--
-- Set associative dcache write-through
--
--- TODO (in no specific order):
---
--- * See list in icache.vhdl
--- * Complete load misses on the cycle when WB data comes instead of
--- at the end of line (this requires dealing with requests coming in
--- while not idle...)
--
library ieee;
use ieee.std_logic_1164.all;
architecture rtl of dcache is
-- BRAM organisation: We never access more than wishbone_data_bits at
-- a time so to save resources we make the array only that wide, and
- -- use consecutive indices for to make a cache "line"
+ -- use consecutive indices to make a cache "line"
--
-- ROW_SIZE is the width in bytes of the BRAM (based on WB, so 64-bits)
constant ROW_SIZE : natural := wishbone_data_bits / 8;
-- which means that the BRAM output is delayed by an extra cycle.
--
-- Thus, the dcache has a 2-stage internal pipeline for cache hits
- -- with no stalls.
+ -- with no stalls. Stores also complete in 2 cycles in most
+ -- circumstances.
+ --
+ -- A request proceeds through the pipeline as follows.
+ --
+ -- Cycle 0: Request is received from loadstore or mmu if either
+ -- d_in.valid or m_in.valid is 1 (not both). In this cycle portions
+ -- of the address are presented to the TLB tag RAM and data RAM
+ -- and the cache tag RAM and data RAM.
+ --
+ -- Clock edge between cycle 0 and cycle 1:
+ -- Request is stored in r0 (assuming r0_full was 0). TLB tag and
+ -- data RAMs are read, and the cache tag RAM is read. (Cache data
+ -- comes out a cycle later due to its output register, giving the
+ -- whole of cycle 1 to read the cache data RAM.)
+ --
+ -- Cycle 1: TLB and cache tag matching is done, the real address
+ -- (RA) for the access is calculated, and the type of operation is
+ -- determined (the OP_* values above). This gives the TLB way for
+ -- a TLB hit, and the cache way for a hit or the way to replace
+ -- for a load miss.
--
- -- All other operations are handled via stalling in the first stage.
+ -- Clock edge between cycle 1 and cycle 2:
+ -- Request is stored in r1 (assuming r1.full was 0)
+ -- The state machine transitions out of IDLE state for a load miss,
+ -- a store, a dcbz, or a non-cacheable load. r1.full is set to 1
+ -- for a load miss, dcbz or non-cacheable load but not a store.
--
- -- The second stage can thus complete a hit at the same time as the
- -- first stage emits a stall for a complex op.
+ -- Cycle 2: Completion signals are asserted for a load hit,
+ -- a store (excluding dcbz), a TLB operation, a conditional
+ -- store which failed due to no matching reservation, or an error
+ -- (cache hit on non-cacheable operation, TLB miss, or protection
+ -- fault).
--
+ -- For a load miss, store, or dcbz, the state machine initiates
+ -- a wishbone cycle, which takes at least 2 cycles. For a store,
+ -- if another store comes in with the same cache tag (therefore
+ -- in the same 4k page), it can be added on to the existing cycle,
+ -- subject to some constraints.
+ -- While r1.full = 1, no new requests can go from r0 to r1, but
+ -- requests can come in to r0 and be satisfied if they are
+ -- cacheable load hits or stores with the same cache tag.
+ --
+ -- Writing to the cache data RAM is done at the clock edge
+ -- at the end of cycle 2 for a store hit (excluding dcbz).
+ -- Stores that miss are not written to the cache data RAM
+ -- but just stored through to memory.
+ -- Dcbz is done like a cache miss, but the wishbone cycle
+ -- is a write rather than a read, and zeroes are written to
+ -- the cache data RAM. Thus dcbz will allocate the line in
+ -- the cache as well as zeroing memory.
+ --
+ -- Since stores are written to the cache data RAM at the end of
+ -- cycle 2, and loads can come in and hit on the data just stored,
+ -- there is a two-stage bypass from store data to load data to
+ -- make sure that loads always see previously-stored data even
+ -- if it has not yet made it to the cache data RAM.
+ --
+ -- Load misses read the requested dword of the cache line first in
+ -- the memory read request and then cycle around through the other
+ -- dwords. The load is completed on the cycle after the requested
+ -- dword comes back from memory (using a forwarding path, rather
+ -- than going via the cache data RAM). We maintain an array of
+ -- valid bits per dword for the line being refilled so that
+ -- subsequent load requests to the same line can be completed as
+ -- soon as the necessary data comes in from memory, without
+ -- waiting for the whole line to be read.
-- Stage 0 register, basically contains just the latched request
type reg_stage_0_t is record
req : Loadstore1ToDcacheType;
- tlbie : std_ulogic;
- doall : std_ulogic;
- tlbld : std_ulogic;
+ tlbie : std_ulogic; -- indicates a tlbie request (from MMU)
+ doall : std_ulogic; -- with tlbie, indicates flush whole TLB
+ tlbld : std_ulogic; -- indicates a TLB load request (from MMU)
mmu_req : std_ulogic; -- indicates source of request
+ d_valid : std_ulogic; -- indicates req.data is valid now
end record;
signal r0 : reg_stage_0_t;
r.mmu_req := '1';
else
r.req := d_in;
+ r.req.data := (others => '0');
r.tlbie := '0';
r.doall := '0';
r.tlbld := '0';
r.mmu_req := '0';
end if;
+ r.d_valid := '0';
if rst = '1' then
r0_full <= '0';
- elsif r1.full = '0' or r0_full = '0' then
+ elsif (r1.full = '0' and d_in.hold = '0') or r0_full = '0' then
r0 <= r;
r0_full <= r.req.valid;
end if;
+ -- Sample data the cycle after a request comes in from loadstore1.
+ -- If another request has come in already then the data will get
+ -- put directly into req.data below.
+ if r0.req.valid = '1' and r.req.valid = '0' and r0.d_valid = '0' and
+ r0.mmu_req = '0' then
+ r0.req.data <= d_in.data;
+ r0.d_valid <= '1';
+ end if;
end if;
end process;
m_out.stall <= '0';
-- Hold off the request in r0 when r1 has an uncompleted request
- r0_stall <= r0_full and r1.full;
- r0_valid <= r0_full and not r1.full;
+ r0_stall <= r0_full and (r1.full or d_in.hold);
+ r0_valid <= r0_full and not r1.full and not d_in.hold;
stall_out <= r0_stall;
-- TLB
-- XXX or if r0.req.nc = '1'
if r0.req.load = '1' then
-- load with reservation
- set_rsrv <= '1';
+ set_rsrv <= r0.req.atomic_last;
else
-- store conditional
- clear_rsrv <= '1';
+ clear_rsrv <= r0.req.atomic_last;
if reservation.valid = '0' or
r0.req.addr(63 downto LINE_OFF_BITS) /= reservation.addr then
cancel_store <= '1';
req.mmu_req := r0.mmu_req;
req.dcbz := r0.req.dcbz;
req.real_addr := ra;
- req.data := r0.req.data;
- req.byte_sel := r0.req.byte_sel;
+ -- Force data to 0 for dcbz
+ if r0.req.dcbz = '1' then
+ req.data := (others => '0');
+ elsif r0.d_valid = '1' then
+ req.data := r0.req.data;
+ else
+ req.data := d_in.data;
+ end if;
+ -- Select all bytes for dcbz and for cacheable loads
+ if r0.req.dcbz = '1' or (r0.req.load = '1' and r0.req.nc = '0') then
+ req.byte_sel := (others => '1');
+ else
+ req.byte_sel := r0.req.byte_sel;
+ end if;
req.hit_way := req_hit_way;
req.same_tag := req_same_tag;
case r1.state is
when IDLE =>
r1.wb.adr <= req.real_addr(r1.wb.adr'left downto 0);
- r1.dcbz <= '0';
+ r1.wb.sel <= req.byte_sel;
+ r1.wb.dat <= req.data;
+ r1.dcbz <= req.dcbz;
-- Keep track of our index and way for subsequent stores.
r1.store_index <= get_index(req.real_addr);
" tag:" & to_hstring(get_tag(req.real_addr));
-- Start the wishbone cycle
- r1.wb.sel <= (others => '1');
r1.wb.we <= '0';
r1.wb.cyc <= '1';
r1.wb.stb <= '1';
r1.write_tag <= '1';
when OP_LOAD_NC =>
- r1.wb.sel <= req.byte_sel;
r1.wb.cyc <= '1';
r1.wb.stb <= '1';
r1.wb.we <= '0';
when OP_STORE_HIT | OP_STORE_MISS =>
if req.dcbz = '0' then
- r1.wb.sel <= req.byte_sel;
- r1.wb.dat <= req.data;
r1.state <= STORE_WAIT_ACK;
r1.acks_pending <= to_unsigned(1, 3);
r1.full <= '0';
else
-- dcbz is handled much like a load miss except
-- that we are writing to memory instead of reading
-
- -- Start the wishbone writes
- r1.wb.sel <= (others => '1');
- r1.wb.dat <= (others => '0');
-
- -- Handle the rest like a load miss
r1.state <= RELOAD_WAIT_ACK;
if req.op = OP_STORE_MISS then
r1.write_tag <= '1';
end if;
- r1.dcbz <= '1';
end if;
r1.wb.we <= '1';
r1.wb.cyc <= '1';
-- complete the request next cycle.
-- Compare the whole address in case the request in
-- r1.req is not the one that started this refill.
- if r1.full = '1' and r1.req.same_tag = '1' and
- ((r1.dcbz = '1' and r1.req.dcbz = '1') or
- (r1.dcbz = '0' and r1.req.op = OP_LOAD_MISS)) and
- r1.store_row = get_row(r1.req.real_addr) then
+ if req.valid = '1' and req.same_tag = '1' and
+ ((r1.dcbz = '1' and req.dcbz = '1') or
+ (r1.dcbz = '0' and req.op = OP_LOAD_MISS)) and
+ r1.store_row = get_row(req.real_addr) then
r1.full <= '0';
r1.slow_valid <= '1';
if r1.mmu_req = '0' then