diff --git a/compiler/base/hierarchy_layout.py b/compiler/base/hierarchy_layout.py
index d350c156..0cbe6e8c 100644
--- a/compiler/base/hierarchy_layout.py
+++ b/compiler/base/hierarchy_layout.py
@@ -63,6 +63,16 @@ class layout():
         self.translate_all(offset)
         return offset
 
+    def offset_x_coordinates(self):
+        """
+        This function is called after everything is placed to
+        shift the origin to the furthest left point. 
+        Y offset is unchanged.
+        """
+        offset = self.find_lowest_coords()
+        self.translate_all(offset.scale(1, 0))
+        return offset
+    
     def get_gate_offset(self, x_offset, height, inv_num):
         """
         Gets the base offset and y orientation of stacked rows of gates
diff --git a/compiler/modules/bank.py b/compiler/modules/bank.py
index 7bddc60a..5dfaa007 100644
--- a/compiler/modules/bank.py
+++ b/compiler/modules/bank.py
@@ -223,10 +223,10 @@ class bank(design.design):
 
         # UPPER LEFT QUADRANT
         # To the left of the bitcell array above the predecoders and control logic
-        x_offset = self.m2_gap + self.port_address.width
+        x_offset = self.m2_gap + self.port_address[port].width
         self.port_address_offsets[port] = vector(-x_offset,
                                                  self.main_bitcell_array_bottom)
-        self.predecoder_height = self.port_address.predecoder_height + self.port_address_offsets[port].y
+        self.predecoder_height = self.port_address[port].predecoder_height + self.port_address_offsets[port].y
 
         # LOWER LEFT QUADRANT
         # Place the col decoder left aligned with wordline driver
@@ -234,7 +234,7 @@ class bank(design.design):
         # control logic to allow control signals to easily pass over in M3
         # by placing 1 1/4 a cell pitch down because both power connections and inputs/outputs
         # may be routed in M3 or M4
-        x_offset = self.central_bus_width[port] + self.port_address.wordline_driver.width
+        x_offset = self.central_bus_width[port] + self.port_address[port].wordline_driver_array.width
         if self.col_addr_size > 0:
             x_offset += self.column_decoder.width + self.col_addr_bus_width
             y_offset = 1.25 * self.dff.height + self.column_decoder.height
@@ -267,7 +267,7 @@ class bank(design.design):
             
         # LOWER RIGHT QUADRANT
         # To the right of the bitcell array
-        x_offset = self.bitcell_array_right + self.port_address.width + self.m2_gap
+        x_offset = self.bitcell_array_right + self.port_address[port].width + self.m2_gap
         self.port_address_offsets[port] = vector(x_offset,
                                                  self.main_bitcell_array_bottom)
 
@@ -278,7 +278,7 @@ class bank(design.design):
         # control logic to allow control signals to easily pass over in M3
         # by placing 1 1/4 a cell pitch down because both power connections and inputs/outputs
         # may be routed in M3 or M4
-        x_offset = self.bitcell_array_right  + self.central_bus_width[port] + self.port_address.wordline_driver.width
+        x_offset = self.bitcell_array_right  + self.central_bus_width[port] + self.port_address[port].wordline_driver_array.width
         if self.col_addr_size > 0:
             x_offset += self.column_decoder.width + self.col_addr_bus_width
             y_offset = self.bitcell_array_top + 1.25 * self.dff.height + self.column_decoder.height
@@ -366,12 +366,13 @@ class bank(design.design):
     def add_modules(self):
         """ Add all the modules using the class loader """
 
-        self.port_address = factory.create(module_type="port_address",
-                                           cols=self.num_cols + self.num_spare_cols,
-                                           rows=self.num_rows)
-        self.add_mod(self.port_address)
-
-        self.num_rbl = len(self.all_ports)
+        self.port_address = []
+        for port in self.all_ports:
+            self.port_address.append(factory.create(module_type="port_address",
+                                                    cols=self.num_cols + self.num_spare_cols,
+                                                    rows=self.num_rows,
+                                                    port=port))
+            self.add_mod(self.port_address[port])
 
         try:
             local_array_size = OPTS.local_array_size
@@ -420,13 +421,10 @@ class bank(design.design):
         # gnd
         temp = self.bitcell_array.get_inouts()
 
-        wordline_names = self.bitcell_array.get_inputs()
-        
-        # Rename the RBL WL to the enable name
-        for port in self.all_ports:
-            rbl_wl_name = self.bitcell_array.get_rbl_wordline_names(port)[port]
-            wordline_names = [x.replace(rbl_wl_name, "wl_en{0}".format(port)) for x in wordline_names]
-        temp.extend(wordline_names)
+        temp.append("rbl_wl0")
+        temp.extend(self.bitcell_array.get_wordline_names())
+        if len(self.all_ports) > 1:
+            temp.append("rbl_wl1")
         
         temp.append("vdd")
         temp.append("gnd")
@@ -486,13 +484,15 @@ class bank(design.design):
         self.port_address_inst = [None] * len(self.all_ports)
         for port in self.all_ports:
             self.port_address_inst[port] = self.add_inst(name="port_address{}".format(port),
-                                                         mod=self.port_address)
+                                                         mod=self.port_address[port])
 
             temp = []
             for bit in range(self.row_addr_size):
                 temp.append("addr{0}_{1}".format(port, bit + self.col_addr_size))
             temp.append("wl_en{}".format(port))
-            temp.extend(self.bitcell_array.get_wordline_names(port))
+            wordline_names = self.bitcell_array.get_wordline_names(port)
+            temp.extend(wordline_names)
+            temp.append("rbl_wl{}".format(port))
             temp.extend(["vdd", "gnd"])
             self.connect_inst(temp)
             
@@ -845,35 +845,51 @@ class bank(design.design):
         self.route_port_address_in(port)
         
         if port % 2:
-            self.route_port_address_right(port)
+            self.route_port_address_out(port, "right")
         else:
-            self.route_port_address_left(port)
+            self.route_port_address_out(port, "left")
             
-    def route_port_address_left(self, port):
+    def route_port_address_out(self, port, side="left"):
         """ Connecting Wordline driver output to Bitcell WL connection  """
 
-        driver_names = ["wl_{}".format(x) for x in range(self.num_rows)]
-        for (driver_name, array_name) in zip(driver_names, self.bitcell_array.get_wordline_names(port)):
+        driver_names = ["wl_{}".format(x) for x in range(self.num_rows)] + ["rbl_wl"]
+        rbl_wl_name = self.bitcell_array.get_rbl_wordline_names(port)[port]
+        for (driver_name, array_name) in zip(driver_names, self.bitcell_array.get_wordline_names(port) + [rbl_wl_name]):
             # The mid guarantees we exit the input cell to the right.
             driver_wl_pin = self.port_address_inst[port].get_pin(driver_name)
-            driver_wl_pos = driver_wl_pin.rc()
+            if side == "left":
+                driver_wl_pos = driver_wl_pin.rc()
+            else:
+                driver_wl_pos = driver_wl_pin.lc()
             bitcell_wl_pin = self.bitcell_array_inst.get_pin(array_name)
-            bitcell_wl_pos = bitcell_wl_pin.lc()
-            mid1 = driver_wl_pos.scale(0, 1) + vector(0.5 * self.port_address_inst[port].rx() + 0.5 * self.bitcell_array_inst.lx(), 0)
-            mid2 = mid1.scale(1, 0) + bitcell_wl_pos.scale(0.5, 1)
-            self.add_path(driver_wl_pin.layer, [driver_wl_pos, mid1, mid2])
-            # Via is non-preferred direction because mid1->mid2 is non-preferred direction
-            self.add_via_stack_center(from_layer=driver_wl_pin.layer,
-                                      to_layer=bitcell_wl_pin.layer,
-                                      offset=mid2,
-                                      directions="nonpref")
-            self.add_path(bitcell_wl_pin.layer, [mid2, bitcell_wl_pos])
+            
+            if side == "left":
+                bitcell_wl_pos = bitcell_wl_pin.lc()
+                port_address_pos = self.port_address_inst[port].rx()
+                bitcell_array_pos = self.bitcell_array_inst.lx()
+            else:
+                bitcell_wl_pos = bitcell_wl_pin.rc()
+                port_address_pos = self.port_address_inst[port].lx()
+                bitcell_array_pos = self.bitcell_array_inst.rx()
+                
+            mid1 = driver_wl_pos.scale(0, 1) + vector(0.5 * port_address_pos + 0.5 * bitcell_array_pos, 0)
+            mid2 = mid1.scale(1, 0) + bitcell_wl_pos.scale(0, 1)
+            if driver_wl_pin.layer != bitcell_wl_pin.layer:
+                self.add_path(driver_wl_pin.layer, [driver_wl_pos, mid1, mid2])
+                self.add_via_stack_center(from_layer=driver_wl_pin.layer,
+                                          to_layer=bitcell_wl_pin.layer,
+                                          offset=mid2)
+                self.add_path(bitcell_wl_pin.layer, [mid2, bitcell_wl_pos])
+            else:
+                self.add_path(bitcell_wl_pin.layer, [driver_wl_pos, mid1, mid2, bitcell_wl_pos])
+                
 
     def route_port_address_right(self, port):
         """ Connecting Wordline driver output to Bitcell WL connection  """
 
-        driver_names = ["wl_{}".format(x) for x in range(self.num_rows)]
-        for (driver_name, array_name) in zip(driver_names, self.bitcell_array.get_wordline_names(port)):
+        driver_names = ["wl_{}".format(x) for x in range(self.num_rows)] + ["rbl_wl"]
+        rbl_wl_name = self.bitcell_array.get_rbl_wordline_names(port)[port]
+        for (driver_name, array_name) in zip(driver_names, self.bitcell_array.get_wordline_names(port) + [rbl_wl_name]):
             # The mid guarantees we exit the input cell to the right.
             driver_wl_pin = self.port_address_inst[port].get_pin(driver_name)
             driver_wl_pos = driver_wl_pin.lc()
@@ -886,6 +902,7 @@ class bank(design.design):
                                       to_layer=bitcell_wl_pin.layer,
                                       offset=mid2)
             self.add_path(bitcell_wl_pin.layer, [mid2, bitcell_wl_pos])
+            
 
     def route_column_address_lines(self, port):
         """ Connecting the select lines of column mux to the address bus """
@@ -1021,10 +1038,6 @@ class bank(design.design):
         connection.append((self.prefix + "p_en_bar{}".format(port),
                            self.port_data_inst[port].get_pin("p_en_bar")))
 
-        rbl_wl_name = self.bitcell_array.get_rbl_wordline_names(port)
-        connection.append((self.prefix + "wl_en{}".format(port),
-                           self.bitcell_array_inst.get_pin(rbl_wl_name[port])))
-
         if port in self.write_ports:
             connection.append((self.prefix + "w_en{}".format(port),
                                self.port_data_inst[port].get_pin("w_en")))
@@ -1045,7 +1058,7 @@ class bank(design.design):
             self.add_via_stack_center(from_layer=pin.layer,
                                       to_layer="m2",
                                       offset=control_pos)
-
+    
         # clk to wordline_driver
         control_signal = self.prefix + "wl_en{}".format(port)
         if port % 2:
diff --git a/compiler/modules/global_bitcell_array.py b/compiler/modules/global_bitcell_array.py
index c26e62f1..ce35455a 100644
--- a/compiler/modules/global_bitcell_array.py
+++ b/compiler/modules/global_bitcell_array.py
@@ -54,6 +54,8 @@ class global_bitcell_array(bitcell_base_array.bitcell_base_array):
         """ Add the modules used in this design """
         self.local_mods = []
 
+        # Special case of a single local array
+        # so it should contain the left and possibly right RBL
         if len(self.column_sizes) == 1:
             la = factory.create(module_type="local_bitcell_array",
                                 rows=self.row_size,
@@ -66,19 +68,21 @@ class global_bitcell_array(bitcell_base_array.bitcell_base_array):
             return
 
         for i, cols in enumerate(self.column_sizes):
-            # Always add the left RBLs to the first subarray and the right RBLs to the last subarray
+            # Always add the left RBLs to the first subarray
             if i == 0:
                 la = factory.create(module_type="local_bitcell_array",
                                     rows=self.row_size,
                                     cols=cols,
                                     rbl=self.rbl,
                                     left_rbl=[0])
+            # Add the right RBL to the last subarray
             elif i == len(self.column_sizes) - 1 and len(self.all_ports) > 1:
                 la = factory.create(module_type="local_bitcell_array",
                                     rows=self.row_size,
                                     cols=cols,
                                     rbl=self.rbl,
                                     right_rbl=[1])
+            # Middle subarrays do not have any RBLs
             else:
                 la = factory.create(module_type="local_bitcell_array",
                                     rows=self.row_size,
@@ -263,3 +267,4 @@ class global_bitcell_array(bitcell_base_array.bitcell_base_array):
         for inst in self.local_insts:
             offsets.extend(inst.lx() + x for x in inst.mod.get_column_offsets())
         return offsets
+    
diff --git a/compiler/modules/hierarchical_decoder.py b/compiler/modules/hierarchical_decoder.py
index ca26993f..21aa1c89 100644
--- a/compiler/modules/hierarchical_decoder.py
+++ b/compiler/modules/hierarchical_decoder.py
@@ -55,9 +55,9 @@ class hierarchical_decoder(design.design):
         self.route_decoder_bus()
         self.route_vdd_gnd()
         
-        self.offset_all_coordinates()
+        self.offset_x_coordinates()
         
-        self.width = self.and_inst[0].rx() + self.m1_space
+        self.width = self.and_inst[0].rx() + 0.5 * self.m1_width
         
         self.add_boundary()
         self.DRC_LVS()
diff --git a/compiler/modules/local_bitcell_array.py b/compiler/modules/local_bitcell_array.py
index d552d77a..af1462c3 100644
--- a/compiler/modules/local_bitcell_array.py
+++ b/compiler/modules/local_bitcell_array.py
@@ -207,9 +207,9 @@ class local_bitcell_array(bitcell_base_array.bitcell_base_array):
                 
                 y_offset = in_pin.cy()
                 if port == 0:
-                    y_offset -= 1.5 * self.m3_pitch
+                    y_offset -= 2 * self.m3_pitch
                 else:
-                    y_offset += 1.5 * self.m3_pitch
+                    y_offset += 2 * self.m3_pitch
                     
                 self.add_layout_pin_segment_center(text=wl_name,
                                                    layer="m3",
diff --git a/compiler/modules/port_address.py b/compiler/modules/port_address.py
index 532463a7..f7e85afd 100644
--- a/compiler/modules/port_address.py
+++ b/compiler/modules/port_address.py
@@ -17,10 +17,11 @@ class port_address(design.design):
     Create the address port (row decoder and wordline driver)..
     """
 
-    def __init__(self, cols, rows, name=""):
+    def __init__(self, cols, rows, port, name=""):
         
         self.num_cols = cols
         self.num_rows = rows
+        self.port = port
         self.addr_size = ceil(log(self.num_rows, 2))
         
         if name == "":
@@ -39,6 +40,7 @@ class port_address(design.design):
         self.add_modules()
         self.create_row_decoder()
         self.create_wordline_driver()
+        self.create_rbl_driver()
         
     def create_layout(self):
         if "li" in layer:
@@ -59,6 +61,8 @@ class port_address(design.design):
 
         for bit in range(self.num_rows):
             self.add_pin("wl_{0}".format(bit), "OUTPUT")
+            
+        self.add_pin("rbl_wl", "OUTPUT")
         
         self.add_pin("vdd", "POWER")
         self.add_pin("gnd", "GROUND")
@@ -71,10 +75,13 @@ class port_address(design.design):
         
     def route_supplies(self):
         """ Propagate all vdd/gnd pins up to this level for all modules """
-        for inst in self.insts:
+        for inst in [self.wordline_driver_array_inst, self.row_decoder_inst]:
             self.copy_power_pins(inst, "vdd")
             self.copy_power_pins(inst, "gnd")
 
+        rbl_vdd_pin = self.rbl_driver_inst.get_pin("vdd")
+        self.add_power_pin("vdd", rbl_vdd_pin.lc())
+
     def route_pins(self):
         for row in range(self.addr_size):
             decoder_name = "addr_{}".format(row)
@@ -82,16 +89,16 @@ class port_address(design.design):
 
         for row in range(self.num_rows):
             driver_name = "wl_{}".format(row)
-            self.copy_layout_pin(self.wordline_driver_inst, driver_name)
+            self.copy_layout_pin(self.wordline_driver_array_inst, driver_name)
 
-        self.copy_layout_pin(self.wordline_driver_inst, "en", "wl_en")
+        self.copy_layout_pin(self.rbl_driver_inst, "Z", "rbl_wl")
             
     def route_internal(self):
         for row in range(self.num_rows):
             # The pre/post is to access the pin from "outside" the cell to avoid DRCs
             decoder_out_pin = self.row_decoder_inst.get_pin("decode_{}".format(row))
             decoder_out_pos = decoder_out_pin.rc()
-            driver_in_pin = self.wordline_driver_inst.get_pin("in_{}".format(row))
+            driver_in_pin = self.wordline_driver_array_inst.get_pin("in_{}".format(row))
             driver_in_pos = driver_in_pin.lc()
             self.add_zjog(self.route_layer, decoder_out_pos, driver_in_pos, var_offset=0.3)
 
@@ -102,6 +109,26 @@ class port_address(design.design):
             self.add_via_stack_center(from_layer=driver_in_pin.layer,
                                       to_layer=self.route_layer,
                                       offset=driver_in_pos)
+
+        # Route the RBL from the enable input
+        en_pin = self.wordline_driver_array_inst.get_pin("en")
+        if self.port == 0:
+            en_pos = en_pin.bc()
+        else:
+            en_pos = en_pin.uc()
+        rbl_in_pin = self.rbl_driver_inst.get_pin("A")
+        rbl_in_pos = rbl_in_pin.center()
+
+        self.add_via_stack_center(from_layer=rbl_in_pin.layer,
+                                  to_layer=en_pin.layer,
+                                  offset=rbl_in_pos)
+        self.add_zjog(layer=en_pin.layer,
+                      start=rbl_in_pos,
+                      end=en_pos,
+                      first_direction="V")
+        self.add_layout_pin_rect_center(text="wl_en",
+                                        layer=en_pin.layer,
+                                        offset=rbl_in_pos)
         
     def add_modules(self):
 
@@ -109,10 +136,33 @@ class port_address(design.design):
                                           num_outputs=self.num_rows)
         self.add_mod(self.row_decoder)
         
-        self.wordline_driver = factory.create(module_type="wordline_driver_array",
-                                              rows=self.num_rows,
-                                              cols=self.num_cols)
-        self.add_mod(self.wordline_driver)
+        self.wordline_driver_array = factory.create(module_type="wordline_driver_array",
+                                                    rows=self.num_rows,
+                                                    cols=self.num_cols)
+        self.add_mod(self.wordline_driver_array)
+
+        try:
+            local_array_size = OPTS.local_array_size
+            driver_size = max(int(self.num_cols / local_array_size), 1)
+        except AttributeError:
+            local_array_size = 0
+            # Defautl to FO4
+            driver_size = max(int(self.num_cols / 4), 1)
+
+        # The polarity must be switched if we have a hierarchical wordline
+        # to compensate for the local array inverters
+        b = factory.create(module_type="bitcell")
+
+        if local_array_size > 0:
+            self.rbl_driver = factory.create(module_type="inv_dec",
+                                             size=driver_size,
+                                             height=b.height)
+        else:
+            self.rbl_driver = factory.create(module_type="buf_dec",
+                                             size=driver_size,
+                                             height=b.height)
+            
+        self.add_mod(self.rbl_driver)
         
     def create_row_decoder(self):
         """  Create the hierarchical row decoder  """
@@ -128,11 +178,24 @@ class port_address(design.design):
         temp.extend(["vdd", "gnd"])
         self.connect_inst(temp)
 
+    def create_rbl_driver(self):
+        """ Create the RBL Wordline Driver """
+
+        self.rbl_driver_inst = self.add_inst(name="rbl_driver",
+                                             mod=self.rbl_driver)
+        
+        temp = []
+        temp.append("wl_en")
+        temp.append("rbl_wl")
+        temp.append("vdd")
+        temp.append("gnd")
+        self.connect_inst(temp)
+
     def create_wordline_driver(self):
         """ Create the Wordline Driver """
-
-        self.wordline_driver_inst = self.add_inst(name="wordline_driver",
-                                                  mod=self.wordline_driver)
+        
+        self.wordline_driver_array_inst = self.add_inst(name="wordline_driver",
+                                                        mod=self.wordline_driver_array)
 
         temp = []
         for row in range(self.num_rows):
@@ -150,11 +213,23 @@ class port_address(design.design):
         """
 
         row_decoder_offset = vector(0, 0)
-        wordline_driver_offset = vector(self.row_decoder.width, 0)
-        self.wordline_driver_inst.place(wordline_driver_offset)
         self.row_decoder_inst.place(row_decoder_offset)
+
+        wordline_driver_array_offset = vector(self.row_decoder_inst.rx(), 0)
+        self.wordline_driver_array_inst.place(wordline_driver_array_offset)
+
+        x_offset = self.wordline_driver_array_inst.rx() - self.rbl_driver.width - self.m1_pitch
+        if self.port == 0:
+            rbl_driver_offset = vector(x_offset,
+                                       0)
+            self.rbl_driver_inst.place(rbl_driver_offset, "MX")
+        else:
+            rbl_driver_offset = vector(x_offset,
+                                       self.wordline_driver_array.height)
+            self.rbl_driver_inst.place(rbl_driver_offset)
+        
         # Pass this up
         self.predecoder_height = self.row_decoder.predecoder_height
 
         self.height = self.row_decoder.height
-        self.width = self.wordline_driver_inst.rx()
+        self.width = self.wordline_driver_array_inst.rx()
diff --git a/compiler/modules/replica_bitcell_array.py b/compiler/modules/replica_bitcell_array.py
index 6c2d54a7..4e1c4df1 100644
--- a/compiler/modules/replica_bitcell_array.py
+++ b/compiler/modules/replica_bitcell_array.py
@@ -305,14 +305,21 @@ class replica_bitcell_array(bitcell_base_array.bitcell_base_array):
 
     def create_layout(self):
 
+        # We will need unused wordlines grounded, so we need to know their layer
+        pin = self.cell.get_pin(self.cell.get_all_wl_names()[0])
+        pin_layer = pin.layer
+        self.unused_pitch = 1.5 * getattr(self, "{}_pitch".format(pin_layer))
+        self.unused_offset = vector(self.unused_pitch, 0)
+
+        # Add extra width on the left and right for the unused WLs
         self.height = (self.row_size + self.extra_rows) * self.dummy_row.height
-        self.width = (self.column_size + self.extra_cols) * self.cell.width
+        self.width = (self.column_size + self.extra_cols) * self.cell.width + 2 * self.unused_pitch
 
         # This is a bitcell x bitcell offset to scale
         self.bitcell_offset = vector(self.cell.width, self.cell.height)
 
-        # Everything is computed with the main array at (0, 0) to start
-        self.bitcell_array_inst.place(offset=[0, 0])
+        # Everything is computed with the main array at (self.unused_pitch, 0) to start
+        self.bitcell_array_inst.place(offset=self.unused_offset)
 
         self.add_replica_columns()
         
@@ -356,7 +363,7 @@ class replica_bitcell_array(bitcell_base_array.bitcell_base_array):
 
         # Grow from left to right, toward the array
         for bit, port in enumerate(self.left_rbl):
-            offset = self.bitcell_offset.scale(-len(self.left_rbl) + bit, -self.rbl[0] - 1)
+            offset = self.bitcell_offset.scale(-len(self.left_rbl) + bit, -self.rbl[0] - 1) + self.unused_offset
             self.replica_col_insts[port].place(offset)
         # Grow to the right of the bitcell array, array outward
         for bit, port in enumerate(self.right_rbl):
@@ -367,11 +374,13 @@ class replica_bitcell_array(bitcell_base_array.bitcell_base_array):
         # Add the dummy rows even if we aren't adding the replica column to this bitcell array
         # These grow up, toward the array
         for bit in range(self.rbl[0]):
-            self.dummy_row_replica_insts[bit].place(offset=self.bitcell_offset.scale(0, -self.rbl[0] + bit + (-self.rbl[0] + bit) % 2),
+            dummy_offset = self.bitcell_offset.scale(0, -self.rbl[0] + bit + (-self.rbl[0] + bit) % 2) + self.unused_offset
+            self.dummy_row_replica_insts[bit].place(offset=dummy_offset,
                                                     mirror="MX" if (-self.rbl[0] + bit) % 2 else "R0")
         # These grow up, away from the array
         for bit in range(self.rbl[1]):
-            self.dummy_row_replica_insts[self.rbl[0] + bit].place(offset=self.bitcell_offset.scale(0, bit + bit % 2) + self.bitcell_array_inst.ul(),
+            dummy_offset = self.bitcell_offset.scale(0, bit + bit % 2) + self.bitcell_array_inst.ul()
+            self.dummy_row_replica_insts[self.rbl[0] + bit].place(offset=dummy_offset,
                                                                   mirror="MX" if bit % 2 else "R0")
         
     def add_end_caps(self):
@@ -386,12 +395,12 @@ class replica_bitcell_array(bitcell_base_array.bitcell_base_array):
         # FIXME: These depend on the array size itself
         # Far bottom dummy row (first row below array IS flipped)
         flip_dummy = (self.rbl[0] + 1) % 2
-        dummy_row_offset = self.bitcell_offset.scale(0, -self.rbl[0] - 1 + flip_dummy)
+        dummy_row_offset = self.bitcell_offset.scale(0, -self.rbl[0] - 1 + flip_dummy) + self.unused_offset
         self.dummy_row_insts[0].place(offset=dummy_row_offset,
                                           mirror="MX" if flip_dummy else "R0")
         # Far left dummy col
         # Shifted down by the number of left RBLs even if we aren't adding replica column to this bitcell array
-        dummy_col_offset = self.bitcell_offset.scale(-len(self.left_rbl) - 1, -self.rbl[0] - 1)
+        dummy_col_offset = self.bitcell_offset.scale(-len(self.left_rbl) - 1, -self.rbl[0] - 1) + self.unused_offset
         self.dummy_col_insts[0].place(offset=dummy_col_offset)
         # Far right dummy col
         # Shifted down by the number of left RBLs even if we aren't adding replica column to this bitcell array
@@ -495,13 +504,13 @@ class replica_bitcell_array(bitcell_base_array.bitcell_base_array):
     def ground_pin(self, inst, name):
         pin = inst.get_pin(name)
         pin_layer = pin.layer
-        layer_pitch = 1.5 * getattr(self, "{}_pitch".format(pin_layer))
+        
         left_pin_loc = vector(self.dummy_col_insts[0].lx(), pin.cy())
         right_pin_loc = vector(self.dummy_col_insts[1].rx(), pin.cy())
 
         # Place the pins a track outside of the array
-        left_loc = left_pin_loc - vector(layer_pitch, 0)
-        right_loc = right_pin_loc + vector(layer_pitch, 0)
+        left_loc = left_pin_loc - vector(self.unused_pitch, 0)
+        right_loc = right_pin_loc + vector(self.unused_pitch, 0)
         self.add_power_pin("gnd", left_loc, directions=("H", "H"))
         self.add_power_pin("gnd", right_loc, directions=("H", "H"))
 
diff --git a/compiler/modules/wordline_buffer_array.py b/compiler/modules/wordline_buffer_array.py
index 7a1bf8d1..45f3ef33 100644
--- a/compiler/modules/wordline_buffer_array.py
+++ b/compiler/modules/wordline_buffer_array.py
@@ -109,12 +109,13 @@ class wordline_buffer_array(design.design):
     def place_drivers(self):
 
         for row in range(self.rows):
+            # These are flipped since we always start with an RBL on the bottom
             if (row % 2):
-                y_offset = self.wl_driver.height * (row + 1)
-                inst_mirror = "MX"
-            else:
                 y_offset = self.wl_driver.height * row
                 inst_mirror = "R0"
+            else:
+                y_offset = self.wl_driver.height * (row + 1)
+                inst_mirror = "MX"
 
             offset = [0, y_offset]
             
diff --git a/compiler/modules/wordline_driver_array.py b/compiler/modules/wordline_driver_array.py
index ac0a0cb2..c334404e 100644
--- a/compiler/modules/wordline_driver_array.py
+++ b/compiler/modules/wordline_driver_array.py
@@ -44,7 +44,7 @@ class wordline_driver_array(design.design):
         self.place_drivers()
         self.route_layout()
         self.route_vdd_gnd()
-        self.offset_all_coordinates()
+        self.offset_x_coordinates()
         self.add_boundary()
         self.DRC_LVS()
         
@@ -60,8 +60,10 @@ class wordline_driver_array(design.design):
         self.add_pin("gnd", "GROUND")
 
     def add_modules(self):
+            
         self.wl_driver = factory.create(module_type="wordline_driver",
-                                        size=self.cols)
+                                        cols=self.cols)
+            
         self.add_mod(self.wl_driver)
         
     def route_vdd_gnd(self):
diff --git a/compiler/options.py b/compiler/options.py
index 8d8e3b42..b9be3999 100644
--- a/compiler/options.py
+++ b/compiler/options.py
@@ -30,6 +30,9 @@ class options(optparse.Values):
     num_r_ports = 0
     num_w_ports = 0
 
+    # By default, use local arrays with a max fanout of 16
+    #local_array_size = 16
+    
     # Write mask size, default will be overwritten with word_size if not user specified
     write_size = None
 
@@ -139,6 +142,7 @@ class options(optparse.Values):
     bank_select = "bank_select"
     bitcell_array = "bitcell_array"
     bitcell = "bitcell"
+    buf_dec = "pbuf"
     column_mux_array = "single_level_column_mux_array"
     control_logic = "control_logic"
     decoder = "hierarchical_decoder"
diff --git a/compiler/pgates/pbuf.py b/compiler/pgates/pbuf.py
index e504b89e..c2398d5f 100644
--- a/compiler/pgates/pbuf.py
+++ b/compiler/pgates/pbuf.py
@@ -96,4 +96,4 @@ class pbuf(pgate.pgate):
                                         offset=a_pin.center(),
                                         width=a_pin.width(),
                                         height=a_pin.height())
-                                        
\ No newline at end of file
+                                        
diff --git a/compiler/pgates/wordline_driver.py b/compiler/pgates/wordline_driver.py
index a8ca76f8..6abf7b20 100644
--- a/compiler/pgates/wordline_driver.py
+++ b/compiler/pgates/wordline_driver.py
@@ -18,9 +18,9 @@ class wordline_driver(design.design):
     This is an AND (or NAND) with configurable drive strength to drive the wordlines.
     It is matched to the bitcell height.
     """
-    def __init__(self, name, size=1, height=None):
+    def __init__(self, name, cols, height=None):
         debug.info(1, "Creating wordline_driver {}".format(name))
-        self.add_comment("size: {}".format(size))
+        self.add_comment("cols: {}".format(cols))
         super().__init__(name)
 
         if height is None:
@@ -28,7 +28,7 @@ class wordline_driver(design.design):
             self.height = b.height
         else:
             self.height = height
-        self.size = size
+        self.cols = cols
 
         self.create_netlist()
         if not OPTS.netlist_only:
@@ -42,10 +42,25 @@ class wordline_driver(design.design):
     def create_modules(self):
         self.nand = factory.create(module_type="nand2_dec",
                                    height=self.height)
-            
-        self.driver = factory.create(module_type="inv_dec",
-                                     size=self.size,
-                                     height=self.nand.height)
+
+        try:
+            local_array_size = OPTS.local_array_size
+            driver_size = max(int(self.cols / local_array_size), 1)
+        except AttributeError:
+            local_array_size = 0
+            # Defautl to FO4
+            driver_size = max(int(self.cols / 4), 1)
+
+        # The polarity must be switched if we have a hierarchical wordline
+        # to compensate for the local array inverters
+        if local_array_size > 0:
+            self.driver = factory.create(module_type="buf_dec",
+                                         size=driver_size,
+                                         height=self.nand.height)
+        else:
+            self.driver = factory.create(module_type="inv_dec",
+                                         size=driver_size,
+                                         height=self.nand.height)
             
         self.add_mod(self.nand)
         self.add_mod(self.driver)
diff --git a/compiler/tests/04_wordline_driver_test.py b/compiler/tests/04_wordline_driver_test.py
index ada65db7..74cfb7a7 100755
--- a/compiler/tests/04_wordline_driver_test.py
+++ b/compiler/tests/04_wordline_driver_test.py
@@ -25,7 +25,7 @@ class wordline_driver_test(openram_test):
 
         # check wordline driver for single port
         debug.info(2, "Checking driver")
-        tx = factory.create(module_type="wordline_driver")
+        tx = factory.create(module_type="wordline_driver", cols=8)
         self.local_check(tx)
 
         globals.end_openram()
diff --git a/compiler/tests/18_port_address_1rw_1r_test.py b/compiler/tests/18_port_address_1rw_1r_test.py
index caf2cb96..42a46614 100755
--- a/compiler/tests/18_port_address_1rw_1r_test.py
+++ b/compiler/tests/18_port_address_1rw_1r_test.py
@@ -27,11 +27,11 @@ class port_address_1rw_1r_test(openram_test):
         globals.setup_bitcell()
         
         debug.info(1, "Port address 16 rows")
-        a = factory.create("port_address", cols=16, rows=16)
+        a = factory.create("port_address", cols=16, rows=16, port=0)
         self.local_check(a)
 
         debug.info(1, "Port address 256 rows")
-        a = factory.create("port_address", cols=256, rows=256)
+        a = factory.create("port_address", cols=256, rows=256, port=1)
         self.local_check(a)
         
         globals.end_openram()
diff --git a/compiler/tests/18_port_address_test.py b/compiler/tests/18_port_address_test.py
index 11da333e..94feedb2 100755
--- a/compiler/tests/18_port_address_test.py
+++ b/compiler/tests/18_port_address_test.py
@@ -21,11 +21,11 @@ class port_address_test(openram_test):
         globals.init_openram(config_file)
 
         debug.info(1, "Port address 16 rows")
-        a = factory.create("port_address", cols=16, rows=16)
+        a = factory.create("port_address", cols=16, rows=16, port=0)
         self.local_check(a)
         
         debug.info(1, "Port address 512 rows")
-        a = factory.create("port_address", cols=256, rows=512)
+        a = factory.create("port_address", cols=256, rows=512, port=0)
         self.local_check(a)
 
         globals.end_openram()