Merge pull request #122 from paulusmack/benh-sprs

Benh sprs

Makefile

@@ -14,7 +14,7 @@ all: $(all)
 	$(GHDL) -a $(GHDLFLAGS) $<
 
 common.o: decode_types.o
-control.o: gpr_hazard.o cr_hazard.o
+control.o: gpr_hazard.o cr_hazard.o common.o
 sim_jtag.o: sim_jtag_socket.o
 core_tb.o: common.o wishbone_types.o core.o soc.o sim_jtag.o
 core.o: common.o wishbone_types.o fetch1.o fetch2.o icache.o decode1.o decode2.o register_file.o cr_file.o execute1.o loadstore1.o dcache.o multiply.o writeback.o core_debug.o divider.o

common.vhdl

@@ -12,15 +12,59 @@ package common is
 
     function decode_spr_num(insn: std_ulogic_vector(31 downto 0)) return spr_num_t;
 
-    constant SPR_LR    : spr_num_t := 8;
+    constant SPR_XER    : spr_num_t := 1;
-    constant SPR_CTR   : spr_num_t := 9;
+    constant SPR_LR     : spr_num_t := 8;
-    constant SPR_TB    : spr_num_t := 268;
+    constant SPR_CTR    : spr_num_t := 9;
+    constant SPR_TB     : spr_num_t := 268;
+    constant SPR_SRR0   : spr_num_t := 26;
+    constant SPR_SRR1   : spr_num_t := 27;
+    constant SPR_HSRR0  : spr_num_t := 314;
+    constant SPR_HSRR1  : spr_num_t := 315;
+    constant SPR_SPRG0  : spr_num_t := 272;
+    constant SPR_SPRG1  : spr_num_t := 273;
+    constant SPR_SPRG2  : spr_num_t := 274;
+    constant SPR_SPRG3  : spr_num_t := 275;
+    constant SPR_SPRG3U : spr_num_t := 259;
+    constant SPR_HSPRG0 : spr_num_t := 304;
+    constant SPR_HSPRG1 : spr_num_t := 305;
 
+    -- GPR indices in the register file (GPR only)
+    subtype gpr_index_t is std_ulogic_vector(4 downto 0);
+
+    -- Extended GPR indice (can hold an SPR)
+    subtype gspr_index_t is std_ulogic_vector(5 downto 0);
+
+    -- Some SPRs are stored in the register file, they use the magic
+    -- GPR numbers above 31.
+    --
+    -- The function fast_spr_num() returns the corresponding fast
+    -- pseudo-GPR number for a given SPR number. The result MSB
+    -- indicates if this is indeed a fast SPR. If clear, then
+    -- the SPR is not stored in the GPR file.
+    --
+    function fast_spr_num(spr: spr_num_t) return gspr_index_t;
+
+    -- Indices conversion functions
+    function gspr_to_gpr(i: gspr_index_t) return gpr_index_t;
+    function gpr_to_gspr(i: gpr_index_t) return gspr_index_t;
+    function gpr_or_spr_to_gspr(g: gpr_index_t; s: gspr_index_t) return gspr_index_t;
+    function is_fast_spr(s: gspr_index_t) return std_ulogic;
+
+    -- The XER is split: the common bits (CA, OV, SO, OV32 and CA32) are
+    -- in the CR file as a kind of CR extension (with a separate write
+    -- control). The rest is stored as a fast SPR.
+    type xer_common_t is record
+	ca : std_ulogic;
+	ca32 : std_ulogic;
+	ov : std_ulogic;
+	ov32 : std_ulogic;
+	so : std_ulogic;
+    end record;
+    constant xerc_init : xer_common_t := (others => '0');
+
+    -- This needs to die...
     type ctrl_t is record
-	lr: std_ulogic_vector(63 downto 0);
-	ctr: std_ulogic_vector(63 downto 0);
 	tb: std_ulogic_vector(63 downto 0);
-	carry: std_ulogic;
     end record;
 
     type Fetch1ToIcacheType is record
@@ -49,6 +93,8 @@ package common is
 	stop_mark : std_ulogic;
 	nia: std_ulogic_vector(63 downto 0);
 	insn: std_ulogic_vector(31 downto 0);
+	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;
     end record;
     constant Decode1ToDecode2Init : Decode1ToDecode2Type := (valid => '0', stop_mark => '0', decode => decode_rom_init, others => (others => '0'));
@@ -57,15 +103,17 @@ package common is
 	valid: std_ulogic;
 	insn_type: insn_type_t;
 	nia: std_ulogic_vector(63 downto 0);
-	write_reg: std_ulogic_vector(4 downto 0);
+	write_reg: gspr_index_t;
-	read_reg1: std_ulogic_vector(4 downto 0);
+	read_reg1: gspr_index_t;
-	read_reg2: std_ulogic_vector(4 downto 0);
+	read_reg2: gspr_index_t;
 	read_data1: std_ulogic_vector(63 downto 0);
 	read_data2: std_ulogic_vector(63 downto 0);
 	read_data3: std_ulogic_vector(63 downto 0);
 	cr: std_ulogic_vector(31 downto 0);
+	xerc: xer_common_t;
 	lr: std_ulogic;
 	rc: std_ulogic;
+	oe: std_ulogic;
 	invert_a: std_ulogic;
 	invert_out: std_ulogic;
 	input_carry: carry_in_t;
@@ -78,23 +126,28 @@ package common is
 	data_len: std_ulogic_vector(3 downto 0);
     end record;
     constant Decode2ToExecute1Init : Decode2ToExecute1Type :=
-	(valid => '0', insn_type => OP_ILLEGAL, lr => '0', rc => '0', invert_a => '0',
+	(valid => '0', insn_type => OP_ILLEGAL, 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', others => (others => '0'));
+	 is_32bit => '0', is_signed => '0', xerc => xerc_init, others => (others => '0'));
 
     type Decode2ToMultiplyType is record
 	valid: std_ulogic;
 	insn_type: insn_type_t;
-	write_reg: std_ulogic_vector(4 downto 0);
+	write_reg: gpr_index_t;
 	data1: std_ulogic_vector(64 downto 0);
 	data2: std_ulogic_vector(64 downto 0);
 	rc: std_ulogic;
+	oe: std_ulogic;
+	is_32bit: std_ulogic;
+	xerc: xer_common_t;
     end record;
-    constant Decode2ToMultiplyInit : Decode2ToMultiplyType := (valid => '0', insn_type => OP_ILLEGAL, rc => '0', others => (others => '0'));
+    constant Decode2ToMultiplyInit : Decode2ToMultiplyType := (valid => '0', insn_type => OP_ILLEGAL, rc => '0',
+							       oe => '0', is_32bit => '0', xerc => xerc_init,
+							       others => (others => '0'));
 
     type Decode2ToDividerType is record
 	valid: std_ulogic;
-	write_reg: std_ulogic_vector(4 downto 0);
+	write_reg: gpr_index_t;
 	dividend: std_ulogic_vector(63 downto 0);
 	divisor: std_ulogic_vector(63 downto 0);
 	is_signed: std_ulogic;
@@ -102,16 +155,21 @@ package common is
 	is_extended: std_ulogic;
 	is_modulus: std_ulogic;
 	rc: std_ulogic;
+	oe: std_ulogic;
+	xerc: xer_common_t;
     end record;
-    constant Decode2ToDividerInit: Decode2ToDividerType := (valid => '0', is_signed => '0', is_32bit => '0', is_extended => '0', is_modulus => '0', rc => '0', others => (others => '0'));
+    constant Decode2ToDividerInit: Decode2ToDividerType := (valid => '0', is_signed => '0', is_32bit => '0',
+							    is_extended => '0', is_modulus => '0',
+							    rc => '0', oe => '0', xerc => xerc_init,
+							    others => (others => '0'));
 
     type Decode2ToRegisterFileType is record
 	read1_enable : std_ulogic;
-	read1_reg : std_ulogic_vector(4 downto 0);
+	read1_reg : gspr_index_t;
 	read2_enable : std_ulogic;
-	read2_reg : std_ulogic_vector(4 downto 0);
+	read2_reg : gspr_index_t;
 	read3_enable : std_ulogic;
-	read3_reg : std_ulogic_vector(4 downto 0);
+	read3_reg : gpr_index_t;
     end record;
 
     type RegisterFileToDecode2Type is record
@@ -126,6 +184,7 @@ package common is
 
     type CrFileToDecode2Type is record
 	read_cr_data   : std_ulogic_vector(31 downto 0);
+	read_xerc_data : xer_common_t;
     end record;
 
     type Execute1ToFetch1Type is record
@@ -140,14 +199,17 @@ package common is
 	addr1 : std_ulogic_vector(63 downto 0);
 	addr2 : std_ulogic_vector(63 downto 0);
 	data : std_ulogic_vector(63 downto 0);		-- data to write, unused for read
-	write_reg : std_ulogic_vector(4 downto 0);	-- read data goes to this register
+	write_reg : gpr_index_t;
 	length : std_ulogic_vector(3 downto 0);
 	byte_reverse : std_ulogic;
 	sign_extend : std_ulogic;			-- do we need to sign extend?
 	update : std_ulogic;				-- is this an update instruction?
-	update_reg : std_ulogic_vector(4 downto 0);	-- if so, the register to update
+	update_reg : gpr_index_t;                      	-- if so, the register to update
+	xerc : xer_common_t;
     end record;
-    constant Decode2ToLoadstore1Init : Decode2ToLoadstore1Type := (valid => '0', load => '0', byte_reverse => '0', sign_extend => '0', update => '0', others => (others => '0'));
+    constant Decode2ToLoadstore1Init : Decode2ToLoadstore1Type := (valid => '0', load => '0', byte_reverse => '0',
+								   sign_extend => '0', update => '0', xerc => xerc_init,
+								   others => (others => '0'));
 
     type Loadstore1ToDcacheType is record
 	valid : std_ulogic;
@@ -155,63 +217,82 @@ package common is
 	nc : std_ulogic;
 	addr : std_ulogic_vector(63 downto 0);
 	data : std_ulogic_vector(63 downto 0);
-	write_reg : std_ulogic_vector(4 downto 0);
+	write_reg : gpr_index_t;
 	length : std_ulogic_vector(3 downto 0);
 	byte_reverse : std_ulogic;
 	sign_extend : std_ulogic;
 	update : std_ulogic;
-	update_reg : std_ulogic_vector(4 downto 0);
+	update_reg : gpr_index_t;
+	xerc : xer_common_t;
     end record;
 
     type DcacheToWritebackType is record
 	valid : std_ulogic;
 	write_enable: std_ulogic;
-	write_reg : std_ulogic_vector(4 downto 0);
+	write_reg : gpr_index_t;
 	write_data : std_ulogic_vector(63 downto 0);
 	write_len : std_ulogic_vector(3 downto 0);
 	write_shift : std_ulogic_vector(2 downto 0);
 	sign_extend : std_ulogic;
 	byte_reverse : std_ulogic;
 	second_word : std_ulogic;
+	xerc : xer_common_t;
     end record;
-    constant DcacheToWritebackInit : DcacheToWritebackType := (valid => '0', write_enable => '0', sign_extend => '0', byte_reverse => '0', second_word => '0', others => (others => '0'));
+    constant DcacheToWritebackInit : DcacheToWritebackType := (valid => '0', write_enable => '0', sign_extend => '0',
+							       byte_reverse => '0', second_word => '0', xerc => xerc_init,
+							       others => (others => '0'));
 
     type Execute1ToWritebackType is record
 	valid: std_ulogic;
 	rc : std_ulogic;
 	write_enable : std_ulogic;
-	write_reg: std_ulogic_vector(4 downto 0);
+	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', others => (others => '0'));
+    constant Execute1ToWritebackInit : Execute1ToWritebackType := (valid => '0', rc => '0', write_enable => '0',
+								   write_cr_enable => '0', sign_extend => '0',
+								   write_xerc_enable => '0', xerc => xerc_init,
+								   others => (others => '0'));
 
     type MultiplyToWritebackType is record
 	valid: std_ulogic;
 
 	write_reg_enable : std_ulogic;
-	write_reg_nr: std_ulogic_vector(4 downto 0);
+	write_reg_nr: gpr_index_t;
 	write_reg_data: std_ulogic_vector(63 downto 0);
+	write_xerc_enable : std_ulogic;
+	xerc : xer_common_t;
 	rc: std_ulogic;
     end record;
-    constant MultiplyToWritebackInit : MultiplyToWritebackType := (valid => '0', write_reg_enable => '0', rc => '0', others => (others => '0'));
+    constant MultiplyToWritebackInit : MultiplyToWritebackType := (valid => '0', write_reg_enable => '0',
+								   rc => '0', write_xerc_enable => '0',
+								   xerc => xerc_init,
+								   others => (others => '0'));
 
     type DividerToWritebackType is record
 	valid: std_ulogic;
 
 	write_reg_enable : std_ulogic;
-	write_reg_nr: std_ulogic_vector(4 downto 0);
+	write_reg_nr: gpr_index_t;
 	write_reg_data: std_ulogic_vector(63 downto 0);
+	write_xerc_enable : std_ulogic;
+	xerc : xer_common_t;
 	rc: std_ulogic;
     end record;
-    constant DividerToWritebackInit : DividerToWritebackType := (valid => '0', write_reg_enable => '0', rc => '0', others => (others => '0'));
+    constant DividerToWritebackInit : DividerToWritebackType := (valid => '0', write_reg_enable => '0',
+								 rc => '0', write_xerc_enable => '0',
+								 xerc => xerc_init,
+								 others => (others => '0'));
 
     type WritebackToRegisterFileType is record
-	write_reg : std_ulogic_vector(4 downto 0);
+	write_reg : gspr_index_t;
 	write_data : std_ulogic_vector(63 downto 0);
 	write_enable : std_ulogic;
     end record;
@@ -221,9 +302,12 @@ package common is
 	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;
+	write_xerc_data : xer_common_t;
     end record;
-    constant WritebackToCrFileInit : WritebackToCrFileType := (write_cr_enable => '0', others => (others => '0'));
+    constant WritebackToCrFileInit : WritebackToCrFileType := (write_cr_enable => '0', write_xerc_enable => '0',
-
+							       write_xerc_data => xerc_init,
+							       others => (others => '0'));
 end common;
 
 package body common is
@@ -231,4 +315,63 @@ package body common is
     begin
 	return to_integer(unsigned(insn(15 downto 11) & insn(20 downto 16)));
     end;
+    function fast_spr_num(spr: spr_num_t) return gspr_index_t is
+       variable n : integer range 0 to 31;
+    begin
+       case spr is
+       when SPR_LR =>
+           n := 0;
+       when SPR_CTR =>
+           n:= 1;
+       when SPR_SRR0 =>
+           n := 2;
+       when SPR_SRR1 =>
+           n := 3;
+       when SPR_HSRR0 =>
+           n := 4;
+       when SPR_HSRR1 =>
+           n := 5;
+       when SPR_SPRG0 =>
+           n := 6;
+       when SPR_SPRG1 =>
+           n := 7;
+       when SPR_SPRG2 =>
+           n := 8;
+       when SPR_SPRG3 | SPR_SPRG3U =>
+           n := 9;
+       when SPR_HSPRG0 =>
+           n := 10;
+       when SPR_HSPRG1 =>
+           n := 11;
+       when SPR_XER =>
+           n := 12;
+       when others =>
+           return "000000";
+       end case;
+       return "1" & std_ulogic_vector(to_unsigned(n, 5));
+    end;
+
+    function gspr_to_gpr(i: gspr_index_t) return gpr_index_t is
+    begin
+	return i(4 downto 0);
+    end;
+
+    function gpr_to_gspr(i: gpr_index_t) return gspr_index_t is
+    begin
+	return "0" & i;
+    end;
+
+    function gpr_or_spr_to_gspr(g: gpr_index_t; s: gspr_index_t) return gspr_index_t is
+    begin
+	if s(5) = '1' then
+	    return s;
+	else
+	    return gpr_to_gspr(g);
+	end if;
+    end;
+
+    function is_fast_spr(s: gspr_index_t) return std_ulogic is
+    begin
+	return s(5);
+    end;
 end common;

control.vhdl

@@ -1,6 +1,9 @@
 library ieee;
 use ieee.std_logic_1164.all;
 
+library work;
+use work.common.all;
+
 entity control is
     generic (
         PIPELINE_DEPTH : natural := 2
@@ -12,20 +15,21 @@ entity control is
         complete_in         : in std_ulogic;
         valid_in            : in std_ulogic;
         flush_in            : in std_ulogic;
+	stall_in            : in std_ulogic;
         sgl_pipe_in         : in std_ulogic;
         stop_mark_in        : in std_ulogic;
 
         gpr_write_valid_in  : in std_ulogic;
-        gpr_write_in        : in std_ulogic_vector(4 downto 0);
+        gpr_write_in        : in gspr_index_t;
 
         gpr_a_read_valid_in : in std_ulogic;
-        gpr_a_read_in       : in std_ulogic_vector(4 downto 0);
+        gpr_a_read_in       : in gspr_index_t;
 
         gpr_b_read_valid_in : in std_ulogic;
-        gpr_b_read_in       : in std_ulogic_vector(4 downto 0);
+        gpr_b_read_in       : in gspr_index_t;
 
         gpr_c_read_valid_in : in std_ulogic;
-        gpr_c_read_in       : in std_ulogic_vector(4 downto 0);
+        gpr_c_read_in       : in gpr_index_t;
 
         cr_read_in          : in std_ulogic;
         cr_write_in         : in std_ulogic;
@@ -61,6 +65,7 @@ begin
             )
         port map (
             clk                => clk,
+	    stall_in           => stall_in,
 
             gpr_write_valid_in => gpr_write_valid,
             gpr_write_in       => gpr_write_in,
@@ -76,6 +81,7 @@ begin
             )
         port map (
             clk                => clk,
+	    stall_in           => stall_in,
 
             gpr_write_valid_in => gpr_write_valid,
             gpr_write_in       => gpr_write_in,
@@ -91,11 +97,12 @@ begin
             )
         port map (
             clk                => clk,
+	    stall_in           => stall_in,
 
             gpr_write_valid_in => gpr_write_valid,
             gpr_write_in       => gpr_write_in,
             gpr_read_valid_in  => gpr_c_read_valid_in,
-            gpr_read_in        => gpr_c_read_in,
+            gpr_read_in        => "0" & gpr_c_read_in,
 
             stall_out          => stall_c_out
             );
@@ -106,6 +113,7 @@ begin
             )
         port map (
             clk                => clk,
+	    stall_in           => stall_in,
 
             cr_read_in         => cr_read_in,
             cr_write_in        => cr_write_valid,
@@ -129,8 +137,8 @@ begin
         v_int := r_int;
 
         -- asynchronous
-        valid_tmp := valid_in and not flush_in;
+        valid_tmp := valid_in and not flush_in and not stall_in;
-        stall_tmp := '0';
+        stall_tmp := stall_in;
 
         if complete_in = '1' then
             v_int.outstanding := r_int.outstanding - 1;

core.vhdl

@@ -76,8 +76,10 @@ architecture behave of core is
     signal icache_stall_out : std_ulogic;
     signal fetch2_stall_in : std_ulogic;
     signal decode1_stall_in : std_ulogic;
+    signal decode2_stall_in : std_ulogic;
     signal decode2_stall_out : std_ulogic;
     signal ex1_icache_inval: std_ulogic;
+    signal ex1_stall_out: std_ulogic;
 
     signal flush: std_ulogic;
 
@@ -184,6 +186,7 @@ begin
         port map (
             clk => clk,
             rst => core_rst,
+	    stall_in => decode2_stall_in,
             stall_out => decode2_stall_out,
             flush_in => flush,
             complete_in => complete,
@@ -198,6 +201,7 @@ begin
             c_in => cr_file_to_decode2,
             c_out => decode2_to_cr_file
             );
+    decode2_stall_in <= ex1_stall_out;
 
     register_file_0: entity work.register_file
         generic map (
@@ -223,6 +227,7 @@ begin
         port map (
             clk => clk,
             flush_out => flush,
+	    stall_out => ex1_stall_out,
             e_in => decode2_to_execute1,
             f_out => execute1_to_fetch1,
             e_out => execute1_to_writeback,

cr_file.vhdl

@@ -18,7 +18,9 @@ end entity cr_file;
 
 architecture behaviour of cr_file is
     signal crs : std_ulogic_vector(31 downto 0) := (others => '0');
-    signal crs_updated : std_ulogic_vector(31 downto 0) := (others => '0');
+    signal crs_updated : std_ulogic_vector(31 downto 0);
+    signal xerc : xer_common_t := xerc_init;
+    signal xerc_updated : xer_common_t;
 begin
     cr_create_0: process(all)
         variable hi, lo : integer := 0;
@@ -35,6 +37,13 @@ begin
         end loop;
 
         crs_updated <= cr_tmp;
+
+	if w_in.write_xerc_enable = '1' then
+	    xerc_updated <= w_in.write_xerc_data;
+	else
+	    xerc_updated <= xerc;
+	end if;
+
     end process;
 
     -- synchronous writes
@@ -43,8 +52,12 @@ begin
         if rising_edge(clk) then
             if w_in.write_cr_enable = '1' then
                 report "Writing " & to_hstring(w_in.write_cr_data) & " to CR mask " & to_hstring(w_in.write_cr_mask);
+		crs <= crs_updated;
             end if;
-            crs <= crs_updated;
+	    if w_in.write_xerc_enable = '1' then
+                report "Writing XERC";
+		xerc <= xerc_updated;
+	    end if;
         end if;
     end process;
 
@@ -56,5 +69,6 @@ begin
             report "Reading CR " & to_hstring(crs_updated);
         end if;
         d_out.read_cr_data <= crs_updated;
+        d_out.read_xerc_data <= xerc_updated;
     end process;
 end architecture behaviour;

cr_hazard.vhdl

@@ -7,7 +7,8 @@ entity cr_hazard is
         PIPELINE_DEPTH : natural := 2
         );
     port(
-        clk         : in std_logic;
+        clk         : in std_ulogic;
+	stall_in    : in std_ulogic;
 
         cr_read_in  : in std_ulogic;
         cr_write_in : in std_ulogic;
@@ -29,7 +30,9 @@ begin
     cr_hazard0: process(clk)
     begin
         if rising_edge(clk) then
-            r <= rin;
+	    if stall_in = '0' then
+		r <= rin;
+	    end if;
         end if;
     end process;
 

dcache.vhdl

@@ -185,6 +185,7 @@ architecture rtl of dcache is
 	length         : std_ulogic_vector(3 downto 0);
 	sign_extend    : std_ulogic;
 	byte_reverse   : std_ulogic;
+	xerc           : xer_common_t;
     end record;
 
     signal r2 : reg_stage_2_t;
@@ -469,6 +470,7 @@ begin
 	d_out.sign_extend <= r2.sign_extend;
 	d_out.byte_reverse <= r2.byte_reverse;
 	d_out.second_word <= '0';
+	d_out.xerc <= r2.xerc;
 
 	-- We have a valid load or store hit or we just completed a slow
 	-- op such as a load miss, a NC load or a store
@@ -518,6 +520,7 @@ begin
 		d_out.sign_extend <= r1.req.sign_extend;
 		d_out.byte_reverse <= r1.req.byte_reverse;
 		d_out.write_len <= r1.req.length;
+		d_out.xerc <= r1.req.xerc;
 	    end if;
 
 	    -- If it's a store or a non-update load form, complete now
@@ -539,6 +542,7 @@ begin
 	    d_out.write_len <= "1000";
 	    d_out.sign_extend <= '0';
 	    d_out.byte_reverse <= '0';
+	    d_out.xerc <= r1.req.xerc;
 
 	    -- If it was a load, this completes the operation (load with
 	    -- update case).

decode1.vhdl

@@ -36,14 +36,14 @@ architecture behaviour of decode1 is
         constant major_decode_rom_array : major_rom_array_t := (
 		--          unit     internal      in1         in2          in3   out   CR   CR   inv  inv  cry   cry  ldst  BR   sgn  upd  rsrv 32b  sgn  rc    lk   sgl
 		--                        op                                            in   out   A   out  in    out  len        ext                                 pipe
-		12 =>       (ALU,    OP_ADD,       RA,         CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '1', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '1'), -- addic
+		12 =>       (ALU,    OP_ADD,       RA,         CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '1', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- addic
-		13 =>       (ALU,    OP_ADD,       RA,         CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '1', NONE, '0', '0', '0', '0', '0', '0', ONE,  '0', '1'), -- addic.
+		13 =>       (ALU,    OP_ADD,       RA,         CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '1', NONE, '0', '0', '0', '0', '0', '0', ONE,  '0', '0'), -- addic.
                 14 =>       (ALU,    OP_ADD,       RA_OR_ZERO, CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- addi
 		15 =>       (ALU,    OP_ADD,       RA_OR_ZERO, CONST_SI_HI, NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- addis
 		28 =>       (ALU,    OP_AND,       NONE,       CONST_UI,    RS,   RA,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', ONE,  '0', '0'), -- andi.
 		29 =>       (ALU,    OP_AND,       NONE,       CONST_UI_HI, RS,   RA,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', ONE,  '0', '0'), -- andis.
 		18 =>       (ALU,    OP_B,         NONE,       CONST_LI,    NONE, NONE, '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '1', '0'), -- b
-		16 =>       (ALU,    OP_BC,        NONE,       CONST_BD,    NONE, NONE, '1', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '1', '0'), -- bc
+		16 =>       (ALU,    OP_BC,        SPR,        CONST_BD,    NONE, SPR , '1', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '1', '0'), -- bc
 		11 =>       (ALU,    OP_CMP,       RA,         CONST_SI,    NONE, NONE, '0', '1', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- cmpi
 		10 =>       (ALU,    OP_CMPL,      RA,         CONST_UI,    NONE, NONE, '0', '1', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- cmpli
 		34 =>       (LDST,   OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '0', '0', '0', '0', NONE, '0', '1'), -- lbz
@@ -106,7 +106,7 @@ architecture behaviour of decode1 is
 		-- addpcis not implemented yet
 		2#001#    =>       (ALU,    OP_ILLEGAL,   NONE,       NONE,        NONE, NONE, '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'),
                 -- bclr, bcctr, bctar
-		2#100#    =>       (ALU,    OP_BCREG,     NONE,       NONE,        NONE, NONE, '1', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '1', '1'),
+		2#100#    =>       (ALU,    OP_BCREG,     SPR,        SPR,         NONE, SPR,  '1', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '1', '0'),
                 -- isync
 		2#111#    =>       (ALU,    OP_ISYNC,     NONE,       NONE,        NONE, NONE, '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '1'),
 		others   => illegal_inst
@@ -133,10 +133,15 @@ architecture behaviour of decode1 is
 		--                       unit    internal      in1         in2          in3   out   CR   CR   inv  inv  cry   cry  ldst  BR   sgn  upd  rsrv 32b  sgn  rc    lk   sgl
 		--                                    op                                            in   out   A   out  in    out  len        ext                                 pipe
 		2#0100001010#  =>       (ALU,    OP_ADD,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- add
-		2#0000001010#  =>       (ALU,    OP_ADD,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- addc
+		2#1100001010#  =>       (ALU,    OP_ADD,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- addo
-		2#0010001010#  =>       (ALU,    OP_ADD,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', CA,   '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- adde
+		2#0000001010#  =>       (ALU,    OP_ADD,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- addc
-		2#0011101010#  =>       (ALU,    OP_ADD,       RA,         CONST_M1,    NONE, RT,   '0', '0', '0', '0', CA,   '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- addme
+		2#1000001010#  =>       (ALU,    OP_ADD,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- addco
-		2#0011001010#  =>       (ALU,    OP_ADD,       RA,         NONE,        NONE, RT,   '0', '0', '0', '0', CA,   '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- addze
+		2#0010001010#  =>       (ALU,    OP_ADD,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', CA,   '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- adde
+		2#1010001010#  =>       (ALU,    OP_ADD,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', CA,   '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- addeo
+		2#0011101010#  =>       (ALU,    OP_ADD,       RA,         CONST_M1,    NONE, RT,   '0', '0', '0', '0', CA,   '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- addme
+		2#1011101010#  =>       (ALU,    OP_ADD,       RA,         CONST_M1,    NONE, RT,   '0', '0', '0', '0', CA,   '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- addmeo
+		2#0011001010#  =>       (ALU,    OP_ADD,       RA,         NONE,        NONE, RT,   '0', '0', '0', '0', CA,   '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- addze
+		2#1011001010#  =>       (ALU,    OP_ADD,       RA,         NONE,        NONE, RT,   '0', '0', '0', '0', CA,   '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- addzeo
 		2#0000011100#  =>       (ALU,    OP_AND,       NONE,       RB,          RS,   RA,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- and
 		2#0000111100#  =>       (ALU,    OP_AND,       NONE,       RB,          RS,   RA,   '0', '0', '1', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- andc
 		-- 2#0011111100# bperm
@@ -156,13 +161,21 @@ architecture behaviour of decode1 is
 		2#0011110110#  =>       (ALU,    OP_NOP,       NONE,       NONE,        NONE, NONE, '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '1'), -- dcbtst
 		-- 2#1111110110# dcbz
 		2#0110001001#  =>       (DIV,    OP_DIV,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- divdeu
+		2#1110001001#  =>       (DIV,    OP_DIV,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- divdeuo
 		2#0110001011#  =>       (DIV,    OP_DIV,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- divweu
+		2#1110001011#  =>       (DIV,    OP_DIV,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- divweuo
 		2#0110101001#  =>       (DIV,    OP_DIV,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- divde
+		2#1110101001#  =>       (DIV,    OP_DIV,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- divdeo
 		2#0110101011#  =>       (DIV,    OP_DIV,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- divwe
+		2#1110101011#  =>       (DIV,    OP_DIV,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- divweo
 		2#0111001001#  =>       (DIV,    OP_DIV,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- divdu
+		2#1111001001#  =>       (DIV,    OP_DIV,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- divduo
 		2#0111001011#  =>       (DIV,    OP_DIV,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- divwu
+		2#1111001011#  =>       (DIV,    OP_DIV,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- divwuo
 		2#0111101001#  =>       (DIV,    OP_DIV,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- divd
+		2#1111101001#  =>       (DIV,    OP_DIV,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- divdo
 		2#0111101011#  =>       (DIV,    OP_DIV,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- divw
+		2#1111101011#  =>       (DIV,    OP_DIV,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- divwo
 		2#0100011100#  =>       (ALU,    OP_XOR,       NONE,       RB,          RS,   RA,   '0', '0', '0', '1', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- eqv
 		2#1110111010#  =>       (ALU,    OP_EXTS,      NONE,       NONE,        RS,   RA,   '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- extsb
 		2#1110011010#  =>       (ALU,    OP_EXTS,      NONE,       NONE,        RS,   RA,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- extsh
@@ -224,13 +237,13 @@ architecture behaviour of decode1 is
 		-- 2#1000000000# mcrxr
 		-- 2#1001000000# mcrxrx
 		2#0000010011#  =>       (ALU,    OP_MFCR,      NONE,       NONE,        NONE, RT,   '1', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- mfcr/mfocrf
-		2#0101010011#  =>       (ALU,    OP_MFSPR,     NONE,       NONE,        NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '1'), -- mfspr
+		2#0101010011#  =>       (ALU,    OP_MFSPR,     SPR,        NONE,        NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- mfspr
 		2#0100001001#  =>       (DIV,    OP_MOD,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '1'), -- modud
 		2#0100001011#  =>       (DIV,    OP_MOD,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '1'), -- moduw
 		2#1100001001#  =>       (DIV,    OP_MOD,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '1'), -- modsd
 		2#1100001011#  =>       (DIV,    OP_MOD,       RA,         RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '1'), -- modsw
 		2#0010010000#  =>       (ALU,    OP_MTCRF,     NONE,       NONE,        RS,   NONE, '0', '1', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- mtcrf/mtocrf
-		2#0111010011#  =>       (ALU,    OP_MTSPR,     NONE,       NONE,        RS,   NONE, '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '1'), -- mtspr
+		2#0111010011#  =>       (ALU,    OP_MTSPR,     NONE,       NONE,        RS,   SPR,  '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- mtspr
 		2#0001001001#  =>       (MUL,    OP_MUL_H64,   RA,         RB,          NONE, RT,   '0', '1', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '1', RC,   '0', '1'), -- mulhd
 		2#0000001001#  =>       (MUL,    OP_MUL_H64,   RA,         RB,          NONE, RT,   '0', '1', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- mulhdu
 		2#0001001011#  =>       (MUL,    OP_MUL_H32,   RA,         RB,          NONE, RT,   '0', '1', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '1', '1', RC,   '0', '1'), -- mulhw
@@ -241,9 +254,12 @@ architecture behaviour of decode1 is
 		2#1001001011#  =>       (MUL,    OP_MUL_H32,   RA,         RB,          NONE, RT,   '0', '1', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '1', '1', RC,   '0', '1'), -- mulhw
 		2#1000001011#  =>       (MUL,    OP_MUL_H32,   RA,         RB,          NONE, RT,   '0', '1', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '1', '0', RC,   '0', '1'), -- mulhwu
 		2#0011101001#  =>       (MUL,    OP_MUL_L64,   RA,         RB,          NONE, RT,   '0', '1', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '1', RC,   '0', '1'), -- mulld
+		2#1011101001#  =>       (MUL,    OP_MUL_L64,   RA,         RB,          NONE, RT,   '0', '1', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '1', RC,   '0', '1'), -- mulldo
 		2#0011101011#  =>       (MUL,    OP_MUL_L64,   RA,         RB,          NONE, RT,   '0', '1', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '1', '1', RC,   '0', '1'), -- mullw
+		2#1011101011#  =>       (MUL,    OP_MUL_L64,   RA,         RB,          NONE, RT,   '0', '1', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '1', '1', RC,   '0', '1'), -- mullwo
 		2#0111011100#  =>       (ALU,    OP_AND,       NONE,       RB,          RS,   RA,   '0', '0', '0', '1', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- nand
 		2#0001101000#  =>       (ALU,    OP_ADD,       RA,         NONE,        NONE, RT,   '0', '0', '1', '0', ONE,  '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- neg
+		2#1001101000#  =>       (ALU,    OP_ADD,       RA,         NONE,        NONE, RT,   '0', '0', '1', '0', ONE,  '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- nego
 		2#0001111100#  =>       (ALU,    OP_OR,        NONE,       RB,          RS,   RA,   '0', '0', '0', '1', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- nor
 		2#0110111100#  =>       (ALU,    OP_OR,        NONE,       RB,          RS,   RA,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- or
 		2#0110011100#  =>       (ALU,    OP_OR,        NONE,       RB,          RS,   RA,   '0', '0', '1', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- orc
@@ -278,10 +294,15 @@ architecture behaviour of decode1 is
 		2#0010110111#  =>       (LDST,   OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '0', '0', NONE, '0', '1'), -- stwux
 		2#0010010111#  =>       (LDST,   OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '0', '0', NONE, '0', '1'), -- stwx
 		2#0000101000#  =>       (ALU,    OP_ADD,       RA,         RB,          NONE, RT,   '0', '0', '1', '0', ONE,  '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- subf
-		2#0000001000#  =>       (ALU,    OP_ADD,       RA,         RB,          NONE, RT,   '0', '0', '1', '0', ONE,  '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- subfc
+		2#1000101000#  =>       (ALU,    OP_ADD,       RA,         RB,          NONE, RT,   '0', '0', '1', '0', ONE,  '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- subfo
-		2#0010001000#  =>       (ALU,    OP_ADD,       RA,         RB,          NONE, RT,   '0', '0', '1', '0', CA,   '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- subfe
+		2#0000001000#  =>       (ALU,    OP_ADD,       RA,         RB,          NONE, RT,   '0', '0', '1', '0', ONE,  '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- subfc
-		2#0011101000#  =>       (ALU,    OP_ADD,       RA,         CONST_M1,    NONE, RT,   '0', '0', '1', '0', CA,   '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- subfme
+		2#1000001000#  =>       (ALU,    OP_ADD,       RA,         RB,          NONE, RT,   '0', '0', '1', '0', ONE,  '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- subfco
-		2#0011001000#  =>       (ALU,    OP_ADD,       RA,         NONE,        NONE, RT,   '0', '0', '1', '0', CA,   '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '1'), -- subfze
+		2#0010001000#  =>       (ALU,    OP_ADD,       RA,         RB,          NONE, RT,   '0', '0', '1', '0', CA,   '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- subfe
+		2#1010001000#  =>       (ALU,    OP_ADD,       RA,         RB,          NONE, RT,   '0', '0', '1', '0', CA,   '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- subfeo
+		2#0011101000#  =>       (ALU,    OP_ADD,       RA,         CONST_M1,    NONE, RT,   '0', '0', '1', '0', CA,   '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- subfme
+		2#1011101000#  =>       (ALU,    OP_ADD,       RA,         CONST_M1,    NONE, RT,   '0', '0', '1', '0', CA,   '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- subfmeo
+		2#0011001000#  =>       (ALU,    OP_ADD,       RA,         NONE,        NONE, RT,   '0', '0', '1', '0', CA,   '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- subfze
+		2#1011001000#  =>       (ALU,    OP_ADD,       RA,         NONE,        NONE, RT,   '0', '0', '1', '0', CA,   '1', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0'), -- subfzeo
 		2#1001010110#  =>       (ALU,    OP_NOP,       NONE,       NONE,        NONE, NONE, '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '1'), -- sync
 		-- 2#0001000100# td
 		2#0000000100#  =>       (ALU,    OP_TW,        RA,         RB,          NONE, NONE, '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '1'), -- tw
@@ -334,6 +355,8 @@ begin
 		v.nia  := f_in.nia;
 		v.insn := f_in.insn;
 		v.stop_mark := f_in.stop_mark;
+		v.ispr1 := (others => '0');
+		v.ispr2 := (others => '0');
 
 		if f_in.valid = '1' then
 			report "Decode insn " & to_hstring(f_in.insn) & " at " & to_hstring(f_in.nia);
@@ -377,6 +400,33 @@ begin
                         v.decode := major_decode_rom_array(to_integer(majorop));
 		end if;
 
+		-- Set ISPR1/ISPR2 when needed
+		if v.decode.insn_type = OP_BC or v.decode.insn_type = OP_BCREG then
+		    -- Branch uses CTR as condition when BO(2) is 0. This is
+		    -- also used to indicate that CTR is modified (they go
+		    -- together).
+		    --
+		    if f_in.insn(23) = '0' then
+			v.ispr1 := fast_spr_num(SPR_CTR);
+		    end if;
+
+		    -- Branch source register is an SPR
+		    if v.decode.insn_type = OP_BCREG then
+			-- TODO: Add TAR
+			if f_in.insn(10) = '0' then
+			    v.ispr2 := fast_spr_num(SPR_LR);
+			else
+			    v.ispr2 := fast_spr_num(SPR_CTR);
+			end if;
+		    end if;
+		elsif v.decode.insn_type = OP_MFSPR or v.decode.insn_type = OP_MTSPR then
+		    v.ispr1 := fast_spr_num(decode_spr_num(f_in.insn));
+		    -- Make slow SPRs single issue
+		    if is_fast_spr(v.ispr1) = '0' then
+			v.decode.sgl_pipe := '1';
+		    end if;
+		end if;
+
 		if flush_in = '1' then
 			v.valid := '0';
 		end if;

decode2.vhdl

@@ -14,6 +14,7 @@ entity decode2 is
 		rst   : in std_ulogic;
 
 		complete_in : in std_ulogic;
+		stall_in : in std_ulogic;
 		stall_out : out std_ulogic;
 
 		stopped_out : out std_ulogic;
@@ -47,30 +48,46 @@ architecture behaviour of decode2 is
 
 	type decode_input_reg_t is record
 		reg_valid : std_ulogic;
-		reg       : std_ulogic_vector(4 downto 0);
+		reg       : gspr_index_t;
 		data      : std_ulogic_vector(63 downto 0);
 	end record;
 
+	type decode_output_reg_t is record
+		reg_valid : std_ulogic;
+		reg       : gspr_index_t;
+	end record;
+
 	function decode_input_reg_a (t : input_reg_a_t; insn_in : std_ulogic_vector(31 downto 0);
-				     reg_data : std_ulogic_vector(63 downto 0)) return decode_input_reg_t is
+				     reg_data : std_ulogic_vector(63 downto 0);
-		variable is_reg : std_ulogic;
+				     ispr : gspr_index_t) return decode_input_reg_t is
 	begin
-		is_reg := '0' when insn_ra(insn_in) = "00000" else '1';
-
 		if t = RA or (t = RA_OR_ZERO and insn_ra(insn_in) /= "00000") then
-			--return (is_reg, insn_ra(insn_in), reg_data);
+		    assert is_fast_spr(ispr) = '0' report "Decode A says GPR but ISPR says SPR:" &
-			return ('1', insn_ra(insn_in), reg_data);
+			to_hstring(ispr) severity failure;
+		    return ('1', gpr_to_gspr(insn_ra(insn_in)), reg_data);
+		elsif t = SPR then
+		    -- ISPR must be either a valid fast SPR number or all 0 for a slow SPR.
+		    -- If it's all 0, we don't treat it as a dependency as slow SPRs
+		    -- operations are single issue.
+		    --
+		    assert is_fast_spr(ispr) =  '1' or ispr = "000000"
+			report "Decode A says SPR but ISPR is invalid:" &
+			to_hstring(ispr) severity failure;
+		    return (is_fast_spr(ispr), ispr, reg_data);
 		else
 			return ('0', (others => '0'), (others => '0'));
 		end if;
 	end;
 
 	function decode_input_reg_b (t : input_reg_b_t; insn_in : std_ulogic_vector(31 downto 0);
-				     reg_data : std_ulogic_vector(63 downto 0)) return decode_input_reg_t is
+				     reg_data : std_ulogic_vector(63 downto 0);
+				     ispr : gspr_index_t) return decode_input_reg_t is
 	begin
 		case t is
 		when RB =>
-			return ('1', insn_rb(insn_in), reg_data);
+		    assert is_fast_spr(ispr) = '0' report "Decode B says GPR but ISPR says SPR:" &
+			to_hstring(ispr) severity failure;
+		    return ('1', gpr_to_gspr(insn_rb(insn_in)), reg_data);
 		when CONST_UI =>
 			return ('0', (others => '0'), std_ulogic_vector(resize(unsigned(insn_ui(insn_in)), 64)));
 		when CONST_SI =>
@@ -91,6 +108,14 @@ architecture behaviour of decode2 is
 			return ('0', (others => '0'), x"00000000000000" & "00" & insn_in(1) & insn_in(15 downto 11));
 		when CONST_SH32 =>
 			return ('0', (others => '0'), x"00000000000000" & "000" & insn_in(15 downto 11));
+		when SPR =>
+		    -- ISPR must be either a valid fast SPR number or all 0 for a slow SPR.
+		    -- If it's all 0, we don't treat it as a dependency as slow SPRs
+		    -- operations are single issue.
+		    assert is_fast_spr(ispr) = '1' or ispr = "000000"
+			report "Decode B says SPR but ISPR is invalid:" &
+			to_hstring(ispr) severity failure;
+		    return (is_fast_spr(ispr), ispr, reg_data);
 		when NONE =>
 			return ('0', (others => '0'), (others => '0'));
 		end case;
@@ -101,21 +126,30 @@ architecture behaviour of decode2 is
 	begin
 		case t is
 		when RS =>
-			return ('1', insn_rs(insn_in), reg_data);
+			return ('1', gpr_to_gspr(insn_rs(insn_in)), reg_data);
 		when NONE =>
 			return ('0', (others => '0'), (others => '0'));
 		end case;
 	end;
 
-	function decode_output_reg (t : output_reg_a_t; insn_in : std_ulogic_vector(31 downto 0)) return std_ulogic_vector is
+	function decode_output_reg (t : output_reg_a_t; insn_in : std_ulogic_vector(31 downto 0);
+				    ispr : gspr_index_t) return decode_output_reg_t is
 	begin
 		case t is
 		when RT =>
-			return insn_rt(insn_in);
+		    return ('1', gpr_to_gspr(insn_rt(insn_in)));
 		when RA =>
-			return insn_ra(insn_in);
+		    return ('1', gpr_to_gspr(insn_ra(insn_in)));
+		when SPR =>
+		    -- ISPR must be either a valid fast SPR number or all 0 for a slow SPR.
+		    -- If it's all 0, we don't treat it as a dependency as slow SPRs
+		    -- operations are single issue.
+		    assert is_fast_spr(ispr) = '1' or ispr = "000000"
+			report "Decode B says SPR but ISPR is invalid:" &
+			to_hstring(ispr) severity failure;
+		    return (is_fast_spr(ispr), ispr);
 		when NONE =>
-			return "00000";
+		    return ('0', "000000");
 		end case;
 	end;
 
@@ -131,22 +165,38 @@ architecture behaviour of decode2 is
 		end case;
 	end;
 
+	-- For now, use "rc" in the decode table to decide whether oe exists.
+	-- This is not entirely correct architecturally: For mulhd and
+	-- mulhdu, the OE field is reserved. It remains to be seen what an
+	-- actual POWER9 does if we set it on those instructions, for now we
+	-- test that further down when assigning to the multiplier oe input.
+	--
+	function decode_oe (t : rc_t; insn_in : std_ulogic_vector(31 downto 0)) return std_ulogic is
+	begin
+		case t is
+		when RC =>
+			return insn_oe(insn_in);
+		when OTHERS =>
+			return '0';
+		end case;
+	end;
+
 	-- issue control signals
 	signal control_valid_in : std_ulogic;
 	signal control_valid_out : std_ulogic;
 	signal control_sgl_pipe : std_logic;
 
 	signal gpr_write_valid : std_ulogic;
-	signal gpr_write : std_ulogic_vector(4 downto 0);
+	signal gpr_write : gspr_index_t;
 
 	signal gpr_a_read_valid : std_ulogic;
-	signal gpr_a_read : std_ulogic_vector(4 downto 0);
+	signal gpr_a_read :gspr_index_t;
 
 	signal gpr_b_read_valid : std_ulogic;
-	signal gpr_b_read : std_ulogic_vector(4 downto 0);
+	signal gpr_b_read : gspr_index_t;
 
 	signal gpr_c_read_valid : std_ulogic;
-	signal gpr_c_read : std_ulogic_vector(4 downto 0);
+	signal gpr_c_read : gpr_index_t;
 
 	signal cr_write_valid : std_ulogic;
 begin
@@ -160,6 +210,7 @@ begin
 
 		complete_in => complete_in,
 		valid_in    => control_valid_in,
+		stall_in    => stall_in,
 		flush_in    => flush_in,
 		sgl_pipe_in => control_sgl_pipe,
 		stop_mark_in => d_in.stop_mark,
@@ -194,8 +245,8 @@ begin
 		end if;
 	end process;
 
-	r_out.read1_reg <= insn_ra(d_in.insn);
+	r_out.read1_reg <= gpr_or_spr_to_gspr(insn_ra(d_in.insn), d_in.ispr1);
-	r_out.read2_reg <= insn_rb(d_in.insn);
+	r_out.read2_reg <= gpr_or_spr_to_gspr(insn_rb(d_in.insn), d_in.ispr2);
 	r_out.read3_reg <= insn_rs(d_in.insn);
 
 	c_out.read <= d_in.decode.input_cr;
@@ -207,6 +258,7 @@ begin
 		variable decoded_reg_a : decode_input_reg_t;
 		variable decoded_reg_b : decode_input_reg_t;
 		variable decoded_reg_c : decode_input_reg_t;
+		variable decoded_reg_o : decode_output_reg_t;
                 variable signed_division: std_ulogic;
                 variable length : std_ulogic_vector(3 downto 0);
 	begin
@@ -223,10 +275,11 @@ begin
 		--v.e.input_cr := d_in.decode.input_cr;
 		--v.m.input_cr := d_in.decode.input_cr;
 		--v.e.output_cr := d_in.decode.output_cr;
-
+    
-		decoded_reg_a := decode_input_reg_a (d_in.decode.input_reg_a, d_in.insn, r_in.read1_data);
+		decoded_reg_a := decode_input_reg_a (d_in.decode.input_reg_a, d_in.insn, r_in.read1_data, d_in.ispr1);
-		decoded_reg_b := decode_input_reg_b (d_in.decode.input_reg_b, d_in.insn, r_in.read2_data);
+		decoded_reg_b := decode_input_reg_b (d_in.decode.input_reg_b, d_in.insn, r_in.read2_data, d_in.ispr2);
 		decoded_reg_c := decode_input_reg_c (d_in.decode.input_reg_c, d_in.insn, r_in.read3_data);
+		decoded_reg_o := decode_output_reg (d_in.decode.output_reg_a, d_in.insn, d_in.ispr1);
 
 		r_out.read1_enable <= decoded_reg_a.reg_valid;
 		r_out.read2_enable <= decoded_reg_b.reg_valid;
@@ -253,9 +306,11 @@ begin
 		v.e.read_reg2 := decoded_reg_b.reg;
 		v.e.read_data2 := decoded_reg_b.data;
                 v.e.read_data3 := decoded_reg_c.data;
-		v.e.write_reg := decode_output_reg(d_in.decode.output_reg_a, d_in.insn);
+		v.e.write_reg := decoded_reg_o.reg;
 		v.e.rc := decode_rc(d_in.decode.rc, d_in.insn);
+		v.e.oe := decode_oe(d_in.decode.rc, d_in.insn);
 		v.e.cr := c_in.read_cr_data;
+		v.e.xerc := c_in.read_xerc_data;
                 v.e.invert_a := d_in.decode.invert_a;
                 v.e.invert_out := d_in.decode.invert_out;
 		v.e.input_carry := d_in.decode.input_carry;
@@ -272,8 +327,13 @@ begin
 		v.m.insn_type := d_in.decode.insn_type;
 		mul_a := decoded_reg_a.data;
 		mul_b := decoded_reg_b.data;
-		v.m.write_reg := decode_output_reg(d_in.decode.output_reg_a, d_in.insn);
+		v.m.write_reg := gspr_to_gpr(decoded_reg_o.reg);
 		v.m.rc := decode_rc(d_in.decode.rc, d_in.insn);
+		v.m.xerc := c_in.read_xerc_data;
+		if v.m.insn_type = OP_MUL_L64 then
+		  v.m.oe := decode_oe(d_in.decode.rc, d_in.insn);
+		end if;
+		v.m.is_32bit := d_in.decode.is_32bit;
 
 		if d_in.decode.is_32bit = '1' then
 			if d_in.decode.is_signed = '1' then
@@ -304,7 +364,7 @@ begin
                 --       s = 1 for signed, 0 for unsigned (for div*)
                 --       t = 1 for 32-bit, 0 for 64-bit
                 --       r = RC bit (record condition code)
-		v.d.write_reg := decode_output_reg(d_in.decode.output_reg_a, d_in.insn);
+		v.d.write_reg := gspr_to_gpr(decoded_reg_o.reg);
                 v.d.is_modulus := not d_in.insn(8);
                 v.d.is_32bit := d_in.insn(2);
                 if d_in.insn(8) = '1' then
@@ -337,13 +397,15 @@ begin
                         end if;
                 end if;
                 v.d.rc := decode_rc(d_in.decode.rc, d_in.insn);
+		v.d.xerc := c_in.read_xerc_data;
+		v.d.oe := decode_oe(d_in.decode.rc, d_in.insn);
 
 		-- load/store unit
-		v.l.update_reg := decoded_reg_a.reg;
+		v.l.update_reg := gspr_to_gpr(decoded_reg_a.reg);
 		v.l.addr1 := decoded_reg_a.data;
 		v.l.addr2 := decoded_reg_b.data;
 		v.l.data := decoded_reg_c.data;
-		v.l.write_reg := decode_output_reg(d_in.decode.output_reg_a, d_in.insn);
+		v.l.write_reg := gspr_to_gpr(decoded_reg_o.reg);
 
 		if d_in.decode.insn_type = OP_LOAD then
 			v.l.load := '1';
@@ -355,13 +417,14 @@ begin
 		v.l.byte_reverse := d_in.decode.byte_reverse;
 		v.l.sign_extend := d_in.decode.sign_extend;
 		v.l.update := d_in.decode.update;
+		v.l.xerc := c_in.read_xerc_data;
 
 		-- issue control
 		control_valid_in <= d_in.valid;
 		control_sgl_pipe <= d_in.decode.sgl_pipe;
 
-		gpr_write_valid <= '1' when d_in.decode.output_reg_a /= NONE else '0';
+		gpr_write_valid <= decoded_reg_o.reg_valid;
-		gpr_write <= decode_output_reg(d_in.decode.output_reg_a, d_in.insn);
+		gpr_write <= decoded_reg_o.reg;
 
 		gpr_a_read_valid <= decoded_reg_a.reg_valid;
 		gpr_a_read <= decoded_reg_a.reg;
@@ -370,7 +433,7 @@ begin
 		gpr_b_read <= decoded_reg_b.reg;
 
 		gpr_c_read_valid <= decoded_reg_c.reg_valid;
-		gpr_c_read <= decoded_reg_c.reg;
+		gpr_c_read <= gspr_to_gpr(decoded_reg_c.reg);
 
                 cr_write_valid <= d_in.decode.output_cr or decode_rc(d_in.decode.rc, d_in.insn);
 

decode_types.vhdl

@@ -21,10 +21,10 @@ package decode_types is
 			 OP_TWI, OP_XOR, OP_SIM_CONFIG
 			 );
 
-    type input_reg_a_t is (NONE, RA, RA_OR_ZERO);
+    type input_reg_a_t is (NONE, RA, RA_OR_ZERO, SPR);
-    type input_reg_b_t is (NONE, RB, CONST_UI, CONST_SI, CONST_SI_HI, CONST_UI_HI, CONST_LI, CONST_BD, CONST_DS, CONST_M1, CONST_SH, CONST_SH32);
+    type input_reg_b_t is (NONE, RB, CONST_UI, CONST_SI, CONST_SI_HI, CONST_UI_HI, CONST_LI, CONST_BD, CONST_DS, CONST_M1, CONST_SH, CONST_SH32, SPR);
     type input_reg_c_t is (NONE, RS);
-    type output_reg_a_t is (NONE, RT, RA);
+    type output_reg_a_t is (NONE, RT, RA, SPR);
     type rc_t is (NONE, ONE, RC);
     type carry_in_t is (ZERO, CA, ONE);
 

divider.vhdl

@@ -20,7 +20,7 @@ architecture behaviour of divider is
     signal div        : unsigned(63 downto 0);
     signal quot       : std_ulogic_vector(63 downto 0);
     signal result     : std_ulogic_vector(63 downto 0);
-    signal sresult    : std_ulogic_vector(63 downto 0);
+    signal sresult    : std_ulogic_vector(64 downto 0);
     signal oresult    : std_ulogic_vector(63 downto 0);
     signal qbit       : std_ulogic;
     signal running    : std_ulogic;
@@ -36,7 +36,8 @@ architecture behaviour of divider is
     signal overflow   : std_ulogic;
     signal ovf32      : std_ulogic;
     signal did_ovf    : std_ulogic;
-
+    signal oe         : std_ulogic;
+    signal xerc       : xer_common_t;
 begin
     divider_0: process(clk)
     begin
@@ -62,6 +63,8 @@ begin
                 is_32bit <= d_in.is_32bit;
                 is_signed <= d_in.is_signed;
                 rc <= d_in.rc;
+                oe <= d_in.oe;
+		xerc <= d_in.xerc;
                 count <= "1111111";
                 running <= '1';
                 overflow <= '0';
@@ -120,13 +123,13 @@ begin
             result <= quot;
         end if;
         if neg_result = '1' then
-            sresult <= std_ulogic_vector(- signed(result));
+            sresult <= std_ulogic_vector(- signed('0' & result));
         else
-            sresult <= result;
+            sresult <= '0' & result;
         end if;
         did_ovf <= '0';
         if is_32bit = '0' then
-            did_ovf <= overflow or (is_signed and (sresult(63) xor neg_result));
+            did_ovf <= overflow or (is_signed and (sresult(64) xor sresult(63)));
         elsif is_signed = '1' then
             if ovf32 = '1' or sresult(32) /= sresult(31) then
                 did_ovf <= '1';
@@ -140,20 +143,32 @@ begin
             -- 32-bit divisions set the top 32 bits of the result to 0
             oresult <= x"00000000" & sresult(31 downto 0);
         else
-            oresult <= sresult;
+            oresult <= sresult(63 downto 0);
         end if;
     end process;
 
     divider_out: process(clk)
     begin
         if rising_edge(clk) then
+	    d_out.valid <= '0';
             d_out.write_reg_data <= oresult;
+	    d_out.write_reg_enable <= '0';
+	    d_out.write_xerc_enable <= '0';
+	    d_out.xerc <= xerc;
             if count = "1000000" then
                 d_out.valid <= '1';
                 d_out.write_reg_enable <= '1';
-            else
+		d_out.write_xerc_enable <= oe;
-                d_out.valid <= '0';
+
-                d_out.write_reg_enable <= '0';
+		-- We must test oe because the RC update code in writeback
+		-- will use the xerc value to set CR0:SO so we must not clobber
+		-- xerc if OE wasn't set.
+		--
+		if oe = '1' then
+		    d_out.xerc.ov <= did_ovf;
+		    d_out.xerc.ov32 <= did_ovf;
+		    d_out.xerc.so <= xerc.so or did_ovf;
+		end if;
             end if;
         end if;
     end process;

execute1.vhdl

@@ -12,10 +12,11 @@ use work.ppc_fx_insns.all;
 
 entity execute1 is
     port (
-	clk   : in std_logic;
+	clk   : in std_ulogic;
 
 	-- asynchronous
 	flush_out : out std_ulogic;
+	stall_out : out std_ulogic;
 
 	e_in  : in Decode2ToExecute1Type;
 
@@ -31,14 +32,15 @@ end entity execute1;
 
 architecture behaviour of execute1 is
     type reg_type is record
-	--f : Execute1ToFetch1Type;
 	e : Execute1ToWritebackType;
+	lr_update : std_ulogic;
+	next_lr : std_ulogic_vector(63 downto 0);
     end record;
 
     signal r, rin : reg_type;
 
-    signal ctrl: ctrl_t := (carry => '0', others => (others => '0'));
+    signal ctrl: ctrl_t := (others => (others => '0'));
-    signal ctrl_tmp: ctrl_t := (carry => '0', others => (others => '0'));
+    signal ctrl_tmp: ctrl_t := (others => (others => '0'));
 
     signal right_shift, rot_clear_left, rot_clear_right: std_ulogic;
     signal rotator_result: std_ulogic_vector(63 downto 0);
@@ -46,17 +48,46 @@ architecture behaviour of execute1 is
     signal logical_result: std_ulogic_vector(63 downto 0);
     signal countzero_result: std_ulogic_vector(63 downto 0);
 
-    function decode_input_carry (carry_sel : carry_in_t; ca_in : std_ulogic) return std_ulogic is
+    procedure set_carry(e: inout Execute1ToWritebackType;
+			carry32 : in std_ulogic;
+			carry : in std_ulogic) is
     begin
-	case carry_sel is
+	e.xerc.ca32 := carry32;
+	e.xerc.ca := carry;
+	e.write_xerc_enable := '1';
+    end;
+
+    procedure set_ov(e: inout Execute1ToWritebackType;
+		     ov   : in std_ulogic;
+		     ov32 : in std_ulogic) is
+    begin
+	e.xerc.ov32 := ov32;
+	e.xerc.ov := ov;
+	if ov = '1' then
+	    e.xerc.so := '1';
+	end if;
+	e.write_xerc_enable := '1';
+    end;
+
+    function calc_ov(msb_a : std_ulogic; msb_b: std_ulogic;
+		     ca: std_ulogic; msb_r: std_ulogic) return std_ulogic is
+    begin
+	return (ca xor msb_r) and not (msb_a xor msb_b);
+    end;
+
+    function decode_input_carry(ic : carry_in_t;
+				xerc : xer_common_t) return std_ulogic is
+    begin
+	case ic is
 	when ZERO =>
 	    return '0';
 	when CA =>
-	    return ca_in;
+	    return xerc.ca;
 	when ONE =>
 	    return '1';
 	end case;
     end;
+
 begin
 
     rotator_0: entity work.rotator
@@ -97,6 +128,12 @@ begin
 	if rising_edge(clk) then
 	    r <= rin;
 	    ctrl <= ctrl_tmp;
+	    assert not (r.lr_update = '1' and e_in.valid = '1')
+		report "LR update collision with valid in EX1"
+		severity failure;
+	    if r.lr_update = '1' then
+		report "LR update to " & to_hstring(r.next_lr);
+	    end if;
 	end if;
     end process;
 
@@ -117,6 +154,7 @@ begin
 	variable bf, bfa : std_ulogic_vector(2 downto 0);
 	variable l : std_ulogic;
 	variable next_nia : std_ulogic_vector(63 downto 0);
+        variable carry_32, carry_64 : std_ulogic;
     begin
 	result := (others => '0');
 	result_with_carry := (others => '0');
@@ -125,7 +163,43 @@ begin
 
 	v := r;
 	v.e := Execute1ToWritebackInit;
-	--v.f := Execute1ToFetch1TypeInit;
+
+	-- XER forwarding. To avoid having to track XER hazards, we
+	-- use the previously latched value.
+	--
+	-- If the XER was modified by a multiply or a divide, those are
+	-- single issue, we'll get the up to date value from decode2 from
+	-- the register file.
+	--
+	-- If it was modified by an instruction older than the previous
+	-- one in EX1, it will have also hit writeback and will be up
+	-- to date in decode2.
+	--
+	-- That leaves us with the case where it was updated by the previous
+	-- instruction in EX1. In that case, we can forward it back here.
+	--
+	-- This will break if we allow pipelining of multiply and divide,
+	-- but ideally, those should go via EX1 anyway and run as a state
+	-- machine from here.
+	--
+	-- One additional hazard to beware of is an XER:SO modifying instruction
+	-- in EX1 followed immediately by a store conditional. Due to our
+	-- writeback latency, the store will go down the LSU with the previous
+	-- XER value, thus the stcx. will set CR0:SO using an obsolete SO value.
+	--
+	-- We will need to handle that if we ever make stcx. not single issue
+	--
+	-- We always pass a valid XER value downto writeback even when
+	-- we aren't updating it, in order for XER:SO -> CR0:SO transfer
+	-- to work for RC instructions.
+	--
+	if r.e.write_xerc_enable = '1' then
+	    v.e.xerc := r.e.xerc;
+	else
+	    v.e.xerc := e_in.xerc;
+	end if;
+
+	v.lr_update := '0';
 
 	ctrl_tmp <= ctrl;
 	-- FIXME: run at 512MHz not core freq
@@ -133,6 +207,7 @@ begin
 
 	terminate_out <= '0';
 	icache_inval <= '0';
+	stall_out <= '0';
 	f_out <= Execute1ToFetch1TypeInit;
 
 	-- Next insn adder used in a couple of places
@@ -163,10 +238,18 @@ begin
 		else
 		    a_inv := not e_in.read_data1;
 		end if;
-		result_with_carry := ppc_adde(a_inv, e_in.read_data2, decode_input_carry(e_in.input_carry, ctrl.carry));
+		result_with_carry := ppc_adde(a_inv, e_in.read_data2,
+					      decode_input_carry(e_in.input_carry, v.e.xerc));
 		result := result_with_carry(63 downto 0);
-		if e_in.output_carry then
+                carry_32 := result(32) xor a_inv(32) xor e_in.read_data2(32);
-		    ctrl_tmp.carry <= result_with_carry(64);
+                carry_64 := result_with_carry(64);
+		if e_in.output_carry = '1' then
+		    set_carry(v.e, carry_32, carry_64);
+		end if;
+		if e_in.oe = '1' then
+		    set_ov(v.e,
+			   calc_ov(a_inv(63), e_in.read_data2(63), carry_64, result_with_carry(63)),
+			   calc_ov(a_inv(31), e_in.read_data2(31), carry_32, result_with_carry(31)));
 		end if;
 		result_en := '1';
 	    when OP_AND | OP_OR | OP_XOR =>
@@ -180,12 +263,15 @@ begin
 		    f_out.redirect_nia <= std_ulogic_vector(signed(e_in.nia) + signed(e_in.read_data2));
 		end if;
 	    when OP_BC =>
+		-- read_data1 is CTR
 		bo := insn_bo(e_in.insn);
 		bi := insn_bi(e_in.insn);
 		if bo(4-2) = '0' then
-		    ctrl_tmp.ctr <= std_ulogic_vector(unsigned(ctrl.ctr) - 1);
+		    result := std_ulogic_vector(unsigned(e_in.read_data1) - 1);
+		    result_en := '1';
+		    v.e.write_reg := fast_spr_num(SPR_CTR);
 		end if;
-		if ppc_bc_taken(bo, bi, e_in.cr, ctrl.ctr) = 1 then
+		if ppc_bc_taken(bo, bi, e_in.cr, e_in.read_data1) = 1 then
 		    f_out.redirect <= '1';
 		    if (insn_aa(e_in.insn)) then
 			f_out.redirect_nia <= std_ulogic_vector(signed(e_in.read_data2));
@@ -194,19 +280,18 @@ begin
 		    end if;
 		end if;
 	    when OP_BCREG =>
-                -- bits 10 and 6 distinguish between bclr, bcctr and bctar
+		-- read_data1 is CTR
+		-- read_data2 is target register (CTR, LR or TAR)
 		bo := insn_bo(e_in.insn);
 		bi := insn_bi(e_in.insn);
 		if bo(4-2) = '0' and e_in.insn(10) = '0' then
-		    ctrl_tmp.ctr <= std_ulogic_vector(unsigned(ctrl.ctr) - 1);
+		    result := std_ulogic_vector(unsigned(e_in.read_data1) - 1);
+		    result_en := '1';
+		    v.e.write_reg := fast_spr_num(SPR_CTR);
 		end if;
-		if ppc_bc_taken(bo, bi, e_in.cr, ctrl.ctr) = 1 then
+		if ppc_bc_taken(bo, bi, e_in.cr, e_in.read_data1) = 1 then
 		    f_out.redirect <= '1';
-		    if e_in.insn(10) = '0' then
+		    f_out.redirect_nia <= e_in.read_data2(63 downto 2) & "00";
-			f_out.redirect_nia <= ctrl.lr(63 downto 2) & "00";
-		    else
-			f_out.redirect_nia <= ctrl.ctr(63 downto 2) & "00";
-		    end if;
 		end if;
 	    when OP_CMPB =>
 		result := ppc_cmpb(e_in.read_data3, e_in.read_data2);
@@ -220,7 +305,7 @@ begin
 		for i in 0 to 7 loop
 		    lo := i*4;
 		    hi := lo + 3;
-		    v.e.write_cr_data(hi downto lo) := ppc_cmp(l, e_in.read_data1, e_in.read_data2);
+		    v.e.write_cr_data(hi downto lo) := ppc_cmp(l, e_in.read_data1, e_in.read_data2, v.e.xerc.so);
 		end loop;
 	    when OP_CMPL =>
 		bf := insn_bf(e_in.insn);
@@ -231,7 +316,7 @@ begin
 		for i in 0 to 7 loop
 		    lo := i*4;
 		    hi := lo + 3;
-		    v.e.write_cr_data(hi downto lo) := ppc_cmpl(l, e_in.read_data1, e_in.read_data2);
+		    v.e.write_cr_data(hi downto lo) := ppc_cmpl(l, e_in.read_data1, e_in.read_data2, v.e.xerc.so);
 		end loop;
 	    when OP_CNTZ =>
 		result := countzero_result;
@@ -269,16 +354,24 @@ begin
 		    v.e.write_cr_data(hi downto lo) := newcrf;
 		end loop;
 	    when OP_MFSPR =>
-		case decode_spr_num(e_in.insn) is
+		if is_fast_spr(e_in.read_reg1) then
-		when SPR_CTR =>
+		    result := e_in.read_data1;
-		    result := ctrl.ctr;
+		    if decode_spr_num(e_in.insn) = SPR_XER then
-		when SPR_LR =>
+			result(63-32) := v.e.xerc.so;
-		    result := ctrl.lr;
+			result(63-33) := v.e.xerc.ov;
-		when SPR_TB =>
+			result(63-34) := v.e.xerc.ca;
-		    result := ctrl.tb;
+			result(63-35 downto 63-43) := "000000000";
-		when others =>
+			result(63-44) := v.e.xerc.ov32;
-		    result := (others => '0');
+			result(63-45) := v.e.xerc.ca32;
-		end case;
+		    end if;
+		else
+		    case decode_spr_num(e_in.insn) is
+		    when SPR_TB =>
+			result := ctrl.tb;
+		    when others =>
+			result := (others => '0');
+		    end case;
+		end if;
 		result_en := '1';
 	    when OP_MFCR =>
 		if e_in.insn(20) = '0' then
@@ -309,13 +402,25 @@ begin
 		end if;
 		v.e.write_cr_data := e_in.read_data3(31 downto 0);
 	    when OP_MTSPR =>
-		case decode_spr_num(e_in.insn) is
+		report "MTSPR to SPR " & integer'image(decode_spr_num(e_in.insn)) &
-		when SPR_CTR =>
+		    "=" & to_hstring(e_in.read_data3);
-		    ctrl_tmp.ctr <= e_in.read_data3;
+		if is_fast_spr(e_in.write_reg) then
-		when SPR_LR =>
+		    result := e_in.read_data3;
-		    ctrl_tmp.lr <= e_in.read_data3;
+		    result_en := '1';
-		when others =>
+		    if decode_spr_num(e_in.insn) = SPR_XER then
-		end case;
+			v.e.xerc.so := e_in.read_data3(63-32);
+			v.e.xerc.ov := e_in.read_data3(63-33);
+			v.e.xerc.ca := e_in.read_data3(63-34);
+			v.e.xerc.ov32 := e_in.read_data3(63-44);
+			v.e.xerc.ca32 := e_in.read_data3(63-45);
+			v.e.write_xerc_enable := '1';
+		    end if;
+		else
+-- TODO: Implement slow SPRs	    
+--		    case decode_spr_num(e_in.insn) is
+--		    when others =>
+--		    end case;
+		end if;
 	    when OP_POPCNTB =>
 		result := ppc_popcntb(e_in.read_data3);
 		result_en := '1';
@@ -334,7 +439,7 @@ begin
 	    when OP_RLC | OP_RLCL | OP_RLCR | OP_SHL | OP_SHR =>
 		result := rotator_result;
 		if e_in.output_carry = '1' then
-		    ctrl_tmp.carry <= rotator_carry;
+		    set_carry(v.e, rotator_carry, rotator_carry);
 		end if;
 		result_en := '1';
 	    when OP_SIM_CONFIG =>
@@ -359,15 +464,36 @@ begin
 		report "illegal";
 	    end case;
 
+	    -- Update LR on the next cycle after a branch link
+	    --
+	    -- WARNING: The LR update isn't tracked by our hazard tracker. This
+	    --          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
-		ctrl_tmp.lr <= next_nia;
+		v.lr_update := '1';
+		v.next_lr := next_nia;
+		v.e.valid := '0';
+		report "Delayed LR update to " & to_hstring(next_nia);
+		stall_out <= '1';
 	    end if;
-
+	elsif r.lr_update = '1' then
+	    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';
 	end if;
 
 	v.e.write_data := result;
 	v.e.write_enable := result_en;
-	v.e.rc := e_in.rc;
+	v.e.rc := e_in.rc and e_in.valid;
 
 	-- Update registers
 	rin <= v;

gpr_hazard.vhdl

@@ -7,12 +7,13 @@ entity gpr_hazard is
         PIPELINE_DEPTH : natural := 2
         );
     port(
-        clk                : in std_logic;
+        clk                : in std_ulogic;
+	stall_in           : in std_ulogic;
 
         gpr_write_valid_in : in std_ulogic;
-        gpr_write_in       : in std_ulogic_vector(4 downto 0);
+        gpr_write_in       : in std_ulogic_vector(5 downto 0);
         gpr_read_valid_in  : in std_ulogic;
-        gpr_read_in        : in std_ulogic_vector(4 downto 0);
+        gpr_read_in        : in std_ulogic_vector(5 downto 0);
 
         stall_out          : out std_ulogic
         );
@@ -20,7 +21,7 @@ end entity gpr_hazard;
 architecture behaviour of gpr_hazard is
     type pipeline_entry_type is record
         valid : std_ulogic;
-        gpr   : std_ulogic_vector(4 downto 0);
+        gpr   : std_ulogic_vector(5 downto 0);
     end record;
     constant pipeline_entry_init : pipeline_entry_type := (valid => '0', gpr => (others => '0'));
 
@@ -32,7 +33,9 @@ begin
     gpr_hazard0: process(clk)
     begin
         if rising_edge(clk) then
-            r <= rin;
+	    if stall_in = '0' then
+		r <= rin;
+	    end if;
         end if;
     end process;
 

helpers.vhdl

@@ -17,8 +17,8 @@ package helpers is
 
     function cmp_one_byte(a, b: std_ulogic_vector(7 downto 0)) return std_ulogic_vector;
 
-    function ppc_signed_compare(a, b: signed(63 downto 0)) return std_ulogic_vector;
+    function ppc_signed_compare(a, b: signed(63 downto 0); so: std_ulogic) return std_ulogic_vector;
-    function ppc_unsigned_compare(a, b: unsigned(63 downto 0)) return std_ulogic_vector;
+    function ppc_unsigned_compare(a, b: unsigned(63 downto 0); so: std_ulogic) return std_ulogic_vector;
 
     function ra_or_zero(ra: std_ulogic_vector(63 downto 0); reg: std_ulogic_vector(4 downto 0)) return std_ulogic_vector;
 
@@ -126,32 +126,32 @@ package body helpers is
         return ret;
     end;
 
-    function ppc_signed_compare(a, b: signed(63 downto 0)) return std_ulogic_vector is
+    function ppc_signed_compare(a, b: signed(63 downto 0); so: std_ulogic) return std_ulogic_vector is
-        variable ret: std_ulogic_vector(3 downto 0);
+        variable ret: std_ulogic_vector(2 downto 0);
     begin
         if a < b then
-            ret := "1000";
+            ret := "100";
         elsif a > b then
-            ret := "0100";
+            ret := "010";
         else
-            ret := "0010";
+            ret := "001";
         end if;
 
-        return ret;
+        return ret & so;
     end;
 
-    function ppc_unsigned_compare(a, b: unsigned(63 downto 0)) return std_ulogic_vector is
+    function ppc_unsigned_compare(a, b: unsigned(63 downto 0); so: std_ulogic) return std_ulogic_vector is
-        variable ret: std_ulogic_vector(3 downto 0);
+        variable ret: std_ulogic_vector(2 downto 0);
     begin
         if a < b then
-            ret := "1000";
+            ret := "100";
         elsif a > b then
-            ret := "0100";
+            ret := "010";
         else
-            ret := "0010";
+            ret := "001";
         end if;
 
-        return ret;
+        return ret & so;
     end;
 
     function ra_or_zero(ra: std_ulogic_vector(63 downto 0); reg: std_ulogic_vector(4 downto 0)) return std_ulogic_vector is

insn_helpers.vhdl

@@ -16,6 +16,7 @@ package insn_helpers is
     function insn_lk (insn_in : std_ulogic_vector) return std_ulogic;
     function insn_aa (insn_in : std_ulogic_vector) return std_ulogic;
     function insn_rc (insn_in : std_ulogic_vector) return std_ulogic;
+    function insn_oe (insn_in : std_ulogic_vector) return std_ulogic;
     function insn_bd (insn_in : std_ulogic_vector) return std_ulogic_vector;
     function insn_bf (insn_in : std_ulogic_vector) return std_ulogic_vector;
     function insn_bfa (insn_in : std_ulogic_vector) return std_ulogic_vector;
@@ -103,6 +104,11 @@ package body insn_helpers is
         return insn_in(0);
     end;
 
+    function insn_oe (insn_in : std_ulogic_vector) return std_ulogic is
+    begin
+        return insn_in(10);
+    end;
+
     function insn_bd (insn_in : std_ulogic_vector) return std_ulogic_vector is
     begin
         return insn_in(15 downto 2);

loadstore1.vhdl

@@ -47,6 +47,7 @@ begin
         v.sign_extend := l_in.sign_extend;
         v.update := l_in.update;
         v.update_reg := l_in.update_reg;
+	v.xerc := l_in.xerc;
 
 	-- XXX Temporary hack. Mark the op as non-cachable if the address
 	-- is the form 0xc-------

multiply.vhdl

@@ -27,8 +27,17 @@ architecture behaviour of multiply is
         data      : signed(129 downto 0);
         write_reg : std_ulogic_vector(4 downto 0);
         rc        : std_ulogic;
+	oe        : std_ulogic;
+	is_32bit  : std_ulogic;
+	xerc      : xer_common_t;
     end record;
-    constant MultiplyPipelineStageInit : multiply_pipeline_stage := (valid => '0', insn_type => OP_ILLEGAL, rc => '0', data => (others => '0'), others => (others => '0'));
+    constant MultiplyPipelineStageInit : multiply_pipeline_stage := (valid => '0',
+								     insn_type => OP_ILLEGAL,
+								     rc => '0', oe => '0',
+								     is_32bit => '0',
+								     xerc => xerc_init,
+								     data => (others => '0'),
+								     others => (others => '0'));
 
     type multiply_pipeline_type is array(0 to PIPELINE_DEPTH-1) of multiply_pipeline_stage;
     constant MultiplyPipelineInit : multiply_pipeline_type := (others => MultiplyPipelineStageInit);
@@ -51,6 +60,7 @@ begin
         variable v : reg_type;
         variable d : std_ulogic_vector(129 downto 0);
         variable d2 : std_ulogic_vector(63 downto 0);
+	variable ov : std_ulogic;
     begin
         v := r;
 
@@ -61,16 +71,26 @@ begin
         v.multiply_pipeline(0).data := signed(m.data1) * signed(m.data2);
         v.multiply_pipeline(0).write_reg := m.write_reg;
         v.multiply_pipeline(0).rc := m.rc;
+        v.multiply_pipeline(0).oe := m.oe;
+        v.multiply_pipeline(0).is_32bit := m.is_32bit;
+        v.multiply_pipeline(0).xerc := m.xerc;
 
         loop_0: for i in 1 to PIPELINE_DEPTH-1 loop
             v.multiply_pipeline(i) := r.multiply_pipeline(i-1);
         end loop;
 
         d := std_ulogic_vector(v.multiply_pipeline(PIPELINE_DEPTH-1).data);
+	ov := '0';
 
+	-- TODO: Handle overflows
         case_0: case v.multiply_pipeline(PIPELINE_DEPTH-1).insn_type is
             when OP_MUL_L64 =>
                 d2 := d(63 downto 0);
+		if v.multiply_pipeline(PIPELINE_DEPTH-1).is_32bit = '1' then
+		    ov := (or d(63 downto 31)) and not (and d(63 downto 31));
+		else
+		    ov := (or d(127 downto 63)) and not (and d(127 downto 63));
+		end if;
             when OP_MUL_H32 =>
                 d2 := d(63 downto 32) & d(63 downto 32);
             when OP_MUL_H64 =>
@@ -82,11 +102,24 @@ begin
 
         m_out.write_reg_data <= d2;
         m_out.write_reg_nr <= v.multiply_pipeline(PIPELINE_DEPTH-1).write_reg;
+	m_out.xerc <= v.multiply_pipeline(PIPELINE_DEPTH-1).xerc;
 
+	-- Generate OV/OV32/SO when OE=1
         if v.multiply_pipeline(PIPELINE_DEPTH-1).valid = '1' then
             m_out.valid <= '1';
             m_out.write_reg_enable <= '1';
             m_out.rc <= v.multiply_pipeline(PIPELINE_DEPTH-1).rc;
+            m_out.write_xerc_enable <= v.multiply_pipeline(PIPELINE_DEPTH-1).oe;
+
+	    -- We must test oe because the RC update code in writeback
+	    -- will use the xerc value to set CR0:SO so we must not clobber
+	    -- xerc if OE wasn't set.
+	    --
+	    if v.multiply_pipeline(PIPELINE_DEPTH-1).oe = '1' then
+		m_out.xerc.ov <= ov;
+		m_out.xerc.ov32 <= ov;
+		m_out.xerc.so <= v.multiply_pipeline(PIPELINE_DEPTH-1).xerc.so or ov;
+	    end if;
         end if;
 
         rin <= v;

ppc_fx_insns.vhdl

@@ -77,10 +77,14 @@ package ppc_fx_insns is
 	function ppc_mulhw (ra, rb: std_ulogic_vector(63 downto 0)) return std_ulogic_vector;
 	function ppc_mulhwu (ra, rb: std_ulogic_vector(63 downto 0)) return std_ulogic_vector;
 
-	function ppc_cmpi (l: std_ulogic; ra: std_ulogic_vector(63 downto 0); si: std_ulogic_vector(15 downto 0)) return std_ulogic_vector;
+	function ppc_cmpi (l: std_ulogic; ra: std_ulogic_vector(63 downto 0); si: std_ulogic_vector(15 downto 0);
-	function ppc_cmp (l: std_ulogic; ra, rb: std_ulogic_vector(63 downto 0)) return std_ulogic_vector;
+                           so: std_ulogic) return std_ulogic_vector;
-	function ppc_cmpli (l: std_ulogic; ra: std_ulogic_vector(63 downto 0); si: std_ulogic_vector(15 downto 0)) return std_ulogic_vector;
+	function ppc_cmp (l: std_ulogic; ra, rb: std_ulogic_vector(63 downto 0);
-	function ppc_cmpl (l: std_ulogic; ra, rb: std_ulogic_vector(63 downto 0)) return std_ulogic_vector;
+                           so: std_ulogic) return std_ulogic_vector;
+	function ppc_cmpli (l: std_ulogic; ra: std_ulogic_vector(63 downto 0); si: std_ulogic_vector(15 downto 0);
+                           so: std_ulogic) return std_ulogic_vector;
+	function ppc_cmpl (l: std_ulogic; ra, rb: std_ulogic_vector(63 downto 0);
+                           so: std_ulogic) return std_ulogic_vector;
 
 	function ppc_cmpb (rs, rb: std_ulogic_vector(63 downto 0)) return std_ulogic_vector;
 
@@ -90,7 +94,6 @@ package ppc_fx_insns is
 	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_bcctr_taken(bo, bi: std_ulogic_vector(4 downto 0); cr: std_ulogic_vector(31 downto 0)) return integer;
 end package ppc_fx_insns;
 
 package body ppc_fx_insns is
@@ -677,7 +680,8 @@ package body ppc_fx_insns is
 		return std_ulogic_vector(tmp(63 downto 32)) & std_ulogic_vector(tmp(63 downto 32));
 	end;
 
-	function ppc_cmpi (l: std_ulogic; ra: std_ulogic_vector(63 downto 0); si: std_ulogic_vector(15 downto 0)) return std_ulogic_vector is
+	function ppc_cmpi (l: std_ulogic; ra: std_ulogic_vector(63 downto 0); si: std_ulogic_vector(15 downto 0);
+                           so: std_ulogic) return std_ulogic_vector is
 		variable tmp: signed(ra'range);
 	begin
 		tmp := signed(ra);
@@ -685,10 +689,11 @@ package body ppc_fx_insns is
 			tmp := resize(signed(ra(31 downto 0)), tmp'length);
 		end if;
 
-		return ppc_signed_compare(tmp, resize(signed(si), tmp'length));
+		return ppc_signed_compare(tmp, resize(signed(si), tmp'length), so);
 	end;
 
-	function ppc_cmp (l: std_ulogic; ra, rb: std_ulogic_vector(63 downto 0)) return std_ulogic_vector is
+	function ppc_cmp (l: std_ulogic; ra, rb: std_ulogic_vector(63 downto 0);
+                          so: std_ulogic) return std_ulogic_vector is
 		variable tmpa, tmpb: signed(ra'range);
 	begin
 		tmpa := signed(ra);
@@ -698,10 +703,11 @@ package body ppc_fx_insns is
 			tmpb := resize(signed(rb(31 downto 0)), ra'length);
 		end if;
 
-		return ppc_signed_compare(tmpa, tmpb);
+		return ppc_signed_compare(tmpa, tmpb, so);
 	end;
 
-	function ppc_cmpli (l: std_ulogic; ra: std_ulogic_vector(63 downto 0); si: std_ulogic_vector(15 downto 0)) return std_ulogic_vector is
+	function ppc_cmpli (l: std_ulogic; ra: std_ulogic_vector(63 downto 0); si: std_ulogic_vector(15 downto 0);
+                            so: std_ulogic) return std_ulogic_vector is
 		variable tmp: unsigned(ra'range);
 	begin
 		tmp := unsigned(ra);
@@ -709,10 +715,11 @@ package body ppc_fx_insns is
 			tmp := resize(unsigned(ra(31 downto 0)), tmp'length);
 		end if;
 
-		return ppc_unsigned_compare(tmp, resize(unsigned(si), tmp'length));
+		return ppc_unsigned_compare(tmp, resize(unsigned(si), tmp'length), so);
 	end;
 
-	function ppc_cmpl (l: std_ulogic; ra, rb: std_ulogic_vector(63 downto 0)) return std_ulogic_vector is
+	function ppc_cmpl (l: std_ulogic; ra, rb: std_ulogic_vector(63 downto 0);
+                           so: std_ulogic) return std_ulogic_vector is
 		variable tmpa, tmpb: unsigned(ra'range);
 	begin
 		tmpa := unsigned(ra);
@@ -722,7 +729,7 @@ package body ppc_fx_insns is
 			tmpb := resize(unsigned(rb(31 downto 0)), ra'length);
 		end if;
 
-		return ppc_unsigned_compare(tmpa, tmpb);
+		return ppc_unsigned_compare(tmpa, tmpb, so);
 	end;
 
 	function ppc_cmpb (rs, rb: std_ulogic_vector(63 downto 0)) return std_ulogic_vector is
@@ -801,21 +808,4 @@ package body ppc_fx_insns is
 		return ret;
 	end;
 
-	function ppc_bcctr_taken(bo, bi: std_ulogic_vector(4 downto 0); cr: std_ulogic_vector(31 downto 0)) return integer is
-		variable crfield: integer;
-		variable crbit_match: std_ulogic;
-		variable cond_ok: std_ulogic;
-		variable ret: integer;
-	begin
-		crfield := to_integer(unsigned(bi));
-		-- BE bit numbering
-		crbit_match := '1' when cr(31-crfield) = bo(4-1) else '0';
-		cond_ok := bo(4-0) or crbit_match;
-		if cond_ok = '1' then
-			ret := 1;
-		else
-			ret := 0;
-		end if;
-		return ret;
-	end;
 end package body ppc_fx_insns;

register_file.vhdl

@@ -23,7 +23,7 @@ entity register_file is
 end entity register_file;
 
 architecture behaviour of register_file is
-    type regfile is array(0 to 31) of std_ulogic_vector(63 downto 0);
+    type regfile is array(0 to 63) of std_ulogic_vector(63 downto 0);
     signal registers : regfile := (others => (others => '0'));
 begin
     -- synchronous writes
@@ -32,7 +32,11 @@ begin
         if rising_edge(clk) then
             if w_in.write_enable = '1' then
                 assert not(is_x(w_in.write_data)) and not(is_x(w_in.write_reg)) severity failure;
-                report "Writing GPR " & to_hstring(w_in.write_reg) & " " & to_hstring(w_in.write_data);
+		if w_in.write_reg(5) = '0' then
+		    report "Writing GPR " & to_hstring(w_in.write_reg) & " " & to_hstring(w_in.write_data);
+		else
+		    report "Writing GSPR " & to_hstring(w_in.write_reg) & " " & to_hstring(w_in.write_data);
+		end if;
                 registers(to_integer(unsigned(w_in.write_reg))) <= w_in.write_data;
             end if;
         end if;
@@ -52,7 +56,7 @@ begin
         end if;
         d_out.read1_data <= registers(to_integer(unsigned(d_in.read1_reg)));
         d_out.read2_data <= registers(to_integer(unsigned(d_in.read2_reg)));
-        d_out.read3_data <= registers(to_integer(unsigned(d_in.read3_reg)));
+        d_out.read3_data <= registers(to_integer(unsigned(gpr_to_gspr(d_in.read3_reg))));
 
         -- Forward any written data
         if w_in.write_enable = '1' then
@@ -62,7 +66,7 @@ begin
             if d_in.read2_reg = w_in.write_reg then
                 d_out.read2_data <= w_in.write_data;
             end if;
-            if d_in.read3_reg = w_in.write_reg then
+            if gpr_to_gspr(d_in.read3_reg) = w_in.write_reg then
                 d_out.read3_data <= w_in.write_data;
             end if;
         end if;

writeback.vhdl

@@ -62,6 +62,8 @@ begin
         variable w : std_ulogic_vector(0 downto 0);
         variable j : integer;
         variable k : unsigned(3 downto 0);
+	variable cf: std_ulogic_vector(3 downto 0);
+	variable xe: xer_common_t;
     begin
         x := "" & e_in.valid;
         y := "" & l_in.valid;
@@ -81,6 +83,11 @@ begin
         z := "" & (d_in.valid and d_in.rc);
         assert (to_integer(unsigned(w)) + to_integer(unsigned(x)) + to_integer(unsigned(y)) + to_integer(unsigned(z))) <= 1 severity failure;
 
+        x := "" & e_in.write_xerc_enable;
+        y := "" & m_in.write_xerc_enable;
+        z := "" & D_in.write_xerc_enable;
+        assert (to_integer(unsigned(x)) + to_integer(unsigned(y)) + to_integer(unsigned(z))) <= 1 severity failure;
+
         w_out <= WritebackToRegisterFileInit;
         c_out <= WritebackToCrFileInit;
 
@@ -96,12 +103,12 @@ begin
         partial_write <= '0';
         sign_extend <= '0';
         second_word <= '0';
-        data_in <= e_in.write_data;
+	xe := e_in.xerc;
 
         if e_in.write_enable = '1' then
             w_out.write_reg <= e_in.write_reg;
-            data_in <= e_in.write_data;
             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;
@@ -113,8 +120,13 @@ begin
             c_out.write_cr_data <= e_in.write_cr_data;
         end if;
 
-        if l_in.write_enable = '1' then
+	if e_in.write_xerc_enable = '1' then
-            w_out.write_reg <= l_in.write_reg;
+            c_out.write_xerc_enable <= '1';
+            c_out.write_xerc_data <= e_in.xerc;
+	end if;
+
+	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);
@@ -127,22 +139,35 @@ begin
             if l_in.valid = '0' and (data_len + byte_offset > 8) then
                 partial_write <= '1';
             end if;
+	    xe := l_in.xerc;
         end if;
 
         if m_in.write_reg_enable = '1' then
             w_out.write_enable <= '1';
-            w_out.write_reg <= m_in.write_reg_nr;
+            w_out.write_reg <= gpr_to_gspr(m_in.write_reg_nr);
             data_in <= m_in.write_reg_data;
             rc <= m_in.rc;
+	    xe := m_in.xerc;
         end if;
 
+	if m_in.write_xerc_enable = '1' then
+            c_out.write_xerc_enable <= '1';
+            c_out.write_xerc_data <= m_in.xerc;
+	end if;
+
         if d_in.write_reg_enable = '1' then
             w_out.write_enable <= '1';
-            w_out.write_reg <= d_in.write_reg_nr;
+            w_out.write_reg <= gpr_to_gspr(d_in.write_reg_nr);
             data_in <= d_in.write_reg_data;
             rc <= d_in.rc;
+	    xe := d_in.xerc;
         end if;
 
+	if d_in.write_xerc_enable = '1' then
+            c_out.write_xerc_enable <= '1';
+            c_out.write_xerc_data <= d_in.xerc;
+	end if;
+
         -- shift and byte-reverse data bytes
         for i in 0 to 7 loop
             k := ('0' & (to_unsigned(i, 3) xor brev_lenm1)) + ('0' & byte_offset);
@@ -193,17 +218,15 @@ begin
         -- deliver to regfile
         w_out.write_data <= data_trimmed;
 
-        -- test value against 0 and set CR0 if requested
+        -- Perform CR0 update for RC forms
         if rc = '1' then
             c_out.write_cr_enable <= '1';
             c_out.write_cr_mask <= num_to_fxm(0);
-            if negative = '1' then
+	    cf(3) := negative;
-                c_out.write_cr_data <= x"80000000";
+	    cf(2) := not negative and not zero;
-            elsif zero = '0' then
+	    cf(1) := zero;
-                c_out.write_cr_data <= x"40000000";
+	    cf(0) := xe.so;
-            else
+	    c_out.write_cr_data(31 downto 28) <= cf;
-                c_out.write_cr_data <= x"20000000";
-            end if;
         end if;
     end process;
 end;