core: Implement a simple branch predictor

This implements a simple branch predictor in the decode1 stage.  If it
sees that the instruction is b or bc and the branch is predicted to be
taken, it sends a flush and redirect upstream (to icache and fetch1)
to redirect fetching to the branch target.  The prediction is sent
downstream with the branch instruction, and execute1 now only sends
a flush/redirect upstream if the prediction was wrong.  Unconditional
branches are always predicted to be taken, and conditional branches
are predicted to be taken if and only if the offset is negative.
Branches that take the branch address from a register (bclr, bcctr)
are predicted not taken, as we don't have any way to predict the
branch address.

Since we can now have a mflr being executed immediately after a bl
or bcl, we now track the update to LR in the hazard tracker, using
the second write register field that is used to track RA updates for
update-form loads and stores.

For those branches that update LR but don't write any other result
(i.e. that don't decrementer CTR), we now write back LR in the same
cycle as the instruction rather than taking a second cycle for the
LR writeback.

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

common.vhdl

@@ -113,8 +113,16 @@ package common is
 	ispr1: gspr_index_t; -- (G)SPR used for branch condition (CTR) or mfspr
 	ispr2: gspr_index_t; -- (G)SPR used for branch target (CTR, LR, TAR)
 	decode: decode_rom_t;
+        br_pred: std_ulogic; -- Branch was predicted to be taken
+    end record;
+    constant Decode1ToDecode2Init : Decode1ToDecode2Type :=
+        (valid => '0', stop_mark => '0', nia => (others => '0'), insn => (others => '0'),
+         ispr1 => (others => '0'), ispr2 => (others => '0'), decode => decode_rom_init, br_pred => '0');
+
+    type Decode1ToFetch1Type is record
+        redirect     : std_ulogic;
+        redirect_nia : std_ulogic_vector(63 downto 0);
     end record;
-    constant Decode1ToDecode2Init : Decode1ToDecode2Type := (valid => '0', stop_mark => '0', nia => (others => '0'), insn => (others => '0'), ispr1 => (others => '0'), ispr2 => (others => '0'), decode => decode_rom_init);
 
     type Decode2ToExecute1Type is record
 	valid: std_ulogic;
@@ -149,12 +157,13 @@ package common is
 	sign_extend : std_ulogic;			-- do we need to sign extend?
 	update : std_ulogic;				-- is this an update instruction?
         reserve : std_ulogic;                           -- set for larx/stcx
+        br_pred : std_ulogic;
     end record;
     constant Decode2ToExecute1Init : Decode2ToExecute1Type :=
 	(valid => '0', unit => NONE, insn_type => OP_ILLEGAL, bypass_data1 => '0', bypass_data2 => '0', bypass_data3 => '0',
          lr => '0', rc => '0', oe => '0', invert_a => '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',
+	 is_32bit => '0', is_signed => '0', xerc => xerc_init, reserve => '0', br_pred => '0',
          byte_reverse => '0', sign_extend => '0', update => '0', nia => (others => '0'), read_data1 => (others => '0'), read_data2 => (others => '0'), read_data3 => (others => '0'), cr => (others => '0'), insn => (others => '0'), data_len => (others => '0'), others => (others => '0'));
 
     type Execute1ToMultiplyType is record

core.vhdl

@@ -48,6 +48,7 @@ architecture behave of core is
 
     -- decode signals
     signal decode1_to_decode2: Decode1ToDecode2Type;
+    signal decode1_to_fetch1: Decode1ToFetch1Type;
     signal decode2_to_execute1: Decode2ToExecute1Type;
 
     -- register file signals
@@ -90,6 +91,8 @@ architecture behave of core is
     signal dcache_stall_out: std_ulogic;
 
     signal flush: std_ulogic;
+    signal decode1_flush: std_ulogic;
+    signal fetch1_flush: std_ulogic;
 
     signal complete: std_ulogic;
     signal terminate: std_ulogic;
@@ -182,14 +185,16 @@ begin
             rst => rst_fetch1,
 	    alt_reset_in => alt_reset_d,
             stall_in => fetch1_stall_in,
-            flush_in => flush,
+            flush_in => fetch1_flush,
 	    stop_in => dbg_core_stop,
+            d_in => decode1_to_fetch1,
             e_in => execute1_to_fetch1,
             i_out => fetch1_to_icache,
             log_out => log_data(42 downto 0)
             );
 
     fetch1_stall_in <= icache_stall_out or decode1_busy;
+    fetch1_flush <= flush or decode1_flush;
 
     icache_0: entity work.icache
         generic map(
@@ -204,7 +209,7 @@ begin
             i_in => fetch1_to_icache,
             i_out => icache_to_decode1,
             m_in => mmu_to_icache,
-            flush_in => flush,
+            flush_in => fetch1_flush,
             inval_in => dbg_icache_rst or ex1_icache_inval,
             stall_in => icache_stall_in,
 	    stall_out => icache_stall_out,
@@ -221,9 +226,11 @@ begin
             rst => rst_dec1,
             stall_in => decode1_stall_in,
             flush_in => flush,
+            flush_out => decode1_flush,
             busy_out => decode1_busy,
             f_in => icache_to_decode1,
             d_out => decode1_to_decode2,
+            f_out => decode1_to_fetch1,
             log_out => log_data(109 downto 97)
             );
 

decode1.vhdl

@@ -8,16 +8,18 @@ use work.decode_types.all;
 
 entity decode1 is
     port (
-        clk      : in std_ulogic;
+        clk       : in std_ulogic;
-        rst      : in std_ulogic;
+        rst       : in std_ulogic;
-
+
-        stall_in : in std_ulogic;
+        stall_in  : in  std_ulogic;
-        flush_in : in std_ulogic;
+        flush_in  : in  std_ulogic;
-        busy_out : out std_ulogic;
+        busy_out  : out std_ulogic;
-
+        flush_out : out std_ulogic;
-        f_in     : in IcacheToDecode1Type;
+
-        d_out    : out Decode1ToDecode2Type;
+        f_in      : in IcacheToDecode1Type;
-        log_out  : out std_ulogic_vector(12 downto 0)
+        f_out     : out Decode1ToFetch1Type;
+        d_out     : out Decode1ToDecode2Type;
+        log_out   : out std_ulogic_vector(12 downto 0)
 	);
 end entity decode1;
 
@@ -385,11 +387,15 @@ begin
 
     decode1_1: process(all)
         variable v : Decode1ToDecode2Type;
+        variable f : Decode1ToFetch1Type;
         variable majorop : major_opcode_t;
         variable op_19_bits: std_ulogic_vector(2 downto 0);
         variable sprn : spr_num_t;
+        variable br_nia    : std_ulogic_vector(61 downto 0);
+        variable br_target : std_ulogic_vector(61 downto 0);
+        variable br_offset : signed(23 downto 0);
     begin
-        v := r;
+        v := Decode1ToDecode2Init;
 
         v.valid := f_in.valid;
         v.nia  := f_in.nia;
@@ -486,14 +492,36 @@ begin
 
         else
             v.decode := major_decode_rom_array(to_integer(majorop));
+        end if;
 
+        -- Branch predictor
+        -- Note bclr, bcctr and bctar are predicted not taken as we have no
+        -- count cache or link stack.
+        br_offset := (others => '0');
+        if majorop = 18 then
+            -- Unconditional branches are always taken
+            v.br_pred := '1';
+            br_offset := signed(f_in.insn(25 downto 2));
+        elsif majorop = 16 then
+            -- Predict backward branches as taken, forward as untaken
+            v.br_pred := f_in.insn(15);
+            br_offset := resize(signed(f_in.insn(15 downto 2)), 24);
+        end if;
+        br_nia := f_in.nia(63 downto 2);
+        if f_in.insn(1) = '1' then
+            br_nia := (others => '0');
         end if;
+        br_target := std_ulogic_vector(signed(br_nia) + br_offset);
+        f.redirect := v.br_pred and f_in.valid and not flush_in and not s.valid;
+        f.redirect_nia := br_target & "00";
 
         -- Update registers
         rin <= v;
 
         -- Update outputs
         d_out <= r;
+        f_out <= f;
+        flush_out <= f.redirect;
     end process;
 
     dec1_log : process(clk)

decode2.vhdl

@@ -358,6 +358,7 @@ begin
         v.e.sign_extend := d_in.decode.sign_extend;
         v.e.update := d_in.decode.update;
         v.e.reserve := d_in.decode.reserve;
+        v.e.br_pred := d_in.br_pred;
 
         -- issue control
         control_valid_in <= d_in.valid;
@@ -371,6 +372,11 @@ begin
         end if;
         update_gpr_write_valid <= d_in.decode.update;
         update_gpr_write_reg <= decoded_reg_a.reg;
+        if v.e.lr = '1' then
+            -- there are no instructions that have both update=1 and lr=1
+            update_gpr_write_valid <= '1';
+            update_gpr_write_reg <= fast_spr_num(SPR_LR);
+        end if;
 
         gpr_a_read_valid <= decoded_reg_a.reg_valid;
         gpr_a_read <= decoded_reg_a.reg;

execute1.vhdl

@@ -305,11 +305,17 @@ begin
         variable exception_nextpc : std_ulogic;
         variable trapval : std_ulogic_vector(4 downto 0);
         variable illegal : std_ulogic;
+        variable is_branch : std_ulogic;
+        variable taken_branch : std_ulogic;
+        variable abs_branch : std_ulogic;
     begin
 	result := (others => '0');
 	result_with_carry := (others => '0');
 	result_en := '0';
 	newcrf := (others => '0');
+        is_branch := '0';
+        taken_branch := '0';
+        abs_branch := '0';
 
 	v := r;
 	v.e := Execute1ToWritebackInit;
@@ -625,12 +631,9 @@ begin
 		result := logical_result;
 		result_en := '1';
 	    when OP_B =>
-		f_out.redirect <= '1';
+                is_branch := '1';
-		if (insn_aa(e_in.insn)) then
+                taken_branch := '1';
-		    f_out.redirect_nia <= b_in;
+                abs_branch := insn_aa(e_in.insn);
-		else
-		    f_out.redirect_nia <= std_ulogic_vector(signed(e_in.nia) + signed(b_in));
-		end if;
 	    when OP_BC =>
 		-- read_data1 is CTR
 		bo := insn_bo(e_in.insn);
@@ -640,14 +643,9 @@ begin
 		    result_en := '1';
 		    v.e.write_reg := fast_spr_num(SPR_CTR);
 		end if;
-		if ppc_bc_taken(bo, bi, e_in.cr, a_in) = 1 then
+                is_branch := '1';
-		    f_out.redirect <= '1';
+		taken_branch := ppc_bc_taken(bo, bi, e_in.cr, a_in);
-		    if (insn_aa(e_in.insn)) then
+                abs_branch := insn_aa(e_in.insn);
-			f_out.redirect_nia <= b_in;
-		    else
-			f_out.redirect_nia <= std_ulogic_vector(signed(e_in.nia) + signed(b_in));
-		    end if;
-		end if;
 	    when OP_BCREG =>
 		-- read_data1 is CTR
 		-- read_data2 is target register (CTR, LR or TAR)
@@ -658,7 +656,7 @@ begin
 		    result_en := '1';
 		    v.e.write_reg := fast_spr_num(SPR_CTR);
 		end if;
-		if ppc_bc_taken(bo, bi, e_in.cr, a_in) = 1 then
+		if ppc_bc_taken(bo, bi, e_in.cr, a_in) = '1' then
 		    f_out.redirect <= '1';
 		    f_out.redirect_nia <= b_in(63 downto 2) & "00";
 		end if;
@@ -903,23 +901,35 @@ begin
 
 	    v.e.rc := e_in.rc and valid_in;
 
+            -- Mispredicted branches cause a redirect
+            if is_branch = '1' and taken_branch /= e_in.br_pred then
+                f_out.redirect <= '1';
+                if taken_branch = '1' then
+                    if abs_branch = '1' then
+                        f_out.redirect_nia <= b_in;
+                    else
+                        f_out.redirect_nia <= std_ulogic_vector(signed(e_in.nia) + signed(b_in));
+                    end if;
+                else
+                    f_out.redirect_nia <= next_nia;
+                end if;
+            end if;
+
 	    -- Update LR on the next cycle after a branch link
-	    --
+	    -- If we're not writing back anything else, we can write back LR
-	    -- WARNING: The LR update isn't tracked by our hazard tracker. This
+            -- this cycle, otherwise we take an extra cycle.
-	    --          will work (well I hope) because it only happens on branches
-	    --          which will flush all decoded instructions. By the time
-	    --          fetch catches up, we'll have the new LR. This will
-	    --          *not* work properly however if we have a branch predictor,
-	    --          in which case the solution would probably be to keep a
-	    --          local cache of the updated LR in execute1 (flushed on
-	    --          exceptions) that is used instead of the value from
-	    --          decode when its content is valid.
 	    if e_in.lr = '1' then
-		v.lr_update := '1';
+                if result_en = '0' then
-		v.next_lr := next_nia;
+                    result_en := '1';
-		v.e.valid := '0';
+                    result := next_nia;
-		report "Delayed LR update to " & to_hstring(next_nia);
+                    v.e.write_reg := fast_spr_num(SPR_LR);
-		v.busy := '1';
+                else
+                    v.lr_update := '1';
+                    v.next_lr := next_nia;
+                    v.e.valid := '0';
+                    report "Delayed LR update to " & to_hstring(next_nia);
+                    v.busy := '1';
+                end if;
 	    end if;
 
         elsif valid_in = '1' then

fetch1.vhdl

@@ -23,6 +23,9 @@ entity fetch1 is
 	-- redirect from execution unit
 	e_in          : in Execute1ToFetch1Type;
 
+        -- redirect from decode1
+        d_in          : in Decode1ToFetch1Type;
+
 	-- Request to icache
 	i_out         : out Fetch1ToIcacheType;
 
@@ -49,7 +52,7 @@ begin
 		report "fetch1 rst:" & std_ulogic'image(rst) &
                     " IR:" & std_ulogic'image(e_in.virt_mode) &
                     " P:" & std_ulogic'image(e_in.priv_mode) &
-		    " R:" & std_ulogic'image(e_in.redirect) &
+		    " R:" & std_ulogic'image(e_in.redirect) & std_ulogic'image(d_in.redirect) &
 		    " S:" & std_ulogic'image(stall_in) &
 		    " T:" & std_ulogic'image(stop_in) &
 		    " nia:" & to_hstring(r_next.nia) &
@@ -83,6 +86,8 @@ begin
 	    v.nia := e_in.redirect_nia;
             v.virt_mode := e_in.virt_mode;
             v.priv_mode := e_in.priv_mode;
+        elsif d_in.redirect = '1' then
+            v.nia := d_in.redirect_nia;
 	elsif stall_in = '0' then
 
 	    -- For debug stop/step to work properly we need a little bit of

ppc_fx_insns.vhdl

@@ -93,7 +93,7 @@ package ppc_fx_insns is
 	function ppc_divd (ra, rb: std_ulogic_vector(63 downto 0)) return std_ulogic_vector;
 	function ppc_divwu (ra, rb: std_ulogic_vector(63 downto 0)) return std_ulogic_vector;
 
-	function ppc_bc_taken(bo, bi: std_ulogic_vector(4 downto 0); cr: std_ulogic_vector(31 downto 0); ctr: std_ulogic_vector(63 downto 0)) return integer;
+	function ppc_bc_taken(bo, bi: std_ulogic_vector(4 downto 0); cr: std_ulogic_vector(31 downto 0); ctr: std_ulogic_vector(63 downto 0)) return std_ulogic;
 end package ppc_fx_insns;
 
 package body ppc_fx_insns is
@@ -785,13 +785,12 @@ package body ppc_fx_insns is
 		return std_ulogic_vector(resize(tmp, ra'length));
 	end;
 
-	function ppc_bc_taken(bo, bi: std_ulogic_vector(4 downto 0); cr: std_ulogic_vector(31 downto 0); ctr: std_ulogic_vector(63 downto 0)) return integer is
+	function ppc_bc_taken(bo, bi: std_ulogic_vector(4 downto 0); cr: std_ulogic_vector(31 downto 0); ctr: std_ulogic_vector(63 downto 0)) return std_ulogic is
 		variable crfield: integer;
 		variable crbit_match: std_ulogic;
 		variable ctr_not_zero: std_ulogic;
 		variable ctr_ok: std_ulogic;
 		variable cond_ok: std_ulogic;
-		variable ret: integer;
 	begin
 		crfield := to_integer(unsigned(bi));
 		-- BE bit numbering
@@ -800,12 +799,7 @@ package body ppc_fx_insns is
 		ctr_not_zero := '1' when ctr /= x"0000000000000001" else '0';
 		ctr_ok := bo(4-2) or (ctr_not_zero xor bo(4-3));
 		cond_ok := bo(4-0) or crbit_match;
-		if ctr_ok = '1' and cond_ok = '1' then
+                return ctr_ok and cond_ok;
-			ret := 1;
-		else
-			ret := 0;
-		end if;
-		return ret;
 	end;
 
 end package body ppc_fx_insns;