16 files changed, 336 insertions, 131 deletions

diff --git a/erts/emulator/beam/arith_instrs.tab b/erts/emulator/beam/arith_instrs.tab
index 7c9cd47e28..3db19df3c4 100644
--- a/erts/emulator/beam/arith_instrs.tab
+++ b/erts/emulator/beam/arith_instrs.tab
@@ -51,8 +51,7 @@ plus.fetch(Op1, Op2) {
 plus.execute(Fail, Live, Dst) {
     if (ERTS_LIKELY(is_both_small(PlusOp1, PlusOp2))) {
         Sint i = signed_val(PlusOp1) + signed_val(PlusOp2);
-        ASSERT(MY_IS_SSMALL(i) == IS_SSMALL(i));
-        if (ERTS_LIKELY(MY_IS_SSMALL(i))) {
+        if (ERTS_LIKELY(IS_SSMALL(i))) {
             $Dst = make_small(i);
             $NEXT0();
         }
@@ -74,8 +73,7 @@ minus.fetch(Op1, Op2) {
 minus.execute(Fail, Live, Dst) {
     if (ERTS_LIKELY(is_both_small(MinusOp1, MinusOp2))) {
         Sint i = signed_val(MinusOp1) - signed_val(MinusOp2);
-        ASSERT(MY_IS_SSMALL(i) == IS_SSMALL(i));
-        if (ERTS_LIKELY(MY_IS_SSMALL(i))) {
+        if (ERTS_LIKELY(IS_SSMALL(i))) {
             $Dst = make_small(i);
             $NEXT0();
         }
@@ -100,8 +98,7 @@ increment.execute(IncrementVal, Live, Dst) {
     increment_val = $IncrementVal;
     if (ERTS_LIKELY(is_small(increment_reg_val))) {
         Sint i = signed_val(increment_reg_val) + increment_val;
-        ASSERT(MY_IS_SSMALL(i) == IS_SSMALL(i));
-        if (ERTS_LIKELY(MY_IS_SSMALL(i))) {
+        if (ERTS_LIKELY(IS_SSMALL(i))) {
             $Dst = make_small(i);
             $NEXT0();
         }
@@ -142,7 +139,7 @@ i_int_div(Fail, Live, Op1, Op2, Dst) {
         $BIF_ERROR_ARITY_2($Fail, BIF_intdiv_2, op1, op2);
     } else if (ERTS_LIKELY(is_both_small(op1, op2))) {
         Sint ires = signed_val(op1) / signed_val(op2);
-        if (ERTS_LIKELY(MY_IS_SSMALL(ires))) {
+        if (ERTS_LIKELY(IS_SSMALL(ires))) {
             $Dst = make_small(ires);
             $NEXT0();
         }
diff --git a/erts/emulator/beam/beam_bif_load.c b/erts/emulator/beam/beam_bif_load.c
index 7f8dc42aca..e48415ecc4 100644
--- a/erts/emulator/beam/beam_bif_load.c
+++ b/erts/emulator/beam/beam_bif_load.c
@@ -259,7 +259,7 @@ struct m {
     Binary* code;
     Eterm module;
     Module* modp;
-    Uint exception;
+    Eterm exception;
 };
 
 static Eterm staging_epilogue(Process* c_p, int, Eterm res, int, struct m*, int, int);
@@ -278,7 +278,7 @@ exception_list(Process* p, Eterm tag, struct m* mp, Sint exceptions)
     Eterm res = NIL;
 
     while (exceptions > 0) {
-	if (mp->exception) {
+	if (is_value(mp->exception)) {
 	    res = CONS(hp, mp->module, res);
 	    hp += 2;
 	    exceptions--;
@@ -379,9 +379,9 @@ finish_loading_1(BIF_ALIST_1)
 
     exceptions = 0;
     for (i = 0; i < n; i++) {
-	p[i].exception = 0;
+	p[i].exception = THE_NON_VALUE;
 	if (p[i].modp->seen) {
-	    p[i].exception = 1;
+	    p[i].exception = am_duplicated;
 	    exceptions++;
 	}
 	p[i].modp->seen = 1;
@@ -415,9 +415,9 @@ finish_loading_1(BIF_ALIST_1)
 
     exceptions = 0;
     for (i = 0; i < n; i++) {
-	p[i].exception = 0;
+	p[i].exception = THE_NON_VALUE;
 	if (p[i].modp->curr.code_hdr && p[i].modp->old.code_hdr) {
-	    p[i].exception = 1;
+	    p[i].exception = am_not_purged;
 	    exceptions++;
 	}
     }
@@ -438,7 +438,7 @@ finish_loading_1(BIF_ALIST_1)
 	    retval = erts_finish_loading(p[i].code, BIF_P, 0, &mod);
 	    ASSERT(retval == NIL || retval == am_on_load);
 	    if (retval == am_on_load) {
-		p[i].exception = 1;
+		p[i].exception = am_on_load;
 		exceptions++;
 	    }
 	}
@@ -469,7 +469,8 @@ staging_epilogue(Process* c_p, int commit, Eterm res, int is_blocking,
 	    erts_commit_staging_code_ix();
 
 	    for (i=0; i < nmods; i++) {
-		if (mods[i].modp->curr.code_hdr) {
+		if (mods[i].modp->curr.code_hdr
+                    && mods[i].exception != am_on_load) {
 		    set_default_trace_pattern(mods[i].module);
 		}
 	      #ifdef HIPE
@@ -686,6 +687,7 @@ BIF_RETTYPE delete_module_1(BIF_ALIST_1)
 	Eterm retval;
 	mod.module = BIF_ARG_1;
 	mod.modp = modp;
+        mod.exception = THE_NON_VALUE;
 	retval = staging_epilogue(BIF_P, success, res, is_blocking, &mod, 1, 0);
 	return retval;
     }
diff --git a/erts/emulator/beam/beam_debug.c b/erts/emulator/beam/beam_debug.c
index 4e91bfffe8..7819e9907d 100644
--- a/erts/emulator/beam/beam_debug.c
+++ b/erts/emulator/beam/beam_debug.c
@@ -629,13 +629,20 @@ print_op(fmtfn_t to, void *to_arg, int op, int size, BeamInstr* addr)
 
     unpacked = ap;
     ap = addr + size;
+
+    /*
+     * In the code below, never use ap[-1], ap[-2], ...
+     * (will not work if the arguments have been packed).
+     *
+     * Instead use unpacked[-1], unpacked[-2], ...
+     */
     switch (op) {
     case op_i_select_val_lins_xfI:
     case op_i_select_val_lins_yfI:
     case op_i_select_val_bins_xfI:
     case op_i_select_val_bins_yfI:
 	{
-	    int n = ap[-1];
+	    int n = unpacked[-1];
 	    int ix = n;
             Sint32* jump_tab = (Sint32 *)(ap + n);
 
@@ -656,7 +663,7 @@ print_op(fmtfn_t to, void *to_arg, int op, int size, BeamInstr* addr)
     case op_i_select_tuple_arity_xfI:
     case op_i_select_tuple_arity_yfI:
         {
-            int n = ap[-1];
+            int n = unpacked[-1];
             int ix = n - 1; /* without sentinel */
             Sint32* jump_tab = (Sint32 *)(ap + n);
 
@@ -698,7 +705,7 @@ print_op(fmtfn_t to, void *to_arg, int op, int size, BeamInstr* addr)
     case op_i_jump_on_val_xfIW:
     case op_i_jump_on_val_yfIW:
 	{
-	    int n = ap[-2];
+	    int n = unpacked[-2];
             Sint32* jump_tab = (Sint32 *) ap;
 
             size += (n+1) / 2;
@@ -712,7 +719,7 @@ print_op(fmtfn_t to, void *to_arg, int op, int size, BeamInstr* addr)
     case op_i_jump_on_val_zero_xfI:
     case op_i_jump_on_val_zero_yfI:
 	{
-	    int n = ap[-1];
+	    int n = unpacked[-1];
             Sint32* jump_tab = (Sint32 *) ap;
 
             size += (n+1) / 2;
diff --git a/erts/emulator/beam/big.h b/erts/emulator/beam/big.h
index 48efce20e7..7556205063 100644
--- a/erts/emulator/beam/big.h
+++ b/erts/emulator/beam/big.h
@@ -70,7 +70,20 @@ typedef Uint  dsize_t;	 /* Vector size type */
 
 /* Check for small */
 #define IS_USMALL(sgn,x)  ((sgn) ? ((x) <= MAX_SMALL+1) : ((x) <= MAX_SMALL))
-#define IS_SSMALL(x)      (((x) >= MIN_SMALL) && ((x) <= MAX_SMALL))
+
+/*
+ * It seems that both clang and gcc will generate sub-optimal code
+ * for the more obvious way to write the range check:
+ *
+ *    #define IS_SSMALL(x)  (((x) >= MIN_SMALL) && ((x) <= MAX_SMALL))
+ *
+ * Note that IS_SSMALL() may be used in the 32-bit emulator with
+ * a Uint64 argument. Therefore, we must test the size of the argument
+ * to ensure that the cast does not discard the high-order 32 bits.
+ */
+#define _IS_SSMALL32(x) (((Uint32) ((((x)) >> (SMALL_BITS-1)) + 1)) < 2)
+#define _IS_SSMALL64(x) (((Uint64) ((((x)) >> (SMALL_BITS-1)) + 1)) < 2)
+#define IS_SSMALL(x) (sizeof(x) == sizeof(Uint32) ? _IS_SSMALL32(x) : _IS_SSMALL64(x))
 
 /* The heap size needed for a bignum */
 #define BIG_NEED_SIZE(x)  ((x) + 1)
diff --git a/erts/emulator/beam/erl_arith.c b/erts/emulator/beam/erl_arith.c
index f2a3e411ec..b6625db0d3 100644
--- a/erts/emulator/beam/erl_arith.c
+++ b/erts/emulator/beam/erl_arith.c
@@ -114,7 +114,7 @@ BIF_RETTYPE intdiv_2(BIF_ALIST_2)
     }
     if (is_both_small(BIF_ARG_1,BIF_ARG_2)){
 	Sint ires = signed_val(BIF_ARG_1) / signed_val(BIF_ARG_2);
-	if (MY_IS_SSMALL(ires))
+	if (IS_SSMALL(ires))
 	    BIF_RET(make_small(ires));
     } 
     BIF_RET(erts_int_div(BIF_P, BIF_ARG_1, BIF_ARG_2));
@@ -340,8 +340,7 @@ erts_mixed_plus(Process* p, Eterm arg1, Eterm arg2)
 		switch ((arg2 & _TAG_IMMED1_MASK) >> _TAG_PRIMARY_SIZE) {
 		case (_TAG_IMMED1_SMALL >> _TAG_PRIMARY_SIZE):
 		    ires = signed_val(arg1) + signed_val(arg2);
-		    ASSERT(MY_IS_SSMALL(ires) == IS_SSMALL(ires));
-		    if (MY_IS_SSMALL(ires)) {
+		    if (IS_SSMALL(ires)) {
 			return make_small(ires);
 		    } else {
 			hp = HAlloc(p, 2);
@@ -486,8 +485,7 @@ erts_mixed_minus(Process* p, Eterm arg1, Eterm arg2)
 		switch ((arg2 & _TAG_IMMED1_MASK) >> _TAG_PRIMARY_SIZE) {
 		case (_TAG_IMMED1_SMALL >> _TAG_PRIMARY_SIZE):
 		    ires = signed_val(arg1) - signed_val(arg2);
-		    ASSERT(MY_IS_SSMALL(ires) == IS_SSMALL(ires));
-		    if (MY_IS_SSMALL(ires)) {
+		    if (IS_SSMALL(ires)) {
 			return make_small(ires);
 		    } else {
 			hp = HAlloc(p, 2);
@@ -1181,8 +1179,7 @@ erts_gc_mixed_plus(Process* p, Eterm* reg, Uint live)
 		switch ((arg2 & _TAG_IMMED1_MASK) >> _TAG_PRIMARY_SIZE) {
 		case (_TAG_IMMED1_SMALL >> _TAG_PRIMARY_SIZE):
 		    ires = signed_val(arg1) + signed_val(arg2);
-		    ASSERT(MY_IS_SSMALL(ires) == IS_SSMALL(ires));
-		    if (MY_IS_SSMALL(ires)) {
+		    if (IS_SSMALL(ires)) {
 			return make_small(ires);
 		    } else {
 			if (ERTS_NEED_GC(p, 2)) {
@@ -1349,8 +1346,7 @@ erts_gc_mixed_minus(Process* p, Eterm* reg, Uint live)
 		switch ((arg2 & _TAG_IMMED1_MASK) >> _TAG_PRIMARY_SIZE) {
 		case (_TAG_IMMED1_SMALL >> _TAG_PRIMARY_SIZE):
 		    ires = signed_val(arg1) - signed_val(arg2);
-		    ASSERT(MY_IS_SSMALL(ires) == IS_SSMALL(ires));
-		    if (MY_IS_SSMALL(ires)) {
+		    if (IS_SSMALL(ires)) {
 			return make_small(ires);
 		    } else {
 			if (ERTS_NEED_GC(p, 2)) {
diff --git a/erts/emulator/beam/erl_gc.c b/erts/emulator/beam/erl_gc.c
index 8344c164fa..97a1ca915f 100644
--- a/erts/emulator/beam/erl_gc.c
+++ b/erts/emulator/beam/erl_gc.c
@@ -337,7 +337,7 @@ erts_heap_sizes(Process* p)
 
     for (i = num_heap_sizes-1; i >= 0; i--) {
 	n += 2;
-	if (!MY_IS_SSMALL(heap_sizes[i])) {
+	if (!IS_SSMALL(heap_sizes[i])) {
 	    big += BIG_UINT_HEAP_SIZE;
 	}
     }
@@ -352,7 +352,7 @@ erts_heap_sizes(Process* p)
 	Eterm num;
 	Sint sz = heap_sizes[i];
 
-	if (MY_IS_SSMALL(sz)) {
+	if (IS_SSMALL(sz)) {
 	    num = make_small(sz);
 	} else {
 	    num = uint_to_big(sz, bigp);
diff --git a/erts/emulator/beam/erl_io_queue.c b/erts/emulator/beam/erl_io_queue.c
index a01b676d39..190ba6bbb9 100644
--- a/erts/emulator/beam/erl_io_queue.c
+++ b/erts/emulator/beam/erl_io_queue.c
@@ -973,9 +973,10 @@ static int iol2v_append_binary(iol2v_state_t *state, Eterm bin_term) {
     UWord binary_size;
 
     Uint byte_offset, bit_offset, bit_size;
+    byte *binary_data;
+
     Eterm *parent_header;
     Eterm parent_binary;
-    byte *binary_data;
 
     ASSERT(state->bytereds_available > state->bytereds_spent);
 
@@ -1001,14 +1002,14 @@ static int iol2v_append_binary(iol2v_state_t *state, Eterm bin_term) {
             erts_emasculate_writable_binary(pb);
         }
 
-        binary_data = pb->bytes;
+        binary_data = &((byte*)pb->bytes)[byte_offset];
     } else {
         ErlHeapBin *hb = (ErlHeapBin*)parent_header;
 
         ASSERT(thing_subtag(*parent_header) == HEAP_BINARY_SUBTAG);
         ASSERT(is_bin_small);
 
-        binary_data = &((unsigned char*)&hb->data)[byte_offset];
+        binary_data = &((byte*)&hb->data)[byte_offset];
     }
 
     if (!is_bin_small && (state->acc_size == 0 || !is_acc_small)) {
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index 938dd1b9ac..61fdf86a56 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -8432,11 +8432,13 @@ erts_start_schedulers(void)
 	erts_atomic_init_nob(&runq_supervisor_sleeping, 0);
 	if (0 != ethr_event_init(&runq_supervision_event))
 	    erts_exit(ERTS_ERROR_EXIT, "Failed to create run-queue supervision event\n");
-	if (0 != ethr_thr_create(&runq_supervisor_tid,
-				 runq_supervisor,
-				 NULL,
-				 &opts))
-	    erts_exit(ERTS_ERROR_EXIT, "Failed to create run-queue supervision thread\n");
+        res = ethr_thr_create(&runq_supervisor_tid,
+                              runq_supervisor,
+                              NULL,
+                              &opts);
+	if (0 != res)
+	    erts_exit(ERTS_ERROR_EXIT, "Failed to create run-queue supervision thread, "
+                      "error = %d\n", res);
 
     }
 
@@ -8471,7 +8473,7 @@ erts_start_schedulers(void)
             opts.suggested_stack_size = erts_dcpu_sched_thread_suggested_stack_size;
 	    res = ethr_thr_create(&esdp->tid,sched_dirty_cpu_thread_func,(void*)esdp,&opts);
 	    if (res != 0)
-		erts_exit(ERTS_ERROR_EXIT, "Failed to create dirty cpu scheduler thread %d\n", ix);
+		erts_exit(ERTS_ERROR_EXIT, "Failed to create dirty cpu scheduler thread %d, error = %d\n", ix, res);
 	}
 	for (ix = 0; ix < erts_no_dirty_io_schedulers; ix++) {
 	    ErtsSchedulerData *esdp = ERTS_DIRTY_IO_SCHEDULER_IX(ix);
@@ -8479,7 +8481,7 @@ erts_start_schedulers(void)
             opts.suggested_stack_size = erts_dio_sched_thread_suggested_stack_size;
 	    res = ethr_thr_create(&esdp->tid,sched_dirty_io_thread_func,(void*)esdp,&opts);
 	    if (res != 0)
-		erts_exit(ERTS_ERROR_EXIT, "Failed to create dirty io scheduler thread %d\n", ix);
+		erts_exit(ERTS_ERROR_EXIT, "Failed to create dirty io scheduler thread %d, error = %d\n", ix, res);
 	}
     }
 
@@ -8489,7 +8491,7 @@ erts_start_schedulers(void)
 
     res = ethr_thr_create(&tid, aux_thread, NULL, &opts);
     if (res != 0)
-	erts_exit(ERTS_ERROR_EXIT, "Failed to create aux thread\n");
+	erts_exit(ERTS_ERROR_EXIT, "Failed to create aux thread, error = %d\n", res);
 
     for (ix = 0; ix < erts_no_poll_threads; ix++) {
         erts_snprintf(opts.name, 16, "%d_poller", ix);
diff --git a/erts/emulator/beam/erl_term.h b/erts/emulator/beam/erl_term.h
index 842802f8d9..6daf043117 100644
--- a/erts/emulator/beam/erl_term.h
+++ b/erts/emulator/beam/erl_term.h
@@ -270,7 +270,6 @@ _ET_DECLARE_CHECKED(Eterm*,list_val,Wterm)
 #define is_byte(x)	(((x) & ((~(Uint)0 << (_TAG_IMMED1_SIZE+8)) + _TAG_IMMED1_MASK)) == _TAG_IMMED1_SMALL)
 #define is_valid_bit_size(x) (((Sint)(x)) >= 0 && ((x) & 0x7F) == _TAG_IMMED1_SMALL)
 #define is_not_valid_bit_size(x) (!is_valid_bit_size((x)))
-#define MY_IS_SSMALL(x) (((Uint) ((((x)) >> (SMALL_BITS-1)) + 1)) < 2)
 #define _unchecked_unsigned_val(x)	((x) >> _TAG_IMMED1_SIZE)
 _ET_DECLARE_CHECKED(Uint,unsigned_val,Eterm)
 #define unsigned_val(x)	_ET_APPLY(unsigned_val,(x))
diff --git a/erts/emulator/beam/external.c b/erts/emulator/beam/external.c
index 60cf09dc07..970158933f 100644
--- a/erts/emulator/beam/external.c
+++ b/erts/emulator/beam/external.c
@@ -3115,7 +3115,7 @@ dec_term(ErtsDistExternal *edep,
 #if defined(ARCH_64)
 		*objp = make_small(sn);
 #else
-		if (MY_IS_SSMALL(sn)) {
+		if (IS_SSMALL(sn)) {
 		    *objp = make_small(sn);
 		} else {
 		    *objp = small_to_big(sn, hp);
diff --git a/erts/emulator/drivers/common/efile_drv.c b/erts/emulator/drivers/common/efile_drv.c
index a1b15a2199..4e1d2f0d7f 100644
--- a/erts/emulator/drivers/common/efile_drv.c
+++ b/erts/emulator/drivers/common/efile_drv.c
@@ -1261,6 +1261,50 @@ static void free_read_line(void *data)
     EF_FREE(d);
 }
 
+void read_file_zero_size(struct t_data* d);
+#define ZERO_FILE_CHUNK (64 * 1024)
+
+/* [ERL-327] Some special files like /proc/... have reported size 0 */
+void read_file_zero_size(struct t_data* d) {
+    size_t total_read_size = 0;
+    size_t allocated_size = ZERO_FILE_CHUNK; /* allocd in invoke_read_file */
+    for (;;) {
+        size_t read_result;
+
+        /* Read until we hit EOF (read less than FILE_SEGMENT_READ) */
+        d->result_ok = efile_read(&d->errInfo,
+                                  EFILE_MODE_READ,
+                                  (int) d->fd,
+                                  (d->c.read_file.binp->orig_bytes +
+                                   total_read_size),
+                                  ZERO_FILE_CHUNK,
+                                  &read_result);
+        if (!d->result_ok) {
+            break;
+        }
+
+        total_read_size += read_result;
+        d->c.read_file.offset += read_result;
+        if (read_result < ZERO_FILE_CHUNK) {
+            break;
+        }
+
+        /* Grow before the next read call */
+        allocated_size = total_read_size + ZERO_FILE_CHUNK;
+        d->c.read_file.binp = driver_realloc_binary(d->c.read_file.binp,
+                                                    allocated_size);
+    }
+
+    /* Finalize the memory usage. Hopefully it was read fully on the first
+     * go, so the binary allocation overhead becomes:
+     * alloc ZERO_FILE_CHUNK (64kb) -> realloc real_size */
+    if (allocated_size != total_read_size) {
+        d->c.read_file.binp = driver_realloc_binary(d->c.read_file.binp,
+                                                    total_read_size);
+    }
+    d->again = 0;
+}
+
 static void invoke_read_file(void *data)
 {
     struct t_data *d = (struct t_data *) data;
@@ -1279,9 +1323,15 @@ static void invoke_read_file(void *data)
 	}
 	d->fd = fd;
 	d->c.read_file.size = (int) size;
-	if (size < 0 || size != d->c.read_file.size ||
-	    ! (d->c.read_file.binp = 
-	       driver_alloc_binary(d->c.read_file.size))) {
+
+        /* For zero sized files allocate a reasonable chunk to attempt reading
+         * anyway. Note: This will eat ZERO_FILE_CHUNK bytes for any 0 file
+         * and free them immediately after (if the file was empty). */
+        ERTS_ASSERT(size >= 0);
+        d->c.read_file.binp = driver_alloc_binary(size != 0 ? (size_t)size
+                                                            : ZERO_FILE_CHUNK);
+
+	if (size < 0 || size != d->c.read_file.size || !d->c.read_file.binp) {
 	    d->result_ok = 0;
 	    d->errInfo.posix_errno = ENOMEM;
 	    goto close;
@@ -1290,6 +1340,11 @@ static void invoke_read_file(void *data)
     }
     /* Invariant: d->c.read_file.size >= d->c.read_file.offset */
     
+    if (d->c.read_file.size == 0) {
+        read_file_zero_size(d);
+        goto close;
+    }
+
     read_size = (size_t) (d->c.read_file.size - d->c.read_file.offset);
     if (! read_size) goto close;
     chop = d->again && read_size >= FILE_SEGMENT_READ*2;
diff --git a/erts/emulator/pcre/README.pcre_update.md b/erts/emulator/pcre/README.pcre_update.md
index 8caf575d31..599e3d0d12 100644
--- a/erts/emulator/pcre/README.pcre_update.md
+++ b/erts/emulator/pcre/README.pcre_update.md
@@ -2,7 +2,7 @@
 
 ## The basic changes to the PCRE library
 
-To work with the Erlang VM, PCRE has been changed in two important ways:
+To work with the Erlang VM, PCRE has been changed in three important ways:
 
 1. The main execution machine in pcre\_exec has been modified so that
 matching can be interrupted and restarted. This functionality utilizes
diff --git a/erts/emulator/test/efile_SUITE.erl b/erts/emulator/test/efile_SUITE.erl
index f0e1bcf04b..08d5597d78 100644
--- a/erts/emulator/test/efile_SUITE.erl
+++ b/erts/emulator/test/efile_SUITE.erl
@@ -19,16 +19,20 @@
 
 -module(efile_SUITE).
 -export([all/0, suite/0]).
--export([iter_max_files/1, async_dist/1]).
+-export([async_dist/1,
+         iter_max_files/1,
+         proc_zero_sized_files/1
+        ]).
 
 -export([do_iter_max_files/2, do_async_dist/1]).
 
 -include_lib("common_test/include/ct.hrl").
+-include_lib("stdlib/include/assert.hrl").
 
 suite() -> [{ct_hooks,[ts_install_cth]}].
 
 all() -> 
-    [iter_max_files, async_dist].
+    [iter_max_files, async_dist, proc_zero_sized_files].
 
 do_async_dist(Dir) ->
     X = 100,
@@ -162,3 +166,44 @@ open_files(Name) ->
             %		  io:format("Error reason: ~p", [_Reason]),
             []
     end.
+
+%% @doc If /proc filesystem exists (no way to know if it is real proc or just
+%% a /proc directory), let's read some zero sized files 500 times each, while
+%% ensuring that response isn't empty << >>
+proc_zero_sized_files(Config) when is_list(Config) ->
+    {Type, Flavor} = os:type(),
+    %% Some files which exist on Linux but might be missing on other systems
+    Inputs = ["/proc/cpuinfo",
+              "/proc/meminfo",
+              "/proc/partitions",
+              "/proc/swaps",
+              "/proc/version",
+              "/proc/uptime",
+              %% curproc is present on freebsd
+              "/proc/curproc/cmdline"],
+    case filelib:is_dir("/proc") of
+        false -> {skip, "/proc not found"}; % skip the test if no /proc
+        _ when Type =:= unix andalso Flavor =:= sunos ->
+            %% SunOS has a /proc, but no zero sized special files
+            {skip, "sunos does not have any zero sized special files"};
+        true ->
+            %% Take away files which do not exist in proc
+            Inputs1 = lists:filter(fun filelib:is_file/1, Inputs),
+
+            %% Fail if none of mentioned files exist in /proc, did we just get
+            %% a normal /proc directory without any special files?
+            ?assertNotEqual([], Inputs1),
+
+            %% For 6 inputs and 500 attempts each this do run anywhere
+            %% between 500 and 3000 function calls.
+            lists:foreach(
+                fun(Filename) -> do_proc_zero_sized(Filename, 500) end,
+                Inputs1)
+    end.
+
+%% @doc Test one file N times to also trigger possible leaking fds and memory
+do_proc_zero_sized(_Filename, 0) -> ok;
+do_proc_zero_sized(Filename, N) ->
+    Data = file:read_file(Filename),
+    ?assertNotEqual(<<>>, Data),
+    do_proc_zero_sized(Filename, N-1).
diff --git a/erts/emulator/test/emulator_smoke.spec b/erts/emulator/test/emulator_smoke.spec
index b2d0de8835..fc98ba6823 100644
--- a/erts/emulator/test/emulator_smoke.spec
+++ b/erts/emulator/test/emulator_smoke.spec
@@ -7,3 +7,4 @@
     [consistency],"Not reliable in October and March"}.
 {cases,'Dir',crypto_SUITE,[t_md5]}.
 {cases,'Dir',float_SUITE,[fpe,cmp_integer]}.
+{cases,'Dir',erts_debug_SUITE,[df]}.
diff --git a/erts/emulator/test/iovec_SUITE.erl b/erts/emulator/test/iovec_SUITE.erl
index 28df36d293..49dc64b0d2 100644
--- a/erts/emulator/test/iovec_SUITE.erl
+++ b/erts/emulator/test/iovec_SUITE.erl
@@ -24,7 +24,8 @@
 
 -export([integer_lists/1, binary_lists/1, empty_lists/1, empty_binary_lists/1,
          mixed_lists/1, improper_lists/1, illegal_lists/1, cons_bomb/1,
-         iolist_to_iovec_idempotence/1, iolist_to_iovec_correctness/1]).
+         sub_binary_lists/1, iolist_to_iovec_idempotence/1,
+         iolist_to_iovec_correctness/1]).
 
 -include_lib("common_test/include/ct.hrl").
 
@@ -34,8 +35,8 @@ suite() ->
 
 all() ->
     [integer_lists, binary_lists, empty_lists, empty_binary_lists, mixed_lists,
-     illegal_lists, improper_lists, cons_bomb, iolist_to_iovec_idempotence,
-     iolist_to_iovec_correctness].
+     sub_binary_lists, illegal_lists, improper_lists, cons_bomb,
+     iolist_to_iovec_idempotence, iolist_to_iovec_correctness].
 
 init_per_suite(Config) ->
     Config.
@@ -46,15 +47,16 @@ end_per_suite(Config) ->
 
 integer_lists(Config) when is_list(Config) ->
     Variations = gen_variations([I || I <- lists:seq(1, 255)]),
+    equivalence_test(fun erlang:iolist_to_iovec/1, Variations).
 
-    equivalence_test(fun erlang:iolist_to_iovec/1, Variations),
-
-    ok.
+sub_binary_lists(Config) when is_list(Config) ->
+    Parent = <<0:256/unit:8, "gazurka">>,
+    <<0:196/unit:8, Child/binary>> = Parent,
+    equivalence_test(fun erlang:iolist_to_iovec/1, gen_variations(Child)).
 
 binary_lists(Config) when is_list(Config) ->
     Variations = gen_variations([<<I:8>> || I <- lists:seq(1, 255)]),
-    equivalence_test(fun erlang:iolist_to_iovec/1, Variations),
-    ok.
+    equivalence_test(fun erlang:iolist_to_iovec/1, Variations).
 
 empty_lists(Config) when is_list(Config) ->
     Variations = gen_variations([[] || _ <- lists:seq(1, 256)]),
@@ -70,8 +72,7 @@ empty_binary_lists(Config) when is_list(Config) ->
 
 mixed_lists(Config) when is_list(Config) ->
     Variations = gen_variations([<<>>, lists:seq(1, 40), <<12, 45, 78>>]),
-    equivalence_test(fun erlang:iolist_to_iovec/1, Variations),
-    ok.
+    equivalence_test(fun erlang:iolist_to_iovec/1, Variations).
 
 illegal_lists(Config) when is_list(Config) ->
     BitStrs = gen_variations(["gurka", <<1:1>>, "gaffel"]),
@@ -82,18 +83,15 @@ illegal_lists(Config) when is_list(Config) ->
     Variations =
         BitStrs ++ BadInts ++ Atoms ++ BadTails,
 
-    illegality_test(fun erlang:iolist_to_iovec/1, Variations),
-
-    ok.
+    illegality_test(fun erlang:iolist_to_iovec/1, Variations).
 
 improper_lists(Config) when is_list(Config) ->
     Variations = [
         [[[[1 | <<2>>] | <<3>>] | <<4>>] | <<5>>],
-        [[<<"test">>, 3] | <<"improper tail">>],
-        [1, 2, 3 | <<"improper tail">>]
+        [[<<1>>, 2] | <<3, 4, 5>>],
+        [1, 2, 3 | <<4, 5>>]
     ],
-    equivalence_test(fun erlang:iolist_to_iovec/1, Variations),
-    ok.
+    equivalence_test(fun erlang:iolist_to_iovec/1, Variations).
 
 cons_bomb(Config) when is_list(Config) ->
     IntBase = gen_variations([I || I <- lists:seq(1, 255)]),
@@ -108,8 +106,7 @@ cons_bomb(Config) when is_list(Config) ->
         end,
 
     Variations = gen_variations([IntBase, BinBase, MixBase], Rounds),
-    equivalence_test(fun erlang:iolist_to_iovec/1, Variations),
-    ok.
+    equivalence_test(fun erlang:iolist_to_iovec/1, Variations).
 
 iolist_to_iovec_idempotence(Config) when is_list(Config) ->
     IntVariations = gen_variations([I || I <- lists:seq(1, 255)]),
@@ -134,11 +131,15 @@ iolist_to_iovec_correctness(Config) when is_list(Config) ->
     ok.
 
 illegality_test(Fun, Variations) ->
-    [{'EXIT',{badarg, _}} = (catch Fun(Variation)) || Variation <- Variations].
+    [{'EXIT',{badarg, _}} = (catch Fun(Variation)) || Variation <- Variations],
+    ok.
 
 equivalence_test(Fun, [Head | _] = Variations) ->
+    %% Check that each variation is equal to the others, and that the sum of
+    %% them is equal to the input.
     Comparand = Fun(Head),
-    [is_iolist_equal(Comparand, Fun(Variation)) || Variation <- Variations],
+    [true = is_iolist_equal(Comparand, Fun(V)) || V <- Variations],
+    true = is_iolist_equal(Variations, Fun(Variations)),
     ok.
 
 is_iolist_equal(A, B) ->
diff --git a/erts/emulator/utils/beam_makeops b/erts/emulator/utils/beam_makeops
index bb31db7eb5..a9b2c8861c 100755
--- a/erts/emulator/utils/beam_makeops
+++ b/erts/emulator/utils/beam_makeops
@@ -80,7 +80,10 @@ my %gen_opnum;
 my %num_specific;
 my %gen_to_spec;
 my %specific_op;
-my %group_size;                 # Group size for specific operators.
+
+# Information about each specific operator. Key is the print name (e.g. get_list_xxy).
+# Value is a hash.
+my %spec_op_info;
 
 my %gen_arity;
 my @gen_arity;
@@ -523,6 +526,37 @@ sub emulator_output {
     my $key;			# Loop variable.
 
     #
+    # Generate code and meta information for all instructions.
+    #
+    foreach $key (keys %specific_op) {
+	foreach (@{$specific_op{$key}}) {
+	    my($name, $hotness, @args) = @$_;
+	    my $sign = join('', @args);
+	    my $print_name = print_name($name, @args);
+
+	    my($size, $code, $pack_spec) = cg_basic($name, @args);
+	    if (defined $code) {
+                $code = "OpCase($print_name):\n$code";
+                push @generated_code, [$hotness,$code,($print_name)];
+	    }
+
+            # Note: Some of the information below will be modified
+            # for combined instructions.
+            my %info = ('size' => $size,
+                        'pack_spec' => $pack_spec,
+                        'adj' => 0,
+                        'args' => \@args);
+            $spec_op_info{$print_name} = \%info;
+	}
+    }
+
+    #
+    # Combine micro instruction into instruction blocks and generate
+    # code for them.
+    #
+    combine_micro_instructions();
+
+    #
     # Information about opcodes (beam_opcodes.c).
     #
     $name = "$outdir/beam_opcodes.c";
@@ -551,14 +585,9 @@ sub emulator_output {
     print "\n";
 
     #
-    # Combine micro instruction into instruction blocks.
-    #
-    combine_micro_instructions();
-
-    #
     # Generate code for specific ops.
     #
-    my($spec_opnum) = 0;
+    my $spec_opnum = 0;
     print "const OpEntry opc[] = {\n";
     foreach $key (sort keys %specific_op) {
 	$gen_to_spec{$key} = $spec_opnum;
@@ -576,35 +605,21 @@ sub emulator_output {
 
 	    # The primitive types should sort before other types.
 
-	    my($sort_key) = $sign;
+	    my $sort_key = $sign;
 	    eval "\$sort_key =~ tr/$genop_types/./";
 	    $sort_key .= ":$sign";
-	    $items{$sort_key} = [$name, $hot, $sign, @args];
+            my $print_name = print_name($name, @args);
+	    $items{$sort_key} = $print_name;
 	}
 
 	#
 	# Now call the generator for the sorted result.
 	#
-	foreach (sort keys %items) {
-	    my($name, $hot, $sign, @args) = @{$items{$_}};
+	foreach my $sort_key (sort keys %items) {
+            my $print_name = $items{$sort_key};
+            my $info = $spec_op_info{$print_name};
+            my(@args) = @{$info->{'args'}};
 	    my $arity = @args;
-	    my($instr) = "${name}_$sign";
-	    $instr =~ s/_$//;
-
-	    #
-	    # Call a generator to calculate size and generate macros
-	    # for the emulator.
-	    #
-	    my($size, $code, $pack) =
-                basic_generator($name, 1, '', 0, undef, @args);
-
-	    #
-	    # Save the generated $code for later.
-	    #
-	    if (defined $code) {
-                $code = "OpCase($instr):\n$code";
-                push @generated_code, [$hot,$code,($instr)];
-	    }
 
 	    #
 	    # Calculate the bit mask which should be used to match this
@@ -626,7 +641,6 @@ sub emulator_output {
 	    }
 
 	    printf "/* %3d */  ", $spec_opnum;
-	    my $print_name = $sign ne '' ? "${name}_$sign" : $name;
 	    my $init = "{";
 	    my $sep = "";
 	    foreach (@bits) {
@@ -634,12 +648,12 @@ sub emulator_output {
 		$sep = ",";
 	    }
 	    $init .= "}";
-            my $adj = 0;
-            if (defined $group_size{$print_name}) {
-                $adj = $size - $group_size{$print_name};
-            }
-	    init_item($print_name, $init, $involves_r, $size, $adj, $pack, $sign);
-	    $op_to_name[$spec_opnum] = $instr;
+            my $adj = $info->{'adj'};
+            my $size = $info->{'size'};
+            my $pack_spec = $info->{'pack_spec'};
+            my $sign = join '', @args;
+	    init_item($print_name, $init, $involves_r, $size, $adj, $pack_spec, $sign);
+	    $op_to_name[$spec_opnum] = $print_name;
 	    $spec_opnum++;
 	}
     }
@@ -835,6 +849,12 @@ sub emulator_output {
     print_code(COLD);
 }
 
+sub print_name {
+    my($name,@args) = @_;
+    my $sign = join '', @args;
+    $sign ne '' ? "${name}_$sign" : $name;
+}
+
 sub init_item {
     my($sep) = "";
 
@@ -1108,8 +1128,9 @@ sub combine_instruction_group {
             my $offset = 0;
             my @rest = @args;
             my @new_subs;
-            my $opcase = $specific;
-            $opcase .= "_" . join '', @args if @args;
+            my $print_name = print_name($specific, @args);
+            my $opcase = $print_name;
+            my $last = $subs[$#subs];
             foreach my $s (@subs) {
                 my $code = $c_code{$s};
                 my(undef,undef,@c_args) = @{$code};
@@ -1117,7 +1138,7 @@ sub combine_instruction_group {
                 foreach (0..$#c_args) {
                     push @first, shift @rest;
                 }
-                my($size,undef) = basic_generator($s, 0, '', 0, undef, @first);
+                my $size = cg_combined_size($s, 1, @first);
                 $offsets{$s} = $offset
                     unless defined $offsets{$s} and $offsets{$s} >= $offset;
                 $offset += $size - 1;
@@ -1126,6 +1147,7 @@ sub combine_instruction_group {
                 push @new_subs, [$opcase,$label,$s,$size-1,@first];
                 $opcase = '';
             }
+            $spec_op_info{$print_name}->{'size'} = $offset + 1;
             $group_size = $offset if $group_size < $offset;
             push @instrs, [$specific_key,@new_subs];
         }
@@ -1162,6 +1184,8 @@ sub combine_instruction_group {
     # Now generate the code for the entire group.
     my $offset = 0;
     my @opcase_labels;
+    my %down;
+    my %up;
     for(my $i = 0; $i < @slots; $i++) {
         my $key = $slots[$i];
 
@@ -1182,11 +1206,11 @@ sub combine_instruction_group {
             my $seen_key = "$label:$next:" . scalar(@first);
             next if $opcase eq '' and $seen{$seen_key};
             $seen{$seen_key} = 1;
+            $seen_key .= $opcase;
 
             if ($opcase ne '') {
                 $gcode .= "OpCase($opcase):\n";
                 push @opcase_labels, $opcase;
-                $group_size{$opcase} = $group_size + 1;
             }
             if ($num_references{$label}) {
                 $gcode .= "$label:\n";
@@ -1204,14 +1228,42 @@ sub combine_instruction_group {
                 $transfer_to_next .= "goto $next;\n\n";
             }
 
-            my(undef,$gen_code) =
-                basic_generator($s, 0, $flags, $offset,
-                                $group_size-$offset-$dec, @first);
+            my($gen_code,$down,$up) =
+                cg_combined_code($s, 1, $flags, $offset,
+                                 $group_size-$offset-$dec, @first);
+            my $spec_label = "$opcase$label";
+            $down{$spec_label} = $down;
+            $up{$spec_label} = $up;
             $gcode .= $gen_code . $transfer_to_next;
         }
         $offset = $order_to_offset{$slots[$i+1]} if $i < $#slots;
     }
 
+    foreach my $print_name (@opcase_labels) {
+        my $info = $spec_op_info{$print_name};
+        $info->{'adj'} = $info->{'size'} - $group_size - 1;
+    }
+
+    #
+    # Assemble pack specifications for all instructions in the group.
+    #
+    foreach my $instr (@instrs) {
+        my(undef,@subs) = @{$instr};
+        my $down = '';
+        my $up = '';
+        for (my $i = 0; $i < @subs; $i++) {
+            my($opcase,$label) = @{$subs[$i]};
+            my $spec_label = "$opcase$label";
+            if (defined $down{$spec_label}) {
+                $down = $down{$spec_label} . $down;
+                $up = $up . $up{$spec_label};
+            }
+        }
+        my $print_name = $subs[0]->[0];
+        my $info = $spec_op_info{$print_name};
+        $info->{'pack_spec'} = build_pack_spec("$down:$up");
+    }
+
     ($group_hotness,"{\n$gcode\n}\n\n",@opcase_labels);
 }
 
@@ -1223,12 +1275,42 @@ sub micro_label {
 
 
 #
-# Basic implementation of instruction in emulator loop
-# (assuming no packing).
+# Basic code generation for one instruction.
 #
 
-sub basic_generator {
-    my($name,$hot,$extra_comments,$offset,$group_size,@args) = @_;
+sub cg_basic {
+    my($name,@args) = @_;
+    my($size,$code,$pack_spec) = code_gen($name, 1, '', 0, undef, @args);
+    $pack_spec = build_pack_spec($pack_spec);
+    ($size,$code,$pack_spec);
+}
+
+#
+# Calculate size for a micro instruction.
+#
+
+sub cg_combined_size {
+    my($name,$pack,@args) = @_;
+    my($size) = code_gen($name, $pack, '', 0, undef, @args);
+    $size;
+}
+
+#
+# Generate code for a micro instruction.
+#
+
+sub cg_combined_code {
+    my($size,$code,$pack_spec) = code_gen(@_);
+    if ($pack_spec eq '') {
+        ($code,'','');
+    } else {
+        my($down,$up) = split /:/, $pack_spec;
+        ($code,$down,$up);
+    }
+}
+
+sub code_gen {
+    my($name,$pack,$extra_comments,$offset,$group_size,@args) = @_;
     my $size = 0;
     my $flags = '';
     my @f;
@@ -1242,8 +1324,8 @@ sub basic_generator {
     #
 
     my $c_code_ref = $c_code{$name};
-    if ($hot and defined $c_code_ref and $name ne 'catch') {
-        ($var_decls, $pack_spec, @args) = do_pack(@args);
+    if ($pack and defined $c_code_ref and $name ne 'catch') {
+        ($var_decls, $pack_spec, @args) = do_pack($offset, @args);
     }
 
     #
@@ -1519,7 +1601,7 @@ sub needs_do_wrapper {
 }
 
 sub do_pack {
-    my(@args) = @_;
+    my($offset,@args) = @_;
     my($packable_args) = 0;
     my @bits_needed;            # Bits needed for each argument.
 
@@ -1569,7 +1651,7 @@ sub do_pack {
     #
     # Nothing to pack unless there are at least 2 packable arguments.
     #
-    return ('', '', @args) if $packable_args < 2;
+    return ('', ':', @args) if $packable_args < 2;
 
     #
     # Determine how many arguments we should pack into each word.
@@ -1644,13 +1726,12 @@ sub do_pack {
 				# beginning).
     my $up = '';		# Pack commands (storing back while
 				# moving forward).
-    my $did_some_packing = 0;   # Nothing packed yet.
 
     # Skip an unpackable argument.
     my $skip_unpackable = sub {
         my($arg) = @_;
 
-        if ($arg_size{$arg} and $did_some_packing) {
+        if ($arg_size{$arg}) {
             # Save the argument on the pack engine's stack.
             my $push = 'g';
             if ($type_bit{$arg} & $type_bit{'q'}) {
@@ -1662,11 +1743,6 @@ sub do_pack {
             }
             $down = "$push${down}";
             $up = "${up}p";
-        } else {
-            # The argument has either zero size (e.g. r(0)),
-            # or is to the left of the first packed argument
-            # and will never be accessed. No need to do
-            # anything.
         }
     };
 
@@ -1700,11 +1776,10 @@ sub do_pack {
             my $this_size = $arg_size{$reg};
             if ($bits_needed[$arg_num]) {
                 $this_size = 0;
-                $did_some_packing = 1;
 
                 if ($ap == 0) {
                     $pack_prefix .= "Eterm $packed_var = " .
-                        arg_offset($size) . ";\n";
+                        arg_offset($size+$offset) . ";\n";
                     $up .= "p";
                     $down = "P$down";
                     $this_size = 1;
@@ -1731,7 +1806,7 @@ sub do_pack {
         $arg_num++;
     }
 
-    my $pack_spec = $down . $up;
+    my $pack_spec = "$down:$up";
     return ($pack_prefix, $pack_spec, @args);
 }
 
@@ -1744,6 +1819,17 @@ sub make_unpack {
     $e;
 }
 
+sub build_pack_spec {
+    my $pack_spec = shift;
+    return '' if $pack_spec eq '';
+    my($down,$up) = split /:/, $pack_spec;
+    while ($down =~ /[gfq]$/ and $up =~ /^p/) {
+        $down = substr($down, 0, -1);
+        $up = substr($up, 1);
+    }
+    "$down$up";
+}
+
 sub quote {
     local($_) = @_;
     return "'$_'" if $_ eq 'try';