644
--- a/erts/emulator/beam/erl_bif_info.c
+++ b/erts/emulator/beam/erl_bif_info.c
@@ -130,6 +130,9 @@ static char erts_system_version[] = ("Erlang/OTP " ERLANG_OTP_RELEASE
 #ifdef USE_SYSTEMTAP
 				     " [systemtap]"
 #endif
+#ifdef SHCOPY
+				     " [sharing-preserving]"
+#endif
 				     "\n");
 
 #define ASIZE(a) (sizeof(a)/sizeof(a[0]))
diff --git a/erts/emulator/beam/erl_bif_trace.c b/erts/emulator/beam/erl_bif_trace.c
index 03f51132b1..4d67e39e7e 100644
--- a/erts/emulator/beam/erl_bif_trace.c
+++ b/erts/emulator/beam/erl_bif_trace.c
@@ -1879,11 +1879,7 @@ new_seq_trace_token(Process* p)
 {
     Eterm* hp;
 
-    if (SEQ_TRACE_TOKEN(p) == NIL
-#ifdef USE_VM_PROBES
-	|| SEQ_TRACE_TOKEN(p) == am_have_dt_utag
-#endif
-	) {
+    if (have_no_seqtrace(SEQ_TRACE_TOKEN(p))) {
 	hp = HAlloc(p, 6);
 	SEQ_TRACE_TOKEN(p) = TUPLE5(hp, make_small(0),		/* Flags  */ 
 				    make_small(0),		/* Label  */
@@ -1903,12 +1899,8 @@ BIF_RETTYPE erl_seq_trace_info(Process *p, Eterm item)
 	BIF_ERROR(p, BADARG);
     }
 
-    if (SEQ_TRACE_TOKEN(p) == NIL
-#ifdef USE_VM_PROBES
-	|| SEQ_TRACE_TOKEN(p) == am_have_dt_utag
-#endif
-	) {
-	if ((item == am_send) || (item == am_receive) || 
+    if (have_no_seqtrace(SEQ_TRACE_TOKEN(p))) {
+	if ((item == am_send)  || (item == am_receive) || 
 	    (item == am_print) || (item == am_timestamp)) {
 	    hp = HAlloc(p,3);
 	    res = TUPLE2(hp, item, am_false);
@@ -1964,11 +1956,7 @@ BIF_RETTYPE seq_trace_info_1(BIF_ALIST_1)
  */
 BIF_RETTYPE seq_trace_print_1(BIF_ALIST_1)    
 {
-    if (SEQ_TRACE_TOKEN(BIF_P) == NIL 
-#ifdef USE_VM_PROBES
-	|| SEQ_TRACE_TOKEN(BIF_P) == am_have_dt_utag
-#endif
-	) {
+    if (have_no_seqtrace(SEQ_TRACE_TOKEN(BIF_P))) {
 	BIF_RET(am_false);
     }
     seq_trace_update_send(BIF_P);
@@ -1987,11 +1975,7 @@ BIF_RETTYPE seq_trace_print_1(BIF_ALIST_1)
  */
 BIF_RETTYPE seq_trace_print_2(BIF_ALIST_2)    
 {
-    if (SEQ_TRACE_TOKEN(BIF_P) == NIL
-#ifdef USE_VM_PROBES
-	|| SEQ_TRACE_TOKEN(BIF_P) == am_have_dt_utag
-#endif
-	) {
+    if (have_no_seqtrace(SEQ_TRACE_TOKEN(BIF_P))) {
 	BIF_RET(am_false);
     }
     if (!(is_atom(BIF_ARG_1) || is_small(BIF_ARG_1))) {
diff --git a/erts/emulator/beam/erl_db.c b/erts/emulator/beam/erl_db.c
index 9ec14ab5ae..3030c1c91a 100644
--- a/erts/emulator/beam/erl_db.c
+++ b/erts/emulator/beam/erl_db.c
@@ -1832,7 +1832,7 @@ BIF_RETTYPE ets_give_away_3(BIF_ALIST_3)
 			     tb->common.id,
 			     from_pid,
 			     BIF_ARG_3), 
-		      0);
+                      0);
     erts_smp_proc_unlock(to_proc, to_locks);
     UnUseTmpHeap(5,BIF_P);
     BIF_RET(am_true);
@@ -3211,7 +3211,7 @@ retry:
 			     tb->common.id,
 			     p->common.id,
 			     heir_data), 
-		      0);
+                      0);
     erts_smp_proc_unlock(to_proc, to_locks);
     return !0;
 }
diff --git a/erts/emulator/beam/erl_db_util.c b/erts/emulator/beam/erl_db_util.c
index 0bf9558ac9..399be6058c 100644
--- a/erts/emulator/beam/erl_db_util.c
+++ b/erts/emulator/beam/erl_db_util.c
@@ -2308,11 +2308,7 @@ restart:
 	    *esp++ = am_true;
 	    break;
 	case matchIsSeqTrace:
-	    if (SEQ_TRACE_TOKEN(c_p) != NIL
-#ifdef USE_VM_PROBES
-		&& SEQ_TRACE_TOKEN(c_p) != am_have_dt_utag
-#endif
-		)
+            if (have_seqtrace(SEQ_TRACE_TOKEN(c_p)))
 		*esp++ = am_true;
 	    else
 		*esp++ = am_false;
@@ -2336,11 +2332,7 @@ restart:
 	    --esp;
 	    break;
 	case matchGetSeqToken:
-	    if (SEQ_TRACE_TOKEN(c_p) == NIL
-#ifdef USE_VM_PROBES
-		|| SEQ_TRACE_TOKEN(c_p) == am_have_dt_utag
-#endif
-		) 
+            if (have_no_seqtrace(SEQ_TRACE_TOKEN(c_p)))
 		*esp++ = NIL;
 	    else {
 		Eterm sender = SEQ_TRACE_TOKEN_SENDER(c_p);
diff --git a/erts/emulator/beam/erl_debug.h b/erts/emulator/beam/erl_debug.h
index f4259e7dae..029320691d 100644
--- a/erts/emulator/beam/erl_debug.h
+++ b/erts/emulator/beam/erl_debug.h
@@ -48,6 +48,7 @@
 #define DEBUG_THREADS      0x0010    /* Thread-related stuff                 */
 #define DEBUG_PROCESSES    0x0020    /* Process creation and removal         */
 #define DEBUG_MEMORY       0x0040    /* Display results of memory checks     */
+#define DEBUG_SHCOPY       0x0080    /* Sharing-preserving copying of terms  */
 
 extern Uint32 verbose;
 
diff --git a/erts/emulator/beam/erl_gc.c b/erts/emulator/beam/erl_gc.c
index 6cb37752bc..c50756d56b 100644
--- a/erts/emulator/beam/erl_gc.c
+++ b/erts/emulator/beam/erl_gc.c
@@ -1166,6 +1166,9 @@ do_minor(Process *p, ErlHeapFragment *live_hf_end,
     char* oh = (char *) OLD_HEAP(p);
     Uint oh_size = (char *) OLD_HTOP(p) - oh;
 
+    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] MINOR GC: %p %p %p %p\n", p->common.id,
+                           HEAP_START(p), HEAP_END(p), OLD_HEAP(p), OLD_HEND(p)));
+
     n_htop = n_heap = (Eterm*) ERTS_HEAP_ALLOC(ERTS_ALC_T_HEAP,
 					       sizeof(Eterm)*new_sz);
 
@@ -1383,6 +1386,9 @@ major_collection(Process* p, ErlHeapFragment *live_hf_end,
     Uint new_sz, stk_sz;
     int adjusted;
 
+    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] MAJOR GC: %p %p %p %p\n", p->common.id,
+                           HEAP_START(p), HEAP_END(p), OLD_HEAP(p), OLD_HEND(p)));
+
     /*
      * Do a fullsweep GC. First figure out the size of the heap
      * to receive all live data.
diff --git a/erts/emulator/beam/erl_init.c b/erts/emulator/beam/erl_init.c
index f396a0a156..296cfdabc3 100644
--- a/erts/emulator/beam/erl_init.c
+++ b/erts/emulator/beam/erl_init.c
@@ -1401,6 +1401,7 @@ erl_start(int argc, char **argv)
 		    case 't': verbose |= DEBUG_THREADS; break;
 		    case 'p': verbose |= DEBUG_PROCESSES; break;
 		    case 'm': verbose |= DEBUG_MESSAGES; break;
+		    case 'c': verbose |= DEBUG_SHCOPY; break;
 		    default : erts_fprintf(stderr,"Unknown verbose option: %c\n",*ch);
 		    }
 		}
@@ -1413,6 +1414,7 @@ erl_start(int argc, char **argv)
 	    if (verbose & DEBUG_THREADS) erts_printf("THREADS ");
 	    if (verbose & DEBUG_PROCESSES) erts_printf("PROCESSES ");
 	    if (verbose & DEBUG_MESSAGES) erts_printf("MESSAGES ");
+	    if (verbose & DEBUG_SHCOPY) erts_printf("SHCOPY ");
             erts_printf("\n");
 #else
 	    erts_fprintf(stderr, "warning: -v (only in debug compiled code)\n");
diff --git a/erts/emulator/beam/erl_message.c b/erts/emulator/beam/erl_message.c
index 79739501a8..1811651a58 100644
--- a/erts/emulator/beam/erl_message.c
+++ b/erts/emulator/beam/erl_message.c
@@ -68,6 +68,7 @@ new_message_buffer(Uint size)
     bp = (ErlHeapFragment*) ERTS_HEAP_ALLOC(ERTS_ALC_T_HEAP_FRAG,
 					    ERTS_HEAP_FRAG_SIZE(size));
     ERTS_INIT_HEAP_FRAG(bp, size, size);
+    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] new message buffer %p\n", erts_get_current_pid(), bp->mem));
     return bp;
 }
 
@@ -249,36 +250,6 @@ erts_realloc_shrink_message(ErtsMessage *mp, Uint sz, Eterm *brefs, Uint brefs_s
 }
 
 void
-erts_link_mbuf_to_proc(Process *proc, ErlHeapFragment *first_bp)
-{
-    if (first_bp) {
-	ErlHeapFragment *bp = first_bp;
-
-	while (1) {
-	    /* Move any off_heap's into the process */
-	    if (bp->off_heap.first != NULL) {
-		struct erl_off_heap_header** next_p = &bp->off_heap.first;
-		while (*next_p != NULL) {
-		    next_p = &((*next_p)->next);
-		}
-		*next_p = MSO(proc).first;
-		MSO(proc).first = bp->off_heap.first;
-		bp->off_heap.first = NULL;
-		OH_OVERHEAD(&(MSO(proc)), bp->off_heap.overhead);
-	    }
-	    MBUF_SIZE(proc) += bp->used_size;
-	    if (!bp->next)
-		break;
-	    bp = bp->next;
-	}
-
-	/* Link the message buffer */
-	bp->next = MBUF(proc);
-	MBUF(proc) = first_bp;
-    }
-}
-
-void
 erts_queue_dist_message(Process *rcvr,
 			ErtsProcLocks *rcvr_locks,
 			ErtsDistExternal *dist_ext,
@@ -343,7 +314,7 @@ erts_queue_dist_message(Process *rcvr,
 		DTRACE_CHARBUF(receiver_name, DTRACE_TERM_BUF_SIZE);
 
 		dtrace_proc_str(rcvr, receiver_name);
-		if (token != NIL && token != am_have_dt_utag) {
+                if (have_seqtrace(token)) {
 		    tok_label = signed_val(SEQ_TRACE_T_LABEL(token));
 		    tok_lastcnt = signed_val(SEQ_TRACE_T_LASTCNT(token));
 		    tok_serial = signed_val(SEQ_TRACE_T_SERIAL(token));
@@ -364,7 +335,7 @@ erts_queue_dist_message(Process *rcvr,
             DTRACE_CHARBUF(receiver_name, DTRACE_TERM_BUF_SIZE);
 
             dtrace_proc_str(rcvr, receiver_name);
-            if (token != NIL && token != am_have_dt_utag) {
+            if (have_seqtrace(token)) {
                 tok_label = signed_val(SEQ_TRACE_T_LABEL(token));
                 tok_lastcnt = signed_val(SEQ_TRACE_T_LASTCNT(token));
                 tok_serial = signed_val(SEQ_TRACE_T_SERIAL(token));
@@ -540,6 +511,36 @@ erts_queue_message(Process* receiver, ErtsProcLocks *receiver_locks,
 	);
 }
 
+void
+erts_link_mbuf_to_proc(Process *proc, ErlHeapFragment *first_bp)
+{
+    if (first_bp) {
+	ErlHeapFragment *bp = first_bp;
+
+	while (1) {
+	    /* Move any off_heap's into the process */
+	    if (bp->off_heap.first != NULL) {
+		struct erl_off_heap_header** next_p = &bp->off_heap.first;
+		while (*next_p != NULL) {
+		    next_p = &((*next_p)->next);
+		}
+		*next_p = MSO(proc).first;
+		MSO(proc).first = bp->off_heap.first;
+		bp->off_heap.first = NULL;
+		OH_OVERHEAD(&(MSO(proc)), bp->off_heap.overhead);
+	    }
+	    MBUF_SIZE(proc) += bp->used_size;
+	    if (!bp->next)
+		break;
+	    bp = bp->next;
+	}
+
+	/* Link the message buffer */
+	bp->next = MBUF(proc);
+	MBUF(proc) = first_bp;
+    }
+}
+
 Uint
 erts_msg_attached_data_size_aux(ErtsMessage *msg)
 {
@@ -663,11 +664,14 @@ erts_send_message(Process* sender,
     Eterm utag = NIL;
 #endif
     erts_aint32_t receiver_state;
+#ifdef SHCOPY_SEND
+    erts_shcopy_t info;
+#endif
     BM_STOP_TIMER(system);
     BM_MESSAGE(message,sender,receiver);
     BM_START_TIMER(send);
 
- #ifdef USE_VM_PROBES
+#ifdef USE_VM_PROBES
     *sender_name = *receiver_name = '\0';
     if (DTRACE_ENABLED(message_send)) {
         erts_snprintf(sender_name, sizeof(DTRACE_CHARBUF_NAME(sender_name)),
@@ -686,48 +690,61 @@ erts_send_message(Process* sender,
 #ifdef USE_VM_PROBES
 	Uint dt_utag_size = 0;
 #endif
-
-	BM_SWAP_TIMER(send,size);
-	msize = size_object(message);
-	BM_SWAP_TIMER(size,send);
-
-#ifdef USE_VM_PROBES
-	if (stoken != am_have_dt_utag) {
-#endif
-
+        BM_SWAP_TIMER(send,size);
+
+        /* SHCOPY corrupts the heap between
+         * copy_shared_calculate, and
+         * copy_shared_perform. (it inserts move_markers like the gc).
+         * Make sure we don't use the heap between those instances.
+         */
+        if (have_seqtrace(stoken)) {
 	    seq_trace_update_send(sender);
 	    seq_trace_output(stoken, message, SEQ_TRACE_SEND, 
 			     receiver->common.id, sender);
 	    seq_trace_size = 6; /* TUPLE5 */
-#ifdef USE_VM_PROBES
-	}
-	if (DT_UTAG_FLAGS(sender) & DT_UTAG_SPREADING) {
-	    dt_utag_size = size_object(DT_UTAG(sender));
-	} else if (stoken == am_have_dt_utag ) {
-	    stoken = NIL;
 	}
+#ifdef USE_VM_PROBES
+        if (DT_UTAG_FLAGS(sender) & DT_UTAG_SPREADING) {
+            dt_utag_size = size_object(DT_UTAG(sender));
+        } else if (stoken == am_have_dt_utag ) {
+            stoken = NIL;
+        }
 #endif
 
-	mp = erts_alloc_message_heap_state(receiver,
-					   &receiver_state,
-					   receiver_locks,
-					   (msize
+#ifdef SHCOPY_SEND
+        INITIALIZE_SHCOPY(info);
+        msize = copy_shared_calculate(message, &info);
+#else
+        msize = size_object(message);
+#endif
+        BM_SWAP_TIMER(size,send);
+
+        mp = erts_alloc_message_heap_state(receiver,
+                                           &receiver_state,
+                                           receiver_locks,
+                                           (msize
 #ifdef USE_VM_PROBES
-					    + dt_utag_size
+                                            + dt_utag_size
 #endif
-					    + seq_trace_size),
-					   &hp,
-					   &ohp);
+                                            + seq_trace_size),
+                                           &hp,
+                                           &ohp);
 
         BM_SWAP_TIMER(send,copy);
+
+#ifdef SHCOPY_SEND
+	if (is_not_immed(message))
+            message = copy_shared_perform(message, msize, &info, &hp, ohp);
+        DESTROY_SHCOPY(info);
+#else
+	if (is_not_immed(message))
+            message = copy_struct(message, msize, &hp, ohp);
+#endif
 	if (is_immed(stoken))
 	    token = stoken;
 	else
 	    token = copy_struct(stoken, seq_trace_size, &hp, ohp);
 
-	if (is_not_immed(message))
-	    message = copy_struct(message, msize, &hp, ohp);
-
 #ifdef USE_VM_PROBES
 	if (DT_UTAG_FLAGS(sender) & DT_UTAG_SPREADING) {
 	    if (is_immed(DT_UTAG(sender)))
@@ -746,7 +763,7 @@ erts_send_message(Process* sender,
 
 #ifdef USE_VM_PROBES
         if (DTRACE_ENABLED(message_send)) {
-	    if (stoken != NIL && stoken != am_have_dt_utag) {
+            if (have_seqtrace(stoken)) {
 		tok_label = signed_val(SEQ_TRACE_T_LABEL(stoken));
 		tok_lastcnt = signed_val(SEQ_TRACE_T_LASTCNT(stoken));
 		tok_serial = signed_val(SEQ_TRACE_T_SERIAL(stoken));
@@ -764,7 +781,12 @@ erts_send_message(Process* sender,
 	}
 	else {
 	    BM_SWAP_TIMER(send,size);
-	    msize = size_object(message);
+#ifdef SHCOPY_SEND
+            INITIALIZE_SHCOPY(info);
+            msize = copy_shared_calculate(message, &info);
+#else
+            msize = size_object(message);
+#endif
 	    BM_SWAP_TIMER(size,send);
 
 	    mp = erts_alloc_message_heap_state(receiver,
@@ -774,8 +796,14 @@ erts_send_message(Process* sender,
 					       &hp,
 					       &ohp);
 	    BM_SWAP_TIMER(send,copy);
-	    if (is_not_immed(message))
-		message = copy_struct(message, msize, &hp, ohp);
+#ifdef SHCOPY_SEND
+            if (is_not_immed(message))
+                message = copy_shared_perform(message, msize, &info, &hp, ohp);
+            DESTROY_SHCOPY(info);
+#else
+            if (is_not_immed(message))
+                message = copy_struct(message, msize, &hp, ohp);
+#endif
 	    BM_MESSAGE_COPIED(msz);
 	    BM_SWAP_TIMER(copy,send);
 	}
@@ -800,6 +828,7 @@ erts_send_message(Process* sender,
     return res;
 }
 
+
 /*
  * This function delivers an EXIT message to a process
  * which is trapping EXITs.
@@ -819,21 +848,29 @@ erts_deliver_exit_message(Eterm from, Process *to, ErtsProcLocks *to_locksp,
     Eterm temptoken;
     ErtsMessage* mp;
     ErlOffHeap *ohp;
-
-    if (token != NIL
-#ifdef USE_VM_PROBES
-	&& token != am_have_dt_utag
+#ifdef SHCOPY_SEND
+    erts_shcopy_t info;
 #endif
-	) {
 
+    if (have_seqtrace(token)) {
 	ASSERT(is_tuple(token));
-	sz_reason = size_object(reason);
 	sz_token = size_object(token);
 	sz_from = size_object(from);
+#ifdef SHCOPY_SEND
+        INITIALIZE_SHCOPY(info);
+        sz_reason = copy_shared_calculate(reason, &info);
+#else
+	sz_reason = size_object(reason);
+#endif
 	mp = erts_alloc_message_heap(to, to_locksp,
 				     sz_reason + sz_from + sz_token + 4,
 				     &hp, &ohp);
+#ifdef SHCOPY_SEND
+        mess = copy_shared_perform(reason, sz_reason, &info, &hp, ohp);
+        DESTROY_SHCOPY(info);
+#else
 	mess = copy_struct(reason, sz_reason, &hp, ohp);
+#endif
 	from_copy = copy_struct(from, sz_from, &hp, ohp);
 	save = TUPLE3(hp, am_EXIT, from_copy, mess);
 	hp += 4;
@@ -842,13 +879,22 @@ erts_deliver_exit_message(Eterm from, Process *to, ErtsProcLocks *to_locksp,
 	temptoken = copy_struct(token, sz_token, &hp, ohp);
 	erts_queue_message(to, to_locksp, mp, save, temptoken);
     } else {
-	sz_reason = size_object(reason);
 	sz_from = IS_CONST(from) ? 0 : size_object(from);
-
+#ifdef SHCOPY_SEND
+        INITIALIZE_SHCOPY(info);
+        sz_reason = copy_shared_calculate(reason, &info);
+#else
+	sz_reason = size_object(reason);
+#endif
 	mp = erts_alloc_message_heap(to, to_locksp,
 				     sz_reason+sz_from+4, &hp, &ohp);
 
+#ifdef SHCOPY_SEND
+        mess = copy_shared_perform(reason, sz_reason, &info, &hp, ohp);
+        DESTROY_SHCOPY(info);
+#else
 	mess = copy_struct(reason, sz_reason, &hp, ohp);
+#endif
 	from_copy = (IS_CONST(from)
 		     ? from
 		     : copy_struct(from, sz_from, &hp, ohp));
@@ -1725,4 +1771,3 @@ void erts_factory_undo(ErtsHeapFactory* factory)
     factory->heap_frags = NULL;
 #endif
 }
-
diff --git a/erts/emulator/beam/erl_message.h b/erts/emulator/beam/erl_message.h
index 740ae46a0f..76387bc34c 100644
--- a/erts/emulator/beam/erl_message.h
+++ b/erts/emulator/beam/erl_message.h
@@ -128,6 +128,13 @@ struct erl_heap_fragment {
 #else
 #endif
 
+#ifdef USE_VM_PROBES
+#define have_no_seqtrace(T) ((T) == NIL || (T) == am_have_dt_utag)
+#else
+#define have_no_seqtrace(T) ((T) == NIL)
+#endif
+#define have_seqtrace(T)    (!have_no_seqtrace(T))
+
 #define ERL_MESSAGE_REF_FIELDS__			\
     ErtsMessage *next;	/* Next message */		\
     union {						\
diff --git a/erts/emulator/beam/erl_nif.c b/erts/emulator/beam/erl_nif.c
index a37cda93ef..2ff509e6d0 100644
--- a/erts/emulator/beam/erl_nif.c
+++ b/erts/emulator/beam/erl_nif.c
@@ -379,9 +379,19 @@ ERL_NIF_TERM enif_make_copy(ErlNifEnv* dst_env, ERL_NIF_TERM src_term)
 {
     Uint sz;
     Eterm* hp;
+#ifdef SHCOPY
+    erts_shcopy_t info;
+    INITIALIZE_SHCOPY(info);
+    sz = copy_shared_calculate(src_term, &info);
+    hp = alloc_heap(dst_env, sz);
+    src_term = copy_shared_perform(src_term, sz, &info, &hp, &MSO(dst_env->proc));
+    DESTROY_SHCOPY(info);
+    return src_term;
+#else
     sz = size_object(src_term);
     hp = alloc_heap(dst_env, sz);
     return copy_struct(src_term, sz, &hp, &MSO(dst_env->proc));
+#endif
 }
 
 
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index 9acce8acb6..5907dd4567 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -10747,6 +10747,10 @@ erl_create_process(Process* parent, /* Parent of process (default group leader).
     Eterm res = THE_NON_VALUE;
     erts_aint32_t state = 0;
     erts_aint32_t prio = (erts_aint32_t) PRIORITY_NORMAL;
+#ifdef SHCOPY_SPAWN
+    erts_shcopy_t info;
+    INITIALIZE_SHCOPY(info);
+#endif
 
 #ifdef ERTS_SMP
     erts_smp_proc_lock(parent, ERTS_PROC_LOCKS_ALL_MINOR);
@@ -10798,7 +10802,11 @@ erl_create_process(Process* parent, /* Parent of process (default group leader).
     BM_COUNT(processes_spawned);
 
     BM_SWAP_TIMER(system,size);
+#ifdef SHCOPY_SPAWN
+    arg_size = copy_shared_calculate(args, &info);
+#else
     arg_size = size_object(args);
+#endif
     BM_SWAP_TIMER(size,system);
     heap_need = arg_size;
 
@@ -10876,7 +10884,12 @@ erl_create_process(Process* parent, /* Parent of process (default group leader).
     BM_MESSAGE(args,p,parent);
     BM_START_TIMER(system);
     BM_SWAP_TIMER(system,copy);
+#ifdef SHCOPY_SPAWN
+    p->arg_reg[2] = copy_shared_perform(args, arg_size, &info, &p->htop, &p->off_heap);
+    DESTROY_SHCOPY(info);
+#else
     p->arg_reg[2] = copy_struct(args, arg_size, &p->htop, &p->off_heap);
+#endif
     BM_MESSAGE_COPIED(arg_size);
     BM_SWAP_TIMER(copy,system);
     p->arity = 3;
@@ -11260,6 +11273,8 @@ static void
 delete_process(Process* p)
 {
     VERBOSE(DEBUG_PROCESSES, ("Removing process: %T\n",p->common.id));
+    VERBOSE(DEBUG_SHCOPY, ("[pid=%T] delete process: %p %p %p %p\n", p->common.id,
+                           HEAP_START(p), HEAP_END(p), OLD_HEAP(p), OLD_HEND(p)));
 
     /* Cleanup psd */
 
@@ -11505,31 +11520,40 @@ send_exit_message(Process *to, ErtsProcLocks *to_locksp,
 {
     ErtsMessage *mp;
     ErlOffHeap *ohp;
-
-    if (token == NIL 
-#ifdef USE_VM_PROBES
-	|| token == am_have_dt_utag
-#endif
-	) {
-	Eterm* hp;
-	Eterm mess;
-
-	mp = erts_alloc_message_heap(to, to_locksp,
-				     term_size, &hp, &ohp);
+    Eterm* hp;
+    Eterm mess;
+#ifdef SHCOPY_SEND
+    erts_shcopy_t info;
+#endif
+
+    if (!have_seqtrace(token)) {
+#ifdef SHCOPY_SEND
+        INITIALIZE_SHCOPY(info);
+        term_size = copy_shared_calculate(exit_term, &info);
+	mp = erts_alloc_message_heap(to, to_locksp, term_size, &hp, &ohp);
+        mess = copy_shared_perform(exit_term, term_size, &info, &hp, ohp);
+        DESTROY_SHCOPY(info);
+#else
+	mp = erts_alloc_message_heap(to, to_locksp, term_size, &hp, &ohp);
 	mess = copy_struct(exit_term, term_size, &hp, ohp);
+#endif
 	erts_queue_message(to, to_locksp, mp, mess, NIL);
     } else {
-	Eterm* hp;
-	Eterm mess;
 	Eterm temp_token;
 	Uint sz_token;
 
 	ASSERT(is_tuple(token));
 	sz_token = size_object(token);
-
-	mp = erts_alloc_message_heap(to, to_locksp,
-				     term_size+sz_token, &hp, &ohp);
+#ifdef SHCOPY_SEND
+        INITIALIZE_SHCOPY(info);
+        term_size = copy_shared_calculate(exit_term, &info);
+	mp = erts_alloc_message_heap(to, to_locksp, term_size, &hp, &ohp);
+        mess = copy_shared_perform(exit_term, term_size, &info, &hp, ohp);
+        DESTROY_SHCOPY(info);
+#else
+	mp = erts_alloc_message_heap(to, to_locksp, term_size+sz_token, &hp, &ohp);
 	mess = copy_struct(exit_term, term_size, &hp, ohp);
+#endif
 	/* the trace token must in this case be updated by the caller */
 	seq_trace_output(token, mess, SEQ_TRACE_SEND, to->common.id, NULL);
 	temp_token = copy_struct(token, sz_token, &hp, ohp);
@@ -11642,11 +11666,7 @@ send_exit_signal(Process *c_p,		/* current process if and only
 
     if ((state & ERTS_PSFLG_TRAP_EXIT)
 	&& (reason != am_kill || (flags & ERTS_XSIG_FLG_IGN_KILL))) {
-	if (is_not_nil(token) 
-#ifdef USE_VM_PROBES
-	    && token != am_have_dt_utag
-#endif
-	    && token_update)
+        if (have_seqtrace(token))
 	    seq_trace_update_send(token_update);
 	if (is_value(exit_tuple))
 	    send_exit_message(rp, rp_locks, exit_tuple, exit_tuple_sz, token);
diff --git a/erts/emulator/beam/erl_trace.c b/erts/emulator/beam/erl_trace.c
index d02f1f7213..b774ba97af 100644
--- a/erts/emulator/beam/erl_trace.c
+++ b/erts/emulator/beam/erl_trace.c
@@ -1046,11 +1046,7 @@ seq_trace_update_send(Process *p)
 {
     Eterm seq_tracer = erts_get_system_seq_tracer();
     ASSERT((is_tuple(SEQ_TRACE_TOKEN(p)) || is_nil(SEQ_TRACE_TOKEN(p))));
-    if ( (p->common.id == seq_tracer) || (SEQ_TRACE_TOKEN(p) == NIL)
-#ifdef USE_VM_PROBES
-	 || (SEQ_TRACE_TOKEN(p) == am_have_dt_utag)
-#endif
-	 ) {
+    if ((p->common.id == seq_tracer) || have_no_seqtrace(SEQ_TRACE_TOKEN(p))) {
 	return 0;
     }
     SEQ_TRACE_TOKEN_SENDER(p) = p->common.id;
diff --git a/erts/emulator/beam/global.h b/erts/emulator/beam/global.h
index 594c0ccf94..98c275a20c 100644
--- a/erts/emulator/beam/global.h
+++ b/erts/emulator/beam/global.h
@@ -380,7 +380,7 @@ extern int bif_reductions;      /* reductions + fcalls (when doing call_bif) */
 extern int stackdump_on_exit;
 
 /*
- * Here is an implementation of a lightweiht stack.
+ * Here is an implementation of a lightweight stack.
  *
  * Use it like this:
  *
@@ -845,6 +845,94 @@ do {\
 } while(0)
 
 
+/*
+ *  An implementation of lightweight unbounded queues,
+ *  using a circular dynamic array.
+ *  It does not include support for change_allocator.
+ *
+ *  Use it like this:
+ *
+ *  DECLARE_EQUEUE(Queue)	(At the start of a block)
+ *  ...
+ *  EQUEUE_PUT(Queue, Term)
+ *  ...
+ *  if (EQUEUE_ISEMPTY(Queue)) {
+ *     Queue is empty
+ *  } else {
+ *     Term = EQUEUE_GET(Stack);
+ *     Process popped Term here
+ *  }
+ *  ...
+ *  DESTROY_EQUEUE(Queue)
+ */
+
+typedef struct {
+    Eterm* start;
+    Eterm* front;
+    Eterm* back;
+    int possibly_empty;
+    Eterm* end;
+    ErtsAlcType_t alloc_type;
+} ErtsEQueue;
+
+#define DEF_EQUEUE_SIZE (16)
+
+void erl_grow_equeue(ErtsEQueue*, Eterm* def_queue);
+#define EQUE_CONCAT(a,b) a##b
+#define EQUE_DEF_QUEUE(q) EQUE_CONCAT(q,_default_equeue)
+
+#define DECLARE_EQUEUE(q)				\
+    UWord EQUE_DEF_QUEUE(q)[DEF_EQUEUE_SIZE];     	\
+    ErtsEQueue q = {					\
+        EQUE_DEF_QUEUE(q), /* start */			\
+        EQUE_DEF_QUEUE(q), /* front */			\
+        EQUE_DEF_QUEUE(q), /* back */			\
+        1,                 /* possibly_empty */		\
+        EQUE_DEF_QUEUE(q) + DEF_EQUEUE_SIZE, /* end */	\
+        ERTS_ALC_T_ESTACK  /* alloc_type */		\
+    }
+
+#define DESTROY_EQUEUE(q)				\
+do {							\
+    if (q.start != EQUE_DEF_QUEUE(q)) {			\
+      erts_free(q.alloc_type, q.start);			\
+    }							\
+} while(0)
+
+#define EQUEUE_PUT_UNCHECKED(q, x)			\
+do {							\
+    q.possibly_empty = 0;				\
+    *(q.back) = (x);                    		\
+    if (++(q.back) == q.end) {				\
+	q.back = q.start;				\
+    }							\
+} while(0)
+
+#define EQUEUE_PUT(q, x)				\
+do {							\
+    if (q.back == q.front && !q.possibly_empty) {	\
+        erl_grow_equeue(&q, EQUE_DEF_QUEUE(q));		\
+    }							\
+    EQUEUE_PUT_UNCHECKED(q, x);				\
+} while(0)
+
+#define EQUEUE_ISEMPTY(q) (q.back == q.front && q.possibly_empty)
+
+ERTS_GLB_INLINE Eterm erts_equeue_get(ErtsEQueue *q);
+
+#if ERTS_GLB_INLINE_INCL_FUNC_DEF
+ERTS_GLB_INLINE Eterm erts_equeue_get(ErtsEQueue *q) {
+    Eterm x;
+    q->possibly_empty = 1;
+    x = *(q->front);
+    if (++(q->front) == q->end) {
+        q->front = q->start;
+    }
+    return x;
+}
+#endif
+#define EQUEUE_GET(q) erts_equeue_get(&(q));
+
 /* binary.c */
 
 void erts_emasculate_writable_binary(ProcBin* pb);
@@ -899,6 +987,12 @@ Eterm erl_is_function(Process* p, Eterm arg1, Eterm arg2);
 /* beam_bif_load.c */
 Eterm erts_check_process_code(Process *c_p, Eterm module, int allow_gc, int *redsp);
 
+typedef struct {
+    Eterm *ptr;
+    Uint   sz;
+} copy_literals_t;
+
+extern copy_literals_t erts_clrange;
 
 /* beam_load.c */
 typedef struct {
@@ -953,12 +1047,67 @@ __decl_noreturn void __noreturn erl_exit(int n, char*, ...);
 __decl_noreturn void __noreturn erl_exit_flush_async(int n, char*, ...);
 void erl_error(char*, va_list);
 
+/* This controls whether sharing-preserving copy is used by Erlang */
+
+#ifdef SHCOPY
+#define SHCOPY_SEND
+#define SHCOPY_SPAWN
+#endif
+
+/* The persistent state while the sharing-preserving copier works */
+
+typedef struct {
+    Eterm  queue_default[DEF_EQUEUE_SIZE];
+    Eterm* queue_start;
+    Eterm* queue_end;
+    ErtsAlcType_t queue_alloc_type;
+    UWord  bitstore_default[DEF_WSTACK_SIZE];
+    UWord* bitstore_start;
+    ErtsAlcType_t bitstore_alloc_type;
+    Eterm  shtable_default[DEF_ESTACK_SIZE];
+    Eterm* shtable_start;
+    ErtsAlcType_t shtable_alloc_type;
+    Uint literal_size;
+    Eterm *range_ptr;
+    Uint  range_sz;
+} erts_shcopy_t;
+
+#define INITIALIZE_SHCOPY(info)                         \
+do {                                                    \
+    info.queue_start = info.queue_default;              \
+    info.bitstore_start = info.bitstore_default;        \
+    info.shtable_start = info.shtable_default;          \
+    info.literal_size = 0;                              \
+    info.range_ptr = erts_clrange.ptr;                  \
+    info.range_sz  = erts_clrange.sz;                   \
+} while(0)
+
+#define DESTROY_SHCOPY(info)                                            \
+do {                                                                    \
+    if (info.queue_start != info.queue_default) {                       \
+        erts_free(info.queue_alloc_type, info.queue_start);             \
+    }                                                                   \
+    if (info.bitstore_start != info.bitstore_default) {                 \
+        erts_free(info.bitstore_alloc_type, info.bitstore_start);       \
+    }                                                                   \
+    if (info.shtable_start != info.shtable_default) {                   \
+        erts_free(info.shtable_alloc_type, info.shtable_start);         \
+    }                                                                   \
+} while(0)
+
 /* copy.c */
 Eterm copy_object_x(Eterm, Process*, Uint);
 #define copy_object(Term, Proc) copy_object_x(Term,Proc,0)
 
 Uint size_object(Eterm);
-Eterm copy_struct(Eterm, Uint, Eterm**, ErlOffHeap*);
+Uint copy_shared_calculate(Eterm, erts_shcopy_t*);
+Eterm copy_shared_perform(Eterm, Uint, erts_shcopy_t*, Eterm**, ErlOffHeap*);
+
+Uint size_shared(Eterm);
+
+Eterm copy_struct_x(Eterm, Uint, Eterm**, ErlOffHeap*, Uint* bsz);
+#define copy_struct(Obj,Sz,HPP,OH) \
+    copy_struct_x(Obj,Sz,HPP,OH,NULL)
 Eterm copy_shallow(Eterm*, Uint, Eterm**, ErlOffHeap*);
 
 void erts_move_multi_frags(Eterm** hpp, ErlOffHeap*, ErlHeapFragment* first,
diff --git a/erts/emulator/beam/utils.c b/erts/emulator/beam/utils.c
index c3735683bb..184477c36b 100644
--- a/erts/emulator/beam/utils.c
+++ b/erts/emulator/beam/utils.c
@@ -258,6 +258,31 @@ erl_grow_pstack(ErtsPStack* s, void* default_pstack, unsigned need_bytes)
     s->psp = s->pstart + sp_offs;
 }
 
+/*
+ * Helper function for the EQUEUE macros defined in global.h.
+ */
+
+void
+erl_grow_equeue(ErtsEQueue* q, Eterm* default_equeue)
+{
+    Uint old_size = (q->end - q->start);
+    Uint new_size = old_size * 2;
+    Uint first_part = (q->end - q->front);
+    Uint second_part = (q->back - q->start);
+    Eterm* new_ptr = erts_alloc(q->alloc_type, new_size*sizeof(Eterm));
+    ASSERT(q->back == q->front);   // of course the queue is full now!
+    if (first_part > 0)
+      sys_memcpy(new_ptr, q->front, first_part*sizeof(Eterm));
+    if (second_part > 0)
+      sys_memcpy(new_ptr+first_part, q->start, second_part*sizeof(Eterm));
+    if (q->start != default_equeue)
+      erts_free(q->alloc_type, q->start);
+    q->start = new_ptr;
+    q->end = q->start + new_size;
+    q->front = q->start;
+    q->back = q->start + old_size;
+}
+
 /* CTYPE macros */
 
 #define LATIN1
diff --git a/erts/emulator/test/code_SUITE.erl b/erts/emulator/test/code_SUITE.erl
index 9f318a38be..1acc4538fb 100644
--- a/erts/emulator/test/code_SUITE.erl
+++ b/erts/emulator/test/code_SUITE.erl
@@ -26,7 +26,8 @@
 	 t_check_process_code_ets/1,
 	 external_fun/1,get_chunk/1,module_md5/1,make_stub/1,
 	 make_stub_many_funs/1,constant_pools/1,constant_refc_binaries/1,
-	 false_dependency/1,coverage/1,fun_confusion/1]).
+	 false_dependency/1,coverage/1,fun_confusion/1,
+         t_copy_literals/1]).
 
 -define(line_trace, 1).
 -include_lib("test_server/include/test_server.hrl").
@@ -38,7 +39,7 @@ all() ->
      t_check_process_code_ets, t_check_old_code, external_fun, get_chunk,
      module_md5, make_stub, make_stub_many_funs,
      constant_pools, constant_refc_binaries, false_dependency,
-     coverage, fun_confusion].
+     coverage, fun_confusion, t_copy_literals].
 
 groups() -> 
     [].
@@ -753,6 +754,80 @@ compile_load(Mod, Src, Ver) ->
     {module,Mod} = code:load_binary(Mod, "fun_confusion.beam", Code1),
     ok.
 
+
+t_copy_literals(Config) when is_list(Config) ->
+    %% Compile the the literals module.
+    Data = ?config(data_dir, Config),
+    File = filename:join(Data, "literals"),
+    {ok,literals,Code} = compile:file(File, [report,binary]),
+    {module,literals} = erlang:load_module(literals, Code),
+
+    N   = 30,
+    Me  = self(),
+    %% reload literals code every 567 ms
+    Rel = spawn_link(fun() -> reloader(literals,Code,567) end),
+    %% add new literal msgs to the loop every 789 ms
+    Sat = spawn_link(fun() -> saturate(Me,789) end),
+    %% run for 10s
+    _   = spawn_link(fun() -> receive after 10000 -> Me ! done end end),
+    ok  = chase_msg(N, Me),
+    %% cleanup
+    Rel ! done,
+    Sat ! done,
+    ok  = flush(),
+    ok.
+
+
+chase_msg(0, Pid) ->
+    chase_loop(Pid);
+chase_msg(N, Master) ->
+    Pid = spawn_link(fun() -> chase_msg(N - 1,Master) end),
+    chase_loop(Pid).
+
+chase_loop(Pid) ->
+    receive
+        done ->
+            Pid ! done,
+            ok;
+        {_From,Msg} ->
+            Pid ! {self(), Msg},
+            ok = traverse(Msg),
+            chase_loop(Pid)
+    end.
+
+saturate(Pid,Time) ->
+    Es = [msg1,msg2,msg3,msg4,msg5],
+    Msg = [literals:E()||E <- Es],
+    Pid ! {self(), Msg},
+    receive
+        done -> ok
+    after Time ->
+              saturate(Pid,Time)
+    end.
+
+traverse([]) -> ok;
+traverse([H|T]) ->
+    ok = traverse(H),
+    traverse(T);
+traverse(T) when is_tuple(T) -> ok;
+traverse(B) when is_binary(B) -> ok;
+traverse(I) when is_integer(I) -> ok;
+traverse(#{ 1 := V1, b := V2 }) ->
+    ok = traverse(V1),
+    ok = traverse(V2),
+    ok.
+
+
+reloader(Mod,Code,Time) ->
+    receive
+        done -> ok
+    after Time ->
+              code:purge(Mod),
+              {module,Mod} = erlang:load_module(Mod, Code),
+              reloader(Mod,Code,Time)
+    end.
+
+
 %% Utilities.
 
 make_sub_binary(Bin) when is_binary(Bin) ->
@@ -775,4 +850,7 @@ bit_sized_binary(Bin0) ->
     BitSize = 8*size(Bin) + 1,
     Bin.
 
+flush() ->
+    receive _ -> flush() after 0 -> ok end.
+
 id(I) -> I.
diff --git a/erts/emulator/test/code_SUITE_data/literals.erl b/erts/emulator/test/code_SUITE_data/literals.erl
index 9802d9d3f9..a36bfe09dd 100644
--- a/erts/emulator/test/code_SUITE_data/literals.erl
+++ b/erts/emulator/test/code_SUITE_data/literals.erl
@@ -20,6 +20,7 @@
 
 -module(literals).
 -export([a/0,b/0,huge_bignum/0,binary/0,unused_binaries/0,bits/0]).
+-export([msg1/0,msg2/0,msg3/0,msg4/0,msg5/0]).
 
 a() ->
     {a,42.0,[7,38877938333399637266518333334747]}.
@@ -101,3 +102,9 @@ unused_binaries() ->
 
 bits() ->
     {bits,<<42:13,?MB_1>>}.
+
+msg1() -> "halloj".
+msg2() -> {"hello","world"}.
+msg3() -> <<"halloj">>.
+msg4() -> #{ 1=> "hello", b => "world"}.
+msg5() -> {1,2,3,4,5,6}.
diff --git a/erts/preloaded/ebin/erlang.beam b/erts/preloaded/ebin/erlang.beam
index 641fac2d26..4f35928db2 100644
--- a/erts/preloaded/ebin/erlang.beam
+++ b/erts/preloaded/ebin/erlang.beam
diff --git a/erts/preloaded/src/erlang.erl b/erts/preloaded/src/erlang.erl
index 6a9ec9c915..0d5176019f 100644
--- a/erts/preloaded/src/erlang.erl
+++ b/erts/preloaded/src/erlang.erl
@@ -91,7 +91,7 @@
 -export([bit_size/1, bitsize/1, bitstring_to_list/1]).
 -export([bump_reductions/1, byte_size/1, call_on_load_function/1]).
 -export([cancel_timer/1, cancel_timer/2, check_old_code/1, check_process_code/2,
-	 check_process_code/3, crc32/1]).
+	 check_process_code/3, copy_literals/2, crc32/1]).
 -export([crc32/2, crc32_combine/3, date/0, decode_packet/3]).
 -export([delete_element/2]).
 -export([delete_module/1, demonitor/1, demonitor/2, display/1]).
@@ -520,6 +520,13 @@ get_cpc_opts([{allow_gc, AllowGC} | Options], Async, _OldAllowGC) ->
 get_cpc_opts([], Async, AllowGC) ->
     {Async, AllowGC}.
 
+%% copy_literals/2
+-spec erlang:copy_literals(Module,Bool) -> 'true' | 'false' | 'aborted' when
+      Module :: module(),
+      Bool :: boolean().
+copy_literals(_Mod, _Bool) ->
+    erlang:nif_error(undefined).
+
 %% crc32/1
 -spec erlang:crc32(Data) -> non_neg_integer() when
       Data :: iodata().
diff --git a/lib/kernel/src/code_server.erl b/lib/kernel/src/code_server.erl
index e461c95d19..68dd21b1d7 100644
--- a/lib/kernel/src/code_server.erl
+++ b/lib/kernel/src/code_server.erl
@@ -1428,16 +1428,18 @@ absname_vr([[X, $:]|Name], _, _AbsBase) ->
 do_purge(Mod0) ->
     Mod = to_atom(Mod0),
     case erlang:check_old_code(Mod) of
-	false ->
-	    false;
-	true ->
-	    Res = check_proc_code(erlang:processes(), Mod, true),
-	    try
-		erlang:purge_module(Mod)
-	    catch
-		_:_ -> ignore
-	    end,
-	    Res
+        false ->
+            false;
+        true ->
+            true = erlang:copy_literals(Mod, true),
+            Res = check_proc_code(erlang:processes(), Mod, true),
+            true = erlang:copy_literals(Mod, false),
+            try
+                erlang:purge_module(Mod)
+            catch
+                _:_ -> ignore
+            end,
+            Res
     end.
 
 %% do_soft_purge(Module)
@@ -1451,10 +1453,13 @@ do_soft_purge(Mod0) ->
 	false ->
 	    true;
 	true ->
+            true = erlang:copy_literals(Mod, true),
 	    case check_proc_code(erlang:processes(), Mod, false) of
 		false ->
+                    true = erlang:copy_literals(Mod, false),
 		    false;
 		true ->
+                    true = erlang:copy_literals(Mod, false),
 		    try
 			erlang:purge_module(Mod)
 		    catch
diff --git a/lib/kernel/src/erts_debug.erl b/lib/kernel/src/erts_debug.erl
index 39308c0043..87f001fdf4 100644
--- a/lib/kernel/src/erts_debug.erl
+++ b/lib/kernel/src/erts_debug.erl
@@ -34,7 +34,8 @@
 -export([breakpoint/2, disassemble/1, display/1, dist_ext_to_term/2,
          dump_monitors/1, dump_links/1, flat_size/1,
          get_internal_state/1, instructions/0, lock_counters/1,
-         map_info/1, same/2, set_internal_state/2]).
+         map_info/1, same/2, set_internal_state/2,
+         size_shared/1, copy_shared/1]).
 
 -spec breakpoint(MFA, Flag) -> non_neg_integer() when
       MFA :: {Module :: module(),
@@ -86,6 +87,18 @@ dump_links(_) ->
 flat_size(_) ->
     erlang:nif_error(undef).
 
+-spec size_shared(Term) -> non_neg_integer() when
+      Term :: term().
+
+size_shared(_) ->
+    erlang:nif_error(undef).
+
+-spec copy_shared(Term) -> term() when
+      Term :: term().
+
+copy_shared(_) ->
+    erlang:nif_error(undef).
+
 -spec get_internal_state(W) -> term() when
       W :: reds_left | node_and_dist_references | monitoring_nodes
          | next_pid | 'DbTable_words' | check_io_debug