core: Add support for floating-point loads and stores

This extends the register file so it can hold FPR values, and
implements the FP loads and stores that do not require conversion
between single and double precision.

We now have the FP, FE0 and FE1 bits in MSR.  FP loads and stores
cause a FP unavailable interrupt if MSR[FP] = 0.

The FPU facilities are optional and their presence is controlled by
the HAS_FPU generic passed down from the top-level board file.  It
defaults to true for all except the A7-35 boards.

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

common.vhdl

@@ -13,8 +13,11 @@ package common is
     constant MSR_SF  : integer := (63 - 0);     -- Sixty-Four bit mode
     constant MSR_EE  : integer := (63 - 48);    -- External interrupt Enable
     constant MSR_PR  : integer := (63 - 49);    -- PRoblem state
+    constant MSR_FP  : integer := (63 - 50);    -- Floating Point available
+    constant MSR_FE0 : integer := (63 - 52);    -- Floating Exception mode
     constant MSR_SE  : integer := (63 - 53);    -- Single-step bit of TE field
     constant MSR_BE  : integer := (63 - 54);    -- Branch trace bit of TE field
+    constant MSR_FE1 : integer := (63 - 55);    -- Floating Exception mode
     constant MSR_IR  : integer := (63 - 58);    -- Instruction Relocation
     constant MSR_DR  : integer := (63 - 59);    -- Data Relocation
     constant MSR_RI  : integer := (63 - 62);    -- Recoverable Interrupt
@@ -53,8 +56,11 @@ package common is
     -- 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)
+    -- Extended GPR index (can hold an SPR or a FPR)
-    subtype gspr_index_t is std_ulogic_vector(5 downto 0);
+    subtype gspr_index_t is std_ulogic_vector(6 downto 0);
+
+    -- FPR indices
+    subtype fpr_index_t is std_ulogic_vector(4 downto 0);
 
     -- Some SPRs are stored in the register file, they use the magic
     -- GPR numbers above 31.
@@ -64,6 +70,9 @@ package common is
     -- indicates if this is indeed a fast SPR. If clear, then
     -- the SPR is not stored in the GPR file.
     --
+    -- FPRs are also stored in the register file, using GSPR
+    -- numbers from 64 to 95.
+    --
     function fast_spr_num(spr: spr_num_t) return gspr_index_t;
 
     -- Indices conversion functions
@@ -71,6 +80,7 @@ package common is
     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;
+    function fpr_to_gspr(f: fpr_index_t) return gspr_index_t;
 
     -- 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
@@ -226,7 +236,7 @@ package common is
 	read2_enable : std_ulogic;
 	read2_reg : gspr_index_t;
 	read3_enable : std_ulogic;
-	read3_reg : gpr_index_t;
+	read3_reg : gspr_index_t;
     end record;
 
     type RegisterFileToDecode2Type is record
@@ -264,7 +274,7 @@ 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 : gpr_index_t;
+	write_reg : gspr_index_t;
 	length : std_ulogic_vector(3 downto 0);
         ci : std_ulogic;                                -- cache-inhibited load/store
 	byte_reverse : std_ulogic;
@@ -282,7 +292,8 @@ package common is
                                                                      sign_extend => '0', update => '0', xerc => xerc_init,
                                                                      reserve => '0', rc => '0', virt_mode => '0', priv_mode => '0',
                                                                      nia => (others => '0'), insn => (others => '0'),
-                                                                     addr1 => (others => '0'), addr2 => (others => '0'), data => (others => '0'), length => (others => '0'),
+                                                                     addr1 => (others => '0'), addr2 => (others => '0'), data => (others => '0'),
+                                                                     write_reg => (others => '0'), length => (others => '0'),
                                                                      mode_32bit => '0', others => (others => '0'));
 
     type Loadstore1ToExecute1Type is record
@@ -369,7 +380,7 @@ package common is
     type Loadstore1ToWritebackType is record
 	valid : std_ulogic;
 	write_enable: std_ulogic;
-	write_reg : gpr_index_t;
+	write_reg : gspr_index_t;
 	write_data : std_ulogic_vector(63 downto 0);
 	xerc : xer_common_t;
         rc : std_ulogic;
@@ -473,10 +484,10 @@ package body common is
            n := 13;
        when others =>
            n := 0;
-           return "000000";
+           return "0000000";
        end case;
        tmp := std_ulogic_vector(to_unsigned(n, 5));
-       return "1" & tmp;
+       return "01" & tmp;
     end;
 
     function gspr_to_gpr(i: gspr_index_t) return gpr_index_t is
@@ -486,7 +497,7 @@ package body common is
 
     function gpr_to_gspr(i: gpr_index_t) return gspr_index_t is
     begin
-	return "0" & i;
+	return "00" & i;
     end;
 
     function gpr_or_spr_to_gspr(g: gpr_index_t; s: gspr_index_t) return gspr_index_t is
@@ -502,4 +513,9 @@ package body common is
     begin
 	return s(5);
     end;
+
+    function fpr_to_gspr(f: fpr_index_t) return gspr_index_t is
+    begin
+        return "10" & f;
+    end;
 end common;

control.vhdl

@@ -34,7 +34,7 @@ entity control is
         gpr_b_read_in       : in gspr_index_t;
 
         gpr_c_read_valid_in : in std_ulogic;
-        gpr_c_read_in       : in gpr_index_t;
+        gpr_c_read_in       : in gspr_index_t;
 
         cr_read_in          : in std_ulogic;
         cr_write_in         : in std_ulogic;
@@ -70,7 +70,6 @@ architecture rtl of control is
     signal gpr_write_valid : std_ulogic := '0';
     signal cr_write_valid  : std_ulogic := '0';
 
-    signal gpr_c_read_in_fmt : std_ulogic_vector(5 downto 0);
 begin
     gpr_hazard0: entity work.gpr_hazard
         generic map (
@@ -122,8 +121,6 @@ begin
             use_bypass         => gpr_bypass_b
             );
 
-    gpr_c_read_in_fmt <= "0" & gpr_c_read_in;
-
     gpr_hazard2: entity work.gpr_hazard
         generic map (
             PIPELINE_DEPTH => PIPELINE_DEPTH
@@ -140,7 +137,7 @@ begin
             gpr_write_in       => gpr_write_in,
             bypass_avail       => gpr_bypassable,
             gpr_read_valid_in  => gpr_c_read_valid_in,
-            gpr_read_in        => gpr_c_read_in_fmt,
+            gpr_read_in        => gpr_c_read_in,
 
             ugpr_write_valid   => update_gpr_write_valid,
             ugpr_write_reg     => update_gpr_write_reg,

core.vhdl

@@ -11,6 +11,7 @@ entity core is
         SIM : boolean := false;
 	DISABLE_FLATTEN : boolean := false;
         EX1_BYPASS : boolean := true;
+        HAS_FPU : boolean := true;
 	ALT_RESET_ADDRESS : std_ulogic_vector(63 downto 0) := (others => '0');
         LOG_LENGTH : natural := 512
         );
@@ -244,6 +245,7 @@ begin
     decode2_0: entity work.decode2
         generic map (
             EX1_BYPASS => EX1_BYPASS,
+            HAS_FPU => HAS_FPU,
             LOG_LENGTH => LOG_LENGTH
             )
         port map (
@@ -267,6 +269,7 @@ begin
     register_file_0: entity work.register_file
         generic map (
             SIM => SIM,
+            HAS_FPU => HAS_FPU,
             LOG_LENGTH => LOG_LENGTH
             )
         port map (
@@ -280,7 +283,7 @@ begin
             dbg_gpr_data => dbg_gpr_data,
 	    sim_dump => terminate,
 	    sim_dump_done => sim_cr_dump,
-            log_out => log_data(255 downto 185)
+            log_out => log_data(255 downto 184)
 	    );
 
     cr_file_0: entity work.cr_file
@@ -294,12 +297,13 @@ begin
             d_out => cr_file_to_decode2,
             w_in => writeback_to_cr_file,
             sim_dump => sim_cr_dump,
-            log_out => log_data(184 downto 172)
+            log_out => log_data(183 downto 171)
             );
 
     execute1_0: entity work.execute1
         generic map (
             EX1_BYPASS => EX1_BYPASS,
+            HAS_FPU => HAS_FPU,
             LOG_LENGTH => LOG_LENGTH
             )
         port map (
@@ -324,6 +328,7 @@ begin
 
     loadstore1_0: entity work.loadstore1
         generic map (
+            HAS_FPU => HAS_FPU,
             LOG_LENGTH => LOG_LENGTH
             )
         port map (
@@ -368,7 +373,7 @@ begin
             stall_out => dcache_stall_out,
             wishbone_in => wishbone_data_in,
             wishbone_out => wishbone_data_out,
-            log_out => log_data(171 downto 152)
+            log_out => log_data(170 downto 151)
             );
 
     writeback_0: entity work.writeback
@@ -381,7 +386,7 @@ begin
             complete_out => complete
             );
 
-    log_data(151 downto 150) <= "00";
+    log_data(150) <= '0';
     log_data(139 downto 135) <= "00000";
 
     debug_0: entity work.core_debug

decode1.vhdl

@@ -72,6 +72,10 @@ architecture behaviour of decode1 is
         10 =>       (ALU,    OP_CMP,       RA,         CONST_UI,    NONE, NONE, '0', '1', '1', '0', ONE,  '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', '0'), -- lbz
         35 =>       (LDST,   OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '1', '0', '0', '0', NONE, '0', '0'), -- lbzu
+        50 =>       (LDST,   OP_FPLOAD,    RA_OR_ZERO, CONST_SI,    NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- lfd
+        51 =>       (LDST,   OP_FPLOAD,    RA_OR_ZERO, CONST_SI,    NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0'), -- lfdu
+--      48 =>       (LDST,   OP_FPLOAD,    RA_OR_ZERO, CONST_SI,    NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '1', '0', NONE, '0', '0'), -- lfs
+--      49 =>       (LDST,   OP_FPLOAD,    RA_OR_ZERO, CONST_SI,    NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '1', '0', NONE, '0', '0'), -- lfsu
         42 =>       (LDST,   OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '1', '0', '0', '0', '0', NONE, '0', '0'), -- lha
         43 =>       (LDST,   OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '1', '1', '0', '0', '0', NONE, '0', '0'), -- lhau
         40 =>       (LDST,   OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- lhz
@@ -87,6 +91,10 @@ architecture behaviour of decode1 is
         17 =>       (ALU,    OP_SC,        NONE,       NONE,        NONE, NONE, '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- sc
         38 =>       (LDST,   OP_STORE,     RA_OR_ZERO, CONST_SI,    RS,   NONE, '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- stb
         39 =>       (LDST,   OP_STORE,     RA_OR_ZERO, CONST_SI,    RS,   NONE, '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '1', '0', '0', '0', NONE, '0', '0'), -- stbu
+        54 =>       (LDST,   OP_FPSTORE,   RA_OR_ZERO, CONST_SI,    FRS,  NONE, '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- stfd
+        55 =>       (LDST,   OP_FPSTORE,   RA_OR_ZERO, CONST_SI,    FRS,  NONE, '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0'), -- stfdu
+--      52 =>       (LDST,   OP_FPSTORE,   RA_OR_ZERO, CONST_SI,    FRS,  NONE, '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '1', '0', NONE, '0', '0'), -- stfs
+--      53 =>       (LDST,   OP_FPSTORE,   RA_OR_ZERO, CONST_SI,    FRS,  NONE, '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '1', '0', NONE, '0', '0'), -- stfsu
         44 =>       (LDST,   OP_STORE,     RA_OR_ZERO, CONST_SI,    RS,   NONE, '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- sth
         45 =>       (LDST,   OP_STORE,     RA_OR_ZERO, CONST_SI,    RS,   NONE, '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '1', '0', '0', '0', NONE, '0', '0'), -- sthu
         36 =>       (LDST,   OP_STORE,     RA_OR_ZERO, CONST_SI,    RS,   NONE, '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- stw
@@ -272,6 +280,12 @@ architecture behaviour of decode1 is
         2#1101110101#  =>       (LDST,   OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- ldcix
         2#0000110101#  =>       (LDST,   OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0'), -- ldux
         2#0000010101#  =>       (LDST,   OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- ldx
+        2#1001010111#  =>       (LDST,   OP_FPLOAD,    RA_OR_ZERO, RB,          NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- lfdx
+        2#1001110111#  =>       (LDST,   OP_FPLOAD,    RA_OR_ZERO, RB,          NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0'), -- lfdux
+        2#1101010111#  =>       (LDST,   OP_FPLOAD,    RA_OR_ZERO, RB,          NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is4B, '0', '1', '0', '0', '0', '0', NONE, '0', '0'), -- lfiwax
+        2#1101110111#  =>       (LDST,   OP_FPLOAD,    RA_OR_ZERO, RB,          NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- lfiwzx
+--      2#1000010111#  =>       (LDST,   OP_FPLOAD,    RA_OR_ZERO, RB,          NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '1', '0', NONE, '0', '0'), -- lfsx
+--      2#1000110111#  =>       (LDST,   OP_FPLOAD,    RA_OR_ZERO, RB,          NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '1', '0', NONE, '0', '0'), -- lfsux
         2#0001110100#  =>       (LDST,   OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '0', '1', '0', '0', NONE, '0', '0'), -- lharx
         2#0101110111#  =>       (LDST,   OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '1', '1', '0', '0', '0', NONE, '0', '0'), -- lhaux
         2#0101010111#  =>       (LDST,   OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '1', '0', '0', '0', '0', NONE, '0', '0'), -- lhax
@@ -350,6 +364,11 @@ architecture behaviour of decode1 is
         2#0011010110#  =>       (LDST,   OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '1', '0', '0', ONE,  '0', '0'), -- stdcx
         2#0010110101#  =>       (LDST,   OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0'), -- stdux
         2#0010010101#  =>       (LDST,   OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- stdx
+        2#1011010111#  =>       (LDST,   OP_FPSTORE,   RA_OR_ZERO, RB,          FRS,  NONE, '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- stfdx
+        2#1011110111#  =>       (LDST,   OP_FPSTORE,   RA_OR_ZERO, RB,          FRS,  NONE, '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0'), -- stfdux
+        2#1111010111#  =>       (LDST,   OP_FPSTORE,   RA_OR_ZERO, RB,          FRS,  NONE, '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- stfiwx
+--      2#1010010111#  =>       (LDST,   OP_FPSTORE,   RA_OR_ZERO, RB,          FRS,  NONE, '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '1', '0', NONE, '0', '0'), -- stfsx
+--      2#1010110111#  =>       (LDST,   OP_FPSTORE,   RA_OR_ZERO, RB,          FRS,  NONE, '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '1', '0', NONE, '0', '0'), -- stfsux
         2#1110010110#  =>       (LDST,   OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is2B, '1', '0', '0', '0', '0', '0', NONE, '0', '0'), -- sthbrx
         2#1110110101#  =>       (LDST,   OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '0', '0', '0', '0', NONE, '0', '0'), -- sthcix
         2#1011010110#  =>       (LDST,   OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '0', '1', '0', '0', ONE,  '0', '0'), -- sthcx

decode2.vhdl

@@ -11,6 +11,7 @@ use work.insn_helpers.all;
 entity decode2 is
     generic (
         EX1_BYPASS : boolean := true;
+        HAS_FPU : boolean := true;
         -- Non-zero to enable log data collection
         LOG_LENGTH : natural := 0
         );
@@ -73,7 +74,7 @@ architecture behaviour of decode2 is
             -- 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"
+            assert is_fast_spr(ispr) =  '1' or ispr = "0000000"
                 report "Decode A says SPR but ISPR is invalid:" &
                 to_hstring(ispr) severity failure;
             return (is_fast_spr(ispr), ispr, reg_data);
@@ -118,7 +119,7 @@ architecture behaviour of decode2 is
                 -- 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"
+                assert is_fast_spr(ispr) = '1' or ispr = "0000000"
                     report "Decode B says SPR but ISPR is invalid:" &
                     to_hstring(ispr) severity failure;
                 ret := (is_fast_spr(ispr), ispr, reg_data);
@@ -137,6 +138,12 @@ architecture behaviour of decode2 is
                 return ('1', gpr_to_gspr(insn_rs(insn_in)), reg_data);
             when RCR =>
                 return ('1', gpr_to_gspr(insn_rcreg(insn_in)), reg_data);
+            when FRS =>
+                if HAS_FPU then
+                    return ('1', fpr_to_gspr(insn_frt(insn_in)), reg_data);
+                else
+                    return ('0', (others => '0'), (others => '0'));
+                end if;
             when NONE =>
                 return ('0', (others => '0'), (others => '0'));
         end case;
@@ -150,16 +157,22 @@ architecture behaviour of decode2 is
                 return ('1', gpr_to_gspr(insn_rt(insn_in)));
             when RA =>
                 return ('1', gpr_to_gspr(insn_ra(insn_in)));
+            when FRT =>
+                if HAS_FPU then
+                    return ('1', fpr_to_gspr(insn_frt(insn_in)));
+                else
+                    return ('0', "0000000");
+                end if;
             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"
+                assert is_fast_spr(ispr) = '1' or ispr = "0000000"
                     report "Decode B says SPR but ISPR is invalid:" &
                     to_hstring(ispr) severity failure;
                 return (is_fast_spr(ispr), ispr);
             when NONE =>
-                return ('0', "000000");
+                return ('0', "0000000");
         end case;
     end;
 
@@ -212,7 +225,7 @@ architecture behaviour of decode2 is
     signal gpr_b_bypass : std_ulogic;
 
     signal gpr_c_read_valid : std_ulogic;
-    signal gpr_c_read : gpr_index_t;
+    signal gpr_c_read : gspr_index_t;
     signal gpr_c_bypass : std_ulogic;
 
     signal cr_write_valid  : std_ulogic;
@@ -284,8 +297,9 @@ begin
                        else gpr_to_gspr(insn_ra(d_in.insn));
     r_out.read2_reg <= d_in.ispr2 when d_in.decode.input_reg_b = SPR
                        else gpr_to_gspr(insn_rb(d_in.insn));
-    r_out.read3_reg <= insn_rcreg(d_in.insn) when d_in.decode.input_reg_c = RCR
+    r_out.read3_reg <= gpr_to_gspr(insn_rcreg(d_in.insn)) when d_in.decode.input_reg_c = RCR
-                       else insn_rs(d_in.insn);
+                       else fpr_to_gspr(insn_frt(d_in.insn)) when d_in.decode.input_reg_c = FRS and HAS_FPU
+                       else gpr_to_gspr(insn_rs(d_in.insn));
 
     c_out.read <= d_in.decode.input_cr;
 
@@ -394,7 +408,7 @@ begin
         gpr_b_read <= decoded_reg_b.reg;
 
         gpr_c_read_valid <= decoded_reg_c.reg_valid;
-        gpr_c_read <= gspr_to_gpr(decoded_reg_c.reg);
+        gpr_c_read <= decoded_reg_c.reg;
 
         cr_write_valid <= d_in.decode.output_cr or decode_rc(d_in.decode.rc, d_in.insn);
         cr_bypass_avail <= '0';

decode_types.vhdl

@@ -10,6 +10,7 @@ package decode_types is
 			 OP_DCBZ, OP_DIV, OP_DIVE, OP_EXTS,
 			 OP_EXTSWSLI, OP_ICBI, OP_ICBT, OP_ISEL, OP_ISYNC,
 			 OP_LOAD, OP_STORE,
+                         OP_FPLOAD, OP_FPSTORE,
 			 OP_MCRXRX, OP_MFCR, OP_MFMSR, OP_MFSPR, OP_MOD,
 			 OP_MTCRF, OP_MTMSRD, OP_MTSPR, OP_MUL_L64,
 			 OP_MUL_H64, OP_MUL_H32, OP_OR,
@@ -24,8 +25,8 @@ package decode_types is
     type input_reg_a_t is (NONE, RA, RA_OR_ZERO, SPR, CIA);
     type input_reg_b_t is (NONE, RB, CONST_UI, CONST_SI, CONST_SI_HI, CONST_UI_HI, CONST_LI, CONST_BD,
                            CONST_DXHI4, CONST_DS, CONST_M1, CONST_SH, CONST_SH32, SPR);
-    type input_reg_c_t is (NONE, RS, RCR);
+    type input_reg_c_t is (NONE, RS, RCR, FRS);
-    type output_reg_a_t is (NONE, RT, RA, SPR);
+    type output_reg_a_t is (NONE, RT, RA, SPR, FRT);
     type rc_t is (NONE, ONE, RC);
     type carry_in_t is (ZERO, CA, OV, ONE);
 

execute1.vhdl

@@ -13,6 +13,7 @@ use work.ppc_fx_insns.all;
 entity execute1 is
     generic (
         EX1_BYPASS : boolean := true;
+        HAS_FPU : boolean := true;
         -- Non-zero to enable log data collection
         LOG_LENGTH : natural := 0
         );
@@ -542,6 +543,9 @@ begin
             ctrl_tmp.msr(MSR_PR) <= '0';
             ctrl_tmp.msr(MSR_SE) <= '0';
             ctrl_tmp.msr(MSR_BE) <= '0';
+            ctrl_tmp.msr(MSR_FP) <= '0';
+            ctrl_tmp.msr(MSR_FE0) <= '0';
+            ctrl_tmp.msr(MSR_FE1) <= '0';
             ctrl_tmp.msr(MSR_IR) <= '0';
             ctrl_tmp.msr(MSR_DR) <= '0';
             ctrl_tmp.msr(MSR_RI) <= '0';
@@ -579,6 +583,18 @@ begin
             ctrl_tmp.srr1(63 - 45) <= '1';
             report "privileged instruction";
 
+        elsif not HAS_FPU and valid_in = '1' and
+            (e_in.insn_type = OP_FPLOAD or e_in.insn_type = OP_FPSTORE) then
+            -- make lfd/stfd/lfs/stfs etc. illegal in no-FPU implementations
+            illegal := '1';
+
+        elsif HAS_FPU and valid_in = '1' and ctrl.msr(MSR_FP) = '0' and
+            (e_in.insn_type = OP_FPLOAD or e_in.insn_type = OP_FPSTORE) then
+            -- generate a floating-point unavailable interrupt
+            exception := '1';
+            v.f.redirect_nia := std_logic_vector(to_unsigned(16#800#, 64));
+            report "FP unavailable interrupt";
+
 	elsif valid_in = '1' and e_in.unit = ALU then
 
 	    report "execute nia " & to_hstring(e_in.nia);
@@ -1225,7 +1241,7 @@ begin
         lv.addr1 := a_in;
         lv.addr2 := b_in;
         lv.data := c_in;
-        lv.write_reg := gspr_to_gpr(e_in.write_reg);
+        lv.write_reg := e_in.write_reg;
         lv.length := e_in.data_len;
         lv.byte_reverse := e_in.byte_reverse xnor ctrl.msr(MSR_LE);
         lv.sign_extend := e_in.sign_extend;

fpga/top-arty.vhdl

@@ -14,6 +14,7 @@ entity toplevel is
         RAM_INIT_FILE      : string   := "firmware.hex";
         RESET_LOW          : boolean  := true;
         CLK_FREQUENCY      : positive := 100000000;
+        HAS_FPU            : boolean  := true;
         USE_LITEDRAM       : boolean  := false;
         NO_BRAM            : boolean  := false;
         DISABLE_FLATTEN_CORE : boolean := false;
@@ -168,6 +169,7 @@ begin
             RAM_INIT_FILE      => RAM_INIT_FILE,
             SIM                => false,
             CLK_FREQ           => CLK_FREQUENCY,
+            HAS_FPU            => HAS_FPU,
             HAS_DRAM           => USE_LITEDRAM,
             DRAM_SIZE          => 256 * 1024 * 1024,
             DRAM_INIT_SIZE     => PAYLOAD_SIZE,

fpga/top-generic.vhdl

@@ -11,6 +11,7 @@ entity toplevel is
 	RESET_LOW     : boolean  := true;
 	CLK_INPUT     : positive := 100000000;
 	CLK_FREQUENCY : positive := 100000000;
+        HAS_FPU       : boolean  := true;
 	DISABLE_FLATTEN_CORE : boolean := false;
         UART_IS_16550 : boolean  := true
 	);
@@ -68,6 +69,7 @@ begin
 	    RAM_INIT_FILE => RAM_INIT_FILE,
 	    SIM           => false,
 	    CLK_FREQ      => CLK_FREQUENCY,
+            HAS_FPU       => HAS_FPU,
 	    DISABLE_FLATTEN_CORE => DISABLE_FLATTEN_CORE,
             UART0_IS_16550     => UART_IS_16550
 	    )

fpga/top-nexys-video.vhdl

@@ -14,6 +14,7 @@ entity toplevel is
 	RAM_INIT_FILE : string   := "firmware.hex";
 	RESET_LOW     : boolean  := true;
 	CLK_FREQUENCY : positive := 100000000;
+        HAS_FPU       : boolean  := true;
 	USE_LITEDRAM  : boolean  := false;
 	NO_BRAM       : boolean  := false;
 	DISABLE_FLATTEN_CORE : boolean := false;
@@ -120,6 +121,7 @@ begin
 	    RAM_INIT_FILE => RAM_INIT_FILE,
 	    SIM           => false,
 	    CLK_FREQ      => CLK_FREQUENCY,
+            HAS_FPU       => HAS_FPU,
 	    HAS_DRAM      => USE_LITEDRAM,
 	    DRAM_SIZE     => 512 * 1024 * 1024,
             DRAM_INIT_SIZE => PAYLOAD_SIZE,

gpr_hazard.vhdl

@@ -2,6 +2,9 @@ library ieee;
 use ieee.std_logic_1164.all;
 use ieee.numeric_std.all;
 
+library work;
+use work.common.all;
+
 entity gpr_hazard is
     generic (
         PIPELINE_DEPTH : natural := 1
@@ -15,13 +18,13 @@ entity gpr_hazard is
         issuing            : in std_ulogic;
 
         gpr_write_valid_in : in std_ulogic;
-        gpr_write_in       : in std_ulogic_vector(5 downto 0);
+        gpr_write_in       : in gspr_index_t;
         bypass_avail       : in std_ulogic;
         gpr_read_valid_in  : in std_ulogic;
-        gpr_read_in        : in std_ulogic_vector(5 downto 0);
+        gpr_read_in        : in gspr_index_t;
 
         ugpr_write_valid   : in std_ulogic;
-        ugpr_write_reg     : in std_ulogic_vector(5 downto 0);
+        ugpr_write_reg     : in gspr_index_t;
 
         stall_out          : out std_ulogic;
         use_bypass         : out std_ulogic
@@ -31,9 +34,9 @@ architecture behaviour of gpr_hazard is
     type pipeline_entry_type is record
         valid  : std_ulogic;
         bypass : std_ulogic;
-        gpr    : std_ulogic_vector(5 downto 0);
+        gpr    : gspr_index_t;
         ugpr_valid : std_ulogic;
-        ugpr   : std_ulogic_vector(5 downto 0);
+        ugpr   : gspr_index_t;
     end record;
     constant pipeline_entry_init : pipeline_entry_type := (valid => '0', bypass => '0', gpr => (others => '0'),
                                                            ugpr_valid => '0', ugpr => (others => '0'));

insn_helpers.vhdl

@@ -37,6 +37,10 @@ package insn_helpers is
     function insn_sh (insn_in : std_ulogic_vector) return std_ulogic_vector;
     function insn_me (insn_in : std_ulogic_vector) return std_ulogic_vector;
     function insn_mb (insn_in : std_ulogic_vector) return std_ulogic_vector;
+    function insn_frt (insn_in : std_ulogic_vector) return std_ulogic_vector;
+    function insn_fra (insn_in : std_ulogic_vector) return std_ulogic_vector;
+    function insn_frb (insn_in : std_ulogic_vector) return std_ulogic_vector;
+    function insn_frc (insn_in : std_ulogic_vector) return std_ulogic_vector;
 end package insn_helpers;
 
 package body insn_helpers is
@@ -214,4 +218,24 @@ package body insn_helpers is
     begin
         return insn_in(5) & insn_in(10 downto 6);
     end;
+
+    function insn_frt(insn_in : std_ulogic_vector) return std_ulogic_vector is
+    begin
+        return insn_in(25 downto 21);
+    end;
+
+    function insn_fra(insn_in : std_ulogic_vector) return std_ulogic_vector is
+    begin
+        return insn_in(20 downto 16);
+    end;
+
+    function insn_frb(insn_in : std_ulogic_vector) return std_ulogic_vector is
+    begin
+        return insn_in(15 downto 11);
+    end;
+
+    function insn_frc(insn_in : std_ulogic_vector) return std_ulogic_vector is
+    begin
+        return insn_in(10 downto 6);
+    end;
 end package body insn_helpers;

loadstore1.vhdl

@@ -5,12 +5,15 @@ use ieee.numeric_std.all;
 library work;
 use work.decode_types.all;
 use work.common.all;
+use work.insn_helpers.all;
+use work.helpers.all;
 
 -- 2 cycle LSU
 -- We calculate the address in the first cycle
 
 entity loadstore1 is
     generic (
+        HAS_FPU : boolean := true;
         -- Non-zero to enable log data collection
         LOG_LENGTH : natural := 0
         );
@@ -58,7 +61,7 @@ architecture behave of loadstore1 is
 	addr         : std_ulogic_vector(63 downto 0);
 	store_data   : std_ulogic_vector(63 downto 0);
 	load_data    : std_ulogic_vector(63 downto 0);
-	write_reg    : gpr_index_t;
+	write_reg    : gspr_index_t;
 	length       : std_ulogic_vector(3 downto 0);
 	byte_reverse : std_ulogic;
 	sign_extend  : std_ulogic;
@@ -431,6 +434,17 @@ begin
                     v.align_intr := v.nc;
                     req := '1';
                     v.dcbz := '1';
+                when OP_FPSTORE =>
+                    if HAS_FPU then
+                        req := '1';
+                    end if;
+                when OP_FPLOAD =>
+                    if HAS_FPU then
+                        v.load := '1';
+                        req := '1';
+                        -- Allow an extra cycle for RA update
+                        v.extra_cycle := l_in.update;
+                    end if;
                 when OP_TLBIE =>
                     mmureq := '1';
                     v.tlbie := '1';
@@ -523,7 +537,7 @@ begin
             l_out.write_data <= r.sprval;
         elsif do_update = '1' then
             l_out.write_enable <= '1';
-            l_out.write_reg <= r.update_reg;
+            l_out.write_reg <= gpr_to_gspr(r.update_reg);
             l_out.write_data <= r.addr;
         else
             l_out.write_enable <= write_enable;

microwatt.core

@@ -132,6 +132,7 @@ targets:
       - disable_flatten_core
       - log_length=2048
       - uart_is_16550
+      - has_fpu
     tools:
       vivado: {part : xc7a100tcsg324-1}
     toplevel : toplevel
@@ -215,6 +216,7 @@ targets:
       - spi_flash_offset=10485760
       - log_length=2048
       - uart_is_16550
+      - has_fpu
     tools:
       vivado: {part : xc7a200tsbg484-1}
     toplevel : toplevel
@@ -231,6 +233,7 @@ targets:
       - spi_flash_offset=10485760
       - log_length=2048
       - uart_is_16550
+      - has_fpu
     generate: [litedram_nexys_video]
     tools:
       vivado: {part : xc7a200tsbg484-1}
@@ -249,6 +252,7 @@ targets:
       - log_length=512
       - uart_is_16550
       - has_uart1
+      - has_fpu=false
     tools:
       vivado: {part : xc7a35ticsg324-1L}
     toplevel : toplevel
@@ -267,6 +271,7 @@ targets:
       - log_length=512
       - uart_is_16550
       - has_uart1
+      - has_fpu=false
     generate: [litedram_arty, liteeth_arty]
     tools:
       vivado: {part : xc7a35ticsg324-1L}
@@ -285,6 +290,7 @@ targets:
       - log_length=2048
       - uart_is_16550
       - has_uart1
+      - has_fpu
     tools:
       vivado: {part : xc7a100ticsg324-1L}
     toplevel : toplevel
@@ -303,6 +309,7 @@ targets:
       - log_length=2048
       - uart_is_16550
       - has_uart1
+      - has_fpu
     generate: [litedram_arty, liteeth_arty]
     tools:
       vivado: {part : xc7a100ticsg324-1L}
@@ -320,6 +327,7 @@ targets:
       - disable_flatten_core
       - log_length=512
       - uart_is_16550
+      - has_fpu=false
     tools:
       vivado: {part : xc7a35tcpg236-1}
     toplevel : toplevel
@@ -380,6 +388,12 @@ parameters:
     paramtype   : generic
     default     : 100000000
 
+  has_fpu:
+    datatype    : bool
+    description : Include a floating-point unit in the core
+    paramtype   : generic
+    default     : true
+
   disable_flatten_core:
     datatype    : bool
     description : Prevent Vivado from flattening the main core components

register_file.vhdl

@@ -8,6 +8,7 @@ use work.common.all;
 entity register_file is
     generic (
         SIM : boolean := false;
+        HAS_FPU : boolean := true;
         -- Non-zero to enable log data collection
         LOG_LENGTH : natural := 0
         );
@@ -28,12 +29,12 @@ entity register_file is
         sim_dump      : in std_ulogic;
         sim_dump_done : out std_ulogic;
 
-        log_out       : out std_ulogic_vector(70 downto 0)
+        log_out       : out std_ulogic_vector(71 downto 0)
         );
 end entity register_file;
 
 architecture behaviour of register_file is
-    type regfile is array(0 to 63) of std_ulogic_vector(63 downto 0);
+    type regfile is array(0 to 127) of std_ulogic_vector(63 downto 0);
     signal registers : regfile := (others => (others => '0'));
     signal rd_port_b : std_ulogic_vector(63 downto 0);
     signal dbg_data : std_ulogic_vector(63 downto 0);
@@ -41,53 +42,73 @@ architecture behaviour of register_file is
 begin
     -- synchronous writes
     register_write_0: process(clk)
+        variable w_addr : gspr_index_t;
     begin
         if rising_edge(clk) then
             if w_in.write_enable = '1' then
-		if w_in.write_reg(5) = '0' then
+                w_addr := w_in.write_reg;
-		    report "Writing GPR " & to_hstring(w_in.write_reg) & " " & to_hstring(w_in.write_data);
+                if HAS_FPU and w_addr(6) = '1' then
-		else
+                    report "Writing FPR " & to_hstring(w_addr(4 downto 0)) & " " & to_hstring(w_in.write_data);
-		    report "Writing GSPR " & to_hstring(w_in.write_reg) & " " & to_hstring(w_in.write_data);
+                else
-		end if;
+                    w_addr(6) := '0';
+                    if w_addr(5) = '0' then
+                        report "Writing GPR " & to_hstring(w_addr) & " " & to_hstring(w_in.write_data);
+                    else
+                        report "Writing GSPR " & to_hstring(w_addr) & " " & to_hstring(w_in.write_data);
+                    end if;
+                end if;
                 assert not(is_x(w_in.write_data)) and not(is_x(w_in.write_reg)) severity failure;
-                registers(to_integer(unsigned(w_in.write_reg))) <= w_in.write_data;
+                registers(to_integer(unsigned(w_addr))) <= w_in.write_data;
             end if;
         end if;
     end process register_write_0;
 
     -- asynchronous reads
     register_read_0: process(all)
-        variable b_addr : gspr_index_t;
+        variable a_addr, b_addr, c_addr : gspr_index_t;
+        variable w_addr : gspr_index_t;
     begin
+        a_addr := d_in.read1_reg;
+        b_addr := d_in.read2_reg;
+        c_addr := d_in.read3_reg;
+        w_addr := w_in.write_reg;
+        if not HAS_FPU then
+            -- Make it obvious that we only want 64 GSPRs for a no-FPU implementation
+            a_addr(6) := '0';
+            b_addr(6) := '0';
+            c_addr(6) := '0';
+            w_addr(6) := '0';
+        end if;
         if d_in.read1_enable = '1' then
-            report "Reading GPR " & to_hstring(d_in.read1_reg) & " " & to_hstring(registers(to_integer(unsigned(d_in.read1_reg))));
+            report "Reading GPR " & to_hstring(a_addr) & " " & to_hstring(registers(to_integer(unsigned(a_addr))));
         end if;
         if d_in.read2_enable = '1' then
-            report "Reading GPR " & to_hstring(d_in.read2_reg) & " " & to_hstring(registers(to_integer(unsigned(d_in.read2_reg))));
+            report "Reading GPR " & to_hstring(b_addr) & " " & to_hstring(registers(to_integer(unsigned(b_addr))));
         end if;
         if d_in.read3_enable = '1' then
-            report "Reading GPR " & to_hstring(d_in.read3_reg) & " " & to_hstring(registers(to_integer(unsigned(d_in.read3_reg))));
+            report "Reading GPR " & to_hstring(c_addr) & " " & to_hstring(registers(to_integer(unsigned(c_addr))));
         end if;
-        d_out.read1_data <= registers(to_integer(unsigned(d_in.read1_reg)));
+        d_out.read1_data <= registers(to_integer(unsigned(a_addr)));
         -- B read port is multiplexed with reads from the debug circuitry
         if d_in.read2_enable = '0' and dbg_gpr_req = '1' and dbg_ack = '0' then
             b_addr := dbg_gpr_addr;
-        else
+            if not HAS_FPU then
-            b_addr := d_in.read2_reg;
+                b_addr(6) := '0';
+            end if;
         end if;
         rd_port_b <= registers(to_integer(unsigned(b_addr)));
         d_out.read2_data <= rd_port_b;
-        d_out.read3_data <= registers(to_integer(unsigned(gpr_to_gspr(d_in.read3_reg))));
+        d_out.read3_data <= registers(to_integer(unsigned(c_addr)));
 
         -- Forward any written data
         if w_in.write_enable = '1' then
-            if d_in.read1_reg = w_in.write_reg then
+            if a_addr = w_addr then
                 d_out.read1_data <= w_in.write_data;
             end if;
-            if d_in.read2_reg = w_in.write_reg then
+            if b_addr = w_addr then
                 d_out.read2_data <= w_in.write_data;
             end if;
-            if gpr_to_gspr(d_in.read3_reg) = w_in.write_reg then
+            if c_addr = w_addr then
                 d_out.read3_data <= w_in.write_data;
             end if;
         end if;
@@ -136,7 +157,7 @@ begin
     end generate;
 
     rf_log: if LOG_LENGTH > 0 generate
-        signal log_data : std_ulogic_vector(70 downto 0);
+        signal log_data : std_ulogic_vector(71 downto 0);
     begin
         reg_log: process(clk)
         begin

scripts/fmt_log/fmt_log.c

@@ -58,7 +58,7 @@ struct log_entry {
 	u64	ls_lo_valid: 1;
 	u64	ls_eo_except: 1;
 	u64	ls_stall_out: 1;
-	u64	pad2: 2;
+	u64	pad2: 1;
 	u64	dc_state: 3;
 	u64	dc_ra_valid: 1;
 	u64	dc_tlb_way: 3;
@@ -74,7 +74,7 @@ struct log_entry {
 	u64	cr_wr_mask: 8;
 	u64	cr_wr_data: 4;
 	u64	cr_wr_enable: 1;
-	u64	reg_wr_reg: 6;
+	u64	reg_wr_reg: 7;
 	u64	reg_wr_enable: 1;
 
 	u64	reg_wr_data;
@@ -90,11 +90,11 @@ const char *ops[64] =
 	"illegal", "nop    ", "add    ", "and    ", "attn   ", "b      ", "bc     ", "bcreg  ",
 	"bperm  ", "cmp    ", "cmpb   ", "cmpeqb ", "cmprb  ", "cntz   ", "crop   ", "darn   ",
 	"dcbf   ", "dcbst  ", "dcbt   ", "dcbtst ", "dcbz   ", "div    ", "dive   ", "exts   ",
-	"extswsl", "icbi   ", "icbt   ", "isel   ", "isync  ", "ld     ", "st     ", "mcrxrx ",
+	"extswsl", "icbi   ", "icbt   ", "isel   ", "isync  ", "ld     ", "st     ", "fpload ",
-	"mfcr   ", "mfmsr  ", "mfspr  ", "mod    ", "mtcrf  ", "mtmsr  ", "mtspr  ", "mull64 ",
+	"fpstore", "mcrxrx ", "mfcr   ", "mfmsr  ", "mfspr  ", "mod    ", "mtcrf  ", "mtmsr  ",
-	"mulh64 ", "mulh32 ", "or     ", "popcnt ", "prty   ", "rfid   ", "rlc    ", "rlcl   ",
+	"mtspr  ", "mull64 ", "mulh64 ", "mulh32 ", "or     ", "popcnt ", "prty   ", "rfid   ",
-	"rlcr   ", "sc     ", "setb   ", "shl    ", "shr    ", "sync   ", "tlbie  ", "trap   ",
+	"rlc    ", "rlcl   ", "rlcr   ", "sc     ", "setb   ", "shl    ", "shr    ", "sync   ",
-	"xor    ", "bcd    ", "addg6s ", "ffail  ", "?60    ", "?61    ", "?62    ", "?63    "
+	"tlbie  ", "trap   ", "xor    ", "bcd    ", "addg6s ", "ffail  ", "?62    ", "?63    "
 };
 
 const char *spr_names[13] =

soc.vhdl

@@ -52,6 +52,7 @@ entity soc is
 	RAM_INIT_FILE      : string;
 	CLK_FREQ           : positive;
 	SIM                : boolean;
+        HAS_FPU            : boolean := true;
 	DISABLE_FLATTEN_CORE : boolean := false;
 	HAS_DRAM           : boolean  := false;
 	DRAM_SIZE          : integer := 0;
@@ -253,6 +254,7 @@ begin
     processor: entity work.core
 	generic map(
 	    SIM => SIM,
+            HAS_FPU => HAS_FPU,
 	    DISABLE_FLATTEN => DISABLE_FLATTEN_CORE,
 	    ALT_RESET_ADDRESS => (23 downto 0 => '0', others => '1'),
             LOG_LENGTH => LOG_LENGTH

writeback.vhdl

@@ -80,7 +80,7 @@ begin
             end if;
 
             if l_in.write_enable = '1' then
-                w_out.write_reg <= gpr_to_gspr(l_in.write_reg);
+                w_out.write_reg <= l_in.write_reg;
                 w_out.write_data <= l_in.write_data;
                 w_out.write_enable <= '1';
             end if;