execute1: Move EXTS* instruction back into execute1

This moves the sign extension done by the extsb, extsh and extsw
instructions back into execute1.  This means that we no longer need
any data formatting in writeback for results coming from execute1,
so this modifies writeback so the data formatter inputs come
directly from the loadstore unit output.  The condition code
updates for RC=1 form instructions are now done on the value from
execute1 rather than the output of the data formatter, which should
help timing.

Signed-off-by: Paul Mackerras <paulus@ozlabs.org>

common.vhdl

@@ -240,16 +240,14 @@ package common is
 	write_enable : std_ulogic;
 	write_reg: gspr_index_t;
 	write_data: std_ulogic_vector(63 downto 0);
-	write_len : std_ulogic_vector(3 downto 0);
 	write_cr_enable : std_ulogic;
 	write_cr_mask : std_ulogic_vector(7 downto 0);
 	write_cr_data : std_ulogic_vector(31 downto 0);
 	write_xerc_enable : std_ulogic;
 	xerc : xer_common_t;
-	sign_extend: std_ulogic;
     end record;
     constant Execute1ToWritebackInit : Execute1ToWritebackType := (valid => '0', rc => '0', write_enable => '0',
-								   write_cr_enable => '0', sign_extend => '0',
+								   write_cr_enable => '0',
 								   write_xerc_enable => '0', xerc => xerc_init,
 								   others => (others => '0'));
 

execute1.vhdl

@@ -192,6 +192,7 @@ begin
         variable sign1, sign2 : std_ulogic;
         variable abs1, abs2 : signed(63 downto 0);
 	variable overflow : std_ulogic;
+	variable negative : std_ulogic;
     begin
 	result := (others => '0');
 	result_with_carry := (others => '0');
@@ -335,8 +336,6 @@ begin
 
 	    v.e.valid := '1';
 	    v.e.write_reg := e_in.write_reg;
-	    v.e.write_len := x"8";
-	    v.e.sign_extend := '0';
 	    v.slow_op_dest := gspr_to_gpr(e_in.write_reg);
 	    v.slow_op_rc := e_in.rc;
 	    v.slow_op_oe := e_in.oe;
@@ -438,10 +437,19 @@ begin
 	    when OP_CNTZ =>
 		result := countzero_result;
 		result_en := '1';
-	    when OP_EXTS =>
-		v.e.write_len := e_in.data_len;
-		v.e.sign_extend := '1';
-		result := e_in.read_data3;
+            when OP_EXTS =>
+                -- note data_len is a 1-hot encoding
+		negative := (e_in.data_len(0) and e_in.read_data3(7)) or
+			    (e_in.data_len(1) and e_in.read_data3(15)) or
+			    (e_in.data_len(2) and e_in.read_data3(31));
+		result := (others => negative);
+		if e_in.data_len(2) = '1' then
+		    result(31 downto 16) := e_in.read_data3(31 downto 16);
+		end if;
+		if e_in.data_len(2) = '1' or e_in.data_len(1) = '1' then
+		    result(15 downto 8) := e_in.read_data3(15 downto 8);
+		end if;
+		result(7 downto 0) := e_in.read_data3(7 downto 0);
 		result_en := '1';
 	    when OP_ISEL =>
 		crbit := to_integer(unsigned(insn_bc(e_in.insn)));
@@ -660,8 +668,6 @@ begin
 	    result_en := '1';
 	    result := r.next_lr;
 	    v.e.write_reg := fast_spr_num(SPR_LR);
-	    v.e.write_len := x"8";
-	    v.e.sign_extend := '0';
 	    v.e.valid := '1';
 	elsif r.mul_in_progress = '1' or r.div_in_progress = '1' then
 	    if (r.mul_in_progress = '1' and multiply_to_x.valid = '1') or
@@ -687,8 +693,6 @@ begin
 		    v.e.xerc.so := v.slow_op_xerc.so or overflow;
 		end if;
 		v.e.valid := '1';
-		v.e.write_len := x"8";
-		v.e.sign_extend := '0';
 	    else
 		stall_out <= '1';
 		v.mul_in_progress := r.mul_in_progress;

writeback.vhdl

@@ -42,7 +42,6 @@ architecture behaviour of writeback is
     signal sign_extend : std_ulogic;
     signal negative : std_ulogic;
     signal second_word : std_ulogic;
-    signal zero : std_ulogic;
 begin
     writeback_0: process(clk)
     begin
@@ -62,6 +61,8 @@ begin
         variable k : unsigned(3 downto 0);
 	variable cf: std_ulogic_vector(3 downto 0);
 	variable xe: xer_common_t;
+        variable zero : std_ulogic;
+        variable sign : std_ulogic;
     begin
         x := "" & e_in.valid;
         y := "" & l_in.valid;
@@ -85,10 +86,7 @@ begin
 
         rc <= '0';
         brev_lenm1 <= "000";
-        byte_offset <= "000";
-        data_len <= x"8";
         partial_write <= '0';
-        sign_extend <= '0';
         second_word <= '0';
 	xe := e_in.xerc;
 	data_in <= (others => '0');
@@ -96,9 +94,6 @@ begin
         if e_in.write_enable = '1' then
             w_out.write_reg <= e_in.write_reg;
             w_out.write_enable <= '1';
-	    data_in <= e_in.write_data;
-            data_len <= unsigned(e_in.write_len);
-            sign_extend <= e_in.sign_extend;
             rc <= e_in.rc;
         end if;
 
@@ -113,12 +108,11 @@ begin
             c_out.write_xerc_data <= e_in.xerc;
 	end if;
 
+        sign_extend <= l_in.sign_extend;
+        data_len <= unsigned(l_in.write_len);
+        byte_offset <= unsigned(l_in.write_shift);
 	if l_in.write_enable = '1' then
             w_out.write_reg <= gpr_to_gspr(l_in.write_reg);
-            data_in <= l_in.write_data;
-            data_len <= unsigned(l_in.write_len);
-            byte_offset <= unsigned(l_in.write_shift);
-            sign_extend <= l_in.sign_extend;
             if l_in.byte_reverse = '1' then
                 brev_lenm1 <= unsigned(l_in.write_len(2 downto 0)) - 1;
             end if;
@@ -138,7 +132,7 @@ begin
         end loop;
         for i in 0 to 7 loop
             j := to_integer(perm(i)) * 8;
-            data_permuted(i * 8 + 7 downto i * 8) <= data_in(j + 7 downto j);
+            data_permuted(i * 8 + 7 downto i * 8) <= l_in.write_data(j + 7 downto j);
         end loop;
 
         -- If the data can arrive split over two cycles, this will be correct
@@ -160,16 +154,12 @@ begin
                 trim_ctl(i) <= '0' & (negative and sign_extend);
             end if;
         end loop;
-	zero <= not negative;
         for i in 0 to 7 loop
             case trim_ctl(i) is
                 when "11" =>
                     data_trimmed(i * 8 + 7 downto i * 8) <= data_latched(i * 8 + 7 downto i * 8);
                 when "10" =>
                     data_trimmed(i * 8 + 7 downto i * 8) <= data_permuted(i * 8 + 7 downto i * 8);
-		    if or data_permuted(i * 8 + 7 downto i * 8) /= '0' then
-			zero <= '0';
-		    end if;
                 when "01" =>
                     data_trimmed(i * 8 + 7 downto i * 8) <= x"FF";
                 when others =>
@@ -178,14 +168,21 @@ begin
         end loop;
 
         -- deliver to regfile
-        w_out.write_data <= data_trimmed;
+        if l_in.write_enable = '1' then
+            w_out.write_data <= data_trimmed;
+        else
+            w_out.write_data <= e_in.write_data;
+        end if;
 
         -- Perform CR0 update for RC forms
+        -- Note that loads never have a form with an RC bit, therefore this can test e_in.write_data
         if rc = '1' then
+            sign := e_in.write_data(63);
+            zero := not (or e_in.write_data);
             c_out.write_cr_enable <= '1';
             c_out.write_cr_mask <= num_to_fxm(0);
-	    cf(3) := negative;
-	    cf(2) := not negative and not zero;
+	    cf(3) := sign;
+	    cf(2) := not sign and not zero;
 	    cf(1) := zero;
 	    cf(0) := xe.so;
 	    c_out.write_cr_data(31 downto 28) <= cf;