xref: /XiangShan/src/main/scala/xiangshan/mem/lsqueue/StoreQueue.scala (revision 38c29594d00da67087523376f6a6f3243679884e)
1/***************************************************************************************
2* Copyright (c) 2024 Beijing Institute of Open Source Chip (BOSC)
3* Copyright (c) 2020-2024 Institute of Computing Technology, Chinese Academy of Sciences
4* Copyright (c) 2020-2021 Peng Cheng Laboratory
5*
6* XiangShan is licensed under Mulan PSL v2.
7* You can use this software according to the terms and conditions of the Mulan PSL v2.
8* You may obtain a copy of Mulan PSL v2 at:
9*          http://license.coscl.org.cn/MulanPSL2
10*
11* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
12* EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
13* MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
14*
15* See the Mulan PSL v2 for more details.
16***************************************************************************************/
17
18package xiangshan.mem
19
20import chisel3._
21import chisel3.util._
22import difftest._
23import difftest.common.DifftestMem
24import org.chipsalliance.cde.config.Parameters
25import utility._
26import utils._
27import xiangshan._
28import xiangshan.cache._
29import xiangshan.cache.{DCacheLineIO, DCacheWordIO, MemoryOpConstants}
30import xiangshan.cache.{CMOReq, CMOResp}
31import xiangshan.backend._
32import xiangshan.backend.rob.{RobLsqIO, RobPtr}
33import xiangshan.backend.Bundles.{DynInst, MemExuOutput}
34import xiangshan.backend.decode.isa.bitfield.{Riscv32BitInst, XSInstBitFields}
35import xiangshan.backend.fu.FuConfig._
36import xiangshan.backend.fu.FuType
37import xiangshan.ExceptionNO._
38
39class SqPtr(implicit p: Parameters) extends CircularQueuePtr[SqPtr](
40  p => p(XSCoreParamsKey).StoreQueueSize
41){
42}
43
44object SqPtr {
45  def apply(f: Bool, v: UInt)(implicit p: Parameters): SqPtr = {
46    val ptr = Wire(new SqPtr)
47    ptr.flag := f
48    ptr.value := v
49    ptr
50  }
51}
52
53class SqEnqIO(implicit p: Parameters) extends MemBlockBundle {
54  val canAccept = Output(Bool())
55  val lqCanAccept = Input(Bool())
56  val needAlloc = Vec(LSQEnqWidth, Input(Bool()))
57  val req = Vec(LSQEnqWidth, Flipped(ValidIO(new DynInst)))
58  val resp = Vec(LSQEnqWidth, Output(new SqPtr))
59}
60
61class DataBufferEntry (implicit p: Parameters)  extends DCacheBundle {
62  val addr   = UInt(PAddrBits.W)
63  val vaddr  = UInt(VAddrBits.W)
64  val data   = UInt(VLEN.W)
65  val mask   = UInt((VLEN/8).W)
66  val wline = Bool()
67  val sqPtr  = new SqPtr
68  val prefetch = Bool()
69  val vecValid = Bool()
70}
71
72class StoreExceptionBuffer(implicit p: Parameters) extends XSModule with HasCircularQueuePtrHelper {
73  // The 1st StorePipelineWidth ports: sta exception generated at s1, except for af
74  // The 2nd StorePipelineWidth ports: sta af generated at s2
75  // The following VecStorePipelineWidth ports: vector st exception
76  // The last port: non-data error generated in SoC
77  val enqPortNum = StorePipelineWidth * 2 + VecStorePipelineWidth + 1
78
79  val io = IO(new Bundle() {
80    val redirect = Flipped(ValidIO(new Redirect))
81    val storeAddrIn = Vec(enqPortNum, Flipped(ValidIO(new LsPipelineBundle())))
82    val flushFrmMaBuf = Input(Bool())
83    val exceptionAddr = new ExceptionAddrIO
84  })
85
86  val req_valid = RegInit(false.B)
87  val req = Reg(new LsPipelineBundle())
88
89  // enqueue
90  // S1:
91  val s1_req = VecInit(io.storeAddrIn.map(_.bits))
92  val s1_valid = VecInit(io.storeAddrIn.map(x =>
93      x.valid && !x.bits.uop.robIdx.needFlush(io.redirect) && ExceptionNO.selectByFu(x.bits.uop.exceptionVec, StaCfg).asUInt.orR
94  ))
95
96  // S2: delay 1 cycle
97  val s2_req = (0 until enqPortNum).map(i =>
98    RegEnable(s1_req(i), s1_valid(i)))
99  val s2_valid = (0 until enqPortNum).map(i =>
100    RegNext(s1_valid(i)) && !s2_req(i).uop.robIdx.needFlush(io.redirect)
101  )
102
103  val s2_enqueue = Wire(Vec(enqPortNum, Bool()))
104  for (w <- 0 until enqPortNum) {
105    s2_enqueue(w) := s2_valid(w)
106  }
107
108  when (req_valid && req.uop.robIdx.needFlush(io.redirect)) {
109    req_valid := s2_enqueue.asUInt.orR
110  }.elsewhen (s2_enqueue.asUInt.orR) {
111    req_valid := req_valid || true.B
112  }
113
114  def selectOldest[T <: LsPipelineBundle](valid: Seq[Bool], bits: Seq[T]): (Seq[Bool], Seq[T]) = {
115    assert(valid.length == bits.length)
116    if (valid.length == 0 || valid.length == 1) {
117      (valid, bits)
118    } else if (valid.length == 2) {
119      val res = Seq.fill(2)(Wire(Valid(chiselTypeOf(bits(0)))))
120      for (i <- res.indices) {
121        res(i).valid := valid(i)
122        res(i).bits := bits(i)
123      }
124      val oldest = Mux(valid(0) && valid(1),
125        Mux(isAfter(bits(0).uop.robIdx, bits(1).uop.robIdx) ||
126          (isNotBefore(bits(0).uop.robIdx, bits(1).uop.robIdx) && bits(0).uop.uopIdx > bits(1).uop.uopIdx), res(1), res(0)),
127        Mux(valid(0) && !valid(1), res(0), res(1)))
128      (Seq(oldest.valid), Seq(oldest.bits))
129    } else {
130      val left = selectOldest(valid.take(valid.length / 2), bits.take(bits.length / 2))
131      val right = selectOldest(valid.takeRight(valid.length - (valid.length / 2)), bits.takeRight(bits.length - (bits.length / 2)))
132      selectOldest(left._1 ++ right._1, left._2 ++ right._2)
133    }
134  }
135
136  val reqSel = selectOldest(s2_enqueue, s2_req)
137
138  when (req_valid) {
139    req := Mux(
140      reqSel._1(0) && (isAfter(req.uop.robIdx, reqSel._2(0).uop.robIdx) || (isNotBefore(req.uop.robIdx, reqSel._2(0).uop.robIdx) && req.uop.uopIdx > reqSel._2(0).uop.uopIdx)),
141      reqSel._2(0),
142      req)
143  } .elsewhen (s2_enqueue.asUInt.orR) {
144    req := reqSel._2(0)
145  }
146
147  io.exceptionAddr.vaddr     := req.fullva
148  io.exceptionAddr.vaNeedExt := req.vaNeedExt
149  io.exceptionAddr.isHyper   := req.isHyper
150  io.exceptionAddr.gpaddr    := req.gpaddr
151  io.exceptionAddr.vstart    := req.uop.vpu.vstart
152  io.exceptionAddr.vl        := req.uop.vpu.vl
153  io.exceptionAddr.isForVSnonLeafPTE := req.isForVSnonLeafPTE
154
155  when(req_valid && io.flushFrmMaBuf) {
156    req_valid := false.B
157  }
158}
159
160// Store Queue
161class StoreQueue(implicit p: Parameters) extends XSModule
162  with HasDCacheParameters
163  with HasCircularQueuePtrHelper
164  with HasPerfEvents
165  with HasVLSUParameters {
166  val io = IO(new Bundle() {
167    val hartId = Input(UInt(hartIdLen.W))
168    val enq = new SqEnqIO
169    val brqRedirect = Flipped(ValidIO(new Redirect))
170    val vecFeedback = Vec(VecLoadPipelineWidth, Flipped(ValidIO(new FeedbackToLsqIO)))
171    val storeAddrIn = Vec(StorePipelineWidth, Flipped(Valid(new LsPipelineBundle))) // store addr, data is not included
172    val storeAddrInRe = Vec(StorePipelineWidth, Input(new LsPipelineBundle())) // store more mmio and exception
173    val storeDataIn = Vec(StorePipelineWidth, Flipped(Valid(new MemExuOutput(isVector = true)))) // store data, send to sq from rs
174    val storeMaskIn = Vec(StorePipelineWidth, Flipped(Valid(new StoreMaskBundle))) // store mask, send to sq from rs
175    val sbuffer = Vec(EnsbufferWidth, Decoupled(new DCacheWordReqWithVaddrAndPfFlag)) // write committed store to sbuffer
176    val sbufferVecDifftestInfo = Vec(EnsbufferWidth, Decoupled(new DynInst)) // The vector store difftest needs is, write committed store to sbuffer
177    val uncacheOutstanding = Input(Bool())
178    val cmoOpReq  = DecoupledIO(new CMOReq)
179    val cmoOpResp = Flipped(DecoupledIO(new CMOResp))
180    val mmioStout = DecoupledIO(new MemExuOutput) // writeback uncached store
181    val vecmmioStout = DecoupledIO(new MemExuOutput(isVector = true))
182    val forward = Vec(LoadPipelineWidth, Flipped(new PipeLoadForwardQueryIO))
183    // TODO: scommit is only for scalar store
184    val rob = Flipped(new RobLsqIO)
185    val uncache = new UncacheWordIO
186    // val refill = Flipped(Valid(new DCacheLineReq ))
187    val exceptionAddr = new ExceptionAddrIO
188    val flushSbuffer = new SbufferFlushBundle
189    val sqEmpty = Output(Bool())
190    val stAddrReadySqPtr = Output(new SqPtr)
191    val stAddrReadyVec = Output(Vec(StoreQueueSize, Bool()))
192    val stDataReadySqPtr = Output(new SqPtr)
193    val stDataReadyVec = Output(Vec(StoreQueueSize, Bool()))
194    val stIssuePtr = Output(new SqPtr)
195    val sqDeqPtr = Output(new SqPtr)
196    val sqFull = Output(Bool())
197    val sqCancelCnt = Output(UInt(log2Up(StoreQueueSize + 1).W))
198    val sqDeq = Output(UInt(log2Ceil(EnsbufferWidth + 1).W))
199    val force_write = Output(Bool())
200    val maControl   = Flipped(new StoreMaBufToSqControlIO)
201  })
202
203  println("StoreQueue: size:" + StoreQueueSize)
204
205  // data modules
206  val uop = Reg(Vec(StoreQueueSize, new DynInst))
207  // val data = Reg(Vec(StoreQueueSize, new LsqEntry))
208  val dataModule = Module(new SQDataModule(
209    numEntries = StoreQueueSize,
210    numRead = EnsbufferWidth,
211    numWrite = StorePipelineWidth,
212    numForward = LoadPipelineWidth
213  ))
214  dataModule.io := DontCare
215  val paddrModule = Module(new SQAddrModule(
216    dataWidth = PAddrBits,
217    numEntries = StoreQueueSize,
218    numRead = EnsbufferWidth,
219    numWrite = StorePipelineWidth,
220    numForward = LoadPipelineWidth
221  ))
222  paddrModule.io := DontCare
223  val vaddrModule = Module(new SQAddrModule(
224    dataWidth = VAddrBits,
225    numEntries = StoreQueueSize,
226    numRead = EnsbufferWidth, // sbuffer; badvaddr will be sent from exceptionBuffer
227    numWrite = StorePipelineWidth,
228    numForward = LoadPipelineWidth
229  ))
230  vaddrModule.io := DontCare
231  val dataBuffer = Module(new DatamoduleResultBuffer(new DataBufferEntry))
232  val difftestBuffer = if (env.EnableDifftest) Some(Module(new DatamoduleResultBuffer(new DynInst))) else None
233  val exceptionBuffer = Module(new StoreExceptionBuffer)
234  exceptionBuffer.io.redirect := io.brqRedirect
235  exceptionBuffer.io.exceptionAddr.isStore := DontCare
236  // vlsu exception!
237  for (i <- 0 until VecStorePipelineWidth) {
238    exceptionBuffer.io.storeAddrIn(StorePipelineWidth * 2 + i).valid               := io.vecFeedback(i).valid && io.vecFeedback(i).bits.feedback(VecFeedbacks.FLUSH) // have exception
239    exceptionBuffer.io.storeAddrIn(StorePipelineWidth * 2 + i).bits                := DontCare
240    exceptionBuffer.io.storeAddrIn(StorePipelineWidth * 2 + i).bits.fullva         := io.vecFeedback(i).bits.vaddr
241    exceptionBuffer.io.storeAddrIn(StorePipelineWidth * 2 + i).bits.vaNeedExt      := io.vecFeedback(i).bits.vaNeedExt
242    exceptionBuffer.io.storeAddrIn(StorePipelineWidth * 2 + i).bits.gpaddr         := io.vecFeedback(i).bits.gpaddr
243    exceptionBuffer.io.storeAddrIn(StorePipelineWidth * 2 + i).bits.uop.uopIdx     := io.vecFeedback(i).bits.uopidx
244    exceptionBuffer.io.storeAddrIn(StorePipelineWidth * 2 + i).bits.uop.robIdx     := io.vecFeedback(i).bits.robidx
245    exceptionBuffer.io.storeAddrIn(StorePipelineWidth * 2 + i).bits.uop.vpu.vstart := io.vecFeedback(i).bits.vstart
246    exceptionBuffer.io.storeAddrIn(StorePipelineWidth * 2 + i).bits.uop.vpu.vl     := io.vecFeedback(i).bits.vl
247    exceptionBuffer.io.storeAddrIn(StorePipelineWidth * 2 + i).bits.isForVSnonLeafPTE := io.vecFeedback(i).bits.isForVSnonLeafPTE
248    exceptionBuffer.io.storeAddrIn(StorePipelineWidth * 2 + i).bits.uop.exceptionVec  := io.vecFeedback(i).bits.exceptionVec
249  }
250
251
252  val debug_paddr = Reg(Vec(StoreQueueSize, UInt((PAddrBits).W)))
253  val debug_vaddr = Reg(Vec(StoreQueueSize, UInt((VAddrBits).W)))
254  val debug_data = Reg(Vec(StoreQueueSize, UInt((XLEN).W)))
255
256  // state & misc
257  val allocated = RegInit(VecInit(List.fill(StoreQueueSize)(false.B))) // sq entry has been allocated
258  val addrvalid = RegInit(VecInit(List.fill(StoreQueueSize)(false.B)))
259  val datavalid = RegInit(VecInit(List.fill(StoreQueueSize)(false.B)))
260  val allvalid  = VecInit((0 until StoreQueueSize).map(i => addrvalid(i) && datavalid(i)))
261  val committed = RegInit(VecInit(List.fill(StoreQueueSize)(false.B))) // inst has been committed by rob
262  val unaligned = RegInit(VecInit(List.fill(StoreQueueSize)(false.B))) // unaligned store
263  val pending = RegInit(VecInit(List.fill(StoreQueueSize)(false.B))) // mmio pending: inst is an mmio inst, it will not be executed until it reachs the end of rob
264  val nc = RegInit(VecInit(List.fill(StoreQueueSize)(false.B))) // nc: inst is a nc inst
265  val mmio = RegInit(VecInit(List.fill(StoreQueueSize)(false.B))) // mmio: inst is an mmio inst
266  val atomic = RegInit(VecInit(List.fill(StoreQueueSize)(false.B)))
267  val prefetch = RegInit(VecInit(List.fill(StoreQueueSize)(false.B))) // need prefetch when committing this store to sbuffer?
268  val isVec = RegInit(VecInit(List.fill(StoreQueueSize)(false.B))) // vector store instruction
269  val vecLastFlow = RegInit(VecInit(List.fill(StoreQueueSize)(false.B))) // last uop the last flow of vector store instruction
270  val vecMbCommit = RegInit(VecInit(List.fill(StoreQueueSize)(false.B))) // vector store committed from merge buffer to rob
271  val vecDataValid = RegInit(VecInit(List.fill(StoreQueueSize)(false.B))) // vector store need write to sbuffer
272  val hasException = RegInit(VecInit(List.fill(StoreQueueSize)(false.B))) // store has exception, should deq but not write sbuffer
273  val waitStoreS2 = RegInit(VecInit(List.fill(StoreQueueSize)(false.B))) // wait for mmio and exception result until store_s2
274  // val vec_robCommit = Reg(Vec(StoreQueueSize, Bool())) // vector store committed by rob
275  // val vec_secondInv = RegInit(VecInit(List.fill(StoreQueueSize)(false.B))) // Vector unit-stride, second entry is invalid
276  val vecExceptionFlag = RegInit(0.U.asTypeOf(Valid(new DynInst)))
277
278  // ptr
279  val enqPtrExt = RegInit(VecInit((0 until io.enq.req.length).map(_.U.asTypeOf(new SqPtr))))
280  val rdataPtrExt = RegInit(VecInit((0 until EnsbufferWidth).map(_.U.asTypeOf(new SqPtr))))
281  val deqPtrExt = RegInit(VecInit((0 until EnsbufferWidth).map(_.U.asTypeOf(new SqPtr))))
282  val cmtPtrExt = RegInit(VecInit((0 until CommitWidth).map(_.U.asTypeOf(new SqPtr))))
283  val addrReadyPtrExt = RegInit(0.U.asTypeOf(new SqPtr))
284  val dataReadyPtrExt = RegInit(0.U.asTypeOf(new SqPtr))
285
286  val enqPtr = enqPtrExt(0).value
287  val deqPtr = deqPtrExt(0).value
288  val cmtPtr = cmtPtrExt(0).value
289
290  val validCount = distanceBetween(enqPtrExt(0), deqPtrExt(0))
291  val allowEnqueue = validCount <= (StoreQueueSize - LSQStEnqWidth).U
292
293  val deqMask = UIntToMask(deqPtr, StoreQueueSize)
294  val enqMask = UIntToMask(enqPtr, StoreQueueSize)
295
296  val commitCount = WireInit(0.U(log2Ceil(CommitWidth + 1).W))
297  val scommit = GatedRegNext(io.rob.scommit)
298
299  // RegNext misalign control for better timing
300  val doMisalignSt = GatedValidRegNext((rdataPtrExt(0).value === deqPtr) && (cmtPtr === deqPtr) && allocated(deqPtr) && datavalid(deqPtr) && unaligned(deqPtr))
301  val finishMisalignSt = GatedValidRegNext(doMisalignSt && io.maControl.control.removeSq && !io.maControl.control.hasException)
302  val misalignBlock = doMisalignSt && !finishMisalignSt
303
304  val mmioReq = Wire(chiselTypeOf(io.uncache.req))
305  val ncReq = Wire(chiselTypeOf(io.uncache.req))
306  val ncResp = Wire(chiselTypeOf(io.uncache.resp))
307  val ncDoReq = Wire(Bool())
308  val ncDoResp = Wire(Bool())
309  val ncReadNextTrigger = Mux(io.uncacheOutstanding, ncDoReq, ncDoResp)
310  // ncDoReq is double RegNexted, as ubuffer data write takes 3 cycles.
311  // TODO lyq: to eliminate coupling by passing signals through ubuffer
312  val ncDeqTrigger = Mux(io.uncacheOutstanding, RegNext(RegNext(ncDoReq)), ncDoResp)
313  val ncPtr = Mux(io.uncacheOutstanding, RegNext(RegNext(io.uncache.req.bits.id)), io.uncache.resp.bits.id)
314
315  // store miss align info
316  io.maControl.storeInfo.data := dataModule.io.rdata(0).data
317  io.maControl.storeInfo.dataReady := doMisalignSt
318  io.maControl.storeInfo.completeSbTrans := doMisalignSt && dataBuffer.io.enq(0).fire
319
320  // store can be committed by ROB
321  io.rob.mmio := DontCare
322  io.rob.uop := DontCare
323
324  // Read dataModule
325  assert(EnsbufferWidth <= 2)
326  // rdataPtrExtNext and rdataPtrExtNext+1 entry will be read from dataModule
327  val rdataPtrExtNext = Wire(Vec(EnsbufferWidth, new SqPtr))
328  rdataPtrExtNext := rdataPtrExt.map(i => i +
329    PopCount(dataBuffer.io.enq.map(_.fire)) +
330    PopCount(ncReadNextTrigger || io.mmioStout.fire || io.vecmmioStout.fire)
331  )
332
333  // deqPtrExtNext traces which inst is about to leave store queue
334  //
335  // io.sbuffer(i).fire is RegNexted, as sbuffer data write takes 2 cycles.
336  // Before data write finish, sbuffer is unable to provide store to load
337  // forward data. As an workaround, deqPtrExt and allocated flag update
338  // is delayed so that load can get the right data from store queue.
339  //
340  // Modify deqPtrExtNext and io.sqDeq with care!
341  val deqPtrExtNext = Wire(Vec(EnsbufferWidth, new SqPtr))
342  deqPtrExtNext := deqPtrExt.map(i =>  i +
343    RegNext(PopCount(VecInit(io.sbuffer.map(_.fire)))) +
344    PopCount(ncDeqTrigger || io.mmioStout.fire || io.vecmmioStout.fire)
345  )
346
347  io.sqDeq := RegNext(
348    RegNext(PopCount(VecInit(io.sbuffer.map(_.fire && !misalignBlock)))) +
349    PopCount(ncDeqTrigger || io.mmioStout.fire || io.vecmmioStout.fire || finishMisalignSt)
350  )
351
352  assert(!RegNext(RegNext(io.sbuffer(0).fire) && (io.mmioStout.fire || io.vecmmioStout.fire)))
353
354  for (i <- 0 until EnsbufferWidth) {
355    dataModule.io.raddr(i) := rdataPtrExtNext(i).value
356    paddrModule.io.raddr(i) := rdataPtrExtNext(i).value
357    vaddrModule.io.raddr(i) := rdataPtrExtNext(i).value
358  }
359
360  /**
361    * Enqueue at dispatch
362    *
363    * Currently, StoreQueue only allows enqueue when #emptyEntries > EnqWidth
364    */
365  io.enq.canAccept := allowEnqueue
366  val canEnqueue = io.enq.req.map(_.valid)
367  val enqCancel = io.enq.req.map(_.bits.robIdx.needFlush(io.brqRedirect))
368  val vStoreFlow = io.enq.req.map(_.bits.numLsElem)
369  val validVStoreFlow = vStoreFlow.zipWithIndex.map{case (vLoadFlowNumItem, index) => Mux(!RegNext(io.brqRedirect.valid) && canEnqueue(index), vLoadFlowNumItem, 0.U)}
370  val validVStoreOffset = vStoreFlow.zip(io.enq.needAlloc).map{case (flow, needAllocItem) => Mux(needAllocItem, flow, 0.U)}
371  val validVStoreOffsetRShift = 0.U +: validVStoreOffset.take(vStoreFlow.length - 1)
372
373  for (i <- 0 until io.enq.req.length) {
374    val sqIdx = enqPtrExt(0) + validVStoreOffsetRShift.take(i + 1).reduce(_ + _)
375    val index = io.enq.req(i).bits.sqIdx
376    val enqInstr = io.enq.req(i).bits.instr.asTypeOf(new XSInstBitFields)
377    when (canEnqueue(i) && !enqCancel(i)) {
378      // The maximum 'numLsElem' number that can be emitted per dispatch port is:
379      //    16 2 2 2 2 2.
380      // Therefore, VecMemLSQEnqIteratorNumberSeq = Seq(16, 2, 2, 2, 2, 2)
381      for (j <- 0 until VecMemLSQEnqIteratorNumberSeq(i)) {
382        when (j.U < validVStoreOffset(i)) {
383          uop((index + j.U).value) := io.enq.req(i).bits
384          // NOTE: the index will be used when replay
385          uop((index + j.U).value).sqIdx := sqIdx + j.U
386          vecLastFlow((index + j.U).value) := Mux((j + 1).U === validVStoreOffset(i), io.enq.req(i).bits.lastUop, false.B)
387          allocated((index + j.U).value) := true.B
388          datavalid((index + j.U).value) := false.B
389          addrvalid((index + j.U).value) := false.B
390          unaligned((index + j.U).value) := false.B
391          committed((index + j.U).value) := false.B
392          pending((index + j.U).value) := false.B
393          prefetch((index + j.U).value) := false.B
394          nc((index + j.U).value) := false.B
395          mmio((index + j.U).value) := false.B
396          isVec((index + j.U).value) :=  FuType.isVStore(io.enq.req(i).bits.fuType)
397          vecMbCommit((index + j.U).value) := false.B
398          vecDataValid((index + j.U).value) := false.B
399          hasException((index + j.U).value) := false.B
400          waitStoreS2((index + j.U).value) := true.B
401          XSError(!io.enq.canAccept || !io.enq.lqCanAccept, s"must accept $i\n")
402          XSError(index.value =/= sqIdx.value, s"must be the same entry $i\n")
403        }
404      }
405    }
406    io.enq.resp(i) := sqIdx
407  }
408  XSDebug(p"(ready, valid): ${io.enq.canAccept}, ${Binary(Cat(io.enq.req.map(_.valid)))}\n")
409
410  /**
411    * Update addr/dataReadyPtr when issue from rs
412    */
413  // update issuePtr
414  val IssuePtrMoveStride = 4
415  require(IssuePtrMoveStride >= 2)
416
417  val addrReadyLookupVec = (0 until IssuePtrMoveStride).map(addrReadyPtrExt + _.U)
418  val addrReadyLookup = addrReadyLookupVec.map(ptr => allocated(ptr.value) &&
419   (mmio(ptr.value) || addrvalid(ptr.value) || vecMbCommit(ptr.value))
420    && ptr =/= enqPtrExt(0))
421  val nextAddrReadyPtr = addrReadyPtrExt + PriorityEncoder(VecInit(addrReadyLookup.map(!_) :+ true.B))
422  addrReadyPtrExt := nextAddrReadyPtr
423
424  val stAddrReadyVecReg = Wire(Vec(StoreQueueSize, Bool()))
425  (0 until StoreQueueSize).map(i => {
426    stAddrReadyVecReg(i) := allocated(i) && (mmio(i) || addrvalid(i) || (isVec(i) && vecMbCommit(i)))
427  })
428  io.stAddrReadyVec := GatedValidRegNext(stAddrReadyVecReg)
429
430  when (io.brqRedirect.valid) {
431    addrReadyPtrExt := Mux(
432      isAfter(cmtPtrExt(0), deqPtrExt(0)),
433      cmtPtrExt(0),
434      deqPtrExtNext(0) // for mmio insts, deqPtr may be ahead of cmtPtr
435    )
436  }
437
438  io.stAddrReadySqPtr := addrReadyPtrExt
439
440  // update
441  val dataReadyLookupVec = (0 until IssuePtrMoveStride).map(dataReadyPtrExt + _.U)
442  val dataReadyLookup = dataReadyLookupVec.map(ptr => allocated(ptr.value) &&
443   (mmio(ptr.value) || datavalid(ptr.value) || vecMbCommit(ptr.value))
444    && ptr =/= enqPtrExt(0))
445  val nextDataReadyPtr = dataReadyPtrExt + PriorityEncoder(VecInit(dataReadyLookup.map(!_) :+ true.B))
446  dataReadyPtrExt := nextDataReadyPtr
447
448  val stDataReadyVecReg = Wire(Vec(StoreQueueSize, Bool()))
449  (0 until StoreQueueSize).map(i => {
450    stDataReadyVecReg(i) := allocated(i) && (mmio(i) || datavalid(i) || (isVec(i) && vecMbCommit(i)))
451  })
452  io.stDataReadyVec := GatedValidRegNext(stDataReadyVecReg)
453
454  when (io.brqRedirect.valid) {
455    dataReadyPtrExt := Mux(
456      isAfter(cmtPtrExt(0), deqPtrExt(0)),
457      cmtPtrExt(0),
458      deqPtrExtNext(0) // for mmio insts, deqPtr may be ahead of cmtPtr
459    )
460  }
461
462  io.stDataReadySqPtr := dataReadyPtrExt
463  io.stIssuePtr := enqPtrExt(0)
464  io.sqDeqPtr := deqPtrExt(0)
465
466  /**
467    * Writeback store from store units
468    *
469    * Most store instructions writeback to regfile in the previous cycle.
470    * However,
471    *   (1) For an mmio instruction with exceptions, we need to mark it as addrvalid
472    * (in this way it will trigger an exception when it reaches ROB's head)
473    * instead of pending to avoid sending them to lower level.
474    *   (2) For an mmio instruction without exceptions, we mark it as pending.
475    * When the instruction reaches ROB's head, StoreQueue sends it to uncache channel.
476    * Upon receiving the response, StoreQueue writes back the instruction
477    * through arbiter with store units. It will later commit as normal.
478    */
479
480  // Write addr to sq
481  for (i <- 0 until StorePipelineWidth) {
482    paddrModule.io.wen(i) := false.B
483    vaddrModule.io.wen(i) := false.B
484    dataModule.io.mask.wen(i) := false.B
485    val stWbIndex = io.storeAddrIn(i).bits.uop.sqIdx.value
486    exceptionBuffer.io.storeAddrIn(i).valid := io.storeAddrIn(i).fire && !io.storeAddrIn(i).bits.miss && !io.storeAddrIn(i).bits.isvec
487    exceptionBuffer.io.storeAddrIn(i).bits := io.storeAddrIn(i).bits
488    // will re-enter exceptionbuffer at store_s2
489    exceptionBuffer.io.storeAddrIn(StorePipelineWidth + i).valid := false.B
490    exceptionBuffer.io.storeAddrIn(StorePipelineWidth + i).bits := 0.U.asTypeOf(new LsPipelineBundle)
491
492    when (io.storeAddrIn(i).fire) {
493      val addr_valid = !io.storeAddrIn(i).bits.miss
494      addrvalid(stWbIndex) := addr_valid //!io.storeAddrIn(i).bits.mmio
495      nc(stWbIndex) := io.storeAddrIn(i).bits.nc
496      // pending(stWbIndex) := io.storeAddrIn(i).bits.mmio
497      unaligned(stWbIndex) := io.storeAddrIn(i).bits.uop.exceptionVec(storeAddrMisaligned) && !io.storeAddrIn(i).bits.isvec
498
499      paddrModule.io.waddr(i) := stWbIndex
500      paddrModule.io.wdata(i) := io.storeAddrIn(i).bits.paddr
501      paddrModule.io.wmask(i) := io.storeAddrIn(i).bits.mask
502      paddrModule.io.wlineflag(i) := io.storeAddrIn(i).bits.wlineflag
503      paddrModule.io.wen(i) := true.B
504
505      vaddrModule.io.waddr(i) := stWbIndex
506      vaddrModule.io.wdata(i) := io.storeAddrIn(i).bits.vaddr
507      vaddrModule.io.wmask(i) := io.storeAddrIn(i).bits.mask
508      vaddrModule.io.wlineflag(i) := io.storeAddrIn(i).bits.wlineflag
509      vaddrModule.io.wen(i) := true.B
510
511      debug_paddr(paddrModule.io.waddr(i)) := paddrModule.io.wdata(i)
512
513      // mmio(stWbIndex) := io.storeAddrIn(i).bits.mmio
514
515      uop(stWbIndex) := io.storeAddrIn(i).bits.uop
516      uop(stWbIndex).debugInfo := io.storeAddrIn(i).bits.uop.debugInfo
517
518      vecDataValid(stWbIndex) := io.storeAddrIn(i).bits.isvec
519
520      XSInfo("store addr write to sq idx %d pc 0x%x miss:%d vaddr %x paddr %x mmio %x isvec %x\n",
521        io.storeAddrIn(i).bits.uop.sqIdx.value,
522        io.storeAddrIn(i).bits.uop.pc,
523        io.storeAddrIn(i).bits.miss,
524        io.storeAddrIn(i).bits.vaddr,
525        io.storeAddrIn(i).bits.paddr,
526        io.storeAddrIn(i).bits.mmio,
527        io.storeAddrIn(i).bits.isvec
528      )
529    }
530
531    // re-replinish mmio, for pma/pmp will get mmio one cycle later
532    val storeAddrInFireReg = RegNext(io.storeAddrIn(i).fire && !io.storeAddrIn(i).bits.miss)
533    //val stWbIndexReg = RegNext(stWbIndex)
534    val stWbIndexReg = RegEnable(stWbIndex, io.storeAddrIn(i).fire)
535    when (storeAddrInFireReg) {
536      pending(stWbIndexReg) := io.storeAddrInRe(i).mmio
537      mmio(stWbIndexReg) := io.storeAddrInRe(i).mmio
538      atomic(stWbIndexReg) := io.storeAddrInRe(i).atomic
539      hasException(stWbIndexReg) := ExceptionNO.selectByFu(uop(stWbIndexReg).exceptionVec, StaCfg).asUInt.orR ||
540                                    TriggerAction.isDmode(uop(stWbIndexReg).trigger) || io.storeAddrInRe(i).af
541      waitStoreS2(stWbIndexReg) := false.B
542    }
543    // dcache miss info (one cycle later than storeIn)
544    // if dcache report a miss in sta pipeline, this store will trigger a prefetch when committing to sbuffer (if EnableAtCommitMissTrigger)
545    when (storeAddrInFireReg) {
546      prefetch(stWbIndexReg) := io.storeAddrInRe(i).miss
547    }
548    // enter exceptionbuffer again
549    when (storeAddrInFireReg) {
550      exceptionBuffer.io.storeAddrIn(StorePipelineWidth + i).valid := io.storeAddrInRe(i).af && !io.storeAddrInRe(i).isvec
551      exceptionBuffer.io.storeAddrIn(StorePipelineWidth + i).bits := RegEnable(io.storeAddrIn(i).bits, io.storeAddrIn(i).fire && !io.storeAddrIn(i).bits.miss)
552      exceptionBuffer.io.storeAddrIn(StorePipelineWidth + i).bits.uop.exceptionVec(storeAccessFault) := io.storeAddrInRe(i).af
553    }
554
555    when(vaddrModule.io.wen(i)){
556      debug_vaddr(vaddrModule.io.waddr(i)) := vaddrModule.io.wdata(i)
557    }
558  }
559
560  // Write data to sq
561  // Now store data pipeline is actually 2 stages
562  for (i <- 0 until StorePipelineWidth) {
563    dataModule.io.data.wen(i) := false.B
564    val stWbIndex = io.storeDataIn(i).bits.uop.sqIdx.value
565    val isVec     = FuType.isVStore(io.storeDataIn(i).bits.uop.fuType)
566    // sq data write takes 2 cycles:
567    // sq data write s0
568    when (io.storeDataIn(i).fire) {
569      // send data write req to data module
570      dataModule.io.data.waddr(i) := stWbIndex
571      dataModule.io.data.wdata(i) := Mux(io.storeDataIn(i).bits.uop.fuOpType === LSUOpType.cbo_zero,
572        0.U,
573        Mux(isVec,
574          io.storeDataIn(i).bits.data,
575          genVWdata(io.storeDataIn(i).bits.data, io.storeDataIn(i).bits.uop.fuOpType(2,0)))
576      )
577      dataModule.io.data.wen(i) := true.B
578
579      debug_data(dataModule.io.data.waddr(i)) := dataModule.io.data.wdata(i)
580
581      XSInfo("store data write to sq idx %d pc 0x%x data %x -> %x\n",
582        io.storeDataIn(i).bits.uop.sqIdx.value,
583        io.storeDataIn(i).bits.uop.pc,
584        io.storeDataIn(i).bits.data,
585        dataModule.io.data.wdata(i)
586      )
587    }
588    // sq data write s1
589    when (
590      RegNext(io.storeDataIn(i).fire) && allocated(RegEnable(stWbIndex, io.storeDataIn(i).fire))
591      // && !RegNext(io.storeDataIn(i).bits.uop).robIdx.needFlush(io.brqRedirect)
592    ) {
593      datavalid(RegEnable(stWbIndex, io.storeDataIn(i).fire)) := true.B
594    }
595  }
596
597  // Write mask to sq
598  for (i <- 0 until StorePipelineWidth) {
599    // sq mask write s0
600    when (io.storeMaskIn(i).fire) {
601      // send data write req to data module
602      dataModule.io.mask.waddr(i) := io.storeMaskIn(i).bits.sqIdx.value
603      dataModule.io.mask.wdata(i) := io.storeMaskIn(i).bits.mask
604      dataModule.io.mask.wen(i) := true.B
605    }
606  }
607
608  /**
609    * load forward query
610    *
611    * Check store queue for instructions that is older than the load.
612    * The response will be valid at the next cycle after req.
613    */
614  // check over all lq entries and forward data from the first matched store
615  for (i <- 0 until LoadPipelineWidth) {
616    // Compare deqPtr (deqPtr) and forward.sqIdx, we have two cases:
617    // (1) if they have the same flag, we need to check range(tail, sqIdx)
618    // (2) if they have different flags, we need to check range(tail, VirtualLoadQueueSize) and range(0, sqIdx)
619    // Forward1: Mux(same_flag, range(tail, sqIdx), range(tail, VirtualLoadQueueSize))
620    // Forward2: Mux(same_flag, 0.U,                   range(0, sqIdx)    )
621    // i.e. forward1 is the target entries with the same flag bits and forward2 otherwise
622    val differentFlag = deqPtrExt(0).flag =/= io.forward(i).sqIdx.flag
623    val forwardMask = io.forward(i).sqIdxMask
624    // all addrvalid terms need to be checked
625    // Real Vaild: all scalar stores, and vector store with (!inactive && !secondInvalid)
626    val addrRealValidVec = WireInit(VecInit((0 until StoreQueueSize).map(j => addrvalid(j) && allocated(j))))
627    // vector store will consider all inactive || secondInvalid flows as valid
628    val addrValidVec = WireInit(VecInit((0 until StoreQueueSize).map(j => addrvalid(j) && allocated(j))))
629    val dataValidVec = WireInit(VecInit((0 until StoreQueueSize).map(j => datavalid(j))))
630    val allValidVec  = WireInit(VecInit((0 until StoreQueueSize).map(j => addrvalid(j) && datavalid(j) && allocated(j))))
631
632    val lfstEnable = Constantin.createRecord("LFSTEnable", LFSTEnable)
633    val storeSetHitVec = Mux(lfstEnable,
634      WireInit(VecInit((0 until StoreQueueSize).map(j => io.forward(i).uop.loadWaitBit && uop(j).robIdx === io.forward(i).uop.waitForRobIdx))),
635      WireInit(VecInit((0 until StoreQueueSize).map(j => uop(j).storeSetHit && uop(j).ssid === io.forward(i).uop.ssid)))
636    )
637
638    val forwardMask1 = Mux(differentFlag, ~deqMask, deqMask ^ forwardMask)
639    val forwardMask2 = Mux(differentFlag, forwardMask, 0.U(StoreQueueSize.W))
640    val canForward1 = forwardMask1 & allValidVec.asUInt
641    val canForward2 = forwardMask2 & allValidVec.asUInt
642    val needForward = Mux(differentFlag, ~deqMask | forwardMask, deqMask ^ forwardMask)
643
644    XSDebug(p"$i f1 ${Binary(canForward1)} f2 ${Binary(canForward2)} " +
645      p"sqIdx ${io.forward(i).sqIdx} pa ${Hexadecimal(io.forward(i).paddr)}\n"
646    )
647
648    // do real fwd query (cam lookup in load_s1)
649    dataModule.io.needForward(i)(0) := canForward1 & vaddrModule.io.forwardMmask(i).asUInt
650    dataModule.io.needForward(i)(1) := canForward2 & vaddrModule.io.forwardMmask(i).asUInt
651
652    vaddrModule.io.forwardMdata(i) := io.forward(i).vaddr
653    vaddrModule.io.forwardDataMask(i) := io.forward(i).mask
654    paddrModule.io.forwardMdata(i) := io.forward(i).paddr
655    paddrModule.io.forwardDataMask(i) := io.forward(i).mask
656
657    // vaddr cam result does not equal to paddr cam result
658    // replay needed
659    // val vpmaskNotEqual = ((paddrModule.io.forwardMmask(i).asUInt ^ vaddrModule.io.forwardMmask(i).asUInt) & needForward) =/= 0.U
660    // val vaddrMatchFailed = vpmaskNotEqual && io.forward(i).valid
661    val vpmaskNotEqual = (
662      (RegEnable(paddrModule.io.forwardMmask(i).asUInt, io.forward(i).valid) ^ RegEnable(vaddrModule.io.forwardMmask(i).asUInt, io.forward(i).valid)) &
663      RegNext(needForward) &
664      GatedRegNext(addrRealValidVec.asUInt)
665    ) =/= 0.U
666    val vaddrMatchFailed = vpmaskNotEqual && RegNext(io.forward(i).valid)
667    when (vaddrMatchFailed) {
668      XSInfo("vaddrMatchFailed: pc %x pmask %x vmask %x\n",
669        RegEnable(io.forward(i).uop.pc, io.forward(i).valid),
670        RegEnable(needForward & paddrModule.io.forwardMmask(i).asUInt, io.forward(i).valid),
671        RegEnable(needForward & vaddrModule.io.forwardMmask(i).asUInt, io.forward(i).valid)
672      );
673    }
674    XSPerfAccumulate("vaddr_match_failed", vpmaskNotEqual)
675    XSPerfAccumulate("vaddr_match_really_failed", vaddrMatchFailed)
676
677    // Fast forward mask will be generated immediately (load_s1)
678    io.forward(i).forwardMaskFast := dataModule.io.forwardMaskFast(i)
679
680    // Forward result will be generated 1 cycle later (load_s2)
681    io.forward(i).forwardMask := dataModule.io.forwardMask(i)
682    io.forward(i).forwardData := dataModule.io.forwardData(i)
683    // If addr match, data not ready, mark it as dataInvalid
684    // load_s1: generate dataInvalid in load_s1 to set fastUop
685    val dataInvalidMask1 = (addrValidVec.asUInt & ~dataValidVec.asUInt & vaddrModule.io.forwardMmask(i).asUInt & forwardMask1.asUInt)
686    val dataInvalidMask2 = (addrValidVec.asUInt & ~dataValidVec.asUInt & vaddrModule.io.forwardMmask(i).asUInt & forwardMask2.asUInt)
687    val dataInvalidMask = dataInvalidMask1 | dataInvalidMask2
688    io.forward(i).dataInvalidFast := dataInvalidMask.orR
689
690    // make chisel happy
691    val dataInvalidMask1Reg = Wire(UInt(StoreQueueSize.W))
692    dataInvalidMask1Reg := RegNext(dataInvalidMask1)
693    // make chisel happy
694    val dataInvalidMask2Reg = Wire(UInt(StoreQueueSize.W))
695    dataInvalidMask2Reg := RegNext(dataInvalidMask2)
696    val dataInvalidMaskReg = dataInvalidMask1Reg | dataInvalidMask2Reg
697
698    // If SSID match, address not ready, mark it as addrInvalid
699    // load_s2: generate addrInvalid
700    val addrInvalidMask1 = (~addrValidVec.asUInt & storeSetHitVec.asUInt & forwardMask1.asUInt)
701    val addrInvalidMask2 = (~addrValidVec.asUInt & storeSetHitVec.asUInt & forwardMask2.asUInt)
702    // make chisel happy
703    val addrInvalidMask1Reg = Wire(UInt(StoreQueueSize.W))
704    addrInvalidMask1Reg := RegNext(addrInvalidMask1)
705    // make chisel happy
706    val addrInvalidMask2Reg = Wire(UInt(StoreQueueSize.W))
707    addrInvalidMask2Reg := RegNext(addrInvalidMask2)
708    val addrInvalidMaskReg = addrInvalidMask1Reg | addrInvalidMask2Reg
709
710    // load_s2
711    io.forward(i).dataInvalid := RegNext(io.forward(i).dataInvalidFast)
712    // check if vaddr forward mismatched
713    io.forward(i).matchInvalid := vaddrMatchFailed
714
715    // data invalid sq index
716    // check whether false fail
717    // check flag
718    val s2_differentFlag = RegNext(differentFlag)
719    val s2_enqPtrExt = RegNext(enqPtrExt(0))
720    val s2_deqPtrExt = RegNext(deqPtrExt(0))
721
722    // addr invalid sq index
723    // make chisel happy
724    val addrInvalidMaskRegWire = Wire(UInt(StoreQueueSize.W))
725    addrInvalidMaskRegWire := addrInvalidMaskReg
726    val addrInvalidFlag = addrInvalidMaskRegWire.orR
727    val hasInvalidAddr = (~addrValidVec.asUInt & needForward).orR
728
729    val addrInvalidSqIdx1 = OHToUInt(Reverse(PriorityEncoderOH(Reverse(addrInvalidMask1Reg))))
730    val addrInvalidSqIdx2 = OHToUInt(Reverse(PriorityEncoderOH(Reverse(addrInvalidMask2Reg))))
731    val addrInvalidSqIdx = Mux(addrInvalidMask2Reg.orR, addrInvalidSqIdx2, addrInvalidSqIdx1)
732
733    // store-set content management
734    //                +-----------------------+
735    //                | Search a SSID for the |
736    //                |    load operation     |
737    //                +-----------------------+
738    //                           |
739    //                           V
740    //                 +-------------------+
741    //                 | load wait strict? |
742    //                 +-------------------+
743    //                           |
744    //                           V
745    //               +----------------------+
746    //            Set|                      |Clean
747    //               V                      V
748    //  +------------------------+   +------------------------------+
749    //  | Waiting for all older  |   | Wait until the corresponding |
750    //  |   stores operations    |   | older store operations       |
751    //  +------------------------+   +------------------------------+
752
753
754
755    when (RegEnable(io.forward(i).uop.loadWaitStrict, io.forward(i).valid)) {
756      io.forward(i).addrInvalidSqIdx := RegEnable((io.forward(i).uop.sqIdx - 1.U), io.forward(i).valid)
757    } .elsewhen (addrInvalidFlag) {
758      io.forward(i).addrInvalidSqIdx.flag := Mux(!s2_differentFlag || addrInvalidSqIdx >= s2_deqPtrExt.value, s2_deqPtrExt.flag, s2_enqPtrExt.flag)
759      io.forward(i).addrInvalidSqIdx.value := addrInvalidSqIdx
760    } .otherwise {
761      // may be store inst has been written to sbuffer already.
762      io.forward(i).addrInvalidSqIdx := RegEnable(io.forward(i).uop.sqIdx, io.forward(i).valid)
763    }
764    io.forward(i).addrInvalid := Mux(RegEnable(io.forward(i).uop.loadWaitStrict, io.forward(i).valid), RegNext(hasInvalidAddr), addrInvalidFlag)
765
766    // data invalid sq index
767    // make chisel happy
768    val dataInvalidMaskRegWire = Wire(UInt(StoreQueueSize.W))
769    dataInvalidMaskRegWire := dataInvalidMaskReg
770    val dataInvalidFlag = dataInvalidMaskRegWire.orR
771
772    val dataInvalidSqIdx1 = OHToUInt(Reverse(PriorityEncoderOH(Reverse(dataInvalidMask1Reg))))
773    val dataInvalidSqIdx2 = OHToUInt(Reverse(PriorityEncoderOH(Reverse(dataInvalidMask2Reg))))
774    val dataInvalidSqIdx = Mux(dataInvalidMask2Reg.orR, dataInvalidSqIdx2, dataInvalidSqIdx1)
775
776    when (dataInvalidFlag) {
777      io.forward(i).dataInvalidSqIdx.flag := Mux(!s2_differentFlag || dataInvalidSqIdx >= s2_deqPtrExt.value, s2_deqPtrExt.flag, s2_enqPtrExt.flag)
778      io.forward(i).dataInvalidSqIdx.value := dataInvalidSqIdx
779    } .otherwise {
780      // may be store inst has been written to sbuffer already.
781      io.forward(i).dataInvalidSqIdx := RegEnable(io.forward(i).uop.sqIdx, io.forward(i).valid)
782    }
783  }
784
785  /**
786    * Memory mapped IO / other uncached operations / CMO
787    *
788    * States:
789    * (1) writeback from store units: mark as pending
790    * (2) when they reach ROB's head, they can be sent to uncache channel
791    * (3) response from uncache channel: mark as datavalidmask.wen
792    * (4) writeback to ROB (and other units): mark as writebacked
793    * (5) ROB commits the instruction: same as normal instructions
794    */
795  //(2) when they reach ROB's head, they can be sent to uncache channel
796  // TODO: CAN NOT deal with vector mmio now!
797  val s_idle :: s_req :: s_resp :: s_wb :: s_wait :: Nil = Enum(5)
798  val mmioState = RegInit(s_idle)
799  val uncacheUop = Reg(new DynInst)
800  val uncacheVAddr = Reg(UInt(VAddrBits.W))
801  val cboFlushedSb = RegInit(false.B)
802  val cmoOpCode = uncacheUop.fuOpType(1, 0)
803  val mmioDoReq = io.uncache.req.fire && !io.uncache.req.bits.nc
804  switch(mmioState) {
805    is(s_idle) {
806      when(RegNext(io.rob.pendingst && uop(deqPtr).robIdx === io.rob.pendingPtr && pending(deqPtr) && allocated(deqPtr) && datavalid(deqPtr) && addrvalid(deqPtr))) {
807        mmioState := s_req
808        uncacheUop := uop(deqPtr)
809        cboFlushedSb := false.B
810      }
811    }
812    is(s_req) {
813      when (mmioDoReq) {
814        mmioState := s_resp
815      }
816    }
817    is(s_resp) {
818      when(io.uncache.resp.fire && !io.uncache.resp.bits.nc) {
819        mmioState := s_wb
820
821        when (io.uncache.resp.bits.nderr) {
822          uncacheUop.exceptionVec(storeAccessFault) := true.B
823        }
824      }
825    }
826    is(s_wb) {
827      when (io.mmioStout.fire || io.vecmmioStout.fire) {
828        when (uncacheUop.exceptionVec(storeAccessFault)) {
829          mmioState := s_idle
830        }.otherwise {
831          mmioState := s_wait
832        }
833      }
834    }
835    is(s_wait) {
836      // A MMIO store can always move cmtPtrExt as it must be ROB head
837      when(scommit > 0.U) {
838        mmioState := s_idle // ready for next mmio
839      }
840    }
841  }
842
843  mmioReq.valid := mmioState === s_req
844  mmioReq.bits := DontCare
845  mmioReq.bits.cmd  := MemoryOpConstants.M_XWR
846  mmioReq.bits.addr := paddrModule.io.rdata(0) // data(deqPtr) -> rdata(0)
847  mmioReq.bits.vaddr:= vaddrModule.io.rdata(0)
848  mmioReq.bits.data := shiftDataToLow(paddrModule.io.rdata(0), dataModule.io.rdata(0).data)
849  mmioReq.bits.mask := shiftMaskToLow(paddrModule.io.rdata(0), dataModule.io.rdata(0).mask)
850  mmioReq.bits.atomic := atomic(GatedRegNext(rdataPtrExtNext(0)).value)
851  mmioReq.bits.nc := false.B
852  mmioReq.bits.id := rdataPtrExt(0).value
853
854  /**
855    * NC Store
856    * (1) req: when it has been commited, it can be sent to lower level.
857    * (2) resp: because SQ data forward is required, it can only be deq when ncResp is received
858    */
859  // TODO: CAN NOT deal with vector nc now!
860  val nc_idle :: nc_req :: nc_resp :: Nil = Enum(3)
861  val ncState = RegInit(nc_idle)
862  val rptr0 = rdataPtrExt(0).value
863  switch(ncState){
864    is(nc_idle) {
865      when(nc(rptr0) && allocated(rptr0) && committed(rptr0) && !mmio(rptr0) && !isVec(rptr0)) {
866        ncState := nc_req
867      }
868    }
869    is(nc_req) {
870      when(ncDoReq) {
871        when(io.uncacheOutstanding) {
872          ncState := nc_idle
873        }.otherwise{
874          ncState := nc_resp
875        }
876      }
877    }
878    is(nc_resp) {
879      when(ncResp.fire) {
880        ncState := nc_idle
881      }
882    }
883  }
884
885  ncDoReq := io.uncache.req.fire && io.uncache.req.bits.nc
886  ncDoResp := ncResp.fire
887
888  ncReq.valid := ncState === nc_req
889  ncReq.bits := DontCare
890  ncReq.bits.cmd  := MemoryOpConstants.M_XWR
891  ncReq.bits.addr := paddrModule.io.rdata(0)
892  ncReq.bits.vaddr:= vaddrModule.io.rdata(0)
893  ncReq.bits.data := shiftDataToLow(paddrModule.io.rdata(0), dataModule.io.rdata(0).data)
894  ncReq.bits.mask := shiftMaskToLow(paddrModule.io.rdata(0), dataModule.io.rdata(0).mask)
895  ncReq.bits.atomic := atomic(GatedRegNext(rdataPtrExtNext(0)).value)
896  ncReq.bits.nc := true.B
897  ncReq.bits.id := rptr0
898
899  ncResp.ready := io.uncache.resp.ready
900  ncResp.valid := io.uncache.resp.fire && io.uncache.resp.bits.nc
901  ncResp.bits <> io.uncache.resp.bits
902  when (ncDeqTrigger) {
903    allocated(ncPtr) := false.B
904    XSDebug("nc fire: ptr %d\n", ncPtr)
905  }
906
907  mmioReq.ready := io.uncache.req.ready
908  ncReq.ready := io.uncache.req.ready && !mmioReq.valid
909  io.uncache.req.valid := mmioReq.valid || ncReq.valid
910  io.uncache.req.bits := Mux(mmioReq.valid, mmioReq.bits, ncReq.bits)
911
912  // CBO op type check can be delayed for 1 cycle,
913  // as uncache op will not start in s_idle
914  val cboMmioAddr = get_block_addr(paddrModule.io.rdata(0))
915  val deqCanDoCbo = GatedRegNext(LSUOpType.isCbo(uop(deqPtr).fuOpType) && allocated(deqPtr) && addrvalid(deqPtr))
916  when (deqCanDoCbo) {
917    // disable uncache channel
918    io.uncache.req.valid := false.B
919
920    when (io.cmoOpReq.fire) {
921      mmioState := s_resp
922    }
923
924    when (mmioState === s_resp) {
925      when (io.cmoOpResp.fire) {
926        mmioState := s_wb
927      }
928    }
929  }
930
931  io.cmoOpReq.valid := deqCanDoCbo && cboFlushedSb && (mmioState === s_req)
932  io.cmoOpReq.bits.opcode  := cmoOpCode
933  io.cmoOpReq.bits.address := cboMmioAddr
934
935  io.cmoOpResp.ready := deqCanDoCbo && (mmioState === s_resp)
936
937  io.flushSbuffer.valid := deqCanDoCbo && !cboFlushedSb && (mmioState === s_req) && !io.flushSbuffer.empty
938
939  when(deqCanDoCbo && !cboFlushedSb && (mmioState === s_req) && io.flushSbuffer.empty) {
940    cboFlushedSb := true.B
941  }
942
943  when(mmioDoReq){
944    // mmio store should not be committed until uncache req is sent
945    pending(deqPtr) := false.B
946
947    XSDebug(
948      p"uncache mmio req: pc ${Hexadecimal(uop(deqPtr).pc)} " +
949      p"addr ${Hexadecimal(io.uncache.req.bits.addr)} " +
950      p"data ${Hexadecimal(io.uncache.req.bits.data)} " +
951      p"op ${Hexadecimal(io.uncache.req.bits.cmd)} " +
952      p"mask ${Hexadecimal(io.uncache.req.bits.mask)}\n"
953    )
954  }
955
956  // (3) response from uncache channel: mark as datavalid
957  io.uncache.resp.ready := true.B
958
959  // (4) scalar store: writeback to ROB (and other units): mark as writebacked
960  io.mmioStout.valid := mmioState === s_wb && !isVec(deqPtr)
961  io.mmioStout.bits.uop := uncacheUop
962  io.mmioStout.bits.uop.exceptionVec := ExceptionNO.selectByFu(uncacheUop.exceptionVec, StaCfg)
963  io.mmioStout.bits.uop.sqIdx := deqPtrExt(0)
964  io.mmioStout.bits.uop.flushPipe := deqCanDoCbo // flush Pipeline to keep order in CMO
965  io.mmioStout.bits.data := shiftDataToLow(paddrModule.io.rdata(0), dataModule.io.rdata(0).data) // dataModule.io.rdata.read(deqPtr)
966  io.mmioStout.bits.isFromLoadUnit := DontCare
967  io.mmioStout.bits.debug.isMMIO := true.B
968  io.mmioStout.bits.debug.isNC := false.B
969  io.mmioStout.bits.debug.paddr := DontCare
970  io.mmioStout.bits.debug.isPerfCnt := false.B
971  io.mmioStout.bits.debug.vaddr := DontCare
972  // Remove MMIO inst from store queue after MMIO request is being sent
973  // That inst will be traced by uncache state machine
974  when (io.mmioStout.fire) {
975    allocated(deqPtr) := false.B
976  }
977
978  exceptionBuffer.io.storeAddrIn.last.valid := io.mmioStout.fire
979  exceptionBuffer.io.storeAddrIn.last.bits := DontCare
980  exceptionBuffer.io.storeAddrIn.last.bits.fullva := vaddrModule.io.rdata.head
981  exceptionBuffer.io.storeAddrIn.last.bits.vaNeedExt := true.B
982  exceptionBuffer.io.storeAddrIn.last.bits.uop := uncacheUop
983
984  // (4) or vector store:
985  // TODO: implement it!
986  io.vecmmioStout := DontCare
987  io.vecmmioStout.valid := false.B //mmioState === s_wb && isVec(deqPtr)
988  io.vecmmioStout.bits.uop := uop(deqPtr)
989  io.vecmmioStout.bits.uop.sqIdx := deqPtrExt(0)
990  io.vecmmioStout.bits.data := shiftDataToLow(paddrModule.io.rdata(0), dataModule.io.rdata(0).data) // dataModule.io.rdata.read(deqPtr)
991  io.vecmmioStout.bits.debug.isMMIO := true.B
992  io.vecmmioStout.bits.debug.isNC   := false.B
993  io.vecmmioStout.bits.debug.paddr := DontCare
994  io.vecmmioStout.bits.debug.isPerfCnt := false.B
995  io.vecmmioStout.bits.debug.vaddr := DontCare
996  // Remove MMIO inst from store queue after MMIO request is being sent
997  // That inst will be traced by uncache state machine
998  when (io.vecmmioStout.fire) {
999    allocated(deqPtr) := false.B
1000  }
1001
1002  /**
1003    * ROB commits store instructions (mark them as committed)
1004    *
1005    * (1) When store commits, mark it as committed.
1006    * (2) They will not be cancelled and can be sent to lower level.
1007    */
1008  XSError(mmioState =/= s_idle && mmioState =/= s_wait && commitCount > 0.U,
1009   "should not commit instruction when MMIO has not been finished\n")
1010
1011  val commitVec = WireInit(VecInit(Seq.fill(CommitWidth)(false.B)))
1012  val needCancel = Wire(Vec(StoreQueueSize, Bool())) // Will be assigned later
1013
1014  if (backendParams.debugEn){ dontTouch(commitVec) }
1015
1016  // TODO: Deal with vector store mmio
1017  for (i <- 0 until CommitWidth) {
1018    // don't mark misalign store as committed
1019    when (
1020      allocated(cmtPtrExt(i).value) &&
1021      !unaligned(cmtPtrExt(i).value) &&
1022      isNotAfter(uop(cmtPtrExt(i).value).robIdx, GatedRegNext(io.rob.pendingPtr)) &&
1023      !needCancel(cmtPtrExt(i).value) &&
1024      (!waitStoreS2(cmtPtrExt(i).value) || isVec(cmtPtrExt(i).value))) {
1025      if (i == 0){
1026        // TODO: fixme for vector mmio
1027        when ((mmioState === s_idle) || (mmioState === s_wait && scommit > 0.U)){
1028          when ((isVec(cmtPtrExt(i).value) && vecMbCommit(cmtPtrExt(i).value)) || !isVec(cmtPtrExt(i).value)) {
1029            committed(cmtPtrExt(0).value) := true.B
1030            commitVec(0) := true.B
1031          }
1032        }
1033      } else {
1034        when ((isVec(cmtPtrExt(i).value) && vecMbCommit(cmtPtrExt(i).value)) || !isVec(cmtPtrExt(i).value)) {
1035          committed(cmtPtrExt(i).value) := commitVec(i - 1) || committed(cmtPtrExt(i).value)
1036          commitVec(i) := commitVec(i - 1)
1037        }
1038      }
1039    }
1040  }
1041
1042  commitCount := PopCount(commitVec)
1043  cmtPtrExt := cmtPtrExt.map(_ + commitCount)
1044
1045  /**
1046   * committed stores will not be cancelled and can be sent to lower level.
1047   *
1048   * 1. Store NC: Read data to uncache
1049   *    implement as above
1050   *
1051   * 2. Store Cache: Read data from data module
1052   *    remove retired insts from sq, add retired store to sbuffer.
1053   *    as store queue grows larger and larger, time needed to read data from data
1054   *    module keeps growing higher. Now we give data read a whole cycle.
1055   */
1056  for (i <- 0 until EnsbufferWidth) {
1057    val ptr = rdataPtrExt(i).value
1058    val mmioStall = if(i == 0) mmio(rdataPtrExt(0).value) else (mmio(rdataPtrExt(i).value) || mmio(rdataPtrExt(i-1).value))
1059    val ncStall = if(i == 0) nc(rdataPtrExt(0).value) else (nc(rdataPtrExt(i).value) || nc(rdataPtrExt(i-1).value))
1060    val exceptionValid = if(i == 0) hasException(rdataPtrExt(0).value) else {
1061      hasException(rdataPtrExt(i).value) || (hasException(rdataPtrExt(i-1).value) && uop(rdataPtrExt(i).value).robIdx === uop(rdataPtrExt(i-1).value).robIdx)
1062    }
1063    val vecNotAllMask = dataModule.io.rdata(i).mask.orR
1064    // Vector instructions that prevent triggered exceptions from being written to the 'databuffer'.
1065    val vecHasExceptionFlagValid = vecExceptionFlag.valid && isVec(ptr) && vecExceptionFlag.bits.robIdx === uop(ptr).robIdx
1066    if (i == 0) {
1067      // use dataBuffer write port 0 to writeback missaligned store out
1068      dataBuffer.io.enq(i).valid := Mux(
1069        doMisalignSt,
1070        io.maControl.control.writeSb,
1071        allocated(ptr) && committed(ptr) && ((!isVec(ptr) && (allvalid(ptr) || hasException(ptr))) || vecMbCommit(ptr)) && !mmioStall && !ncStall
1072      )
1073    } else {
1074      dataBuffer.io.enq(i).valid := Mux(
1075        doMisalignSt,
1076        false.B,
1077        allocated(ptr) && committed(ptr) && ((!isVec(ptr) && (allvalid(ptr) || hasException(ptr))) || vecMbCommit(ptr)) && !mmioStall && !ncStall
1078      )
1079    }
1080    // Note that store data/addr should both be valid after store's commit
1081    assert(!dataBuffer.io.enq(i).valid || allvalid(ptr) || doMisalignSt || hasException(ptr) || (allocated(ptr) && vecMbCommit(ptr)))
1082    dataBuffer.io.enq(i).bits.addr     := Mux(doMisalignSt, io.maControl.control.paddr, paddrModule.io.rdata(i))
1083    dataBuffer.io.enq(i).bits.vaddr    := Mux(doMisalignSt, io.maControl.control.vaddr, vaddrModule.io.rdata(i))
1084    dataBuffer.io.enq(i).bits.data     := Mux(doMisalignSt, io.maControl.control.wdata, dataModule.io.rdata(i).data)
1085    dataBuffer.io.enq(i).bits.mask     := Mux(doMisalignSt, io.maControl.control.wmask, dataModule.io.rdata(i).mask)
1086    dataBuffer.io.enq(i).bits.wline    := Mux(doMisalignSt, false.B, paddrModule.io.rlineflag(i))
1087    dataBuffer.io.enq(i).bits.sqPtr    := rdataPtrExt(i)
1088    dataBuffer.io.enq(i).bits.prefetch := Mux(doMisalignSt, false.B, prefetch(ptr))
1089    // when scalar has exception, will also not write into sbuffer
1090    dataBuffer.io.enq(i).bits.vecValid := Mux(doMisalignSt, true.B, (!isVec(ptr) || (vecDataValid(ptr) && vecNotAllMask)) && !exceptionValid && !vecHasExceptionFlagValid)
1091//    dataBuffer.io.enq(i).bits.vecValid := (!isVec(ptr) || vecDataValid(ptr)) && !hasException(ptr)
1092  }
1093
1094  // Send data stored in sbufferReqBitsReg to sbuffer
1095  for (i <- 0 until EnsbufferWidth) {
1096    io.sbuffer(i).valid := dataBuffer.io.deq(i).valid
1097    dataBuffer.io.deq(i).ready := io.sbuffer(i).ready
1098    io.sbuffer(i).bits := DontCare
1099    io.sbuffer(i).bits.cmd   := MemoryOpConstants.M_XWR
1100    io.sbuffer(i).bits.addr  := dataBuffer.io.deq(i).bits.addr
1101    io.sbuffer(i).bits.vaddr := dataBuffer.io.deq(i).bits.vaddr
1102    io.sbuffer(i).bits.data  := dataBuffer.io.deq(i).bits.data
1103    io.sbuffer(i).bits.mask  := dataBuffer.io.deq(i).bits.mask
1104    io.sbuffer(i).bits.wline := dataBuffer.io.deq(i).bits.wline && dataBuffer.io.deq(i).bits.vecValid
1105    io.sbuffer(i).bits.prefetch := dataBuffer.io.deq(i).bits.prefetch
1106    io.sbuffer(i).bits.vecValid := dataBuffer.io.deq(i).bits.vecValid
1107    // io.sbuffer(i).fire is RegNexted, as sbuffer data write takes 2 cycles.
1108    // Before data write finish, sbuffer is unable to provide store to load
1109    // forward data. As an workaround, deqPtrExt and allocated flag update
1110    // is delayed so that load can get the right data from store queue.
1111    val ptr = dataBuffer.io.deq(i).bits.sqPtr.value
1112    when (RegNext(io.sbuffer(i).fire && !doMisalignSt)) {
1113      allocated(RegEnable(ptr, io.sbuffer(i).fire)) := false.B
1114      XSDebug("sbuffer "+i+" fire: ptr %d\n", ptr)
1115    }
1116  }
1117
1118  // All vector instruction uop normally dequeue, but the Uop after the exception is raised does not write to the 'sbuffer'.
1119  // Flags are used to record whether there are any exceptions when the queue is displayed.
1120  // This is determined each time a write is made to the 'databuffer', prevent subsequent uop of the same instruction from writing to the 'dataBuffer'.
1121  val vecCommitHasException = (0 until EnsbufferWidth).map{ i =>
1122    val ptr = rdataPtrExt(i).value
1123    val mmioStall = if(i == 0) mmio(rdataPtrExt(0).value) else (mmio(rdataPtrExt(i).value) || mmio(rdataPtrExt(i-1).value))
1124    val ncStall = if(i == 0) nc(rdataPtrExt(0).value) else (nc(rdataPtrExt(i).value) || nc(rdataPtrExt(i-1).value))
1125    val exceptionVliad      = isVec(ptr) && hasException(ptr) && dataBuffer.io.enq(i).fire
1126    (exceptionVliad, uop(ptr), vecLastFlow(ptr))
1127  }
1128
1129  val vecCommitHasExceptionValid      = vecCommitHasException.map(_._1)
1130  val vecCommitHasExceptionUop        = vecCommitHasException.map(_._2)
1131  val vecCommitHasExceptionLastFlow   = vecCommitHasException.map(_._3)
1132  val vecCommitHasExceptionValidOR    = vecCommitHasExceptionValid.reduce(_ || _)
1133  // Just select the last Uop tah has an exception.
1134  val vecCommitHasExceptionSelectUop  = ParallelPosteriorityMux(vecCommitHasExceptionValid, vecCommitHasExceptionUop)
1135  // If the last flow with an exception is the LastFlow of this instruction, the flag is not set.
1136  // compare robidx to select the last flow
1137  require(EnsbufferWidth == 2, "The vector store exception handle process only support EnsbufferWidth == 2 yet.")
1138  val robidxEQ = dataBuffer.io.enq(0).valid && dataBuffer.io.enq(1).valid &&
1139    uop(rdataPtrExt(0).value).robIdx === uop(rdataPtrExt(1).value).robIdx
1140  val robidxNE = dataBuffer.io.enq(0).valid && dataBuffer.io.enq(1).valid && (
1141    uop(rdataPtrExt(0).value).robIdx =/= uop(rdataPtrExt(1).value).robIdx
1142  )
1143  val onlyCommit0 = dataBuffer.io.enq(0).valid && !dataBuffer.io.enq(1).valid
1144
1145  val vecCommitLastFlow =
1146    // robidx equal => check if 1 is last flow
1147    robidxEQ && vecCommitHasExceptionLastFlow(1) ||
1148    // robidx not equal => 0 must be the last flow, just check if 1 is last flow when 1 has exception
1149    robidxNE && (vecCommitHasExceptionValid(1) && vecCommitHasExceptionLastFlow(1) || !vecCommitHasExceptionValid(1)) ||
1150    onlyCommit0 && vecCommitHasExceptionLastFlow(0)
1151
1152
1153  val vecExceptionFlagCancel  = (0 until EnsbufferWidth).map{ i =>
1154    val ptr = rdataPtrExt(i).value
1155    val vecLastFlowCommit = vecLastFlow(ptr) && (uop(ptr).robIdx === vecExceptionFlag.bits.robIdx) && dataBuffer.io.enq(i).fire
1156    vecLastFlowCommit
1157  }.reduce(_ || _)
1158
1159  // When a LastFlow with an exception instruction is commited, clear the flag.
1160  when(!vecExceptionFlag.valid && vecCommitHasExceptionValidOR && !vecCommitLastFlow) {
1161    vecExceptionFlag.valid  := true.B
1162    vecExceptionFlag.bits   := vecCommitHasExceptionSelectUop
1163  }.elsewhen(vecExceptionFlag.valid && vecExceptionFlagCancel) {
1164    vecExceptionFlag.valid  := false.B
1165    vecExceptionFlag.bits   := 0.U.asTypeOf(new DynInst)
1166  }
1167
1168  // A dumb defensive code. The flag should not be placed for a long period of time.
1169  // A relatively large timeout period, not have any special meaning.
1170  // If an assert appears and you confirm that it is not a Bug: Increase the timeout or remove the assert.
1171  TimeOutAssert(vecExceptionFlag.valid, 3000, "vecExceptionFlag timeout, Plase check for bugs or add timeouts.")
1172
1173  // Initialize when unenabled difftest.
1174  for (i <- 0 until EnsbufferWidth) {
1175    io.sbufferVecDifftestInfo(i) := DontCare
1176  }
1177  // Consistent with the logic above.
1178  // Only the vector store difftest required signal is separated from the rtl code.
1179  if (env.EnableDifftest) {
1180    for (i <- 0 until EnsbufferWidth) {
1181      val ptr = rdataPtrExt(i).value
1182      difftestBuffer.get.io.enq(i).valid := dataBuffer.io.enq(i).valid
1183      difftestBuffer.get.io.enq(i).bits := uop(ptr)
1184    }
1185    for (i <- 0 until EnsbufferWidth) {
1186      io.sbufferVecDifftestInfo(i).valid := difftestBuffer.get.io.deq(i).valid
1187      difftestBuffer.get.io.deq(i).ready := io.sbufferVecDifftestInfo(i).ready
1188
1189      io.sbufferVecDifftestInfo(i).bits := difftestBuffer.get.io.deq(i).bits
1190    }
1191
1192    // commit cbo.inval to difftest
1193    val cmoInvalEvent = DifftestModule(new DiffCMOInvalEvent)
1194    cmoInvalEvent.coreid := io.hartId
1195    cmoInvalEvent.valid  := io.mmioStout.fire && deqCanDoCbo && LSUOpType.isCboInval(uop(deqPtr).fuOpType)
1196    cmoInvalEvent.addr   := cboMmioAddr
1197  }
1198
1199  (1 until EnsbufferWidth).foreach(i => when(io.sbuffer(i).fire) { assert(io.sbuffer(i - 1).fire) })
1200  if (coreParams.dcacheParametersOpt.isEmpty) {
1201    for (i <- 0 until EnsbufferWidth) {
1202      val ptr = deqPtrExt(i).value
1203      val ram = DifftestMem(64L * 1024 * 1024 * 1024, 8)
1204      val wen = allocated(ptr) && committed(ptr) && !mmio(ptr)
1205      val waddr = ((paddrModule.io.rdata(i) - "h80000000".U) >> 3).asUInt
1206      val wdata = Mux(paddrModule.io.rdata(i)(3), dataModule.io.rdata(i).data(127, 64), dataModule.io.rdata(i).data(63, 0))
1207      val wmask = Mux(paddrModule.io.rdata(i)(3), dataModule.io.rdata(i).mask(15, 8), dataModule.io.rdata(i).mask(7, 0))
1208      when (wen) {
1209        ram.write(waddr, wdata.asTypeOf(Vec(8, UInt(8.W))), wmask.asBools)
1210      }
1211    }
1212  }
1213
1214  // Read vaddr for mem exception
1215  io.exceptionAddr.vaddr     := exceptionBuffer.io.exceptionAddr.vaddr
1216  io.exceptionAddr.vaNeedExt := exceptionBuffer.io.exceptionAddr.vaNeedExt
1217  io.exceptionAddr.isHyper   := exceptionBuffer.io.exceptionAddr.isHyper
1218  io.exceptionAddr.gpaddr    := exceptionBuffer.io.exceptionAddr.gpaddr
1219  io.exceptionAddr.vstart    := exceptionBuffer.io.exceptionAddr.vstart
1220  io.exceptionAddr.vl        := exceptionBuffer.io.exceptionAddr.vl
1221  io.exceptionAddr.isForVSnonLeafPTE := exceptionBuffer.io.exceptionAddr.isForVSnonLeafPTE
1222
1223  // vector commit or replay from
1224  val vecCommittmp = Wire(Vec(StoreQueueSize, Vec(VecStorePipelineWidth, Bool())))
1225  val vecCommit = Wire(Vec(StoreQueueSize, Bool()))
1226  for (i <- 0 until StoreQueueSize) {
1227    val fbk = io.vecFeedback
1228    for (j <- 0 until VecStorePipelineWidth) {
1229      vecCommittmp(i)(j) := fbk(j).valid && (fbk(j).bits.isCommit || fbk(j).bits.isFlush) &&
1230        uop(i).robIdx === fbk(j).bits.robidx && uop(i).uopIdx === fbk(j).bits.uopidx && allocated(i)
1231    }
1232    vecCommit(i) := vecCommittmp(i).reduce(_ || _)
1233
1234    when (vecCommit(i)) {
1235      vecMbCommit(i) := true.B
1236    }
1237  }
1238
1239  // misprediction recovery / exception redirect
1240  // invalidate sq term using robIdx
1241  for (i <- 0 until StoreQueueSize) {
1242    needCancel(i) := uop(i).robIdx.needFlush(io.brqRedirect) && allocated(i) && !committed(i) &&
1243      (!isVec(i) || !(uop(i).robIdx === io.brqRedirect.bits.robIdx))
1244    when (needCancel(i)) {
1245      allocated(i) := false.B
1246    }
1247  }
1248
1249 /**
1250* update pointers
1251**/
1252  val enqCancelValid = canEnqueue.zip(io.enq.req).map{case (v , x) =>
1253    v && x.bits.robIdx.needFlush(io.brqRedirect)
1254  }
1255  val enqCancelNum = enqCancelValid.zip(io.enq.req).map{case (v, req) =>
1256    Mux(v, req.bits.numLsElem, 0.U)
1257  }
1258  val lastEnqCancel = RegEnable(enqCancelNum.reduce(_ + _), io.brqRedirect.valid) // 1 cycle after redirect
1259
1260  val lastCycleCancelCount = PopCount(RegEnable(needCancel, io.brqRedirect.valid)) // 1 cycle after redirect
1261  val lastCycleRedirect = RegNext(io.brqRedirect.valid) // 1 cycle after redirect
1262  val enqNumber = validVStoreFlow.reduce(_ + _)
1263
1264  val lastlastCycleRedirect=RegNext(lastCycleRedirect)// 2 cycle after redirect
1265  val redirectCancelCount = RegEnable(lastCycleCancelCount + lastEnqCancel, 0.U, lastCycleRedirect) // 2 cycle after redirect
1266
1267  when (lastlastCycleRedirect) {
1268    // we recover the pointers in 2 cycle after redirect for better timing
1269    enqPtrExt := VecInit(enqPtrExt.map(_ - redirectCancelCount))
1270  }.otherwise {
1271    // lastCycleRedirect.valid or nornal case
1272    // when lastCycleRedirect.valid, enqNumber === 0.U, enqPtrExt will not change
1273    enqPtrExt := VecInit(enqPtrExt.map(_ + enqNumber))
1274  }
1275  assert(!(lastCycleRedirect && enqNumber =/= 0.U))
1276
1277  exceptionBuffer.io.flushFrmMaBuf := finishMisalignSt
1278  // special case (store miss align) in updating ptr
1279  when (doMisalignSt) {
1280    when (!finishMisalignSt) {
1281      // dont move deqPtr and rdataPtr until all split store has been written to sb
1282      deqPtrExtNext := deqPtrExt
1283      rdataPtrExtNext := rdataPtrExt
1284    } .otherwise {
1285      // remove this unaligned store from sq
1286      allocated(deqPtr) := false.B
1287      committed(deqPtr) := true.B
1288      cmtPtrExt := cmtPtrExt.map(_ + 1.U)
1289      deqPtrExtNext := deqPtrExt.map(_ + 1.U)
1290      rdataPtrExtNext := rdataPtrExt.map(_ + 1.U)
1291    }
1292  }
1293
1294  deqPtrExt := deqPtrExtNext
1295  rdataPtrExt := rdataPtrExtNext
1296
1297  // val dequeueCount = Mux(io.sbuffer(1).fire, 2.U, Mux(io.sbuffer(0).fire || io.mmioStout.fire, 1.U, 0.U))
1298
1299  // If redirect at T0, sqCancelCnt is at T2
1300  io.sqCancelCnt := redirectCancelCount
1301  val ForceWriteUpper = Wire(UInt(log2Up(StoreQueueSize + 1).W))
1302  ForceWriteUpper := Constantin.createRecord(s"ForceWriteUpper_${p(XSCoreParamsKey).HartId}", initValue = 60)
1303  val ForceWriteLower = Wire(UInt(log2Up(StoreQueueSize + 1).W))
1304  ForceWriteLower := Constantin.createRecord(s"ForceWriteLower_${p(XSCoreParamsKey).HartId}", initValue = 55)
1305
1306  val valid_cnt = PopCount(allocated)
1307  io.force_write := RegNext(Mux(valid_cnt >= ForceWriteUpper, true.B, valid_cnt >= ForceWriteLower && io.force_write), init = false.B)
1308
1309  // io.sqempty will be used by sbuffer
1310  // We delay it for 1 cycle for better timing
1311  // When sbuffer need to check if it is empty, the pipeline is blocked, which means delay io.sqempty
1312  // for 1 cycle will also promise that sq is empty in that cycle
1313  io.sqEmpty := RegNext(
1314    enqPtrExt(0).value === deqPtrExt(0).value &&
1315    enqPtrExt(0).flag === deqPtrExt(0).flag
1316  )
1317  // perf counter
1318  QueuePerf(StoreQueueSize, validCount, !allowEnqueue)
1319  val vecValidVec = WireInit(VecInit((0 until StoreQueueSize).map(i => allocated(i) && isVec(i))))
1320  QueuePerf(StoreQueueSize, PopCount(vecValidVec), !allowEnqueue)
1321  io.sqFull := !allowEnqueue
1322  XSPerfAccumulate("mmioCycle", mmioState =/= s_idle) // lq is busy dealing with uncache req
1323  XSPerfAccumulate("mmioCnt", mmioDoReq)
1324  XSPerfAccumulate("mmio_wb_success", io.mmioStout.fire || io.vecmmioStout.fire)
1325  XSPerfAccumulate("mmio_wb_blocked", (io.mmioStout.valid && !io.mmioStout.ready) || (io.vecmmioStout.valid && !io.vecmmioStout.ready))
1326  XSPerfAccumulate("validEntryCnt", distanceBetween(enqPtrExt(0), deqPtrExt(0)))
1327  XSPerfAccumulate("cmtEntryCnt", distanceBetween(cmtPtrExt(0), deqPtrExt(0)))
1328  XSPerfAccumulate("nCmtEntryCnt", distanceBetween(enqPtrExt(0), cmtPtrExt(0)))
1329
1330  val perfValidCount = distanceBetween(enqPtrExt(0), deqPtrExt(0))
1331  val perfEvents = Seq(
1332    ("mmioCycle      ", mmioState =/= s_idle),
1333    ("mmioCnt        ", mmioDoReq),
1334    ("mmio_wb_success", io.mmioStout.fire || io.vecmmioStout.fire),
1335    ("mmio_wb_blocked", (io.mmioStout.valid && !io.mmioStout.ready) || (io.vecmmioStout.valid && !io.vecmmioStout.ready)),
1336    ("stq_1_4_valid  ", (perfValidCount < (StoreQueueSize.U/4.U))),
1337    ("stq_2_4_valid  ", (perfValidCount > (StoreQueueSize.U/4.U)) & (perfValidCount <= (StoreQueueSize.U/2.U))),
1338    ("stq_3_4_valid  ", (perfValidCount > (StoreQueueSize.U/2.U)) & (perfValidCount <= (StoreQueueSize.U*3.U/4.U))),
1339    ("stq_4_4_valid  ", (perfValidCount > (StoreQueueSize.U*3.U/4.U))),
1340  )
1341  generatePerfEvent()
1342
1343  // debug info
1344  XSDebug("enqPtrExt %d:%d deqPtrExt %d:%d\n", enqPtrExt(0).flag, enqPtr, deqPtrExt(0).flag, deqPtr)
1345
1346  def PrintFlag(flag: Bool, name: String): Unit = {
1347    when(flag) {
1348      XSDebug(false, true.B, name)
1349    }.otherwise {
1350      XSDebug(false, true.B, " ")
1351    }
1352  }
1353
1354  for (i <- 0 until StoreQueueSize) {
1355    XSDebug(s"$i: pc %x va %x pa %x data %x ",
1356      uop(i).pc,
1357      debug_vaddr(i),
1358      debug_paddr(i),
1359      debug_data(i)
1360    )
1361    PrintFlag(allocated(i), "a")
1362    PrintFlag(allocated(i) && addrvalid(i), "a")
1363    PrintFlag(allocated(i) && datavalid(i), "d")
1364    PrintFlag(allocated(i) && committed(i), "c")
1365    PrintFlag(allocated(i) && pending(i), "p")
1366    PrintFlag(allocated(i) && mmio(i), "m")
1367    XSDebug(false, true.B, "\n")
1368  }
1369
1370}
1371