dcache/mainpipe/MainPipe.scala

/***************************************************************************************
* Copyright (c) 2020-2021 Institute of Computing Technology, Chinese Academy of Sciences
* Copyright (c) 2020-2021 Peng Cheng Laboratory
*
* XiangShan is licensed under Mulan PSL v2.
* You can use this software according to the terms and conditions of the Mulan PSL v2.
* You may obtain a copy of Mulan PSL v2 at:
*          http://license.coscl.org.cn/MulanPSL2
*
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
* EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
* MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
*
* See the Mulan PSL v2 for more details.
***************************************************************************************/

package xiangshan.cache

import chipsalliance.rocketchip.config.Parameters
import chisel3._
import chisel3.util._
import freechips.rocketchip.tilelink.ClientStates._
import freechips.rocketchip.tilelink.MemoryOpCategories._
import freechips.rocketchip.tilelink.TLPermissions._
import freechips.rocketchip.tilelink.{ClientMetadata, ClientStates, TLPermissions}
import utils._

class MainPipeReq(implicit p: Parameters) extends DCacheBundle {
  val miss = Bool() // only amo miss will refill in main pipe
  val miss_id = UInt(log2Up(cfg.nMissEntries).W)
  val miss_param = UInt(TLPermissions.bdWidth.W)
  val miss_dirty = Bool()

  val probe = Bool()
  val probe_param = UInt(TLPermissions.bdWidth.W)
  val probe_need_data = Bool()

  // request info
  // reqs from Store, AMO use this
  // probe does not use this
  val source = UInt(sourceTypeWidth.W)
  val cmd = UInt(M_SZ.W)
  // if dcache size > 32KB, vaddr is also needed for store
  // vaddr is used to get extra index bits
  val vaddr  = UInt(VAddrBits.W)
  // must be aligned to block
  val addr   = UInt(PAddrBits.W)

  // store
  val store_data = UInt((cfg.blockBytes * 8).W)
  val store_mask = UInt(cfg.blockBytes.W)

  // which word does amo work on?
  val word_idx = UInt(log2Up(cfg.blockBytes * 8 / DataBits).W)
  val amo_data   = UInt(DataBits.W)
  val amo_mask   = UInt((DataBits / 8).W)

  // replace
  val replace = Bool()
  val replace_way_en = UInt(DCacheWays.W)

  val id = UInt(reqIdWidth.W)

  def isLoad: Bool = source === LOAD_SOURCE.U
  def isStore: Bool = source === STORE_SOURCE.U
  def isAMO: Bool = source === AMO_SOURCE.U

  def convertStoreReq(store: DCacheLineReq): MainPipeReq = {
    val req = Wire(new MainPipeReq)
    req := DontCare
    req.miss := false.B
    req.miss_dirty := false.B
    req.probe := false.B
    req.probe_need_data := false.B
    req.source := STORE_SOURCE.U
    req.cmd := store.cmd
    req.addr := store.addr
    req.vaddr := store.vaddr
    req.store_data := store.data
    req.store_mask := store.mask
    req.replace := false.B
    req.id := store.id
    req
  }
}

class MainPipe(implicit p: Parameters) extends DCacheModule {
  val metaBits = (new Meta).getWidth
  val encMetaBits = cacheParams.tagCode.width((new MetaAndTag).getWidth) - tagBits

  val io = IO(new Bundle() {
    // probe queue
    val probe_req = Flipped(DecoupledIO(new MainPipeReq))
    // store miss go to miss queue
    val miss_req = DecoupledIO(new MissReq)
    // store buffer
    val store_req = Flipped(DecoupledIO(new DCacheLineReq))
    val store_replay_resp = ValidIO(new DCacheLineResp)
    val store_hit_resp = ValidIO(new DCacheLineResp)
    val release_update = ValidIO(new ReleaseUpdate)
    // atmoics
    val atomic_req = Flipped(DecoupledIO(new MainPipeReq))
    val atomic_resp = ValidIO(new AtomicsResp)
    // replace
    val replace_req = Flipped(DecoupledIO(new MainPipeReq))
    val replace_resp = ValidIO(UInt(log2Up(cfg.nMissEntries).W))
    // write-back queue
    val wb = DecoupledIO(new WritebackReq)

    val data_read = DecoupledIO(new L1BankedDataReadLineReq)
    val data_resp = Input(Vec(DCacheBanks, new L1BankedDataReadResult()))
    val data_write = DecoupledIO(new L1BankedDataWriteReq)

    val meta_read = DecoupledIO(new MetaReadReq)
    val meta_resp = Input(Vec(nWays, UInt(encMetaBits.W)))
    val meta_write = DecoupledIO(new MetaWriteReq)

    val tag_read = DecoupledIO(new TagReadReq)
    val tag_resp = Input(Vec(nWays, UInt(tagBits.W)))
    val tag_write = DecoupledIO(new TagWriteReq)

    // update state vec in replacement algo
    val replace_access = ValidIO(new ReplacementAccessBundle)
    // find the way to be replaced
    val replace_way = new ReplacementWayReqIO

    val status = new Bundle() {
      val s0_set = ValidIO(UInt(idxBits.W))
      val s1, s2, s3 = ValidIO(new Bundle() {
        val set = UInt(idxBits.W)
        val way_en = UInt(nWays.W)
      })
    }

    // lrsc locked block should block probe
    val lrsc_locked_block = Output(Valid(UInt(PAddrBits.W)))
    val invalid_resv_set = Input(Bool())
    val update_resv_set = Output(Bool())
  })

  // meta array is made of regs, so meta write or read should always be ready
  assert(RegNext(io.meta_read.ready))
  assert(RegNext(io.meta_write.ready))

  val s1_s0_set_conflict, s2_s0_set_conlict, s3_s0_set_conflict = Wire(Bool())
  val set_conflict = s1_s0_set_conflict || s2_s0_set_conlict || s3_s0_set_conflict
  val s1_ready, s2_ready, s3_ready = Wire(Bool())

  // convert store req to main pipe req, and select a req from store and probe
  val store_req = Wire(DecoupledIO(new MainPipeReq))
  store_req.bits := (new MainPipeReq).convertStoreReq(io.store_req.bits)
  store_req.valid := io.store_req.valid
  io.store_req.ready := store_req.ready

  // s0: read meta and tag
  val req = Wire(DecoupledIO(new MainPipeReq))
  arbiter(
    in = Seq(
      store_req,
      io.probe_req,
      io.atomic_req,
      io.replace_req
    ),
    out = req,
    name = Some("main_pipe_req")
  )
  val s0_req = req.bits
  val s0_idx = get_idx(s0_req.vaddr)
  val s0_can_go = io.meta_read.ready && io.tag_read.ready && s1_ready && !set_conflict
  val s0_fire = req.valid && s0_can_go

  val bank_write = VecInit((0 until DCacheBanks).map(i => get_mask_of_bank(i, s0_req.store_mask).orR)).asUInt
  val bank_full_write = VecInit((0 until DCacheBanks).map(i => get_mask_of_bank(i, s0_req.store_mask).andR)).asUInt
  val banks_full_overwrite = bank_full_write.andR

  val banked_store_rmask = bank_write & ~bank_full_write
  val banked_full_rmask = ~0.U(DCacheBanks.W)
  val banked_none_rmask = 0.U(DCacheBanks.W)

  val store_need_data = !s0_req.probe && s0_req.isStore && banked_store_rmask.orR
  val probe_need_data = s0_req.probe
  val amo_need_data = !s0_req.probe && s0_req.isAMO
  val miss_need_data = s0_req.miss
  val replace_need_data = s0_req.replace

  val banked_need_data = store_need_data || probe_need_data || amo_need_data || miss_need_data || replace_need_data

  val s0_banked_rmask = Mux(store_need_data, banked_store_rmask,
    Mux(probe_need_data || amo_need_data || miss_need_data || replace_need_data,
      banked_full_rmask,
      banked_none_rmask
    ))

  // generate wmask here and use it in stage 2
  val banked_store_wmask = bank_write
  val banked_full_wmask = ~0.U(DCacheBanks.W)
  val banked_none_wmask = 0.U(DCacheBanks.W)

  // s1: read data
  val s1_valid = RegInit(false.B)
  val s1_need_data = RegEnable(banked_need_data, s0_fire)
  val s1_req = RegEnable(s0_req, s0_fire)
  val s1_banked_rmask = RegEnable(s0_banked_rmask, s0_fire)
  val s1_banked_store_wmask = RegEnable(banked_store_wmask, s0_fire)
  val s1_can_go = s2_ready && (io.data_read.ready || !s1_need_data)
  val s1_fire = s1_valid && s1_can_go
  val s1_idx = get_idx(s1_req.vaddr)
  when (s0_fire) {
    s1_valid := true.B
  }.elsewhen (s1_fire) {
    s1_valid := false.B
  }
  s1_ready := !s1_valid || s1_can_go
  s1_s0_set_conflict := s1_valid && s0_idx === s1_idx

  def getMeta(encMeta: UInt): UInt = {
    require(encMeta.getWidth == encMetaBits)
    encMeta(metaBits - 1, 0)
  }

  val tag_resp = Wire(Vec(nWays, UInt(tagBits.W)))
  val ecc_meta_resp = Wire(Vec(nWays, UInt(encMetaBits.W)))
  tag_resp := Mux(RegNext(s0_fire), io.tag_resp, RegNext(tag_resp))
  ecc_meta_resp := Mux(RegNext(s0_fire), io.meta_resp, RegNext(ecc_meta_resp))
  val meta_resp = ecc_meta_resp.map(getMeta(_))

  def wayMap[T <: Data](f: Int => T) = VecInit((0 until nWays).map(f))
  val s1_tag_eq_way = wayMap((w: Int) => tag_resp(w) === get_tag(s1_req.addr)).asUInt
  val s1_tag_match_way = wayMap((w: Int) => s1_tag_eq_way(w) && Meta(meta_resp(w)).coh.isValid()).asUInt
  val s1_tag_match = s1_tag_match_way.orR

  val s1_hit_tag = Mux(s1_tag_match, Mux1H(s1_tag_match_way, wayMap(w => tag_resp(w))), get_tag(s1_req.addr))
  val s1_hit_coh = ClientMetadata(Mux(s1_tag_match, Mux1H(s1_tag_match_way, wayMap(w => meta_resp(w))), 0.U))

  // replacement policy
  val s1_repl_way_en = WireInit(0.U(nWays.W))
  s1_repl_way_en := Mux(RegNext(s0_fire), UIntToOH(io.replace_way.way), RegNext(s1_repl_way_en))
  val s1_repl_tag = Mux1H(s1_repl_way_en, wayMap(w => tag_resp(w)))
  val s1_repl_coh = Mux1H(s1_repl_way_en, wayMap(w => meta_resp(w))).asTypeOf(new ClientMetadata)

  val s1_need_replacement = (s1_req.miss || s1_req.isStore && !s1_req.probe) && !s1_tag_match
  val s1_way_en = Mux(s1_req.replace, s1_req.replace_way_en, Mux(s1_need_replacement, s1_repl_way_en, s1_tag_match_way))
  val s1_tag = Mux(s1_req.replace, get_tag(s1_req.addr), Mux(s1_need_replacement, s1_repl_tag, s1_hit_tag))
  val s1_coh = Mux(
    s1_req.replace,
    Mux1H(s1_req.replace_way_en, meta_resp.map(ClientMetadata(_))),
    Mux(s1_need_replacement, s1_repl_coh, s1_hit_coh)
  )

  val s1_has_permission = s1_hit_coh.onAccess(s1_req.cmd)._1
  val s1_hit = s1_tag_match && s1_has_permission
  val s1_pregen_can_go_to_mq = !s1_req.replace && !s1_req.probe && !s1_req.miss && (s1_req.isStore || s1_req.isAMO) && !s1_hit

  // s2: select data, return resp if this is a store miss
  val s2_valid = RegInit(false.B)
  val s2_req = RegEnable(s1_req, s1_fire)
  val s2_tag_match = RegEnable(s1_tag_match, s1_fire)
  val s2_hit_coh = RegEnable(s1_hit_coh, s1_fire)
  val (s2_has_permission, _, s2_new_hit_coh) = s2_hit_coh.onAccess(s2_req.cmd)
  val s2_repl_way_en = RegEnable(s1_repl_way_en, s1_fire)
  val s2_repl_tag = RegEnable(s1_repl_tag, s1_fire)
  val s2_repl_coh = RegEnable(s1_repl_coh, s1_fire)
  val s2_need_replacement = RegEnable(s1_need_replacement, s1_fire)
  val s2_idx = get_idx(s2_req.vaddr)
  val s2_way_en = RegEnable(s1_way_en, s1_fire)
  val s2_tag = RegEnable(s1_tag, s1_fire)
  val s2_coh = RegEnable(s1_coh, s1_fire)
  val s2_banked_store_wmask = RegEnable(s1_banked_store_wmask, s1_fire)

  val s2_hit = s2_tag_match && s2_has_permission
  val s2_amo_hit = s2_hit && !s2_req.probe && !s2_req.miss && s2_req.isAMO
  val s2_store_hit = s2_hit && !s2_req.probe && !s2_req.miss && s2_req.isStore

  s2_s0_set_conlict := s2_valid && s0_idx === s2_idx

  // For a store req, it either hits and goes to s3, or miss and enter miss queue immediately
  val s2_can_go_to_s3 = (s2_req.replace || s2_req.probe || s2_req.miss || (s2_req.isStore || s2_req.isAMO) && s2_hit) && s3_ready
  val s2_can_go_to_mq = RegEnable(s1_pregen_can_go_to_mq, s1_fire)
  assert(RegNext(!(s2_valid && s2_can_go_to_s3 && s2_can_go_to_mq)))
  val s2_can_go = s2_can_go_to_s3 || s2_can_go_to_mq
  val s2_fire = s2_valid && s2_can_go
  val s2_fire_to_s3 = s2_valid && s2_can_go_to_s3
  when (s1_fire) {
    s2_valid := true.B
  }.elsewhen (s2_fire) {
    s2_valid := false.B
  }
  s2_ready := !s2_valid || s2_can_go
  val replay = !io.miss_req.ready

  val data_resp = Wire(io.data_resp.cloneType)
  data_resp := Mux(RegNext(s1_fire), io.data_resp, RegNext(data_resp))
  val s2_store_data_merged = Wire(Vec(DCacheBanks, UInt(DCacheSRAMRowBits.W)))

  def mergePutData(old_data: UInt, new_data: UInt, wmask: UInt): UInt = {
    val full_wmask = FillInterleaved(8, wmask)
    ((~full_wmask & old_data) | (full_wmask & new_data))
  }

  val s2_data = WireInit(VecInit((0 until DCacheBanks).map(i => {
    val decoded = cacheParams.dataCode.decode(data_resp(i).asECCData())
    // assert(!RegNext(s2_valid && s2_hit && decoded.uncorrectable))
    // TODO: trigger ecc error
    data_resp(i).raw_data
  })))

  for (i <- 0 until DCacheBanks) {
    val old_data = s2_data(i)
    val new_data = get_data_of_bank(i, s2_req.store_data)
    // for amo hit, we should use read out SRAM data
    // do not merge with store data
    val wmask = Mux(s2_amo_hit, 0.U(wordBytes.W), get_mask_of_bank(i, s2_req.store_mask))
    s2_store_data_merged(i) := mergePutData(old_data, new_data, wmask)
  }

  val s2_data_word = s2_store_data_merged(s2_req.word_idx)

  // s3: write data, meta and tag
  val s3_valid = RegInit(false.B)
  val s3_req = RegEnable(s2_req, s2_fire_to_s3)
  val s3_idx = get_idx(s3_req.vaddr)
  val s3_tag = RegEnable(s2_tag, s2_fire_to_s3)
  val s3_tag_match = RegEnable(s2_tag_match, s2_fire_to_s3)
  val s3_coh = RegEnable(s2_coh, s2_fire_to_s3)
  val s3_hit = RegEnable(s2_hit, s2_fire_to_s3)
  val s3_amo_hit = RegEnable(s2_amo_hit, s2_fire_to_s3)
  val s3_store_hit = RegEnable(s2_store_hit, s2_fire_to_s3)
  val s3_hit_coh = RegEnable(s2_hit_coh, s2_fire_to_s3)
  val s3_new_hit_coh = RegEnable(s2_new_hit_coh, s2_fire_to_s3)
  val s3_way_en = RegEnable(s2_way_en, s2_fire_to_s3)
  val s3_banked_store_wmask = RegEnable(s2_banked_store_wmask, s2_fire_to_s3)
  val s3_store_data_merged = RegEnable(s2_store_data_merged, s2_fire_to_s3)
  val s3_data_word = RegEnable(s2_data_word, s2_fire_to_s3)
  val s3_data = RegEnable(s2_data, s2_fire_to_s3)
  val (probe_has_dirty_data, probe_shrink_param, probe_new_coh) = s3_coh.onProbe(s3_req.probe_param)
  val s3_need_replacement = RegEnable(s2_need_replacement, s2_fire_to_s3)

  val miss_update_meta = s3_req.miss
  val probe_update_meta = s3_req.probe && s3_tag_match && s3_coh =/= probe_new_coh
  val store_update_meta = s3_req.isStore && !s3_req.probe && s3_hit_coh =/= s3_new_hit_coh
  val amo_update_meta = s3_req.isAMO && !s3_req.probe && s3_hit_coh =/= s3_new_hit_coh
  val amo_wait_amoalu = s3_req.isAMO && s3_req.cmd =/= M_XLR && s3_req.cmd =/= M_XSC
  val update_meta = (miss_update_meta || probe_update_meta || store_update_meta || amo_update_meta) && !s3_req.replace

  def missCohGen(cmd: UInt, param: UInt, dirty: Bool) = {
    val c = categorize(cmd)
    MuxLookup(Cat(c, param, dirty), Nothing, Seq(
      //(effect param) -> (next)
      Cat(rd, toB, false.B)  -> Branch,
      Cat(rd, toB, true.B)   -> Branch,
      Cat(rd, toT, false.B)  -> Trunk,
      Cat(rd, toT, true.B)   -> Dirty,
      Cat(wi, toT, false.B)  -> Trunk,
      Cat(wi, toT, true.B)   -> Dirty,
      Cat(wr, toT, false.B)  -> Dirty,
      Cat(wr, toT, true.B)   -> Dirty))
  }
  val miss_new_coh = ClientMetadata(missCohGen(s3_req.cmd, s3_req.miss_param, s3_req.miss_dirty))

  val new_coh = Mux(
    miss_update_meta,
    miss_new_coh,
    Mux(
      probe_update_meta,
      probe_new_coh,
      Mux(
        store_update_meta || amo_update_meta,
        s3_new_hit_coh,
        ClientMetadata.onReset
      )
    )
  )

  // LR, SC and AMO
  val debug_sc_fail_addr = RegInit(0.U)
  val debug_sc_fail_cnt  = RegInit(0.U(8.W))

  val lrsc_count = RegInit(0.U(log2Ceil(lrscCycles).W))
  val lrsc_valid = lrsc_count > lrscBackoff.U
  val lrsc_addr  = Reg(UInt())
  val s3_lr = !s3_req.probe && s3_req.isAMO && s3_req.cmd === M_XLR
  val s3_sc = !s3_req.probe && s3_req.isAMO && s3_req.cmd === M_XSC
  val s3_lrsc_addr_match = lrsc_valid && lrsc_addr === get_block_addr(s3_req.addr)
  val s3_sc_fail = s3_sc && !s3_lrsc_addr_match
  val s3_sc_resp = Mux(s3_sc_fail, 1.U, 0.U)

  val s3_can_do_amo = (s3_req.miss && !s3_req.probe && s3_req.source === AMO_SOURCE.U) || s3_amo_hit
  val s3_can_do_amo_write = s3_can_do_amo && isWrite(s3_req.cmd) && !s3_sc_fail

  when (s3_valid && (s3_lr || s3_sc)) {
    when (s3_can_do_amo && s3_lr) {
      lrsc_count := (lrscCycles - 1).U
      lrsc_addr := get_block_addr(s3_req.addr)
    } .otherwise {
      lrsc_count := 0.U
    }
  } .elsewhen (lrsc_count > 0.U) {
    lrsc_count := lrsc_count - 1.U
  }

  io.lrsc_locked_block.valid := lrsc_valid
  io.lrsc_locked_block.bits  := lrsc_addr

  // When we update update_resv_set, block all probe req in the next cycle
  // It should give Probe reservation set addr compare an independent cycle,
  // which will lead to better timing
  io.update_resv_set := s3_valid && s3_lr && s3_can_do_amo

  // when we release this block,
  // we invalidate this reservation set
  when (io.invalid_resv_set) {
    lrsc_count := 0.U
  }

  when (s3_valid) {
    when (s3_req.addr === debug_sc_fail_addr) {
      when (s3_sc_fail) {
        debug_sc_fail_cnt := debug_sc_fail_cnt + 1.U
      } .elsewhen (s3_sc) {
        debug_sc_fail_cnt := 0.U
      }
    } .otherwise {
      when (s3_sc_fail) {
        debug_sc_fail_addr := s3_req.addr
        debug_sc_fail_cnt  := 1.U
      }
    }
  }
  assert(debug_sc_fail_cnt < 100.U, "L1DCache failed too many SCs in a row")


  val banked_amo_wmask = UIntToOH(s3_req.word_idx)
//  val banked_wmask = s3_banked_store_wmask
  val banked_wmask = Mux(
    s3_req.miss,
    banked_full_wmask,
    Mux(
      s3_store_hit,
      s3_banked_store_wmask,
      Mux(
        s3_can_do_amo_write,
        banked_amo_wmask,
        banked_none_wmask
      )
    )
  )
  val update_data = banked_wmask.asUInt.orR

  // generate write data
  // AMO hits
  val s3_s_amoalu = RegInit(false.B)
  val do_amoalu = amo_wait_amoalu && s3_valid && !s3_s_amoalu
  val amoalu   = Module(new AMOALU(wordBits))
  amoalu.io.mask := s3_req.amo_mask
  amoalu.io.cmd  := s3_req.cmd
  amoalu.io.lhs  := s3_data_word
  amoalu.io.rhs  := s3_req.amo_data

  // merge amo write data
//  val amo_bitmask = FillInterleaved(8, s3_req.amo_mask)
  val s3_amo_data_merged = Wire(Vec(DCacheBanks, UInt(DCacheSRAMRowBits.W)))
  val s3_sc_data_merged = Wire(Vec(DCacheBanks, UInt(DCacheSRAMRowBits.W)))
  for (i <- 0 until DCacheBanks) {
    val old_data = s3_store_data_merged(i)
    val new_data = amoalu.io.out
    val wmask = Mux(
      s3_req.word_idx === i.U,
      ~0.U(wordBytes.W),
      0.U(wordBytes.W)
    )
    s3_amo_data_merged(i) := mergePutData(old_data, new_data, wmask)
//    s3_sc_data_merged(i) := amo_bitmask & s3_req.amo_data | ~amo_bitmask & old_data
    s3_sc_data_merged(i) := mergePutData(old_data, s3_req.amo_data,
      Mux(s3_req.word_idx === i.U && !s3_sc_fail, s3_req.amo_mask, 0.U(wordBytes.W))
    )
  }
  val s3_amo_data_merged_reg = RegEnable(s3_amo_data_merged, do_amoalu)
  when(do_amoalu){
    s3_s_amoalu := true.B
  }

  val miss_wb = s3_req.miss && s3_need_replacement && s3_coh.state =/= ClientStates.Nothing
  val probe_wb = s3_req.probe
  val replace_wb = s3_req.replace
  val need_wb = miss_wb || probe_wb || replace_wb

  val (_, miss_shrink_param, _) = s3_coh.onCacheControl(M_FLUSH)
  val writeback_param = Mux(probe_wb, probe_shrink_param, miss_shrink_param)
  val writeback_data = if (dcacheParameters.alwaysReleaseData) {
    s3_tag_match && s3_req.probe && s3_req.probe_need_data ||
      s3_coh === ClientStates.Dirty || (miss_wb || replace_wb) && s3_coh.state =/= ClientStates.Nothing
  } else {
    s3_tag_match && s3_req.probe && s3_req.probe_need_data || s3_coh === ClientStates.Dirty
  }

  val s3_probe_can_go = s3_req.probe && io.wb.ready && (io.meta_write.ready || !probe_update_meta)
  val s3_store_can_go = s3_req.isStore && !s3_req.probe && (io.meta_write.ready || !store_update_meta) && (io.data_write.ready || !update_data)
  val s3_amo_can_go = s3_amo_hit && (io.meta_write.ready || !amo_update_meta) && (io.data_write.ready || !update_data) && (s3_s_amoalu || !amo_wait_amoalu)
  val s3_miss_can_go = s3_req.miss &&
    (io.meta_write.ready || !amo_update_meta) &&
    (io.data_write.ready || !update_data) &&
    (s3_s_amoalu || !amo_wait_amoalu) &&
    io.tag_write.ready &&
    io.wb.ready
  val s3_replace_nothing = s3_req.replace && s3_coh.state === ClientStates.Nothing
  val s3_replace_can_go = s3_req.replace && (s3_replace_nothing || io.wb.ready)
  val s3_can_go = s3_probe_can_go || s3_store_can_go || s3_amo_can_go || s3_miss_can_go || s3_replace_can_go
  val s3_fire = s3_valid && s3_can_go
  when (s2_fire_to_s3) {
    s3_valid := true.B
  }.elsewhen (s3_fire) {
    s3_valid := false.B
  }
  s3_ready := !s3_valid || s3_can_go
  s3_s0_set_conflict := s3_valid && s3_idx === s0_idx
  assert(RegNext(!s3_valid || !(s3_req.isStore && !s3_req.probe) || s3_hit)) // miss store should never come to s3

  when(s3_fire) {
    s3_s_amoalu := false.B
  }

  req.ready := s0_can_go

  io.meta_read.valid := req.valid && s1_ready && !set_conflict
  io.meta_read.bits.idx := get_idx(s0_req.vaddr)
  io.meta_read.bits.way_en := Mux(s0_req.replace, s0_req.replace_way_en, ~0.U(nWays.W))

  io.tag_read.valid := req.valid && s1_ready && !set_conflict && !s0_req.replace
  io.tag_read.bits.idx := get_idx(s0_req.vaddr)
  io.tag_read.bits.way_en := ~0.U(nWays.W)

  io.data_read.valid := s1_valid && s1_need_data && s2_ready
  io.data_read.bits.rmask := s1_banked_rmask
  io.data_read.bits.way_en := s1_way_en
  io.data_read.bits.addr := s1_req.vaddr

  io.miss_req.valid := s2_valid && s2_can_go_to_mq
  val miss_req = io.miss_req.bits
  miss_req := DontCare
  miss_req.source := s2_req.source
  miss_req.cmd := s2_req.cmd
  miss_req.addr := s2_req.addr
  miss_req.vaddr := s2_req.vaddr
  miss_req.way_en := s2_way_en
  miss_req.store_data := s2_req.store_data
  miss_req.store_mask := s2_req.store_mask
  miss_req.word_idx := s2_req.word_idx
  miss_req.amo_data := s2_req.amo_data
  miss_req.amo_mask := s2_req.amo_mask
  miss_req.req_coh := s2_hit_coh
  miss_req.replace_coh := s2_repl_coh
  miss_req.replace_tag := s2_repl_tag
  miss_req.id := s2_req.id
  miss_req.cancel := false.B

  io.store_replay_resp.valid := s2_valid && s2_can_go_to_mq && replay && s2_req.isStore
  io.store_replay_resp.bits.data := DontCare
  io.store_replay_resp.bits.miss := true.B
  io.store_replay_resp.bits.replay := true.B
  io.store_replay_resp.bits.id := s2_req.id

  io.store_hit_resp.valid := s3_valid && s3_store_can_go
  io.store_hit_resp.bits.data := DontCare
  io.store_hit_resp.bits.miss := false.B
  io.store_hit_resp.bits.replay := false.B
  io.store_hit_resp.bits.id := s3_req.id

  io.release_update.valid := s3_valid && (s3_store_can_go || s3_amo_can_go) && s3_hit && update_data
  io.release_update.bits.addr := s3_req.addr
  io.release_update.bits.mask := Mux(s3_store_hit, s3_banked_store_wmask, banked_amo_wmask)
  io.release_update.bits.data := Mux(
    amo_wait_amoalu,
    s3_amo_data_merged_reg,
    Mux(
      s3_sc,
      s3_sc_data_merged,
      s3_store_data_merged
    )
  ).asUInt

  val atomic_hit_resp = Wire(new AtomicsResp)
  atomic_hit_resp.data := Mux(s3_sc, s3_sc_resp, s3_data_word)
  atomic_hit_resp.miss := false.B
  atomic_hit_resp.miss_id := s3_req.miss_id
  atomic_hit_resp.replay := false.B
  atomic_hit_resp.ack_miss_queue := s3_req.miss
  atomic_hit_resp.id := lrsc_valid
  val atomic_replay_resp = Wire(new AtomicsResp)
  atomic_replay_resp.data := DontCare
  atomic_replay_resp.miss := true.B
  atomic_replay_resp.miss_id := DontCare
  atomic_replay_resp.replay := true.B
  atomic_replay_resp.ack_miss_queue := false.B
  atomic_replay_resp.id := DontCare
  val atomic_replay_resp_valid = s2_valid && s2_can_go_to_mq && replay && s2_req.isAMO
  val atomic_hit_resp_valid = s3_valid && (s3_amo_can_go || s3_miss_can_go && s3_req.isAMO)
  io.atomic_resp.valid := atomic_replay_resp_valid || atomic_hit_resp_valid
  io.atomic_resp.bits := Mux(atomic_replay_resp_valid, atomic_replay_resp, atomic_hit_resp)

  io.replace_resp.valid := s3_fire && s3_req.replace
  io.replace_resp.bits := s3_req.miss_id

  io.meta_write.valid := s3_fire && update_meta
  io.meta_write.bits.idx := s3_idx
  io.meta_write.bits.way_en := s3_way_en
  io.meta_write.bits.tag := get_tag(s3_req.addr)
  io.meta_write.bits.meta.coh := new_coh

  io.tag_write.valid := s3_fire && s3_req.miss
  io.tag_write.bits.idx := s3_idx
  io.tag_write.bits.way_en := s3_way_en
  io.tag_write.bits.tag := get_tag(s3_req.addr)

  io.data_write.valid := s3_fire && update_data
  io.data_write.bits.way_en := s3_way_en
  io.data_write.bits.addr := s3_req.vaddr
  io.data_write.bits.wmask := banked_wmask
  io.data_write.bits.data := Mux(
    amo_wait_amoalu,
    s3_amo_data_merged_reg,
    Mux(
      s3_sc,
      s3_sc_data_merged,
      s3_store_data_merged
    )
  )
  assert(RegNext(!io.meta_write.valid || !s3_req.replace))
  assert(RegNext(!io.tag_write.valid || !s3_req.replace))
  assert(RegNext(!io.data_write.valid || !s3_req.replace))

  io.wb.valid := s3_valid && (
    // replace
    s3_req.replace && !s3_replace_nothing ||
    // probe can go to wbq
    s3_req.probe && (io.meta_write.ready || !probe_update_meta) ||
      // amo miss can go to wbq
      s3_req.miss &&
        (io.meta_write.ready || !amo_update_meta) &&
        (io.data_write.ready || !update_data) &&
        (s3_s_amoalu || !amo_wait_amoalu) &&
        io.tag_write.ready
    ) && need_wb
  io.wb.bits.addr := get_block_addr(Cat(s3_tag, get_untag(s3_req.vaddr)))
  io.wb.bits.param := writeback_param
  io.wb.bits.voluntary := s3_req.miss || s3_req.replace
  io.wb.bits.hasData := writeback_data
  io.wb.bits.dirty := s3_coh === ClientStates.Dirty
  io.wb.bits.data := s3_data.asUInt()
  io.wb.bits.delay_release := s3_req.replace
  io.wb.bits.miss_id := s3_req.miss_id

  io.replace_access.valid := RegNext(s1_fire && (s1_req.isAMO || s1_req.isStore) && !s1_req.probe && s1_tag_match)
  io.replace_access.bits.set := s2_idx
  io.replace_access.bits.way := RegNext(OHToUInt(s1_way_en))

  io.replace_way.set.valid := RegNext(s0_fire)
  io.replace_way.set.bits := s1_idx

  // TODO: consider block policy of a finer granularity
  io.status.s0_set.valid := req.valid
  io.status.s0_set.bits := get_idx(s0_req.vaddr)
  io.status.s1.valid := s1_valid
  io.status.s1.bits.set := s1_idx
  io.status.s1.bits.way_en := s1_way_en
  io.status.s2.valid := s2_valid && !s2_req.replace
  io.status.s2.bits.set := s2_idx
  io.status.s2.bits.way_en := s2_way_en
  io.status.s3.valid := s3_valid && !s3_req.replace
  io.status.s3.bits.set := s3_idx
  io.status.s3.bits.way_en := s3_way_en

  val perfinfo = IO(new Bundle(){
    val perfEvents = Output(new PerfEventsBundle(2))
  })
  val perfEvents = Seq(
    ("dcache_mp_req                    ", s0_fire                                                                     ),
    ("dcache_mp_total_penalty          ", (PopCount(VecInit(Seq(s0_fire, s1_valid, s2_valid, s3_valid))))             ),
  )

  for (((perf_out,(perf_name,perf)),i) <- perfinfo.perfEvents.perf_events.zip(perfEvents).zipWithIndex) {
    perf_out.incr_step := RegNext(perf)
  }
}