Updating modexp core to v 0.50. This version contains a working core that can perform sign and verify with big keys/values. The core builds ok in Altera and Xilinx FPGA tools. This commit also includes a new testgenerator capable of generating testbench for modexp with autgenerated test data of different lengths. The README has been updated with status and implementation results in for different FPGA devices.

2 files changed, 261 insertions, 99 deletions

diff --git a/src/rtl/modexp.v b/src/rtl/modexp.v
index d64a556..9f3d4f0 100644
--- a/src/rtl/modexp.v
+++ b/src/rtl/modexp.v
@@ -71,35 +71,38 @@ module modexp(
   localparam ADDR_NAME0          = 8'h00;
   localparam ADDR_NAME1          = 8'h01;
   localparam ADDR_VERSION        = 8'h02;
-  localparam ADDR_CTRL           = 8'h03;
 
+  localparam ADDR_CTRL           = 8'h08;
   localparam CTRL_INIT_BIT       = 0;
   localparam CTRL_NEXT_BIT       = 1;
 
   localparam ADDR_STATUS         = 8'h09;
+  localparam STATUS_READY_BIT    = 0;
+
+  localparam ADDR_MODULUS_LENGTH  = 8'h20;
+  localparam ADDR_EXPONENT_LENGTH = 8'h21;
+  localparam ADDR_LENGTH          = 8'h22; // Should be deprecated.
 
-  localparam ADDR_MODSIZE        = 8'h20;
-  localparam ADDR_LENGTH         = 8'h21;
-  localparam ADDR_EXPONENT       = 8'h22;
 
   localparam MODULUS_PREFIX      = 4'h1;
   localparam ADDR_MODULUS_START  = 8'h00;
   localparam ADDR_MODULUS_END    = 8'hff;
 
+
   localparam EXPONENT_PREFIX     = 4'h2;
   localparam ADDR_EXPONENT_START = 8'h00;
   localparam ADDR_EXPONENT_END   = 8'hff;
 
+
   localparam MESSAGE_PREFIX      = 4'h3;
   localparam MESSAGE_START       = 8'h00;
   localparam MESSAGE_END         = 8'hff;
 
+
   localparam RESULT_PREFIX       = 4'h4;
   localparam RESULT_START        = 8'h00;
   localparam RESULT_END          = 8'hff;
 
-  localparam LENGTH_PREFIX       = 4'h5;
-
   localparam DEFAULT_MODLENGTH   = 8'h80;
   localparam DEFAULT_EXPLENGTH   = 8'h80;
 
@@ -107,7 +110,7 @@ module modexp(
   localparam MONTPROD_SELECT_X_NR   = 3'h1;
   localparam MONTPROD_SELECT_Z_P    = 3'h2;
   localparam MONTPROD_SELECT_P_P    = 3'h3;
-  localparam MONTPROD_SELECT_ONE_Z  = 3'h4;
+  localparam MONTPROD_SELECT_Z_ONE  = 3'h4;
 
   localparam MONTPROD_DEST_Z        = 2'b00;
   localparam MONTPROD_DEST_P        = 2'b01;
@@ -124,53 +127,36 @@ module modexp(
   localparam CTRL_CALCULATE_ZN   = 4'h8;
   localparam CTRL_DONE           = 4'h9;
 
+  localparam EXPONATION_MODE_SECRET_SECURE = 1'b0;
+  localparam EXPONATION_MODE_PUBLIC_FAST   = 1'b1; //for rsa, c=M^65537 etc, there is no need to slow down to hide the exponent
+
+
   localparam CORE_NAME0          = 32'h6d6f6465; // "mode"
   localparam CORE_NAME1          = 32'h78702020; // "xp  "
-  localparam CORE_VERSION        = 32'h302e3031; // "0.01"
+  localparam CORE_VERSION        = 32'h302e3530; // "0.50"
 
 
   //----------------------------------------------------------------
   // Registers including update variables and write enable.
   //----------------------------------------------------------------
-  reg [07 : 0]  modulus_mem_int_rd_addr;
-  wire [31 : 0] modulus_mem_int_rd_data;
-  wire [31 : 0] modulus_mem_api_rd_data;
-  reg           modulus_mem_api_we;
-
-  reg [07 : 0]  message_mem_int_rd_addr;
-  wire [31 : 0] message_mem_int_rd_data;
-  wire [31 : 0] message_mem_api_rd_data;
-  reg           message_mem_api_we;
-
-  reg [07 : 0]  exponent_mem_int_rd_addr;
-  wire [31 : 0] exponent_mem_int_rd_data;
-  wire [31 : 0] exponent_mem_api_rd_data;
-  reg           exponent_mem_api_we;
+  reg [07 : 0] exponent_length_reg;
+  reg [07 : 0] exponent_length_new;
+  reg          exponent_length_we;
 
-  wire [31 : 0] result_mem_api_rd_data;
-  reg  [07 : 0] result_mem_int_rd_addr;
-  wire [31 : 0] result_mem_int_rd_data;
-  reg  [07 : 0] result_mem_int_wr_addr;
-  reg  [31 : 0] result_mem_int_wr_data;
-  reg           result_mem_int_we;
-
-  reg  [07 : 0] p_mem_rd0_addr;
-  wire [31 : 0] p_mem_rd0_data;
-  reg  [07 : 0] p_mem_rd1_addr;
-  wire [31 : 0] p_mem_rd1_data;
-  reg  [07 : 0] p_mem_wr_addr;
-  reg  [31 : 0] p_mem_wr_data;
-  reg           p_mem_we;
+  reg [07 : 0] modulus_length_reg;
+  reg [07 : 0] modulus_length_new;
+  reg          modulus_length_we;
 
   reg [07 : 0] length_reg;
+  reg [07 : 0] length_new;
   reg [07 : 0] length_m1_reg;
-  reg [07 : 0] length_new;    //TODO: API should write length!!!
-  reg [07 : 0] length_m1_new; //TODO: API should write length-1 when writing length!!!
-  reg          length_we;     //TODO: API should enable length_we!!!
+  reg [07 : 0] length_m1_new;
+  reg          length_we;
 
   reg          start_reg;
-  reg          start_new; //TODO: API should start operations!!!
-  reg          start_we;  //TODO: API should start operations!!!
+  reg          start_new;
+  reg          start_we;
+  reg          clear_start;
 
   reg          ready_reg;
   reg          ready_new;
@@ -187,22 +173,61 @@ module modexp(
   reg [3 : 0]  modexp_ctrl_new;
   reg          modexp_ctrl_we;
 
-  reg [31 : 0] one;
+  reg [31 : 0] one_reg;
   reg [31 : 0] one_new;
+  reg [31 : 0] b_one_reg;
+  reg [31 : 0] b_one_new;
 
   reg [12 : 0] loop_counter_reg;
   reg [12 : 0] loop_counter_new;
   reg          loop_counter_we;
+
   reg [07 : 0] E_word_index;
   reg [04 : 0] E_bit_index;
   reg          last_iteration;
+
   reg          ei_reg;
   reg          ei_new;
   reg          ei_we;
 
+  reg          exponation_mode_reg;
+  reg          exponation_mode_new;
+  reg          exponation_mode_we;
+
+
   //----------------------------------------------------------------
   // Wires.
   //----------------------------------------------------------------
+  reg [07 : 0]  modulus_mem_int_rd_addr;
+  wire [31 : 0] modulus_mem_int_rd_data;
+  wire [31 : 0] modulus_mem_api_rd_data;
+  reg           modulus_mem_api_we;
+
+  reg [07 : 0]  message_mem_int_rd_addr;
+  wire [31 : 0] message_mem_int_rd_data;
+  wire [31 : 0] message_mem_api_rd_data;
+  reg           message_mem_api_we;
+
+  reg [07 : 0]  exponent_mem_int_rd_addr;
+  wire [31 : 0] exponent_mem_int_rd_data;
+  wire [31 : 0] exponent_mem_api_rd_data;
+  reg           exponent_mem_api_we;
+
+  wire [31 : 0] result_mem_api_rd_data;
+  reg  [07 : 0] result_mem_int_rd_addr;
+  wire [31 : 0] result_mem_int_rd_data;
+  reg  [07 : 0] result_mem_int_wr_addr;
+  reg  [31 : 0] result_mem_int_wr_data;
+  reg           result_mem_int_we;
+
+  reg  [07 : 0] p_mem_rd0_addr;
+  wire [31 : 0] p_mem_rd0_data;
+  reg  [07 : 0] p_mem_rd1_addr;
+  wire [31 : 0] p_mem_rd1_data;
+  reg  [07 : 0] p_mem_wr_addr;
+  reg  [31 : 0] p_mem_wr_data;
+  reg           p_mem_we;
+
   reg [31 : 0]  tmp_read_data;
   //reg           tmp_error;
 
@@ -225,7 +250,7 @@ module modexp(
 
   reg            residue_calculate;
   wire           residue_ready;
-  reg  [14 : 0]  residue_nn;
+  reg [14 : 0]   residue_nn;
   reg  [07 : 0]  residue_length;
   wire [07 : 0]  residue_opa_rd_addr;
   wire [31 : 0]  residue_opa_rd_data;
@@ -238,6 +263,11 @@ module modexp(
   reg  [07 : 0] residue_mem_montprod_read_addr;
   wire [31 : 0] residue_mem_montprod_read_data;
 
+  reg           residue_valid_reg;
+  reg           residue_valid_new;
+  reg           invalidate_residue;
+  reg           residue_valid_int_validated;
+
   //----------------------------------------------------------------
   // Concurrent connectivity for ports etc.
   //----------------------------------------------------------------
@@ -368,18 +398,37 @@ module modexp(
     begin
       if (!reset_n)
         begin
+          exponent_length_reg <= DEFAULT_EXPLENGTH;
+          modulus_length_reg  <= DEFAULT_MODLENGTH;
+          start_reg           <= 1'b0;
           ready_reg           <= 1'b1;
           montprod_select_reg <= MONTPROD_SELECT_ONE_NR;
           montprod_dest_reg   <= MONTPROD_DEST_NOWHERE;
           modexp_ctrl_reg     <= CTRL_IDLE;
-          one                 <= 32'h0;
+          one_reg             <= 32'h0;
+          b_one_reg           <= 32'h0;
           length_reg          <= 8'h0;
           length_m1_reg       <= 8'h0;
           loop_counter_reg    <= 13'b0;
           ei_reg              <= 1'b0;
+          residue_valid_reg   <= 1'b0;
+          exponation_mode_reg <= EXPONATION_MODE_SECRET_SECURE;
         end
       else
         begin
+          one_reg <= one_new;
+          b_one_reg <= b_one_new;
+          residue_valid_reg <= residue_valid_new;
+
+          if (exponent_length_we)
+            exponent_length_reg <= exponent_length_new;
+
+          if (modulus_length_we)
+            modulus_length_reg <= modulus_length_new;
+
+          if (start_we)
+            start_reg <= start_new;
+
           if (ready_we)
             ready_reg <= ready_new;
 
@@ -394,8 +443,8 @@ module modexp(
 
           if (length_we)
             begin
-              length_reg    <= write_data[7 : 0];
-              length_m1_reg <= write_data[7 : 0] - 8'h1;
+              length_reg    <= length_new;
+              length_m1_reg <= length_m1_new;
             end
 
           if (loop_counter_we)
@@ -404,7 +453,8 @@ module modexp(
           if (ei_we)
             ei_reg <= ei_new;
 
-          one <= one_new;
+          if (exponation_mode_we)
+            exponation_mode_reg <= exponation_mode_new;
         end
     end // reg_update
 
@@ -416,12 +466,35 @@ module modexp(
   //----------------------------------------------------------------
   always @*
     begin : api
+      modulus_length_we   = 1'b0;
+      exponent_length_we  = 1'b0;
+      start_new           = 1'b0;
+      start_we            = 1'b0;
       modulus_mem_api_we  = 1'b0;
       exponent_mem_api_we = 1'b0;
       message_mem_api_we  = 1'b0;
       length_we           = 1'b0;
+      invalidate_residue  = 1'b0;
+
+      //TODO: Add API code to enable fast exponation for working with public exponents.
+      exponation_mode_we  = 1'b0;
+
+      exponation_mode_new = EXPONATION_MODE_SECRET_SECURE;
+      modulus_length_new  = write_data[7 : 0];
+      exponent_length_new = write_data[7 : 0];
+      length_new          = write_data[7 : 0];
+      length_m1_new       = write_data[7 : 0] - 8'h1;
+
       tmp_read_data       = 32'h00000000;
 
+      // We need to be able to clear start bit after we have
+      // started a modexp operation.
+      if (clear_start)
+        begin
+          start_new = 1'b0;
+          start_we  = 1'b1;
+        end
+
       if (cs)
         begin
           case (address[11 : 8])
@@ -430,8 +503,26 @@ module modexp(
                 if (we)
                   begin
                     case (address[7 : 0])
+                      ADDR_CTRL:
+                        begin
+                          start_new = write_data[0];
+                          start_we  = 1'b1;
+                        end
+
                       ADDR_LENGTH:
-                        length_we = 1'b1;
+                        begin
+                          length_we = 1'b1;
+                        end
+
+                      ADDR_MODULUS_LENGTH:
+                        begin
+                          modulus_length_we = 1'b1;
+                        end
+
+                      ADDR_EXPONENT_LENGTH:
+                        begin
+                          exponent_length_we = 1'b1;
+                        end
 
                       default:
                         begin
@@ -451,10 +542,16 @@ module modexp(
                         tmp_read_data = CORE_VERSION;
 
                       ADDR_CTRL:
-                        tmp_read_data = 32'h00000000;
+                        tmp_read_data = {31'h00000000, start_reg};
 
                       ADDR_STATUS:
-                        tmp_read_data = 32'h00000000;
+                        tmp_read_data = {31'h00000000, ready_reg};
+
+                      ADDR_MODULUS_LENGTH:
+                        tmp_read_data = {24'h000000, modulus_length_reg};
+
+                      ADDR_EXPONENT_LENGTH:
+                        tmp_read_data = {24'h000000, exponent_length_reg};
 
                       ADDR_LENGTH:
                         tmp_read_data = {24'h000000, length_reg};
@@ -471,6 +568,7 @@ module modexp(
                 if (we)
                   begin
                     modulus_mem_api_we = 1'b1;
+                    invalidate_residue = 1'b1;
                   end
                 else
                   begin
@@ -522,15 +620,19 @@ module modexp(
   // generates the big integer one ( 00... 01 )
   //----------------------------------------------------------------
   always @*
-    begin : one_process;
+    begin : one_process
+      one_new = 32'h00000000;
+      b_one_new = 32'h00000000;
       if (montprod_opa_addr == length_m1_reg)
         one_new = 32'h00000001;
-      else
-        one_new = 32'h00000000;
+      if (montprod_opb_addr == length_m1_reg)
+        b_one_new = 32'h00000001;
     end
 
+
   //----------------------------------------------------------------
-  // read mux for modulus, since it is being addressed by two sources
+  // Read mux for modulus. Needed since it is being
+  // addressed by two sources.
   //----------------------------------------------------------------
   always @*
     begin : modulus_mem_reader_process
@@ -540,21 +642,38 @@ module modexp(
         modulus_mem_int_rd_addr = montprod_opm_addr;
     end
 
+
   //----------------------------------------------------------------
-  // feeds residue calculator
+  // Feeds residue calculator.
   //----------------------------------------------------------------
   always @*
-    begin : residue_process;
-      residue_nn = { 1'b0, length_reg, 6'h0 }; //N*2, N=length*32, *32 = shl5, *64 = shl6
+    begin : residue_process
+      //N*2, N=length*32, *32 = shl5, *64 = shl6
+      residue_nn = { 1'b0, length_reg, 6'h0 };
       residue_length = length_reg;
       residue_opm_data = modulus_mem_int_rd_data;
     end
 
+
+  //----------------------------------------------------------------
+  // Sets the register that decides if residue is valid or not.
+  //----------------------------------------------------------------
+  always @*
+    begin : residue_valid_process
+      if (invalidate_residue)
+        residue_valid_new = 1'b0;
+      else if ( residue_valid_int_validated == 1'b1)
+        residue_valid_new = 1'b1;
+     else
+        residue_valid_new = residue_valid_reg;
+    end
+
+
   //----------------------------------------------------------------
   // montprod_op_select
   //
   // Select operands used during montprod calculations depending
-  // on what operation we want to do
+  // on what operation we want to do.
   //----------------------------------------------------------------
   always @*
     begin : montprod_op_select
@@ -572,7 +691,7 @@ module modexp(
       case (montprod_select_reg)
         MONTPROD_SELECT_ONE_NR:
           begin
-            montprod_opa_data       = one;
+            montprod_opa_data       = one_reg;
             montprod_opb_data       = residue_mem_montprod_read_data;
           end
 
@@ -594,10 +713,10 @@ module modexp(
             montprod_opb_data       = p_mem_rd1_data;
           end
 
-        MONTPROD_SELECT_ONE_Z:
+        MONTPROD_SELECT_Z_ONE:
           begin
-            montprod_opa_data       = one;
-            montprod_opb_data       = result_mem_int_rd_data;
+            montprod_opa_data       = result_mem_int_rd_data;
+            montprod_opb_data       = b_one_reg;
           end
 
         default:
@@ -612,7 +731,7 @@ module modexp(
   //----------------------------------------------------------------
   // memory write mux
   //
-  // direct memory write signals to correct memory
+  // Direct memory write signals to correct memory.
   //----------------------------------------------------------------
   always @*
     begin : memory_write_process
@@ -643,11 +762,12 @@ module modexp(
   //----------------------------------------------------------------
   // loop_counter
   //
-  // Calculate the loop counter and related variables
+  // Calculate the loop counter and related variables.
   //----------------------------------------------------------------
   always @*
     begin : loop_counters_process
-      E_bit_index      = loop_counter_reg[ 04 : 0 ];
+      loop_counter_new = 13'b0;
+      loop_counter_we  = 1'b0;
 
       if (loop_counter_reg == { length_m1_reg, 5'b11111 })
         last_iteration = 1'b1;
@@ -669,7 +789,7 @@ module modexp(
 
         default:
           begin
-            loop_counter_new = 13'b0;
+            loop_counter_new = loop_counter_reg;
             loop_counter_we  = 1'b0;
           end
 
@@ -680,12 +800,15 @@ module modexp(
   //----------------------------------------------------------------
   // exponent
   //
-  // reads the exponent
+  // Reads the exponent.
   //----------------------------------------------------------------
   always @*
     begin : exponent_process
-      // accessing new instead of reg - pick up update at CTRL_ITERATE_NEW to remove a pipeline stall
-      E_word_index  = loop_counter_new[ 12 : 5 ];
+      // Accessing new instead of reg - pick up update at
+      // CTRL_ITERATE_NEW to remove a pipeline stall.
+      E_word_index  = length_m1_reg - loop_counter_new[ 12 : 5 ];
+
+      E_bit_index   = loop_counter_reg[ 04 : 0 ];
 
       exponent_mem_int_rd_addr = E_word_index;
 
@@ -705,48 +828,70 @@ module modexp(
   //----------------------------------------------------------------
   always @*
     begin
-      ready_new           = 0;
-      ready_we            = 0;
+      ready_new           = 1'b0;
+      ready_we            = 1'b0;
       montprod_select_new = MONTPROD_SELECT_ONE_NR;
       montprod_select_we  = 0;
       montprod_dest_new   = MONTPROD_DEST_NOWHERE;
       montprod_dest_we    = 0;
       montprod_calc       = 0;
       modexp_ctrl_new     = CTRL_IDLE;
-      modexp_ctrl_we      = 0;
+      modexp_ctrl_we      = 1'b0;
+      clear_start         = 1'b0;
 
       residue_calculate = 1'b0;
 
+      residue_valid_int_validated = 1'b0;
+
       case (modexp_ctrl_reg)
         CTRL_IDLE:
           begin
-            ready_new           = 0;
-            ready_we            = 1;
-            if (start_reg == 1'b1)
+            if (start_reg)
               begin
-                modexp_ctrl_new = CTRL_RESIDUE;
-                modexp_ctrl_we  = 1;
-                residue_calculate = 1'b1;
+                ready_new       = 1'b0;
+                ready_we        = 1'b1;
+                modexp_ctrl_new = CTRL_DONE;
+                modexp_ctrl_we  = 1'b1;
+
+                if (residue_valid_reg)
+                  begin
+                    //residue has alrady been calculated, start with MONTPROD( 1, Nr, MODULUS )
+                    montprod_select_new = MONTPROD_SELECT_ONE_NR;
+                    montprod_select_we  = 1;
+                    montprod_dest_new   = MONTPROD_DEST_Z;
+                    montprod_dest_we    = 1;
+                    montprod_calc       = 1;
+                    modexp_ctrl_new     = CTRL_CALCULATE_Z0;
+                    modexp_ctrl_we      = 1;
+                  end
+                else
+                  begin
+                    //modulus has been written and residue (Nr) must be calculated
+                    modexp_ctrl_new = CTRL_RESIDUE;
+                    modexp_ctrl_we  = 1;
+                    residue_calculate = 1'b1;
+                  end
               end
           end
 
         CTRL_RESIDUE:
           begin
-            if (residue_ready == 1'b1)
+            if (residue_ready)
               begin
                 montprod_select_new = MONTPROD_SELECT_ONE_NR;
                 montprod_select_we  = 1;
                 montprod_dest_new   = MONTPROD_DEST_Z;
                 montprod_dest_we    = 1;
                 montprod_calc       = 1;
-                modexp_ctrl_new = CTRL_CALCULATE_Z0;
-                modexp_ctrl_we  = 1;
+                modexp_ctrl_new     = CTRL_CALCULATE_Z0;
+                modexp_ctrl_we      = 1;
+                residue_valid_int_validated = 1'b1; //update registers telling residue is valid
               end
           end
 
         CTRL_CALCULATE_Z0:
           begin
-            if (montprod_ready == 1'b1)
+            if (montprod_ready)
               begin
                 montprod_select_new = MONTPROD_SELECT_X_NR;
                 montprod_select_we  = 1;
@@ -774,20 +919,28 @@ module modexp(
             montprod_dest_new   = MONTPROD_DEST_Z;
             montprod_dest_we    = 1;
             montprod_calc       = 1;
-            modexp_ctrl_new = CTRL_ITERATE_Z_P;
-            modexp_ctrl_we  = 1;
+            modexp_ctrl_new     = CTRL_ITERATE_Z_P;
+            modexp_ctrl_we      = 1;
+
+            if (ei_new == 1'b0 && exponation_mode_reg == EXPONATION_MODE_PUBLIC_FAST)
+              begin
+                //Skip the fake montgomery calculation, exponation_mode_reg optimizing for speed not blinding.
+                montprod_select_new = MONTPROD_SELECT_P_P;
+                montprod_dest_new   = MONTPROD_DEST_P;
+                modexp_ctrl_new     = CTRL_ITERATE_P_P;
+              end
           end
 
         CTRL_ITERATE_Z_P:
-            if (montprod_ready == 1'b1)
+            if (montprod_ready)
               begin
                 montprod_select_new = MONTPROD_SELECT_P_P;
                 montprod_select_we  = 1;
                 montprod_dest_new   = MONTPROD_DEST_P;
                 montprod_dest_we    = 1;
                 montprod_calc       = 1;
-                modexp_ctrl_new = CTRL_ITERATE_P_P;
-                modexp_ctrl_we  = 1;
+                modexp_ctrl_new     = CTRL_ITERATE_P_P;
+                modexp_ctrl_we      = 1;
               end
 
         CTRL_ITERATE_P_P:
@@ -799,24 +952,26 @@ module modexp(
 
         CTRL_ITERATE_END:
           begin
-            if (last_iteration == 1'b0)
+            if (!last_iteration)
               begin
                 modexp_ctrl_new = CTRL_ITERATE;
                 modexp_ctrl_we  = 1;
               end
             else
               begin
-                montprod_select_new = MONTPROD_SELECT_ONE_Z;
+                montprod_select_new = MONTPROD_SELECT_Z_ONE;
                 montprod_select_we  = 1;
+                montprod_dest_new   = MONTPROD_DEST_Z;
+                montprod_dest_we    = 1;
                 montprod_calc       = 1;
-                modexp_ctrl_new = CTRL_CALCULATE_ZN;
-                modexp_ctrl_we  = 1;
+                modexp_ctrl_new     = CTRL_CALCULATE_ZN;
+                modexp_ctrl_we      = 1;
               end
           end
 
         CTRL_CALCULATE_ZN:
           begin
-            if (montprod_ready == 1'b1)
+            if (montprod_ready)
               begin
                 modexp_ctrl_new = CTRL_DONE;
                 modexp_ctrl_we  = 1;
@@ -825,10 +980,11 @@ module modexp(
 
         CTRL_DONE:
           begin
-            ready_new           = 1;
-            ready_we            = 1;
-            modexp_ctrl_new     = CTRL_IDLE;
-            modexp_ctrl_we      = 1;
+            clear_start     = 1'b1;
+            ready_new       = 1'b1;
+            ready_we        = 1'b1;
+            modexp_ctrl_new = CTRL_IDLE;
+            modexp_ctrl_we  = 1;
           end
 
         default:
diff --git a/src/rtl/montprod.v b/src/rtl/montprod.v
index 32bbdec..2ed1f5d 100644
--- a/src/rtl/montprod.v
+++ b/src/rtl/montprod.v
@@ -63,6 +63,7 @@ module montprod(
   //----------------------------------------------------------------
   // Internal constant and parameter definitions.
   //----------------------------------------------------------------
+
   localparam CTRL_IDLE           = 4'h0;
   localparam CTRL_INIT_S         = 4'h1;
   localparam CTRL_LOOP_INIT      = 4'h2;
@@ -217,9 +218,11 @@ module montprod(
         SMUX_SHR:
           s_mem_new = shr_adiv2;
       endcase
-      $display("SMUX%x: %x", s_mux_reg, s_mem_new);
+      if (DEBUG)
+        $display("SMUX%x: %x", s_mux_reg, s_mem_new);
     end
 
+
   //----------------------------------------------------------------
   // reg_update
   //
@@ -286,7 +289,8 @@ module montprod(
            //opa_addr will point to length-1 to get A LSB.
            //s_read_addr will point to length-1
            q = s_mem_read_data[0] ^ (opa_data[0] & b);
-           $display("s_mem_read_data: %x opa_data %x b %x q %x B_bit_index_reg %x", s_mem_read_data, opa_data, b, q, B_bit_index_reg);
+           if (DEBUG)
+             $display("s_mem_read_data: %x opa_data %x b %x q %x B_bit_index_reg %x", s_mem_read_data, opa_data, b, q, B_bit_index_reg);
         end
    end
 
@@ -296,6 +300,7 @@ module montprod(
   //----------------------------------------------------------------
   always @*
    begin : loop_counter_process
+     loop_counter_new  = loop_counter;
       length_m1        = length - 1'b1;
       loop_counter_dec = loop_counter - 1'b1;
       B_word_index     = loop_counter[12:5];
@@ -560,7 +565,8 @@ module montprod(
 
         CTRL_EMIT_S:
            begin
-             $display("EMIT_S word_index: %d", word_index);
+             if (DEBUG)
+               $display("EMIT_S word_index: %d", word_index);
              if (word_index_prev == 8'h0)
                begin
                  montprod_ctrl_new = CTRL_DONE;