fetch/icache: Fit icache in BRAM

The goal is to have the icache fit in BRAM by latching the output
into a register. In order to avoid timing issues , we need to give
the BRAM a full cycle on reads, and thus we souce the BRAM address
directly from fetch1 latched NIA.

(Note: This will be problematic if/when we want to hash the address,
we'll probably be better off having fetch1 latch a fully hashed address
along with the normal one, so the icache can use the former to address
the BRAM and pass the latter along)

One difficulty is that we cannot really stall the icache without adding
more combo logic that would break the "one full cycle" BRAM model. This
means that on stalls from decode, by the time we stall fetch1, it has
already gone to the next address, which the icache is already latching.

We work around this by having a "stash" buffer in fetch2 that will stash
away the icache output on a stall, and override the output of the icache
with the content of the stash buffer when unstalling.

This requires a rewrite of the stop/step debug logic as well. We now
do most of the hard work in fetch1 which makes more sense.

Note: Vivado is still not inferring an built-in output register for the
BRAMs. I don't want to add another cycle... I don't fully understand why
it wouldn't be able to treat current_row as such but clearly it won't. At
least the timing seems good enough now for 100Mhz, possibly more.

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
jtag-port
Benjamin Herrenschmidt 5 years ago
parent 3589f92d5a
commit d415e5544a

@ -12,17 +12,16 @@ package common is
carry: std_ulogic;
end record;

type Fetch1ToFetch2Type is record
nia: std_ulogic_vector(63 downto 0);
end record;

type Fetch2ToIcacheType is record
type Fetch1ToIcacheType is record
req: std_ulogic;
addr: std_ulogic_vector(63 downto 0);
stop_mark: std_ulogic;
nia: std_ulogic_vector(63 downto 0);
end record;

type IcacheToFetch2Type is record
ack: std_ulogic;
valid: std_ulogic;
stop_mark: std_ulogic;
nia: std_ulogic_vector(63 downto 0);
insn: std_ulogic_vector(31 downto 0);
end record;


@ -33,11 +33,10 @@ end core;

architecture behave of core is
-- fetch signals
signal fetch1_to_fetch2: Fetch1ToFetch2Type;
signal fetch2_to_decode1: Fetch2ToDecode1Type;

-- icache signals
signal fetch2_to_icache : Fetch2ToIcacheType;
signal fetch1_to_icache : Fetch1ToIcacheType;
signal icache_to_fetch2 : IcacheToFetch2Type;

-- decode signals
@ -74,8 +73,8 @@ architecture behave of core is

-- local signals
signal fetch1_stall_in : std_ulogic;
signal icache_stall_out : std_ulogic;
signal fetch2_stall_in : std_ulogic;
signal fetch2_stall_out : std_ulogic;
signal decode1_stall_in : std_ulogic;
signal decode2_stall_out : std_ulogic;

@ -107,27 +106,12 @@ begin
rst => core_rst,
stall_in => fetch1_stall_in,
flush_in => flush,
e_in => execute1_to_fetch1,
f_out => fetch1_to_fetch2
);

fetch1_stall_in <= fetch2_stall_out or decode2_stall_out;

fetch2_0: entity work.fetch2
port map (
clk => clk,
rst => core_rst,
stall_in => fetch2_stall_in,
stall_out => fetch2_stall_out,
flush_in => flush,
i_in => icache_to_fetch2,
i_out => fetch2_to_icache,
stop_in => dbg_core_stop,
f_in => fetch1_to_fetch2,
f_out => fetch2_to_decode1
e_in => execute1_to_fetch1,
i_out => fetch1_to_icache
);

fetch2_stall_in <= decode2_stall_out;
fetch1_stall_in <= icache_stall_out or decode2_stall_out;

icache_0: entity work.icache
generic map(
@ -137,13 +121,27 @@ begin
port map(
clk => clk,
rst => icache_rst,
i_in => fetch2_to_icache,
i_in => fetch1_to_icache,
i_out => icache_to_fetch2,
flush_in => flush,
stall_out => icache_stall_out,
wishbone_out => wishbone_insn_out,
wishbone_in => wishbone_insn_in
);

icache_rst <= rst or dbg_icache_rst;
icache_rst <= rst or dbg_icache_rst;

fetch2_0: entity work.fetch2
port map (
clk => clk,
rst => core_rst,
stall_in => fetch2_stall_in,
flush_in => flush,
i_in => icache_to_fetch2,
f_out => fetch2_to_decode1
);

fetch2_stall_in <= decode2_stall_out;

decode1_0: entity work.decode1
port map (
@ -274,7 +272,7 @@ begin
icache_rst => dbg_icache_rst,
terminate => terminate,
core_stopped => dbg_core_is_stopped,
nia => fetch1_to_fetch2.nia,
nia => fetch1_to_icache.nia,
terminated_out => terminated_out
);


@ -91,15 +91,15 @@ begin
reg_write: process(clk)
begin
if rising_edge(clk) then
-- Reset the 1-cycle "do" signals
do_step <= '0';
do_reset <= '0';
do_icreset <= '0';

if (rst) then
stopping <= '0';
terminated <= '0';
else
-- Reset the 1-cycle "do" signals
do_step <= '0';
do_reset <= '0';
do_icreset <= '0';

-- Edge detect on dmi_req for 1-shot pulses
dmi_req_1 <= dmi_req;
if dmi_req = '1' and dmi_req_1 = '0' then

@ -16,51 +16,111 @@ entity fetch1 is
-- Control inputs:
stall_in : in std_ulogic;
flush_in : in std_ulogic;
stop_in : in std_ulogic;

-- redirect from execution unit
e_in : in Execute1ToFetch1Type;

-- fetch data out
f_out : out Fetch1ToFetch2Type
-- Request to icache
i_out : out Fetch1ToIcacheType
);
end entity fetch1;

architecture behaviour of fetch1 is
signal r, r_next : Fetch1ToFetch2Type;
type stop_state_t is (RUNNING, STOPPED, RESTARTING);
type reg_internal_t is record
stop_state: stop_state_t;
end record;
signal r, r_next : Fetch1ToIcacheType;
signal r_int, r_next_int : reg_internal_t;
begin

regs : process(clk)
begin
if rising_edge(clk) then
if rst = '1' or e_in.redirect = '1' or stall_in = '0' then
r <= r_next;
if r /= r_next then
report "fetch1 rst:" & std_ulogic'image(rst) &
" R:" & std_ulogic'image(e_in.redirect) &
" S:" & std_ulogic'image(stall_in) &
" T:" & std_ulogic'image(stop_in) &
" nia:" & to_hstring(r_next.nia) &
" SM:" & std_ulogic'image(r_next.stop_mark);
end if;
r <= r_next;
r_int <= r_next_int;
end if;
end process;

comb : process(all)
variable v : Fetch1ToFetch2Type;
variable v : Fetch1ToIcacheType;
variable v_int : reg_internal_t;
variable increment : boolean;
begin
v := r;
v_int := r_int;

if rst = '1' then
v.nia := RESET_ADDRESS;
v_int.stop_state := RUNNING;
elsif e_in.redirect = '1' then
v.nia := e_in.redirect_nia;
else
v.nia := std_logic_vector(unsigned(v.nia) + 4);
elsif stall_in = '0' then

-- For debug stop/step to work properly we need a little bit of
-- trickery here. If we just stop incrementing and send stop marks
-- when stop_in is set, then we'll increment on the cycle it clears
-- and end up never executing the instruction we were stopped on.
--
-- Avoid this along with the opposite issue when stepping (stop is
-- cleared for only one cycle) is handled by the state machine below
--
-- By default, increment addresses
increment := true;
case v_int.stop_state is
when RUNNING =>
-- If we are running and stop_in is set, then stop incrementing,
-- we are now stopped.
if stop_in = '1' then
increment := false;
v_int.stop_state := STOPPED;
end if;
when STOPPED =>
-- When stopped, never increment. If stop is cleared, go to state
-- "restarting" but still don't increment that cycle. stop_in is
-- now 0 so we'll send the NIA down without a stop mark.
increment := false;
if stop_in = '0' then
v_int.stop_state := RESTARTING;
end if;
when RESTARTING =>
-- We have just sent the NIA down, we can start incrementing again.
-- If stop_in is still not set, go back to running normally.
-- If stop_in is set again (that was a one-cycle "step"), go
-- back to "stopped" state which means we'll stop incrementing
-- on the next cycle. This ensures we increment the PC once after
-- sending one instruction without a stop mark. Since stop_in is
-- now set, the new PC will be sent with a stop mark and thus not
-- executed.
if stop_in = '0' then
v_int.stop_state := RUNNING;
else
v_int.stop_state := STOPPED;
end if;
end case;

if increment then
v.nia := std_logic_vector(unsigned(v.nia) + 4);
end if;
end if;

v.req := not rst;
v.stop_mark := stop_in;

r_next <= v;
r_next_int <= v_int;

-- Update outputs to the icache
f_out <= r;

report "fetch1 rst:" & std_ulogic'image(rst) &
" R:" & std_ulogic'image(e_in.redirect) &
" S:" & std_ulogic'image(stall_in) &
" nia_next:" & to_hstring(r_next.nia) &
" nia:" & to_hstring(r.nia);
i_out <= r;

end process;


@ -12,55 +12,108 @@ entity fetch2 is
rst : in std_ulogic;

stall_in : in std_ulogic;
stall_out : out std_ulogic;

flush_in : in std_ulogic;
stop_in : in std_ulogic;

-- Results from icache
i_in : in IcacheToFetch2Type;
i_out : out Fetch2ToIcacheType;

f_in : in Fetch1ToFetch2Type;

-- Output to decode
f_out : out Fetch2ToDecode1Type
);
end entity fetch2;

architecture behaviour of fetch2 is

-- The icache cannot stall, so we need to stash a cycle
-- of output from it when we stall.
type reg_internal_type is record
stash : IcacheToFetch2Type;
stash_valid : std_ulogic;
stopped : std_ulogic;
end record;

signal r_int, rin_int : reg_internal_type;
signal r, rin : Fetch2ToDecode1Type;

begin
regs : process(clk)
begin
if rising_edge(clk) then

if (r /= rin) then
report "fetch2 rst:" & std_ulogic'image(rst) &
" S:" & std_ulogic'image(stall_in) &
" F:" & std_ulogic'image(flush_in) &
" T:" & std_ulogic'image(rin.stop_mark) &
" V:" & std_ulogic'image(rin.valid) &
" nia:" & to_hstring(rin.nia);
end if;

-- Output state remains unchanged on stall, unless we are flushing
if rst = '1' or flush_in = '1' or stall_in = '0' then
r <= rin;
end if;

-- Internal state is updated on every clock
r_int <= rin_int;
end if;
end process;

comb : process(all)
variable v : Fetch2ToDecode1Type;
variable v : Fetch2ToDecode1Type;
variable v_int : reg_internal_type;
variable v_i_in : IcacheToFetch2Type;
begin
v := r;
v_int := r_int;

-- asynchronous icache lookup
i_out.req <= '1';
i_out.addr <= f_in.nia;
v.valid := i_in.ack;
v.nia := f_in.nia;
v.insn := i_in.insn;
stall_out <= stop_in or not i_in.ack;
-- If stalling, stash away the current input from the icache
if stall_in = '1' and v_int.stash_valid = '0' then
v_int.stash := i_in;
v_int.stash_valid := '1';
end if;

-- If unstalling, source input from the stash and invalidate it,
-- otherwise source normally from the icache.
--
v_i_in := i_in;
if v_int.stash_valid = '1' and stall_in = '0' then
v_i_in := v_int.stash;
v_int.stash_valid := '0';
end if;

v.valid := v_i_in.valid;
v.stop_mark := v_i_in.stop_mark;
v.nia := v_i_in.nia;
v.insn := v_i_in.insn;

-- Clear stash internal valid bit on flush. We still mark
-- the stash itself as valid since we still want to override
-- whatever comes form icache when unstalling, but we'll
-- override it with something invalid.
--
if flush_in = '1' then
v_int.stash.valid := '0';
end if;

-- If we are flushing or the instruction comes with a stop mark
-- we tag it as invalid so it doesn't get decoded and executed
if flush_in = '1' or v.stop_mark = '1' then

if flush_in = '1' or stop_in = '1' then
v.valid := '0';
end if;
v.stop_mark := stop_in;

-- Clear stash on reset
if rst = '1' then
v_int.stash_valid := '0';
end if;

-- Update registers
rin <= v;
rin_int <= v_int;

-- Update outputs
f_out <= r;
end process;

end architecture behaviour;

@ -19,9 +19,12 @@ entity icache is
clk : in std_ulogic;
rst : in std_ulogic;

i_in : in Fetch2ToIcacheType;
i_in : in Fetch1ToIcacheType;
i_out : out IcacheToFetch2Type;

stall_out : out std_ulogic;
flush_in : in std_ulogic;

wishbone_out : out wishbone_master_out;
wishbone_in : in wishbone_slave_out
);
@ -59,113 +62,194 @@ architecture rtl of icache is
subtype cacheline_tag_type is std_logic_vector(TAG_BITS-1 downto 0);
type cacheline_tag_array is array(0 to NUM_LINES-1) of cacheline_tag_type;

signal cachelines : cacheline_array := (others => (others => '0'));
signal tags : cacheline_tag_array := (others => (others => '0'));
signal tags_valid : std_ulogic_vector(NUM_LINES-1 downto 0) := (others => '0');

-- Storage. Hopefully "cachelines" is a BRAM, the rest is LUTs
signal cachelines : cacheline_array;
signal tags : cacheline_tag_array;
signal tags_valid : std_ulogic_vector(NUM_LINES-1 downto 0);
attribute ram_style : string;
attribute ram_style of cachelines : signal is "block";

attribute ram_decomp : string;
attribute ram_decomp of cachelines : signal is "power";

-- Cache reload state machine
type state_type is (IDLE, WAIT_ACK);

type reg_internal_type is record
state : state_type;
w : wishbone_master_out;
store_index : integer range 0 to (NUM_LINES-1);
store_word : integer range 0 to (LINE_SIZE-1);
-- Cache hit state (1 cycle BRAM access)
hit_line : cacheline_type;
hit_nia : std_ulogic_vector(63 downto 0);
hit_smark : std_ulogic;
hit_valid : std_ulogic;

-- Cache miss state (reload state machine)
state : state_type;
wb : wishbone_master_out;
store_index : integer range 0 to (NUM_LINES-1);
store_mask : std_ulogic_vector(LINE_SIZE_DW-1 downto 0);
end record;

signal r : reg_internal_type;

signal read_index : integer range 0 to NUM_LINES-1;
signal read_tag : std_ulogic_vector(63-OFFSET_BITS-INDEX_BITS downto 0);
signal read_miss : boolean;
-- Async signals decoding incoming requests
signal req_index : integer range 0 to NUM_LINES-1;
signal req_tag : std_ulogic_vector(TAG_BITS-1 downto 0);
signal req_word : integer range 0 to LINE_SIZE_DW*2-1;
signal req_is_hit : std_ulogic;

-- Return the cache line index (tag index) for an address
function get_index(addr: std_ulogic_vector(63 downto 0)) return integer is
begin
return to_integer(unsigned(addr((OFFSET_BITS+INDEX_BITS-1) downto OFFSET_BITS)));
end;

function get_word(addr: std_ulogic_vector(63 downto 0); data: cacheline_type) return std_ulogic_vector is
variable word : integer;
-- Return the word index in a cache line for an address
function get_word(addr: std_ulogic_vector(63 downto 0)) return integer is
begin
return to_integer(unsigned(addr(OFFSET_BITS-1 downto 2)));
end;

-- Read a word in a cache line for an address
function read_word(word: integer; data: cacheline_type) return std_ulogic_vector is
begin
word := to_integer(unsigned(addr(OFFSET_BITS-1 downto 2)));
return data((word+1)*32-1 downto word*32);
return data((word+1)*32-1 downto word*32);
end;

-- Calculate the tag value from the address
function get_tag(addr: std_ulogic_vector(63 downto 0)) return std_ulogic_vector is
begin
return addr(63 downto OFFSET_BITS+INDEX_BITS);
end;

begin
assert ispow2(LINE_SIZE) report "LINE_SIZE not power of 2" severity FAILURE;
assert ispow2(NUM_LINES) report "NUM_LINES not power of 2" severity FAILURE;

icache_read : process(all)
icache_comb : process(all)
begin
read_index <= get_index(i_in.addr);
read_tag <= get_tag(i_in.addr);
read_miss <= false;

i_out.ack <= '0';
i_out.insn <= get_word(i_in.addr, cachelines(read_index));

if i_in.req = '1' then
if (tags_valid(read_index) = '1') and (tags(read_index) = read_tag) then
-- report hit asynchronously
i_out.ack <= '1';
else
read_miss <= true;
end if;
end if;
-- Calculate next index and tag index
req_index <= get_index(i_in.nia);
req_tag <= get_tag(i_in.nia);
req_word <= get_word(i_in.nia);

-- Test if pending request is a hit
if tags(req_index) = req_tag then
req_is_hit <= tags_valid(req_index);
else
req_is_hit <= '0';
end if;

-- Output instruction from current cache line
--
-- Note: This is a mild violation of our design principle of having pipeline
-- stages output from a clean latch. In this case we output the result
-- of a mux. The alternative would be output an entire cache line
-- which I prefer not to do just yet.
--
i_out.valid <= r.hit_valid;
i_out.insn <= read_word(get_word(r.hit_nia), r.hit_line);
i_out.nia <= r.hit_nia;
i_out.stop_mark <= r.hit_smark;

-- This needs to match the latching of a new request in icache_hit
stall_out <= not req_is_hit;

-- Wishbone requests output (from the cache miss reload machine)
wishbone_out <= r.wb;
end process;

wishbone_out <= r.w;
icache_hit : process(clk)
begin
if rising_edge(clk) then
-- Assume we have nothing valid first
r.hit_valid <= '0';

-- Are we free to latch a new request ?
--
-- Note: this test needs to match the equation for generating stall_out
--
if i_in.req = '1' and req_is_hit = '1' and flush_in = '0' then
-- Read the cache line (BRAM read port) and remember the NIA
r.hit_line <= cachelines(req_index);
r.hit_nia <= i_in.nia;
r.hit_smark <= i_in.stop_mark;
r.hit_valid <= '1';

report "cache hit nia:" & to_hstring(i_in.nia) &
" SM:" & std_ulogic'image(i_in.stop_mark) &
" idx:" & integer'image(req_index) &
" tag:" & to_hstring(req_tag);
end if;

icache_write : process(clk)
-- Flush requested ? discard...
if flush_in then
r.hit_valid <= '0';
end if;
end if;
end process;

icache_miss : process(clk)
variable store_dword : std_ulogic_vector(OFFSET_BITS-4 downto 0);
begin
if rising_edge(clk) then
if rst = '1' then
tags_valid <= (others => '0');
r.store_mask <= (others => '0');
r.state <= IDLE;
r.w.cyc <= '0';
r.w.stb <= '0';
end if;
r.wb.cyc <= '0';
r.wb.stb <= '0';

r.w.dat <= (others => '0');
r.w.sel <= "11111111";
r.w.we <= '0';
-- We only ever do reads on wishbone
r.wb.dat <= (others => '0');
r.wb.sel <= "11111111";
r.wb.we <= '0';
end if;

-- State machine
case r.state is
when IDLE =>
if read_miss = true then
-- We need to read a cache line
if i_in.req = '1' and req_is_hit = '0' then

report "cache miss nia:" & to_hstring(i_in.nia) &
" SM:" & std_ulogic'image(i_in.stop_mark) &
" idx:" & integer'image(req_index) &
" tag:" & to_hstring(req_tag);

r.state <= WAIT_ACK;
r.store_word <= 0;
r.store_index <= read_index;
r.store_mask <= (0 => '1', others => '0');
r.store_index <= req_index;

tags(read_index) <= read_tag;
tags_valid(read_index) <= '0';
-- Force misses while reloading that line
tags_valid(req_index) <= '0';
tags(req_index) <= req_tag;

r.w.adr <= i_in.addr(63 downto OFFSET_BITS) & (OFFSET_BITS-1 downto 0 => '0');
r.w.cyc <= '1';
r.w.stb <= '1';
-- Prep for first dword read
r.wb.adr <= i_in.nia(63 downto OFFSET_BITS) & (OFFSET_BITS-1 downto 0 => '0');
r.wb.cyc <= '1';
r.wb.stb <= '1';
end if;
when WAIT_ACK =>
if wishbone_in.ack = '1' then
cachelines(r.store_index)((r.store_word+1)*64-1 downto ((r.store_word)*64)) <= wishbone_in.dat;
r.store_word <= r.store_word + 1;
-- Store the current dword in both the cache
for i in 0 to LINE_SIZE_DW-1 loop
if r.store_mask(i) = '1' then
cachelines(r.store_index)(63 + i*64 downto i*64) <= wishbone_in.dat;
end if;
end loop;

if r.store_word = (LINE_SIZE_DW-1) then
-- That was the last word ? We are done
if r.store_mask(LINE_SIZE_DW-1) = '1' then
r.state <= IDLE;
tags_valid(r.store_index) <= '1';
r.w.cyc <= '0';
r.w.stb <= '0';
r.wb.cyc <= '0';
r.wb.stb <= '0';
else
r.w.adr(OFFSET_BITS-1 downto 3) <= std_ulogic_vector(to_unsigned(r.store_word+1, OFFSET_BITS-3));
store_dword := r.wb.adr(OFFSET_BITS-1 downto 3);
store_dword := std_ulogic_vector(unsigned(store_dword) + 1);
r.wb.adr(OFFSET_BITS-1 downto 3) <= store_dword;
end if;
-- Advance to next word
r.store_mask <= r.store_mask(LINE_SIZE_DW-2 downto 0) & '0';
end if;
end case;
end if;

@ -12,7 +12,7 @@ architecture behave of icache_tb is
signal clk : std_ulogic;
signal rst : std_ulogic;

signal i_out : Fetch2ToIcacheType;
signal i_out : Fetch1ToIcacheType;
signal i_in : IcacheToFetch2Type;

signal wb_bram_in : wishbone_master_out;
@ -30,6 +30,7 @@ begin
rst => rst,
i_in => i_out,
i_out => i_in,
flush_in => '0',
wishbone_out => wb_bram_in,
wishbone_in => wb_bram_out
);
@ -66,16 +67,16 @@ begin
stim: process
begin
i_out.req <= '0';
i_out.addr <= (others => '0');
i_out.nia <= (others => '0');

wait for 4*clk_period;

i_out.req <= '1';
i_out.addr <= x"0000000000000004";
i_out.nia <= x"0000000000000004";

wait for 30*clk_period;

assert i_in.ack = '1';
assert i_in.valid = '1';
assert i_in.insn = x"00000001";

i_out.req <= '0';
@ -84,31 +85,31 @@ begin

-- hit
i_out.req <= '1';
i_out.addr <= x"0000000000000008";
wait for clk_period/2;
assert i_in.ack = '1';
i_out.nia <= x"0000000000000008";
wait for clk_period;
assert i_in.valid = '1';
assert i_in.insn = x"00000002";
wait for clk_period/2;
wait for clk_period;

-- another miss
i_out.req <= '1';
i_out.addr <= x"0000000000000040";
i_out.nia <= x"0000000000000040";

wait for 30*clk_period;

assert i_in.ack = '1';
assert i_in.valid = '1';
assert i_in.insn = x"00000010";

-- test something that aliases
i_out.req <= '1';
i_out.addr <= x"0000000000000100";
wait for clk_period/2;
assert i_in.ack = '0';
wait for clk_period/2;
i_out.nia <= x"0000000000000100";
wait for clk_period;
assert i_in.valid = '0';
wait for clk_period;

wait for 30*clk_period;

assert i_in.ack = '1';
assert i_in.valid = '1';
assert i_in.insn = x"00000040";

i_out.req <= '0';

Loading…
Cancel
Save