1
0
rocket-chip/rocket/src/main/scala/icache.scala

299 lines
11 KiB
Scala
Raw Normal View History

package rocket
2012-10-08 05:15:54 +02:00
import Chisel._
import uncore._
import Util._
2014-09-01 22:28:58 +02:00
case object ECCCode extends Field[Option[Code]]
2014-09-01 22:28:58 +02:00
abstract trait L1CacheParameters extends CacheParameters with CoreParameters {
2014-12-07 12:09:54 +01:00
val outerDataBeats = params(TLDataBeats)
2015-03-01 02:00:05 +01:00
val outerDataBits = params(TLDataBits)
val code = params(ECCCode).getOrElse(new IdentityCode)
val refillCyclesPerBeat = outerDataBits/rowBits
2014-12-17 23:28:05 +01:00
val refillCycles = refillCyclesPerBeat*outerDataBeats
2014-09-01 22:28:58 +02:00
}
2014-09-01 22:28:58 +02:00
abstract trait FrontendParameters extends L1CacheParameters
abstract class FrontendBundle extends Bundle with FrontendParameters
abstract class FrontendModule extends Module with FrontendParameters
2014-09-01 22:28:58 +02:00
class FrontendReq extends CoreBundle {
val pc = UInt(width = vaddrBits+1)
}
2014-09-01 22:28:58 +02:00
class FrontendResp extends CoreBundle {
val pc = UInt(width = vaddrBits+1) // ID stage PC
val data = Vec.fill(coreFetchWidth) (Bits(width = coreInstBits))
val mask = Bits(width = coreFetchWidth)
2012-10-10 06:35:03 +02:00
val xcpt_if = Bool()
}
class CPUFrontendIO extends Bundle {
2013-08-12 19:39:11 +02:00
val req = Valid(new FrontendReq)
val resp = Decoupled(new FrontendResp).flip
val btb_resp = Valid(new BTBResp).flip
val btb_update = Valid(new BTBUpdate)
val bht_update = Valid(new BHTUpdate)
val ras_update = Valid(new RASUpdate)
val invalidate = Bool(OUTPUT)
2012-10-10 06:35:03 +02:00
}
class Frontend(btb_updates_out_of_order: Boolean = false) extends FrontendModule
2012-10-10 06:35:03 +02:00
{
val io = new Bundle {
val cpu = new CPUFrontendIO().flip
val ptw = new TLBPTWIO()
val mem = new HeaderlessUncachedTileLinkIO
2012-10-10 06:35:03 +02:00
}
val btb = Module(new BTB(btb_updates_out_of_order))
2013-08-12 19:39:11 +02:00
val icache = Module(new ICache)
val tlb = Module(new TLB)
2012-10-10 06:35:03 +02:00
2013-09-13 02:55:58 +02:00
val s1_pc_ = Reg(UInt())
val s1_pc = s1_pc_ & SInt(-coreInstBytes) // discard PC LSBS (this propagates down the pipeline)
2013-08-12 19:39:11 +02:00
val s1_same_block = Reg(Bool())
2013-08-16 00:28:15 +02:00
val s2_valid = Reg(init=Bool(true))
val s2_pc = Reg(init=UInt(START_ADDR))
2014-04-02 00:01:27 +02:00
val s2_btb_resp_valid = Reg(init=Bool(false))
val s2_btb_resp_bits = Reg(btb.io.resp.bits.clone)
2013-08-16 00:28:15 +02:00
val s2_xcpt_if = Reg(init=Bool(false))
2012-10-10 06:35:03 +02:00
val msb = vaddrBits-1
val lsb = log2Up(coreFetchWidth*coreInstBytes)
val btbTarget = Cat(btb.io.resp.bits.target(msb), btb.io.resp.bits.target)
val ntpc_0 = s1_pc + UInt(coreInstBytes*coreFetchWidth)
val ntpc = Cat(s1_pc(msb) & ntpc_0(msb), ntpc_0(msb,lsb), Bits(0,lsb)) // unsure
2012-10-10 06:35:03 +02:00
val icmiss = s2_valid && !icache.io.resp.valid
val predicted_npc = Mux(btb.io.resp.bits.taken, btbTarget, ntpc)
2013-08-12 19:39:11 +02:00
val npc = Mux(icmiss, s2_pc, predicted_npc).toUInt
val s0_same_block = !icmiss && !io.cpu.req.valid && !btb.io.resp.bits.taken && ((ntpc & rowBytes) === (s1_pc & rowBytes))
2012-10-10 06:35:03 +02:00
val stall = io.cpu.resp.valid && !io.cpu.resp.ready
2012-10-10 06:35:03 +02:00
when (!stall) {
s1_same_block := s0_same_block && !tlb.io.resp.miss
2013-09-13 02:55:58 +02:00
s1_pc_ := npc
2012-10-10 06:35:03 +02:00
s2_valid := !icmiss
when (!icmiss) {
s2_pc := s1_pc
2014-04-02 00:01:27 +02:00
s2_btb_resp_valid := btb.io.resp.valid
when (btb.io.resp.valid) { s2_btb_resp_bits := btb.io.resp.bits }
s2_xcpt_if := tlb.io.resp.xcpt_if
}
2011-11-05 04:52:21 +01:00
}
2012-10-10 06:35:03 +02:00
when (io.cpu.req.valid) {
s1_same_block := Bool(false)
2013-09-13 02:55:58 +02:00
s1_pc_ := io.cpu.req.bits.pc
2012-10-10 06:35:03 +02:00
s2_valid := Bool(false)
2011-11-05 04:52:21 +01:00
}
2012-10-10 06:35:03 +02:00
btb.io.req.valid := !stall && !icmiss
btb.io.req.bits.addr := s1_pc
btb.io.btb_update := io.cpu.btb_update
btb.io.bht_update := io.cpu.bht_update
btb.io.ras_update := io.cpu.ras_update
btb.io.invalidate := io.cpu.invalidate || io.ptw.invalidate
2012-10-10 06:35:03 +02:00
tlb.io.ptw <> io.ptw
2012-10-10 06:35:03 +02:00
tlb.io.req.valid := !stall && !icmiss
tlb.io.req.bits.vpn := s1_pc >> UInt(pgIdxBits)
2013-08-12 19:39:11 +02:00
tlb.io.req.bits.asid := UInt(0)
tlb.io.req.bits.passthrough := Bool(false)
2012-10-10 06:35:03 +02:00
tlb.io.req.bits.instruction := Bool(true)
2015-03-14 10:49:07 +01:00
tlb.io.req.bits.store := Bool(false)
2012-10-10 06:35:03 +02:00
icache.io.mem <> io.mem
icache.io.req.valid := !stall && !s0_same_block
2012-10-10 06:35:03 +02:00
icache.io.req.bits.idx := Mux(io.cpu.req.valid, io.cpu.req.bits.pc, npc)
icache.io.invalidate := io.cpu.invalidate
2012-10-10 06:35:03 +02:00
icache.io.req.bits.ppn := tlb.io.resp.ppn
icache.io.req.bits.kill := io.cpu.req.valid || tlb.io.resp.miss || icmiss || io.ptw.invalidate
icache.io.resp.ready := !stall && !s1_same_block
2012-10-10 06:35:03 +02:00
io.cpu.resp.valid := s2_valid && (s2_xcpt_if || icache.io.resp.valid)
io.cpu.resp.bits.pc := s2_pc
var fetch_data:Bits = null
require (coreFetchWidth <= 4)
if (coreFetchWidth == 4) {
fetch_data = icache.io.resp.bits.datablock
} else {
fetch_data = icache.io.resp.bits.datablock >> (s2_pc(log2Up(rowBytes)-1,log2Up(coreFetchWidth*coreInstBytes)) << log2Up(coreFetchWidth*coreInstBits))
}
for (i <- 0 until coreFetchWidth) {
io.cpu.resp.bits.data(i) := fetch_data(i*coreInstBits+coreInstBits-1, i*coreInstBits)
}
2014-11-03 10:13:22 +01:00
val all_ones = UInt((1 << (coreFetchWidth+1))-1)
val msk_pc = if (coreFetchWidth == 1) all_ones else all_ones << s2_pc(log2Up(coreFetchWidth) -1+2,2)
io.cpu.resp.bits.mask := Mux(s2_btb_resp_valid, msk_pc & s2_btb_resp_bits.mask, msk_pc)
2012-10-10 06:35:03 +02:00
io.cpu.resp.bits.xcpt_if := s2_xcpt_if
2014-04-02 00:01:27 +02:00
io.cpu.btb_resp.valid := s2_btb_resp_valid
io.cpu.btb_resp.bits := s2_btb_resp_bits
2012-10-10 06:35:03 +02:00
}
class ICacheReq extends FrontendBundle {
val idx = UInt(width = pgIdxBits)
val ppn = UInt(width = ppnBits) // delayed one cycle
2013-08-12 19:39:11 +02:00
val kill = Bool() // delayed one cycle
}
class ICacheResp extends FrontendBundle {
2014-09-01 22:28:58 +02:00
val data = Bits(width = coreInstBits)
val datablock = Bits(width = rowBits)
2013-08-12 19:39:11 +02:00
}
class ICache extends FrontendModule
2012-10-10 06:35:03 +02:00
{
val io = new Bundle {
2013-08-12 19:39:11 +02:00
val req = Valid(new ICacheReq).flip
val resp = Decoupled(new ICacheResp)
val invalidate = Bool(INPUT)
val mem = new HeaderlessUncachedTileLinkIO
2012-01-25 00:13:49 +01:00
}
require(isPow2(nSets) && isPow2(nWays))
2014-09-01 22:28:58 +02:00
require(isPow2(coreInstBytes))
require(pgIdxBits >= untagBits)
2012-01-25 00:13:49 +01:00
2013-09-10 19:51:35 +02:00
val s_ready :: s_request :: s_refill_wait :: s_refill :: Nil = Enum(UInt(), 4)
2013-08-16 00:28:15 +02:00
val state = Reg(init=s_ready)
2013-08-12 19:39:11 +02:00
val invalidated = Reg(Bool())
2012-10-10 06:35:03 +02:00
val stall = !io.resp.ready
val rdy = Bool()
2013-08-16 00:28:15 +02:00
val s2_valid = Reg(init=Bool(false))
val s2_addr = Reg(UInt(width = paddrBits))
val s2_any_tag_hit = Bool()
2013-08-16 00:28:15 +02:00
val s1_valid = Reg(init=Bool(false))
val s1_pgoff = Reg(UInt(width = pgIdxBits))
2013-08-12 19:39:11 +02:00
val s1_addr = Cat(io.req.bits.ppn, s1_pgoff).toUInt
val s1_tag = s1_addr(tagBits+untagBits-1,untagBits)
2012-10-10 06:35:03 +02:00
val s0_valid = io.req.valid || s1_valid && stall
val s0_pgoff = Mux(s1_valid && stall, s1_pgoff, io.req.bits.idx)
2012-10-10 06:35:03 +02:00
s1_valid := io.req.valid && rdy || s1_valid && stall && !io.req.bits.kill
2012-10-10 06:35:03 +02:00
when (io.req.valid && rdy) {
s1_pgoff := io.req.bits.idx
2011-11-05 04:52:21 +01:00
}
2012-10-10 06:35:03 +02:00
s2_valid := s1_valid && rdy && !io.req.bits.kill || io.resp.valid && stall
2012-10-10 06:35:03 +02:00
when (s1_valid && rdy && !stall) {
s2_addr := s1_addr
2012-10-10 06:35:03 +02:00
}
val s2_tag = s2_addr(tagBits+untagBits-1,untagBits)
val s2_idx = s2_addr(untagBits-1,blockOffBits)
val s2_offset = s2_addr(blockOffBits-1,0)
2012-10-10 06:35:03 +02:00
val s2_hit = s2_valid && s2_any_tag_hit
val s2_miss = s2_valid && !s2_any_tag_hit
rdy := state === s_ready && !s2_miss
val ser = Module(new FlowThroughSerializer(
io.mem.grant.bits,
refillCyclesPerBeat))
2014-12-07 12:09:54 +01:00
ser.io.in <> io.mem.grant
val (refill_cnt, refill_wrap) = Counter(ser.io.out.fire(), refillCycles) //TODO Zero width wire
val refill_done = state === s_refill && refill_wrap
val refill_valid = ser.io.out.valid
val refill_bits = ser.io.out.bits
ser.io.out.ready := Bool(true)
2012-10-10 06:35:03 +02:00
val repl_way = if (isDM) UInt(0) else LFSR16(s2_miss)(log2Up(nWays)-1,0)
val entagbits = code.width(tagBits)
val tag_array = Mem(Bits(width = entagbits*nWays), nSets, seqRead = true)
2013-08-12 19:39:11 +02:00
val tag_raddr = Reg(UInt())
2012-10-10 06:35:03 +02:00
when (refill_done) {
val wmask = FillInterleaved(entagbits, if (isDM) Bits(1) else UIntToOH(repl_way))
val tag = code.encode(s2_tag).toUInt
tag_array.write(s2_idx, Fill(nWays, tag), wmask)
}
// /*.else*/when (s0_valid) { // uncomment ".else" to infer 6T SRAM
.elsewhen (s0_valid) {
tag_raddr := s0_pgoff(untagBits-1,blockOffBits)
}
val vb_array = Reg(init=Bits(0, nSets*nWays))
2012-10-10 06:35:03 +02:00
when (refill_done && !invalidated) {
vb_array := vb_array.bitSet(Cat(repl_way, s2_idx), Bool(true))
}
when (io.invalidate) {
vb_array := Bits(0)
2012-10-10 06:35:03 +02:00
invalidated := Bool(true)
}
val s2_disparity = Vec.fill(nWays){Bool()}
for (i <- 0 until nWays)
2013-08-12 19:39:11 +02:00
when (s2_valid && s2_disparity(i)) { vb_array := vb_array.bitSet(Cat(UInt(i), s2_idx), Bool(false)) }
2012-10-10 06:35:03 +02:00
val s1_tag_match = Vec.fill(nWays){Bool()}
val s2_tag_hit = Vec.fill(nWays){Bool()}
val s2_dout = Vec.fill(nWays){Reg(Bits())}
2012-11-25 07:00:43 +01:00
for (i <- 0 until nWays) {
val s1_vb = !io.invalidate && vb_array(Cat(UInt(i), s1_pgoff(untagBits-1,blockOffBits))).toBool
2013-08-12 19:39:11 +02:00
val s2_vb = Reg(Bool())
val s2_tag_disparity = Reg(Bool())
val s2_tag_match = Reg(Bool())
val tag_out = tag_array(tag_raddr)(entagbits*(i+1)-1, entagbits*i)
2012-10-10 06:35:03 +02:00
when (s1_valid && rdy && !stall) {
s2_vb := s1_vb
s2_tag_disparity := code.decode(tag_out).error
s2_tag_match := s1_tag_match(i)
2012-10-10 06:35:03 +02:00
}
s1_tag_match(i) := tag_out(tagBits-1,0) === s1_tag
s2_tag_hit(i) := s2_vb && s2_tag_match
s2_disparity(i) := s2_vb && (s2_tag_disparity || code.decode(s2_dout(i)).error)
2012-10-10 06:35:03 +02:00
}
s2_any_tag_hit := s2_tag_hit.reduceLeft(_||_) && !s2_disparity.reduceLeft(_||_)
for (i <- 0 until nWays) {
val data_array = Mem(Bits(width = code.width(rowBits)), nSets*refillCycles, seqRead = true)
2013-08-12 19:39:11 +02:00
val s1_raddr = Reg(UInt())
when (refill_valid && repl_way === UInt(i)) {
val e_d = code.encode(refill_bits.payload.data)
if(refillCycles > 1) data_array(Cat(s2_idx, refill_bits.payload.addr_beat)) := e_d
2014-12-07 12:09:54 +01:00
else data_array(s2_idx) := e_d
}
// /*.else*/when (s0_valid) { // uncomment ".else" to infer 6T SRAM
.elsewhen (s0_valid) {
s1_raddr := s0_pgoff(untagBits-1,blockOffBits-(if(refillCycles > 1) refill_cnt.getWidth else 0))
2012-10-10 06:35:03 +02:00
}
// if s1_tag_match is critical, replace with partial tag check
when (s1_valid && rdy && !stall && (Bool(isDM) || s1_tag_match(i))) { s2_dout(i) := data_array(s1_raddr) }
2012-01-25 01:51:30 +01:00
}
2014-09-01 22:28:58 +02:00
val s2_dout_word = s2_dout.map(x => (x >> (s2_offset(log2Up(rowBytes)-1,log2Up(coreInstBytes)) << log2Up(coreInstBits)))(coreInstBits-1,0))
2012-10-10 06:35:03 +02:00
io.resp.bits.data := Mux1H(s2_tag_hit, s2_dout_word)
io.resp.bits.datablock := Mux1H(s2_tag_hit, s2_dout)
val ack_q = Module(new Queue(new LogicalNetworkIO(new Finish), 1))
ack_q.io.enq.valid := refill_done && refill_bits.payload.requiresAck()
ack_q.io.enq.bits.payload := refill_bits.payload.makeFinish()
ack_q.io.enq.bits.header.dst := refill_bits.header.src
2012-03-07 00:47:19 +01:00
// output signals
2012-10-10 06:35:03 +02:00
io.resp.valid := s2_hit
io.mem.acquire.valid := (state === s_request) && ack_q.io.enq.ready
io.mem.acquire.bits := GetBlock(addr_block = s2_addr >> UInt(blockOffBits))
io.mem.finish <> ack_q.io.deq
// control state machine
switch (state) {
is (s_ready) {
2012-10-10 06:35:03 +02:00
when (s2_miss) { state := s_request }
2012-03-06 09:31:44 +01:00
invalidated := Bool(false)
}
2012-10-10 06:35:03 +02:00
is (s_request) {
when (io.mem.acquire.ready && ack_q.io.enq.ready) { state := s_refill_wait }
}
is (s_refill_wait) {
2013-01-22 02:18:23 +01:00
when (io.mem.grant.valid) { state := s_refill }
}
is (s_refill) {
2012-10-10 06:35:03 +02:00
when (refill_done) { state := s_ready }
}
2012-10-10 06:35:03 +02:00
}
}