You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
microwatt/dcache.vhdl

1145 lines
40 KiB
VHDL

--
-- 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;
use ieee.numeric_std.all;
library work;
use work.utils.all;
use work.common.all;
use work.helpers.all;
use work.wishbone_types.all;
entity dcache is
generic (
-- Line size in bytes
LINE_SIZE : positive := 64;
-- Number of lines in a set
NUM_LINES : positive := 32;
-- Number of ways
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
NUM_WAYS : positive := 4;
-- L1 DTLB entries per set
TLB_SET_SIZE : positive := 64;
-- L1 DTLB number of sets
TLB_NUM_WAYS : positive := 2;
-- L1 DTLB log_2(page_size)
TLB_LG_PGSZ : positive := 12
);
port (
clk : in std_ulogic;
rst : in std_ulogic;
d_in : in Loadstore1ToDcacheType;
d_out : out DcacheToLoadstore1Type;
stall_out : out std_ulogic;
wishbone_out : out wishbone_master_out;
wishbone_in : in wishbone_slave_out
);
end entity dcache;
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"
--
-- ROW_SIZE is the width in bytes of the BRAM (based on WB, so 64-bits)
constant ROW_SIZE : natural := wishbone_data_bits / 8;
-- ROW_PER_LINE is the number of row (wishbone transactions) in a line
constant ROW_PER_LINE : natural := LINE_SIZE / ROW_SIZE;
-- BRAM_ROWS is the number of rows in BRAM needed to represent the full
-- dcache
constant BRAM_ROWS : natural := NUM_LINES * ROW_PER_LINE;
-- Bit fields counts in the address
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
-- REAL_ADDR_BITS is the number of real address bits that we store
constant REAL_ADDR_BITS : positive := 56;
-- ROW_BITS is the number of bits to select a row
constant ROW_BITS : natural := log2(BRAM_ROWS);
-- ROW_LINEBITS is the number of bits to select a row within a line
constant ROW_LINEBITS : natural := log2(ROW_PER_LINE);
-- LINE_OFF_BITS is the number of bits for the offset in a cache line
constant LINE_OFF_BITS : natural := log2(LINE_SIZE);
-- ROW_OFF_BITS is the number of bits for the offset in a row
constant ROW_OFF_BITS : natural := log2(ROW_SIZE);
-- INDEX_BITS is the number if bits to select a cache line
constant INDEX_BITS : natural := log2(NUM_LINES);
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
-- SET_SIZE_BITS is the log base 2 of the set size
constant SET_SIZE_BITS : natural := LINE_OFF_BITS + INDEX_BITS;
-- TAG_BITS is the number of bits of the tag part of the address
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
constant TAG_BITS : natural := REAL_ADDR_BITS - SET_SIZE_BITS;
-- WAY_BITS is the number of bits to select a way
constant WAY_BITS : natural := log2(NUM_WAYS);
-- Example of layout for 32 lines of 64 bytes:
--
-- .. tag |index| line |
-- .. | row | |
-- .. | |---| | ROW_LINEBITS (3)
-- .. | |--- - --| LINE_OFF_BITS (6)
-- .. | |- --| ROW_OFF_BITS (3)
-- .. |----- ---| | ROW_BITS (8)
-- .. |-----| | INDEX_BITS (5)
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
-- .. --------| | TAG_BITS (45)
subtype row_t is integer range 0 to BRAM_ROWS-1;
subtype index_t is integer range 0 to NUM_LINES-1;
subtype way_t is integer range 0 to NUM_WAYS-1;
-- The cache data BRAM organized as described above for each way
subtype cache_row_t is std_ulogic_vector(wishbone_data_bits-1 downto 0);
-- The cache tags LUTRAM has a row per set. Vivado is a pain and will
-- not handle a clean (commented) definition of the cache tags as a 3d
-- memory. For now, work around it by putting all the tags
subtype cache_tag_t is std_logic_vector(TAG_BITS-1 downto 0);
-- type cache_tags_set_t is array(way_t) of cache_tag_t;
-- type cache_tags_array_t is array(index_t) of cache_tags_set_t;
constant TAG_RAM_WIDTH : natural := TAG_BITS * NUM_WAYS;
subtype cache_tags_set_t is std_logic_vector(TAG_RAM_WIDTH-1 downto 0);
type cache_tags_array_t is array(index_t) of cache_tags_set_t;
-- The cache valid bits
subtype cache_way_valids_t is std_ulogic_vector(NUM_WAYS-1 downto 0);
type cache_valids_t is array(index_t) of cache_way_valids_t;
-- Storage. Hopefully "cache_rows" is a BRAM, the rest is LUTs
signal cache_tags : cache_tags_array_t;
signal cache_valids : cache_valids_t;
attribute ram_style : string;
attribute ram_style of cache_tags : signal is "distributed";
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
-- L1 TLB.
constant TLB_SET_BITS : natural := log2(TLB_SET_SIZE);
constant TLB_WAY_BITS : natural := log2(TLB_NUM_WAYS);
constant TLB_EA_TAG_BITS : natural := 64 - (TLB_LG_PGSZ + TLB_SET_BITS);
constant TLB_TAG_WAY_BITS : natural := TLB_NUM_WAYS * TLB_EA_TAG_BITS;
constant TLB_PTE_BITS : natural := 64;
constant TLB_PTE_WAY_BITS : natural := TLB_NUM_WAYS * TLB_PTE_BITS;
subtype tlb_way_t is integer range 0 to TLB_NUM_WAYS - 1;
subtype tlb_index_t is integer range 0 to TLB_SET_SIZE - 1;
subtype tlb_way_valids_t is std_ulogic_vector(TLB_NUM_WAYS-1 downto 0);
type tlb_valids_t is array(tlb_index_t) of tlb_way_valids_t;
subtype tlb_tag_t is std_ulogic_vector(TLB_EA_TAG_BITS - 1 downto 0);
subtype tlb_way_tags_t is std_ulogic_vector(TLB_TAG_WAY_BITS-1 downto 0);
type tlb_tags_t is array(tlb_index_t) of tlb_way_tags_t;
subtype tlb_pte_t is std_ulogic_vector(TLB_PTE_BITS - 1 downto 0);
subtype tlb_way_ptes_t is std_ulogic_vector(TLB_PTE_WAY_BITS-1 downto 0);
type tlb_ptes_t is array(tlb_index_t) of tlb_way_ptes_t;
type hit_way_set_t is array(tlb_way_t) of way_t;
signal dtlb_valids : tlb_valids_t;
signal dtlb_tags : tlb_tags_t;
signal dtlb_ptes : tlb_ptes_t;
attribute ram_style of dtlb_tags : signal is "distributed";
attribute ram_style of dtlb_ptes : signal is "distributed";
signal r0 : Loadstore1ToDcacheType;
-- Type of operation on a "valid" input
type op_t is (OP_NONE,
OP_LOAD_HIT, -- Cache hit on load
OP_LOAD_MISS, -- Load missing cache
OP_LOAD_NC, -- Non-cachable load
OP_BAD, -- BAD: Cache hit on NC load/store
OP_STORE_HIT, -- Store hitting cache
OP_STORE_MISS); -- Store missing cache
-- Cache state machine
type state_t is (IDLE, -- Normal load hit processing
RELOAD_WAIT_ACK, -- Cache reload wait ack
FINISH_LD_MISS, -- Extra cycle after load miss
STORE_WAIT_ACK, -- Store wait ack
NC_LOAD_WAIT_ACK);-- Non-cachable load wait ack
--
-- Dcache operations:
--
-- In order to make timing, we use the BRAMs with an output buffer,
-- 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.
--
-- All other operations are handled via stalling in the first stage.
--
-- The second stage can thus complete a hit at the same time as the
-- first stage emits a stall for a complex op.
--
-- First stage register, contains state for stage 1 of load hits
-- and for the state machine used by all other operations
--
type reg_stage_1_t is record
-- Latch the complete request from ls1
req : Loadstore1ToDcacheType;
-- Cache hit state
hit_way : way_t;
hit_load_valid : std_ulogic;
-- Data buffer for "slow" read ops (load miss and NC loads).
slow_data : std_ulogic_vector(63 downto 0);
slow_valid : std_ulogic;
-- Signal to complete a failed stcx.
stcx_fail : std_ulogic;
-- Cache miss state (reload state machine)
state : state_t;
wb : wishbone_master_out;
store_way : way_t;
store_row : row_t;
store_index : index_t;
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
-- Signals to complete with error
error_done : std_ulogic;
tlb_miss : std_ulogic;
-- completion signal for tlbie
tlbie_done : std_ulogic;
end record;
signal r1 : reg_stage_1_t;
-- Reservation information
--
type reservation_t is record
valid : std_ulogic;
addr : std_ulogic_vector(63 downto LINE_OFF_BITS);
end record;
signal reservation : reservation_t;
-- Async signals on incoming request
signal req_index : index_t;
signal req_row : row_t;
signal req_hit_way : way_t;
signal req_tag : cache_tag_t;
signal req_op : op_t;
signal req_data : std_ulogic_vector(63 downto 0);
signal req_laddr : std_ulogic_vector(63 downto 0);
dcache: Trim one cycle from the load hit path Currently we don't get the result from a load that hits in the dcache until the fourth cycle after the instruction was presented to loadstore1. This trims this back to 3 cycles by taking the low order bits of the address generated in loadstore1 into dcache directly (not via the output register of loadstore1) and using them to address the read port of the dcache data RAM. We use the lower 12 address bits here in the expectation that any reasonable data cache design will have a set size of 4kB or less in order to avoid the aliasing problems that can arise with a virtually-indexed physically-tagged cache if the set size is greater than the smallest page size provided by the MMU. With this we can get rid of r2 and drive the signals going to writeback from r1, since the load hit data is now available one cycle earlier. We need a multiplexer on the read address of the data cache RAM in order to handle the second doubleword of an unaligned access. One small complication is that we now need an extra cycle in the case of an unaligned load which misses in the data cache and which reads the 2nd-last and last doublewords of a cache line. This is the reason for the PRE_NEXT_DWORD state; if we just go straight to NEXT_DWORD then we end up having the write of the last doubleword of the cache line and the read of that same doubleword occurring in the same cycle, which means we read stale data rather than the just-fetched data. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
signal early_req_row : row_t;
signal cancel_store : std_ulogic;
signal set_rsrv : std_ulogic;
signal clear_rsrv : std_ulogic;
-- Cache RAM interface
type cache_ram_out_t is array(way_t) of cache_row_t;
signal cache_out : cache_ram_out_t;
-- PLRU output interface
type plru_out_t is array(index_t) of std_ulogic_vector(WAY_BITS-1 downto 0);
signal plru_victim : plru_out_t;
signal replace_way : way_t;
-- Wishbone read/write/cache write formatting signals
signal bus_sel : std_ulogic_vector(7 downto 0);
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
-- TLB signals
signal tlb_tag_way : tlb_way_tags_t;
signal tlb_pte_way : tlb_way_ptes_t;
signal tlb_valid_way : tlb_way_valids_t;
signal tlb_req_index : tlb_index_t;
signal tlb_hit : std_ulogic;
signal tlb_hit_way : tlb_way_t;
signal pte : tlb_pte_t;
signal ra : std_ulogic_vector(REAL_ADDR_BITS - 1 downto 0);
signal valid_ra : std_ulogic;
-- TLB PLRU output interface
type tlb_plru_out_t is array(tlb_index_t) of std_ulogic_vector(TLB_WAY_BITS-1 downto 0);
signal tlb_plru_victim : tlb_plru_out_t;
--
-- Helper functions to decode incoming requests
--
-- Return the cache line index (tag index) for an address
function get_index(addr: std_ulogic_vector(63 downto 0)) return index_t is
begin
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
return to_integer(unsigned(addr(SET_SIZE_BITS - 1 downto LINE_OFF_BITS)));
end;
-- Return the cache row index (data memory) for an address
function get_row(addr: std_ulogic_vector(63 downto 0)) return row_t is
begin
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
return to_integer(unsigned(addr(SET_SIZE_BITS - 1 downto ROW_OFF_BITS)));
end;
-- Returns whether this is the last row of a line
function is_last_row_addr(addr: wishbone_addr_type) return boolean is
constant ones : std_ulogic_vector(ROW_LINEBITS-1 downto 0) := (others => '1');
begin
return addr(LINE_OFF_BITS-1 downto ROW_OFF_BITS) = ones;
end;
-- Returns whether this is the last row of a line
function is_last_row(row: row_t) return boolean is
variable row_v : std_ulogic_vector(ROW_BITS-1 downto 0);
constant ones : std_ulogic_vector(ROW_LINEBITS-1 downto 0) := (others => '1');
begin
row_v := std_ulogic_vector(to_unsigned(row, ROW_BITS));
return row_v(ROW_LINEBITS-1 downto 0) = ones;
end;
-- Return the address of the next row in the current cache line
function next_row_addr(addr: wishbone_addr_type) return std_ulogic_vector is
variable row_idx : std_ulogic_vector(ROW_LINEBITS-1 downto 0);
variable result : wishbone_addr_type;
begin
-- Is there no simpler way in VHDL to generate that 3 bits adder ?
row_idx := addr(LINE_OFF_BITS-1 downto ROW_OFF_BITS);
row_idx := std_ulogic_vector(unsigned(row_idx) + 1);
result := addr;
result(LINE_OFF_BITS-1 downto ROW_OFF_BITS) := row_idx;
return result;
end;
-- Return the next row in the current cache line. We use a dedicated
-- function in order to limit the size of the generated adder to be
-- only the bits within a cache line (3 bits with default settings)
--
function next_row(row: row_t) return row_t is
variable row_v : std_ulogic_vector(ROW_BITS-1 downto 0);
variable row_idx : std_ulogic_vector(ROW_LINEBITS-1 downto 0);
variable result : std_ulogic_vector(ROW_BITS-1 downto 0);
begin
row_v := std_ulogic_vector(to_unsigned(row, ROW_BITS));
row_idx := row_v(ROW_LINEBITS-1 downto 0);
row_v(ROW_LINEBITS-1 downto 0) := std_ulogic_vector(unsigned(row_idx) + 1);
return to_integer(unsigned(row_v));
end;
-- Get the tag value from the address
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
function get_tag(addr: std_ulogic_vector(REAL_ADDR_BITS - 1 downto 0)) return cache_tag_t is
begin
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
return addr(REAL_ADDR_BITS - 1 downto SET_SIZE_BITS);
end;
-- Read a tag from a tag memory row
function read_tag(way: way_t; tagset: cache_tags_set_t) return cache_tag_t is
begin
return tagset((way+1) * TAG_BITS - 1 downto way * TAG_BITS);
end;
-- Write a tag to tag memory row
procedure write_tag(way: in way_t; tagset: inout cache_tags_set_t;
tag: cache_tag_t) is
begin
tagset((way+1) * TAG_BITS - 1 downto way * TAG_BITS) := tag;
end;
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
-- Read a TLB tag from a TLB tag memory row
function read_tlb_tag(way: tlb_way_t; tags: tlb_way_tags_t) return tlb_tag_t is
variable j : integer;
begin
j := way * TLB_EA_TAG_BITS;
return tags(j + TLB_EA_TAG_BITS - 1 downto j);
end;
-- Write a TLB tag to a TLB tag memory row
procedure write_tlb_tag(way: tlb_way_t; tags: inout tlb_way_tags_t;
tag: tlb_tag_t) is
variable j : integer;
begin
j := way * TLB_EA_TAG_BITS;
tags(j + TLB_EA_TAG_BITS - 1 downto j) := tag;
end;
-- Read a PTE from a TLB PTE memory row
function read_tlb_pte(way: tlb_way_t; ptes: tlb_way_ptes_t) return tlb_pte_t is
variable j : integer;
begin
j := way * TLB_PTE_BITS;
return ptes(j + TLB_PTE_BITS - 1 downto j);
end;
procedure write_tlb_pte(way: tlb_way_t; ptes: inout tlb_way_ptes_t; newpte: tlb_pte_t) is
variable j : integer;
begin
j := way * TLB_PTE_BITS;
ptes(j + TLB_PTE_BITS - 1 downto j) := newpte;
end;
begin
assert LINE_SIZE mod ROW_SIZE = 0 report "LINE_SIZE not multiple of ROW_SIZE" severity FAILURE;
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;
assert ispow2(ROW_PER_LINE) report "ROW_PER_LINE not power of 2" severity FAILURE;
assert (ROW_BITS = INDEX_BITS + ROW_LINEBITS)
report "geometry bits don't add up" severity FAILURE;
assert (LINE_OFF_BITS = ROW_OFF_BITS + ROW_LINEBITS)
report "geometry bits don't add up" severity FAILURE;
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
assert (REAL_ADDR_BITS = TAG_BITS + INDEX_BITS + LINE_OFF_BITS)
report "geometry bits don't add up" severity FAILURE;
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
assert (REAL_ADDR_BITS = TAG_BITS + ROW_BITS + ROW_OFF_BITS)
report "geometry bits don't add up" severity FAILURE;
assert (64 = wishbone_data_bits)
report "Can't yet handle a wishbone width that isn't 64-bits" severity FAILURE;
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
-- Latch the request in r0 as long as we're not stalling
stage_0 : process(clk)
begin
if rising_edge(clk) then
if rst = '1' then
r0.valid <= '0';
elsif stall_out = '0' then
r0 <= d_in;
end if;
end if;
end process;
-- TLB
-- Operates in the second cycle on the request latched in r0.
-- TLB updates write the entry at the end of the second cycle.
tlb_read : process(clk)
variable index : tlb_index_t;
begin
if rising_edge(clk) then
if stall_out = '1' then
-- keep reading the same thing while stalled
index := tlb_req_index;
else
index := to_integer(unsigned(d_in.addr(TLB_LG_PGSZ + TLB_SET_BITS - 1
downto TLB_LG_PGSZ)));
end if;
tlb_valid_way <= dtlb_valids(index);
tlb_tag_way <= dtlb_tags(index);
tlb_pte_way <= dtlb_ptes(index);
end if;
end process;
-- Generate TLB PLRUs
maybe_tlb_plrus: if TLB_NUM_WAYS > 1 generate
begin
tlb_plrus: for i in 0 to TLB_SET_SIZE - 1 generate
-- TLB PLRU interface
signal tlb_plru_acc : std_ulogic_vector(TLB_WAY_BITS-1 downto 0);
signal tlb_plru_acc_en : std_ulogic;
signal tlb_plru_out : std_ulogic_vector(TLB_WAY_BITS-1 downto 0);
begin
tlb_plru : entity work.plru
generic map (
BITS => TLB_WAY_BITS
)
port map (
clk => clk,
rst => rst,
acc => tlb_plru_acc,
acc_en => tlb_plru_acc_en,
lru => tlb_plru_out
);
process(tlb_req_index, tlb_hit, tlb_hit_way, tlb_plru_out)
begin
-- PLRU interface
if tlb_hit = '1' and tlb_req_index = i then
tlb_plru_acc_en <= '1';
else
tlb_plru_acc_en <= '0';
end if;
tlb_plru_acc <= std_ulogic_vector(to_unsigned(tlb_hit_way, TLB_WAY_BITS));
tlb_plru_victim(i) <= tlb_plru_out;
end process;
end generate;
end generate;
tlb_search : process(all)
variable hitway : tlb_way_t;
variable hit : std_ulogic;
variable eatag : tlb_tag_t;
begin
tlb_req_index <= to_integer(unsigned(r0.addr(TLB_LG_PGSZ + TLB_SET_BITS - 1
downto TLB_LG_PGSZ)));
hitway := 0;
hit := '0';
eatag := r0.addr(63 downto TLB_LG_PGSZ + TLB_SET_BITS);
for i in tlb_way_t loop
if tlb_valid_way(i) = '1' and
read_tlb_tag(i, tlb_tag_way) = eatag then
hitway := i;
hit := '1';
end if;
end loop;
tlb_hit <= hit and r0.valid;
tlb_hit_way <= hitway;
pte <= read_tlb_pte(hitway, tlb_pte_way);
valid_ra <= tlb_hit or not r0.virt_mode;
if r0.virt_mode = '1' then
ra <= pte(REAL_ADDR_BITS - 1 downto TLB_LG_PGSZ) &
r0.addr(TLB_LG_PGSZ - 1 downto 0);
else
ra <= r0.addr(REAL_ADDR_BITS - 1 downto 0);
end if;
end process;
tlb_update : process(clk)
variable tlbie : std_ulogic;
variable tlbia : std_ulogic;
variable tlbwe : std_ulogic;
variable repl_way : tlb_way_t;
variable eatag : tlb_tag_t;
variable tagset : tlb_way_tags_t;
variable pteset : tlb_way_ptes_t;
begin
if rising_edge(clk) then
tlbie := '0';
tlbia := '0';
tlbwe := '0';
if r0.valid = '1' and stall_out = '0' and r0.tlbie = '1' then
if r0.addr(11 downto 10) /= "00" then
tlbia := '1';
elsif r0.addr(9) = '1' then
tlbwe := '1';
else
tlbie := '1';
end if;
end if;
if rst = '1' or tlbia = '1' then
-- clear all valid bits at once
for i in tlb_index_t loop
dtlb_valids(i) <= (others => '0');
end loop;
elsif tlbie = '1' then
if tlb_hit = '1' then
dtlb_valids(tlb_req_index)(tlb_hit_way) <= '0';
end if;
elsif tlbwe = '1' then
if tlb_hit = '1' then
repl_way := tlb_hit_way;
else
repl_way := to_integer(unsigned(tlb_plru_victim(tlb_req_index)));
end if;
eatag := r0.addr(63 downto TLB_LG_PGSZ + TLB_SET_BITS);
tagset := tlb_tag_way;
write_tlb_tag(repl_way, tagset, eatag);
dtlb_tags(tlb_req_index) <= tagset;
pteset := tlb_pte_way;
write_tlb_pte(repl_way, pteset, r0.data);
dtlb_ptes(tlb_req_index) <= pteset;
dtlb_valids(tlb_req_index)(repl_way) <= '1';
end if;
end if;
end process;
-- Generate PLRUs
maybe_plrus: if NUM_WAYS > 1 generate
begin
plrus: for i in 0 to NUM_LINES-1 generate
-- PLRU interface
signal plru_acc : std_ulogic_vector(WAY_BITS-1 downto 0);
signal plru_acc_en : std_ulogic;
signal plru_out : std_ulogic_vector(WAY_BITS-1 downto 0);
begin
plru : entity work.plru
generic map (
BITS => WAY_BITS
)
port map (
clk => clk,
rst => rst,
acc => plru_acc,
acc_en => plru_acc_en,
lru => plru_out
);
process(req_index, req_op, req_hit_way, plru_out)
begin
-- PLRU interface
if (req_op = OP_LOAD_HIT or
req_op = OP_STORE_HIT) and req_index = i then
plru_acc_en <= '1';
else
plru_acc_en <= '0';
end if;
plru_acc <= std_ulogic_vector(to_unsigned(req_hit_way, WAY_BITS));
plru_victim(i) <= plru_out;
end process;
end generate;
end generate;
-- Cache request parsing and hit detection
dcache_request : process(all)
variable is_hit : std_ulogic;
variable hit_way : way_t;
variable op : op_t;
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
variable opsel : std_ulogic_vector(2 downto 0);
variable go : std_ulogic;
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
variable s_hit : std_ulogic;
variable s_tag : cache_tag_t;
variable s_pte : tlb_pte_t;
variable s_ra : std_ulogic_vector(REAL_ADDR_BITS - 1 downto 0);
variable hit_set : std_ulogic_vector(TLB_NUM_WAYS - 1 downto 0);
variable hit_way_set : hit_way_set_t;
begin
-- Extract line, row and tag from request
req_index <= get_index(r0.addr);
req_row <= get_row(r0.addr);
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
req_tag <= get_tag(ra);
-- Only do anything if not being stalled by stage 1
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
go := r0.valid and not stall_out and not r0.tlbie;
-- Calculate address of beginning of cache line, will be
-- used for cache miss processing if needed
--
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
req_laddr <= (63 downto REAL_ADDR_BITS => '0') &
ra(REAL_ADDR_BITS - 1 downto LINE_OFF_BITS) &
(LINE_OFF_BITS-1 downto 0 => '0');
-- Test if pending request is a hit on any way
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
-- In order to make timing in virtual mode, when we are using the TLB,
-- we compare each way with each of the real addresses from each way of
-- the TLB, and then decide later which match to use.
hit_way := 0;
is_hit := '0';
if r0.virt_mode = '1' then
for j in tlb_way_t loop
hit_way_set(j) := 0;
s_hit := '0';
s_pte := read_tlb_pte(j, tlb_pte_way);
s_ra := s_pte(REAL_ADDR_BITS - 1 downto TLB_LG_PGSZ) &
r0.addr(TLB_LG_PGSZ - 1 downto 0);
s_tag := get_tag(s_ra);
for i in way_t loop
if go = '1' and cache_valids(req_index)(i) = '1' and
read_tag(i, cache_tags(req_index)) = s_tag and
tlb_valid_way(j) = '1' then
hit_way_set(j) := i;
s_hit := '1';
end if;
end loop;
hit_set(j) := s_hit;
end loop;
if tlb_hit = '1' then
is_hit := hit_set(tlb_hit_way);
hit_way := hit_way_set(tlb_hit_way);
end if;
else
s_tag := get_tag(r0.addr(REAL_ADDR_BITS - 1 downto 0));
for i in way_t loop
if go = '1' and cache_valids(req_index)(i) = '1' and
read_tag(i, cache_tags(req_index)) = s_tag then
hit_way := i;
is_hit := '1';
end if;
end loop;
end if;
-- The way that matched on a hit
req_hit_way <= hit_way;
-- The way to replace on a miss
replace_way <= to_integer(unsigned(plru_victim(req_index)));
-- Combine the request and cache his status to decide what
-- operation needs to be done
--
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
op := OP_NONE;
if go = '1' then
if valid_ra = '1' then
opsel := r0.load & r0.nc & is_hit;
case opsel is
when "101" => op := OP_LOAD_HIT;
when "100" => op := OP_LOAD_MISS;
when "110" => op := OP_LOAD_NC;
when "001" => op := OP_STORE_HIT;
when "000" => op := OP_STORE_MISS;
when "010" => op := OP_STORE_MISS;
when "011" => op := OP_BAD;
when "111" => op := OP_BAD;
when others => op := OP_NONE;
end case;
else
op := OP_BAD;
end if;
end if;
req_op <= op;
-- Version of the row number that is valid one cycle earlier
dcache: Trim one cycle from the load hit path Currently we don't get the result from a load that hits in the dcache until the fourth cycle after the instruction was presented to loadstore1. This trims this back to 3 cycles by taking the low order bits of the address generated in loadstore1 into dcache directly (not via the output register of loadstore1) and using them to address the read port of the dcache data RAM. We use the lower 12 address bits here in the expectation that any reasonable data cache design will have a set size of 4kB or less in order to avoid the aliasing problems that can arise with a virtually-indexed physically-tagged cache if the set size is greater than the smallest page size provided by the MMU. With this we can get rid of r2 and drive the signals going to writeback from r1, since the load hit data is now available one cycle earlier. We need a multiplexer on the read address of the data cache RAM in order to handle the second doubleword of an unaligned access. One small complication is that we now need an extra cycle in the case of an unaligned load which misses in the data cache and which reads the 2nd-last and last doublewords of a cache line. This is the reason for the PRE_NEXT_DWORD state; if we just go straight to NEXT_DWORD then we end up having the write of the last doubleword of the cache line and the read of that same doubleword occurring in the same cycle, which means we read stale data rather than the just-fetched data. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
-- in the cases where we need to read the cache data BRAM.
-- If we're stalling then we need to keep reading the last
-- row requested.
if stall_out = '0' then
early_req_row <= get_row(d_in.addr);
dcache: Trim one cycle from the load hit path Currently we don't get the result from a load that hits in the dcache until the fourth cycle after the instruction was presented to loadstore1. This trims this back to 3 cycles by taking the low order bits of the address generated in loadstore1 into dcache directly (not via the output register of loadstore1) and using them to address the read port of the dcache data RAM. We use the lower 12 address bits here in the expectation that any reasonable data cache design will have a set size of 4kB or less in order to avoid the aliasing problems that can arise with a virtually-indexed physically-tagged cache if the set size is greater than the smallest page size provided by the MMU. With this we can get rid of r2 and drive the signals going to writeback from r1, since the load hit data is now available one cycle earlier. We need a multiplexer on the read address of the data cache RAM in order to handle the second doubleword of an unaligned access. One small complication is that we now need an extra cycle in the case of an unaligned load which misses in the data cache and which reads the 2nd-last and last doublewords of a cache line. This is the reason for the PRE_NEXT_DWORD state; if we just go straight to NEXT_DWORD then we end up having the write of the last doubleword of the cache line and the read of that same doubleword occurring in the same cycle, which means we read stale data rather than the just-fetched data. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
else
early_req_row <= req_row;
dcache: Trim one cycle from the load hit path Currently we don't get the result from a load that hits in the dcache until the fourth cycle after the instruction was presented to loadstore1. This trims this back to 3 cycles by taking the low order bits of the address generated in loadstore1 into dcache directly (not via the output register of loadstore1) and using them to address the read port of the dcache data RAM. We use the lower 12 address bits here in the expectation that any reasonable data cache design will have a set size of 4kB or less in order to avoid the aliasing problems that can arise with a virtually-indexed physically-tagged cache if the set size is greater than the smallest page size provided by the MMU. With this we can get rid of r2 and drive the signals going to writeback from r1, since the load hit data is now available one cycle earlier. We need a multiplexer on the read address of the data cache RAM in order to handle the second doubleword of an unaligned access. One small complication is that we now need an extra cycle in the case of an unaligned load which misses in the data cache and which reads the 2nd-last and last doublewords of a cache line. This is the reason for the PRE_NEXT_DWORD state; if we just go straight to NEXT_DWORD then we end up having the write of the last doubleword of the cache line and the read of that same doubleword occurring in the same cycle, which means we read stale data rather than the just-fetched data. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
end if;
end process;
-- Wire up wishbone request latch out of stage 1
wishbone_out <= r1.wb;
-- Generate stalls from stage 1 state machine
stall_out <= '1' when r1.state /= IDLE else '0';
-- Handle load-with-reservation and store-conditional instructions
reservation_comb: process(all)
begin
cancel_store <= '0';
set_rsrv <= '0';
clear_rsrv <= '0';
if stall_out = '0' and r0.valid = '1' and r0.reserve = '1' then
-- XXX generate alignment interrupt if address is not aligned
-- XXX or if r0.nc = '1'
if r0.load = '1' then
-- load with reservation
set_rsrv <= '1';
else
-- store conditional
clear_rsrv <= '1';
if reservation.valid = '0' or
r0.addr(63 downto LINE_OFF_BITS) /= reservation.addr then
cancel_store <= '1';
end if;
end if;
end if;
end process;
reservation_reg: process(clk)
begin
if rising_edge(clk) then
if rst = '1' or clear_rsrv = '1' then
reservation.valid <= '0';
elsif set_rsrv = '1' then
reservation.valid <= '1';
reservation.addr <= r0.addr(63 downto LINE_OFF_BITS);
end if;
end if;
end process;
-- Return data for loads & completion control logic
--
writeback_control: process(all)
begin
-- The mux on d_out.data defaults to the normal load hit case.
d_out.valid <= '0';
d_out.data <= cache_out(r1.hit_way);
d_out.store_done <= '0';
dcache: Implement data TLB This adds a TLB to dcache, providing the ability to translate addresses for loads and stores. No protection mechanism has been implemented yet. The MSR_DR bit controls whether addresses are translated through the TLB. The TLB is a fixed-pagesize, set-associative cache. Currently the page size is 4kB and the TLB is 2-way set associative with 64 entries per set. This implements the tlbie instruction. RB bits 10 and 11 control whether the whole TLB is invalidated (if either bit is 1) or just a single entry corresponding to the effective page number in bits 12-63 of RB. As an extension until we get a hardware page table walk, a tlbie instruction with RB bits 9-11 set to 001 will load an entry into the TLB. The TLB entry value is in RS in the format of a radix PTE. Currently there is no proper handling of TLB misses. The load or store will not be performed but no interrupt is generated. In order to make timing at 100MHz on the Arty A7-100, we compare the real address from each way of the TLB with the tag from each way of the cache in parallel (requiring # TLB ways * # cache ways comparators). Then the result is selected based on which way hit in the TLB. That avoids a timing path going through the TLB EA comparators, the multiplexer that selects the RA, and the cache tag comparators. The hack where addresses of the form 0xc------- are marked as cache-inhibited is kept for now but restricted to real-mode accesses. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
d_out.error <= '0';
d_out.tlb_miss <= '0';
-- 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
--
-- Note: the load hit is delayed by one cycle. However it can still
-- not collide with r.slow_valid (well unless I miscalculated) because
-- slow_valid can only be set on a subsequent request and not on its
-- first cycle (the state machine must have advanced), which makes
-- slow_valid at least 2 cycles from the previous hit_load_valid.
--
-- Sanity: Only one of these must be set in any given cycle
assert (r1.slow_valid and r1.stcx_fail) /= '1' report
"unexpected slow_valid collision with stcx_fail"
severity FAILURE;
dcache: Trim one cycle from the load hit path Currently we don't get the result from a load that hits in the dcache until the fourth cycle after the instruction was presented to loadstore1. This trims this back to 3 cycles by taking the low order bits of the address generated in loadstore1 into dcache directly (not via the output register of loadstore1) and using them to address the read port of the dcache data RAM. We use the lower 12 address bits here in the expectation that any reasonable data cache design will have a set size of 4kB or less in order to avoid the aliasing problems that can arise with a virtually-indexed physically-tagged cache if the set size is greater than the smallest page size provided by the MMU. With this we can get rid of r2 and drive the signals going to writeback from r1, since the load hit data is now available one cycle earlier. We need a multiplexer on the read address of the data cache RAM in order to handle the second doubleword of an unaligned access. One small complication is that we now need an extra cycle in the case of an unaligned load which misses in the data cache and which reads the 2nd-last and last doublewords of a cache line. This is the reason for the PRE_NEXT_DWORD state; if we just go straight to NEXT_DWORD then we end up having the write of the last doubleword of the cache line and the read of that same doubleword occurring in the same cycle, which means we read stale data rather than the just-fetched data. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
assert ((r1.slow_valid or r1.stcx_fail) and r1.hit_load_valid) /= '1' report
"unexpected hit_load_delayed collision with slow_valid"
severity FAILURE;
dcache: Trim one cycle from the load hit path Currently we don't get the result from a load that hits in the dcache until the fourth cycle after the instruction was presented to loadstore1. This trims this back to 3 cycles by taking the low order bits of the address generated in loadstore1 into dcache directly (not via the output register of loadstore1) and using them to address the read port of the dcache data RAM. We use the lower 12 address bits here in the expectation that any reasonable data cache design will have a set size of 4kB or less in order to avoid the aliasing problems that can arise with a virtually-indexed physically-tagged cache if the set size is greater than the smallest page size provided by the MMU. With this we can get rid of r2 and drive the signals going to writeback from r1, since the load hit data is now available one cycle earlier. We need a multiplexer on the read address of the data cache RAM in order to handle the second doubleword of an unaligned access. One small complication is that we now need an extra cycle in the case of an unaligned load which misses in the data cache and which reads the 2nd-last and last doublewords of a cache line. This is the reason for the PRE_NEXT_DWORD state; if we just go straight to NEXT_DWORD then we end up having the write of the last doubleword of the cache line and the read of that same doubleword occurring in the same cycle, which means we read stale data rather than the just-fetched data. Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
-- Load hit case is the standard path
if r1.hit_load_valid = '1' then
report "completing load hit";
d_out.valid <= '1';
end if;