diff --git a/parmys/parmys-plugin/core/subtractor.cc b/parmys/parmys-plugin/core/subtractor.cc
index 03c259f95ad..18a09cd51fd 100644
--- a/parmys/parmys-plugin/core/subtractor.cc
+++ b/parmys/parmys-plugin/core/subtractor.cc
@@ -501,7 +501,7 @@ void split_adder_for_sub(nnode_t *nodeo, int a, int b, int sizea, int sizeb, int
 
     if ((flag == 0 || count > 1) && !configuration.adder_cin_global) {
         // connect the a[0] of first adder node to ground, and b[0] of first adder node to vcc
-        connect_nodes(netlist->gnd_node, 0, node[0], 0);
+        connect_nodes(netlist->vcc_node, 0, node[0], 0);
         connect_nodes(netlist->vcc_node, 0, node[0], sizea);
         // hang the first sumout
         node[0]->output_pins[1] = allocate_npin();
@@ -516,14 +516,25 @@ void split_adder_for_sub(nnode_t *nodeo, int a, int b, int sizea, int sizeb, int
 
     // for normal subtraction: if any input pins beside intial cin is NULL, it should connect to unconn
     // for unary subtraction: the first number should has the number of a input pins connected to gnd. The others are as same as normal subtraction
+    int tail = count -1;
     for (int i = 0; i < count; i++) {
         num = node[i]->num_input_pins;
         for (int j = 0; j < num - 1; j++) {
             if (node[i]->input_pins[j] == NULL) {
                 if (nodeo->num_input_port_sizes != 3 && i * sizea + j < a)
                     connect_nodes(netlist->gnd_node, 0, node[i], j);
-                else
-                    connect_nodes(netlist->pad_node, 0, node[i], j);
+                else{
+                    if (i == count - 1){
+                        if (j == 0){
+                            connect_nodes(netlist->gnd_node, 0, node[i], j);
+                        }
+                        else if (j == 1)
+                            connect_nodes(netlist->vcc_node, 0, node[i], j);
+                    }
+                    else{
+                        connect_nodes(netlist->pad_node, 0, node[i], j);
+                    }
+                }
             }
         }
     }
@@ -554,12 +565,15 @@ void split_adder_for_sub(nnode_t *nodeo, int a, int b, int sizea, int sizeb, int
         }
     }
 
+
     if (count > 1 || configuration.adder_cin_global) {
         // remap the output pins of each adder to nodeo
         for (int i = offset; i < count; i++) {
             for (int j = 0; j < node[i]->num_output_pins - 1; j++) {
-                if ((i * sizea + j - offset) < nodeo->num_output_pins)
+                if ((i * sizea + j - offset) < nodeo->num_output_pins){
                     remap_pin_to_new_node(nodeo->output_pins[i * sizea + j - offset], node[i], j + 1);
+                    nodeo->output_pins[i * sizea + j - offset] = NULL;
+                }
                 else {
                     node[i]->output_pins[j + 1] = allocate_npin();
                     // Pad outputs with a unique and descriptive name to avoid collisions.
@@ -573,7 +587,7 @@ void split_adder_for_sub(nnode_t *nodeo, int a, int b, int sizea, int sizeb, int
     node[count - 1]->output_pins[0]->name = append_string("", "%s~dummy_output~%d~%d", node[(count - 1)]->name, (count - 1), 0);
     // connect_nodes(node[count - 1], (node[(count - 1)]->num_output_pins - 1), netlist->gnd_node, 0);
     // }
-
+    
     /* Freeing the old node! */
     cleanup_sub_old_node(nodeo, netlist);
 
@@ -581,7 +595,6 @@ void split_adder_for_sub(nnode_t *nodeo, int a, int b, int sizea, int sizeb, int
     vtr::free(not_node);
     return;
 }
-
 /*-------------------------------------------------------------------------
  * (function: iterate_adders_for_sub)
  *
@@ -628,27 +641,51 @@ void iterate_adders_for_sub(netlist_t *netlist)
 
             if (num >= min_threshold_adder) {
                 // how many subtractors base on a can split
-                if ((a + 1) % sizea == 0)
-                    counta = (a + offset) / sizea;
-                else
-                    counta = (a + 1) / sizea + 1;
-                // how many subtractors base on b can split
-                if ((b + 1) % sizeb == 0)
-                    countb = (b + offset) / sizeb;
-                else
-                    countb = (b + 1) / sizeb + 1;
-                // how many subtractors need to be split
-                if (counta >= countb)
-                    count = counta;
+                if (num >= min_threshold_adder && num >= min_add) {
+                    // if the first cin in a chain is fed by a global input (offset = 0) the adder width is the
+                    // input width + 1 (to pass the last cout -> sumout) divided by size of the adder input ports
+                    // otherwise (offset = 1) a dummy adder is added to the chain to feed the first cin with gnd
+                    // how many adders a can split
+                    counta = (a + 1) / sizea + offset;
+                    // how many adders b can split
+                    countb = (b + 1) / sizeb + offset;
+                    // how many adders need to be split
+                    if (counta >= countb)
+                        count = counta;
+                    else
+                        count = countb;
+                    subchaintotal++;
+                    split_adder_for_sub(node, a, b, sizea, sizeb, 1, 1, count, netlist);
+                }
+                // Store the node into processed_adder_list if the threshold is bigger than num
                 else
-                    count = countb;
-                subchaintotal++;
-
-                split_adder_for_sub(node, a, b, sizea, sizeb, 1, 1, count, netlist);
+                    processed_adder_list = insert_in_vptr_list(processed_adder_list, node);
             }
-            // Store the node into processed_adder_list if the threshold is bigger than num
             else
-                processed_adder_list = insert_in_vptr_list(processed_adder_list, node);
+                    processed_adder_list = insert_in_vptr_list(processed_adder_list, node);
+            // if (num >= min_threshold_adder) {
+            //     // how many subtractors base on a can split
+            //     if ((a + 1) % sizea == 0)
+            //         counta = (a + offset) / sizea;
+            //     else
+            //         counta = (a + 1) / sizea + 1;
+            //     // how many subtractors base on b can split
+            //     if ((b + 1) % sizeb == 0)
+            //         countb = (b + offset) / sizeb;
+            //     else
+            //         countb = (b + 1) / sizeb + 1;
+            //     // how many subtractors need to be split
+            //     if (counta >= countb)
+            //         count = counta;
+            //     else
+            //         count = countb;
+            //     subchaintotal++;
+
+            //     split_adder_for_sub(node, a, b, sizea, sizeb, 1, 1, count, netlist);
+            // }
+            // // Store the node into processed_adder_list if the threshold is bigger than num
+            // else
+            //     processed_adder_list = insert_in_vptr_list(processed_adder_list, node);
         }
     }
 
diff --git a/parmys/parmys-plugin/netlist/netlist_utils.cc b/parmys/parmys-plugin/netlist/netlist_utils.cc
index 9c3fb060b4f..b2092d81f56 100644
--- a/parmys/parmys-plugin/netlist/netlist_utils.cc
+++ b/parmys/parmys-plugin/netlist/netlist_utils.cc
@@ -1392,7 +1392,7 @@ void equalize_ports_size(nnode_t *&node, uintptr_t traverse_mark_number, netlist
         return;
 
     /* new port size */
-    int new_out_size = port_a_size;
+    int new_out_size = port_y_size;
 
     /* creating the new node */
     nnode_t *new_node = (port_b_size == -1) ? make_1port_gate(node->type, port_a_size, new_out_size, node, traverse_mark_number)
@@ -1469,4 +1469,4 @@ void delete_npin(npin_t *pin)
     }
     // CLEAN UP
     free_npin(pin);
-}
\ No newline at end of file
+}
diff --git a/parmys/parmys-plugin/parmys.cc b/parmys/parmys-plugin/parmys.cc
index 9e0338586e7..d03181ae7ca 100644
--- a/parmys/parmys-plugin/parmys.cc
+++ b/parmys/parmys-plugin/parmys.cc
@@ -817,6 +817,17 @@ struct ParMYSPass : public Pass {
         log("\nTechmap Time: ");
         log_time(techmap_time);
         log("\n--------------------------------------------------------------------\n");
+
+        fprintf(stderr, "[BLIF-OUT] POs=%d\n", odin_netlist->num_top_output_nodes);
+        for (int k=0; k<odin_netlist->num_top_output_nodes; ++k) {
+            npin_t* in = odin_netlist->top_output_nodes[k]->input_pins[0];
+            fprintf(stderr, "  PO[%d]: net=%s driver=%s.pin%d\n",
+                k,
+                in->net && in->net->name ? in->net->name : "(noname)",
+                (in->net && in->net->driver_pins[0]->node &&
+                in->net->driver_pins[0]->node->name) ? in->net->driver_pins[0]->node->name : "(null)",
+                in->net ? in->net->driver_pins[0]->pin_node_idx : -1);
+        }
     }
 
     static void report(netlist_t *odin_netlist)
diff --git a/vtr_flow/misc/yosys/synthesis.tcl b/vtr_flow/misc/yosys/synthesis.tcl
index 61d0c4a0bb6..6104ae4e303 100644
--- a/vtr_flow/misc/yosys/synthesis.tcl
+++ b/vtr_flow/misc/yosys/synthesis.tcl
@@ -80,12 +80,16 @@ techmap -map +/parmys/aldffe2dff.v
 
 opt -full
 
+write_verilog -noexpr pre_parmys.v
+
 # Separate options for Parmys execution (Verilog or SystemVerilog)
 if {$env(PARSER) == "default" || $env(PARSER) == "slang"} {
     # For Verilog, use -nopass for a simpler, faster flow
     parmys -a QQQ -nopass -c CCC YYY
 } 
 
+write_verilog -noexpr post_parmys.v
+
 opt -full
 
 techmap 
@@ -99,4 +103,6 @@ stat
 
 hierarchy -check -auto-top -purge_lib
 
+write_verilog -noexpr post_everything.v
+
 write_blif -true + vcc -false + gnd -undef + unconn -blackbox ZZZ