core: Use a busy signal rather than a stall

This changes the instruction dependency tracking so that we can
generate a "busy" signal from execute1 and loadstore1 which comes
along one cycle later than the current "stall" signal.  This will
enable us to signal busy cycles only when we need to from loadstore1.

The "busy" signal from execute1/loadstore1 indicates "I didn't take
the thing you gave me on this cycle", as distinct from the previous
stall signal which meant "I took that but don't give me anything
next cycle".  That means that decode2 proactively gives execute1
a new instruction as soon as it has taken the previous one (assuming
there is a valid instruction available from decode1), and that then
sits in decode2's output until execute1 can take it.  So instructions
are issued by decode2 somewhat earlier than they used to be.

Decode2 now only signals a stall upstream when its output buffer is
full, meaning that we can fill up bubbles in the upstream pipe while a
long instruction is executing.  This gives a small boost in
performance.

This also adds dependency tracking for rA updates by update-form
load/store instructions.

The GPR and CR hazard detection machinery now has one extra stage,
which may not be strictly necessary.  Some of the code now really
only applies to PIPELINE_DEPTH=1.

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

common.vhdl

@@ -244,6 +244,7 @@ package common is
                                                                      others => (others => '0'));
 
     type Loadstore1ToExecute1Type is record
+        busy : std_ulogic;
         exception : std_ulogic;
         invalid : std_ulogic;
         perm_error : std_ulogic;

control.vhdl

@@ -15,7 +15,8 @@ entity control is
         complete_in         : in std_ulogic;
         valid_in            : in std_ulogic;
         flush_in            : in std_ulogic;
-	stall_in            : in std_ulogic;
+	busy_in             : in std_ulogic;
+        deferred            : in std_ulogic;
         sgl_pipe_in         : in std_ulogic;
         stop_mark_in        : in std_ulogic;
 
@@ -23,6 +24,9 @@ entity control is
         gpr_write_in        : in gspr_index_t;
         gpr_bypassable      : in std_ulogic;
 
+        update_gpr_write_valid : in std_ulogic;
+        update_gpr_write_reg : in gspr_index_t;
+
         gpr_a_read_valid_in : in std_ulogic;
         gpr_a_read_in       : in gspr_index_t;
 
@@ -72,7 +76,11 @@ begin
             )
         port map (
             clk                => clk,
-	    stall_in           => stall_in,
+            busy_in            => busy_in,
+	    deferred           => deferred,
+            complete_in        => complete_in,
+            flush_in           => flush_in,
+            issuing            => valid_out,
 
             gpr_write_valid_in => gpr_write_valid,
             gpr_write_in       => gpr_write_in,
@@ -80,6 +88,9 @@ begin
             gpr_read_valid_in  => gpr_a_read_valid_in,
             gpr_read_in        => gpr_a_read_in,
 
+            ugpr_write_valid   => update_gpr_write_valid,
+            ugpr_write_reg     => update_gpr_write_reg,
+
             stall_out          => stall_a_out,
             use_bypass         => gpr_bypass_a
             );
@@ -90,7 +101,11 @@ begin
             )
         port map (
             clk                => clk,
-	    stall_in           => stall_in,
+            busy_in            => busy_in,
+	    deferred           => deferred,
+            complete_in        => complete_in,
+            flush_in           => flush_in,
+            issuing            => valid_out,
 
             gpr_write_valid_in => gpr_write_valid,
             gpr_write_in       => gpr_write_in,
@@ -98,6 +113,9 @@ begin
             gpr_read_valid_in  => gpr_b_read_valid_in,
             gpr_read_in        => gpr_b_read_in,
 
+            ugpr_write_valid   => update_gpr_write_valid,
+            ugpr_write_reg     => update_gpr_write_reg,
+
             stall_out          => stall_b_out,
             use_bypass         => gpr_bypass_b
             );
@@ -110,7 +128,11 @@ begin
             )
         port map (
             clk                => clk,
-	    stall_in           => stall_in,
+            busy_in            => busy_in,
+	    deferred           => deferred,
+            complete_in        => complete_in,
+            flush_in           => flush_in,
+            issuing            => valid_out,
 
             gpr_write_valid_in => gpr_write_valid,
             gpr_write_in       => gpr_write_in,
@@ -118,6 +140,9 @@ begin
             gpr_read_valid_in  => gpr_c_read_valid_in,
             gpr_read_in        => gpr_c_read_in_fmt,
 
+            ugpr_write_valid   => update_gpr_write_valid,
+            ugpr_write_reg     => update_gpr_write_reg,
+
             stall_out          => stall_c_out,
             use_bypass         => gpr_bypass_c
             );
@@ -128,7 +153,11 @@ begin
             )
         port map (
             clk                => clk,
-	    stall_in           => stall_in,
+            busy_in            => busy_in,
+	    deferred           => deferred,
+            complete_in        => complete_in,
+            flush_in           => flush_in,
+            issuing            => valid_out,
 
             cr_read_in         => cr_read_in,
             cr_write_in        => cr_write_valid,
@@ -139,7 +168,8 @@ begin
     control0: process(clk)
     begin
         if rising_edge(clk) then
-            assert r_int.outstanding >= 0 and r_int.outstanding <= (PIPELINE_DEPTH+1) report "Outstanding bad " & integer'image(r_int.outstanding) severity failure;
+            assert rin_int.outstanding >= 0 and rin_int.outstanding <= (PIPELINE_DEPTH+1)
+                report "Outstanding bad " & integer'image(rin_int.outstanding) severity failure;
             r_int <= rin_int;
         end if;
     end process;
@@ -152,17 +182,18 @@ begin
         v_int := r_int;
 
         -- asynchronous
-        valid_tmp := valid_in and not flush_in and not stall_in;
-        stall_tmp := stall_in;
+        valid_tmp := valid_in and not flush_in;
+        stall_tmp := '0';
 
-        if complete_in = '1' then
+        if flush_in = '1' then
+            -- expect to see complete_in next cycle
+            v_int.outstanding := 1;
+        elsif complete_in = '1' then
             v_int.outstanding := r_int.outstanding - 1;
         end if;
 
         if rst = '1' then
-            v_int.state := IDLE;
-            v_int.outstanding := 0;
-            stall_tmp := '0';
+            v_int := reg_internal_init;
             valid_tmp := '0';
         end if;
 
@@ -227,7 +258,9 @@ begin
         end if;
 
         if valid_tmp = '1' then
-            v_int.outstanding := v_int.outstanding + 1;
+            if deferred = '0' then
+                v_int.outstanding := v_int.outstanding + 1;
+            end if;
             gpr_write_valid <= gpr_write_valid_in;
             cr_write_valid <= cr_write_in;
         else
@@ -237,7 +270,7 @@ begin
 
         -- update outputs
         valid_out <= valid_tmp;
-        stall_out <= stall_tmp;
+        stall_out <= stall_tmp or deferred;
 
         -- update registers
         rin_int <= v_int;

core.vhdl

@@ -82,11 +82,10 @@ architecture behave of core is
     signal icache_stall_out : std_ulogic;
     signal icache_stall_in : std_ulogic;
     signal decode1_stall_in : std_ulogic;
-    signal decode2_stall_in : std_ulogic;
+    signal decode2_busy_in : std_ulogic;
     signal decode2_stall_out : std_ulogic;
     signal ex1_icache_inval: std_ulogic;
-    signal ex1_stall_out: std_ulogic;
-    signal ls1_stall_out: std_ulogic;
+    signal ex1_busy_out: std_ulogic;
     signal dcache_stall_out: std_ulogic;
 
     signal flush: std_ulogic;
@@ -235,7 +234,7 @@ begin
         port map (
             clk => clk,
             rst => rst_dec2,
-	    stall_in => decode2_stall_in,
+	    busy_in => decode2_busy_in,
             stall_out => decode2_stall_out,
             flush_in => flush,
             complete_in => complete,
@@ -248,7 +247,7 @@ begin
             c_out => decode2_to_cr_file,
             log_out => log_data(119 downto 110)
             );
-    decode2_stall_in <= ex1_stall_out or ls1_stall_out;
+    decode2_busy_in <= ex1_busy_out;
 
     register_file_0: entity work.register_file
         generic map (
@@ -289,7 +288,7 @@ begin
             clk => clk,
             rst => rst_ex1,
             flush_out => flush,
-	    stall_out => ex1_stall_out,
+	    busy_out => ex1_busy_out,
             e_in => decode2_to_execute1,
             l_in => loadstore1_to_execute1,
             ext_irq_in => ext_irq,
@@ -317,7 +316,6 @@ begin
             m_out => loadstore1_to_mmu,
             m_in => mmu_to_loadstore1,
             dc_stall => dcache_stall_out,
-            stall_out => ls1_stall_out,
             log_out => log_data(149 downto 140)
             );
 

cr_hazard.vhdl

@@ -4,11 +4,15 @@ use ieee.numeric_std.all;
 
 entity cr_hazard is
     generic (
-        PIPELINE_DEPTH : natural := 2
+        PIPELINE_DEPTH : natural := 1
         );
     port(
         clk         : in std_ulogic;
-	stall_in    : in std_ulogic;
+        busy_in     : in std_ulogic;
+        deferred    : in std_ulogic;
+        complete_in : in std_ulogic;
+        flush_in    : in std_ulogic;
+        issuing     : in std_ulogic;
 
         cr_read_in  : in std_ulogic;
         cr_write_in : in std_ulogic;
@@ -22,7 +26,7 @@ architecture behaviour of cr_hazard is
     end record;
     constant pipeline_entry_init : pipeline_entry_type := (valid => '0');
 
-    type pipeline_t is array(0 to PIPELINE_DEPTH-1) of pipeline_entry_type;
+    type pipeline_t is array(0 to PIPELINE_DEPTH) of pipeline_entry_type;
     constant pipeline_t_init : pipeline_t := (others => pipeline_entry_init);
 
     signal r, rin : pipeline_t := pipeline_t_init;
@@ -30,9 +34,7 @@ begin
     cr_hazard0: process(clk)
     begin
         if rising_edge(clk) then
-	    if stall_in = '0' then
-		r <= rin;
-	    end if;
+            r <= rin;
         end if;
     end process;
 
@@ -41,22 +43,23 @@ begin
     begin
         v := r;
 
-        stall_out <= '0';
-        loop_0: for i in 0 to PIPELINE_DEPTH-1 loop
-            if (r(i).valid = cr_read_in) then
-                stall_out <= '1';
-            end if;
-        end loop;
-
-        v(0).valid := cr_write_in;
-        loop_1: for i in 0 to PIPELINE_DEPTH-2 loop
-            -- propagate to next slot
-            v(i+1) := r(i);
-        end loop;
+        -- XXX assumes PIPELINE_DEPTH = 1
+        if complete_in = '1' then
+            v(1).valid := '0';
+        end if;
+        stall_out <= cr_read_in and (v(0).valid or v(1).valid);
 
-        -- asynchronous output
-        if cr_read_in = '0' then
-            stall_out <= '0';
+        -- XXX assumes PIPELINE_DEPTH = 1
+        if busy_in = '0' then
+            v(1) := r(0);
+            v(0).valid := '0';
+        end if;
+        if deferred = '0' and issuing = '1' then
+            v(0).valid := cr_write_in;
+        end if;
+        if flush_in = '1' then
+            v(0).valid := '0';
+            v(1).valid := '0';
         end if;
 
         -- update registers

decode2.vhdl

@@ -17,7 +17,7 @@ entity decode2 is
         rst   : in std_ulogic;
 
         complete_in : in std_ulogic;
-        stall_in : in std_ulogic;
+        busy_in   : in std_ulogic;
         stall_out : out std_ulogic;
 
         stopped_out : out std_ulogic;
@@ -45,6 +45,8 @@ architecture behaviour of decode2 is
 
     signal r, rin : reg_type;
 
+    signal deferred : std_ulogic;
+
     signal log_data : std_ulogic_vector(9 downto 0);
 
     type decode_input_reg_t is record
@@ -200,6 +202,9 @@ architecture behaviour of decode2 is
     signal gpr_write : gspr_index_t;
     signal gpr_bypassable  : std_ulogic;
 
+    signal update_gpr_write_valid : std_ulogic;
+    signal update_gpr_write_reg : gspr_index_t;
+
     signal gpr_a_read_valid : std_ulogic;
     signal gpr_a_read :gspr_index_t;
     signal gpr_a_bypass : std_ulogic;
@@ -224,7 +229,8 @@ begin
 
             complete_in => complete_in,
             valid_in    => control_valid_in,
-            stall_in    => stall_in,
+            busy_in     => busy_in,
+            deferred    => deferred,
             flush_in    => flush_in,
             sgl_pipe_in => control_sgl_pipe,
             stop_mark_in => d_in.stop_mark,
@@ -233,6 +239,9 @@ begin
             gpr_write_in       => gpr_write,
             gpr_bypassable     => gpr_bypassable,
 
+            update_gpr_write_valid => update_gpr_write_valid,
+            update_gpr_write_reg => update_gpr_write_reg,
+
             gpr_a_read_valid_in  => gpr_a_read_valid,
             gpr_a_read_in        => gpr_a_read,
 
@@ -254,13 +263,17 @@ begin
             gpr_bypass_c => gpr_c_bypass
             );
 
+    deferred <= r.e.valid and busy_in;
+
     decode2_0: process(clk)
     begin
         if rising_edge(clk) then
-            if rin.e.valid = '1' then
-                report "execute " & to_hstring(rin.e.nia);
+            if rst = '1' or flush_in = '1' or deferred = '0' then
+                if rin.e.valid = '1' then
+                    report "execute " & to_hstring(rin.e.nia);
+                end if;
+                r <= rin;
             end if;
-            r <= rin;
         end if;
     end process;
 
@@ -358,6 +371,8 @@ begin
         if EX1_BYPASS and d_in.decode.unit = ALU then
             gpr_bypassable <= '1';
         end if;
+        update_gpr_write_valid <= d_in.decode.update;
+        update_gpr_write_reg <= decoded_reg_a.reg;
 
         gpr_a_read_valid <= decoded_reg_a.reg_valid;
         gpr_a_read <= decoded_reg_a.reg;
@@ -375,7 +390,7 @@ begin
             v.e.insn_type := OP_ILLEGAL;
         end if;
 
-        if rst = '1' then
+        if rst = '1' or flush_in = '1' then
             v.e := Decode2ToExecute1Init;
         end if;
 

execute1.vhdl

@@ -20,7 +20,7 @@ entity execute1 is
 
 	-- asynchronous
 	flush_out : out std_ulogic;
-	stall_out : out std_ulogic;
+	busy_out : out std_ulogic;
 
 	e_in  : in Decode2ToExecute1Type;
         l_in  : in Loadstore1ToExecute1Type;
@@ -48,6 +48,8 @@ end entity execute1;
 architecture behaviour of execute1 is
     type reg_type is record
 	e : Execute1ToWritebackType;
+        busy: std_ulogic;
+        terminate: std_ulogic;
 	lr_update : std_ulogic;
 	next_lr : std_ulogic_vector(63 downto 0);
 	mul_in_progress : std_ulogic;
@@ -62,7 +64,7 @@ architecture behaviour of execute1 is
         log_addr_spr : std_ulogic_vector(31 downto 0);
     end record;
     constant reg_type_init : reg_type :=
-        (e => Execute1ToWritebackInit, lr_update => '0',
+        (e => Execute1ToWritebackInit, busy => '0', lr_update => '0', terminate => '0',
          mul_in_progress => '0', div_in_progress => '0', cntz_in_progress => '0',
          slow_op_insn => OP_ILLEGAL, slow_op_rc => '0', slow_op_oe => '0', slow_op_xerc => xerc_init,
          next_lr => (others => '0'), ldst_nia => (others => '0'), others => (others => '0'));
@@ -71,6 +73,7 @@ architecture behaviour of execute1 is
 
     signal a_in, b_in, c_in : std_ulogic_vector(63 downto 0);
 
+    signal valid_in : std_ulogic;
     signal ctrl: ctrl_t := (irq_state => WRITE_SRR0, others => (others => '0'));
     signal ctrl_tmp: ctrl_t := (irq_state => WRITE_SRR0, others => (others => '0'));
     signal right_shift, rot_clear_left, rot_clear_right: std_ulogic;
@@ -241,6 +244,11 @@ begin
     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;
 
+    busy_out <= l_in.busy or r.busy;
+    valid_in <= e_in.valid and not busy_out;
+
+    terminate_out <= r.terminate;
+
     execute1_0: process(clk)
     begin
 	if rising_edge(clk) then
@@ -251,7 +259,7 @@ begin
             else
                 r <= rin;
                 ctrl <= ctrl_tmp;
-                assert not (r.lr_update = '1' and e_in.valid = '1')
+                assert not (r.lr_update = '1' and valid_in = '1')
                     report "LR update collision with valid in EX1"
                     severity failure;
                 if r.lr_update = '1' then
@@ -423,9 +431,9 @@ begin
 	    end if;
 	end if;
 
-	terminate_out <= '0';
+	v.terminate := '0';
 	icache_inval <= '0';
-	stall_out <= '0';
+	v.busy := '0';
 	f_out <= Execute1ToFetch1TypeInit;
         -- send MSR[IR] and ~MSR[PR] up to fetch1
         f_out.virt_mode <= ctrl.msr(MSR_IR);
@@ -463,10 +471,10 @@ begin
             f_out.virt_mode <= '0';
             f_out.priv_mode <= '1';
 	    f_out.redirect_nia <= ctrl.irq_nia;
-	    v.e.valid := e_in.valid;
+            v.e.valid := '1';
 	    report "Writing SRR1: " & to_hstring(ctrl.srr1);
 
-	elsif irq_valid = '1' and e_in.valid = '1' then
+	elsif irq_valid = '1' and valid_in = '1' then
 	    -- we need two cycles to write srr0 and 1
 	    -- will need more when we have to write HEIR
             -- Don't deliver the interrupt until we have a valid instruction
@@ -474,7 +482,7 @@ begin
 	    exception := '1';
 	    ctrl_tmp.srr1 <= msr_copy(ctrl.msr);
 
-        elsif e_in.valid = '1' and ctrl.msr(MSR_PR) = '1' and
+        elsif valid_in = '1' and ctrl.msr(MSR_PR) = '1' and
             instr_is_privileged(e_in.insn_type, e_in.insn) then
             -- generate a program interrupt
             exception := '1';
@@ -484,7 +492,7 @@ begin
             ctrl_tmp.srr1(63 - 45) <= '1';
             report "privileged instruction";
             
-	elsif e_in.valid = '1' and e_in.unit = ALU then
+	elsif valid_in = '1' and e_in.unit = ALU then
 
 	    report "execute nia " & to_hstring(e_in.nia);
 
@@ -519,7 +527,7 @@ begin
                 -- check bits 1-10 of the instruction to make sure it's attn
                 -- if not then it is illegal
                 if e_in.insn(10 downto 1) = "0100000000" then
-                    terminate_out <= '1';
+                    v.terminate := '1';
                     report "ATTN";
                 else
                     illegal := '1';
@@ -674,7 +682,7 @@ begin
             when OP_CNTZ =>
                 v.e.valid := '0';
                 v.cntz_in_progress := '1';
-                stall_out <= '1';
+                v.busy := '1';
             when OP_EXTS =>
                 -- note data_len is a 1-hot encoding
 		negative := (e_in.data_len(0) and c_in(7)) or
@@ -876,21 +884,21 @@ begin
 	    when OP_MUL_L64 | OP_MUL_H64 | OP_MUL_H32 =>
 		v.e.valid := '0';
 		v.mul_in_progress := '1';
-		stall_out <= '1';
+		v.busy := '1';
 		x_to_multiply.valid <= '1';
 
 	    when OP_DIV | OP_DIVE | OP_MOD =>
 		v.e.valid := '0';
 		v.div_in_progress := '1';
-		stall_out <= '1';
+		v.busy := '1';
 		x_to_divider.valid <= '1';
 
             when others =>
-		terminate_out <= '1';
+		v.terminate := '1';
 		report "illegal";
 	    end case;
 
-	    v.e.rc := e_in.rc and e_in.valid;
+	    v.e.rc := e_in.rc and valid_in;
 
 	    -- Update LR on the next cycle after a branch link
 	    --
@@ -908,10 +916,10 @@ begin
 		v.next_lr := next_nia;
 		v.e.valid := '0';
 		report "Delayed LR update to " & to_hstring(next_nia);
-		stall_out <= '1';
+		v.busy := '1';
 	    end if;
 
-        elsif e_in.valid = '1' then
+        elsif valid_in = '1' then
             -- instruction for other units, i.e. LDST
             v.ldst_nia := e_in.nia;
             v.e.valid := '0';
@@ -967,7 +975,7 @@ begin
 		end if;
 		v.e.valid := '1';
 	    else
-		stall_out <= '1';
+		v.busy := '1';
 		v.mul_in_progress := r.mul_in_progress;
 		v.div_in_progress := r.div_in_progress;
 	    end if;
@@ -988,7 +996,8 @@ begin
                 v.e.exc_write_data := next_nia;
             end if;
             ctrl_tmp.irq_state <= WRITE_SRR1;
-            v.e.valid := '1';
+            v.busy := '1';
+            v.e.valid := '0';
 	end if;
 
 	v.e.write_data := result;
@@ -1020,7 +1029,6 @@ begin
             v.e.exc_write_data := r.ldst_nia;
             report "ldst exception writing srr0=" & to_hstring(r.ldst_nia);
             ctrl_tmp.irq_state <= WRITE_SRR1;
-            v.e.valid := '1';   -- complete the original load or store
         end if;
 
         -- Outputs to loadstore1 (async)
@@ -1072,7 +1080,7 @@ begin
                         r.e.write_enable &
                         r.e.valid &
                         f_out.redirect &
-                        stall_out &
+                        r.busy &
                         flush_out;
         end if;
     end process;

gpr_hazard.vhdl

@@ -4,11 +4,15 @@ use ieee.numeric_std.all;
 
 entity gpr_hazard is
     generic (
-        PIPELINE_DEPTH : natural := 2
+        PIPELINE_DEPTH : natural := 1
         );
     port(
         clk                : in std_ulogic;
-	stall_in           : in std_ulogic;
+        busy_in            : in std_ulogic;
+        deferred           : in std_ulogic;
+        complete_in        : in std_ulogic;
+        flush_in           : in std_ulogic;
+        issuing            : in std_ulogic;
 
         gpr_write_valid_in : in std_ulogic;
         gpr_write_in       : in std_ulogic_vector(5 downto 0);
@@ -16,6 +20,9 @@ entity gpr_hazard is
         gpr_read_valid_in  : in std_ulogic;
         gpr_read_in        : in std_ulogic_vector(5 downto 0);
 
+        ugpr_write_valid   : in std_ulogic;
+        ugpr_write_reg     : in std_ulogic_vector(5 downto 0);
+
         stall_out          : out std_ulogic;
         use_bypass         : out std_ulogic
         );
@@ -25,10 +32,13 @@ architecture behaviour of gpr_hazard is
         valid  : std_ulogic;
         bypass : std_ulogic;
         gpr    : std_ulogic_vector(5 downto 0);
+        ugpr_valid : std_ulogic;
+        ugpr   : std_ulogic_vector(5 downto 0);
     end record;
-    constant pipeline_entry_init : pipeline_entry_type := (valid => '0', bypass => '0', gpr => (others => '0'));
+    constant pipeline_entry_init : pipeline_entry_type := (valid => '0', bypass => '0', gpr => (others => '0'),
+                                                           ugpr_valid => '0', ugpr => (others => '0'));
 
-    type pipeline_t is array(0 to PIPELINE_DEPTH-1) of pipeline_entry_type;
+    type pipeline_t is array(0 to PIPELINE_DEPTH) of pipeline_entry_type;
     constant pipeline_t_init : pipeline_t := (others => pipeline_entry_init);
 
     signal r, rin : pipeline_t := pipeline_t_init;
@@ -45,50 +55,46 @@ begin
     begin
         v := r;
 
+        if complete_in = '1' then
+            v(PIPELINE_DEPTH).valid := '0';
+            v(PIPELINE_DEPTH).ugpr_valid := '0';
+        end if;
+
         stall_out <= '0';
         use_bypass <= '0';
         if gpr_read_valid_in = '1' then
-            if r(0).valid = '1' and r(0).gpr = gpr_read_in then
-                if r(0).bypass = '1' and stall_in = '0' then
-                    use_bypass <= '1';
-                else
-                    stall_out <= '1';
-                end if;
-            end if;
-            loop_0: for i in 1 to PIPELINE_DEPTH-1 loop
-                if r(i).valid = '1' and r(i).gpr = gpr_read_in then
+            loop_0: for i in 0 to PIPELINE_DEPTH loop
+                if v(i).valid = '1' and r(i).gpr = gpr_read_in then
                     if r(i).bypass = '1' then
                         use_bypass <= '1';
                     else
                         stall_out <= '1';
                     end if;
                 end if;
+                if v(i).ugpr_valid = '1' and r(i).ugpr = gpr_read_in then
+                    stall_out <= '1';
+                end if;
             end loop;
         end if;
 
-        if stall_in = '0' then
+        -- XXX assumes PIPELINE_DEPTH = 1
+        if busy_in = '0' then
+            v(1) := v(0);
+            v(0).valid := '0';
+            v(0).ugpr_valid := '0';
+        end if;
+        if deferred = '0' and issuing = '1' then
             v(0).valid  := gpr_write_valid_in;
             v(0).bypass := bypass_avail;
             v(0).gpr    := gpr_write_in;
-            loop_1: for i in 1 to PIPELINE_DEPTH-1 loop
-                -- propagate to next slot
-                v(i).valid  := r(i-1).valid;
-                v(i).bypass := r(i-1).bypass;
-                v(i).gpr    := r(i-1).gpr;
-            end loop;
-
-        else
-            -- stage 0 stalled, so stage 1 becomes empty
-            loop_1b: for i in 1 to PIPELINE_DEPTH-1 loop
-                -- propagate to next slot
-                if i = 1 then
-                    v(i).valid := '0';
-                else
-                    v(i).valid  := r(i-1).valid;
-                    v(i).bypass := r(i-1).bypass;
-                    v(i).gpr    := r(i-1).gpr;
-                end if;
-            end loop;
+            v(0).ugpr_valid := ugpr_write_valid;
+            v(0).ugpr   := ugpr_write_reg;
+        end if;
+        if flush_in = '1' then
+            v(0).valid := '0';
+            v(0).ugpr_valid := '0';
+            v(1).valid := '0';
+            v(1).ugpr_valid := '0';
         end if;
 
         -- update registers

loadstore1.vhdl

@@ -25,7 +25,6 @@ entity loadstore1 is
         m_in  : in MmuToLoadstore1Type;
 
         dc_stall  : in std_ulogic;
-        stall_out : out std_ulogic;
 
         log_out : out std_ulogic_vector(9 downto 0)
         );
@@ -47,6 +46,7 @@ architecture behave of loadstore1 is
                      );
 
     type reg_stage_t is record
+        busy         : std_ulogic;
         -- latch most of the input request
         load         : std_ulogic;
         tlbie        : std_ulogic;
@@ -123,6 +123,7 @@ begin
         if rising_edge(clk) then
             if rst = '1' then
                 r.state <= IDLE;
+                r.busy <= '0';
             else
                 r <= rin;
             end if;
@@ -499,6 +500,7 @@ begin
         l_out.store_done <= d_in.store_done;
 
         -- update exception info back to execute1
+        e_out.busy <= r.busy;
         e_out.exception <= exception;
         e_out.instr_fault <= r.instr_fault;
         e_out.invalid <= m_in.invalid;
@@ -513,7 +515,7 @@ begin
             end if;
         end if;
 
-        stall_out <= stall;
+        v.busy := stall;
 
         -- Update registers
         rin <= v;
@@ -523,7 +525,7 @@ begin
     ls1_log: process(clk)
     begin
         if rising_edge(clk) then
-            log_data <= stall_out &
+            log_data <= r.busy &
                         e_out.exception &
                         l_out.valid &
                         m_out.valid &

writeback.vhdl

@@ -22,27 +22,33 @@ end entity writeback;
 
 architecture behaviour of writeback is
 begin
-    writeback_1: process(all)
+    writeback_0: process(clk)
         variable x : std_ulogic_vector(0 downto 0);
         variable y : std_ulogic_vector(0 downto 0);
         variable w : std_ulogic_vector(0 downto 0);
+    begin
+        if rising_edge(clk) then
+            -- Do consistency checks only on the clock edge
+            x(0) := e_in.valid;
+            y(0) := l_in.valid;
+            assert (to_integer(unsigned(x)) + to_integer(unsigned(y))) <= 1 severity failure;
+
+            x(0) := e_in.write_enable or e_in.exc_write_enable;
+            y(0) := l_in.write_enable;
+            assert (to_integer(unsigned(x)) + to_integer(unsigned(y))) <= 1 severity failure;
+
+            w(0) := e_in.write_cr_enable;
+            x(0) := (e_in.write_enable and e_in.rc);
+            assert (to_integer(unsigned(w)) + to_integer(unsigned(x))) <= 1 severity failure;
+        end if;
+    end process;
+
+    writeback_1: process(all)
 	variable cf: std_ulogic_vector(3 downto 0);
         variable zero : std_ulogic;
         variable sign : std_ulogic;
 	variable scf  : std_ulogic_vector(3 downto 0);
     begin
-        x(0) := e_in.valid;
-        y(0) := l_in.valid;
-        assert (to_integer(unsigned(x)) + to_integer(unsigned(y))) <= 1 severity failure;
-
-        x(0) := e_in.write_enable or e_in.exc_write_enable;
-        y(0) := l_in.write_enable;
-        assert (to_integer(unsigned(x)) + to_integer(unsigned(y))) <= 1 severity failure;
-
-        w(0) := e_in.write_cr_enable;
-        x(0) := (e_in.write_enable and e_in.rc);
-        assert (to_integer(unsigned(w)) + to_integer(unsigned(x))) <= 1 severity failure;
-
         w_out <= WritebackToRegisterFileInit;
         c_out <= WritebackToCrFileInit;