1
0

clean up caches

- remove incompatible blocking D$
- remove direct-mapped nonblocking cache
This commit is contained in:
Andrew Waterman 2012-02-12 20:32:06 -08:00
parent 08b6517a23
commit 25ecfb9bbc
3 changed files with 40 additions and 793 deletions

View File

@ -1,517 +0,0 @@
package Top {
import Chisel._
import Node._;
import Constants._;
import scala.math._;
// interface between D$ and processor/DTLB
class ioDmem(view: List[String] = null) extends Bundle(view) {
val req_kill = Bool(INPUT);
val req_val = Bool(INPUT);
val req_rdy = Bool(OUTPUT);
val req_cmd = Bits(4, INPUT);
val req_type = Bits(3, INPUT);
val req_idx = Bits(PGIDX_BITS, INPUT);
val req_ppn = Bits(PPN_BITS, INPUT);
val req_data = Bits(64, INPUT);
val req_tag = Bits(DCACHE_TAG_BITS, INPUT);
val xcpt_ma_ld = Bool(OUTPUT); // misaligned load
val xcpt_ma_st = Bool(OUTPUT); // misaligned store
val resp_miss = Bool(OUTPUT);
val resp_nack = Bool(OUTPUT);
val resp_val = Bool(OUTPUT);
val resp_replay = Bool(OUTPUT);
val resp_data = Bits(64, OUTPUT);
val resp_data_subword = Bits(64, OUTPUT);
val resp_tag = Bits(DCACHE_TAG_BITS, OUTPUT);
}
// interface between D$ and next level in memory hierarchy
class ioDCache(view: List[String] = null) extends Bundle(view) {
val req_addr = UFix(PADDR_BITS - OFFSET_BITS, INPUT);
val req_tag = UFix(DMEM_TAG_BITS, INPUT);
val req_val = Bool(INPUT);
val req_rdy = Bool(OUTPUT);
val req_wdata = Bits(MEM_DATA_BITS, INPUT);
val req_rw = Bool(INPUT);
val resp_data = Bits(MEM_DATA_BITS, OUTPUT);
val resp_tag = Bits(DMEM_TAG_BITS, OUTPUT);
val resp_val = Bool(OUTPUT);
}
class ioDCacheDM extends Bundle() {
val cpu = new ioDmem();
val mem = new ioDCache().flip();
}
class ioDCacheHella extends Bundle() {
val cpu = new ioDmem();
val mem = new ioDCache().flip();
}
class rocketDCacheStoreGen extends Component {
val io = new Bundle {
val req_type = Bits(3, INPUT);
val req_addr_lsb = Bits(3, INPUT);
val req_data = Bits(64, INPUT);
val store_wmask = Bits(64, OUTPUT);
val store_data = Bits(64, OUTPUT);
}
// generate write mask and store data signals based on store type and address LSBs
val wmask_b =
Mux(io.req_addr_lsb === UFix(0, 3), Bits("b0000_0001", 8),
Mux(io.req_addr_lsb === UFix(1, 3), Bits("b0000_0010", 8),
Mux(io.req_addr_lsb === UFix(2, 3), Bits("b0000_0100", 8),
Mux(io.req_addr_lsb === UFix(3, 3), Bits("b0000_1000", 8),
Mux(io.req_addr_lsb === UFix(4, 3), Bits("b0001_0000", 8),
Mux(io.req_addr_lsb === UFix(5, 3), Bits("b0010_0000", 8),
Mux(io.req_addr_lsb === UFix(6, 3), Bits("b0100_0000", 8),
Mux(io.req_addr_lsb === UFix(7, 3), Bits("b1000_0000", 8),
UFix(0, 8)))))))));
val wmask_h =
Mux(io.req_addr_lsb(2,1) === UFix(0, 2), Bits("b0000_0011", 8),
Mux(io.req_addr_lsb(2,1) === UFix(1, 2), Bits("b0000_1100", 8),
Mux(io.req_addr_lsb(2,1) === UFix(2, 2), Bits("b0011_0000", 8),
Mux(io.req_addr_lsb(2,1) === UFix(3, 2), Bits("b1100_0000", 8),
UFix(0, 8)))));
val wmask_w =
Mux(io.req_addr_lsb(2) === UFix(0, 1), Bits("b0000_1111", 8),
Mux(io.req_addr_lsb(2) === UFix(1, 1), Bits("b1111_0000", 8),
UFix(0, 8)));
val wmask_d =
Bits("b1111_1111", 8);
val store_wmask_byte =
Mux(io.req_type === MT_B, wmask_b,
Mux(io.req_type === MT_H, wmask_h,
Mux(io.req_type === MT_W, wmask_w,
Mux(io.req_type === MT_D, wmask_d,
UFix(0, 8)))));
val store_wmask_d = Cat(Fill(8, store_wmask_byte(7)),
Fill(8, store_wmask_byte(6)),
Fill(8, store_wmask_byte(5)),
Fill(8, store_wmask_byte(4)),
Fill(8, store_wmask_byte(3)),
Fill(8, store_wmask_byte(2)),
Fill(8, store_wmask_byte(1)),
Fill(8, store_wmask_byte(0)));
io.store_wmask := store_wmask_d;
io.store_data :=
Mux(io.req_type === MT_B, Fill(8, io.req_data( 7,0)),
Mux(io.req_type === MT_H, Fill(4, io.req_data(15,0)),
Mux(io.req_type === MT_W, Fill(2, io.req_data(31,0)),
Mux(io.req_type === MT_D, io.req_data,
UFix(0, 64)))));
}
// state machine to flush (write back dirty lines, invalidate clean ones) the D$
class rocketDCacheDM_flush(lines: Int) extends Component {
val io = new ioDCacheDM();
val dcache = new rocketDCacheDM(lines);
val addrbits = PADDR_BITS;
val indexbits = ceil(log10(lines)/log10(2)).toInt;
val offsetbits = 6;
val tagmsb = addrbits - 1;
val taglsb = indexbits+offsetbits;
val tagbits = tagmsb-taglsb+1;
val indexmsb = taglsb-1;
val indexlsb = offsetbits;
val offsetmsb = indexlsb-1;
val offsetlsb = 3;
val flush_count = Reg(resetVal = UFix(0, indexbits));
val flush_resp_count = Reg(resetVal = UFix(0, indexbits));
val flushing = Reg(resetVal = Bool(false));
val flush_waiting = Reg(resetVal = Bool(false));
val r_cpu_req_tag = Reg() { Bits() }
when (io.cpu.req_val && io.cpu.req_rdy && (io.cpu.req_cmd === M_FLA))
{
r_cpu_req_tag := io.cpu.req_tag;
flushing := Bool(true);
flush_waiting := Bool(true);
}
when (dcache.io.cpu.req_rdy && (flush_count === ~Bits(0, indexbits))) {
flushing := Bool(false);
}
when (dcache.io.cpu.resp_val && (dcache.io.cpu.resp_tag === r_cpu_req_tag) && (flush_resp_count === ~Bits(0, indexbits))) {
flush_waiting := Bool(false);
}
when (flushing && dcache.io.cpu.req_rdy) {
flush_count := flush_count + UFix(1,1);
}
when (flush_waiting && dcache.io.cpu.resp_val && (dcache.io.cpu.resp_tag === r_cpu_req_tag)) {
flush_resp_count := flush_resp_count + UFix(1,1);
}
dcache.io.cpu.req_val := (io.cpu.req_val && (io.cpu.req_cmd != M_FLA) && !flush_waiting) || flushing;
dcache.io.cpu.req_cmd := Mux(flushing, M_FLA, io.cpu.req_cmd);
dcache.io.cpu.req_idx := Mux(flushing, Cat(flush_count, Bits(0,offsetbits)), io.cpu.req_idx);
dcache.io.cpu.req_ppn := Mux(flushing, UFix(0,PPN_BITS), io.cpu.req_ppn);
dcache.io.cpu.req_tag := Mux(flushing, r_cpu_req_tag, io.cpu.req_tag);
dcache.io.cpu.req_type := io.cpu.req_type;
dcache.io.cpu.req_data <> io.cpu.req_data;
dcache.io.cpu.req_kill := io.cpu.req_kill && !flush_waiting;
dcache.io.mem <> io.mem;
io.cpu.xcpt_ma_ld := dcache.io.cpu.xcpt_ma_ld;
io.cpu.xcpt_ma_st := dcache.io.cpu.xcpt_ma_st;
io.cpu.req_rdy := dcache.io.cpu.req_rdy && !flush_waiting;
io.cpu.resp_miss := dcache.io.cpu.resp_miss;
io.cpu.resp_nack := dcache.io.cpu.resp_nack;
io.cpu.resp_data := dcache.io.cpu.resp_data;
io.cpu.resp_tag := dcache.io.cpu.resp_tag;
io.cpu.resp_val := dcache.io.cpu.resp_val &
!(flush_waiting && (io.cpu.resp_tag === r_cpu_req_tag) && (flush_count != ~Bits(0, addrbits)));
}
class rocketDCacheDM(lines: Int) extends Component {
val io = new ioDCacheDM();
val addrbits = PADDR_BITS;
val indexbits = ceil(log10(lines)/log10(2)).toInt;
val offsetbits = 6; // 64 byte cache lines = 2^6 bytes
val tagmsb = PADDR_BITS-1;
val taglsb = indexbits+offsetbits;
val tagbits = tagmsb-taglsb+1;
val indexmsb = taglsb-1;
val indexlsb = offsetbits;
val offsetmsb = indexlsb-1;
val offsetlsb = 3;
val s_reset :: s_ready :: s_replay_load :: s_write_amo :: s_start_writeback :: s_writeback :: s_req_refill :: s_refill :: s_resolve_miss :: Nil = Enum(9) { UFix() };
val state = Reg(resetVal = s_reset);
// idx arrives one clock cycle prior to ppn b/c of DTLB
val r_cpu_req_idx = Reg(resetVal = Bits(0, PGIDX_BITS));
val r_cpu_req_ppn = Reg(resetVal = Bits(0, PPN_BITS));
val r_cpu_req_val = Reg(resetVal = Bool(false));
val r_cpu_req_cmd = Reg(resetVal = Bits(0,4));
val r_cpu_req_type = Reg(resetVal = Bits(0,3));
val r_cpu_req_tag = Reg() { Bits() }
val r_cpu_resp_val = Reg(resetVal = Bool(false));
val r_amo_data = Reg(resetVal = Bits(0,64));
val p_store_data = Reg(resetVal = Bits(0,64));
val p_store_idx = Reg(resetVal = Bits(0,PGIDX_BITS));
val p_store_type = Reg(resetVal = Bits(0,3));
val p_store_valid = Reg(resetVal = Bool(false));
val req_store = (io.cpu.req_cmd === M_XWR);
val req_load = (io.cpu.req_cmd === M_XRD);
val req_flush = (io.cpu.req_cmd === M_FLA);
val req_amo = io.cpu.req_cmd(3).toBool;
val r_req_load = (r_cpu_req_cmd === M_XRD);
val r_req_store = (r_cpu_req_cmd === M_XWR);
val r_req_flush = (r_cpu_req_cmd === M_FLA);
val r_req_amo = r_cpu_req_cmd(3).toBool;
when (io.cpu.req_val && io.cpu.req_rdy) {
r_cpu_req_idx := io.cpu.req_idx;
r_cpu_req_cmd := io.cpu.req_cmd;
r_cpu_req_type := io.cpu.req_type;
r_cpu_req_tag := io.cpu.req_tag;
}
when ((state === s_ready) && r_cpu_req_val && !io.cpu.req_kill) {
r_cpu_req_ppn := io.cpu.req_ppn;
}
when (io.cpu.req_rdy) {
r_cpu_req_val := io.cpu.req_val;
}
otherwise {
r_cpu_req_val := Bool(false);
}
when (((state === s_resolve_miss) && (r_req_load || r_req_amo)) || (state === s_replay_load)) {
r_cpu_resp_val := Bool(true);
}
otherwise {
r_cpu_resp_val := Bool(false);
}
// refill counter
val rr_count = Reg(resetVal = UFix(0,2));
val rr_count_next = rr_count + UFix(1);
when (((state === s_refill) && io.mem.resp_val) || ((state === s_writeback) && io.mem.req_rdy)) {
rr_count := rr_count_next;
}
// tag array
val tag_addr =
Mux((state === s_ready), io.cpu.req_idx(PGIDX_BITS-1,offsetbits),
r_cpu_req_idx(PGIDX_BITS-1,offsetbits)).toUFix;
val tag_we =
((state === s_refill) && io.mem.resp_val && (rr_count === UFix(3,2))) ||
((state === s_resolve_miss) && r_req_flush);
val tag_array = Mem(lines, r_cpu_req_ppn);
tag_array.setReadLatency(1);
tag_array.setTarget('inst);
val tag_rdata = tag_array.rw(tag_addr, r_cpu_req_ppn, tag_we);
// valid bit array
val vb_array = Reg(resetVal = Bits(0, lines));
when (tag_we && !r_req_flush) {
vb_array := vb_array.bitSet(r_cpu_req_idx(PGIDX_BITS-1,offsetbits).toUFix, UFix(1,1));
}
when (tag_we && r_req_flush) {
vb_array := vb_array.bitSet(r_cpu_req_idx(PGIDX_BITS-1,offsetbits).toUFix, UFix(0,1));
}
val vb_rdata = vb_array(r_cpu_req_idx(PGIDX_BITS-1,offsetbits).toUFix).toBool;
val tag_valid = r_cpu_req_val && vb_rdata;
val tag_match = (tag_rdata === io.cpu.req_ppn);
val tag_hit = tag_valid && tag_match;
val miss = r_cpu_req_val && (!vb_rdata || !tag_match);
// load/store addresses conflict if they are to any part of the same 64 bit word
val addr_match = (r_cpu_req_idx(PGIDX_BITS-1,offsetlsb) === p_store_idx(PGIDX_BITS-1,offsetlsb));
val ldst_conflict = tag_valid && tag_match && (r_req_load || r_req_amo) && p_store_valid && addr_match;
val store_hit = r_cpu_req_val && !io.cpu.req_kill && tag_hit && r_req_store ;
// write the pending store data when the cache is idle, when the next command isn't a load
// or when there's a load to the same address (in which case there's a 2 cycle delay:
// once cycle to write the store data and another to read the data back)
val drain_store =
((store_hit || p_store_valid) && (!io.cpu.req_val || req_store || req_flush)) ||
(p_store_valid && (miss || ldst_conflict));
// write pending store data from a store which missed
// after the cache line refill has completed
val resolve_store = (state === s_resolve_miss) && r_req_store;
// pending store data
when (io.cpu.req_val && io.cpu.req_rdy && req_store) {
p_store_idx := io.cpu.req_idx;
p_store_data := io.cpu.req_data;
p_store_type := io.cpu.req_type;
}
when (store_hit && !drain_store) {
p_store_valid := Bool(true);
}
when (drain_store) {
p_store_valid := Bool(false);
}
// AMO operand
when (io.cpu.req_val && io.cpu.req_rdy && req_amo) {
r_amo_data := io.cpu.req_data;
}
// dirty bit array
val db_array = Reg(resetVal = Bits(0, lines));
val tag_dirty = db_array(r_cpu_req_idx(PGIDX_BITS-1,offsetbits).toUFix).toBool;
when ((r_cpu_req_val && !io.cpu.req_kill && tag_hit && r_req_store) || resolve_store) {
db_array := db_array.bitSet(p_store_idx(PGIDX_BITS-1,offsetbits).toUFix, UFix(1,1));
}
when (state === s_write_amo) {
db_array := db_array.bitSet(r_cpu_req_idx(PGIDX_BITS-1,offsetbits).toUFix, UFix(1,1));
}
when (tag_we) {
db_array := db_array.bitSet(r_cpu_req_idx(PGIDX_BITS-1,offsetbits).toUFix, UFix(0,1));
}
// generate write mask and data signals for stores and amos
val storegen = new rocketDCacheStoreGen();
storegen.io.req_addr_lsb := p_store_idx(2,0);
storegen.io.req_data := p_store_data;
storegen.io.req_type := p_store_type;
val store_data = Fill(2, storegen.io.store_data);
val store_wmask_d = storegen.io.store_wmask;
val store_wmask = Mux(p_store_idx(offsetlsb).toBool, Cat(store_wmask_d, Bits(0,64)), Cat(Bits(0,64), store_wmask_d));
// ALU for AMOs
val amo_alu = new rocketDCacheAmoALU();
val amo_alu_out = Cat(amo_alu.io.result,amo_alu.io.result);
val amo_wmask =
Mux(r_cpu_req_type === MT_D, ~Bits(0,8),
Mux(r_cpu_req_idx(2).toBool, Cat(~Bits(0,4), Bits(0,4)),
Cat(Bits(0,4), ~Bits(0,4))));
val amo_store_wmask_d = Cat(Fill(8, amo_wmask(7)),
Fill(8, amo_wmask(6)),
Fill(8, amo_wmask(5)),
Fill(8, amo_wmask(4)),
Fill(8, amo_wmask(3)),
Fill(8, amo_wmask(2)),
Fill(8, amo_wmask(1)),
Fill(8, amo_wmask(0)));
val amo_store_wmask = Mux(r_cpu_req_idx(offsetlsb).toBool, Cat(amo_store_wmask_d, Bits(0,64)), Cat(Bits(0,64), amo_store_wmask_d));
// data array
val data_addr =
Mux(drain_store || resolve_store, p_store_idx(PGIDX_BITS-1, offsetmsb-1),
Mux((state === s_writeback) && io.mem.req_rdy, Cat(r_cpu_req_idx(PGIDX_BITS-1, offsetbits), rr_count_next),
Mux((state === s_start_writeback) || (state === s_writeback) || (state === s_refill), Cat(r_cpu_req_idx(PGIDX_BITS-1, offsetbits), rr_count),
Mux((state === s_resolve_miss) || (state === s_replay_load) || (state === s_write_amo), r_cpu_req_idx(PGIDX_BITS-1, offsetmsb-1),
io.cpu.req_idx(PGIDX_BITS-1, offsetmsb-1))))).toUFix;
val data_wdata =
Mux((state === s_refill), io.mem.resp_data,
Mux((state === s_write_amo), amo_alu_out,
store_data));
val data_we =
((state === s_refill) && io.mem.resp_val) ||
(state === s_write_amo) ||
drain_store || resolve_store;
val data_wmask =
Mux((state === s_refill), ~Bits(0,128),
Mux((state === s_write_amo), amo_store_wmask,
store_wmask));
val data_array = Mem(lines*4, data_wdata);
data_array.setReadLatency(1);
data_array.setTarget('inst);
val data_array_rdata = data_array.rw(data_addr, data_wdata, data_we, data_wmask);
val resp_data = Mux(r_cpu_req_idx(offsetlsb).toBool, data_array_rdata(127, 64), data_array_rdata(63,0));
val r_resp_data = Reg(resp_data);
amo_alu.io.cmd := r_cpu_req_cmd;
amo_alu.io.wmask := amo_wmask;
amo_alu.io.lhs := Mux(r_cpu_resp_val, resp_data, r_resp_data).toUFix;
amo_alu.io.rhs := r_amo_data.toUFix;
// signal a load miss when the data isn't present in the cache and when it's in the pending store data register
// (causes the cache to block for 2 cycles and the load or amo instruction is replayed)
val load_miss =
!io.cpu.req_kill &&
(state === s_ready) && r_cpu_req_val && (r_req_load || r_req_amo) && (!tag_hit || (p_store_valid && addr_match));
// output signals
// busy when there's a load to the same address as a pending store, or on a cache miss, or when executing a flush
io.cpu.req_rdy := (state === s_ready) && !io.cpu.req_kill && !ldst_conflict && (!r_cpu_req_val || (tag_hit && !(r_req_flush || r_req_amo)));
io.cpu.resp_val := !io.cpu.req_kill &&
((state === s_ready) && tag_hit && (r_req_load || r_req_amo) && !(p_store_valid && addr_match)) ||
((state === s_resolve_miss) && r_req_flush) ||
r_cpu_resp_val;
val misaligned =
(((r_cpu_req_type === MT_H) || (r_cpu_req_type === MT_HU)) && r_cpu_req_idx(0).toBool) ||
(((r_cpu_req_type === MT_W) || (r_cpu_req_type === MT_WU)) && (r_cpu_req_idx(1,0) != Bits(0,2))) ||
((r_cpu_req_type === MT_D) && (r_cpu_req_idx(2,0) != Bits(0,3)));
io.cpu.xcpt_ma_ld := r_cpu_req_val && (r_req_load || r_req_amo) && misaligned;
io.cpu.xcpt_ma_st := r_cpu_req_val && (r_req_store || r_req_amo) && misaligned;
io.cpu.resp_miss := load_miss;
io.cpu.resp_nack := Bool(false)
io.cpu.resp_tag := r_cpu_req_tag
io.cpu.resp_data := resp_data;
io.mem.req_val := (state === s_req_refill) || (state === s_writeback);
io.mem.req_rw := (state === s_writeback);
io.mem.req_wdata := data_array_rdata;
io.mem.req_tag := UFix(0);
io.mem.req_addr :=
Mux(state === s_writeback, Cat(tag_rdata, r_cpu_req_idx(PGIDX_BITS-1, offsetbits)),
Cat(r_cpu_req_ppn, r_cpu_req_idx(PGIDX_BITS-1, offsetbits))).toUFix;
// control state machine
switch (state) {
is (s_reset) {
state := s_ready;
}
is (s_ready) {
when (io.cpu.req_kill) {
state := s_ready;
}
when (ldst_conflict) {
state := s_replay_load;
}
when (!r_cpu_req_val || (tag_hit && !(r_req_flush || r_req_amo))) {
state := s_ready;
}
when (tag_hit && r_req_amo) {
state := s_write_amo;
}
when (tag_valid & tag_dirty) {
state := s_start_writeback;
}
when (r_req_flush) {
state := s_resolve_miss;
}
otherwise {
state := s_req_refill;
}
}
is (s_replay_load) {
state := s_ready;
}
is (s_write_amo) {
state := s_ready;
}
is (s_start_writeback) {
state := s_writeback;
}
is (s_writeback) {
when (io.mem.req_rdy && (rr_count === UFix(3,2))) {
when (r_req_flush) {
state := s_resolve_miss;
}
otherwise {
state := s_req_refill;
}
}
}
is (s_req_refill)
{
when (io.mem.req_rdy) { state := s_refill; }
}
is (s_refill) {
when (io.mem.resp_val && (rr_count === UFix(3,2))) { state := s_resolve_miss; }
}
is (s_resolve_miss) {
when (r_req_amo) {
state := s_write_amo;
}
state := s_ready;
}
}
}
class rocketDCacheAmoALU extends Component {
val io = new Bundle {
val cmd = Bits(4, INPUT);
val wmask = Bits(8, INPUT);
val lhs = UFix(64, INPUT);
val rhs = UFix(64, INPUT);
val result = UFix(64, OUTPUT);
}
// val signed_cmp = (op === M_XA_MIN) || (op === M_XA_MAX);
// val sub = (op === M_XA_MIN) || (op === M_XA_MINU) ||
// (op === M_XA_MAX) || (op === M_XA_MAXU);
val adder_lhs = Cat(io.lhs(63,32),io.wmask(3) & io.lhs(31), io.lhs(30,0)).toUFix;
val adder_rhs = Cat(io.rhs(63,32),io.wmask(3) & io.rhs(31), io.rhs(30,0)).toUFix;
// val adder_rhs = Cat(Mux(sub, ~io.rhs, io.rhs), sub).toUFix;
// val sum = adder_lhs + adder_rhs;
// val adder_out = sum(64,1);
val adder_out = adder_lhs + adder_rhs;
val alu_out = Wire() { UFix() };
switch (io.cmd) {
// is (M_XA_ADD) { alu_out := adder_out; }
is (M_XA_SWAP) { alu_out := io.rhs; }
is (M_XA_AND) { alu_out := io.lhs & io.rhs; }
is (M_XA_OR) { alu_out := io.lhs | io.rhs; }
}
alu_out := adder_out;
io.result := alu_out;
}
}

View File

@ -650,282 +650,46 @@ class AMOALU extends Component {
io.out := Mux(word, Cat(out(31,0), out(31,0)).toUFix, out)
}
class HellaCacheDM extends Component {
val io = new ioDCacheHella()
// interface between D$ and processor/DTLB
class ioDmem(view: List[String] = null) extends Bundle(view) {
val req_kill = Bool(INPUT);
val req_val = Bool(INPUT);
val req_rdy = Bool(OUTPUT);
val req_cmd = Bits(4, INPUT);
val req_type = Bits(3, INPUT);
val req_idx = Bits(PGIDX_BITS, INPUT);
val req_ppn = Bits(PPN_BITS, INPUT);
val req_data = Bits(64, INPUT);
val req_tag = Bits(DCACHE_TAG_BITS, INPUT);
val xcpt_ma_ld = Bool(OUTPUT); // misaligned load
val xcpt_ma_st = Bool(OUTPUT); // misaligned store
val resp_miss = Bool(OUTPUT);
val resp_nack = Bool(OUTPUT);
val resp_val = Bool(OUTPUT);
val resp_replay = Bool(OUTPUT);
val resp_data = Bits(64, OUTPUT);
val resp_data_subword = Bits(64, OUTPUT);
val resp_tag = Bits(DCACHE_TAG_BITS, OUTPUT);
}
val lines = 1 << IDX_BITS
val addrbits = PADDR_BITS
val indexbits = log2up(lines)
val offsetbits = OFFSET_BITS
val tagmsb = PADDR_BITS-1
val taglsb = indexbits+offsetbits
val tagbits = tagmsb-taglsb+1
val indexmsb = taglsb-1
val indexlsb = offsetbits
val offsetmsb = indexlsb-1
val offsetlsb = log2up(CPU_DATA_BITS/8)
val ramindexlsb = log2up(MEM_DATA_BITS/8)
val early_nack = Reg { Bool() }
val r_cpu_req_val_ = Reg(io.cpu.req_val && io.cpu.req_rdy, resetVal = Bool(false))
val r_cpu_req_val = r_cpu_req_val_ && !io.cpu.req_kill && !early_nack
val r_cpu_req_idx = Reg() { Bits() }
val r_cpu_req_cmd = Reg() { Bits() }
val r_cpu_req_type = Reg() { Bits() }
val r_cpu_req_tag = Reg() { Bits() }
val r_amo_replay_data = Reg() { Bits() }
val p_store_valid = Reg(resetVal = Bool(false))
val p_store_data = Reg() { Bits() }
val p_store_idx = Reg() { Bits() }
val p_store_cmd = Reg() { Bits() }
val p_store_type = Reg() { Bits() }
val r_replay_amo = Reg(resetVal = Bool(false))
val req_store = (io.cpu.req_cmd === M_XWR)
val req_load = (io.cpu.req_cmd === M_XRD)
val req_amo = io.cpu.req_cmd(3).toBool
val req_read = req_load || req_amo
val req_write = req_store || req_amo
val r_req_load = (r_cpu_req_cmd === M_XRD)
val r_req_store = (r_cpu_req_cmd === M_XWR)
val r_req_flush = (r_cpu_req_cmd === M_FLA)
val r_req_fence = (r_cpu_req_cmd === M_FENCE)
val r_req_amo = r_cpu_req_cmd(3).toBool
val r_req_read = r_req_load || r_req_amo
val r_req_write = r_req_store || r_req_amo
val r_req_readwrite = r_req_read || r_req_write
// replay unit
val replayer = new ReplayUnit()
val replay_amo_val = replayer.io.data_req.valid && replayer.io.data_req.bits.cmd(3).toBool
when (io.cpu.req_val) {
r_cpu_req_idx := io.cpu.req_idx
r_cpu_req_cmd := io.cpu.req_cmd
r_cpu_req_type := io.cpu.req_type
r_cpu_req_tag := io.cpu.req_tag
}
when (replay_amo_val) {
r_cpu_req_idx := Cat(replayer.io.data_req.bits.idx, replayer.io.data_req.bits.offset)
r_cpu_req_cmd := replayer.io.data_req.bits.cmd
r_cpu_req_type := replayer.io.data_req.bits.typ
r_amo_replay_data := replayer.io.data_req.bits.data
}
val cpu_req_data = Mux(r_replay_amo, r_amo_replay_data, io.cpu.req_data)
// refill counter
val rr_count = Reg(resetVal = UFix(0, log2up(REFILL_CYCLES)))
val rr_count_next = rr_count + UFix(1)
when (io.mem.resp_val) { rr_count := rr_count_next }
val misaligned =
(((r_cpu_req_type === MT_H) || (r_cpu_req_type === MT_HU)) && (r_cpu_req_idx(0) != Bits(0))) ||
(((r_cpu_req_type === MT_W) || (r_cpu_req_type === MT_WU)) && (r_cpu_req_idx(1,0) != Bits(0))) ||
((r_cpu_req_type === MT_D) && (r_cpu_req_idx(2,0) != Bits(0)));
io.cpu.xcpt_ma_ld := r_cpu_req_val_ && r_req_read && misaligned
io.cpu.xcpt_ma_st := r_cpu_req_val_ && r_req_write && misaligned
// tags
val meta = new MetaDataArray(lines)
val meta_arb = (new Arbiter(3)) { new MetaArrayReq() }
meta_arb.io.out <> meta.io.req
// data
val data = new DataArray(lines)
val data_arb = (new Arbiter(5)) { new DataArrayReq() }
data_arb.io.out <> data.io.req
// writeback unit
val wb = new WritebackUnit
val wb_arb = (new Arbiter(2)) { new WritebackReq() }
wb_arb.io.out <> wb.io.req
wb.io.data_req.bits.inner_req <> data_arb.io.in(3).bits
wb.io.data_req.ready := data_arb.io.in(3).ready
data_arb.io.in(3).valid := wb.io.data_req.valid
wb.io.data_resp <> data.io.resp
// cpu tag check
meta_arb.io.in(2).valid := io.cpu.req_val
meta_arb.io.in(2).bits.idx := io.cpu.req_idx(indexmsb,indexlsb)
meta_arb.io.in(2).bits.rw := Bool(false)
meta_arb.io.in(2).bits.data.valid := Bool(false) // don't care
meta_arb.io.in(2).bits.data.dirty := Bool(false) // don't care
meta_arb.io.in(2).bits.data.tag := UFix(0) // don't care
val early_tag_nack = !meta_arb.io.in(2).ready
val cpu_req_tag = Cat(io.cpu.req_ppn, r_cpu_req_idx)(tagmsb,taglsb)
val tag_match = meta.io.resp.valid && (meta.io.resp.tag === cpu_req_tag)
val tag_hit = r_cpu_req_val && tag_match
val tag_miss = r_cpu_req_val && !tag_match
val dirty = meta.io.resp.valid && meta.io.resp.dirty
// refill response
val block_during_refill = !io.mem.resp_val && (rr_count != UFix(0))
data_arb.io.in(0).valid := io.mem.resp_val || block_during_refill
data_arb.io.in(0).bits.offset := rr_count
data_arb.io.in(0).bits.rw := !block_during_refill
data_arb.io.in(0).bits.wmask := ~UFix(0, MEM_DATA_BITS/8)
data_arb.io.in(0).bits.data := io.mem.resp_data
// load hits
data_arb.io.in(4).bits.offset := io.cpu.req_idx(offsetmsb,ramindexlsb)
data_arb.io.in(4).bits.idx := io.cpu.req_idx(indexmsb,indexlsb)
data_arb.io.in(4).bits.rw := Bool(false)
data_arb.io.in(4).bits.wmask := UFix(0) // don't care
data_arb.io.in(4).bits.data := io.mem.resp_data // don't care
data_arb.io.in(4).valid := io.cpu.req_val && req_read
val early_load_nack = req_read && !data_arb.io.in(4).ready
// store hits and AMO hits and misses use a pending store register.
// we nack new stores if a pending store can't retire for some reason.
// we drain a pending store if the CPU performs a store or a
// conflictig load, or if the cache is idle, or after a miss.
val p_store_idx_match = p_store_valid && (r_cpu_req_idx(indexmsb,indexlsb) === p_store_idx(indexmsb,indexlsb))
val p_store_offset_match = (r_cpu_req_idx(indexlsb-1,offsetlsb) === p_store_idx(indexlsb-1,offsetlsb))
val p_store_match = r_cpu_req_val && r_req_read && p_store_idx_match && p_store_offset_match
val drain_store_val = (p_store_valid && (!io.cpu.req_val || !req_read || Reg(tag_miss))) || p_store_match
data_arb.io.in(2).bits.offset := p_store_idx(offsetmsb,ramindexlsb)
data_arb.io.in(2).bits.idx := p_store_idx(indexmsb,indexlsb)
data_arb.io.in(2).bits.rw := Bool(true)
data_arb.io.in(2).valid := drain_store_val
val drain_store = drain_store_val && data_arb.io.in(2).ready
val p_store_rdy = !p_store_valid || drain_store
val p_amo = Reg(tag_hit && r_req_amo && p_store_rdy && !p_store_match || r_replay_amo, resetVal = Bool(false))
p_store_valid := !p_store_rdy || (tag_hit && r_req_store) || p_amo
// writeback
val wb_rdy = wb_arb.io.in(1).ready && !p_store_idx_match
wb_arb.io.in(1).valid := tag_miss && r_req_readwrite && dirty && !p_store_idx_match
wb_arb.io.in(1).bits.ppn := meta.io.resp.tag
wb_arb.io.in(1).bits.idx := r_cpu_req_idx(indexmsb,indexlsb)
// tag update after a miss or a store to an exclusive clean line.
val clear_valid = tag_miss && r_req_readwrite && meta.io.resp.valid && (!dirty || wb_rdy)
val set_dirty = tag_hit && !meta.io.resp.dirty && r_req_write
meta.io.state_req.valid := clear_valid || set_dirty
meta.io.state_req.bits.rw := Bool(true)
meta.io.state_req.bits.idx := r_cpu_req_idx(indexmsb,indexlsb)
meta.io.state_req.bits.data.tag := UFix(0) // don't care
meta.io.state_req.bits.data.valid := tag_match
meta.io.state_req.bits.data.dirty := tag_match
// pending store data, also used for AMO RHS
val amoalu = new AMOALU
when (tag_hit && r_req_write && p_store_rdy || r_replay_amo) {
p_store_idx := r_cpu_req_idx
p_store_type := r_cpu_req_type
p_store_cmd := r_cpu_req_cmd
p_store_data := cpu_req_data
}
when (p_amo) {
p_store_data := amoalu.io.out
}
// miss handling
val mshr = new MSHRFile()
mshr.io.req_val := tag_miss && r_req_readwrite && (!dirty || wb_rdy) && (!r_req_write || replayer.io.sdq_enq.ready)
mshr.io.req_ppn := cpu_req_tag
mshr.io.req_idx := r_cpu_req_idx(indexmsb,indexlsb)
mshr.io.req_tag := r_cpu_req_tag
mshr.io.req_offset := r_cpu_req_idx(offsetmsb,0)
mshr.io.req_cmd := r_cpu_req_cmd
mshr.io.req_type := r_cpu_req_type
mshr.io.req_sdq_id := replayer.io.sdq_id
mshr.io.mem_resp_val := io.mem.resp_val && (~rr_count === UFix(0))
mshr.io.mem_resp_tag := io.mem.resp_tag
mshr.io.mem_req <> wb.io.refill_req
mshr.io.meta_req.bits.inner_req <> meta_arb.io.in(1).bits
mshr.io.meta_req.ready := meta_arb.io.in(1).ready
meta_arb.io.in(1).valid := mshr.io.meta_req.valid
mshr.io.replay <> replayer.io.replay
replayer.io.sdq_enq.valid := tag_miss && r_req_write && (!dirty || wb_rdy) && mshr.io.req_rdy
replayer.io.sdq_enq.bits := cpu_req_data
data_arb.io.in(0).bits.idx := mshr.io.mem_resp_idx
// replays
val replay = replayer.io.data_req.bits
val stall_replay = r_replay_amo || p_amo || p_store_valid
val replay_val = replayer.io.data_req.valid && !stall_replay
val replay_rdy = data_arb.io.in(1).ready
data_arb.io.in(1).bits.offset := replay.offset(offsetmsb,ramindexlsb)
data_arb.io.in(1).bits.idx := replay.idx
data_arb.io.in(1).bits.rw := replay.cmd === M_XWR
data_arb.io.in(1).valid := replay_val
replayer.io.data_req.ready := replay_rdy && !stall_replay
r_replay_amo := replay_amo_val && replay_rdy && !stall_replay
// store write mask generation.
// assumes store replays are higher-priority than pending stores.
val maskgen = new StoreMaskGen
val store_offset = Mux(!replay_val, p_store_idx(offsetmsb,0), replay.offset)
maskgen.io.typ := Mux(!replay_val, p_store_type, replay.typ)
maskgen.io.addr := store_offset(offsetlsb-1,0)
val store_wmask_wide = maskgen.io.wmask << Cat(store_offset(ramindexlsb-1,offsetlsb), Bits(0, log2up(CPU_DATA_BITS/8))).toUFix
val store_data = Mux(!replay_val, p_store_data, replay.data)
val store_data_wide = Fill(MEM_DATA_BITS/CPU_DATA_BITS, store_data)
data_arb.io.in(1).bits.data := store_data_wide
data_arb.io.in(1).bits.wmask := store_wmask_wide
data_arb.io.in(2).bits.data := store_data_wide
data_arb.io.in(2).bits.wmask := store_wmask_wide
// load data subword mux/sign extension.
// subword loads are delayed by one cycle.
val loadgen = new LoadDataGen
val loadgen_use_replay = Reg(replay_val && replay_rdy)
loadgen.io.typ := Mux(loadgen_use_replay, Reg(replay.typ), r_cpu_req_type)
loadgen.io.addr := Mux(loadgen_use_replay, Reg(replay.offset), r_cpu_req_idx)(ramindexlsb-1,0)
loadgen.io.din := data.io.resp
amoalu.io.cmd := p_store_cmd
amoalu.io.typ := p_store_type
amoalu.io.lhs := loadgen.io.r_dout.toUFix
amoalu.io.rhs := p_store_data.toUFix
early_nack := early_tag_nack || early_load_nack || r_cpu_req_val && r_req_amo || replay_amo_val || r_replay_amo
// reset and flush unit
val flusher = new FlushUnit(lines)
val flushed = Reg(resetVal = Bool(true))
val flush_rdy = mshr.io.fence_rdy && wb_rdy && !p_store_valid
flushed := flushed && !r_cpu_req_val || r_cpu_req_val && r_req_flush && flush_rdy && flusher.io.req.ready
flusher.io.req.valid := r_cpu_req_val && r_req_flush && flush_rdy && !flushed
flusher.io.wb_req <> wb_arb.io.in(0)
flusher.io.meta_req.bits.inner_req <> meta_arb.io.in(0).bits
flusher.io.meta_req.ready := meta_arb.io.in(0).ready
meta_arb.io.in(0).valid := flusher.io.meta_req.valid
flusher.io.meta_resp <> meta.io.resp
flusher.io.resp.ready := Bool(true) // we don't respond to flush requests
// we usually nack rather than reporting that the cache is not ready.
// fences and flushes are the exceptions.
val pending_fence = Reg(resetVal = Bool(false))
pending_fence := (r_cpu_req_val && r_req_fence || pending_fence) && !flush_rdy
val nack_hit = p_store_match || r_req_write && !p_store_rdy
val nack_miss = dirty && !wb_rdy || !mshr.io.req_rdy || r_req_write && !replayer.io.sdq_enq.ready
val nack_flush = !flush_rdy && (r_req_fence || r_req_flush) ||
!flushed && r_req_flush
val nack = early_nack || r_req_readwrite && Mux(tag_match, nack_hit, nack_miss) || nack_flush
io.cpu.req_rdy := flusher.io.req.ready && !(r_cpu_req_val_ && r_req_flush) && !pending_fence
io.cpu.resp_nack := r_cpu_req_val_ && !io.cpu.req_kill && nack
io.cpu.resp_val := (tag_hit && !nack_hit && r_req_read) || replayer.io.cpu_resp_val
io.cpu.resp_replay := replayer.io.cpu_resp_val
io.cpu.resp_miss := tag_miss && !nack_miss && r_req_read
io.cpu.resp_tag := Mux(replayer.io.cpu_resp_val, replayer.io.cpu_resp_tag, r_cpu_req_tag)
io.cpu.resp_data := loadgen.io.dout
io.cpu.resp_data_subword := loadgen.io.r_dout_subword
wb.io.mem_req.ready := io.mem.req_rdy
io.mem.req_val := wb.io.mem_req.valid
io.mem.req_rw := wb.io.mem_req.bits.rw
io.mem.req_wdata := wb.io.mem_req_data
io.mem.req_tag := wb.io.mem_req.bits.tag.toUFix
io.mem.req_addr := wb.io.mem_req.bits.addr
// interface between D$ and next level in memory hierarchy
class ioDCache(view: List[String] = null) extends Bundle(view) {
val req_addr = UFix(PADDR_BITS - OFFSET_BITS, INPUT);
val req_tag = UFix(DMEM_TAG_BITS, INPUT);
val req_val = Bool(INPUT);
val req_rdy = Bool(OUTPUT);
val req_wdata = Bits(MEM_DATA_BITS, INPUT);
val req_rw = Bool(INPUT);
val resp_data = Bits(MEM_DATA_BITS, OUTPUT);
val resp_tag = Bits(DMEM_TAG_BITS, OUTPUT);
val resp_val = Bool(OUTPUT);
}
class HellaCacheAssoc extends Component {
val io = new ioDCacheHella()
class HellaCache extends Component {
val io = new Bundle {
val cpu = new ioDmem()
val mem = new ioDCache().flip
}
val lines = 1 << IDX_BITS
val addrbits = PADDR_BITS

View File

@ -17,7 +17,7 @@ class Top() extends Component {
val cpu = new rocketProc();
val icache = new rocketICache(128, 2); // 128 sets x 2 ways
val icache_pf = new rocketIPrefetcher();
val dcache = new HellaCacheAssoc();
val dcache = new HellaCache();
val arbiter = new rocketMemArbiter();
arbiter.io.mem <> io.mem;
@ -37,9 +37,9 @@ class Top() extends Component {
object top_main {
def main(args: Array[String]) = {
// Can turn off --debug and --vcd when done with debugging to improve emulator performance
// val cpu_args = args ++ Array("--target-dir", "generated-src","--debug","--vcd");
val cpu_args = args ++ Array("--target-dir", "generated-src","--debug","--vcd");
// val cpu_args = args ++ Array("--target-dir", "generated-src", "--debug");
val cpu_args = args ++ Array("--target-dir", "generated-src");
// val cpu_args = args ++ Array("--target-dir", "generated-src");
// Set variables based off of command flags
// for(a <- args) {
// a match {