xiangshan/frontend/IFU.scala

/***************************************************************************************
* Copyright (c) 2024 Beijing Institute of Open Source Chip (BOSC)
* Copyright (c) 2020-2024 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.frontend

import chisel3._
import chisel3.util._
import org.chipsalliance.cde.config.Parameters
import utility._
import utility.ChiselDB
import xiangshan._
import xiangshan.backend.GPAMemEntry
import xiangshan.cache.mmu._
import xiangshan.frontend.icache._

trait HasInstrMMIOConst extends HasXSParameter with HasIFUConst {
  def mmioBusWidth = 64
  def mmioBusBytes = mmioBusWidth / 8
  def maxInstrLen  = 32
}

trait HasIFUConst extends HasXSParameter {
  def addrAlign(addr: UInt, bytes: Int, highest: Int): UInt =
    Cat(addr(highest - 1, log2Ceil(bytes)), 0.U(log2Ceil(bytes).W))
  def fetchQueueSize = 2

  def getBasicBlockIdx(pc: UInt, start: UInt): UInt = {
    val byteOffset = pc - start
    (byteOffset - instBytes.U)(log2Ceil(PredictWidth), instOffsetBits)
  }
}

class IfuToFtqIO(implicit p: Parameters) extends XSBundle {
  val pdWb = Valid(new PredecodeWritebackBundle)
}

class IfuToBackendIO(implicit p: Parameters) extends XSBundle {
  // write to backend gpaddr mem
  val gpaddrMem_wen   = Output(Bool())
  val gpaddrMem_waddr = Output(UInt(log2Ceil(FtqSize).W)) // Ftq Ptr
  // 2 gpaddrs, correspond to startAddr & nextLineAddr in bundle FtqICacheInfo
  // TODO: avoid cross page entry in Ftq
  val gpaddrMem_wdata = Output(new GPAMemEntry)
}

class FtqInterface(implicit p: Parameters) extends XSBundle {
  val fromFtq = Flipped(new FtqToIfuIO)
  val toFtq   = new IfuToFtqIO
}

class UncacheInterface(implicit p: Parameters) extends XSBundle {
  val fromUncache = Flipped(DecoupledIO(new InsUncacheResp))
  val toUncache   = DecoupledIO(new InsUncacheReq)
}

class NewIFUIO(implicit p: Parameters) extends XSBundle {
  val ftqInter        = new FtqInterface
  val icacheInter     = Flipped(new IFUICacheIO)
  val icacheStop      = Output(Bool())
  val icachePerfInfo  = Input(new ICachePerfInfo)
  val toIbuffer       = Decoupled(new FetchToIBuffer)
  val toBackend       = new IfuToBackendIO
  val uncacheInter    = new UncacheInterface
  val frontendTrigger = Flipped(new FrontendTdataDistributeIO)
  val rob_commits     = Flipped(Vec(CommitWidth, Valid(new RobCommitInfo)))
  val iTLBInter       = new TlbRequestIO
  val pmp             = new ICachePMPBundle
  val mmioCommitRead  = new mmioCommitRead
  val csr_fsIsOff     = Input(Bool())
}

// record the situation in which fallThruAddr falls into
// the middle of an RVI inst
class LastHalfInfo(implicit p: Parameters) extends XSBundle {
  val valid    = Bool()
  val middlePC = UInt(VAddrBits.W)
  def matchThisBlock(startAddr: UInt) = valid && middlePC === startAddr
}

class IfuToPreDecode(implicit p: Parameters) extends XSBundle {
  val data            = if (HasCExtension) Vec(PredictWidth + 1, UInt(16.W)) else Vec(PredictWidth, UInt(32.W))
  val frontendTrigger = new FrontendTdataDistributeIO
  val pc              = Vec(PredictWidth, UInt(VAddrBits.W))
}

class IfuToPredChecker(implicit p: Parameters) extends XSBundle {
  val ftqOffset  = Valid(UInt(log2Ceil(PredictWidth).W))
  val jumpOffset = Vec(PredictWidth, UInt(XLEN.W))
  val target     = UInt(VAddrBits.W)
  val instrRange = Vec(PredictWidth, Bool())
  val instrValid = Vec(PredictWidth, Bool())
  val pds        = Vec(PredictWidth, new PreDecodeInfo)
  val pc         = Vec(PredictWidth, UInt(VAddrBits.W))
  val fire_in    = Bool()
}

class FetchToIBufferDB extends Bundle {
  val start_addr   = UInt(39.W)
  val instr_count  = UInt(32.W)
  val exception    = Bool()
  val is_cache_hit = Bool()
}

class IfuWbToFtqDB extends Bundle {
  val start_addr        = UInt(39.W)
  val is_miss_pred      = Bool()
  val miss_pred_offset  = UInt(32.W)
  val checkJalFault     = Bool()
  val checkRetFault     = Bool()
  val checkTargetFault  = Bool()
  val checkNotCFIFault  = Bool()
  val checkInvalidTaken = Bool()
}

class NewIFU(implicit p: Parameters) extends XSModule
    with HasICacheParameters
    with HasXSParameter
    with HasIFUConst
    with HasPdConst
    with HasCircularQueuePtrHelper
    with HasPerfEvents
    with HasTlbConst {
  val io                       = IO(new NewIFUIO)
  val (toFtq, fromFtq)         = (io.ftqInter.toFtq, io.ftqInter.fromFtq)
  val fromICache               = io.icacheInter.resp
  val (toUncache, fromUncache) = (io.uncacheInter.toUncache, io.uncacheInter.fromUncache)

  def isCrossLineReq(start: UInt, end: UInt): Bool = start(blockOffBits) ^ end(blockOffBits)

  def numOfStage = 3
  // equal lower_result overflow bit
  def PcCutPoint = (VAddrBits / 4) - 1
  def CatPC(low: UInt, high: UInt, high1: UInt): UInt =
    Mux(
      low(PcCutPoint),
      Cat(high1, low(PcCutPoint - 1, 0)),
      Cat(high, low(PcCutPoint - 1, 0))
    )
  def CatPC(lowVec: Vec[UInt], high: UInt, high1: UInt): Vec[UInt] = VecInit(lowVec.map(CatPC(_, high, high1)))
  require(numOfStage > 1, "BPU numOfStage must be greater than 1")
  val topdown_stages = RegInit(VecInit(Seq.fill(numOfStage)(0.U.asTypeOf(new FrontendTopDownBundle))))
  // bubble events in IFU, only happen in stage 1
  val icacheMissBubble = Wire(Bool())
  val itlbMissBubble   = Wire(Bool())

  // only driven by clock, not valid-ready
  topdown_stages(0) := fromFtq.req.bits.topdown_info
  for (i <- 1 until numOfStage) {
    topdown_stages(i) := topdown_stages(i - 1)
  }
  when(icacheMissBubble) {
    topdown_stages(1).reasons(TopDownCounters.ICacheMissBubble.id) := true.B
  }
  when(itlbMissBubble) {
    topdown_stages(1).reasons(TopDownCounters.ITLBMissBubble.id) := true.B
  }
  io.toIbuffer.bits.topdown_info := topdown_stages(numOfStage - 1)
  when(fromFtq.topdown_redirect.valid) {
    // only redirect from backend, IFU redirect itself is handled elsewhere
    when(fromFtq.topdown_redirect.bits.debugIsCtrl) {
      /*
      for (i <- 0 until numOfStage) {
        topdown_stages(i).reasons(TopDownCounters.ControlRedirectBubble.id) := true.B
      }
      io.toIbuffer.bits.topdown_info.reasons(TopDownCounters.ControlRedirectBubble.id) := true.B
       */
      when(fromFtq.topdown_redirect.bits.ControlBTBMissBubble) {
        for (i <- 0 until numOfStage) {
          topdown_stages(i).reasons(TopDownCounters.BTBMissBubble.id) := true.B
        }
        io.toIbuffer.bits.topdown_info.reasons(TopDownCounters.BTBMissBubble.id) := true.B
      }.elsewhen(fromFtq.topdown_redirect.bits.TAGEMissBubble) {
        for (i <- 0 until numOfStage) {
          topdown_stages(i).reasons(TopDownCounters.TAGEMissBubble.id) := true.B
        }
        io.toIbuffer.bits.topdown_info.reasons(TopDownCounters.TAGEMissBubble.id) := true.B
      }.elsewhen(fromFtq.topdown_redirect.bits.SCMissBubble) {
        for (i <- 0 until numOfStage) {
          topdown_stages(i).reasons(TopDownCounters.SCMissBubble.id) := true.B
        }
        io.toIbuffer.bits.topdown_info.reasons(TopDownCounters.SCMissBubble.id) := true.B
      }.elsewhen(fromFtq.topdown_redirect.bits.ITTAGEMissBubble) {
        for (i <- 0 until numOfStage) {
          topdown_stages(i).reasons(TopDownCounters.ITTAGEMissBubble.id) := true.B
        }
        io.toIbuffer.bits.topdown_info.reasons(TopDownCounters.ITTAGEMissBubble.id) := true.B
      }.elsewhen(fromFtq.topdown_redirect.bits.RASMissBubble) {
        for (i <- 0 until numOfStage) {
          topdown_stages(i).reasons(TopDownCounters.RASMissBubble.id) := true.B
        }
        io.toIbuffer.bits.topdown_info.reasons(TopDownCounters.RASMissBubble.id) := true.B
      }
    }.elsewhen(fromFtq.topdown_redirect.bits.debugIsMemVio) {
      for (i <- 0 until numOfStage) {
        topdown_stages(i).reasons(TopDownCounters.MemVioRedirectBubble.id) := true.B
      }
      io.toIbuffer.bits.topdown_info.reasons(TopDownCounters.MemVioRedirectBubble.id) := true.B
    }.otherwise {
      for (i <- 0 until numOfStage) {
        topdown_stages(i).reasons(TopDownCounters.OtherRedirectBubble.id) := true.B
      }
      io.toIbuffer.bits.topdown_info.reasons(TopDownCounters.OtherRedirectBubble.id) := true.B
    }
  }

  class TlbExept(implicit p: Parameters) extends XSBundle {
    val pageFault   = Bool()
    val accessFault = Bool()
    val mmio        = Bool()
  }

  val preDecoder = Module(new PreDecode)

  val predChecker     = Module(new PredChecker)
  val frontendTrigger = Module(new FrontendTrigger)
  val (checkerIn, checkerOutStage1, checkerOutStage2) =
    (predChecker.io.in, predChecker.io.out.stage1Out, predChecker.io.out.stage2Out)

  /**
    ******************************************************************************
    * IFU Stage 0
    * - send cacheline fetch request to ICacheMainPipe
    ******************************************************************************
    */

  val f0_valid      = fromFtq.req.valid
  val f0_ftq_req    = fromFtq.req.bits
  val f0_doubleLine = fromFtq.req.bits.crossCacheline
  val f0_vSetIdx    = VecInit(get_idx(f0_ftq_req.startAddr), get_idx(f0_ftq_req.nextlineStart))
  val f0_fire       = fromFtq.req.fire

  val f0_flush, f1_flush, f2_flush, f3_flush                                     = WireInit(false.B)
  val from_bpu_f0_flush, from_bpu_f1_flush, from_bpu_f2_flush, from_bpu_f3_flush = WireInit(false.B)

  from_bpu_f0_flush := fromFtq.flushFromBpu.shouldFlushByStage2(f0_ftq_req.ftqIdx) ||
    fromFtq.flushFromBpu.shouldFlushByStage3(f0_ftq_req.ftqIdx)

  val wb_redirect, mmio_redirect, backend_redirect = WireInit(false.B)
  val f3_wb_not_flush                              = WireInit(false.B)

  backend_redirect := fromFtq.redirect.valid
  f3_flush         := backend_redirect || (wb_redirect && !f3_wb_not_flush)
  f2_flush         := backend_redirect || mmio_redirect || wb_redirect
  f1_flush         := f2_flush || from_bpu_f1_flush
  f0_flush         := f1_flush || from_bpu_f0_flush

  val f1_ready, f2_ready, f3_ready = WireInit(false.B)

  fromFtq.req.ready := f1_ready && io.icacheInter.icacheReady

  when(wb_redirect) {
    when(f3_wb_not_flush) {
      topdown_stages(2).reasons(TopDownCounters.BTBMissBubble.id) := true.B
    }
    for (i <- 0 until numOfStage - 1) {
      topdown_stages(i).reasons(TopDownCounters.BTBMissBubble.id) := true.B
    }
  }

  /** <PERF> f0 fetch bubble */

  XSPerfAccumulate("fetch_bubble_ftq_not_valid", !fromFtq.req.valid && fromFtq.req.ready)
  // XSPerfAccumulate("fetch_bubble_pipe_stall",    f0_valid && toICache(0).ready && toICache(1).ready && !f1_ready )
  // XSPerfAccumulate("fetch_bubble_icache_0_busy",   f0_valid && !toICache(0).ready  )
  // XSPerfAccumulate("fetch_bubble_icache_1_busy",   f0_valid && !toICache(1).ready  )
  XSPerfAccumulate("fetch_flush_backend_redirect", backend_redirect)
  XSPerfAccumulate("fetch_flush_wb_redirect", wb_redirect)
  XSPerfAccumulate("fetch_flush_bpu_f1_flush", from_bpu_f1_flush)
  XSPerfAccumulate("fetch_flush_bpu_f0_flush", from_bpu_f0_flush)

  /**
    ******************************************************************************
    * IFU Stage 1
    * - calculate pc/half_pc/cut_ptr for every instruction
    ******************************************************************************
    */

  val f1_valid   = RegInit(false.B)
  val f1_ftq_req = RegEnable(f0_ftq_req, f0_fire)
  // val f1_situation  = RegEnable(f0_situation,  f0_fire)
  val f1_doubleLine = RegEnable(f0_doubleLine, f0_fire)
  val f1_vSetIdx    = RegEnable(f0_vSetIdx, f0_fire)
  val f1_fire       = f1_valid && f2_ready

  f1_ready := f1_fire || !f1_valid

  from_bpu_f1_flush := fromFtq.flushFromBpu.shouldFlushByStage3(f1_ftq_req.ftqIdx) && f1_valid
  // from_bpu_f1_flush := false.B

  when(f1_flush)(f1_valid := false.B)
    .elsewhen(f0_fire && !f0_flush)(f1_valid := true.B)
    .elsewhen(f1_fire)(f1_valid := false.B)

  val f1_pc_high       = f1_ftq_req.startAddr(VAddrBits - 1, PcCutPoint)
  val f1_pc_high_plus1 = f1_pc_high + 1.U

  /**
   * In order to reduce power consumption, avoid calculating the full PC value in the first level.
   * code of original logic, this code has been deprecated
   * val f1_pc                 = VecInit(f1_pc_lower_result.map{ i =>
   *  Mux(i(f1_pc_adder_cut_point), Cat(f1_pc_high_plus1,i(f1_pc_adder_cut_point-1,0)), Cat(f1_pc_high,i(f1_pc_adder_cut_point-1,0)))})
   */
  val f1_pc_lower_result = VecInit((0 until PredictWidth).map(i =>
    Cat(0.U(1.W), f1_ftq_req.startAddr(PcCutPoint - 1, 0)) + (i * 2).U
  )) // cat with overflow bit

  val f1_pc = CatPC(f1_pc_lower_result, f1_pc_high, f1_pc_high_plus1)

  val f1_half_snpc_lower_result = VecInit((0 until PredictWidth).map(i =>
    Cat(0.U(1.W), f1_ftq_req.startAddr(PcCutPoint - 1, 0)) + ((i + 2) * 2).U
  )) // cat with overflow bit
  val f1_half_snpc = CatPC(f1_half_snpc_lower_result, f1_pc_high, f1_pc_high_plus1)

  if (env.FPGAPlatform) {
    val f1_pc_diff        = VecInit((0 until PredictWidth).map(i => f1_ftq_req.startAddr + (i * 2).U))
    val f1_half_snpc_diff = VecInit((0 until PredictWidth).map(i => f1_ftq_req.startAddr + ((i + 2) * 2).U))

    XSError(
      f1_pc.zip(f1_pc_diff).map { case (a, b) => a.asUInt =/= b.asUInt }.reduce(_ || _),
      "f1_half_snpc adder cut fail"
    )
    XSError(
      f1_half_snpc.zip(f1_half_snpc_diff).map { case (a, b) => a.asUInt =/= b.asUInt }.reduce(_ || _),
      "f1_half_snpc adder cut fail"
    )
  }

  val f1_cut_ptr = if (HasCExtension)
    VecInit((0 until PredictWidth + 1).map(i => Cat(0.U(2.W), f1_ftq_req.startAddr(blockOffBits - 1, 1)) + i.U))
  else VecInit((0 until PredictWidth).map(i => Cat(0.U(2.W), f1_ftq_req.startAddr(blockOffBits - 1, 2)) + i.U))

  /**
    ******************************************************************************
    * IFU Stage 2
    * - icache response data (latched for pipeline stop)
    * - generate exceprion bits for every instruciton (page fault/access fault/mmio)
    * - generate predicted instruction range (1 means this instruciton is in this fetch packet)
    * - cut data from cachlines to packet instruction code
    * - instruction predecode and RVC expand
    ******************************************************************************
    */

  val icacheRespAllValid = WireInit(false.B)

  val f2_valid   = RegInit(false.B)
  val f2_ftq_req = RegEnable(f1_ftq_req, f1_fire)
  // val f2_situation  = RegEnable(f1_situation,  f1_fire)
  val f2_doubleLine = RegEnable(f1_doubleLine, f1_fire)
  val f2_vSetIdx    = RegEnable(f1_vSetIdx, f1_fire)
  val f2_fire       = f2_valid && f3_ready && icacheRespAllValid

  f2_ready := f2_fire || !f2_valid
  // TODO: addr compare may be timing critical
  val f2_icache_all_resp_wire =
    fromICache.valid &&
      fromICache.bits.vaddr(0) === f2_ftq_req.startAddr &&
      (fromICache.bits.doubleline && fromICache.bits.vaddr(1) === f2_ftq_req.nextlineStart || !f2_doubleLine)
  val f2_icache_all_resp_reg = RegInit(false.B)

  icacheRespAllValid := f2_icache_all_resp_reg || f2_icache_all_resp_wire

  icacheMissBubble := io.icacheInter.topdownIcacheMiss
  itlbMissBubble   := io.icacheInter.topdownItlbMiss

  io.icacheStop := !f3_ready

  when(f2_flush)(f2_icache_all_resp_reg := false.B)
    .elsewhen(f2_valid && f2_icache_all_resp_wire && !f3_ready)(f2_icache_all_resp_reg := true.B)
    .elsewhen(f2_fire && f2_icache_all_resp_reg)(f2_icache_all_resp_reg := false.B)

  when(f2_flush)(f2_valid := false.B)
    .elsewhen(f1_fire && !f1_flush)(f2_valid := true.B)
    .elsewhen(f2_fire)(f2_valid := false.B)

  val f2_exception_in     = fromICache.bits.exception
  val f2_backendException = fromICache.bits.backendException
  // paddr and gpaddr of [startAddr, nextLineAddr]
  val f2_paddrs            = fromICache.bits.paddr
  val f2_gpaddr            = fromICache.bits.gpaddr
  val f2_isForVSnonLeafPTE = fromICache.bits.isForVSnonLeafPTE

  // FIXME: raise af if one fetch block crosses the cacheable-noncacheable boundary, might not correct
  val f2_mmio_mismatch_exception = VecInit(Seq(
    ExceptionType.none, // mark the exception only on the second line
    Mux(
      // not double-line, skip check
      !fromICache.bits.doubleline || (
        // is double-line, ask for consistent pmp_mmio and itlb_pbmt value
        fromICache.bits.pmp_mmio(0) === fromICache.bits.pmp_mmio(1) &&
          fromICache.bits.itlb_pbmt(0) === fromICache.bits.itlb_pbmt(1)
      ),
      ExceptionType.none,
      ExceptionType.af
    )
  ))

  // merge exceptions
  val f2_exception = ExceptionType.merge(f2_exception_in, f2_mmio_mismatch_exception)

  // we need only the first port, as the second is asked to be the same
  val f2_pmp_mmio  = fromICache.bits.pmp_mmio(0)
  val f2_itlb_pbmt = fromICache.bits.itlb_pbmt(0)

  /**
    * reduce the number of registers, origin code
    * f2_pc = RegEnable(f1_pc, f1_fire)
    */
  val f2_pc_lower_result = RegEnable(f1_pc_lower_result, f1_fire)
  val f2_pc_high         = RegEnable(f1_pc_high, f1_fire)
  val f2_pc_high_plus1   = RegEnable(f1_pc_high_plus1, f1_fire)
  val f2_pc              = CatPC(f2_pc_lower_result, f2_pc_high, f2_pc_high_plus1)

  val f2_cut_ptr      = RegEnable(f1_cut_ptr, f1_fire)
  val f2_resend_vaddr = RegEnable(f1_ftq_req.startAddr + 2.U, f1_fire)

  def isNextLine(pc: UInt, startAddr: UInt) =
    startAddr(blockOffBits) ^ pc(blockOffBits)

  def isLastInLine(pc: UInt) =
    pc(blockOffBits - 1, 0) === "b111110".U

  val f2_foldpc = VecInit(f2_pc.map(i => XORFold(i(VAddrBits - 1, 1), MemPredPCWidth)))
  val f2_jump_range =
    Fill(PredictWidth, !f2_ftq_req.ftqOffset.valid) | Fill(PredictWidth, 1.U(1.W)) >> ~f2_ftq_req.ftqOffset.bits
  require(
    isPow2(PredictWidth),
    "If PredictWidth does not satisfy the power of 2," +
      "expression: Fill(PredictWidth, 1.U(1.W)) >> ~f2_ftq_req.ftqOffset.bits is not right !!"
  )
  val f2_ftr_range = Fill(PredictWidth, f2_ftq_req.ftqOffset.valid) | Fill(PredictWidth, 1.U(1.W)) >> ~getBasicBlockIdx(
    f2_ftq_req.nextStartAddr,
    f2_ftq_req.startAddr
  )
  val f2_instr_range = f2_jump_range & f2_ftr_range
  val f2_exception_vec = VecInit((0 until PredictWidth).map(i =>
    MuxCase(
      ExceptionType.none,
      Seq(
        !isNextLine(f2_pc(i), f2_ftq_req.startAddr)                   -> f2_exception(0),
        (isNextLine(f2_pc(i), f2_ftq_req.startAddr) && f2_doubleLine) -> f2_exception(1)
      )
    )
  ))
  val f2_perf_info = io.icachePerfInfo

  def cut(cacheline: UInt, cutPtr: Vec[UInt]): Vec[UInt] = {
    require(HasCExtension)
    // if(HasCExtension){
    val result  = Wire(Vec(PredictWidth + 1, UInt(16.W)))
    val dataVec = cacheline.asTypeOf(Vec(blockBytes, UInt(16.W))) // 32 16-bit data vector
    (0 until PredictWidth + 1).foreach(i =>
      result(i) := dataVec(cutPtr(i)) // the max ptr is 3*blockBytes/4-1
    )
    result
    // } else {
    //   val result   = Wire(Vec(PredictWidth, UInt(32.W)) )
    //   val dataVec  = cacheline.asTypeOf(Vec(blockBytes * 2/ 4, UInt(32.W)))
    //   (0 until PredictWidth).foreach( i =>
    //     result(i) := dataVec(cutPtr(i))
    //   )
    //   result
    // }
  }

  /* NOTE: the following `Cat(_data, _data)` *is* intentional.
   * Explanation:
   * In the old design, IFU is responsible for selecting requested data from two adjacent cachelines,
   *    so IFU has to receive 2*64B (2cacheline * 64B) data from ICache, and do `Cat(_data(1), _data(0))` here.
   * However, a fetch block is 34B at max, sending 2*64B is quiet a waste of power.
   * In current design (2024.06~), ICacheDataArray is responsible for selecting data from two adjacent cachelines,
   *    so IFU only need to receive 40B (5bank * 8B) valid data, and use only one port is enough.
   * For example, when pc falls on the 6th bank in cacheline0(so this is a doubleline request):
   *                                MSB                                         LSB
   *                  cacheline 1 || 1-7 | 1-6 | 1-5 | 1-4 | 1-3 | 1-2 | 1-1 | 1-0 ||
   *                  cacheline 0 || 0-7 | 0-6 | 0-5 | 0-4 | 0-3 | 0-2 | 0-1 | 0-0 ||
   *    and ICacheDataArray will respond:
   *         fromICache.bits.data || 0-7 | 0-6 | xxx | xxx | xxx | 1-2 | 1-1 | 1-0 ||
   *    therefore simply make a copy of the response and `Cat` together, and obtain the requested data from centre:
   *          f2_data_2_cacheline || 0-7 | 0-6 | xxx | xxx | xxx | 1-2 | 1-1 | 1-0 | 0-7 | 0-6 | xxx | xxx | xxx | 1-2 | 1-1 | 1-0 ||
   *                                             requested data: ^-----------------------------^
   * For another example, pc falls on the 1st bank in cacheline 0, we have:
   *         fromICache.bits.data || xxx | xxx | 0-5 | 0-4 | 0-3 | 0-2 | 0-1 | xxx ||
   *          f2_data_2_cacheline || xxx | xxx | 0-5 | 0-4 | 0-3 | 0-2 | 0-1 | xxx | xxx | xxx | 0-5 | 0-4 | 0-3 | 0-2 | 0-1 | xxx ||
   *                                                                           requested data: ^-----------------------------^
   * Each "| x-y |" block is a 8B bank from cacheline(x).bank(y)
   * Please also refer to:
   * - DataArray selects data:
   * https://github.com/OpenXiangShan/XiangShan/blob/d4078d6edbfb4611ba58c8b0d1d8236c9115dbfc/src/main/scala/xiangshan/frontend/icache/ICache.scala#L355-L381
   * https://github.com/OpenXiangShan/XiangShan/blob/d4078d6edbfb4611ba58c8b0d1d8236c9115dbfc/src/main/scala/xiangshan/frontend/icache/ICache.scala#L149-L161
   * - ICache respond to IFU:
   * https://github.com/OpenXiangShan/XiangShan/blob/d4078d6edbfb4611ba58c8b0d1d8236c9115dbfc/src/main/scala/xiangshan/frontend/icache/ICacheMainPipe.scala#L473
   */
  val f2_data_2_cacheline = Cat(fromICache.bits.data, fromICache.bits.data)

  val f2_cut_data = cut(f2_data_2_cacheline, f2_cut_ptr)

  /** predecode (include RVC expander) */
  // preDecoderRegIn.data := f2_reg_cut_data
  // preDecoderRegInIn.frontendTrigger := io.frontendTrigger
  // preDecoderRegInIn.csrTriggerEnable := io.csrTriggerEnable
  // preDecoderRegIn.pc  := f2_pc

  val preDecoderIn = preDecoder.io.in
  preDecoderIn.valid                := f2_valid
  preDecoderIn.bits.data            := f2_cut_data
  preDecoderIn.bits.frontendTrigger := io.frontendTrigger
  preDecoderIn.bits.pc              := f2_pc
  val preDecoderOut = preDecoder.io.out

  // val f2_expd_instr     = preDecoderOut.expInstr
  val f2_instr        = preDecoderOut.instr
  val f2_pd           = preDecoderOut.pd
  val f2_jump_offset  = preDecoderOut.jumpOffset
  val f2_hasHalfValid = preDecoderOut.hasHalfValid
  /* if there is a cross-page RVI instruction, and the former page has no exception,
   * whether it has exception is actually depends on the latter page
   */
  val f2_crossPage_exception_vec = VecInit((0 until PredictWidth).map { i =>
    Mux(
      isLastInLine(f2_pc(i)) && !f2_pd(i).isRVC && f2_doubleLine && !ExceptionType.hasException(f2_exception(0)),
      f2_exception(1),
      ExceptionType.none
    )
  })
  XSPerfAccumulate("fetch_bubble_icache_not_resp", f2_valid && !icacheRespAllValid)

  /**
    ******************************************************************************
    * IFU Stage 3
    * - handle MMIO instruciton
    *  -send request to Uncache fetch Unit
    *  -every packet include 1 MMIO instruction
    *  -MMIO instructions will stop fetch pipeline until commiting from RoB
    *  -flush to snpc (send ifu_redirect to Ftq)
    * - Ibuffer enqueue
    * - check predict result in Frontend (jalFault/retFault/notCFIFault/invalidTakenFault/targetFault)
    * - handle last half RVI instruction
    ******************************************************************************
    */

  val expanders = Seq.fill(PredictWidth)(Module(new RVCExpander))

  val f3_valid   = RegInit(false.B)
  val f3_ftq_req = RegEnable(f2_ftq_req, f2_fire)
  // val f3_situation      = RegEnable(f2_situation,  f2_fire)
  val f3_doubleLine = RegEnable(f2_doubleLine, f2_fire)
  val f3_fire       = io.toIbuffer.fire

  val f3_cut_data = RegEnable(f2_cut_data, f2_fire)

  val f3_exception        = RegEnable(f2_exception, f2_fire)
  val f3_pmp_mmio         = RegEnable(f2_pmp_mmio, f2_fire)
  val f3_itlb_pbmt        = RegEnable(f2_itlb_pbmt, f2_fire)
  val f3_backendException = RegEnable(f2_backendException, f2_fire)

  val f3_instr = RegEnable(f2_instr, f2_fire)

  expanders.zipWithIndex.foreach { case (expander, i) =>
    expander.io.in      := f3_instr(i)
    expander.io.fsIsOff := io.csr_fsIsOff
  }
  // Use expanded instruction only when input is legal.
  // Otherwise use origin illegal RVC instruction.
  val f3_expd_instr = VecInit(expanders.map { expander: RVCExpander =>
    Mux(expander.io.ill, expander.io.in, expander.io.out.bits)
  })
  val f3_ill = VecInit(expanders.map(_.io.ill))

  val f3_pd_wire                 = RegEnable(f2_pd, f2_fire)
  val f3_pd                      = WireInit(f3_pd_wire)
  val f3_jump_offset             = RegEnable(f2_jump_offset, f2_fire)
  val f3_exception_vec           = RegEnable(f2_exception_vec, f2_fire)
  val f3_crossPage_exception_vec = RegEnable(f2_crossPage_exception_vec, f2_fire)

  val f3_pc_lower_result = RegEnable(f2_pc_lower_result, f2_fire)
  val f3_pc_high         = RegEnable(f2_pc_high, f2_fire)
  val f3_pc_high_plus1   = RegEnable(f2_pc_high_plus1, f2_fire)
  val f3_pc              = CatPC(f3_pc_lower_result, f3_pc_high, f3_pc_high_plus1)

  val f3_pc_last_lower_result_plus2 = RegEnable(f2_pc_lower_result(PredictWidth - 1) + 2.U, f2_fire)
  val f3_pc_last_lower_result_plus4 = RegEnable(f2_pc_lower_result(PredictWidth - 1) + 4.U, f2_fire)
  // val f3_half_snpc      = RegEnable(f2_half_snpc,   f2_fire)

  /**
    ***********************************************************************
    * Half snpc(i) is larger than pc(i) by 4. Using pc to calculate half snpc may be a good choice.
    ***********************************************************************
    */
  val f3_half_snpc = Wire(Vec(PredictWidth, UInt(VAddrBits.W)))
  for (i <- 0 until PredictWidth) {
    if (i == (PredictWidth - 2)) {
      f3_half_snpc(i) := CatPC(f3_pc_last_lower_result_plus2, f3_pc_high, f3_pc_high_plus1)
    } else if (i == (PredictWidth - 1)) {
      f3_half_snpc(i) := CatPC(f3_pc_last_lower_result_plus4, f3_pc_high, f3_pc_high_plus1)
    } else {
      f3_half_snpc(i) := f3_pc(i + 2)
    }
  }

  val f3_instr_range       = RegEnable(f2_instr_range, f2_fire)
  val f3_foldpc            = RegEnable(f2_foldpc, f2_fire)
  val f3_hasHalfValid      = RegEnable(f2_hasHalfValid, f2_fire)
  val f3_paddrs            = RegEnable(f2_paddrs, f2_fire)
  val f3_gpaddr            = RegEnable(f2_gpaddr, f2_fire)
  val f3_isForVSnonLeafPTE = RegEnable(f2_isForVSnonLeafPTE, f2_fire)
  val f3_resend_vaddr      = RegEnable(f2_resend_vaddr, f2_fire)

  // Expand 1 bit to prevent overflow when assert
  val f3_ftq_req_startAddr     = Cat(0.U(1.W), f3_ftq_req.startAddr)
  val f3_ftq_req_nextStartAddr = Cat(0.U(1.W), f3_ftq_req.nextStartAddr)
  // brType, isCall and isRet generation is delayed to f3 stage
  val f3Predecoder = Module(new F3Predecoder)

  f3Predecoder.io.in.instr := f3_instr

  f3_pd.zipWithIndex.map { case (pd, i) =>
    pd.brType := f3Predecoder.io.out.pd(i).brType
    pd.isCall := f3Predecoder.io.out.pd(i).isCall
    pd.isRet  := f3Predecoder.io.out.pd(i).isRet
  }

  val f3PdDiff = f3_pd_wire.zip(f3_pd).map { case (a, b) => a.asUInt =/= b.asUInt }.reduce(_ || _)
  XSError(f3_valid && f3PdDiff, "f3 pd diff")

  when(f3_valid && !f3_ftq_req.ftqOffset.valid) {
    assert(
      f3_ftq_req_startAddr + (2 * PredictWidth).U >= f3_ftq_req_nextStartAddr,
      s"More tha ${2 * PredictWidth} Bytes fetch is not allowed!"
    )
  }

  /*** MMIO State Machine***/
  val f3_mmio_data          = Reg(Vec(2, UInt(16.W)))
  val mmio_is_RVC           = RegInit(false.B)
  val mmio_resend_addr      = RegInit(0.U(PAddrBits.W))
  val mmio_resend_exception = RegInit(0.U(ExceptionType.width.W))
  // NOTE: we dont use GPAddrBits here, refer to ICacheMainPipe.scala L43-48 and PR#3795
  val mmio_resend_gpaddr            = RegInit(0.U(PAddrBitsMax.W))
  val mmio_resend_isForVSnonLeafPTE = RegInit(false.B)

  // last instuction finish
  val is_first_instr = RegInit(true.B)

  /*** Determine whether the MMIO instruction is executable based on the previous prediction block ***/
  io.mmioCommitRead.mmioFtqPtr := RegNext(f3_ftq_req.ftqIdx - 1.U)

  val m_idle :: m_waitLastCmt :: m_sendReq :: m_waitResp :: m_sendTLB :: m_tlbResp :: m_sendPMP :: m_resendReq :: m_waitResendResp :: m_waitCommit :: m_commited :: Nil =
    Enum(11)
  val mmio_state = RegInit(m_idle)

  // do mmio fetch only when pmp/pbmt shows it is a uncacheable address and no exception occurs
  /* FIXME: we do not distinguish pbmt is NC or IO now
   *        but we actually can do speculative execution if pbmt is NC, maybe fix this later for performance
   */
  val f3_req_is_mmio =
    f3_valid && (f3_pmp_mmio || Pbmt.isUncache(f3_itlb_pbmt)) && !ExceptionType.hasException(f3_exception)
  val mmio_commit = VecInit(io.rob_commits.map { commit =>
    commit.valid && commit.bits.ftqIdx === f3_ftq_req.ftqIdx && commit.bits.ftqOffset === 0.U
  }).asUInt.orR
  val f3_mmio_req_commit = f3_req_is_mmio && mmio_state === m_commited

  val f3_mmio_to_commit      = f3_req_is_mmio && mmio_state === m_waitCommit
  val f3_mmio_to_commit_next = RegNext(f3_mmio_to_commit)
  val f3_mmio_can_go         = f3_mmio_to_commit && !f3_mmio_to_commit_next

  val fromFtqRedirectReg = Wire(fromFtq.redirect.cloneType)
  fromFtqRedirectReg.bits := RegEnable(
    fromFtq.redirect.bits,
    0.U.asTypeOf(fromFtq.redirect.bits),
    fromFtq.redirect.valid
  )
  fromFtqRedirectReg.valid := RegNext(fromFtq.redirect.valid, init = false.B)
  val mmioF3Flush           = RegNext(f3_flush, init = false.B)
  val f3_ftq_flush_self     = fromFtqRedirectReg.valid && RedirectLevel.flushItself(fromFtqRedirectReg.bits.level)
  val f3_ftq_flush_by_older = fromFtqRedirectReg.valid && isBefore(fromFtqRedirectReg.bits.ftqIdx, f3_ftq_req.ftqIdx)

  val f3_need_not_flush = f3_req_is_mmio && fromFtqRedirectReg.valid && !f3_ftq_flush_self && !f3_ftq_flush_by_older

  /**
    **********************************************************************************
    * We want to defer instruction fetching when encountering MMIO instructions to ensure that the MMIO region is not negatively impacted.
    * This is the exception when the first instruction is an MMIO instruction.
    **********************************************************************************
    */
  when(is_first_instr && f3_fire) {
    is_first_instr := false.B
  }

  when(f3_flush && !f3_req_is_mmio)(f3_valid := false.B)
    .elsewhen(mmioF3Flush && f3_req_is_mmio && !f3_need_not_flush)(f3_valid := false.B)
    .elsewhen(f2_fire && !f2_flush)(f3_valid := true.B)
    .elsewhen(io.toIbuffer.fire && !f3_req_is_mmio)(f3_valid := false.B)
    .elsewhen(f3_req_is_mmio && f3_mmio_req_commit)(f3_valid := false.B)

  val f3_mmio_use_seq_pc = RegInit(false.B)

  val (redirect_ftqIdx, redirect_ftqOffset) = (fromFtqRedirectReg.bits.ftqIdx, fromFtqRedirectReg.bits.ftqOffset)
  val redirect_mmio_req =
    fromFtqRedirectReg.valid && redirect_ftqIdx === f3_ftq_req.ftqIdx && redirect_ftqOffset === 0.U

  when(RegNext(f2_fire && !f2_flush) && f3_req_is_mmio)(f3_mmio_use_seq_pc := true.B)
    .elsewhen(redirect_mmio_req)(f3_mmio_use_seq_pc := false.B)

  f3_ready := (io.toIbuffer.ready && (f3_mmio_req_commit || !f3_req_is_mmio)) || !f3_valid

  // mmio state machine
  switch(mmio_state) {
    is(m_idle) {
      when(f3_req_is_mmio) {
        // in idempotent spaces, we can send request directly (i.e. can do speculative fetch)
        mmio_state := Mux(f3_itlb_pbmt === Pbmt.nc, m_sendReq, m_waitLastCmt)
      }
    }

    is(m_waitLastCmt) {
      when(is_first_instr) {
        mmio_state := m_sendReq
      }.otherwise {
        mmio_state := Mux(io.mmioCommitRead.mmioLastCommit, m_sendReq, m_waitLastCmt)
      }
    }

    is(m_sendReq) {
      mmio_state := Mux(toUncache.fire, m_waitResp, m_sendReq)
    }

    is(m_waitResp) {
      when(fromUncache.fire) {
        val isRVC      = fromUncache.bits.data(1, 0) =/= 3.U
        val needResend = !isRVC && f3_paddrs(0)(2, 1) === 3.U
        mmio_state      := Mux(needResend, m_sendTLB, m_waitCommit)
        mmio_is_RVC     := isRVC
        f3_mmio_data(0) := fromUncache.bits.data(15, 0)
        f3_mmio_data(1) := fromUncache.bits.data(31, 16)
      }
    }

    is(m_sendTLB) {
      mmio_state := Mux(io.iTLBInter.req.fire, m_tlbResp, m_sendTLB)
    }

    is(m_tlbResp) {
      when(io.iTLBInter.resp.fire) {
        // we are using a blocked tlb, so resp.fire must have !resp.bits.miss
        assert(!io.iTLBInter.resp.bits.miss, "blocked mode iTLB miss when resp.fire")
        val tlb_exception = ExceptionType.fromTlbResp(io.iTLBInter.resp.bits)
        // if itlb re-check respond pbmt mismatch with previous check, must be access fault
        val pbmt_mismatch_exception = Mux(
          io.iTLBInter.resp.bits.pbmt(0) =/= f3_itlb_pbmt,
          ExceptionType.af,
          ExceptionType.none
        )
        val exception = ExceptionType.merge(tlb_exception, pbmt_mismatch_exception)
        // if tlb has exception, abort checking pmp, just send instr & exception to ibuffer and wait for commit
        mmio_state := Mux(ExceptionType.hasException(exception), m_waitCommit, m_sendPMP)
        // also save itlb response
        mmio_resend_addr              := io.iTLBInter.resp.bits.paddr(0)
        mmio_resend_exception         := exception
        mmio_resend_gpaddr            := io.iTLBInter.resp.bits.gpaddr(0)
        mmio_resend_isForVSnonLeafPTE := io.iTLBInter.resp.bits.isForVSnonLeafPTE(0)
      }
    }

    is(m_sendPMP) {
      val pmp_exception = ExceptionType.fromPMPResp(io.pmp.resp)
      // if pmp re-check respond mismatch with previous check, must be access fault
      val mmio_mismatch_exception = Mux(
        io.pmp.resp.mmio =/= f3_pmp_mmio,
        ExceptionType.af,
        ExceptionType.none
      )
      val exception = ExceptionType.merge(pmp_exception, mmio_mismatch_exception)
      // if pmp has exception, abort sending request, just send instr & exception to ibuffer and wait for commit
      mmio_state := Mux(ExceptionType.hasException(exception), m_waitCommit, m_resendReq)
      // also save pmp response
      mmio_resend_exception := exception
    }

    is(m_resendReq) {
      mmio_state := Mux(toUncache.fire, m_waitResendResp, m_resendReq)
    }

    is(m_waitResendResp) {
      when(fromUncache.fire) {
        mmio_state      := m_waitCommit
        f3_mmio_data(1) := fromUncache.bits.data(15, 0)
      }
    }

    is(m_waitCommit) {
      // in idempotent spaces, we can skip waiting for commit (i.e. can do speculative fetch)
      // but we do not skip m_waitCommit state, as other signals (e.g. f3_mmio_can_go relies on this)
      mmio_state := Mux(mmio_commit || f3_itlb_pbmt === Pbmt.nc, m_commited, m_waitCommit)
    }

    // normal mmio instruction
    is(m_commited) {
      mmio_state                    := m_idle
      mmio_is_RVC                   := false.B
      mmio_resend_addr              := 0.U
      mmio_resend_exception         := ExceptionType.none
      mmio_resend_gpaddr            := 0.U
      mmio_resend_isForVSnonLeafPTE := false.B
    }
  }

  // Exception or flush by older branch prediction
  // Condition is from RegNext(fromFtq.redirect), 1 cycle after backend rediect
  when(f3_ftq_flush_self || f3_ftq_flush_by_older) {
    mmio_state                    := m_idle
    mmio_is_RVC                   := false.B
    mmio_resend_addr              := 0.U
    mmio_resend_exception         := ExceptionType.none
    mmio_resend_gpaddr            := 0.U
    mmio_resend_isForVSnonLeafPTE := false.B
    f3_mmio_data.map(_ := 0.U)
  }

  toUncache.valid     := ((mmio_state === m_sendReq) || (mmio_state === m_resendReq)) && f3_req_is_mmio
  toUncache.bits.addr := Mux(mmio_state === m_resendReq, mmio_resend_addr, f3_paddrs(0))
  fromUncache.ready   := true.B

  // send itlb request in m_sendTLB state
  io.iTLBInter.req.valid                   := (mmio_state === m_sendTLB) && f3_req_is_mmio
  io.iTLBInter.req.bits.size               := 3.U
  io.iTLBInter.req.bits.vaddr              := f3_resend_vaddr
  io.iTLBInter.req.bits.debug.pc           := f3_resend_vaddr
  io.iTLBInter.req.bits.cmd                := TlbCmd.exec
  io.iTLBInter.req.bits.isPrefetch         := false.B
  io.iTLBInter.req.bits.kill               := false.B // IFU use itlb for mmio, doesn't need sync, set it to false
  io.iTLBInter.req.bits.no_translate       := false.B
  io.iTLBInter.req.bits.fullva             := 0.U
  io.iTLBInter.req.bits.checkfullva        := false.B
  io.iTLBInter.req.bits.hyperinst          := DontCare
  io.iTLBInter.req.bits.hlvx               := DontCare
  io.iTLBInter.req.bits.memidx             := DontCare
  io.iTLBInter.req.bits.debug.robIdx       := DontCare
  io.iTLBInter.req.bits.debug.isFirstIssue := DontCare
  io.iTLBInter.req.bits.pmp_addr           := DontCare
  // whats the difference between req_kill and req.bits.kill?
  io.iTLBInter.req_kill := false.B
  // wait for itlb response in m_tlbResp state
  io.iTLBInter.resp.ready := (mmio_state === m_tlbResp) && f3_req_is_mmio

  io.pmp.req.valid     := (mmio_state === m_sendPMP) && f3_req_is_mmio
  io.pmp.req.bits.addr := mmio_resend_addr
  io.pmp.req.bits.size := 3.U
  io.pmp.req.bits.cmd  := TlbCmd.exec

  val f3_lastHalf = RegInit(0.U.asTypeOf(new LastHalfInfo))

  val f3_predecode_range = VecInit(preDecoderOut.pd.map(inst => inst.valid)).asUInt
  val f3_mmio_range      = VecInit((0 until PredictWidth).map(i => if (i == 0) true.B else false.B))
  val f3_instr_valid     = Wire(Vec(PredictWidth, Bool()))

  /*** prediction result check   ***/
  checkerIn.ftqOffset  := f3_ftq_req.ftqOffset
  checkerIn.jumpOffset := f3_jump_offset
  checkerIn.target     := f3_ftq_req.nextStartAddr
  checkerIn.instrRange := f3_instr_range.asTypeOf(Vec(PredictWidth, Bool()))
  checkerIn.instrValid := f3_instr_valid.asTypeOf(Vec(PredictWidth, Bool()))
  checkerIn.pds        := f3_pd
  checkerIn.pc         := f3_pc
  checkerIn.fire_in    := RegNext(f2_fire, init = false.B)

  /*** handle half RVI in the last 2 Bytes  ***/

  def hasLastHalf(idx: UInt) =
    // !f3_pd(idx).isRVC && checkerOutStage1.fixedRange(idx) && f3_instr_valid(idx) && !checkerOutStage1.fixedTaken(idx) && !checkerOutStage2.fixedMissPred(idx) && ! f3_req_is_mmio
    !f3_pd(idx).isRVC && checkerOutStage1.fixedRange(idx) && f3_instr_valid(idx) && !checkerOutStage1.fixedTaken(
      idx
    ) && !f3_req_is_mmio

  val f3_last_validIdx = ParallelPosteriorityEncoder(checkerOutStage1.fixedRange)

  val f3_hasLastHalf    = hasLastHalf((PredictWidth - 1).U)
  val f3_false_lastHalf = hasLastHalf(f3_last_validIdx)
  val f3_false_snpc     = f3_half_snpc(f3_last_validIdx)

  val f3_lastHalf_mask    = VecInit((0 until PredictWidth).map(i => if (i == 0) false.B else true.B)).asUInt
  val f3_lastHalf_disable = RegInit(false.B)

  when(f3_flush || (f3_fire && f3_lastHalf_disable)) {
    f3_lastHalf_disable := false.B
  }

  when(f3_flush) {
    f3_lastHalf.valid := false.B
  }.elsewhen(f3_fire) {
    f3_lastHalf.valid    := f3_hasLastHalf && !f3_lastHalf_disable
    f3_lastHalf.middlePC := f3_ftq_req.nextStartAddr
  }

  f3_instr_valid := Mux(f3_lastHalf.valid, f3_hasHalfValid, VecInit(f3_pd.map(inst => inst.valid)))

  /*** frontend Trigger  ***/
  frontendTrigger.io.pds  := f3_pd
  frontendTrigger.io.pc   := f3_pc
  frontendTrigger.io.data := f3_cut_data

  frontendTrigger.io.frontendTrigger := io.frontendTrigger

  val f3_triggered       = frontendTrigger.io.triggered
  val f3_toIbuffer_valid = f3_valid && (!f3_req_is_mmio || f3_mmio_can_go) && !f3_flush

  /*** send to Ibuffer  ***/
  io.toIbuffer.valid          := f3_toIbuffer_valid
  io.toIbuffer.bits.instrs    := f3_expd_instr
  io.toIbuffer.bits.valid     := f3_instr_valid.asUInt
  io.toIbuffer.bits.enqEnable := checkerOutStage1.fixedRange.asUInt & f3_instr_valid.asUInt
  io.toIbuffer.bits.pd        := f3_pd
  io.toIbuffer.bits.ftqPtr    := f3_ftq_req.ftqIdx
  io.toIbuffer.bits.pc        := f3_pc
  // Find last using PriorityMux
  io.toIbuffer.bits.isLastInFtqEntry := Reverse(PriorityEncoderOH(Reverse(io.toIbuffer.bits.enqEnable))).asBools
  io.toIbuffer.bits.ftqOffset.zipWithIndex.map { case (a, i) =>
    a.bits := i.U; a.valid := checkerOutStage1.fixedTaken(i) && !f3_req_is_mmio
  }
  io.toIbuffer.bits.foldpc        := f3_foldpc
  io.toIbuffer.bits.exceptionType := ExceptionType.merge(f3_exception_vec, f3_crossPage_exception_vec)
  // backendException only needs to be set for the first instruction.
  // Other instructions in the same block may have pf or af set,
  // which is a side effect of the first instruction and actually not necessary.
  io.toIbuffer.bits.backendException := (0 until PredictWidth).map {
    case 0 => f3_backendException
    case _ => false.B
  }
  io.toIbuffer.bits.crossPageIPFFix := f3_crossPage_exception_vec.map(ExceptionType.hasException)
  io.toIbuffer.bits.illegalInstr    := f3_ill
  io.toIbuffer.bits.triggered       := f3_triggered

  when(f3_lastHalf.valid) {
    io.toIbuffer.bits.enqEnable := checkerOutStage1.fixedRange.asUInt & f3_instr_valid.asUInt & f3_lastHalf_mask
    io.toIbuffer.bits.valid     := f3_lastHalf_mask & f3_instr_valid.asUInt
  }

  /** to backend */
  // f3_gpaddr is valid iff gpf is detected
  io.toBackend.gpaddrMem_wen := f3_toIbuffer_valid && Mux(
    f3_req_is_mmio,
    mmio_resend_exception === ExceptionType.gpf,
    f3_exception.map(_ === ExceptionType.gpf).reduce(_ || _)
  )
  io.toBackend.gpaddrMem_waddr        := f3_ftq_req.ftqIdx.value
  io.toBackend.gpaddrMem_wdata.gpaddr := Mux(f3_req_is_mmio, mmio_resend_gpaddr, f3_gpaddr)
  io.toBackend.gpaddrMem_wdata.isForVSnonLeafPTE := Mux(
    f3_req_is_mmio,
    mmio_resend_isForVSnonLeafPTE,
    f3_isForVSnonLeafPTE
  )

  // Write back to Ftq
  val f3_cache_fetch     = f3_valid && !(f2_fire && !f2_flush)
  val finishFetchMaskReg = RegNext(f3_cache_fetch)

  val mmioFlushWb        = Wire(Valid(new PredecodeWritebackBundle))
  val f3_mmio_missOffset = Wire(ValidUndirectioned(UInt(log2Ceil(PredictWidth).W)))
  f3_mmio_missOffset.valid := f3_req_is_mmio
  f3_mmio_missOffset.bits  := 0.U

  // Send mmioFlushWb back to FTQ 1 cycle after uncache fetch return
  // When backend redirect, mmio_state reset after 1 cycle.
  // In this case, mask .valid to avoid overriding backend redirect
  mmioFlushWb.valid := (f3_req_is_mmio && mmio_state === m_waitCommit && RegNext(fromUncache.fire) &&
    f3_mmio_use_seq_pc && !f3_ftq_flush_self && !f3_ftq_flush_by_older)
  mmioFlushWb.bits.pc := f3_pc
  mmioFlushWb.bits.pd := f3_pd
  mmioFlushWb.bits.pd.zipWithIndex.map { case (instr, i) => instr.valid := f3_mmio_range(i) }
  mmioFlushWb.bits.ftqIdx     := f3_ftq_req.ftqIdx
  mmioFlushWb.bits.ftqOffset  := f3_ftq_req.ftqOffset.bits
  mmioFlushWb.bits.misOffset  := f3_mmio_missOffset
  mmioFlushWb.bits.cfiOffset  := DontCare
  mmioFlushWb.bits.target     := Mux(mmio_is_RVC, f3_ftq_req.startAddr + 2.U, f3_ftq_req.startAddr + 4.U)
  mmioFlushWb.bits.jalTarget  := DontCare
  mmioFlushWb.bits.instrRange := f3_mmio_range

  val mmioRVCExpander = Module(new RVCExpander)
  mmioRVCExpander.io.in      := Mux(f3_req_is_mmio, Cat(f3_mmio_data(1), f3_mmio_data(0)), 0.U)
  mmioRVCExpander.io.fsIsOff := io.csr_fsIsOff

  /** external predecode for MMIO instruction */
  when(f3_req_is_mmio) {
    val inst         = Cat(f3_mmio_data(1), f3_mmio_data(0))
    val currentIsRVC = isRVC(inst)

    val brType :: isCall :: isRet :: Nil = brInfo(inst)
    val jalOffset                        = jal_offset(inst, currentIsRVC)
    val brOffset                         = br_offset(inst, currentIsRVC)

    io.toIbuffer.bits.instrs(0) := Mux(mmioRVCExpander.io.ill, mmioRVCExpander.io.in, mmioRVCExpander.io.out.bits)

    io.toIbuffer.bits.pd(0).valid  := true.B
    io.toIbuffer.bits.pd(0).isRVC  := currentIsRVC
    io.toIbuffer.bits.pd(0).brType := brType
    io.toIbuffer.bits.pd(0).isCall := isCall
    io.toIbuffer.bits.pd(0).isRet  := isRet

    io.toIbuffer.bits.exceptionType(0)   := mmio_resend_exception
    io.toIbuffer.bits.crossPageIPFFix(0) := ExceptionType.hasException(mmio_resend_exception)
    io.toIbuffer.bits.illegalInstr(0)    := mmioRVCExpander.io.ill

    io.toIbuffer.bits.enqEnable := f3_mmio_range.asUInt

    mmioFlushWb.bits.pd(0).valid  := true.B
    mmioFlushWb.bits.pd(0).isRVC  := currentIsRVC
    mmioFlushWb.bits.pd(0).brType := brType
    mmioFlushWb.bits.pd(0).isCall := isCall
    mmioFlushWb.bits.pd(0).isRet  := isRet
  }

  mmio_redirect := (f3_req_is_mmio && mmio_state === m_waitCommit && RegNext(fromUncache.fire) && f3_mmio_use_seq_pc)

  XSPerfAccumulate("fetch_bubble_ibuffer_not_ready", io.toIbuffer.valid && !io.toIbuffer.ready)

  /**
    ******************************************************************************
    * IFU Write Back Stage
    * - write back predecode information to Ftq to update
    * - redirect if found fault prediction
    * - redirect if has false hit last half (last PC is not start + 32 Bytes, but in the midle of an notCFI RVI instruction)
    ******************************************************************************
    */
  val wb_enable  = RegNext(f2_fire && !f2_flush) && !f3_req_is_mmio && !f3_flush
  val wb_valid   = RegNext(wb_enable, init = false.B)
  val wb_ftq_req = RegEnable(f3_ftq_req, wb_enable)

  val wb_check_result_stage1 = RegEnable(checkerOutStage1, wb_enable)
  val wb_check_result_stage2 = checkerOutStage2
  val wb_instr_range         = RegEnable(io.toIbuffer.bits.enqEnable, wb_enable)

  val wb_pc_lower_result = RegEnable(f3_pc_lower_result, wb_enable)
  val wb_pc_high         = RegEnable(f3_pc_high, wb_enable)
  val wb_pc_high_plus1   = RegEnable(f3_pc_high_plus1, wb_enable)
  val wb_pc              = CatPC(wb_pc_lower_result, wb_pc_high, wb_pc_high_plus1)

  // val wb_pc             = RegEnable(f3_pc, wb_enable)
  val wb_pd          = RegEnable(f3_pd, wb_enable)
  val wb_instr_valid = RegEnable(f3_instr_valid, wb_enable)

  /* false hit lastHalf */
  val wb_lastIdx        = RegEnable(f3_last_validIdx, wb_enable)
  val wb_false_lastHalf = RegEnable(f3_false_lastHalf, wb_enable) && wb_lastIdx =/= (PredictWidth - 1).U
  val wb_false_target   = RegEnable(f3_false_snpc, wb_enable)

  val wb_half_flush  = wb_false_lastHalf
  val wb_half_target = wb_false_target

  /* false oversize */
  val lastIsRVC = wb_instr_range.asTypeOf(Vec(PredictWidth, Bool())).last && wb_pd.last.isRVC
  val lastIsRVI = wb_instr_range.asTypeOf(Vec(PredictWidth, Bool()))(PredictWidth - 2) && !wb_pd(PredictWidth - 2).isRVC
  val lastTaken = wb_check_result_stage1.fixedTaken.last

  f3_wb_not_flush := wb_ftq_req.ftqIdx === f3_ftq_req.ftqIdx && f3_valid && wb_valid

  /** if a req with a last half but miss predicted enters in wb stage, and this cycle f3 stalls,
    * we set a flag to notify f3 that the last half flag need not to be set.
    */
  // f3_fire is after wb_valid
  when(wb_valid && RegNext(f3_hasLastHalf, init = false.B)
    && wb_check_result_stage2.fixedMissPred(PredictWidth - 1) && !f3_fire && !RegNext(
      f3_fire,
      init = false.B
    ) && !f3_flush) {
    f3_lastHalf_disable := true.B
  }

  // wb_valid and f3_fire are in same cycle
  when(wb_valid && RegNext(f3_hasLastHalf, init = false.B)
    && wb_check_result_stage2.fixedMissPred(PredictWidth - 1) && f3_fire) {
    f3_lastHalf.valid := false.B
  }

  val checkFlushWb = Wire(Valid(new PredecodeWritebackBundle))
  val checkFlushWbjalTargetIdx = ParallelPriorityEncoder(VecInit(wb_pd.zip(wb_instr_valid).map { case (pd, v) =>
    v && pd.isJal
  }))
  val checkFlushWbTargetIdx = ParallelPriorityEncoder(wb_check_result_stage2.fixedMissPred)
  checkFlushWb.valid   := wb_valid
  checkFlushWb.bits.pc := wb_pc
  checkFlushWb.bits.pd := wb_pd
  checkFlushWb.bits.pd.zipWithIndex.map { case (instr, i) => instr.valid := wb_instr_valid(i) }
  checkFlushWb.bits.ftqIdx          := wb_ftq_req.ftqIdx
  checkFlushWb.bits.ftqOffset       := wb_ftq_req.ftqOffset.bits
  checkFlushWb.bits.misOffset.valid := ParallelOR(wb_check_result_stage2.fixedMissPred) || wb_half_flush
  checkFlushWb.bits.misOffset.bits := Mux(
    wb_half_flush,
    wb_lastIdx,
    ParallelPriorityEncoder(wb_check_result_stage2.fixedMissPred)
  )
  checkFlushWb.bits.cfiOffset.valid := ParallelOR(wb_check_result_stage1.fixedTaken)
  checkFlushWb.bits.cfiOffset.bits  := ParallelPriorityEncoder(wb_check_result_stage1.fixedTaken)
  checkFlushWb.bits.target := Mux(
    wb_half_flush,
    wb_half_target,
    wb_check_result_stage2.fixedTarget(checkFlushWbTargetIdx)
  )
  checkFlushWb.bits.jalTarget  := wb_check_result_stage2.jalTarget(checkFlushWbjalTargetIdx)
  checkFlushWb.bits.instrRange := wb_instr_range.asTypeOf(Vec(PredictWidth, Bool()))

  toFtq.pdWb := Mux(wb_valid, checkFlushWb, mmioFlushWb)

  wb_redirect := checkFlushWb.bits.misOffset.valid && wb_valid

  /*write back flush type*/
  val checkFaultType    = wb_check_result_stage2.faultType
  val checkJalFault     = wb_valid && checkFaultType.map(_.isjalFault).reduce(_ || _)
  val checkRetFault     = wb_valid && checkFaultType.map(_.isRetFault).reduce(_ || _)
  val checkTargetFault  = wb_valid && checkFaultType.map(_.istargetFault).reduce(_ || _)
  val checkNotCFIFault  = wb_valid && checkFaultType.map(_.notCFIFault).reduce(_ || _)
  val checkInvalidTaken = wb_valid && checkFaultType.map(_.invalidTakenFault).reduce(_ || _)

  XSPerfAccumulate("predecode_flush_jalFault", checkJalFault)
  XSPerfAccumulate("predecode_flush_retFault", checkRetFault)
  XSPerfAccumulate("predecode_flush_targetFault", checkTargetFault)
  XSPerfAccumulate("predecode_flush_notCFIFault", checkNotCFIFault)
  XSPerfAccumulate("predecode_flush_incalidTakenFault", checkInvalidTaken)

  XSDebug(
    checkRetFault,
    "startAddr:%x  nextstartAddr:%x  taken:%d    takenIdx:%d\n",
    wb_ftq_req.startAddr,
    wb_ftq_req.nextStartAddr,
    wb_ftq_req.ftqOffset.valid,
    wb_ftq_req.ftqOffset.bits
  )

  /** performance counter */
  val f3_perf_info = RegEnable(f2_perf_info, f2_fire)
  val f3_req_0     = io.toIbuffer.fire
  val f3_req_1     = io.toIbuffer.fire && f3_doubleLine
  val f3_hit_0     = io.toIbuffer.fire && f3_perf_info.bank_hit(0)
  val f3_hit_1     = io.toIbuffer.fire && f3_doubleLine & f3_perf_info.bank_hit(1)
  val f3_hit       = f3_perf_info.hit
  val perfEvents = Seq(
    ("frontendFlush                ", wb_redirect),
    ("ifu_req                      ", io.toIbuffer.fire),
    ("ifu_miss                     ", io.toIbuffer.fire && !f3_perf_info.hit),
    ("ifu_req_cacheline_0          ", f3_req_0),
    ("ifu_req_cacheline_1          ", f3_req_1),
    ("ifu_req_cacheline_0_hit      ", f3_hit_1),
    ("ifu_req_cacheline_1_hit      ", f3_hit_1),
    ("only_0_hit                   ", f3_perf_info.only_0_hit && io.toIbuffer.fire),
    ("only_0_miss                  ", f3_perf_info.only_0_miss && io.toIbuffer.fire),
    ("hit_0_hit_1                  ", f3_perf_info.hit_0_hit_1 && io.toIbuffer.fire),
    ("hit_0_miss_1                 ", f3_perf_info.hit_0_miss_1 && io.toIbuffer.fire),
    ("miss_0_hit_1                 ", f3_perf_info.miss_0_hit_1 && io.toIbuffer.fire),
    ("miss_0_miss_1                ", f3_perf_info.miss_0_miss_1 && io.toIbuffer.fire)
  )
  generatePerfEvent()

  XSPerfAccumulate("ifu_req", io.toIbuffer.fire)
  XSPerfAccumulate("ifu_miss", io.toIbuffer.fire && !f3_hit)
  XSPerfAccumulate("ifu_req_cacheline_0", f3_req_0)
  XSPerfAccumulate("ifu_req_cacheline_1", f3_req_1)
  XSPerfAccumulate("ifu_req_cacheline_0_hit", f3_hit_0)
  XSPerfAccumulate("ifu_req_cacheline_1_hit", f3_hit_1)
  XSPerfAccumulate("frontendFlush", wb_redirect)
  XSPerfAccumulate("only_0_hit", f3_perf_info.only_0_hit && io.toIbuffer.fire)
  XSPerfAccumulate("only_0_miss", f3_perf_info.only_0_miss && io.toIbuffer.fire)
  XSPerfAccumulate("hit_0_hit_1", f3_perf_info.hit_0_hit_1 && io.toIbuffer.fire)
  XSPerfAccumulate("hit_0_miss_1", f3_perf_info.hit_0_miss_1 && io.toIbuffer.fire)
  XSPerfAccumulate("miss_0_hit_1", f3_perf_info.miss_0_hit_1 && io.toIbuffer.fire)
  XSPerfAccumulate("miss_0_miss_1", f3_perf_info.miss_0_miss_1 && io.toIbuffer.fire)
  XSPerfAccumulate("hit_0_except_1", f3_perf_info.hit_0_except_1 && io.toIbuffer.fire)
  XSPerfAccumulate("miss_0_except_1", f3_perf_info.miss_0_except_1 && io.toIbuffer.fire)
  XSPerfAccumulate("except_0", f3_perf_info.except_0 && io.toIbuffer.fire)
  XSPerfHistogram(
    "ifu2ibuffer_validCnt",
    PopCount(io.toIbuffer.bits.valid & io.toIbuffer.bits.enqEnable),
    io.toIbuffer.fire,
    0,
    PredictWidth + 1,
    1
  )

  val hartId                     = p(XSCoreParamsKey).HartId
  val isWriteFetchToIBufferTable = Constantin.createRecord(s"isWriteFetchToIBufferTable$hartId")
  val isWriteIfuWbToFtqTable     = Constantin.createRecord(s"isWriteIfuWbToFtqTable$hartId")
  val fetchToIBufferTable        = ChiselDB.createTable(s"FetchToIBuffer$hartId", new FetchToIBufferDB)
  val ifuWbToFtqTable            = ChiselDB.createTable(s"IfuWbToFtq$hartId", new IfuWbToFtqDB)

  val fetchIBufferDumpData = Wire(new FetchToIBufferDB)
  fetchIBufferDumpData.start_addr  := f3_ftq_req.startAddr
  fetchIBufferDumpData.instr_count := PopCount(io.toIbuffer.bits.enqEnable)
  fetchIBufferDumpData.exception := (f3_perf_info.except_0 && io.toIbuffer.fire) || (f3_perf_info.hit_0_except_1 && io.toIbuffer.fire) || (f3_perf_info.miss_0_except_1 && io.toIbuffer.fire)
  fetchIBufferDumpData.is_cache_hit := f3_hit

  val ifuWbToFtqDumpData = Wire(new IfuWbToFtqDB)
  ifuWbToFtqDumpData.start_addr        := wb_ftq_req.startAddr
  ifuWbToFtqDumpData.is_miss_pred      := checkFlushWb.bits.misOffset.valid
  ifuWbToFtqDumpData.miss_pred_offset  := checkFlushWb.bits.misOffset.bits
  ifuWbToFtqDumpData.checkJalFault     := checkJalFault
  ifuWbToFtqDumpData.checkRetFault     := checkRetFault
  ifuWbToFtqDumpData.checkTargetFault  := checkTargetFault
  ifuWbToFtqDumpData.checkNotCFIFault  := checkNotCFIFault
  ifuWbToFtqDumpData.checkInvalidTaken := checkInvalidTaken

  fetchToIBufferTable.log(
    data = fetchIBufferDumpData,
    en = isWriteFetchToIBufferTable.orR && io.toIbuffer.fire,
    site = "IFU" + p(XSCoreParamsKey).HartId.toString,
    clock = clock,
    reset = reset
  )
  ifuWbToFtqTable.log(
    data = ifuWbToFtqDumpData,
    en = isWriteIfuWbToFtqTable.orR && checkFlushWb.valid,
    site = "IFU" + p(XSCoreParamsKey).HartId.toString,
    clock = clock,
    reset = reset
  )

}