core: Restore bypass path from execute1

This changes the bypass path.  Previously it went from after
execute1's output to after decode2's output.  Now it goes from before
execute1's output register to before decode2's output register.  The
reason is that the new path will be simpler to manage when there are
possibly multiple instructions in flight.  This means that the
bypassing can be managed inside decode2 and control.

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

common.vhdl

@@ -137,6 +137,7 @@ package common is
         valid : std_ulogic;
     end record;
     constant instr_tag_init : instr_tag_t := (tag => 0, valid => '0');
+    function tag_match(tag1 : instr_tag_t; tag2 : instr_tag_t) return boolean;
 
     type irq_state_t is (WRITE_SRR0, WRITE_SRR1);
 
@@ -203,6 +204,12 @@ package common is
         redirect_nia : std_ulogic_vector(63 downto 0);
     end record;
 
+    type bypass_data_t is record
+        tag  : instr_tag_t;
+        data : std_ulogic_vector(63 downto 0);
+    end record;
+    constant bypass_data_init : bypass_data_t := (tag => instr_tag_init, data => (others => '0'));
+
     type Decode2ToExecute1Type is record
 	valid: std_ulogic;
         unit : unit_t;
@@ -217,9 +224,6 @@ package common is
 	read_data1: std_ulogic_vector(63 downto 0);
 	read_data2: std_ulogic_vector(63 downto 0);
 	read_data3: std_ulogic_vector(63 downto 0);
-        bypass_data1: std_ulogic;
-        bypass_data2: std_ulogic;
-        bypass_data3: std_ulogic;
 	cr: std_ulogic_vector(31 downto 0);
         bypass_cr : std_ulogic;
 	xerc: xer_common_t;
@@ -250,7 +254,7 @@ package common is
     end record;
     constant Decode2ToExecute1Init : Decode2ToExecute1Type :=
 	(valid => '0', unit => NONE, fac => NONE, insn_type => OP_ILLEGAL, instr_tag => instr_tag_init,
-         write_reg_enable => '0', bypass_data1 => '0', bypass_data2 => '0', bypass_data3 => '0',
+         write_reg_enable => '0',
          bypass_cr => '0', lr => '0', br_abs => '0', rc => '0', oe => '0', invert_a => '0', addm1 => '0',
 	 invert_out => '0', input_carry => ZERO, output_carry => '0', input_cr => '0', output_cr => '0',
 	 is_32bit => '0', is_signed => '0', xerc => xerc_init, reserve => '0', br_pred => '0',
@@ -644,4 +648,9 @@ package body common is
     begin
         return "10" & f;
     end;
+
+    function tag_match(tag1 : instr_tag_t; tag2 : instr_tag_t) return boolean is
+    begin
+        return tag1.valid = '1' and tag2.valid = '1' and tag1.tag = tag2.tag;
+    end;
 end common;

control.vhdl

@@ -6,6 +6,7 @@ use work.common.all;
 
 entity control is
     generic (
+        EX1_BYPASS : boolean := true;
         PIPELINE_DEPTH : natural := 2
         );
     port (
@@ -23,7 +24,6 @@ entity control is
 
         gpr_write_valid_in  : in std_ulogic;
         gpr_write_in        : in gspr_index_t;
-        gpr_bypassable      : in std_ulogic;
 
         gpr_a_read_valid_in : in std_ulogic;
         gpr_a_read_in       : in gspr_index_t;
@@ -34,6 +34,8 @@ entity control is
         gpr_c_read_valid_in : in std_ulogic;
         gpr_c_read_in       : in gspr_index_t;
 
+        execute_next_tag    : in instr_tag_t;
+
         cr_read_in          : in std_ulogic;
         cr_write_in         : in std_ulogic;
         cr_bypassable       : in std_ulogic;
@@ -81,19 +83,11 @@ architecture rtl of control is
 
     signal instr_tag  : instr_tag_t;
 
-    signal gpr_tag_a : instr_tag_t;
-    signal gpr_tag_b : instr_tag_t;
-    signal gpr_tag_c : instr_tag_t;
     signal gpr_tag_stall : std_ulogic;
 
     signal curr_tag : tag_number_t;
     signal next_tag : tag_number_t;
 
-    function tag_match(tag1 : instr_tag_t; tag2 : instr_tag_t) return boolean is
-    begin
-        return tag1.valid = '1' and tag2.valid = '1' and tag1.tag = tag2.tag;
-    end;
-
 begin
     cr_hazard0: entity work.cr_hazard
         generic map (
@@ -115,10 +109,6 @@ begin
             use_bypass         => cr_bypass
             );
 
-    gpr_bypass_a <= '0';
-    gpr_bypass_b <= '0';
-    gpr_bypass_c <= '0';
-
     control0: process(clk)
     begin
         if rising_edge(clk) then
@@ -165,6 +155,9 @@ begin
         variable tag_s : instr_tag_t;
         variable tag_t : instr_tag_t;
         variable incr_tag : tag_number_t;
+        variable byp_a : std_ulogic;
+        variable byp_b : std_ulogic;
+        variable byp_c : std_ulogic;
     begin
         tag_a := instr_tag_init;
         for i in tag_number_t loop
@@ -196,10 +189,27 @@ begin
         if tag_match(tag_c, complete_in) then
             tag_c.valid := '0';
         end if;
-        gpr_tag_a <= tag_a;
-        gpr_tag_b <= tag_b;
-        gpr_tag_c <= tag_c;
-        gpr_tag_stall <= tag_a.valid or tag_b.valid or tag_c.valid;
+
+        byp_a := '0';
+        if EX1_BYPASS and tag_match(execute_next_tag, tag_a) then
+            byp_a := '1';
+        end if;
+        byp_b := '0';
+        if EX1_BYPASS and tag_match(execute_next_tag, tag_b) then
+            byp_b := '1';
+        end if;
+        byp_c := '0';
+        if EX1_BYPASS and tag_match(execute_next_tag, tag_c) then
+            byp_c := '1';
+        end if;
+
+        gpr_bypass_a <= byp_a;
+        gpr_bypass_b <= byp_b;
+        gpr_bypass_c <= byp_c;
+
+        gpr_tag_stall <= (tag_a.valid and not byp_a) or
+                         (tag_b.valid and not byp_b) or
+                         (tag_c.valid and not byp_c);
 
         incr_tag := curr_tag;
         instr_tag.tag <= curr_tag;

core.vhdl

@@ -67,6 +67,7 @@ architecture behave of core is
     -- execute signals
     signal execute1_to_writeback: Execute1ToWritebackType;
     signal execute1_to_fetch1: Execute1ToFetch1Type;
+    signal execute1_bypass: bypass_data_t;
 
     -- load store signals
     signal execute1_to_loadstore1: Execute1ToLoadstore1Type;
@@ -273,6 +274,7 @@ begin
             r_out => decode2_to_register_file,
             c_in => cr_file_to_decode2,
             c_out => decode2_to_cr_file,
+            execute_bypass => execute1_bypass,
             log_out => log_data(119 downto 110)
             );
     decode2_busy_in <= ex1_busy_out;
@@ -330,6 +332,7 @@ begin
             f_out => execute1_to_fetch1,
             fp_out => execute1_to_fpu,
             e_out => execute1_to_writeback,
+            bypass_data => execute1_bypass,
 	    icache_inval => ex1_icache_inval,
             dbg_msr_out => msr,
             terminate_out => terminate,

decode2.vhdl

@@ -37,6 +37,8 @@ entity decode2 is
         c_in  : in CrFileToDecode2Type;
         c_out : out Decode2ToCrFileType;
 
+        execute_bypass   : in bypass_data_t;
+
         log_out : out std_ulogic_vector(9 downto 0)
 	);
 end entity decode2;
@@ -285,7 +287,6 @@ architecture behaviour of decode2 is
 
     signal gpr_write_valid : std_ulogic;
     signal gpr_write : gspr_index_t;
-    signal gpr_bypassable  : std_ulogic;
 
     signal gpr_a_read_valid : std_ulogic;
     signal gpr_a_read       : gspr_index_t;
@@ -308,6 +309,7 @@ architecture behaviour of decode2 is
 begin
     control_0: entity work.control
 	generic map (
+            EX1_BYPASS => EX1_BYPASS,
             PIPELINE_DEPTH => 1
             )
 	port map (
@@ -325,7 +327,6 @@ begin
 
             gpr_write_valid_in => gpr_write_valid,
             gpr_write_in       => gpr_write,
-            gpr_bypassable     => gpr_bypassable,
 
             gpr_a_read_valid_in  => gpr_a_read_valid,
             gpr_a_read_in        => gpr_a_read,
@@ -336,6 +337,8 @@ begin
             gpr_c_read_valid_in  => gpr_c_read_valid,
             gpr_c_read_in        => gpr_c_read,
 
+            execute_next_tag     => execute_bypass.tag,
+
             cr_read_in           => d_in.decode.input_cr,
             cr_write_in          => cr_write_valid,
             cr_bypass            => cr_bypass,
@@ -457,13 +460,7 @@ begin
         v.e.fac := d_in.decode.facility;
         v.e.instr_tag := instr_tag;
         v.e.read_reg1 := decoded_reg_a.reg;
-        v.e.read_data1 := decoded_reg_a.data;
-        v.e.bypass_data1 := gpr_a_bypass;
         v.e.read_reg2 := decoded_reg_b.reg;
-        v.e.read_data2 := decoded_reg_b.data;
-        v.e.bypass_data2 := gpr_b_bypass;
-        v.e.read_data3 := decoded_reg_c.data;
-        v.e.bypass_data3 := gpr_c_bypass;
         v.e.write_reg := decoded_reg_o.reg;
         v.e.write_reg_enable := decoded_reg_o.reg_valid;
         v.e.rc := decode_rc(d_in.decode.rc, d_in.insn);
@@ -499,16 +496,32 @@ begin
             end if;
         end if;
 
+        -- See if any of the operands can get their value via the bypass path.
+        case gpr_a_bypass is
+            when '1' =>
+                v.e.read_data1 := execute_bypass.data;
+            when others =>
+                v.e.read_data1 := decoded_reg_a.data;
+        end case;
+        case gpr_b_bypass is
+            when '1' =>
+                v.e.read_data2 := execute_bypass.data;
+            when others =>
+                v.e.read_data2 := decoded_reg_b.data;
+        end case;
+        case gpr_c_bypass is
+            when '1' =>
+                v.e.read_data3 := execute_bypass.data;
+            when others =>
+                v.e.read_data3 := decoded_reg_c.data;
+        end case;
+
         -- issue control
         control_valid_in <= d_in.valid;
         control_sgl_pipe <= d_in.decode.sgl_pipe;
 
         gpr_write_valid <= v.e.write_reg_enable;
         gpr_write <= decoded_reg_o.reg;
-        gpr_bypassable <= '0';
-        if EX1_BYPASS and d_in.decode.unit = ALU then
-            gpr_bypassable <= '1';
-        end if;
 
         gpr_a_read_valid <= decoded_reg_a.reg_valid;
         gpr_a_read <= decoded_reg_a.reg;
@@ -554,9 +567,9 @@ begin
                             r.e.valid &
                             stopped_out &
                             stall_out &
-                            r.e.bypass_data3 &
-                            r.e.bypass_data2 &
-                            r.e.bypass_data1;
+                            gpr_a_bypass &
+                            gpr_b_bypass &
+                            gpr_c_bypass;
             end if;
         end process;
         log_out <= log_data;

execute1.vhdl

@@ -37,6 +37,7 @@ entity execute1 is
         fp_out : out Execute1ToFPUType;
 
 	e_out : out Execute1ToWritebackType;
+        bypass_data : out bypass_data_t;
 
         dbg_msr_out : out std_ulogic_vector(63 downto 0);
 
@@ -283,9 +284,9 @@ begin
     dbg_msr_out <= ctrl.msr;
     log_rd_addr <= r.log_addr_spr;
 
-    a_in <= r.e.write_data when EX1_BYPASS and e_in.bypass_data1 = '1' else e_in.read_data1;
-    b_in <= r.e.write_data when EX1_BYPASS and e_in.bypass_data2 = '1' else e_in.read_data2;
-    c_in <= r.e.write_data when EX1_BYPASS and e_in.bypass_data3 = '1' else e_in.read_data3;
+    a_in <= e_in.read_data1;
+    b_in <= e_in.read_data2;
+    c_in <= e_in.read_data3;
 
     busy_out <= l_in.busy or r.busy or fp_in.busy;
     valid_in <= e_in.valid and not busy_out;
@@ -1270,6 +1271,10 @@ begin
 	v.e.write_enable := current.write_reg_enable and v.e.valid and not exception;
         v.e.rc := current.rc and v.e.valid and not exception;
 
+        bypass_data.tag.valid <= current.instr_tag.valid and current.write_reg_enable and v.e.valid;
+        bypass_data.tag.tag <= current.instr_tag.tag;
+        bypass_data.data <= v.e.write_data;
+
         -- Defer completion for one cycle when redirecting.
         -- This also ensures r.busy = 1 when ctrl.irq_state = WRITE_SRR1
         if v.redirect = '1' then