riscv/rocket-chip
1
0
Fork 0
Code Releases Activity

hellacache returns!

Browse Source 
but AMOs are unimplemented.
This commit is contained in:
Andrew Waterman 2011-12-12 06:49:16 -08:00
parent 0ea2704b80
commit 56c4f44c2a
12 changed files with 487 additions and 342 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
rocket/src/main/scala/arbiter.scala
View File

@ -9,7 +9,7 @@ class ioMem() extends Bundle
val req_val = Bool('output); val req_val = Bool('output);
val req_rdy = Bool('input); val req_rdy = Bool('input);
val req_rw = Bool('output); val req_rw = Bool('output);
val req_addr = UFix(PADDR_BITS, 'output); val req_addr = UFix(PADDR_BITS - OFFSET_BITS, 'output);
val req_wdata = Bits(MEM_DATA_BITS, 'output); val req_wdata = Bits(MEM_DATA_BITS, 'output);
val req_tag = Bits(MEM_TAG_BITS, 'output); val req_tag = Bits(MEM_TAG_BITS, 'output);

2
rocket/src/main/scala/consts.scala
View File

@ -186,7 +186,7 @@ object Constants
// rocketNBDCacheDM parameters // rocketNBDCacheDM parameters
val CPU_DATA_BITS = 64; val CPU_DATA_BITS = 64;
val CPU_TAG_BITS = 11; val CPU_TAG_BITS = 5;
val DCACHE_TAG_BITS = 1 + CPU_TAG_BITS; val DCACHE_TAG_BITS = 1 + CPU_TAG_BITS;
val OFFSET_BITS = 6; // log2(cache line size in bytes) val OFFSET_BITS = 6; // log2(cache line size in bytes)
val NMSHR = 2; // number of primary misses val NMSHR = 2; // number of primary misses

4
rocket/src/main/scala/cpu.scala
View File

@ -66,7 +66,6 @@ class rocketProc extends Component
ctrl.io.imem.resp_val := io.imem.resp_val; ctrl.io.imem.resp_val := io.imem.resp_val;
dpath.io.imem.resp_data := io.imem.resp_data; dpath.io.imem.resp_data := io.imem.resp_data;
ctrl.io.xcpt_itlb := itlb.io.cpu.exception; ctrl.io.xcpt_itlb := itlb.io.cpu.exception;
// ctrl.io.itlb_miss := itlb.io.cpu.resp_miss;
io.imem.itlb_miss := itlb.io.cpu.resp_miss; io.imem.itlb_miss := itlb.io.cpu.resp_miss;
// connect DTLB to D$ arbiter, ctrl+dpath // connect DTLB to D$ arbiter, ctrl+dpath
@ -95,7 +94,7 @@ class rocketProc extends Component
arb.io.cpu.req_val := ctrl.io.dmem.req_val; arb.io.cpu.req_val := ctrl.io.dmem.req_val;
arb.io.cpu.req_cmd := ctrl.io.dmem.req_cmd; arb.io.cpu.req_cmd := ctrl.io.dmem.req_cmd;
arb.io.cpu.req_type := ctrl.io.dmem.req_type; arb.io.cpu.req_type := ctrl.io.dmem.req_type;
arb.io.cpu.req_nack := ctrl.io.dpath.killm; arb.io.cpu.req_kill := ctrl.io.dmem.req_kill;
arb.io.cpu.req_idx := dpath.io.dmem.req_addr(PGIDX_BITS-1,0); arb.io.cpu.req_idx := dpath.io.dmem.req_addr(PGIDX_BITS-1,0);
arb.io.cpu.req_ppn := dtlb.io.cpu.resp_ppn; arb.io.cpu.req_ppn := dtlb.io.cpu.resp_ppn;
arb.io.cpu.req_data := dpath.io.dmem.req_data; arb.io.cpu.req_data := dpath.io.dmem.req_data;
@ -106,6 +105,7 @@ class rocketProc extends Component
dpath.io.dmem.resp_val := arb.io.cpu.resp_val; dpath.io.dmem.resp_val := arb.io.cpu.resp_val;
dpath.io.dmem.resp_tag := arb.io.cpu.resp_tag; dpath.io.dmem.resp_tag := arb.io.cpu.resp_tag;
dpath.io.dmem.resp_data := arb.io.cpu.resp_data; dpath.io.dmem.resp_data := arb.io.cpu.resp_data;
dpath.io.dmem.resp_data_subword := io.dmem.resp_data_subword;
io.console.bits := dpath.io.console.bits; io.console.bits := dpath.io.console.bits;
io.console.valid := dpath.io.console.valid; io.console.valid := dpath.io.console.valid;

16
rocket/src/main/scala/ctrl.scala
View File

@ -75,7 +75,7 @@ class ioCtrlAll extends Bundle()
val dpath = new ioCtrlDpath(); val dpath = new ioCtrlDpath();
val console = new ioConsole(List("rdy")); val console = new ioConsole(List("rdy"));
val imem = new ioImem(List("req_val", "req_rdy", "resp_val")).flip(); val imem = new ioImem(List("req_val", "req_rdy", "resp_val")).flip();
val dmem = new ioDmem(List("req_val", "req_rdy", "req_cmd", "req_type", "resp_miss", "resp_nack")).flip(); val dmem = new ioDmem(List("req_val", "req_kill", "req_rdy", "req_cmd", "req_type", "resp_miss", "resp_nack")).flip();
val host = new ioHost(List("start")); val host = new ioHost(List("start"));
val dtlb_val = Bool('output) val dtlb_val = Bool('output)
val dtlb_rdy = Bool('input); val dtlb_rdy = Bool('input);
@ -358,6 +358,7 @@ class rocketCtrl extends Component
val mem_reg_xcpt_fpu = Reg(resetVal = Bool(false)); val mem_reg_xcpt_fpu = Reg(resetVal = Bool(false));
val mem_reg_xcpt_syscall = Reg(resetVal = Bool(false)); val mem_reg_xcpt_syscall = Reg(resetVal = Bool(false));
val mem_reg_replay = Reg(resetVal = Bool(false)); val mem_reg_replay = Reg(resetVal = Bool(false));
val mem_reg_kill_dmem = Reg(resetVal = Bool(false));
when (!io.dpath.stalld) { when (!io.dpath.stalld) {
when (io.dpath.killf) { when (io.dpath.killf) {
@ -527,7 +528,8 @@ class rocketCtrl extends Component
io.dpath.badvaddr_wen := io.xcpt_dtlb_ld || io.xcpt_dtlb_st; io.dpath.badvaddr_wen := io.xcpt_dtlb_ld || io.xcpt_dtlb_st;
// replay mem stage PC on a DTLB miss // replay mem stage PC on a DTLB miss
val mem_hazard = io.dtlb_miss || io.dmem.resp_nack val mem_hazard = io.dtlb_miss || io.dmem.resp_nack;
val mem_kill_dmem = io.dtlb_miss || mem_exception || mem_reg_kill_dmem;
val replay_mem = mem_hazard || mem_reg_replay; val replay_mem = mem_hazard || mem_reg_replay;
val kill_mem = mem_hazard || mem_exception; val kill_mem = mem_hazard || mem_exception;
@ -541,10 +543,11 @@ class rocketCtrl extends Component
val ex_hazard = io.dmem.resp_miss || mem_reg_privileged || mem_reg_flush_inst val ex_hazard = io.dmem.resp_miss || mem_reg_privileged || mem_reg_flush_inst
val mem_kill_ex = kill_mem || take_pc_mem val mem_kill_ex = kill_mem || take_pc_mem
val kill_ex = mem_kill_ex || ex_hazard || !(io.dmem.req_rdy && io.dtlb_rdy) && ex_reg_mem_val val kill_ex = mem_kill_ex || ex_hazard || !(io.dmem.req_rdy && io.dtlb_rdy) && ex_reg_mem_val
val kill_dtlb = mem_kill_ex || ex_hazard || !io.dmem.req_rdy val ex_kill_dtlb = mem_kill_ex || ex_hazard || !io.dmem.req_rdy
val kill_dmem = mem_kill_ex || ex_hazard || !io.dtlb_rdy val ex_kill_dmem = mem_kill_ex || ex_hazard || !io.dtlb_rdy
mem_reg_replay <== kill_ex && !mem_kill_ex mem_reg_replay <== kill_ex && !mem_kill_ex
mem_reg_kill_dmem <== ex_kill_dmem
io.dpath.sel_pc := io.dpath.sel_pc :=
Mux(replay_mem, PC_MEM, // dtlb miss Mux(replay_mem, PC_MEM, // dtlb miss
@ -664,8 +667,9 @@ class rocketCtrl extends Component
io.dpath.irq_disable := mem_reg_inst_di && !kill_mem; io.dpath.irq_disable := mem_reg_inst_di && !kill_mem;
io.dpath.irq_enable := mem_reg_inst_ei && !kill_mem; io.dpath.irq_enable := mem_reg_inst_ei && !kill_mem;
io.dtlb_val := ex_reg_mem_val && !kill_dtlb; io.dtlb_val := ex_reg_mem_val && !ex_kill_dtlb;
io.dmem.req_val := ex_reg_mem_val && !kill_dmem; io.dmem.req_val := ex_reg_mem_val;
io.dmem.req_kill := mem_kill_dmem;
io.dmem.req_cmd := ex_reg_mem_cmd; io.dmem.req_cmd := ex_reg_mem_cmd;
io.dmem.req_type := ex_reg_mem_type; io.dmem.req_type := ex_reg_mem_type;
io.dpath.ex_mem_type:= ex_reg_mem_type io.dpath.ex_mem_type:= ex_reg_mem_type

25
rocket/src/main/scala/dcache.scala
View File

@ -7,7 +7,7 @@ import scala.math._;
// interface between D$ and processor/DTLB // interface between D$ and processor/DTLB
class ioDmem(view: List[String] = null) extends Bundle(view) { class ioDmem(view: List[String] = null) extends Bundle(view) {
val req_nack = Bool('input); val req_kill = Bool('input);
val req_val = Bool('input); val req_val = Bool('input);
val req_rdy = Bool('output); val req_rdy = Bool('output);
val req_cmd = Bits(4, 'input); val req_cmd = Bits(4, 'input);
@ -22,12 +22,13 @@ class ioDmem(view: List[String] = null) extends Bundle(view) {
val resp_nack = Bool('output); val resp_nack = Bool('output);
val resp_val = Bool('output); val resp_val = Bool('output);
val resp_data = Bits(64, 'output); val resp_data = Bits(64, 'output);
val resp_data_subword = Bits(64, 'output);
val resp_tag = Bits(DCACHE_TAG_BITS, 'output); val resp_tag = Bits(DCACHE_TAG_BITS, 'output);
} }
// interface between D$ and next level in memory hierarchy // interface between D$ and next level in memory hierarchy
class ioDcache(view: List[String] = null) extends Bundle(view) { class ioDcache(view: List[String] = null) extends Bundle(view) {
val req_addr = UFix(PADDR_BITS, 'input); val req_addr = UFix(PADDR_BITS - OFFSET_BITS, 'input);
val req_tag = UFix(DMEM_TAG_BITS, 'input); val req_tag = UFix(DMEM_TAG_BITS, 'input);
val req_val = Bool('input); val req_val = Bool('input);
val req_rdy = Bool('output); val req_rdy = Bool('output);
@ -156,7 +157,7 @@ class rocketDCacheDM_flush(lines: Int) extends Component {
dcache.io.cpu.req_tag := Mux(flushing, r_cpu_req_tag, io.cpu.req_tag); 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_type := io.cpu.req_type;
dcache.io.cpu.req_data ^^ io.cpu.req_data; dcache.io.cpu.req_data ^^ io.cpu.req_data;
dcache.io.cpu.req_nack := io.cpu.req_nack && !flush_waiting; dcache.io.cpu.req_kill := io.cpu.req_kill && !flush_waiting;
dcache.io.mem ^^ io.mem; dcache.io.mem ^^ io.mem;
io.cpu.xcpt_ma_ld := dcache.io.cpu.xcpt_ma_ld; io.cpu.xcpt_ma_ld := dcache.io.cpu.xcpt_ma_ld;
@ -220,7 +221,7 @@ class rocketDCacheDM(lines: Int) extends Component {
r_cpu_req_tag <== io.cpu.req_tag; r_cpu_req_tag <== io.cpu.req_tag;
} }
when ((state === s_ready) && r_cpu_req_val && !io.cpu.req_nack) { when ((state === s_ready) && r_cpu_req_val && !io.cpu.req_kill) {
r_cpu_req_ppn <== io.cpu.req_ppn; r_cpu_req_ppn <== io.cpu.req_ppn;
} }
when (io.cpu.req_rdy) { when (io.cpu.req_rdy) {
@ -273,7 +274,7 @@ class rocketDCacheDM(lines: Int) extends Component {
// load/store addresses conflict if they are to any part of the same 64 bit word // 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 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 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_nack && tag_hit && r_req_store ; 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 // 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: // or when there's a load to the same address (in which case there's a 2 cycle delay:
@ -307,7 +308,7 @@ class rocketDCacheDM(lines: Int) extends Component {
// dirty bit array // dirty bit array
val db_array = Reg(resetVal = Bits(0, lines)); val db_array = Reg(resetVal = Bits(0, lines));
val tag_dirty = db_array(r_cpu_req_idx(PGIDX_BITS-1,offsetbits).toUFix).toBool; val tag_dirty = db_array(r_cpu_req_idx(PGIDX_BITS-1,offsetbits).toUFix).toBool;
when ((r_cpu_req_val && !io.cpu.req_nack && tag_hit && r_req_store) || resolve_store) { 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)); db_array <== db_array.bitSet(p_store_idx(PGIDX_BITS-1,offsetbits).toUFix, UFix(1,1));
} }
when (state === s_write_amo) { when (state === s_write_amo) {
@ -383,13 +384,13 @@ class rocketDCacheDM(lines: Int) extends Component {
// signal a load miss when the data isn't present in the cache and when it's in the pending store data register // 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) // (causes the cache to block for 2 cycles and the load or amo instruction is replayed)
val load_miss = val load_miss =
!io.cpu.req_nack && !io.cpu.req_kill &&
(state === s_ready) && r_cpu_req_val && (r_req_load || r_req_amo) && (!tag_hit || (p_store_valid && addr_match)); (state === s_ready) && r_cpu_req_val && (r_req_load || r_req_amo) && (!tag_hit || (p_store_valid && addr_match));
// output signals // 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 // 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_nack && !ldst_conflict && (!r_cpu_req_val || (tag_hit && !(r_req_flush || r_req_amo))); 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_nack && 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_ready) && tag_hit && (r_req_load || r_req_amo) && !(p_store_valid && addr_match)) ||
((state === s_resolve_miss) && r_req_flush) || ((state === s_resolve_miss) && r_req_flush) ||
r_cpu_resp_val; r_cpu_resp_val;
@ -412,8 +413,8 @@ class rocketDCacheDM(lines: Int) extends Component {
io.mem.req_wdata := data_array_rdata; io.mem.req_wdata := data_array_rdata;
io.mem.req_tag := UFix(0); io.mem.req_tag := UFix(0);
io.mem.req_addr := io.mem.req_addr :=
Mux(state === s_writeback, Cat(tag_rdata, r_cpu_req_idx(PGIDX_BITS-1, offsetbits), rr_count), 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), Bits(0,2))).toUFix; Cat(r_cpu_req_ppn, r_cpu_req_idx(PGIDX_BITS-1, offsetbits))).toUFix;
// control state machine // control state machine
switch (state) { switch (state) {
@ -421,7 +422,7 @@ class rocketDCacheDM(lines: Int) extends Component {
state <== s_ready; state <== s_ready;
} }
is (s_ready) { is (s_ready) {
when (io.cpu.req_nack) { when (io.cpu.req_kill) {
state <== s_ready; state <== s_ready;
} }
when (ldst_conflict) { when (ldst_conflict) {

51
rocket/src/main/scala/dpath.scala
View File

@ -13,6 +13,7 @@ class ioDpathDmem extends Bundle()
val resp_val = Bool('input); val resp_val = Bool('input);
val resp_tag = Bits(CPU_TAG_BITS, 'input); val resp_tag = Bits(CPU_TAG_BITS, 'input);
val resp_data = Bits(64, 'input); val resp_data = Bits(64, 'input);
val resp_data_subword = Bits(64, 'input);
} }
class ioDpathImem extends Bundle() class ioDpathImem extends Bundle()
@ -113,9 +114,6 @@ class rocketDpath extends Component
val r_dmem_resp_val = Reg(resetVal = Bool(false)); val r_dmem_resp_val = Reg(resetVal = Bool(false));
val r_dmem_resp_waddr = Reg(resetVal = UFix(0,5)); val r_dmem_resp_waddr = Reg(resetVal = UFix(0,5));
val r_dmem_resp_pos = Reg(resetVal = UFix(0,3));
val r_dmem_resp_type = Reg(resetVal = Bits(0,3));
val r_dmem_resp_data = Reg(resetVal = Bits(0,64));
// instruction fetch stage // instruction fetch stage
val if_pc_plus4 = if_reg_pc + UFix(4); val if_pc_plus4 = if_reg_pc + UFix(4);
@ -194,52 +192,22 @@ class rocketDpath extends Component
Mux(io.ctrl.sel_wa === WA_RA, RA, Mux(io.ctrl.sel_wa === WA_RA, RA,
UFix(0, 5))))); UFix(0, 5)))));
// moved this here to avoid having to do forward declaration
// FIXME: cleanup
// 64/32 bit load handling (in mem stage)
val dmem_resp_pos = io.dmem.resp_tag(7,5).toUFix;
val dmem_resp_type = io.dmem.resp_tag(10,8);
val mem_dmem_resp_data_w =
Mux(dmem_resp_pos(2).toBool, io.dmem.resp_data(63, 32), io.dmem.resp_data(31, 0));
val mem_dmem_resp_data =
Mux(dmem_resp_type === MT_D, io.dmem.resp_data,
Mux(dmem_resp_type === MT_W, Cat(Fill(32, mem_dmem_resp_data_w(31)), mem_dmem_resp_data_w),
Cat(UFix(0,32), mem_dmem_resp_data_w)));
// crossbar/sign extension for 8/16 bit loads (in writeback stage)
val dmem_resp_data_h =
Mux(r_dmem_resp_pos(1).toBool, r_dmem_resp_data(31, 16), r_dmem_resp_data(15, 0));
val dmem_resp_data_b =
Mux(r_dmem_resp_pos(0).toBool, dmem_resp_data_h(15, 8), dmem_resp_data_h(7, 0));
val dmem_resp_data_final =
Mux(r_dmem_resp_type === MT_B, Cat(Fill(56, dmem_resp_data_b(7)), dmem_resp_data_b),
Mux(r_dmem_resp_type === MT_BU, Cat(UFix(0, 56), dmem_resp_data_b),
Mux(r_dmem_resp_type === MT_H, Cat(Fill(48, dmem_resp_data_h(15)), dmem_resp_data_h),
Mux(r_dmem_resp_type === MT_HU, Cat(UFix(0, 48), dmem_resp_data_h),
Mux((r_dmem_resp_type === MT_W) ||
(r_dmem_resp_type === MT_WU) ||
(r_dmem_resp_type === MT_D), r_dmem_resp_data,
UFix(0,64))))));
// bypass muxes // bypass muxes
val id_rs1 = val id_rs1 =
Mux(io.ctrl.div_wb, div_result, Mux(io.ctrl.div_wb, div_result,
Mux(io.ctrl.mul_wb, mul_result, Mux(io.ctrl.mul_wb, mul_result,
Mux(id_raddr1 != UFix(0, 5) && ex_reg_ctrl_wen && id_raddr1 === ex_reg_waddr, ex_wdata, Mux(id_raddr1 != UFix(0, 5) && ex_reg_ctrl_wen && id_raddr1 === ex_reg_waddr, ex_wdata,
Mux(id_raddr1 != UFix(0, 5) && mem_reg_ctrl_wen && id_raddr1 === mem_reg_waddr, mem_reg_wdata, Mux(id_raddr1 != UFix(0, 5) && mem_reg_ctrl_wen && id_raddr1 === mem_reg_waddr, mem_reg_wdata,
Mux(id_raddr1 != UFix(0, 5) && io.ctrl.mem_load && id_raddr1 === mem_reg_waddr, mem_dmem_resp_data, Mux(id_raddr1 != UFix(0, 5) && io.ctrl.mem_load && id_raddr1 === mem_reg_waddr, io.dmem.resp_data,
Mux(id_raddr1 != UFix(0, 5) && r_dmem_resp_val && id_raddr1 === r_dmem_resp_waddr, dmem_resp_data_final, Mux(id_raddr1 != UFix(0, 5) && r_dmem_resp_val && id_raddr1 === r_dmem_resp_waddr, io.dmem.resp_data_subword,
Mux(id_raddr1 != UFix(0, 5) && wb_reg_ctrl_wen && id_raddr1 === wb_reg_waddr, wb_reg_wdata, Mux(id_raddr1 != UFix(0, 5) && wb_reg_ctrl_wen && id_raddr1 === wb_reg_waddr, wb_reg_wdata,
id_rdata1))))))); id_rdata1)))))));
val id_rs2 = val id_rs2 =
Mux(id_raddr2 != UFix(0, 5) && ex_reg_ctrl_wen && id_raddr2 === ex_reg_waddr, ex_wdata, Mux(id_raddr2 != UFix(0, 5) && ex_reg_ctrl_wen && id_raddr2 === ex_reg_waddr, ex_wdata,
Mux(id_raddr2 != UFix(0, 5) && mem_reg_ctrl_wen && id_raddr2 === mem_reg_waddr, mem_reg_wdata, Mux(id_raddr2 != UFix(0, 5) && mem_reg_ctrl_wen && id_raddr2 === mem_reg_waddr, mem_reg_wdata,
Mux(id_raddr2 != UFix(0, 5) && io.ctrl.mem_load && id_raddr2 === mem_reg_waddr, mem_dmem_resp_data, Mux(id_raddr2 != UFix(0, 5) && io.ctrl.mem_load && id_raddr2 === mem_reg_waddr, io.dmem.resp_data,
Mux(id_raddr2 != UFix(0, 5) && r_dmem_resp_val && id_raddr2 === r_dmem_resp_waddr, dmem_resp_data_final, Mux(id_raddr2 != UFix(0, 5) && r_dmem_resp_val && id_raddr2 === r_dmem_resp_waddr, io.dmem.resp_data_subword,
Mux(id_raddr2 != UFix(0, 5) && wb_reg_ctrl_wen && id_raddr2 === wb_reg_waddr, wb_reg_wdata, Mux(id_raddr2 != UFix(0, 5) && wb_reg_ctrl_wen && id_raddr2 === wb_reg_waddr, wb_reg_wdata,
id_rdata2))))); id_rdata2)))));
@ -328,7 +296,7 @@ class rocketDpath extends Component
// other signals (req_val, req_rdy) connect to control module // other signals (req_val, req_rdy) connect to control module
io.dmem.req_addr := ex_alu_out(VADDR_BITS-1,0); io.dmem.req_addr := ex_alu_out(VADDR_BITS-1,0);
io.dmem.req_data := ex_reg_rs2; io.dmem.req_data := ex_reg_rs2;
io.dmem.req_tag := Cat(io.ctrl.ex_mem_type, io.dmem.req_addr(2,0), ex_reg_waddr).toUFix; io.dmem.req_tag := ex_reg_waddr;
// processor control regfile read // processor control regfile read
pcr.io.r.en := ex_reg_ctrl_ren_pcr | ex_reg_ctrl_eret; pcr.io.r.en := ex_reg_ctrl_ren_pcr | ex_reg_ctrl_eret;
@ -393,10 +361,7 @@ class rocketDpath extends Component
// writeback stage // writeback stage
r_dmem_resp_val <== io.dmem.resp_val; r_dmem_resp_val <== io.dmem.resp_val;
r_dmem_resp_waddr <== io.dmem.resp_tag(4,0).toUFix; r_dmem_resp_waddr <== io.dmem.resp_tag.toUFix
r_dmem_resp_pos <== dmem_resp_pos;
r_dmem_resp_type <== dmem_resp_type;
r_dmem_resp_data <== mem_dmem_resp_data;
wb_reg_waddr <== mem_reg_waddr; wb_reg_waddr <== mem_reg_waddr;
wb_reg_wdata <== mem_reg_wdata; wb_reg_wdata <== mem_reg_wdata;
@ -418,7 +383,7 @@ class rocketDpath extends Component
rfile.io.w1.addr := r_dmem_resp_waddr; rfile.io.w1.addr := r_dmem_resp_waddr;
rfile.io.w1.en := r_dmem_resp_val; rfile.io.w1.en := r_dmem_resp_val;
rfile.io.w1.data := dmem_resp_data_final; rfile.io.w1.data := io.dmem.resp_data_subword;
io.ctrl.wb_waddr := wb_reg_waddr; io.ctrl.wb_waddr := wb_reg_waddr;

18
rocket/src/main/scala/icache.scala
View File

@ -21,7 +21,7 @@ class ioImem(view: List[String] = null) extends Bundle (view)
// interface between I$ and memory (128 bits wide) // interface between I$ and memory (128 bits wide)
class ioIcache(view: List[String] = null) extends Bundle (view) class ioIcache(view: List[String] = null) extends Bundle (view)
{ {
val req_addr = UFix(PADDR_BITS, 'input); val req_addr = UFix(PADDR_BITS - OFFSET_BITS, 'input);
val req_val = Bool('input); val req_val = Bool('input);
val req_rdy = Bool('output); val req_rdy = Bool('output);
val resp_data = Bits(MEM_DATA_BITS, 'output); val resp_data = Bits(MEM_DATA_BITS, 'output);
@ -43,7 +43,7 @@ class rocketICacheDM(lines: Int) extends Component {
val addrbits = PADDR_BITS; val addrbits = PADDR_BITS;
val indexbits = ceil(log10(lines)/log10(2)).toInt; val indexbits = ceil(log10(lines)/log10(2)).toInt;
val offsetbits = 6; val offsetbits = OFFSET_BITS;
val tagmsb = addrbits - 1; val tagmsb = addrbits - 1;
val taglsb = indexbits+offsetbits; val taglsb = indexbits+offsetbits;
val tagbits = addrbits-taglsb; val tagbits = addrbits-taglsb;
@ -57,8 +57,8 @@ class rocketICacheDM(lines: Int) extends Component {
val s_reset :: s_ready :: s_request :: s_refill_wait :: s_refill :: s_resolve_miss :: Nil = Enum(6) { UFix() }; val s_reset :: s_ready :: s_request :: s_refill_wait :: s_refill :: s_resolve_miss :: Nil = Enum(6) { UFix() };
val state = Reg(resetVal = s_reset); val state = Reg(resetVal = s_reset);
val r_cpu_req_idx = Reg(resetVal = Bits(0, PGIDX_BITS)); val r_cpu_req_idx = Reg { Bits(width = PGIDX_BITS) }
val r_cpu_req_ppn = Reg(resetVal = Bits(0, PPN_BITS)); val r_cpu_req_ppn = Reg { Bits(width = PPN_BITS) }
val r_cpu_req_val = Reg(resetVal = Bool(false)); val r_cpu_req_val = Reg(resetVal = Bool(false));
when (io.cpu.req_val && io.cpu.req_rdy) { when (io.cpu.req_val && io.cpu.req_rdy) {
@ -114,13 +114,11 @@ class rocketICacheDM(lines: Int) extends Component {
io.cpu.resp_val := !io.cpu.itlb_miss && (state === s_ready) && r_cpu_req_val && tag_valid && tag_match; io.cpu.resp_val := !io.cpu.itlb_miss && (state === s_ready) && r_cpu_req_val && tag_valid && tag_match;
io.cpu.req_rdy := !io.cpu.itlb_miss && (state === s_ready) && (!r_cpu_req_val || (tag_valid && tag_match)); io.cpu.req_rdy := !io.cpu.itlb_miss && (state === s_ready) && (!r_cpu_req_val || (tag_valid && tag_match));
val word_mux = (new MuxN(REFILL_CYCLES)) { Bits(width = databits) } val test = Wire { Bits(width = MEM_DATA_BITS) }
word_mux.io.sel := r_cpu_req_idx(offsetmsb - rf_cnt_bits, offsetlsb).toUFix test <== data_array_rdata
for (i <- 0 to MEM_DATA_BITS/databits-1) { word_mux.io.in(i) := data_array_rdata((i+1)*databits-1, i*databits) } io.cpu.resp_data := Slice(MEM_DATA_BITS/databits, test, r_cpu_req_idx(offsetmsb-rf_cnt_bits,offsetlsb))
io.cpu.resp_data := word_mux.io.out
io.mem.req_val := (state === s_request); io.mem.req_val := (state === s_request);
io.mem.req_addr := Cat(r_cpu_req_ppn, r_cpu_req_idx(PGIDX_BITS-1, offsetbits), Bits(0, rf_cnt_bits)).toUFix; io.mem.req_addr := Cat(r_cpu_req_ppn, r_cpu_req_idx(indexmsb,indexlsb)).toUFix
// control state machine // control state machine
switch (state) { switch (state) {

6
rocket/src/main/scala/icache_prefetch.scala
View File

@ -7,7 +7,7 @@ import scala.math._;
class ioIPrefetcherMem(view: List[String] = null) extends Bundle (view) class ioIPrefetcherMem(view: List[String] = null) extends Bundle (view)
{ {
val req_addr = UFix(PADDR_BITS, 'output); val req_addr = UFix(PADDR_BITS - OFFSET_BITS, 'output);
val req_val = Bool('output); val req_val = Bool('output);
val req_rdy = Bool('input); val req_rdy = Bool('input);
val req_tag = Bits(IMEM_TAG_BITS, 'output); val req_tag = Bits(IMEM_TAG_BITS, 'output);
@ -29,8 +29,8 @@ class rocketIPrefetcher extends Component() {
val state = Reg(resetVal = s_invalid); val state = Reg(resetVal = s_invalid);
val demand_miss = io.icache.req_val & io.icache.req_rdy; val demand_miss = io.icache.req_val & io.icache.req_rdy;
val prefetch_addr = Reg() { UFix(width = PADDR_BITS) }; val prefetch_addr = Reg() { UFix(width = io.icache.req_addr.width) };
when (demand_miss) { prefetch_addr <== io.icache.req_addr + UFix(REFILL_CYCLES); } when (demand_miss) { prefetch_addr <== io.icache.req_addr + UFix(1); }
val addr_match = (prefetch_addr === io.icache.req_addr); val addr_match = (prefetch_addr === io.icache.req_addr);
val hit = (state != s_invalid) & (state != s_req_wait) & addr_match; val hit = (state != s_invalid) & (state != s_req_wait) & addr_match;

606
rocket/src/main/scala/nbdcache.scala
View File

@ -5,64 +5,104 @@ import Node._;
import Constants._; import Constants._;
import scala.math._; import scala.math._;
class rocketNBDCacheStoreGen extends Component { class StoreMaskGen extends Component {
val io = new Bundle { val io = new Bundle {
val req_type = Bits(3, 'input) val typ = Bits(3, 'input)
val req_addr_lsb = Bits(3, 'input) val addr = Bits(3, 'input)
val req_data = Bits(64, 'input) val wmask = Bits(8, 'output)
val store_wmask = Bits(8, 'output)
val store_data = Bits(64, 'output)
} }
// generate write mask and store data signals based on store type and address LSBs val mask = Wire { Bits(width = io.wmask.width) }
val wmask = Wire { Bits(8) } switch (io.typ(1,0))
switch (io.req_type(1,0))
{ {
is (MT_B) { wmask <== Bits( 1,1) << io.req_addr_lsb(2,0).toUFix } is (MT_B) { mask <== Bits( 1,1) << io.addr(2,0).toUFix }
is (MT_H) { wmask <== Bits( 3,2) << Cat(io.req_addr_lsb(2,1), Bits(0,1)).toUFix } is (MT_H) { mask <== Bits( 3,2) << Cat(io.addr(2,1), Bits(0,1)).toUFix }
is (MT_W) { wmask <== Bits( 15,4) << Cat(io.req_addr_lsb(2,2), Bits(0,2)).toUFix } is (MT_W) { mask <== Bits( 15,4) << Cat(io.addr(2,2), Bits(0,2)).toUFix }
otherwise { wmask <== Bits(255,8) } // MT_D otherwise { mask <== Bits(255,8) } // MT_D
} }
io.wmask := mask
val data = Wire { Bits(64) }
switch (io.req_type(1,0))
{
is (MT_B) { data <== Fill(8, io.req_data( 7,0)) }
is (MT_H) { data <== Fill(4, io.req_data(15,0)) }
is (MT_W) { data <== Fill(2, io.req_data(31,0)) }
otherwise { data <== io.req_data } // MT_D
}
io.store_wmask := wmask
io.store_data := data
} }
class DataMemCmd extends Bundle { class StoreDataGen extends Component {
val offset = Bits(width = OFFSET_BITS) val io = new Bundle {
val cmd = Bits(width = 4) val typ = Bits(3, 'input)
val typ = Bits(width = 3) val addr = Bits(3, 'input)
val din = Bits(64, 'input)
val dout = Bits(64, 'output)
}
val data = Wire { Bits(width = io.din.width) }
switch (io.typ(1,0))
{
is (MT_B) { data <== Fill(8, io.din( 7,0)) }
is (MT_H) { data <== Fill(4, io.din(15,0)) }
is (MT_W) { data <== Fill(2, io.din(31,0)) }
otherwise { data <== io.din } // MT_D
}
io.dout := data
}
class LoadDataGen extends Component {
val io = new Bundle {
val typ = Bits(3, 'input)
val addr = Bits(3, 'input)
val din = Bits(64, 'input)
val dout = Bits(64, 'output)
val dout_subword = Bits(64, 'output)
}
val shifted = io.din >> Cat(io.addr(2), Bits(0, 5)).toUFix
val extended = Wire { Bits(width = io.din.width) }
switch (io.typ)
{
is(MT_W) { extended <== Cat(Fill(32, shifted(31)), shifted(31,0)) }
is(MT_WU) { extended <== Cat(Bits(0, 32), shifted(31,0)) }
otherwise { extended <== shifted }
}
val shifted_subword = Reg(extended) >> Cat(Reg(io.addr(1,0)), Bits(0, 3)).toUFix
val extended_subword = Wire { Bits(width = io.din.width) }
switch (Reg(io.typ))
{
is (MT_B) { extended_subword <== Cat(Fill(56, shifted_subword( 7)), shifted_subword( 7, 0)) }
is (MT_BU) { extended_subword <== Cat(Bits(0, 56), shifted_subword( 7, 0)) }
is (MT_H) { extended_subword <== Cat(Fill(48, shifted_subword(15)), shifted_subword(15, 0)) }
is (MT_HU) { extended_subword <== Cat(Bits(0, 48), shifted_subword(15, 0)) }
otherwise { extended_subword <== shifted_subword }
}
io.dout := extended
io.dout_subword := extended_subword
} }
class RPQEntry extends Bundle { class RPQEntry extends Bundle {
val cmd = new DataMemCmd() val offset = Bits(width = OFFSET_BITS)
val sdq_id = UFix(width = ceil(log(NSDQ)/log(2)).toInt) val cmd = Bits(width = 4)
val typ = Bits(width = 3)
val sdq_id = UFix(width = log2up(NSDQ))
val tag = Bits(width = CPU_TAG_BITS)
} }
class Replay extends Bundle { class Replay extends Bundle {
val idx = Bits(width = IDX_BITS) val idx = Bits(width = IDX_BITS)
val cmd = new DataMemCmd() val offset = Bits(width = OFFSET_BITS)
val sdq_id = UFix(width = ceil(log(NSDQ)/log(2)).toInt) val cmd = Bits(width = 4)
val typ = Bits(width = 3)
val sdq_id = UFix(width = log2up(NSDQ))
val tag = Bits(width = CPU_TAG_BITS)
} }
class DataReq extends Bundle { class DataReq extends Bundle {
val idx = Bits(width = IDX_BITS) val idx = Bits(width = IDX_BITS)
val cmd = new DataMemCmd() val offset = Bits(width = IDX_BITS)
val cmd = Bits(width = 4)
val typ = Bits(width = 3)
val data = Bits(width = CPU_DATA_BITS) val data = Bits(width = CPU_DATA_BITS)
} }
class DataArrayReq extends Bundle { class DataArrayReq extends Bundle {
val idx = Bits(width = IDX_BITS) val idx = Bits(width = IDX_BITS)
val offset = Bits(width = ceil(log(REFILL_CYCLES)/log(2)).toInt) val offset = Bits(width = log2up(REFILL_CYCLES))
val rw = Bool() val rw = Bool()
val wmask = Bits(width = MEM_DATA_BITS/8) val wmask = Bits(width = MEM_DATA_BITS/8)
val data = Bits(width = MEM_DATA_BITS) val data = Bits(width = MEM_DATA_BITS)
@ -70,7 +110,8 @@ class DataArrayReq extends Bundle {
class MemReq extends Bundle { class MemReq extends Bundle {
val rw = Bool() val rw = Bool()
val addr = Bits(width = PPN_BITS+IDX_BITS) val addr = UFix(width = PPN_BITS+IDX_BITS)
val tag = Bits(width = DMEM_TAG_BITS)
} }
class WritebackReq extends Bundle { class WritebackReq extends Bundle {
@ -90,7 +131,7 @@ class MetaArrayReq extends Bundle {
val data = new MetaData() val data = new MetaData()
} }
class MSHR extends Component { class MSHR(id: Int) extends Component {
val io = new Bundle { val io = new Bundle {
val req_pri_val = Bool('input) val req_pri_val = Bool('input)
val req_pri_rdy = Bool('output) val req_pri_rdy = Bool('output)
@ -98,10 +139,14 @@ class MSHR extends Component {
val req_sec_rdy = Bool('output) val req_sec_rdy = Bool('output)
val req_ppn = Bits(PPN_BITS, 'input) val req_ppn = Bits(PPN_BITS, 'input)
val req_idx = Bits(IDX_BITS, 'input) val req_idx = Bits(IDX_BITS, 'input)
val req_cmd = new RPQEntry().asInput val req_offset = Bits(width = OFFSET_BITS)
val req_cmd = Bits(width = 4)
val req_type = Bits(width = 3)
val req_sdq_id = UFix(width = log2up(NSDQ))
val req_tag = Bits(CPU_TAG_BITS, 'input) val req_tag = Bits(CPU_TAG_BITS, 'input)
val idx_match = Bool('output) val idx_match = Bool('output)
val idx = Bits(IDX_BITS, 'output)
val tag = Bits(PPN_BITS, 'output) val tag = Bits(PPN_BITS, 'output)
val mem_resp_val = Bool('input) val mem_resp_val = Bool('input)
@ -115,17 +160,21 @@ class MSHR extends Component {
val requested = Reg { Bool() } val requested = Reg { Bool() }
val refilled = Reg { Bool() } val refilled = Reg { Bool() }
val ppn = Reg { Bits() } val ppn = Reg { Bits() }
val idx = Reg { Bits() } val idx_ = Reg { Bits() }
val req_load = (io.req_cmd.cmd.cmd === M_XRD) || (io.req_cmd.cmd.cmd === M_PFR) val req_load = (io.req_cmd === M_XRD) || (io.req_cmd === M_PFR)
val req_use_rpq = (io.req_cmd.cmd.cmd != M_PFR) && (io.req_cmd.cmd.cmd != M_PFW) val req_use_rpq = (io.req_cmd != M_PFR) && (io.req_cmd != M_PFW)
val next_dirty = io.req_pri_val && io.req_pri_rdy && !req_load || io.req_sec_val && io.req_sec_rdy && (!req_load || dirty) val next_dirty = io.req_pri_val && io.req_pri_rdy && !req_load || io.req_sec_val && io.req_sec_rdy && (!req_load || dirty)
val sec_rdy = io.idx_match && !refilled && (dirty || !requested || req_load) val sec_rdy = io.idx_match && !refilled && (dirty || !requested || req_load)
val rpq = (new queueSimplePF(NRPQ)) { new RPQEntry() } val rpq = (new queueSimplePF(NRPQ)) { new RPQEntry() }
rpq.io.q_reset := Bool(false) rpq.io.q_reset := Bool(false)
rpq.io.enq.valid := (io.req_pri_val && io.req_pri_rdy || io.req_sec_val && sec_rdy) && req_use_rpq rpq.io.enq.valid := (io.req_pri_val && io.req_pri_rdy || io.req_sec_val && sec_rdy) && req_use_rpq
io.req_cmd ^^ rpq.io.enq.bits rpq.io.enq.bits.offset := io.req_offset
rpq.io.enq.bits.cmd := io.req_cmd
rpq.io.enq.bits.typ := io.req_type
rpq.io.enq.bits.sdq_id := io.req_sdq_id
rpq.io.enq.bits.tag := io.req_tag
rpq.io.deq.ready := io.replay.ready && refilled rpq.io.deq.ready := io.replay.ready && refilled
when (io.req_pri_val && io.req_pri_rdy) { when (io.req_pri_val && io.req_pri_rdy) {
@ -133,7 +182,7 @@ class MSHR extends Component {
requested <== Bool(false) requested <== Bool(false)
refilled <== Bool(false) refilled <== Bool(false)
ppn <== io.req_ppn ppn <== io.req_ppn
idx <== io.req_idx idx_ <== io.req_idx
} }
when (io.mem_req.valid && io.mem_req.ready) { when (io.mem_req.valid && io.mem_req.ready) {
requested <== Bool(true) requested <== Bool(true)
@ -146,14 +195,15 @@ class MSHR extends Component {
} }
dirty <== next_dirty dirty <== next_dirty
io.idx_match := valid && (idx === io.req_idx) io.idx_match := valid && (idx_ === io.req_idx)
io.idx := idx_
io.tag := ppn io.tag := ppn
io.req_pri_rdy := !valid io.req_pri_rdy := !valid
io.req_sec_rdy := sec_rdy && rpq.io.enq.ready io.req_sec_rdy := sec_rdy && rpq.io.enq.ready
io.meta_req.valid := valid && refilled && !rpq.io.deq.valid io.meta_req.valid := valid && refilled && !rpq.io.deq.valid
io.meta_req.bits.rw := Bool(true) io.meta_req.bits.rw := Bool(true)
io.meta_req.bits.idx := idx io.meta_req.bits.idx := idx_
io.meta_req.bits.data.valid := Bool(true) io.meta_req.bits.data.valid := Bool(true)
io.meta_req.bits.data.dirty := dirty io.meta_req.bits.data.dirty := dirty
io.meta_req.bits.data.tag := ppn io.meta_req.bits.data.tag := ppn
@ -161,11 +211,15 @@ class MSHR extends Component {
io.mem_req.valid := valid && !requested io.mem_req.valid := valid && !requested
//io.mem_req.bits.itm := next_dirty //io.mem_req.bits.itm := next_dirty
io.mem_req.bits.rw := Bool(false) io.mem_req.bits.rw := Bool(false)
io.mem_req.bits.addr := Cat(ppn, idx) io.mem_req.bits.addr := Cat(ppn, idx_).toUFix
io.mem_req.bits.tag := Bits(id)
io.replay.valid := rpq.io.deq.valid && refilled io.replay.valid := rpq.io.deq.valid && refilled
io.replay.bits.idx := idx io.replay.bits.idx := idx_
rpq.io.deq.bits.cmd ^^ io.replay.bits.cmd io.replay.bits.tag := rpq.io.deq.bits.tag
io.replay.bits.offset := rpq.io.deq.bits.offset
io.replay.bits.cmd := rpq.io.deq.bits.cmd
io.replay.bits.typ := rpq.io.deq.bits.typ
io.replay.bits.sdq_id := rpq.io.deq.bits.sdq_id io.replay.bits.sdq_id := rpq.io.deq.bits.sdq_id
} }
@ -173,14 +227,18 @@ class MSHRFile extends Component {
val io = new Bundle { val io = new Bundle {
val req_val = Bool('input) val req_val = Bool('input)
val req_rdy = Bool('output) val req_rdy = Bool('output)
val req_cmd = (new DataMemCmd).asInput val req_ppn = Bits(PPN_BITS, 'input)
val req_ppn = Bits(PADDR_BITS, 'input)
val req_idx = Bits(IDX_BITS, 'input) val req_idx = Bits(IDX_BITS, 'input)
val req_data = Bits(64, 'input) val req_offset = Bits(OFFSET_BITS, 'input)
val req_cmd = Bits(4, 'input)
val req_type = Bits(3, 'input)
val req_data = Bits(CPU_DATA_BITS, 'input)
val req_tag = Bits(CPU_TAG_BITS, 'input) val req_tag = Bits(CPU_TAG_BITS, 'input)
val req_sdq_id = UFix(log2up(NSDQ), 'input)
val mem_resp_val = Bool('input) val mem_resp_val = Bool('input)
val mem_resp_tag = Bits(DMEM_TAG_BITS, 'input) val mem_resp_tag = Bits(DMEM_TAG_BITS, 'input)
val mem_resp_idx = Bits(IDX_BITS, 'output)
val mem_req = (new ioDecoupled) { new MemReq() }.flip() val mem_req = (new ioDecoupled) { new MemReq() }.flip()
val meta_req = (new ioDecoupled) { new MetaArrayReq() }.flip() val meta_req = (new ioDecoupled) { new MetaArrayReq() }.flip()
@ -192,39 +250,42 @@ class MSHRFile extends Component {
val sec_rdy = Wire { Bool() } val sec_rdy = Wire { Bool() }
val tag_mux = new Mux1H(NMSHR, PPN_BITS) val tag_mux = new Mux1H(NMSHR, PPN_BITS)
val mem_resp_idx_mux = new Mux1H(NMSHR, IDX_BITS)
val meta_req_arb = (new Arbiter(NMSHR)) { new MetaArrayReq() } val meta_req_arb = (new Arbiter(NMSHR)) { new MetaArrayReq() }
val mem_req_arb = (new Arbiter(NMSHR)) { new MemReq() } val mem_req_arb = (new Arbiter(NMSHR)) { new MemReq() }
val replay_arb = (new Arbiter(NMSHR)) { new RPQEntry() } val replay_arb = (new Arbiter(NMSHR)) { new Replay() }
val alloc_arb = (new Arbiter(NMSHR)) { Bool() } val alloc_arb = (new Arbiter(NMSHR)) { Bool() }
alloc_arb.io.out.ready := io.req_val && !idx_match
val tag_match = tag_mux.io.out === io.req_ppn val tag_match = tag_mux.io.out === io.req_ppn
for (i <- 0 to NMSHR-1) { for (i <- 0 to NMSHR-1) {
val mshr = new MSHR() val mshr = new MSHR(i)
val rpqe = new RPQEntry().asInput
rpqe.cmd.offset <== io.req_cmd.offset
rpqe.cmd.cmd <== io.req_cmd.cmd
rpqe.cmd.typ <== io.req_cmd.typ
rpqe.sdq_id <== UFix(0)
tag_mux.io.sel(i) := mshr.io.idx_match tag_mux.io.sel(i) := mshr.io.idx_match
tag_mux.io.in(i) := mshr.io.tag tag_mux.io.in(i) := mshr.io.tag
alloc_arb.io.in(i).valid := mshr.io.req_pri_rdy alloc_arb.io.in(i).valid := mshr.io.req_pri_rdy
mshr.io.req_pri_val := io.req_val && !idx_match && alloc_arb.io.in(i).ready mshr.io.req_pri_val := alloc_arb.io.in(i).ready
mshr.io.req_sec_val := io.req_val && tag_match mshr.io.req_sec_val := io.req_val && tag_match
mshr.io.req_ppn := io.req_ppn mshr.io.req_ppn := io.req_ppn
mshr.io.req_idx := io.req_idx
mshr.io.req_tag := io.req_tag mshr.io.req_tag := io.req_tag
rpqe ^^ mshr.io.req_cmd mshr.io.req_idx := io.req_idx
mshr.io.req_offset := io.req_offset
mshr.io.req_cmd := io.req_cmd
mshr.io.req_type := io.req_type
mshr.io.req_sdq_id := io.req_sdq_id
mshr.io.meta_req <> meta_req_arb.io.in(i) mshr.io.meta_req <> meta_req_arb.io.in(i)
mshr.io.mem_req <> mem_req_arb.io.in(i) mshr.io.mem_req <> mem_req_arb.io.in(i)
mshr.io.replay <> replay_arb.io.in(i) mshr.io.replay <> replay_arb.io.in(i)
mshr.io.mem_resp_val := io.mem_resp_val && (UFix(i) === io.mem_resp_tag) val mem_resp_val = io.mem_resp_val && (UFix(i) === io.mem_resp_tag)
mshr.io.mem_resp_val := mem_resp_val
mem_resp_idx_mux.io.sel(i) := mem_resp_val
mem_resp_idx_mux.io.in(i) := mshr.io.idx
when (mshr.io.req_pri_rdy) { pri_rdy <== Bool(true) } when (mshr.io.req_pri_rdy) { pri_rdy <== Bool(true) }
when (mshr.io.req_sec_rdy) { sec_rdy <== Bool(true) } when (mshr.io.req_sec_rdy) { sec_rdy <== Bool(true) }
@ -239,80 +300,101 @@ class MSHRFile extends Component {
replay_arb.io.out ^^ io.replay replay_arb.io.out ^^ io.replay
io.req_rdy := Mux(idx_match, tag_match && sec_rdy, pri_rdy) io.req_rdy := Mux(idx_match, tag_match && sec_rdy, pri_rdy)
io.mem_resp_idx := mem_resp_idx_mux.io.out
} }
class StoreDataUnit extends Component { class ReplayUnit extends Component {
val io = new Bundle { val io = new Bundle {
val sdq_enq = (new ioDecoupled) { Bits(width = CPU_DATA_BITS) } val sdq_enq = (new ioDecoupled) { Bits(width = CPU_DATA_BITS) }
val sdq_id = UFix(width = ceil(log(NSDQ)/log(2)).toInt, dir = 'output) val sdq_id = UFix(log2up(NSDQ), 'output)
val replay = (new ioDecoupled) { new Replay() } val replay = (new ioDecoupled) { new Replay() }
val data_req = (new ioDecoupled) { new DataReq() }.flip() val data_req = (new ioDecoupled) { new DataReq() }.flip()
val cpu_resp_val = Bool('output)
val cpu_resp_tag = Bits(CPU_TAG_BITS, 'output)
} }
val cmdq = (new queueSimplePF(2)) { new Replay() } val sdq_val = Reg(resetVal = UFix(0, NSDQ))
val dataq = (new queueSimplePF(2)) { Bits(width = CPU_DATA_BITS) } val sdq_allocator = new priorityEncoder(NSDQ)
sdq_allocator.io.in := ~sdq_val
val sdq_alloc_id = sdq_allocator.io.out.toUFix
val next_dataq_enq_rdy = !dataq.io.deq.valid || dataq.io.enq.ready && (!dataq.io.enq.valid || dataq.io.deq.ready) val replay_retry = Wire { Bool() }
val next_dataq_enq_val = io.replay.valid && next_dataq_enq_rdy && (io.replay.bits.cmd.cmd != M_XRD) && cmdq.io.enq.ready val replay_val = Reg(io.replay.valid || replay_retry, resetVal = Bool(false))
dataq.io.enq.valid := Reg(next_dataq_enq_val, resetVal = Bool(false)) replay_retry <== replay_val && !io.data_req.ready
dataq.io.enq.bits := sdq_dout
dataq.io.deq.ready := io.data_req.ready && (cmdq.io.deq.bits.cmd.cmd != M_XRD)
cmdq.io.enq.valid := io.replay.valid && ((io.replay.bits.cmd.cmd === M_XRD) || next_dataq_enq_rdy) val rp = Reg { new Replay() }
io.replay.bits ^^ cmdq.io.enq.bits when (io.replay.valid && io.replay.ready) { rp <== io.replay.bits }
cmdq.io.deq.ready := io.data_req.ready && ((cmdq.io.deq.bits.cmd.cmd === M_XRD) || dataq.io.deq.valid)
val sdq = Mem4(NSDQ, io.sdq_enq.bits); val rp_amo = rp.cmd(3).toBool
sdq.setReadLatency(1); val rp_store = (rp.cmd === M_XWR)
sdq.setTarget('inst); val rp_load = (rp.cmd === M_XRD)
val sdq_addr = Mux(next_dataq_enq_val, io.replay.bits.sdq_id, io.sdq_id) val rp_write = rp_store || rp_amo
val rp_read = rp_load || rp_amo
val sdq_ren_new = io.replay.valid && (io.replay.bits.cmd != M_XRD)
val sdq_ren_retry = replay_retry && rp_write
val sdq_ren = sdq_ren_new || sdq_ren_retry
val sdq_wen = io.sdq_enq.valid && io.sdq_enq.ready val sdq_wen = io.sdq_enq.valid && io.sdq_enq.ready
val sdq_dout = sdq.rw(sdq_addr, io.sdq_enq.bits, sdq_wen, cs = next_dataq_enq_val || sdq_wen); val sdq_addr = Mux(sdq_ren_retry, rp.sdq_id, Mux(sdq_ren_new, io.replay.bits.sdq_id, sdq_alloc_id))
val sdq_val = Reg(resetVal = Bits(0, ceil(log(NSDQ)/log(2)).toInt))
when (next_dataq_enq_val) { sdq_val <== sdq_val.bitSet(io.replay.bits.sdq_id, Bool(false)) }
when (sdq_wen) { sdq_val <== sdq_val.bitSet(io.sdq_id, Bool(true)) }
def priority_enc(in: Bits, n: Int = 0): Bits = if (in.width == n-1) UFix(n-1) else if(in(n) == Bool(true)) UFix(n) else priority_enc(in, n+1) val sdq = Mem4(NSDQ, io.sdq_enq.bits)
io.sdq_id := priority_enc(~sdq_val) sdq.setReadLatency(0)
io.sdq_enq.ready := ((~sdq_val) != UFix(0)) && !next_dataq_enq_val sdq.setTarget('inst)
io.replay.ready := cmdq.io.enq.ready && next_dataq_enq_rdy val sdq_dout = sdq.rw(sdq_addr, io.sdq_enq.bits, sdq_wen, cs = sdq_ren || sdq_wen)
io.data_req.valid := cmdq.io.deq.valid && ((cmdq.io.deq.bits.cmd.cmd === M_XRD) || dataq.io.deq.valid)
io.data_req.bits.idx := cmdq.io.deq.bits.idx val sdq_free = replay_val && !replay_retry && rp_write
cmdq.io.deq.bits.cmd ^^ io.data_req.bits.cmd sdq_val <== sdq_val & ~(sdq_free.toUFix << rp.sdq_id) | (sdq_wen.toUFix << sdq_alloc_id)
io.data_req.bits.data := dataq.io.deq.bits
io.sdq_enq.ready := (~sdq_val != UFix(0)) && !sdq_ren
io.sdq_id := sdq_alloc_id
io.replay.ready := !replay_retry
io.data_req.valid := replay_val
io.data_req.bits.idx := rp.idx
io.data_req.bits.offset := rp.offset
io.data_req.bits.cmd := rp.cmd
io.data_req.bits.typ := rp.typ
io.data_req.bits.data := sdq_dout
io.cpu_resp_val := Reg(replay_val && !replay_retry && rp_read, resetVal = Bool(false))
io.cpu_resp_tag := Reg(rp.tag)
} }
class WritebackUnit extends Component { class WritebackUnit extends Component {
val io = new Bundle { val io = new Bundle {
val req = (new ioDecoupled) { new WritebackReq() } val req = (new ioDecoupled) { new WritebackReq() }
val data_req = (new ioDecoupled) { new DataReq() }.flip() val data_req = (new ioDecoupled) { new DataArrayReq() }.flip()
val data_resp = Bits(width = MEM_DATA_BITS, dir = 'input) val data_resp = Bits(MEM_DATA_BITS, 'input)
val mem_req = (new ioDecoupled) { new MemReq() }.flip() val mem_req = (new ioDecoupled) { new MemReq() }.flip()
val mem_req_data = Bits(MEM_DATA_BITS, 'output)
} }
val wbq = (new queueSimplePF(REFILL_CYCLES)) { Bits(width = MEM_DATA_BITS) } val wbq = (new queueSimplePF(REFILL_CYCLES)) { Bits(width = MEM_DATA_BITS) }
val valid = Reg(resetVal = Bool(false)) val valid = Reg(resetVal = Bool(false))
val cnt = Reg() { UFix(width = ceil(log(REFILL_CYCLES)/log(2)).toInt) } val cnt = Reg() { UFix(width = log2up(REFILL_CYCLES+1)) }
val addr = Reg() { new WritebackReq() } val addr = Reg() { new WritebackReq() }
wbq.io.enq.valid := valid && Reg(io.data_req.valid && io.data_req.ready) wbq.io.enq.valid := valid && Reg(io.data_req.valid && io.data_req.ready)
wbq.io.enq.bits := io.data_resp wbq.io.enq.bits := io.data_resp
wbq.io.deq.ready := io.mem_req.ready && (~cnt === UFix(0)) wbq.io.deq.ready := io.mem_req.ready && (cnt === UFix(REFILL_CYCLES))
when (io.req.valid && io.req.ready) { valid <== Bool(true); cnt <== UFix(0); addr <== io.req.bits } when (io.req.valid && io.req.ready) { valid <== Bool(true); cnt <== UFix(0); addr <== io.req.bits }
when (io.data_req.valid && io.data_req.ready) { cnt <== cnt + UFix(1) } when (io.data_req.valid && io.data_req.ready) { cnt <== cnt + UFix(1) }
when ((~cnt === UFix(0)) && !wbq.io.deq.valid) { valid <== Bool(false) } when ((cnt === UFix(REFILL_CYCLES)) && !wbq.io.deq.valid) { valid <== Bool(false) }
io.req.ready := !valid io.req.ready := !valid
io.data_req.valid := valid && wbq.io.enq.ready io.data_req.valid := valid && wbq.io.enq.ready
io.data_req.bits.idx := addr.idx io.data_req.bits.idx := addr.idx
io.data_req.bits.cmd.offset := cnt * UFix(MEM_DATA_BITS/8) io.data_req.bits.offset := cnt
io.data_req.bits.cmd.cmd := M_XRD io.data_req.bits.rw := Bool(false)
io.data_req.bits.cmd.typ := UFix(0) io.data_req.bits.wmask := Bits(0)
io.data_req.bits.data := wbq.io.deq.bits io.data_req.bits.data := Bits(0)
io.mem_req.valid := wbq.io.deq.valid && (~cnt === UFix(0)) io.mem_req.valid := wbq.io.deq.valid && (cnt === UFix(REFILL_CYCLES))
io.mem_req.bits.rw := Bool(true) io.mem_req.bits.rw := Bool(true)
io.mem_req.bits.addr := Cat(addr.ppn, addr.idx) io.mem_req.bits.addr := Cat(addr.ppn, addr.idx).toUFix
io.mem_req.bits.tag := Bits(0)
io.mem_req_data := wbq.io.deq.bits
} }
class FlushUnit(lines: Int) extends Component { class FlushUnit(lines: Int) extends Component {
@ -324,14 +406,13 @@ class FlushUnit(lines: Int) extends Component {
val wb_req = (new ioDecoupled) { new WritebackReq() }.flip() val wb_req = (new ioDecoupled) { new WritebackReq() }.flip()
} }
val s_reset0 :: s_reset :: s_ready :: s_meta_read :: s_meta_wait :: s_meta_write :: s_done :: Nil = Enum(7) { UFix() } val s_reset :: s_ready :: s_meta_read :: s_meta_wait :: s_meta_write :: s_done :: Nil = Enum(6) { UFix() }
val state = Reg(resetVal = s_reset0) val state = Reg(resetVal = s_reset)
val tag = Reg() { Bits(width = CPU_TAG_BITS) } val tag = Reg() { Bits() }
val cnt = Reg() { UFix(width = ceil(log(lines)/log(2)).toInt) } val cnt = Reg(resetVal = UFix(0, log2up(lines)))
val next_cnt = cnt + UFix(1) val next_cnt = cnt + UFix(1)
switch (state) { switch (state) {
is(s_reset0) { state <== s_reset; cnt <== UFix(0) }
is(s_reset) { when (io.meta_req.ready) { state <== Mux(~cnt === UFix(0), s_ready, s_reset); cnt <== next_cnt } } is(s_reset) { when (io.meta_req.ready) { state <== Mux(~cnt === UFix(0), s_ready, s_reset); cnt <== next_cnt } }
is(s_ready) { when (io.req.valid) { state <== s_meta_read; tag <== io.req.bits } } is(s_ready) { when (io.req.valid) { state <== s_meta_read; tag <== io.req.bits } }
is(s_meta_read) { when (io.meta_req.ready) { state <== s_meta_wait } } is(s_meta_read) { when (io.meta_req.ready) { state <== s_meta_wait } }
@ -357,13 +438,25 @@ class MetaDataArray(lines: Int) extends Component {
val io = new Bundle { val io = new Bundle {
val req = (new ioDecoupled) { new MetaArrayReq() } val req = (new ioDecoupled) { new MetaArrayReq() }
val resp = (new MetaData).asOutput() val resp = (new MetaData).asOutput()
val state_req = (new ioDecoupled) { new MetaArrayReq() }
} }
val array = Mem4(lines, io.resp) val vd_array = Mem4(lines, Bits(width = 2))
array.setReadLatency(1) vd_array.setReadLatency(0)
array.setTarget('inst) val vd_wdata1 = Cat(io.req.bits.data.valid, io.req.bits.data.dirty)
val rdata = array.rw(io.req.bits.idx, io.req.bits.data, io.req.valid && io.req.bits.rw, cs = io.req.valid) val vd_rdata1 = vd_array.rw(io.req.bits.idx, vd_wdata1, io.req.valid && io.req.bits.rw)
rdata ^^ io.resp val vd_wdata2 = Cat(io.state_req.bits.data.valid, io.req.bits.data.dirty)
vd_array.write(io.state_req.bits.idx, vd_wdata2, io.state_req.valid && io.state_req.bits.rw)
val tag_array = Mem4(lines, io.resp.tag)
tag_array.setReadLatency(0)
tag_array.setTarget('inst)
val tag_rdata = tag_array.rw(io.req.bits.idx, io.req.bits.data.tag, io.req.valid && io.req.bits.rw, cs = io.req.valid)
io.resp.valid := vd_rdata1(1).toBool
io.resp.dirty := vd_rdata1(0).toBool
io.resp.tag := tag_rdata
io.req.ready := Bool(true) io.req.ready := Bool(true)
} }
@ -373,19 +466,14 @@ class DataArray(lines: Int) extends Component {
val resp = Bits(width = MEM_DATA_BITS, dir = 'output) val resp = Bits(width = MEM_DATA_BITS, dir = 'output)
} }
val wmask_array = Vec(MEM_DATA_BITS/8) { Wire() { Bits(width = MEM_DATA_BITS) } } val wmask = FillInterleaved(8, io.req.bits.wmask)
wmask_array(0) <== Fill(8, io.req.bits.wmask(0))
for (i <- 1 to MEM_DATA_BITS/8-1) {
wmask_array(i) <== Cat(Fill(8, io.req.bits.wmask(i)), wmask_array(i-1)(8*(i+1)-1, 8*i))
}
val wmask = wmask_array(MEM_DATA_BITS/8-1)
val array = Mem4(lines*REFILL_CYCLES, io.resp) val array = Mem4(lines*REFILL_CYCLES, io.resp)
array.setReadLatency(1) array.setReadLatency(0)
array.setTarget('inst) array.setTarget('inst)
val addr = Cat(io.req.bits.idx, io.req.bits.offset) val addr = Cat(io.req.bits.idx, io.req.bits.offset)
val rdata = array.rw(addr, io.req.bits.data, io.req.valid && io.req.bits.rw, wmask, cs = io.req.valid) val rdata = array.rw(addr, io.req.bits.data, io.req.valid && io.req.bits.rw, wmask, cs = io.req.valid)
rdata ^^ io.resp io.resp := rdata
io.req.ready := Bool(true) io.req.ready := Bool(true)
} }
@ -413,7 +501,7 @@ class rocketNBDCacheAMOALU extends Component {
val less = Mux(cmp_lhs === cmp_rhs, cmp_diff, Mux(signed, cmp_lhs, cmp_rhs)) val less = Mux(cmp_lhs === cmp_rhs, cmp_diff, Mux(signed, cmp_lhs, cmp_rhs))
val cmp_out = Mux(min === less, io.lhs, io.rhs) val cmp_out = Mux(min === less, io.lhs, io.rhs)
val alu_out = Wire() { UFix() }; val alu_out = Wire { UFix(width = io.result.width) };
switch (io.cmd) { switch (io.cmd) {
is (M_XA_ADD) { alu_out <== adder_out } is (M_XA_ADD) { alu_out <== adder_out }
is (M_XA_SWAP) { alu_out <== io.rhs } is (M_XA_SWAP) { alu_out <== io.rhs }
@ -425,7 +513,7 @@ class rocketNBDCacheAMOALU extends Component {
io.result := alu_out io.result := alu_out
} }
// XXX broken for CPU_DATA_WIDTH != 64 // XXX broken for CPU_DATA_BITS != 64
class AMOUnit extends Component { class AMOUnit extends Component {
val io = new Bundle { val io = new Bundle {
val req = (new ioDecoupled) { new DataReq() } val req = (new ioDecoupled) { new DataReq() }
@ -436,17 +524,21 @@ class AMOUnit extends Component {
} }
val valid = Reg(resetVal = Bool(false)) val valid = Reg(resetVal = Bool(false))
val r_cmd = Reg() { new DataMemCmd() } val r_cmd = Reg() { Bits() }
val r_idx = Reg() { Bits(width = IDX_BITS) } val r_offset = Reg() { Bits() }
val r_lhs = Reg() { Bits(width = 64) } val r_type = Reg() { Bits() }
val r_rhs = Reg() { Bits(width = 64) } val r_idx = Reg() { Bits() }
val r_wmask = Reg() { Bits(width = 64/8) } val r_lhs = Reg() { Bits() }
val r_rhs = Reg() { Bits() }
val r_wmask = Reg() { Bits() }
when (io.req.valid && io.req.ready) { when (io.req.valid && io.req.ready) {
valid <== Bool(true); valid <== Bool(true);
r_idx <== io.req.bits.idx r_idx <== io.req.bits.idx
r_lhs <== io.lhs; r_lhs <== io.lhs;
r_rhs <== io.rhs; r_rhs <== io.rhs;
r_cmd <== io.req.bits.cmd; r_cmd <== io.req.bits.cmd;
r_type <== io.req.bits.typ;
r_offset <== io.req.bits.offset;
r_wmask <== io.wmask r_wmask <== io.wmask
} }
when (io.data_req.valid && io.data_req.ready) { when (io.data_req.valid && io.data_req.ready) {
@ -454,7 +546,7 @@ class AMOUnit extends Component {
} }
val alu = new rocketNBDCacheAMOALU val alu = new rocketNBDCacheAMOALU
alu.io.cmd := r_cmd.cmd alu.io.cmd := r_cmd
alu.io.wmask := r_wmask alu.io.wmask := r_wmask
alu.io.lhs := r_lhs alu.io.lhs := r_lhs
alu.io.rhs := r_rhs alu.io.rhs := r_rhs
@ -462,7 +554,9 @@ class AMOUnit extends Component {
io.req.ready := !valid io.req.ready := !valid
io.data_req.valid := valid io.data_req.valid := valid
io.data_req.bits.idx := r_idx io.data_req.bits.idx := r_idx
r_cmd ^^ io.data_req.bits.cmd io.data_req.bits.cmd := r_cmd
io.data_req.bits.typ := r_type
io.data_req.bits.offset := r_offset
io.data_req.bits.data := alu.io.result io.data_req.bits.data := alu.io.result
} }
@ -470,7 +564,7 @@ class HellaCache(lines: Int) extends Component {
val io = new ioDCacheDM(); val io = new ioDCacheDM();
val addrbits = PADDR_BITS; val addrbits = PADDR_BITS;
val indexbits = ceil(log(lines)/log(2)).toInt; val indexbits = log2up(lines);
val offsetbits = OFFSET_BITS; val offsetbits = OFFSET_BITS;
val tagmsb = PADDR_BITS-1; val tagmsb = PADDR_BITS-1;
val taglsb = indexbits+offsetbits; val taglsb = indexbits+offsetbits;
@ -478,21 +572,33 @@ class HellaCache(lines: Int) extends Component {
val indexmsb = taglsb-1; val indexmsb = taglsb-1;
val indexlsb = offsetbits; val indexlsb = offsetbits;
val offsetmsb = indexlsb-1; val offsetmsb = indexlsb-1;
val offsetlsb = ceil(log(CPU_DATA_BITS/8)/log(2)).toInt; val offsetlsb = log2up(CPU_DATA_BITS/8);
val rf_cnt_bits = ceil(log(REFILL_CYCLES)/log(2)).toInt val ramindexlsb = log2up(MEM_DATA_BITS/8);
val early_nack = Reg { Bool() }
val r_cpu_req_val_ = Reg(io.cpu.req_val, resetVal = Bool(false)) val r_cpu_req_val_ = Reg(io.cpu.req_val, 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_idx = Reg() { Bits() }
val r_cpu_req_cmd = Reg() { Bits() } val r_cpu_req_cmd = Reg() { Bits() }
val r_cpu_req_type = Reg() { Bits() } val r_cpu_req_type = Reg() { Bits() }
val r_cpu_req_tag = Reg() { Bits() } val r_cpu_req_tag = Reg() { Bits() }
val r_cpu_req_data = Reg() { Bits() }
val r2_cpu_req_val = Reg(r_cpu_req_val, resetVal = Bool(false))
val r2_cpu_req_ppn = Reg(io.cpu.req_ppn)
val r2_cpu_req_idx = Reg(r_cpu_req_idx)
val r2_cpu_req_cmd = Reg(r_cpu_req_cmd)
val r2_cpu_req_type = Reg(r_cpu_req_type)
val r2_cpu_req_tag = Reg(r_cpu_req_tag)
val p_store_valid = Reg(resetVal = Bool(false)) val p_store_valid = Reg(resetVal = Bool(false))
val p_store_data = Reg() { Bits() } val p_store_data = Reg() { Bits() }
val p_store_idx = Reg() { Bits() } val p_store_idx = Reg() { Bits() }
val p_store_type = Reg() { Bits() } val p_store_type = Reg() { Bits() }
val r_cpu_req_val = r_cpu_req_val_ && !io.cpu.req_nack val store_data_wide = Wire { Bits(width = MEM_DATA_BITS) }
val store_wmask_wide = Wire { Bits(width = MEM_DATA_BITS) }
val req_store = (io.cpu.req_cmd === M_XWR) val req_store = (io.cpu.req_cmd === M_XWR)
val req_load = (io.cpu.req_cmd === M_XRD) || (io.cpu.req_cmd === M_PRD) val req_load = (io.cpu.req_cmd === M_XRD) || (io.cpu.req_cmd === M_PRD)
val req_flush = (io.cpu.req_cmd === M_FLA) val req_flush = (io.cpu.req_cmd === M_FLA)
@ -505,12 +611,23 @@ class HellaCache(lines: Int) extends Component {
val r_req_amo = r_cpu_req_cmd(3).toBool val r_req_amo = r_cpu_req_cmd(3).toBool
val r_req_read = r_req_load || r_req_amo val r_req_read = r_req_load || r_req_amo
val r_req_write = r_req_store || r_req_amo val r_req_write = r_req_store || r_req_amo
val r2_req_load = (r2_cpu_req_cmd === M_XRD)
val r2_req_store = (r2_cpu_req_cmd === M_XWR)
val r2_req_amo = r2_cpu_req_cmd(3).toBool
val r2_req_write = r2_req_store || r2_req_amo
val nack_wb = Wire { Bool() }
val nack_mshr = Wire { Bool() }
val nack_sdq = Wire { Bool() }
when (io.cpu.req_val) { when (io.cpu.req_val) {
r_cpu_req_idx <== io.cpu.req_idx r_cpu_req_idx <== io.cpu.req_idx
r_cpu_req_cmd <== Mux(req_load, M_XRD, io.cpu.req_cmd) r_cpu_req_cmd <== Mux(req_load, M_XRD, io.cpu.req_cmd)
r_cpu_req_type <== io.cpu.req_type r_cpu_req_type <== io.cpu.req_type
r_cpu_req_tag <== io.cpu.req_tag r_cpu_req_tag <== io.cpu.req_tag
when (req_write) {
r_cpu_req_data <== io.cpu.req_data
}
} }
// tags // tags
@ -520,13 +637,17 @@ class HellaCache(lines: Int) extends Component {
// data // data
val data = new DataArray(lines) val data = new DataArray(lines)
val data_arb = (new Arbiter(3)) { new DataArrayReq() } val data_arb = (new Arbiter(5)) { new DataArrayReq() }
data_arb.io.out <> data.io.req data_arb.io.out <> data.io.req
// writeback unit // writeback unit
val wb = new WritebackUnit val wb = new WritebackUnit
val mem_arb = (new Arbiter(2)) { new MemReq() }
val wb_arb = (new Arbiter(2)) { new WritebackReq() } val wb_arb = (new Arbiter(2)) { new WritebackReq() }
wb_arb.io.out <> wb.io.req wb_arb.io.out <> wb.io.req
wb.io.data_req <> data_arb.io.in(1)
wb.io.data_resp <> data.io.resp
wb.io.mem_req <> mem_arb.io.in(0)
// reset and flush unit // reset and flush unit
val flusher = new FlushUnit(lines) val flusher = new FlushUnit(lines)
@ -536,119 +657,168 @@ class HellaCache(lines: Int) extends Component {
flusher.io.meta_resp <> meta.io.resp flusher.io.meta_resp <> meta.io.resp
// cpu tag check // cpu tag check
val meta_req = new MetaArrayReq().asInput
meta_req.idx <== io.cpu.req_idx
meta_req.rw <== Bool(false)
meta_arb.io.in(2).valid := io.cpu.req_val meta_arb.io.in(2).valid := io.cpu.req_val
meta_req ^^ meta_arb.io.in(2).bits 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 early_tag_nack = !meta_arb.io.in(2).ready
val tag_match = meta.io.resp.valid && (meta.io.resp.tag === io.cpu.req_ppn) val tag_match = meta.io.resp.valid && (meta.io.resp.tag === io.cpu.req_ppn)
val hit = r_cpu_req_val && tag_match val tag_hit = r_cpu_req_val && tag_match
val miss = r_cpu_req_val && !tag_match val tag_miss = r_cpu_req_val && !tag_match
// refill counter // refill counter
val rr_count = Reg(resetVal = UFix(0, rf_cnt_bits)); val rr_count = Reg(resetVal = UFix(0, log2up(REFILL_CYCLES)))
val rr_count_next = rr_count + UFix(1); val rr_count_next = rr_count + UFix(1)
when (io.mem.resp_val) { rr_count <== rr_count_next } when (io.mem.resp_val) { rr_count <== rr_count_next }
// refill response // refill response
val rr = new DataArrayReq().asInput val block_during_refill = !io.mem.resp_val && (rr_count != UFix(0))
rr.offset <== rr_count data_arb.io.in(0).valid := io.mem.resp_val || block_during_refill
rr.idx <== Bits(0) /* TODO: get this from MSHR file */ data_arb.io.in(0).bits.offset := rr_count
rr.rw <== Bool(true) data_arb.io.in(0).bits.rw := !block_during_refill
rr.wmask <== ~UFix(0) data_arb.io.in(0).bits.wmask := ~UFix(0, MEM_DATA_BITS/8)
rr.data <== io.mem.resp_data data_arb.io.in(0).bits.data := io.mem.resp_data
data_arb.io.in(0).valid := io.mem.resp_val
rr ^^ data_arb.io.in(0).bits // writeback
val wb_rdy = wb_arb.io.in(1).ready
wb_arb.io.in(1).valid := r_cpu_req_val && !tag_match && meta.io.resp.dirty
wb_arb.io.in(1).bits.ppn := meta.io.resp.tag
wb_arb.io.in(1).bits.idx := r_cpu_req_idx(indexmsb,indexlsb)
// load hits // load hits
val load = new DataArrayReq().asInput data_arb.io.in(4).bits.offset := io.cpu.req_idx(offsetmsb,ramindexlsb)
load.offset <== io.cpu.req_idx(offsetmsb, offsetlsb+rf_cnt_bits) data_arb.io.in(4).bits.idx := io.cpu.req_idx(indexmsb,indexlsb)
load.idx <== io.cpu.req_idx(indexmsb, indexlsb) data_arb.io.in(4).bits.rw := Bool(false)
load.rw <== Bool(false) data_arb.io.in(4).bits.wmask := UFix(0) // don't care
load.wmask <== ~UFix(0) // don't care data_arb.io.in(4).bits.data := io.mem.resp_data // don't care
load.data <== io.mem.resp_data // don't care data_arb.io.in(4).valid := io.cpu.req_val && req_read
data_arb.io.in(2).valid := io.cpu.req_val && req_read val early_load_nack = req_read && !data_arb.io.in(4).ready
load ^^ data_arb.io.in(2).bits
val early_load_nack = req_read && !data_arb.io.in(2).ready
// load/store addresses conflict if they are to any part of the same word
p_store_valid <== Bool(false)
val p_store_match = r_req_read && p_store_valid && (r_cpu_req_idx(indexlsb-1,offsetlsb) === p_store_idx(indexlsb-1,offsetlsb))
// store hits. // store hits.
// we nack new stores if a pending store can't retire for some reason. // 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 // we drain a pending store if the CPU performs a store or a
// conflictig load, or if the cache misses or is idle. // conflictig load, or if the cache is idle, or after a miss.
val store = new DataArrayReq().asInput val p_store_match = r_cpu_req_val && r_req_read && p_store_valid && (r_cpu_req_idx(indexlsb-1,offsetlsb) === p_store_idx(indexlsb-1,offsetlsb))
store.offset <== p_store_idx(offsetmsb, offsetlsb+rf_cnt_bits) val drain_store_val = (p_store_valid && (!io.cpu.req_val || req_store || Reg(tag_miss))) || p_store_match
store.idx <== p_store_idx(indexmsb, indexlsb) data_arb.io.in(2).bits.offset := p_store_idx(offsetmsb,ramindexlsb)
store.rw <== Bool(true) data_arb.io.in(2).bits.idx := p_store_idx(indexmsb,indexlsb)
store.wmask <== UFix(0) // TODO data_arb.io.in(2).bits.rw := Bool(true)
store.data <== UFix(0) // TODO data_arb.io.in(2).bits.wmask := store_wmask_wide
data_arb.io.in(1).valid := p_store_valid && (miss || !io.cpu.req_val || req_store || (r_cpu_req_val && p_store_match)) data_arb.io.in(2).bits.data := store_data_wide
val early_store_nack = req_write && p_store_valid && !data_arb.io.in(1).ready data_arb.io.in(2).valid := drain_store_val
val drain_store_rdy = data_arb.io.in(2).ready
val drain_store = drain_store_val && drain_store_rdy
val p_store_notready = p_store_valid && !drain_store
p_store_valid <== p_store_notready || (tag_hit && r_req_store)
// tag update after a miss or a store to an exclusive clean line. // tag update after a miss or a store to an exclusive clean line.
// we don't look at the meta ready signal because the only requester val clear_valid = tag_miss && !r_req_flush && meta.io.resp.valid
// with higher precedence is the flush unit, which nacks us anyway. val set_dirty = tag_hit && !meta.io.resp.dirty && r_req_write
val meta_update = new MetaArrayReq().asInput meta.io.state_req.valid := clear_valid || set_dirty
meta_update.idx <== r_cpu_req_idx meta.io.state_req.bits.idx := r_cpu_req_idx(indexmsb,indexlsb)
meta_update.rw <== Bool(true) meta.io.state_req.bits.data.valid := tag_match
meta_update.data.valid <== tag_match meta.io.state_req.bits.data.dirty := tag_match
meta_update.data.dirty <== tag_match
meta_update.data.tag <== io.cpu.req_ppn
meta_req.data <== meta_update.data // don't care
meta_arb.io.in(1).valid := miss && wb_arb.io.in(1).ready || hit && r_req_write
meta_update ^^ meta_arb.io.in(1).bits
// pending store data, also used for AMO RHS // pending store data, also used for AMO RHS
when (io.cpu.req_val && req_store && !early_store_nack) { val storegen = new StoreDataGen
p_store_idx <== io.cpu.req_idx; storegen.io.typ := r_cpu_req_type
p_store_type <== io.cpu.req_type; storegen.io.addr := r_cpu_req_idx(offsetlsb-1, 0)
} storegen.io.din := r_cpu_req_data
when (io.cpu.req_val && req_write && !early_store_nack) { when (tag_hit && r_req_store && !p_store_notready) {
p_store_data <== io.cpu.req_data p_store_idx <== r_cpu_req_idx
p_store_type <== r_cpu_req_type
p_store_data <== storegen.io.dout
} }
/*val mshr = new MSHRFile() // miss handling
mshr.io.req_val := r_cpu_req_val val mshr = new MSHRFile
val replayer = new ReplayUnit
mshr.io.req_val := tag_miss && !r_req_flush && !nack_mshr
mshr.io.req_ppn := io.cpu.req_ppn mshr.io.req_ppn := io.cpu.req_ppn
mshr.io.req_idx := r_cpu_req_idx(PGIDX_BITS-1, offsetbits) mshr.io.req_idx := r_cpu_req_idx(indexmsb,indexlsb)
mshr.io.req_cmd.offset := r_cpu_req_idx(offsetbits-1, 0) mshr.io.req_data := p_store_data
mshr.io.req_cmd.cmd := r_cpu_req_cmd mshr.io.req_tag := r_cpu_req_tag
mshr.io.req_cmd.typ := r_cpu_req_type*/ 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
mshr.io.mem_resp_tag := io.mem.resp_tag
mshr.io.mem_req <> mem_arb.io.in(1)
mshr.io.meta_req <> meta_arb.io.in(1)
mshr.io.replay <> replayer.io.replay
replayer.io.sdq_enq.valid := tag_miss && r_req_write && !nack_sdq
replayer.io.sdq_enq.bits := storegen.io.dout
data_arb.io.in(0).bits.idx := mshr.io.mem_resp_idx
// signal a load miss when the data isn't present in the cache and when it's in the pending store data register // replays
// (causes the cache to block for 2 cycles and the load or amo instruction is replayed) val replay = replayer.io.data_req.bits
val early_nack = early_tag_nack || early_load_nack || early_store_nack data_arb.io.in(3).bits.offset := replay.offset(offsetmsb,ramindexlsb)
val nack = Reg(early_nack) || p_store_match || !flusher.io.req.ready data_arb.io.in(3).bits.idx := replay.idx
val load_miss = !nack && miss && r_req_read data_arb.io.in(3).bits.rw := replay.cmd === M_XWR
val resp_val = (!nack && hit && r_req_read) || flusher.io.resp.valid data_arb.io.in(3).bits.wmask := store_wmask_wide
data_arb.io.in(3).bits.data := store_data_wide
data_arb.io.in(3).valid := replayer.io.data_req.valid
replayer.io.data_req.ready := data_arb.io.in(3).ready
// store write mask generation.
// assumes pending stores are higher-priority than store replays.
val maskgen = new StoreMaskGen
val store_offset = Mux(drain_store_val, p_store_idx(offsetmsb,0), replay.offset)
maskgen.io.typ := Mux(drain_store_val, p_store_type, replay.typ)
maskgen.io.addr := store_offset(offsetlsb-1,0)
store_wmask_wide <== maskgen.io.wmask << Cat(store_offset(ramindexlsb-1,offsetlsb), Bits(0, log2up(CPU_DATA_BITS/8))).toUFix
val store_data = Mux(drain_store_val, p_store_data, replay.data)
store_data_wide <== Fill(MEM_DATA_BITS/CPU_DATA_BITS, store_data)
// load data subword mux/sign extension.
// assumes load replays are higher-priority than load hits.
// subword loads are delayed by one cycle.
val loadgen = new LoadDataGen
val loadgen_use_replay = Reg(replayer.io.data_req.valid)
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)(offsetlsb-1,0)
loadgen.io.din := Slice(MEM_DATA_BITS/CPU_DATA_BITS, data.io.resp, r_cpu_req_idx(ramindexlsb-1,offsetlsb).toUFix)
early_nack <== early_tag_nack || early_load_nack
val nack_miss_wb = meta.io.resp.dirty && !wb_rdy
val nack_miss_mshr = !mshr.io.req_rdy
val nack_miss_sdq = r_req_write && !replayer.io.sdq_enq.ready
nack_wb <== nack_miss_mshr || nack_miss_sdq || p_store_notready || p_store_match
nack_mshr <== nack_miss_wb || nack_miss_sdq || p_store_notready || p_store_match
nack_sdq <== nack_miss_wb || nack_miss_mshr || p_store_notready || p_store_match
val nack_for_flush = r_req_flush && !flusher.io.req.ready
val nack = p_store_match || r_req_store && p_store_notready || early_nack ||
tag_miss && !r_req_flush && (nack_miss_wb || nack_miss_mshr || nack_miss_sdq || p_store_notready)
// report that cache is always ready. we nack instead. // report that cache is always ready. we nack instead.
io.cpu.req_rdy := Bool(true) io.cpu.req_rdy := Bool(true)
io.cpu.resp_nack := r_cpu_req_val_ && nack io.cpu.resp_nack := r_cpu_req_val_ && !io.cpu.req_kill && nack
io.cpu.resp_val := resp_val io.cpu.resp_val := (tag_hit && !nack && r_req_read) || flusher.io.resp.valid || replayer.io.cpu_resp_val
io.cpu.resp_miss := tag_miss && !nack && r_req_read
io.cpu.resp_tag := Mux(replayer.io.cpu_resp_val, replayer.io.cpu_resp_tag, Mux(flusher.io.resp.valid, flusher.io.resp.bits, r_cpu_req_tag))
io.cpu.resp_data := loadgen.io.dout
io.cpu.resp_data_subword := loadgen.io.dout_subword
val misaligned = 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_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(0,2))) || (((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(0,3))); ((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_ld := r_cpu_req_val_ && !io.cpu.req_kill && r_req_read && misaligned
io.cpu.xcpt_ma_st := r_cpu_req_val_ && r_req_write && misaligned io.cpu.xcpt_ma_st := r_cpu_req_val_ && !io.cpu.req_kill && r_req_write && misaligned
io.cpu.resp_miss := load_miss mem_arb.io.out.ready := io.mem.req_rdy
io.cpu.resp_tag := flusher.io.resp.bits io.mem.req_val := mem_arb.io.out.valid
io.cpu.resp_data := Bits(0) io.mem.req_rw := mem_arb.io.out.bits.rw
io.mem.req_wdata := wb.io.mem_req_data
io.mem.req_val := Bool(false) io.mem.req_tag := mem_arb.io.out.bits.tag.toUFix
io.mem.req_rw := Bool(false) io.mem.req_addr := mem_arb.io.out.bits.addr
io.mem.req_wdata := Bits(0)
io.mem.req_tag := UFix(0)
io.mem.req_addr := UFix(0)
} }
} }

2
rocket/src/main/scala/ptw.scala
View File

@ -27,7 +27,7 @@ class rocketDmemArbiter extends Component
io.mem.req_ppn := Mux(r_ptw_req_val, r_ptw_req_ppn, io.cpu.req_ppn); io.mem.req_ppn := Mux(r_ptw_req_val, r_ptw_req_ppn, io.cpu.req_ppn);
io.mem.req_data := io.cpu.req_data; io.mem.req_data := io.cpu.req_data;
io.mem.req_tag := Cat(io.cpu.req_tag, io.ptw.req_val); io.mem.req_tag := Cat(io.cpu.req_tag, io.ptw.req_val);
io.mem.req_nack := io.cpu.req_nack; io.mem.req_kill := io.cpu.req_kill;
io.ptw.req_rdy := io.mem.req_rdy; io.ptw.req_rdy := io.mem.req_rdy;
io.cpu.req_rdy := io.mem.req_rdy && !io.ptw.req_val; io.cpu.req_rdy := io.mem.req_rdy && !io.ptw.req_val;

6
rocket/src/main/scala/queues.scala
View File

@ -19,7 +19,7 @@ class ioQueueCtrl(addr_sz: Int) extends Bundle()
class queueCtrl(entries: Int) extends Component class queueCtrl(entries: Int) extends Component
{ {
val addr_sz = ceil(log(entries)/log(2)).toInt val addr_sz = log2up(entries)
override val io = new ioQueueCtrl(addr_sz); override val io = new ioQueueCtrl(addr_sz);
// Enqueue and dequeue pointers // Enqueue and dequeue pointers
@ -117,7 +117,7 @@ class ioQueueCtrlFlow(addr_sz: Int) extends Bundle() /* IOqueueCtrl */
class queueCtrlFlow(entries: Int) extends Component class queueCtrlFlow(entries: Int) extends Component
{ {
val addr_sz = ceil(log(entries)/log(2)).toInt val addr_sz = log2up(entries)
override val io = new ioQueueCtrlFlow(addr_sz); override val io = new ioQueueCtrlFlow(addr_sz);
// Enqueue and dequeue pointers // Enqueue and dequeue pointers
@ -186,7 +186,7 @@ class ioQueueDpathFlow[T <: Data](addr_sz: Int)(data: => T) extends Bundle()
class queueDpathFlow[T <: Data](entries: Int)(data: => T) extends Component class queueDpathFlow[T <: Data](entries: Int)(data: => T) extends Component
{ {
val addr_sz = ceil(log(entries)/log(2)).toInt val addr_sz = log2up(entries)
override val io = new ioQueueDpathFlow(addr_sz)(data); override val io = new ioQueueDpathFlow(addr_sz)(data);
val ram = Mem(entries, io.wen, io.waddr, io.enq_bits); val ram = Mem(entries, io.wen, io.waddr, io.enq_bits);
val rout = ram(io.raddr); val rout = ram(io.raddr);

55
rocket/src/main/scala/util.scala
View File

@ -5,20 +5,32 @@ import Chisel._
import Node._; import Node._;
import scala.math._; import scala.math._;
class MuxN[T <: Data](n: Int)(data: => T) extends Component { object log2up
val io = new Bundle { {
val sel = Bits(width = ceil(log(n)/log(2)).toInt) def apply(in: Int) = ceil(log(in)/log(2)).toInt
val in = Vec(n) { data }.asInput() }
val out = data.asOutput()
}
val out = Vec(n) { Wire() { data } } object Slice
out(0) <== io.in(0) {
for (i <- 1 to n-1) { def apply(n: Int, in: Bits, sel: Bits) =
out(i) <== Mux(io.sel === UFix(i), io.in(i), out(i-1)) {
val w = in.width / n
var out = in(w-1, 0) & Fill(w, sel === UFix(0))
for (i <- 1 until n)
out = out | (in((i+1)*w-1, i*w) & Fill(w, sel === Bits(i)))
if (n > 1) out else in
} }
}
out(n-1) ^^ io.out object FillInterleaved
{
def apply(n: Int, in: Bits) =
{
var out = Fill(n, in(0))
for (i <- 1 until in.width)
out = Cat(Fill(n, in(i)), out)
out
}
} }
class Mux1H(n: Int, w: Int) extends Component class Mux1H(n: Int, w: Int) extends Component
@ -30,13 +42,10 @@ class Mux1H(n: Int, w: Int) extends Component
} }
if (n > 1) { if (n > 1) {
val out = Vec(n) { Wire() { Bits(width = w) } } var out = io.in(0) & Fill(w, io.sel(0))
out(0) <== io.in(0) & Fill(w, io.sel(0)) for (i <- 1 to n-1)
for (i <- 1 to n-1) { out = out | (io.in(i) & Fill(w, io.sel(i)))
out(i) <== out(i-1) | (io.in(i) & Fill(w, io.sel(i))) io.out := out
}
io.out := out(n-1)
} else { } else {
io.out := io.in(0) io.out := io.in(0)
} }
@ -56,7 +65,6 @@ class ioArbiter[T <: Data](n: Int)(data: => T) extends Bundle {
class Arbiter[T <: Data](n: Int)(data: => T) extends Component { class Arbiter[T <: Data](n: Int)(data: => T) extends Component {
val io = new ioArbiter(n)(data) val io = new ioArbiter(n)(data)
val dout = Vec(n) { Wire() { data } }
val vout = Wire { Bool() } val vout = Wire { Bool() }
io.in(0).ready := io.out.ready io.in(0).ready := io.out.ready
@ -64,10 +72,9 @@ class Arbiter[T <: Data](n: Int)(data: => T) extends Component {
io.in(i).ready := !io.in(i-1).valid && io.in(i-1).ready io.in(i).ready := !io.in(i-1).valid && io.in(i-1).ready
} }
dout(0) <== io.in(n-1).bits var dout = io.in(n-1).bits
for (i <- 1 to n-1) { for (i <- 1 to n-1)
dout(i) <== Mux(io.in(n-1-i).valid, io.in(n-1-i).bits, dout(i-1)) dout = Mux(io.in(n-1-i).valid, io.in(n-1-i).bits, dout)
}
for (i <- 0 to n-2) { for (i <- 0 to n-2) {
when (io.in(i).valid) { vout <== Bool(true) } when (io.in(i).valid) { vout <== Bool(true) }
@ -75,7 +82,7 @@ class Arbiter[T <: Data](n: Int)(data: => T) extends Component {
vout <== io.in(n-1).valid vout <== io.in(n-1).valid
vout ^^ io.out.valid vout ^^ io.out.valid
dout(n-1) ^^ io.out.bits dout ^^ io.out.bits
} }
class ioPriorityDecoder(in_width: Int, out_width: Int) extends Bundle class ioPriorityDecoder(in_width: Int, out_width: Int) extends Bundle