stnolting · stnolting · Apr 21, 2024 · Apr 20, 2024 · Apr 20, 2024 · Apr 21, 2024
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -29,6 +29,7 @@ mimpid = 0x01040312 -> Version 01.04.03.12 -> v1.4.3.12
 
 | Date | Version | Comment | Link |
 |:----:|:-------:|:--------|:----:|
+| 21.04.2024 | 1.9.8.5 | rtl cleanups and (area) optimizations | [#880](https://github.com/stnolting/neorv32/pull/880) |
 | 16.04.2024 | 1.9.8.4 | :warning: use a 4-bit FIRQ select instead of a 16-bit FIRQ mask for DMA auto-trigger configuration | [#877](https://github.com/stnolting/neorv32/pull/877) |
 | 15.04.2024 | 1.9.8.3 | :warning: simplify XBUS gateway logic and configuration generics; only "pipelined Wishbone" protocol is supported now | [#876](https://github.com/stnolting/neorv32/pull/876) |
 | 14.04.2024 | 1.9.8.2 | :warning: rename SLINK data interface registers; minor CPU control logic/area optimizations | [#874](https://github.com/stnolting/neorv32/pull/874) |

diff --git a/rtl/core/neorv32_cfs.vhd b/rtl/core/neorv32_cfs.vhd
@@ -209,7 +209,7 @@ begin
   cfs_reg_rd(0) <= bin_to_gray_f(cfs_reg_wr(0)); -- convert binary to gray code
   cfs_reg_rd(1) <= gray_to_bin_f(cfs_reg_wr(1)); -- convert gray to binary code
   cfs_reg_rd(2) <= bit_rev_f(cfs_reg_wr(2)); -- bit reversal
-  cfs_reg_rd(3) <= bswap32_f(cfs_reg_wr(3)); -- byte swap (endianness conversion)
+  cfs_reg_rd(3) <= bswap_f(cfs_reg_wr(3)); -- byte swap (endianness conversion)
 
 
 end neorv32_cfs_rtl;
diff --git a/rtl/core/neorv32_cpu_cp_bitmanip.vhd b/rtl/core/neorv32_cpu_cp_bitmanip.vhd
diff --git a/rtl/core/neorv32_cpu_cp_shifter.vhd b/rtl/core/neorv32_cpu_cp_shifter.vhd
@@ -53,7 +53,7 @@ architecture neorv32_cpu_cp_shifter_rtl of neorv32_cpu_cp_shifter is
   -- barrel shifter --
   type bs_level_t is array (index_size_f(XLEN) downto 0) of std_ulogic_vector(XLEN-1 downto 0);
   signal bs_level  : bs_level_t;
-  signal bs_mask   : std_ulogic;
+  signal bs_sign   : std_ulogic;
   signal bs_start  : std_ulogic;
   signal bs_result : std_ulogic_vector(XLEN-1 downto 0);
 
@@ -109,14 +109,14 @@ begin
   if FAST_SHIFT_EN generate
 
     -- input layer: convert left shifts to right shifts by bit-reversal --
-    bs_level(index_size_f(XLEN)) <= bit_rev_f(rs1_i) when (ctrl_i.ir_funct3(2) = '0') else rs1_i;
-    bs_mask <= rs1_i(XLEN-1) and ctrl_i.ir_funct12(10); -- MSBs mask for arithmetic/logic shifts
+    bs_level(0) <= bit_rev_f(rs1_i) when (ctrl_i.ir_funct3(2) = '0') else rs1_i;
+    bs_sign <= rs1_i(XLEN-1) and ctrl_i.ir_funct12(10); -- sign extension for arithmetic shifts
 
-    -- shifter layers: right-shifts only --
+    -- mux layers: right-shifts only --
     barrel_shifter_core:
-    for i in index_size_f(XLEN)-1 downto 0 generate
-      bs_level(i)(XLEN-1 downto XLEN-(2**i)) <= (others => bs_mask)               when (shamt_i(i) = '1') else bs_level(i+1)(XLEN-1 downto XLEN-(2**i));
-      bs_level(i)((XLEN-(2**i))-1 downto 0)  <= bs_level(i+1)(XLEN-1 downto 2**i) when (shamt_i(i) = '1') else bs_level(i+1)((XLEN-(2**i))-1 downto 0);
+    for i in 0 to index_size_f(XLEN)-1 generate
+      bs_level(i+1)(XLEN-1 downto XLEN-(2**i)) <= (others => bs_sign)             when (shamt_i(i) = '1') else bs_level(i)(XLEN-1 downto XLEN-(2**i));
+      bs_level(i+1)((XLEN-(2**i))-1 downto 0)  <= bs_level(i)(XLEN-1 downto 2**i) when (shamt_i(i) = '1') else bs_level(i)((XLEN-(2**i))-1 downto 0);
     end generate;
 
     -- pipeline register --
@@ -127,7 +127,7 @@ begin
         bs_result <= (others => '0');
       elsif rising_edge(clk_i) then -- this register stage can be moved by the register balancing
         bs_start  <= start_i;
-        bs_result <= bs_level(0);
+        bs_result <= bs_level(index_size_f(XLEN));
       end if;
     end process barrel_shifter_buf;
 

diff --git a/rtl/core/neorv32_cpu_regfile.vhd b/rtl/core/neorv32_cpu_regfile.vhd
@@ -57,10 +57,8 @@ architecture neorv32_cpu_regfile_rtl of neorv32_cpu_regfile is
 
   -- access --
   signal rf_we     : std_ulogic; -- write enable
-  signal rf_we_sel : std_ulogic_vector((2**addr_bits_c)-1 downto 0); -- one-hot write enable
   signal rd_zero   : std_ulogic; -- writing to x0?
   signal opa_addr  : std_ulogic_vector(4 downto 0); -- rs1/rd address
-  signal rd_addr   : std_ulogic_vector(4 downto 0); -- rd address
   signal rs3_addr  : std_ulogic_vector(4 downto 0); -- rs3 address
   signal rs4_addr  : std_ulogic_vector(4 downto 0); -- rs4 address
 
@@ -101,16 +99,6 @@ begin
   register_file_asic:
   if RST_EN generate
 
-    -- "write" to x0 if no write access --
-    rd_addr <= ctrl_i.rf_rd(addr_bits_c-1 downto 0) when (ctrl_i.rf_wb_en = '1') else (others => '0');
-
-    -- write enable decoder --
-    we_decode: process(rd_addr)
-    begin
-      rf_we_sel <= (others => '0');
-      rf_we_sel(to_integer(unsigned(rd_addr(addr_bits_c-1 downto 0)))) <= '1';
-    end process we_decode;
-
     -- individual registers --
     reg_gen:
     for i in 1 to (2**addr_bits_c)-1 generate
@@ -119,7 +107,7 @@ begin
         if (rstn_i = '0') then
           reg_file(i) <= (others => '0');
         elsif rising_edge(clk_i) then
-          if (rf_we_sel(i) = '1') then
+          if (unsigned(ctrl_i.rf_rd(addr_bits_c-1 downto 0)) = to_unsigned(i, addr_bits_c)) and (ctrl_i.rf_wb_en = '1') then
             reg_file(i) <= rd_i;
           end if;
         end if;

diff --git a/rtl/core/neorv32_dma.vhd b/rtl/core/neorv32_dma.vhd
@@ -319,7 +319,7 @@ begin
   -- -------------------------------------------------------------------------------------------
 
   -- endianness conversion --
-  align_end <= dma_rsp_i.data when (config.endian = '0') else bswap32_f(dma_rsp_i.data);
+  align_end <= dma_rsp_i.data when (config.endian = '0') else bswap_f(dma_rsp_i.data);
 
   -- source data alignment --
   src_align: process(rstn_i, clk_i)

diff --git a/rtl/core/neorv32_package.vhd b/rtl/core/neorv32_package.vhd
@@ -29,7 +29,7 @@ package neorv32_package is
 
   -- Architecture Constants -----------------------------------------------------------------
   -- -------------------------------------------------------------------------------------------
-  constant hw_version_c : std_ulogic_vector(31 downto 0) := x"01090804"; -- hardware version
+  constant hw_version_c : std_ulogic_vector(31 downto 0) := x"01090805"; -- hardware version
   constant archid_c     : natural := 19; -- official RISC-V architecture ID
   constant XLEN         : natural := 32; -- native data path width
 
@@ -694,7 +694,7 @@ package neorv32_package is
   function to_hstring32_f(input : std_ulogic_vector(31 downto 0)) return string;
   function bit_rev_f(input : std_ulogic_vector) return std_ulogic_vector;
   function is_power_of_two_f(input : natural) return boolean;
-  function bswap32_f(input : std_ulogic_vector) return std_ulogic_vector;
+  function bswap_f(input : std_ulogic_vector) return std_ulogic_vector;
   function popcount_f(input : std_ulogic_vector) return natural;
   function leading_zeros_f(input : std_ulogic_vector) return natural;
   impure function mem32_init_f(init : mem32_t; depth : natural) return mem32_t;
@@ -1034,7 +1034,7 @@ package body neorv32_package is
     end case;
   end function su_undefined_f;
 
-  -- Convert std_ulogic_vector to lowercase HEX char ----------------------------------------
+  -- Convert std_ulogic_vector to lowercase hex char ----------------------------------------
   -- -------------------------------------------------------------------------------------------
   function to_hexchar_f(input : std_ulogic_vector(3 downto 0)) return character is
     variable hex_v : string(1 to 16);
@@ -1073,41 +1073,42 @@ package body neorv32_package is
   -- Test if input number is a power of two -------------------------------------------------
   -- -------------------------------------------------------------------------------------------
   function is_power_of_two_f(input : natural) return boolean is
-    variable tmp : unsigned(31 downto 0);
+    variable tmp_v : unsigned(31 downto 0);
   begin
     if (input = 0) then
       return false;
     elsif (input = 1) then
       return true;
     else
-      tmp := to_unsigned(input, 32);
-      if ((tmp and (tmp - 1)) = 0) then
+      tmp_v := to_unsigned(input, 32);
+      if ((tmp_v and (tmp_v - 1)) = 0) then
         return true;
       else
         return false;
       end if;
     end if;
   end function is_power_of_two_f;
 
-  -- Swap all bytes of a 32-bit word (endianness conversion) --------------------------------
+  -- Swap all bytes of a N*8-bit word (endianness conversion) -------------------------------
   -- -------------------------------------------------------------------------------------------
-  function bswap32_f(input : std_ulogic_vector) return std_ulogic_vector is
+  function bswap_f(input : std_ulogic_vector) return std_ulogic_vector is
     variable output_v : std_ulogic_vector(input'range);
+    variable j        : natural range 0 to input'length/8;
   begin
-    output_v(07 downto 00) := input(31 downto 24);
-    output_v(15 downto 08) := input(23 downto 16);
-    output_v(23 downto 16) := input(15 downto 08);
-    output_v(31 downto 24) := input(07 downto 00);
+    for i in 0 to (input'length/8)-1 loop
+      j := ((input'length/8) - 1) - i;
+      output_v(i*8+7 downto i*8+0) := input(j*8+7 downto j*8+0);
+    end loop;
     return output_v;
-  end function bswap32_f;
+  end function bswap_f;
 
   -- Population count (number of set bits) --------------------------------------------------
   -- -------------------------------------------------------------------------------------------
   function popcount_f(input : std_ulogic_vector) return natural is
     variable cnt_v : natural range 0 to input'length;
   begin
     cnt_v := 0;
-    for i in input'length-1 downto 0 loop
+    for i in 0 to input'length-1 loop
       if (input(i) = '1') then
         cnt_v := cnt_v + 1;
       end if;

diff --git a/rtl/core/neorv32_xip.vhd b/rtl/core/neorv32_xip.vhd
@@ -318,7 +318,7 @@ begin
 
       when S_BUSY => -- wait for PHY to complete operation
       -- ------------------------------------------------------------
-        xip_rsp_o.data <= bswap32_f(phy_if.rdata); -- convert incrementing byte-read to little-endian
+        xip_rsp_o.data <= bswap_f(phy_if.rdata); -- convert incrementing byte-read to little-endian
         if (phy_if.busy = '0') then
           xip_rsp_o.ack     <= '1';
           arbiter.state_nxt <= S_IDLE;