backend,mem: add Store Sets memory dependence predictor (#796)

* LoadQueue: send stFtqIdx via rollback request

* It will make it possible for setore set to update its SSIT

* StoreSet: setup store set update req

* StoreSet: add store set identifier table (SSIT)

* StoreSet: add last fetched store table (LFST)

* StoreSet: put SSIT into decode stage

* StoreSet: put LFST into dispatch1

* Future work: optimize timing

* RS: store rs now supports delayed issue

* StoreSet: add perf counter

* StoreSet: fix SSIT update logic

* StoreSet: delay LFST update input for 1 cycle

* StoreSet: fix LFST update logic

* StoreSet: fix LFST raddr width

* StoreSet: do not force store in ss issue in order

Classic store set requires store in the same store set issue in seq.
However, in current micro-architecture, such restrict will lead to
severe perf lost. We choose to disable it until we find another way
to fix it.

* StoreSet: support ooo store in the same store set

* StoreSet: fix store set merge logic

* StoreSet: check earlier store when read LFST

* If store-load pair is in the same dispatch bundle, loadWaitBit should
also be set for load

* StoreSet: increase default SSIT flush period

* StoreSet: fix LFST read logic

* Fix commit c0e541d1

* StoreSet: add StoreSetEnable parameter

* RSFeedback: add source type

* StoreQueue: split store addr and store data

* StoreQueue: update ls forward logic

* Now it supports splited addr and data

* Chore: force assign name for load/store unit

* RS: add rs'support for store a-d split

* StoreQueue: fix stlf logic

* StoreQueue: fix addr wb sq update logic

* AtomicsUnit: support splited a/d

* Parameters: disable store set by default

* WaitTable: wait table will not cause store delay

* WaitTable: recover default reset period to 2^17

* Fix dev-stad merge conflict

* StoreSet: enable storeset

* RS: disable store rs delay logic

CI perf shows that current delay logic will cause perf loss. Disable
unnecessary delay logic will help.

To be more specific, `io.readyVec` caused the problem. It will be
updated in future commits.

* RS: opt select logic with load delay (ldWait)

* StoreSet: disable 2-bit lwt
Co-authored-by: NZhangZifei <zhangzifei20z@ict.ac.cn>

src/main/scala/xiangshan/Bundle.scala

@@ -3,7 +3,7 @@ package xiangshan
 import chisel3._
 import chisel3.util._
 import xiangshan.backend.roq.RoqPtr
-import xiangshan.backend.decode.{ImmUnion, WaitTableParameters, XDecode}
+import xiangshan.backend.decode.{ImmUnion, XDecode}
 import xiangshan.mem.{LqPtr, SqPtr}
 import xiangshan.frontend.PreDecodeInfoForDebug
 import xiangshan.frontend.PreDecodeInfo
@@ -23,13 +23,13 @@ import chipsalliance.rocketchip.config.Parameters
 import xiangshan.backend.ftq.FtqPtr
 
 // Fetch FetchWidth x 32-bit insts from Icache
-class FetchPacket(implicit p: Parameters) extends XSBundle with WaitTableParameters {
+class FetchPacket(implicit p: Parameters) extends XSBundle {
   val instrs = Vec(PredictWidth, UInt(32.W))
   val mask = UInt(PredictWidth.W)
   val pdmask = UInt(PredictWidth.W)
   // val pc = UInt(VAddrBits.W)
   val pc = Vec(PredictWidth, UInt(VAddrBits.W))
-  val foldpc = Vec(PredictWidth, UInt(WaitTableAddrWidth.W))
+  val foldpc = Vec(PredictWidth, UInt(MemPredPCWidth.W))
   val pd = Vec(PredictWidth, new PreDecodeInfo)
   val ipf = Bool()
   val acf = Bool()
@@ -179,16 +179,18 @@ class CfiUpdateInfo(implicit p: Parameters) extends XSBundle with HasBPUParamete
 }
 
 // Dequeue DecodeWidth insts from Ibuffer
-class CtrlFlow(implicit p: Parameters) extends XSBundle with WaitTableParameters {
+class CtrlFlow(implicit p: Parameters) extends XSBundle {
   val instr = UInt(32.W)
   val pc = UInt(VAddrBits.W)
-  val foldpc = UInt(WaitTableAddrWidth.W)
+  val foldpc = UInt(MemPredPCWidth.W)
   val exceptionVec = ExceptionVec()
   val intrVec = Vec(12, Bool())
   val pd = new PreDecodeInfo
   val pred_taken = Bool()
   val crossPageIPFFix = Bool()
+  val storeSetHit = Bool() // inst has been allocated an store set
   val loadWaitBit = Bool() // load inst should not be executed until all former store addr calcuated
+  val ssid = UInt(SSIDWidth.W)
   val ftqPtr = new FtqPtr
   val ftqOffset = UInt(log2Up(PredictWidth).W)
 }
@@ -325,6 +327,11 @@ class MicroOp(implicit p: Parameters) extends CfCtrl {
   }
 }
 
+class MicroOpRbExt(implicit p: Parameters) extends XSBundle {
+  val uop = new MicroOp
+  val flag = UInt(1.W)
+}
+
 class Redirect(implicit p: Parameters) extends XSBundle {
   val roqIdx = new RoqPtr
   val ftqIdx = new FtqPtr
@@ -333,6 +340,8 @@ class Redirect(implicit p: Parameters) extends XSBundle {
   val interrupt = Bool()
   val cfiUpdate = new CfiUpdateInfo
 
+  val stFtqIdx = new FtqPtr // for load violation predict
+  val stFtqOffset = UInt(log2Up(PredictWidth).W)
 
   // def isUnconditional() = RedirectLevel.isUnconditional(level)
   def flushItself() = RedirectLevel.flushItself(level)
@@ -467,10 +476,18 @@ class SfenceBundle(implicit p: Parameters) extends XSBundle {
   }
 }
 
-class WaitTableUpdateReq(implicit p: Parameters) extends XSBundle with WaitTableParameters {
+// Bundle for load violation predictor updating
+class MemPredUpdateReq(implicit p: Parameters) extends XSBundle  {
   val valid = Bool()
-  val waddr = UInt(WaitTableAddrWidth.W)
+
+  // wait table update
+  val waddr = UInt(MemPredPCWidth.W)
   val wdata = Bool() // true.B by default
+
+  // store set update
+  // by default, ldpc/stpc should be xor folded
+  val ldpc = UInt(MemPredPCWidth.W)
+  val stpc = UInt(MemPredPCWidth.W)
 }
 
 class PerfInfoIO extends Bundle {

src/main/scala/xiangshan/Parameters.scala

@@ -285,7 +285,21 @@ trait HasXSParameter {
       blockBytes = L2BlockSize,
       nEntries = dcacheParameters.nMissEntries * 2 // TODO: this is too large
     ),
-  )
+  )  
+  
+  // load violation predict
+  val ResetTimeMax2Pow = 20 //1078576
+  val ResetTimeMin2Pow = 10 //1024
+  // wait table parameters
+  val WaitTableSize = 1024
+  val MemPredPCWidth = log2Up(WaitTableSize)
+  val LWTUse2BitCounter = true
+  // store set parameters
+  val SSITSize = WaitTableSize
+  val LFSTSize = 32
+  val SSIDWidth = log2Up(LFSTSize)
+  val LFSTWidth = 4
+  val StoreSetEnable = true // LWT will be disabled if SS is enabled
 
   val loadExuConfigs = coreParams.loadExuConfigs
   val storeExuConfigs = coreParams.storeExuConfigs

src/main/scala/xiangshan/backend/CtrlBlock.scala

@@ -5,7 +5,7 @@ import chisel3._
 import chisel3.util._
 import utils._
 import xiangshan._
-import xiangshan.backend.decode.{DecodeStage, ImmUnion, WaitTableParameters}
+import xiangshan.backend.decode.{DecodeStage, ImmUnion}
 import xiangshan.backend.rename.{BusyTable, Rename}
 import xiangshan.backend.dispatch.Dispatch
 import xiangshan.backend.exu._
@@ -38,13 +38,13 @@ class CtrlToFpBlockIO(implicit p: Parameters) extends XSBundle {
 class CtrlToLsBlockIO(implicit p: Parameters) extends XSBundle {
   val enqIqCtrl = Vec(exuParameters.LsExuCnt, DecoupledIO(new MicroOp))
   val enqLsq = Flipped(new LsqEnqIO)
-  val waitTableUpdate = Vec(StorePipelineWidth, Input(new WaitTableUpdateReq))
+  val memPredUpdate = Vec(StorePipelineWidth, Input(new MemPredUpdateReq))
   val redirect = ValidIO(new Redirect)
   val flush = Output(Bool())
 }
 
 class RedirectGenerator(implicit p: Parameters) extends XSModule
-  with HasCircularQueuePtrHelper with WaitTableParameters with HasFtqHelper {
+  with HasCircularQueuePtrHelper with HasFtqHelper {
   val numRedirect = exuParameters.JmpCnt + exuParameters.AluCnt
   val io = IO(new Bundle() {
     val exuMispredict = Vec(numRedirect, Flipped(ValidIO(new ExuOutput)))
@@ -54,7 +54,8 @@ class RedirectGenerator(implicit p: Parameters) extends XSModule
     val stage2FtqRead = new FtqRead
     val stage2Redirect = ValidIO(new Redirect)
     val stage3Redirect = ValidIO(new Redirect)
-    val waitTableUpdate = Output(new WaitTableUpdateReq)
+    val memPredUpdate = Output(new MemPredUpdateReq)
+    val memPredFtqRead = new FtqRead // read req send form stage 2
   })
   /*
         LoadQueue  Jump  ALU0  ALU1  ALU2  ALU3   exception    Stage1
@@ -143,12 +144,26 @@ class RedirectGenerator(implicit p: Parameters) extends XSModule
     )
   )
 
-  // update waittable if load violation redirect triggered
-  io.waitTableUpdate.valid := RegNext(s1_isReplay && s1_redirect_valid_reg, init = false.B)
-  io.waitTableUpdate.waddr := RegNext(XORFold(real_pc(VAddrBits-1, 1), WaitTableAddrWidth))
-  io.waitTableUpdate.wdata := true.B
+  // get pc from ftq
+  io.memPredFtqRead.ptr := s1_redirect_bits_reg.stFtqIdx
+  // valid only if redirect is caused by load violation
+  // store_pc is used to update store set
+  val memPredFtqRead = io.memPredFtqRead.entry
+  val store_pc = GetPcByFtq(memPredFtqRead.ftqPC, RegNext(s1_redirect_bits_reg).stFtqOffset,
+    memPredFtqRead.lastPacketPC.valid,
+    memPredFtqRead.lastPacketPC.bits
+  )
+
+  // update load violation predictor if load violation redirect triggered
+  io.memPredUpdate.valid := RegNext(s1_isReplay && s1_redirect_valid_reg, init = false.B)
+  // update wait table
+  io.memPredUpdate.waddr := RegNext(XORFold(real_pc(VAddrBits-1, 1), MemPredPCWidth))
+  io.memPredUpdate.wdata := true.B
+  // update store set
+  io.memPredUpdate.ldpc := RegNext(XORFold(real_pc(VAddrBits-1, 1), MemPredPCWidth))
+  // store pc is ready 1 cycle after s1_isReplay is judged
+  io.memPredUpdate.stpc := XORFold(store_pc(VAddrBits-1, 1), MemPredPCWidth)
 
-  io.stage2FtqRead.ptr := s1_redirect_bits_reg.ftqIdx
 
   val s2_br_mask = RegEnable(ftqRead.br_mask, enable = s1_redirect_valid_reg)
   val s2_sawNotTakenBranch = RegEnable(VecInit((0 until PredictWidth).map{ i =>
@@ -242,7 +257,8 @@ class CtrlBlock(implicit p: Parameters) extends XSModule
     init = false.B
   )
   loadReplay.bits := RegEnable(io.fromLsBlock.replay.bits, io.fromLsBlock.replay.valid)
-  VecInit(ftq.io.ftqRead.tail.dropRight(1)) <> redirectGen.io.stage1FtqRead
+  VecInit(ftq.io.ftqRead.tail.dropRight(2)) <> redirectGen.io.stage1FtqRead
+  ftq.io.ftqRead.dropRight(1).last <> redirectGen.io.memPredFtqRead
   ftq.io.cfiRead <> redirectGen.io.stage2FtqRead
   redirectGen.io.exuMispredict <> exuRedirect
   redirectGen.io.loadReplay <> loadReplay
@@ -288,10 +304,10 @@ class CtrlBlock(implicit p: Parameters) extends XSModule
 
   decode.io.in <> io.frontend.cfVec
   // currently, we only update wait table when isReplay
-  decode.io.waitTableUpdate(0) <> RegNext(redirectGen.io.waitTableUpdate)
-  decode.io.waitTableUpdate(1) := DontCare
-  decode.io.waitTableUpdate(1).valid := false.B
-  // decode.io.waitTableUpdate <> io.toLsBlock.waitTableUpdate
+  decode.io.memPredUpdate(0) <> RegNext(redirectGen.io.memPredUpdate)
+  decode.io.memPredUpdate(1) := DontCare
+  decode.io.memPredUpdate(1).valid := false.B
+  // decode.io.memPredUpdate <> io.toLsBlock.memPredUpdate
   decode.io.csrCtrl := RegNext(io.csrCtrl)
 
 
@@ -335,6 +351,8 @@ class CtrlBlock(implicit p: Parameters) extends XSModule
   dispatch.io.numExist <> io.fromIntBlock.numExist ++ io.fromFpBlock.numExist ++ io.fromLsBlock.numExist
   dispatch.io.enqIQCtrl <> io.toIntBlock.enqIqCtrl ++ io.toFpBlock.enqIqCtrl ++ io.toLsBlock.enqIqCtrl
 //  dispatch.io.enqIQData <> io.toIntBlock.enqIqData ++ io.toFpBlock.enqIqData ++ io.toLsBlock.enqIqData
+  dispatch.io.csrCtrl <> io.csrCtrl
+  dispatch.io.storeIssue <> io.fromLsBlock.stIn
 
 
   fpBusyTable.io.flush := flushReg

src/main/scala/xiangshan/backend/MemBlock.scala

@@ -17,6 +17,7 @@ import xiangshan.backend.regfile.RfReadPort
 import utils._
 
 class LsBlockToCtrlIO(implicit p: Parameters) extends XSBundle {
+  val stIn = Vec(exuParameters.StuCnt, ValidIO(new ExuInput))
   val stOut = Vec(exuParameters.StuCnt, ValidIO(new ExuOutput))
   val numExist = Vec(exuParameters.LsExuCnt, Output(UInt(log2Ceil(IssQueSize).W)))
   val replay = ValidIO(new Redirect)
@@ -246,7 +247,7 @@ class MemBlockImp(outer: MemBlock) extends LazyModuleImp(outer)
 
     // update waittable
     // TODO: read pc
-    io.fromCtrlBlock.waitTableUpdate(i) := DontCare
+    io.fromCtrlBlock.memPredUpdate(i) := DontCare
     lsq.io.needReplayFromRS(i)    <> loadUnits(i).io.lsq.needReplayFromRS
   }
 
@@ -265,6 +266,8 @@ class MemBlockImp(outer: MemBlock) extends LazyModuleImp(outer)
     stu.io.stin        <> rs.io.deq
     stu.io.lsq         <> lsq.io.storeIn(i)
 
+    // Lsq to load unit's rs
+    rs.io.stIssuePtr := lsq.io.issuePtrExt
     // rs.io.storeData <> lsq.io.storeDataIn(i)
     lsq.io.storeDataIn(i) := rs.io.stData
 
@@ -272,6 +275,10 @@ class MemBlockImp(outer: MemBlock) extends LazyModuleImp(outer)
     lsq.io.storeIssue(i).valid := rs.io.deq.valid
     lsq.io.storeIssue(i).bits := rs.io.deq.bits
 
+    // sync issue info to store set LFST
+    io.toCtrlBlock.stIn(i).valid := rs.io.deq.valid
+    io.toCtrlBlock.stIn(i).bits := rs.io.deq.bits
+
     io.toCtrlBlock.stOut(i).valid := stu.io.stout.valid
     io.toCtrlBlock.stOut(i).bits  := stu.io.stout.bits
     stu.io.stout.ready := true.B

src/main/scala/xiangshan/backend/decode/DecodeStage.scala

@@ -11,7 +11,7 @@ class DecodeStage(implicit p: Parameters) extends XSModule {
     // from Ibuffer
     val in = Vec(DecodeWidth, Flipped(DecoupledIO(new CtrlFlow)))
     // from memblock
-    val waitTableUpdate = Vec(StorePipelineWidth, Input(new WaitTableUpdateReq))
+    val memPredUpdate = Vec(StorePipelineWidth, Input(new MemPredUpdateReq))
     // to DecBuffer
     val out = Vec(DecodeWidth, DecoupledIO(new CfCtrl))
     // waitable ctrl
@@ -19,7 +19,12 @@ class DecodeStage(implicit p: Parameters) extends XSModule {
   })
 
   val decoders = Seq.fill(DecodeWidth)(Module(new DecodeUnit))
+
+  // basic wait table load violation predictor (for debug only)
   val waittable = Module(new WaitTable)
+  // store set load violation predictor stage 1: SSIT look up
+  val ssit = Module(new SSIT)
+
   for (i <- 0 until DecodeWidth) {
     decoders(i).io.enq.ctrl_flow <> io.in(i).bits
 
@@ -27,18 +32,26 @@ class DecodeStage(implicit p: Parameters) extends XSModule {
     waittable.io.raddr(i) := io.in(i).bits.foldpc
     decoders(i).io.enq.ctrl_flow.loadWaitBit := waittable.io.rdata(i)
 
+    // read SSIT, get SSID
+    ssit.io.raddr(i) := io.in(i).bits.foldpc
+    decoders(i).io.enq.ctrl_flow.storeSetHit := ssit.io.rdata(i).valid
+    decoders(i).io.enq.ctrl_flow.ssid := ssit.io.rdata(i).ssid
+
     io.out(i).valid      := io.in(i).valid
     io.out(i).bits       := decoders(i).io.deq.cf_ctrl
     io.in(i).ready       := io.out(i).ready
   }
-
   for (i <- 0 until StorePipelineWidth) {
-    waittable.io.update(i) <> RegNext(io.waitTableUpdate(i))
+    waittable.io.update(i) <> RegNext(io.memPredUpdate(i))
   }
   waittable.io.csrCtrl <> io.csrCtrl
+  ssit.io.update <> RegNext(io.memPredUpdate(0))
+  ssit.io.csrCtrl <> io.csrCtrl
 
   val loadWaitBitSet = PopCount(io.out.map(o => o.fire() && o.bits.cf.loadWaitBit))
   XSPerfAccumulate("loadWaitBitSet", loadWaitBitSet)
+  val storeSetHit = PopCount(io.out.map(o => o.fire() && o.bits.cf.storeSetHit))
+  XSPerfAccumulate("storeset_ssit_hit", storeSetHit)
 
   val hasValid = VecInit(io.in.map(_.valid)).asUInt.orR
   XSPerfAccumulate("utilization", PopCount(io.in.map(_.valid)))

src/main/scala/xiangshan/backend/decode/StoreSet.scala

+package xiangshan.backend.decode
+
+import chipsalliance.rocketchip.config.Parameters
+import chisel3._
+import chisel3.util._
+import xiangshan._
+import utils._
+import xiangshan.mem.{LqPtr, SqPtr}
+import xiangshan.backend.roq.RoqPtr
+
+// store set load violation predictor
+// See "Memory Dependence Prediction using Store Sets" for details
+
+// Store Set Identifier Table Entry
+class SSITEntry(implicit p: Parameters) extends XSBundle {
+  val valid = Bool()
+  val isload = Bool()
+  val ssid = UInt(SSIDWidth.W) // store set identifier
+}
+
+// Store Set Identifier Table
+class SSIT(implicit p: Parameters) extends XSModule {
+  val io = IO(new Bundle {
+    val raddr = Vec(DecodeWidth, Input(UInt(MemPredPCWidth.W))) // xor hashed decode pc(VaddrBits-1, 1)
+    val rdata = Vec(DecodeWidth, Output(new SSITEntry))
+    val update = Input(new MemPredUpdateReq) // RegNext should be added outside
+    val csrCtrl = Input(new CustomCSRCtrlIO)
+  })
+
+  // TODO: use MemTemplate
+  val valid = RegInit(VecInit(Seq.fill(SSITSize)(false.B)))
+  val isload = Reg(Vec(SSITSize, Bool()))
+  val ssid = Reg(Vec(SSITSize, UInt(SSIDWidth.W)))
+
+  val resetCounter = RegInit(0.U(ResetTimeMax2Pow.W))
+  resetCounter := resetCounter + 1.U
+
+  // read SSIT in decode stage
+  for (i <- 0 until DecodeWidth) {
+    // io.rdata(i) := (data(io.raddr(i))(1) || io.csrCtrl.no_spec_load) && !io.csrCtrl.lvpred_disable
+    io.rdata(i).valid := valid(io.raddr(i))
+    io.rdata(i).isload := isload(io.raddr(i))
+    io.rdata(i).ssid := ssid(io.raddr(i))
+  }
+
+  // update SSIT if load violation redirect is detected
+
+  // update stage -1
+  // when io.update.valid, we should RegNext() it for at least 1 cycle
+  // outside of SSIT.
+
+  // update stage 0
+  // RegNext(io.update) while reading SSIT entry for necessary information
+  val memPredUpdateReqValid = RegNext(io.update.valid)
+  val memPredUpdateReqReg = RegEnable(io.update, enable = io.update.valid)
+  // load has already been assigned with a store set
+  val loadAssigned = RegNext(valid(io.update.ldpc))
+  val loadOldSSID = RegNext(ssid(io.update.ldpc))
+  // store has already been assigned with a store set
+  val storeAssigned = RegNext(valid(io.update.stpc))
+  val storeOldSSID = RegNext(ssid(io.update.stpc))
+  // both the load and the store have already been assigned store sets
+  // but load's store set ID is smaller
+  val winnerSSID = Mux(loadOldSSID < storeOldSSID, loadOldSSID, storeOldSSID)
+
+  // for now we just use lowest bits of ldpc as store set id
+  val ssidAllocate = memPredUpdateReqReg.ldpc(SSIDWidth-1, 0)
+
+  // update stage 1
+  when(memPredUpdateReqValid){
+    switch (Cat(loadAssigned, storeAssigned)) {
+      // 1. "If neither the load nor the store has been assigned a store set, 
+      // one is allocated and assigned to both instructions."
+      is ("b00".U(2.W)) {
+        valid(memPredUpdateReqReg.ldpc) := true.B
+        isload(memPredUpdateReqReg.ldpc) := true.B
+        ssid(memPredUpdateReqReg.ldpc) := ssidAllocate
+        valid(memPredUpdateReqReg.stpc) := true.B
+        isload(memPredUpdateReqReg.stpc) := false.B
+        ssid(memPredUpdateReqReg.stpc) := ssidAllocate
+      }
+      // 2. "If the load has been assigned a store set, but the store has not, 
+      // the store is assigned the load’s store set."
+      is ("b10".U(2.W)) {
+        valid(memPredUpdateReqReg.stpc) := true.B
+        isload(memPredUpdateReqReg.stpc) := false.B
+        ssid(memPredUpdateReqReg.stpc) := loadOldSSID
+      }
+      // 3. "If the store has been assigned a store set, but the load has not, 
+      // the load is assigned the store’s store set."
+      is ("b01".U(2.W)) {
+        valid(memPredUpdateReqReg.ldpc) := true.B
+        isload(memPredUpdateReqReg.ldpc) := true.B
+        ssid(memPredUpdateReqReg.ldpc) := storeOldSSID
+      }
+      // 4. "If both the load and the store have already been assigned store sets, 
+      // one of the two store sets is declared the "winner". 
+      // The instruction belonging to the loser’s store set is assigned the winner’s store set."
+      is ("b11".U(2.W)) {
+        valid(memPredUpdateReqReg.ldpc) := true.B
+        isload(memPredUpdateReqReg.ldpc) := true.B
+        ssid(memPredUpdateReqReg.ldpc) := winnerSSID
+        valid(memPredUpdateReqReg.stpc) := true.B
+        isload(memPredUpdateReqReg.stpc) := false.B
+        ssid(memPredUpdateReqReg.stpc) := winnerSSID
+      }
+    }
+  }
+
+  XSPerfAccumulate("ssit_update_lxsx", memPredUpdateReqValid && !loadAssigned && !storeAssigned)
+  XSPerfAccumulate("ssit_update_lysx", memPredUpdateReqValid && loadAssigned && !storeAssigned)
+  XSPerfAccumulate("ssit_update_lxsy", memPredUpdateReqValid && !loadAssigned && storeAssigned)
+  XSPerfAccumulate("ssit_update_lysy", memPredUpdateReqValid && loadAssigned && storeAssigned)
+
+  // reset period: ResetTimeMax2Pow
+  when(resetCounter(ResetTimeMax2Pow-1, ResetTimeMin2Pow)(RegNext(io.csrCtrl.waittable_timeout))) {
+    for (j <- 0 until SSITSize) {
+      valid(j) := 0.U
+    }
+    resetCounter:= 0.U
+  }
+
+  // debug
+  for (i <- 0 until StorePipelineWidth) {
+    when (memPredUpdateReqReg.valid) {
+      XSDebug("%d: SSIT update: load pc %x store pc %x\n", GTimer(), memPredUpdateReqReg.ldpc, memPredUpdateReqReg.stpc)
+      XSDebug("%d: SSIT update: load valid %b ssid %x  store valid %b ssid %x\n", GTimer(), loadAssigned, loadOldSSID, storeAssigned,storeOldSSID)
+    }
+  }
+}
+
+
+// Last Fetched Store Table Entry
+class LFSTEntry(implicit p: Parameters) extends XSBundle  {
+  val valid = Bool()
+  val sqIdx = new SqPtr
+  val roqIdx = new RoqPtr
+}
+
+class DispatchToLFST(implicit p: Parameters) extends XSBundle  {
+  val sqIdx = new SqPtr
+  val roqIdx = new RoqPtr
+  val ssid = UInt(SSIDWidth.W)
+}
+
+class LookupLFST(implicit p: Parameters) extends XSBundle  {
+  val raddr = Vec(DecodeWidth, Input(UInt(SSIDWidth.W))) // use ssid to llokup LFST
+  val ren = Vec(DecodeWidth, Input(Bool())) // ren iff uop.cf.storeSetHit
+  val rdata = Vec(DecodeWidth, Output(Bool()))
+}
+
+// Last Fetched Store Table
+class LFST(implicit p: Parameters) extends XSModule  {
+  val io = IO(new Bundle {
+    val lookup = new LookupLFST
+    // val update = Input(new MemPredUpdateReq) // RegNext should be added outside
+    // when redirect, mark canceled store as invalid
+    val redirect = Input(Valid(new Redirect))
+    val flush = Input(Bool())
+    // when store is dispatched, mark it as valid
+    val dispatch = Vec(RenameWidth, Flipped(Valid(new DispatchToLFST)))
+    // when store issued, mark store as invalid
+    val storeIssue = Vec(exuParameters.StuCnt, Flipped(Valid(new ExuInput)))
+    val csrCtrl = Input(new CustomCSRCtrlIO)
+  })
+
+  // TODO: use MemTemplate
+  val validVec = RegInit(VecInit(Seq.fill(LFSTSize)(VecInit(Seq.fill(LFSTWidth)(false.B)))))
+  val sqIdxVec = Reg(Vec(LFSTSize, Vec(LFSTWidth, new SqPtr)))
+  val roqIdxVec = Reg(Vec(LFSTSize, Vec(LFSTWidth, new RoqPtr)))
+  val allocPtr = RegInit(VecInit(Seq.fill(LFSTSize)(0.U(log2Up(LFSTWidth).W))))
+  val valid = Wire(Vec(LFSTSize, Bool()))
+  (0 until LFSTSize).map(i => {
+    valid(i) := validVec(i).asUInt.orR
+  })
+
+  // read LFST in rename stage
+  for (i <- 0 until DecodeWidth) {
+    // If store-load pair is in the same dispatch bundle, loadWaitBit should also be set for load 
+    val hitInDispatchBundle = if(i > 0){
+      (0 until i).map(j =>
+        io.dispatch(j).valid && io.dispatch(j).bits.ssid === io.lookup.raddr(i)
+      ).reduce(_||_)
+    } else {
+      false.B
+    }
+    // Check if store set is valid in LFST
+    io.lookup.rdata(i) := (
+        (valid(io.lookup.raddr(i)) || hitInDispatchBundle) && io.lookup.ren(i) || 
+        io.csrCtrl.no_spec_load // set loadWaitBit for all loads
+      ) && !io.csrCtrl.lvpred_disable
+  }
+
+  // when store is issued, mark it as invalid
+  (0 until exuParameters.StuCnt).map(i => {
+    // TODO: opt timing
+    (0 until LFSTWidth).map(j => {
+      when(io.storeIssue(i).valid && io.storeIssue(i).bits.uop.sqIdx.asUInt === sqIdxVec(io.storeIssue(i).bits.uop.cf.ssid)(j).asUInt){
+        validVec(io.storeIssue(i).bits.uop.cf.ssid)(j) := false.B
+      }
+    })
+  })
+
+  // when store is dispatched, mark it as valid
+  (0 until RenameWidth).map(i => {
+    when(io.dispatch(i).valid){
+      val waddr = io.dispatch(i).bits.ssid
+      val wptr = allocPtr(waddr)
+      allocPtr(waddr) := allocPtr(waddr) + 1.U
+      validVec(waddr)(wptr) := true.B
+      sqIdxVec(waddr)(wptr) := io.dispatch(i).bits.sqIdx
+      roqIdxVec(waddr)(wptr) := io.dispatch(i).bits.roqIdx
+    }
+  })
+
+  // when redirect, cancel store influenced
+  (0 until LFSTSize).map(i => {
+    (0 until LFSTWidth).map(j => {
+      when(roqIdxVec(i)(j).needFlush(io.redirect, io.flush)){
+        validVec(i)(j) := false.B
+      }
+    })
+  })
+}
\ No newline at end of file

src/main/scala/xiangshan/backend/decode/WaitTable.scala

@@ -6,19 +6,12 @@ import chisel3.util._
 import xiangshan._
 import utils._
 
-trait WaitTableParameters {
-  val WaitTableSize = 1024
-  val WaitTableAddrWidth = log2Up(WaitTableSize)
-  val ResetTimeMax2Pow = 20 //1078576
-  val ResetTimeMin2Pow = 10 //1024
-}
-
 // 21264-like wait table
-class WaitTable(implicit p: Parameters) extends XSModule with WaitTableParameters {
+class WaitTable(implicit p: Parameters) extends XSModule {
   val io = IO(new Bundle {
-    val raddr = Vec(DecodeWidth, Input(UInt(WaitTableAddrWidth.W))) // decode pc(VaddrBits-1, 1)
+    val raddr = Vec(DecodeWidth, Input(UInt(MemPredPCWidth.W))) // decode pc(VaddrBits-1, 1)
     val rdata = Vec(DecodeWidth, Output(Bool())) // loadWaitBit
-    val update = Vec(StorePipelineWidth, Input(new WaitTableUpdateReq)) // RegNext should be added outside
+    val update = Vec(StorePipelineWidth, Input(new MemPredUpdateReq)) // RegNext should be added outside
     val csrCtrl = Input(new CustomCSRCtrlIO)
   })
 
@@ -28,7 +21,7 @@ class WaitTable(implicit p: Parameters) extends XSModule with WaitTableParameter
 
   // read ports
   for (i <- 0 until DecodeWidth) {
-    io.rdata(i) := (data(io.raddr(i))(1) || io.csrCtrl.no_spec_load) && !io.csrCtrl.lvpred_disable
+    io.rdata(i) := (data(io.raddr(i))(LWTUse2BitCounter.B.asUInt) || io.csrCtrl.no_spec_load) && !io.csrCtrl.lvpred_disable
   }
 
   // write ports (with priority)

src/main/scala/xiangshan/backend/dispatch/Dispatch.scala

@@ -50,6 +50,9 @@ class Dispatch(implicit p: Parameters) extends XSModule {
       val fpIndex = Vec(exuParameters.FpExuCnt, Output(UInt(log2Ceil((NRFpReadPorts - exuParameters.StuCnt) / 3).W)))
       // ls: hardwired to (0, 1, 2, 4)
     }
+    val csrCtrl = Input(new CustomCSRCtrlIO)
+    // LFST state sync
+    val storeIssue = Vec(StorePipelineWidth, Flipped(Valid(new ExuInput)))
     val ctrlInfo = new Bundle {
       val roqFull   = Output(Bool())
       val intdqFull = Output(Bool())
@@ -80,6 +83,10 @@ class Dispatch(implicit p: Parameters) extends XSModule {
   dispatch1.io.toFpDq <> fpDq.io.enq
   dispatch1.io.toLsDq <> lsDq.io.enq
   dispatch1.io.allocPregs <> io.allocPregs
+  dispatch1.io.csrCtrl <> io.csrCtrl
+  dispatch1.io.storeIssue <> io.storeIssue
+  dispatch1.io.redirect <> io.redirect
+  dispatch1.io.flush <> io.flush
 
   // dispatch queue: queue uops and dispatch them to different reservation stations or issue queues
   // it may cancel the uops

src/main/scala/xiangshan/backend/dispatch/Dispatch1.scala

@@ -9,6 +9,7 @@ import xiangshan.backend.roq.{RoqEnqIO, RoqPtr}
 import xiangshan.backend.rename.RenameBypassInfo
 import xiangshan.mem.LsqEnqIO
 import xiangshan.backend.fu.HasExceptionNO
+import xiangshan.backend.decode.{LFST, DispatchToLFST, LookupLFST}
 
 
 class PreDispatchInfo(implicit p: Parameters) extends XSBundle {
@@ -44,9 +45,30 @@ class Dispatch1(implicit p: Parameters) extends XSModule with HasExceptionNO {
       val needAlloc = Vec(RenameWidth, Output(Bool()))
       val req = Vec(RenameWidth, ValidIO(new MicroOp))
     }
+    // to store set LFST
+    val lfst = Vec(RenameWidth, Valid(new DispatchToLFST))
+    // flush or replay, for LFST
+    val redirect = Flipped(ValidIO(new Redirect))
+    val flush = Input(Bool())
+    // LFST ctrl
+    val csrCtrl = Input(new CustomCSRCtrlIO)
+    // LFST state sync
+    val storeIssue = Vec(StorePipelineWidth, Flipped(Valid(new ExuInput)))
   })
 
 
+  /**
+    * Store set LFST lookup
+    */
+  // store set LFST lookup may start from rename for better timing
+
+  val lfst = Module(new LFST)
+  lfst.io.redirect <> RegNext(io.redirect)
+  lfst.io.flush <> RegNext(io.flush)
+  lfst.io.storeIssue <> RegNext(io.storeIssue)
+  lfst.io.csrCtrl <> RegNext(io.csrCtrl)
+  lfst.io.dispatch := io.lfst
+
   /**
     * Part 1: choose the target dispatch queue and the corresponding write ports
     */
@@ -124,8 +146,54 @@ class Dispatch1(implicit p: Parameters) extends XSModule with HasExceptionNO {
 //    XSError(io.fromRename(i).valid && updatedUop(i).roqIdx.asUInt =/= io.enqRoq.resp(i).asUInt, "they should equal")
     updatedUop(i).lqIdx  := io.enqLsq.resp(i).lqIdx
     updatedUop(i).sqIdx  := io.enqLsq.resp(i).sqIdx
+
+    // lookup store set LFST
+    lfst.io.lookup.raddr(i) := updatedUop(i).cf.ssid
+    lfst.io.lookup.ren(i) := updatedUop(i).cf.storeSetHit
+
+    // override load delay ctrl signal with store set result
+    if(StoreSetEnable) {
+      // updatedUop(i).cf.loadWaitBit := lfst.io.lookup.rdata(i) // classic store set
+      updatedUop(i).cf.loadWaitBit := lfst.io.lookup.rdata(i) && !isStore(i) // store set lite
+      // updatedUop(i).cf.loadWaitBit := lfst.io.lookup.rdata(i) && io.fromRename(i).bits.cf.loadWaitBit && !isStore(i) // 2-bit store set
+    } else {
+      updatedUop(i).cf.loadWaitBit := io.fromRename(i).bits.cf.loadWaitBit && !isStore(i) // wait table does not require store to be delayed
+    }
+
+    // update store set LFST
+    io.lfst(i).valid := io.fromRename(i).valid && updatedUop(i).cf.storeSetHit && isStore(i)
+    // or io.fromRename(i).ready && updatedUop(i).cf.storeSetHit && isStore(i), which is much slower
+    io.lfst(i).bits.roqIdx := updatedUop(i).roqIdx
+    io.lfst(i).bits.sqIdx := updatedUop(i).sqIdx
+    io.lfst(i).bits.ssid := updatedUop(i).cf.ssid
   }
 
+  // store set perf count
+  XSPerfAccumulate("waittable_load_wait", PopCount((0 until RenameWidth).map(i => 
+    io.fromRename(i).fire() && io.fromRename(i).bits.cf.loadWaitBit && !isStore(i) && isLs(i)
+  )))
+  XSPerfAccumulate("storeset_load_wait", PopCount((0 until RenameWidth).map(i => 
+    io.fromRename(i).fire() && updatedUop(i).cf.loadWaitBit && !isStore(i) && isLs(i)
+  )))
+  XSPerfAccumulate("storeset_store_wait", PopCount((0 until RenameWidth).map(i => 
+    io.fromRename(i).fire() && updatedUop(i).cf.loadWaitBit && isStore(i)
+  )))
+  XSPerfAccumulate("loadwait_diffmat_sywy", PopCount((0 until RenameWidth).map(i => 
+    io.fromRename(i).fire() && updatedUop(i).cf.loadWaitBit && io.fromRename(i).bits.cf.loadWaitBit &&
+    !isStore(i) && isLs(i)
+  )))
+  XSPerfAccumulate("loadwait_diffmat_sywx", PopCount((0 until RenameWidth).map(i => 
+    io.fromRename(i).fire() && updatedUop(i).cf.loadWaitBit && !io.fromRename(i).bits.cf.loadWaitBit &&
+    !isStore(i) && isLs(i)
+  )))
+  XSPerfAccumulate("loadwait_diffmat_sxwy", PopCount((0 until RenameWidth).map(i => 
+    io.fromRename(i).fire() && !updatedUop(i).cf.loadWaitBit && io.fromRename(i).bits.cf.loadWaitBit &&
+    !isStore(i) && isLs(i)
+  )))
+  XSPerfAccumulate("loadwait_diffmat_sxwx", PopCount((0 until RenameWidth).map(i => 
+    io.fromRename(i).fire() && !updatedUop(i).cf.loadWaitBit && !io.fromRename(i).bits.cf.loadWaitBit &&
+    !isStore(i) && isLs(i)
+  )))
 
   /**
     * Part 3:

src/main/scala/xiangshan/backend/ftq/Ftq.scala

@@ -108,8 +108,8 @@ class Ftq(implicit p: Parameters) extends XSModule with HasCircularQueuePtrHelpe
     val frontendRedirect = Flipped(ValidIO(new Redirect))
     // exu write back, update info
     val exuWriteback = Vec(exuParameters.JmpCnt + exuParameters.AluCnt, Flipped(ValidIO(new ExuOutput)))
-    // pc read reqs (0: jump/auipc 1~6: mispredict/load replay 7: exceptions)
-    val ftqRead = Vec(1 + 6 + 1, Flipped(new FtqRead))
+    // pc read reqs (0: jump/auipc 1~6: mispredict/load replay 7: store pc for store set update 8: exceptions)
+    val ftqRead = Vec(1 + 6 + 1 + 1, Flipped(new FtqRead))
     val cfiRead = Flipped(new FtqRead)
     val bpuInfo = new Bundle {
       val bpRight = Output(UInt(XLEN.W))
@@ -131,7 +131,7 @@ class Ftq(implicit p: Parameters) extends XSModule with HasCircularQueuePtrHelpe
 
   val real_fire = io.enq.fire() && !stage2Flush && !stage3Flush
 
-  val ftq_pc_mem = Module(new SyncDataModuleTemplate(new Ftq_4R_SRAMEntry, FtqSize, 9, 1))
+  val ftq_pc_mem = Module(new SyncDataModuleTemplate(new Ftq_4R_SRAMEntry, FtqSize, 10, 1))
   ftq_pc_mem.io.wen(0) := real_fire
   ftq_pc_mem.io.waddr(0) := tailPtr.value
   ftq_pc_mem.io.wdata(0).ftqPC := io.enq.bits.ftqPC

src/main/scala/xiangshan/backend/issue/ReservationStation.scala

@@ -105,7 +105,7 @@ class ReservationStation
     val stData = if (exuCfg == StExeUnitCfg) ValidIO(new StoreDataBundle) else null
     val srcRegValue = Input(Vec(srcNum, UInt(srcLen.W)))
 
-    val stIssuePtr = if (exuCfg == LdExeUnitCfg) Input(new SqPtr()) else null
+    val stIssuePtr = if (exuCfg == LdExeUnitCfg || exuCfg == StExeUnitCfg) Input(new SqPtr()) else null
 
     val fpRegValue = if (exuCfg == StExeUnitCfg) Input(UInt(srcLen.W)) else null
     val jumpPc = if(exuCfg == JumpExeUnitCfg) Input(UInt(VAddrBits.W)) else null
@@ -165,7 +165,7 @@ class ReservationStation
     c.valid := i.valid
     c.bits  := i.bits.uop
   }
-  if (exuCfg == LdExeUnitCfg) {
+  if (exuCfg == LdExeUnitCfg || exuCfg == StExeUnitCfg) {
     ctrl.io.stIssuePtr := RegNext(io.stIssuePtr)
   }
   if (exuCfg == StExeUnitCfg) {
@@ -541,7 +541,7 @@ class ReservationStationCtrl
     val listen = Output(Vec(srcNum, Vec(iqSize, Vec(fastPortsCnt + slowPortsCnt, Bool()))))
     val enqSrcReady = Output(Vec(srcNum, Bool()))
 
-    val stIssuePtr = if (exuCfg == LdExeUnitCfg) Input(new SqPtr()) else null
+    val stIssuePtr = if (exuCfg == LdExeUnitCfg || exuCfg == StExeUnitCfg) Input(new SqPtr()) else null
   })
 
   val selValid = io.sel.valid
@@ -619,6 +619,7 @@ class ReservationStationCtrl
   if (exuCfg == LdExeUnitCfg) {
     val ldWait = Reg(Vec(iqSize, Bool()))
     val sqIdx  = Reg(Vec(iqSize, new SqPtr()))
+    val ldWaitUpdated = WireInit(ldWait)
     ldWait.zip(sqIdx).map{ case (lw, sq) =>
       when (!isAfter(sq, io.stIssuePtr)) {
         lw := true.B
@@ -626,11 +627,12 @@ class ReservationStationCtrl
     }
     when (enqEn) {
       ldWait(enqPtr) := !enqUop.cf.loadWaitBit
+      ldWaitUpdated(enqPtr) := !enqUop.cf.loadWaitBit
       sqIdx(enqPtr)  := enqUop.sqIdx
     }
     ldWait.suggestName(s"${this.name}_ldWait")
     sqIdx.suggestName(s"${this.name}_sqIdx")
-    io.readyVec := srcQueueWire.map(Cat(_).andR).zip(ldWait).map{ case (s, l) => s&l }
+    io.readyVec := srcQueueWire.map(Cat(_).andR).zip(ldWaitUpdated).map{ case (s, l) => s&l }
   }
 
   val redirectHit = io.redirectVec(selPtr)

src/main/scala/xiangshan/frontend/IFU.scala

@@ -9,7 +9,6 @@ import xiangshan.cache._
 import chisel3.experimental.chiselName
 import freechips.rocketchip.tile.HasLazyRoCC
 import xiangshan.backend.ftq.FtqPtr
-import xiangshan.backend.decode.WaitTableParameters
 import system.L1CacheErrorInfo
 
 trait HasInstrMMIOConst extends HasXSParameter with HasIFUConst{
@@ -99,7 +98,7 @@ class PrevHalfInstr(implicit p: Parameters) extends XSBundle {
 }
 
 @chiselName
-class IFU(implicit p: Parameters) extends XSModule with HasIFUConst with HasCircularQueuePtrHelper with WaitTableParameters
+class IFU(implicit p: Parameters) extends XSModule with HasIFUConst with HasCircularQueuePtrHelper
 {
   val io = IO(new IFUIO)
   val bpu = BPU(EnableBPU)
@@ -518,7 +517,7 @@ class IFU(implicit p: Parameters) extends XSModule with HasIFUConst with HasCirc
   fetchPacketWire.instrs := expandedInstrs
 
   fetchPacketWire.pc := if4_pd.pc
-  fetchPacketWire.foldpc := if4_pd.pc.map(i => XORFold(i(VAddrBits-1,1), WaitTableAddrWidth))
+  fetchPacketWire.foldpc := if4_pd.pc.map(i => XORFold(i(VAddrBits-1,1), MemPredPCWidth))
 
   fetchPacketWire.pdmask := if4_pd.mask
   fetchPacketWire.pd := if4_pd.pd

src/main/scala/xiangshan/frontend/Ibuffer.scala

@@ -6,7 +6,6 @@ import chisel3.util._
 import xiangshan._
 import utils._
 import xiangshan.backend.ftq.FtqPtr
-import xiangshan.backend.decode.WaitTableParameters
 
 class IbufPtr(implicit p: Parameters) extends CircularQueuePtr[IbufPtr](
   p => p(XSCoreParamsKey).IBufSize
@@ -24,10 +23,10 @@ class IBufferIO(implicit p: Parameters) extends XSBundle {
 class Ibuffer(implicit p: Parameters) extends XSModule with HasCircularQueuePtrHelper {
   val io = IO(new IBufferIO)
 
-  class IBufEntry(implicit p: Parameters) extends XSBundle with WaitTableParameters {
+  class IBufEntry(implicit p: Parameters) extends XSBundle {
     val inst = UInt(32.W)
     val pc = UInt(VAddrBits.W)
-    val foldpc = UInt(WaitTableAddrWidth.W)
+    val foldpc = UInt(MemPredPCWidth.W)
     val pd = new PreDecodeInfo
     val ipf = Bool()
     val acf = Bool()
@@ -125,6 +124,8 @@ class Ibuffer(implicit p: Parameters) extends XSModule with HasCircularQueuePtrH
     io.out(i).bits.crossPageIPFFix := outWire.crossPageIPFFix
     io.out(i).bits.foldpc := outWire.foldpc
     io.out(i).bits.loadWaitBit := DontCare
+    io.out(i).bits.storeSetHit := DontCare
+    io.out(i).bits.ssid := DontCare
   }
   val next_head_vec = VecInit(head_vec.map(_ + numDeq))
   ibuf.io.raddr := VecInit(next_head_vec.map(_.value))

src/main/scala/xiangshan/mem/lsqueue/LoadQueue.scala

@@ -11,6 +11,7 @@ import xiangshan.cache.{DCacheLineIO, DCacheWordIO, MemoryOpConstants, TlbReques
 import xiangshan.mem._
 import xiangshan.backend.roq.RoqLsqIO
 import xiangshan.backend.fu.HasExceptionNO
+import xiangshan.backend.ftq.FtqPtr
 
 
 class LqPtr(implicit p: Parameters) extends CircularQueuePtr[LqPtr](
@@ -455,12 +456,12 @@ class LoadQueue(implicit p: Parameters) extends XSModule
     ((lqViolation, lqViolationUop), (wbViolation, wbViolationUop), (l1Violation, l1ViolationUop))
   }
 
-  def rollbackSel(a: Valid[MicroOp], b: Valid[MicroOp]): ValidIO[MicroOp] = {
+  def rollbackSel(a: Valid[MicroOpRbExt], b: Valid[MicroOpRbExt]): ValidIO[MicroOpRbExt] = {
     Mux(
       a.valid,
       Mux(
         b.valid,
-        Mux(isAfter(a.bits.roqIdx, b.bits.roqIdx), b, a), // a,b both valid, sel oldest
+        Mux(isAfter(a.bits.uop.roqIdx, b.bits.uop.roqIdx), b, a), // a,b both valid, sel oldest
         a // sel a
       ),
       b // sel b
@@ -474,21 +475,29 @@ class LoadQueue(implicit p: Parameters) extends XSModule
   // S2: select rollback (part1) and generate rollback request
   // rollback check
   // Wb/L1 rollback seq check is done in s2
-  val rollbackWb = Wire(Vec(StorePipelineWidth, Valid(new MicroOp)))
-  val rollbackL1 = Wire(Vec(StorePipelineWidth, Valid(new MicroOp)))
-  val rollbackL1Wb = Wire(Vec(StorePipelineWidth*2, Valid(new MicroOp)))
+  val rollbackWb = Wire(Vec(StorePipelineWidth, Valid(new MicroOpRbExt)))
+  val rollbackL1 = Wire(Vec(StorePipelineWidth, Valid(new MicroOpRbExt)))
+  val rollbackL1Wb = Wire(Vec(StorePipelineWidth*2, Valid(new MicroOpRbExt)))
   // Lq rollback seq check is done in s3 (next stage), as getting rollbackLq MicroOp is slow
-  val rollbackLq = Wire(Vec(StorePipelineWidth, Valid(new MicroOp)))
+  val rollbackLq = Wire(Vec(StorePipelineWidth, Valid(new MicroOpRbExt)))
+  // store ftq index for store set update
+  val stFtqIdxS2 = Wire(Vec(StorePipelineWidth, new FtqPtr))
+  val stFtqOffsetS2 = Wire(Vec(StorePipelineWidth, UInt(log2Up(PredictWidth).W)))
   for (i <- 0 until StorePipelineWidth) {
     val detectedRollback = detectRollback(i)
     rollbackLq(i).valid := detectedRollback._1._1 && RegNext(io.storeIn(i).valid)
-    rollbackLq(i).bits := detectedRollback._1._2
+    rollbackLq(i).bits.uop := detectedRollback._1._2
+    rollbackLq(i).bits.flag := i.U
     rollbackWb(i).valid := detectedRollback._2._1 && RegNext(io.storeIn(i).valid)
-    rollbackWb(i).bits := detectedRollback._2._2
+    rollbackWb(i).bits.uop := detectedRollback._2._2
+    rollbackWb(i).bits.flag := i.U
     rollbackL1(i).valid := detectedRollback._3._1 && RegNext(io.storeIn(i).valid)
-    rollbackL1(i).bits := detectedRollback._3._2
+    rollbackL1(i).bits.uop := detectedRollback._3._2
+    rollbackL1(i).bits.flag := i.U
     rollbackL1Wb(2*i) := rollbackL1(i)
     rollbackL1Wb(2*i+1) := rollbackWb(i)
+    stFtqIdxS2(i) := RegNext(io.storeIn(i).bits.uop.cf.ftqPtr)
+    stFtqOffsetS2(i) := RegNext(io.storeIn(i).bits.uop.cf.ftqOffset)
   }
 
   val rollbackL1WbSelected = ParallelOperation(rollbackL1Wb, rollbackSel)
@@ -505,18 +514,23 @@ class LoadQueue(implicit p: Parameters) extends XSModule
 
   // FIXME: this is ugly
   val rollbackValidVec = Seq(rollbackL1WbVReg, rollbackLq0VReg, rollbackLq1VReg)
-  val rollbackUopVec = Seq(rollbackL1WbReg, rollbackLq0Reg, rollbackLq1Reg)
+  val rollbackUopExtVec = Seq(rollbackL1WbReg, rollbackLq0Reg, rollbackLq1Reg)
 
   // select uop in parallel
-  val mask = getAfterMask(rollbackValidVec, rollbackUopVec)
+  val mask = getAfterMask(rollbackValidVec, rollbackUopExtVec.map(i => i.uop))
   val oneAfterZero = mask(1)(0)
-  val rollbackUop = Mux(oneAfterZero && mask(2)(0),
-    rollbackUopVec(0),
-    Mux(!oneAfterZero && mask(2)(1), rollbackUopVec(1), rollbackUopVec(2)))
+  val rollbackUopExt = Mux(oneAfterZero && mask(2)(0),
+    rollbackUopExtVec(0),
+    Mux(!oneAfterZero && mask(2)(1), rollbackUopExtVec(1), rollbackUopExtVec(2)))
+  val stFtqIdxS3 = RegNext(stFtqIdxS2)
+  val stFtqOffsetS3 = RegNext(stFtqOffsetS2)
+  val rollbackUop = rollbackUopExt.uop
+  val rollbackStFtqIdx = stFtqIdxS3(rollbackUopExt.flag)
+  val rollbackStFtqOffset = stFtqOffsetS3(rollbackUopExt.flag)
 
   // check if rollback request is still valid in parallel
   val rollbackValidVecChecked = Wire(Vec(3, Bool()))
-  for(((v, uop), idx) <- rollbackValidVec.zip(rollbackUopVec).zipWithIndex) {
+  for(((v, uop), idx) <- rollbackValidVec.zip(rollbackUopExtVec.map(i => i.uop)).zipWithIndex) {
     rollbackValidVecChecked(idx) := v && 
       (!lastCycleRedirect.valid || isBefore(uop.roqIdx, lastCycleRedirect.bits.roqIdx)) &&
       (!lastlastCycleRedirect.valid || isBefore(uop.roqIdx, lastlastCycleRedirect.bits.roqIdx))
@@ -524,7 +538,9 @@ class LoadQueue(implicit p: Parameters) extends XSModule
 
   io.rollback.bits.roqIdx := rollbackUop.roqIdx
   io.rollback.bits.ftqIdx := rollbackUop.cf.ftqPtr
+  io.rollback.bits.stFtqIdx := rollbackStFtqIdx
   io.rollback.bits.ftqOffset := rollbackUop.cf.ftqOffset
+  io.rollback.bits.stFtqOffset := rollbackStFtqOffset
   io.rollback.bits.level := RedirectLevel.flush
   io.rollback.bits.interrupt := DontCare
   io.rollback.bits.cfiUpdate := DontCare