xref: /XiangShan/src/main/scala/xiangshan/mem/lsqueue/StoreQueue.scala (revision f7063a43ab34da917ba6c670d21871314340c550)
1/***************************************************************************************
2* Copyright (c) 2020-2021 Institute of Computing Technology, Chinese Academy of Sciences
3* Copyright (c) 2020-2021 Peng Cheng Laboratory
4*
5* XiangShan is licensed under Mulan PSL v2.
6* You can use this software according to the terms and conditions of the Mulan PSL v2.
7* You may obtain a copy of Mulan PSL v2 at:
8*          http://license.coscl.org.cn/MulanPSL2
9*
10* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
11* EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
12* MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
13*
14* See the Mulan PSL v2 for more details.
15***************************************************************************************/
16
17package xiangshan.mem
18
19import chisel3._
20import chisel3.util._
21import difftest._
22import difftest.common.DifftestMem
23import org.chipsalliance.cde.config.Parameters
24import utility._
25import utils._
26import xiangshan._
27import xiangshan.cache._
28import xiangshan.cache.{DCacheLineIO, DCacheWordIO, MemoryOpConstants}
29import xiangshan.backend._
30import xiangshan.backend.rob.{RobLsqIO, RobPtr}
31import xiangshan.backend.Bundles.{DynInst, MemExuOutput}
32import xiangshan.backend.decode.isa.bitfield.{Riscv32BitInst, XSInstBitFields}
33
34class SqPtr(implicit p: Parameters) extends CircularQueuePtr[SqPtr](
35  p => p(XSCoreParamsKey).StoreQueueSize
36){
37}
38
39object SqPtr {
40  def apply(f: Bool, v: UInt)(implicit p: Parameters): SqPtr = {
41    val ptr = Wire(new SqPtr)
42    ptr.flag := f
43    ptr.value := v
44    ptr
45  }
46}
47
48class SqEnqIO(implicit p: Parameters) extends MemBlockBundle {
49  val canAccept = Output(Bool())
50  val lqCanAccept = Input(Bool())
51  val needAlloc = Vec(LSQEnqWidth, Input(Bool()))
52  val req = Vec(LSQEnqWidth, Flipped(ValidIO(new DynInst)))
53  val resp = Vec(LSQEnqWidth, Output(new SqPtr))
54}
55
56class DataBufferEntry (implicit p: Parameters)  extends DCacheBundle {
57  val addr   = UInt(PAddrBits.W)
58  val vaddr  = UInt(VAddrBits.W)
59  val data   = UInt(VLEN.W)
60  val mask   = UInt((VLEN/8).W)
61  val wline = Bool()
62  val sqPtr  = new SqPtr
63  val prefetch = Bool()
64}
65
66// Store Queue
67class StoreQueue(implicit p: Parameters) extends XSModule
68  with HasDCacheParameters with HasCircularQueuePtrHelper with HasPerfEvents {
69  val io = IO(new Bundle() {
70    val hartId = Input(UInt(hartIdLen.W))
71    val enq = new SqEnqIO
72    val brqRedirect = Flipped(ValidIO(new Redirect))
73    val storeAddrIn = Vec(StorePipelineWidth, Flipped(Valid(new LsPipelineBundle))) // store addr, data is not included
74    val vecStoreAddrIn = Vec(StorePipelineWidth, Flipped(Valid(new LsPipelineBundle))) // vec store addr, data is not include , from vsFlowQueue
75    val storeAddrInRe = Vec(StorePipelineWidth, Input(new LsPipelineBundle())) // store more mmio and exception
76    val storeDataIn = Vec(StorePipelineWidth, Flipped(Valid(new MemExuOutput))) // store data, send to sq from rs
77    val storeMaskIn = Vec(StorePipelineWidth, Flipped(Valid(new StoreMaskBundle))) // store mask, send to sq from rs
78    val sbuffer = Vec(EnsbufferWidth, Decoupled(new DCacheWordReqWithVaddrAndPfFlag)) // write committed store to sbuffer
79    val uncacheOutstanding = Input(Bool())
80    val mmioStout = DecoupledIO(new MemExuOutput) // writeback uncached store
81    val forward = Vec(LoadPipelineWidth, Flipped(new PipeLoadForwardQueryIO))
82    val rob = Flipped(new RobLsqIO)
83    val uncache = new UncacheWordIO
84    // val refill = Flipped(Valid(new DCacheLineReq ))
85    val exceptionAddr = new ExceptionAddrIO
86    val sqEmpty = Output(Bool())
87    val stAddrReadySqPtr = Output(new SqPtr)
88    val stAddrReadyVec = Output(Vec(StoreQueueSize, Bool()))
89    val stDataReadySqPtr = Output(new SqPtr)
90    val stDataReadyVec = Output(Vec(StoreQueueSize, Bool()))
91    val stIssuePtr = Output(new SqPtr)
92    val sqDeqPtr = Output(new SqPtr)
93    val sqFull = Output(Bool())
94    val sqCancelCnt = Output(UInt(log2Up(StoreQueueSize + 1).W))
95    val sqDeq = Output(UInt(log2Ceil(EnsbufferWidth + 1).W))
96    val force_write = Output(Bool())
97    val vecStoreRetire = Flipped(ValidIO(new SqPtr))
98  })
99
100  println("StoreQueue: size:" + StoreQueueSize)
101
102  // data modules
103  val uop = Reg(Vec(StoreQueueSize, new DynInst))
104  // val data = Reg(Vec(StoreQueueSize, new LsqEntry))
105  val dataModule = Module(new SQDataModule(
106    numEntries = StoreQueueSize,
107    numRead = EnsbufferWidth,
108    numWrite = StorePipelineWidth,
109    numForward = LoadPipelineWidth
110  ))
111  dataModule.io := DontCare
112  val paddrModule = Module(new SQAddrModule(
113    dataWidth = PAddrBits,
114    numEntries = StoreQueueSize,
115    numRead = EnsbufferWidth,
116    numWrite = StorePipelineWidth,
117    numForward = LoadPipelineWidth
118  ))
119  paddrModule.io := DontCare
120  val vaddrModule = Module(new SQAddrModule(
121    dataWidth = VAddrBits,
122    numEntries = StoreQueueSize,
123    numRead = EnsbufferWidth + 1, // sbuffer + badvaddr 1 (TODO)
124    numWrite = StorePipelineWidth,
125    numForward = LoadPipelineWidth
126  ))
127  val gpaddrModule = Module(new SQAddrModule(
128    dataWidth = GPAddrBits,
129    numEntries = StoreQueueSize,
130    numRead = EnsbufferWidth + 1,
131    numWrite = StorePipelineWidth,
132    numForward = LoadPipelineWidth
133  ))
134  vaddrModule.io := DontCare
135  gpaddrModule.io := DontCare
136  val dataBuffer = Module(new DatamoduleResultBuffer(new DataBufferEntry))
137  val debug_paddr = Reg(Vec(StoreQueueSize, UInt((PAddrBits).W)))
138  val debug_vaddr = Reg(Vec(StoreQueueSize, UInt((VAddrBits).W)))
139  val debug_data = Reg(Vec(StoreQueueSize, UInt((XLEN).W)))
140
141  // state & misc
142  val allocated = RegInit(VecInit(List.fill(StoreQueueSize)(false.B))) // sq entry has been allocated
143  val addrvalid = RegInit(VecInit(List.fill(StoreQueueSize)(false.B))) // non-mmio addr is valid
144  val datavalid = RegInit(VecInit(List.fill(StoreQueueSize)(false.B))) // non-mmio data is valid
145  val allvalid  = VecInit((0 until StoreQueueSize).map(i => addrvalid(i) && datavalid(i))) // non-mmio data & addr is valid
146  val committed = Reg(Vec(StoreQueueSize, Bool())) // inst has been committed by rob
147  val pending = Reg(Vec(StoreQueueSize, Bool())) // mmio pending: inst is an mmio inst, it will not be executed until it reachs the end of rob
148  val mmio = Reg(Vec(StoreQueueSize, Bool())) // mmio: inst is an mmio inst
149  val atomic = Reg(Vec(StoreQueueSize, Bool()))
150  val prefetch = RegInit(VecInit(List.fill(StoreQueueSize)(false.B))) // need prefetch when committing this store to sbuffer?
151  val vec = Reg(Vec(StoreQueueSize, Bool()))
152  val vecAddrvalid = Reg(Vec(StoreQueueSize, Bool())) // TODO
153
154  // ptr
155  val enqPtrExt = RegInit(VecInit((0 until io.enq.req.length).map(_.U.asTypeOf(new SqPtr))))
156  val rdataPtrExt = RegInit(VecInit((0 until EnsbufferWidth).map(_.U.asTypeOf(new SqPtr))))
157  val deqPtrExt = RegInit(VecInit((0 until EnsbufferWidth).map(_.U.asTypeOf(new SqPtr))))
158  val cmtPtrExt = RegInit(VecInit((0 until CommitWidth).map(_.U.asTypeOf(new SqPtr))))
159  val addrReadyPtrExt = RegInit(0.U.asTypeOf(new SqPtr))
160  val dataReadyPtrExt = RegInit(0.U.asTypeOf(new SqPtr))
161  val validCounter = RegInit(0.U(log2Ceil(VirtualLoadQueueSize + 1).W))
162
163  val enqPtr = enqPtrExt(0).value
164  val deqPtr = deqPtrExt(0).value
165  val cmtPtr = cmtPtrExt(0).value
166
167  val validCount = distanceBetween(enqPtrExt(0), deqPtrExt(0))
168  val allowEnqueue = validCount <= (StoreQueueSize - LSQStEnqWidth).U
169
170  val deqMask = UIntToMask(deqPtr, StoreQueueSize)
171  val enqMask = UIntToMask(enqPtr, StoreQueueSize)
172
173  val commitCount = RegNext(io.rob.scommit)
174
175  // store can be committed by ROB
176  io.rob.mmio := DontCare
177  io.rob.uop := DontCare
178
179  // Read dataModule
180  assert(EnsbufferWidth <= 2)
181  // rdataPtrExtNext and rdataPtrExtNext+1 entry will be read from dataModule
182  val rdataPtrExtNext = WireInit(Mux(dataBuffer.io.enq(1).fire,
183    VecInit(rdataPtrExt.map(_ + 2.U)),
184    Mux(dataBuffer.io.enq(0).fire || io.mmioStout.fire || io.vecStoreRetire.valid,
185      VecInit(rdataPtrExt.map(_ + 1.U)),
186      rdataPtrExt
187    )
188  ))
189
190  // deqPtrExtNext traces which inst is about to leave store queue
191  //
192  // io.sbuffer(i).fire is RegNexted, as sbuffer data write takes 2 cycles.
193  // Before data write finish, sbuffer is unable to provide store to load
194  // forward data. As an workaround, deqPtrExt and allocated flag update
195  // is delayed so that load can get the right data from store queue.
196  //
197  // Modify deqPtrExtNext and io.sqDeq with care!
198  val deqPtrExtNext = Mux(RegNext(io.sbuffer(1).fire),
199    VecInit(deqPtrExt.map(_ + 2.U)),
200    Mux(RegNext(io.sbuffer(0).fire) || io.mmioStout.fire || io.vecStoreRetire.valid,
201      VecInit(deqPtrExt.map(_ + 1.U)),
202      deqPtrExt
203    )
204  )
205  io.sqDeq := RegNext(Mux(RegNext(io.sbuffer(1).fire), 2.U,
206    Mux(RegNext(io.sbuffer(0).fire) || io.mmioStout.fire || io.vecStoreRetire.valid, 1.U, 0.U)
207  ))
208  assert(!RegNext(RegNext(io.sbuffer(0).fire) && io.mmioStout.fire))
209
210  for (i <- 0 until EnsbufferWidth) {
211    dataModule.io.raddr(i) := rdataPtrExtNext(i).value
212    paddrModule.io.raddr(i) := rdataPtrExtNext(i).value
213    vaddrModule.io.raddr(i) := rdataPtrExtNext(i).value
214    gpaddrModule.io.raddr(i) := rdataPtrExtNext(i).value
215  }
216
217  // no inst will be committed 1 cycle before tval update
218  vaddrModule.io.raddr(EnsbufferWidth) := (cmtPtrExt(0) + commitCount).value
219  gpaddrModule.io.raddr(EnsbufferWidth) := (cmtPtrExt(0) + commitCount).value
220  /**
221    * Enqueue at dispatch
222    *
223    * Currently, StoreQueue only allows enqueue when #emptyEntries > EnqWidth
224    */
225  io.enq.canAccept := allowEnqueue
226  val canEnqueue = io.enq.req.map(_.valid)
227  val enqCancel = io.enq.req.map(_.bits.robIdx.needFlush(io.brqRedirect))
228  for (i <- 0 until io.enq.req.length) {
229    val offset = if (i == 0) 0.U else PopCount(io.enq.needAlloc.take(i))
230    val sqIdx = enqPtrExt(offset)
231    val index = io.enq.req(i).bits.sqIdx.value
232    val enqInstr = io.enq.req(i).bits.instr.asTypeOf(new XSInstBitFields)
233    when (canEnqueue(i) && !enqCancel(i)) {
234      uop(index) := io.enq.req(i).bits
235      // NOTE: the index will be used when replay
236      uop(index).sqIdx := sqIdx
237      allocated(index) := true.B
238      datavalid(index) := false.B
239      addrvalid(index) := false.B
240      committed(index) := false.B
241      pending(index) := false.B
242      prefetch(index) := false.B
243      mmio(index) := false.B
244      vec(index) := enqInstr.isVecStore // check vector store by the encoding of inst
245      vecAddrvalid(index) := false.B//TODO
246
247      XSError(!io.enq.canAccept || !io.enq.lqCanAccept, s"must accept $i\n")
248      XSError(index =/= sqIdx.value, s"must be the same entry $i\n")
249    }
250    io.enq.resp(i) := sqIdx
251  }
252  XSDebug(p"(ready, valid): ${io.enq.canAccept}, ${Binary(Cat(io.enq.req.map(_.valid)))}\n")
253
254  /**
255    * Update addr/dataReadyPtr when issue from rs
256    */
257  // update issuePtr
258  val IssuePtrMoveStride = 4
259  require(IssuePtrMoveStride >= 2)
260
261  val addrReadyLookupVec = (0 until IssuePtrMoveStride).map(addrReadyPtrExt + _.U)
262  val addrReadyLookup = addrReadyLookupVec.map(ptr => allocated(ptr.value) && (mmio(ptr.value) || addrvalid(ptr.value) || (vec(ptr.value) && vecAddrvalid(ptr.value))) && ptr =/= enqPtrExt(0))
263  val nextAddrReadyPtr = addrReadyPtrExt + PriorityEncoder(VecInit(addrReadyLookup.map(!_) :+ true.B))
264  addrReadyPtrExt := nextAddrReadyPtr
265
266  (0 until StoreQueueSize).map(i => {
267    io.stAddrReadyVec(i) := RegNext(allocated(i) && (mmio(i) || addrvalid(i)))
268  })
269
270  when (io.brqRedirect.valid) {
271    addrReadyPtrExt := Mux(
272      isAfter(cmtPtrExt(0), deqPtrExt(0)),
273      cmtPtrExt(0),
274      deqPtrExtNext(0) // for mmio insts, deqPtr may be ahead of cmtPtr
275    )
276  }
277
278  io.stAddrReadySqPtr := addrReadyPtrExt
279
280  // update
281  val dataReadyLookupVec = (0 until IssuePtrMoveStride).map(dataReadyPtrExt + _.U)
282  val dataReadyLookup = dataReadyLookupVec.map(ptr => allocated(ptr.value) && (mmio(ptr.value) || datavalid(ptr.value) || vec(ptr.value)) && ptr =/= enqPtrExt(0)) // TODO : flag of vector store data valid not add yet
283  val nextDataReadyPtr = dataReadyPtrExt + PriorityEncoder(VecInit(dataReadyLookup.map(!_) :+ true.B))
284  dataReadyPtrExt := nextDataReadyPtr
285
286  (0 until StoreQueueSize).map(i => {
287    io.stDataReadyVec(i) := RegNext(allocated(i) && (mmio(i) || datavalid(i)))
288  })
289
290  when (io.brqRedirect.valid) {
291    dataReadyPtrExt := Mux(
292      isAfter(cmtPtrExt(0), deqPtrExt(0)),
293      cmtPtrExt(0),
294      deqPtrExtNext(0) // for mmio insts, deqPtr may be ahead of cmtPtr
295    )
296  }
297
298  io.stDataReadySqPtr := dataReadyPtrExt
299  io.stIssuePtr := enqPtrExt(0)
300  io.sqDeqPtr := deqPtrExt(0)
301
302  /**
303    * Writeback store from store units
304    *
305    * Most store instructions writeback to regfile in the previous cycle.
306    * However,
307    *   (1) For an mmio instruction with exceptions, we need to mark it as addrvalid
308    * (in this way it will trigger an exception when it reaches ROB's head)
309    * instead of pending to avoid sending them to lower level.
310    *   (2) For an mmio instruction without exceptions, we mark it as pending.
311    * When the instruction reaches ROB's head, StoreQueue sends it to uncache channel.
312    * Upon receiving the response, StoreQueue writes back the instruction
313    * through arbiter with store units. It will later commit as normal.
314    */
315
316  // Write addr to sq
317  for (i <- 0 until StorePipelineWidth) {
318    paddrModule.io.wen(i) := false.B
319    vaddrModule.io.wen(i) := false.B
320    gpaddrModule.io.wen(i) := false.B
321    dataModule.io.mask.wen(i) := false.B
322    val stWbIndex = io.storeAddrIn(i).bits.uop.sqIdx.value
323    when (io.storeAddrIn(i).fire) {
324      val addr_valid = !io.storeAddrIn(i).bits.miss
325      addrvalid(stWbIndex) := addr_valid //!io.storeAddrIn(i).bits.mmio
326      // pending(stWbIndex) := io.storeAddrIn(i).bits.mmio
327
328      paddrModule.io.waddr(i) := stWbIndex
329      paddrModule.io.wdata(i) := io.storeAddrIn(i).bits.paddr
330      paddrModule.io.wmask(i) := io.storeAddrIn(i).bits.mask
331      paddrModule.io.wlineflag(i) := io.storeAddrIn(i).bits.wlineflag
332      paddrModule.io.wen(i) := true.B
333
334      vaddrModule.io.waddr(i) := stWbIndex
335      vaddrModule.io.wdata(i) := io.storeAddrIn(i).bits.vaddr
336      vaddrModule.io.wmask(i) := io.storeAddrIn(i).bits.mask
337      vaddrModule.io.wlineflag(i) := io.storeAddrIn(i).bits.wlineflag
338      vaddrModule.io.wen(i) := true.B
339
340      gpaddrModule.io.waddr(i) := stWbIndex
341      gpaddrModule.io.wdata(i) := io.storeAddrIn(i).bits.gpaddr
342      gpaddrModule.io.wmask(i)  := io.storeAddrIn(i).bits.mask
343      gpaddrModule.io.wlineflag(i) := io.storeAddrIn(i).bits.wlineflag
344      gpaddrModule.io.wen(i) := true.B
345
346      debug_paddr(paddrModule.io.waddr(i)) := paddrModule.io.wdata(i)
347
348      // mmio(stWbIndex) := io.storeAddrIn(i).bits.mmio
349
350      uop(stWbIndex) := io.storeAddrIn(i).bits.uop
351      uop(stWbIndex).debugInfo := io.storeAddrIn(i).bits.uop.debugInfo
352      XSInfo("store addr write to sq idx %d pc 0x%x miss:%d vaddr %x paddr %x mmio %x\n",
353        io.storeAddrIn(i).bits.uop.sqIdx.value,
354        io.storeAddrIn(i).bits.uop.pc,
355        io.storeAddrIn(i).bits.miss,
356        io.storeAddrIn(i).bits.vaddr,
357        io.storeAddrIn(i).bits.paddr,
358        io.storeAddrIn(i).bits.mmio
359      )
360    }
361
362    // re-replinish mmio, for pma/pmp will get mmio one cycle later
363    val storeAddrInFireReg = RegNext(io.storeAddrIn(i).fire && !io.storeAddrIn(i).bits.miss)
364    val stWbIndexReg = RegNext(stWbIndex)
365    when (storeAddrInFireReg) {
366      pending(stWbIndexReg) := io.storeAddrInRe(i).mmio
367      mmio(stWbIndexReg) := io.storeAddrInRe(i).mmio
368      atomic(stWbIndexReg) := io.storeAddrInRe(i).atomic
369    }
370    // dcache miss info (one cycle later than storeIn)
371    // if dcache report a miss in sta pipeline, this store will trigger a prefetch when committing to sbuffer (if EnableAtCommitMissTrigger)
372    when (storeAddrInFireReg) {
373      prefetch(stWbIndexReg) := io.storeAddrInRe(i).miss
374    }
375
376    when(vaddrModule.io.wen(i)){
377      debug_vaddr(vaddrModule.io.waddr(i)) := vaddrModule.io.wdata(i)
378    }
379    // TODO :  When lastElem issue to stu or inactivative issue in vsFlowQueue, set vector store addr ready
380    val vecStWbIndex = io.vecStoreAddrIn(i).bits.uop.sqIdx.value
381    when(io.vecStoreAddrIn(i).fire){
382      vecAddrvalid(vecStWbIndex) := true.B
383    }
384  }
385
386  // Write data to sq
387  // Now store data pipeline is actually 2 stages
388  for (i <- 0 until StorePipelineWidth) {
389    dataModule.io.data.wen(i) := false.B
390    val stWbIndex = io.storeDataIn(i).bits.uop.sqIdx.value
391    // sq data write takes 2 cycles:
392    // sq data write s0
393    when (io.storeDataIn(i).fire) {
394      // send data write req to data module
395      dataModule.io.data.waddr(i) := stWbIndex
396      dataModule.io.data.wdata(i) := Mux(io.storeDataIn(i).bits.uop.fuOpType === LSUOpType.cbo_zero,
397        0.U,
398        genWdata(io.storeDataIn(i).bits.data, io.storeDataIn(i).bits.uop.fuOpType(1,0))
399      )
400      dataModule.io.data.wen(i) := true.B
401
402      debug_data(dataModule.io.data.waddr(i)) := dataModule.io.data.wdata(i)
403
404      XSInfo("store data write to sq idx %d pc 0x%x data %x -> %x\n",
405        io.storeDataIn(i).bits.uop.sqIdx.value,
406        io.storeDataIn(i).bits.uop.pc,
407        io.storeDataIn(i).bits.data,
408        dataModule.io.data.wdata(i)
409      )
410    }
411    // sq data write s1
412    when (
413      RegNext(io.storeDataIn(i).fire)
414      // && !RegNext(io.storeDataIn(i).bits.uop).robIdx.needFlush(io.brqRedirect)
415    ) {
416      datavalid(RegNext(stWbIndex)) := true.B
417    }
418  }
419
420  // Write mask to sq
421  for (i <- 0 until StorePipelineWidth) {
422    // sq mask write s0
423    when (io.storeMaskIn(i).fire) {
424      // send data write req to data module
425      dataModule.io.mask.waddr(i) := io.storeMaskIn(i).bits.sqIdx.value
426      dataModule.io.mask.wdata(i) := io.storeMaskIn(i).bits.mask
427      dataModule.io.mask.wen(i) := true.B
428    }
429  }
430
431  /**
432    * load forward query
433    *
434    * Check store queue for instructions that is older than the load.
435    * The response will be valid at the next cycle after req.
436    */
437  // check over all lq entries and forward data from the first matched store
438  for (i <- 0 until LoadPipelineWidth) {
439    // Compare deqPtr (deqPtr) and forward.sqIdx, we have two cases:
440    // (1) if they have the same flag, we need to check range(tail, sqIdx)
441    // (2) if they have different flags, we need to check range(tail, VirtualLoadQueueSize) and range(0, sqIdx)
442    // Forward1: Mux(same_flag, range(tail, sqIdx), range(tail, VirtualLoadQueueSize))
443    // Forward2: Mux(same_flag, 0.U,                   range(0, sqIdx)    )
444    // i.e. forward1 is the target entries with the same flag bits and forward2 otherwise
445    val differentFlag = deqPtrExt(0).flag =/= io.forward(i).sqIdx.flag
446    val forwardMask = io.forward(i).sqIdxMask
447    // all addrvalid terms need to be checked
448    val addrValidVec = WireInit(VecInit((0 until StoreQueueSize).map(j => addrvalid(j) && allocated(j))))
449    val dataValidVec = WireInit(VecInit((0 until StoreQueueSize).map(j => datavalid(j))))
450    val allValidVec  = WireInit(VecInit((0 until StoreQueueSize).map(j => addrvalid(j) && datavalid(j) && allocated(j))))
451
452    val lfstEnable = Constantin.createRecord("LFSTEnable", LFSTEnable.B).orR
453    val storeSetHitVec = Mux(lfstEnable,
454      WireInit(VecInit((0 until StoreQueueSize).map(j => io.forward(i).uop.loadWaitBit && uop(j).robIdx === io.forward(i).uop.waitForRobIdx))),
455      WireInit(VecInit((0 until StoreQueueSize).map(j => uop(j).storeSetHit && uop(j).ssid === io.forward(i).uop.ssid)))
456    )
457
458    val forwardMask1 = Mux(differentFlag, ~deqMask, deqMask ^ forwardMask)
459    val forwardMask2 = Mux(differentFlag, forwardMask, 0.U(StoreQueueSize.W))
460    val canForward1 = forwardMask1 & allValidVec.asUInt
461    val canForward2 = forwardMask2 & allValidVec.asUInt
462    val needForward = Mux(differentFlag, ~deqMask | forwardMask, deqMask ^ forwardMask)
463
464    XSDebug(p"$i f1 ${Binary(canForward1)} f2 ${Binary(canForward2)} " +
465      p"sqIdx ${io.forward(i).sqIdx} pa ${Hexadecimal(io.forward(i).paddr)}\n"
466    )
467
468    // do real fwd query (cam lookup in load_s1)
469    dataModule.io.needForward(i)(0) := canForward1 & vaddrModule.io.forwardMmask(i).asUInt
470    dataModule.io.needForward(i)(1) := canForward2 & vaddrModule.io.forwardMmask(i).asUInt
471
472    vaddrModule.io.forwardMdata(i) := io.forward(i).vaddr
473    vaddrModule.io.forwardDataMask(i) := io.forward(i).mask
474    paddrModule.io.forwardMdata(i) := io.forward(i).paddr
475    paddrModule.io.forwardDataMask(i) := io.forward(i).mask
476    gpaddrModule.io.forwardMdata(i) := io.forward(i).gpaddr
477    gpaddrModule.io.forwardDataMask(i) := io.forward(i).mask
478
479
480    // vaddr cam result does not equal to paddr cam result
481    // replay needed
482    // val vpmaskNotEqual = ((paddrModule.io.forwardMmask(i).asUInt ^ vaddrModule.io.forwardMmask(i).asUInt) & needForward) =/= 0.U
483    // val vaddrMatchFailed = vpmaskNotEqual && io.forward(i).valid
484    val vpmaskNotEqual = (
485      (RegNext(paddrModule.io.forwardMmask(i).asUInt) ^ RegNext(vaddrModule.io.forwardMmask(i).asUInt)) &
486      RegNext(needForward) &
487      RegNext(addrValidVec.asUInt)
488    ) =/= 0.U
489    val vaddrMatchFailed = vpmaskNotEqual && RegNext(io.forward(i).valid)
490    when (vaddrMatchFailed) {
491      XSInfo("vaddrMatchFailed: pc %x pmask %x vmask %x\n",
492        RegNext(io.forward(i).uop.pc),
493        RegNext(needForward & paddrModule.io.forwardMmask(i).asUInt),
494        RegNext(needForward & vaddrModule.io.forwardMmask(i).asUInt)
495      );
496    }
497    XSPerfAccumulate("vaddr_match_failed", vpmaskNotEqual)
498    XSPerfAccumulate("vaddr_match_really_failed", vaddrMatchFailed)
499
500    // Fast forward mask will be generated immediately (load_s1)
501    io.forward(i).forwardMaskFast := dataModule.io.forwardMaskFast(i)
502
503    // Forward result will be generated 1 cycle later (load_s2)
504    io.forward(i).forwardMask := dataModule.io.forwardMask(i)
505    io.forward(i).forwardData := dataModule.io.forwardData(i)
506    // If addr match, data not ready, mark it as dataInvalid
507    // load_s1: generate dataInvalid in load_s1 to set fastUop
508    val dataInvalidMask1 = (addrValidVec.asUInt & ~dataValidVec.asUInt & vaddrModule.io.forwardMmask(i).asUInt & forwardMask1.asUInt)
509    val dataInvalidMask2 = (addrValidVec.asUInt & ~dataValidVec.asUInt & vaddrModule.io.forwardMmask(i).asUInt & forwardMask2.asUInt)
510    val dataInvalidMask = dataInvalidMask1 | dataInvalidMask2
511    io.forward(i).dataInvalidFast := dataInvalidMask.orR
512
513    // make chisel happy
514    val dataInvalidMask1Reg = Wire(UInt(StoreQueueSize.W))
515    dataInvalidMask1Reg := RegNext(dataInvalidMask1)
516    // make chisel happy
517    val dataInvalidMask2Reg = Wire(UInt(StoreQueueSize.W))
518    dataInvalidMask2Reg := RegNext(dataInvalidMask2)
519    val dataInvalidMaskReg = dataInvalidMask1Reg | dataInvalidMask2Reg
520
521    // If SSID match, address not ready, mark it as addrInvalid
522    // load_s2: generate addrInvalid
523    val addrInvalidMask1 = (~addrValidVec.asUInt & storeSetHitVec.asUInt & forwardMask1.asUInt)
524    val addrInvalidMask2 = (~addrValidVec.asUInt & storeSetHitVec.asUInt & forwardMask2.asUInt)
525    // make chisel happy
526    val addrInvalidMask1Reg = Wire(UInt(StoreQueueSize.W))
527    addrInvalidMask1Reg := RegNext(addrInvalidMask1)
528    // make chisel happy
529    val addrInvalidMask2Reg = Wire(UInt(StoreQueueSize.W))
530    addrInvalidMask2Reg := RegNext(addrInvalidMask2)
531    val addrInvalidMaskReg = addrInvalidMask1Reg | addrInvalidMask2Reg
532
533    // load_s2
534    io.forward(i).dataInvalid := RegNext(io.forward(i).dataInvalidFast)
535    // check if vaddr forward mismatched
536    io.forward(i).matchInvalid := vaddrMatchFailed
537
538    // data invalid sq index
539    // check whether false fail
540    // check flag
541    val s2_differentFlag = RegNext(differentFlag)
542    val s2_enqPtrExt = RegNext(enqPtrExt(0))
543    val s2_deqPtrExt = RegNext(deqPtrExt(0))
544
545    // addr invalid sq index
546    // make chisel happy
547    val addrInvalidMaskRegWire = Wire(UInt(StoreQueueSize.W))
548    addrInvalidMaskRegWire := addrInvalidMaskReg
549    val addrInvalidFlag = addrInvalidMaskRegWire.orR
550    val hasInvalidAddr = (~addrValidVec.asUInt & needForward).orR
551
552    val addrInvalidSqIdx1 = OHToUInt(Reverse(PriorityEncoderOH(Reverse(addrInvalidMask1Reg))))
553    val addrInvalidSqIdx2 = OHToUInt(Reverse(PriorityEncoderOH(Reverse(addrInvalidMask2Reg))))
554    val addrInvalidSqIdx = Mux(addrInvalidMask2Reg.orR, addrInvalidSqIdx2, addrInvalidSqIdx1)
555
556    // store-set content management
557    //                +-----------------------+
558    //                | Search a SSID for the |
559    //                |    load operation     |
560    //                +-----------------------+
561    //                           |
562    //                           V
563    //                 +-------------------+
564    //                 | load wait strict? |
565    //                 +-------------------+
566    //                           |
567    //                           V
568    //               +----------------------+
569    //            Set|                      |Clean
570    //               V                      V
571    //  +------------------------+   +------------------------------+
572    //  | Waiting for all older  |   | Wait until the corresponding |
573    //  |   stores operations    |   | older store operations       |
574    //  +------------------------+   +------------------------------+
575
576
577
578    when (RegNext(io.forward(i).uop.loadWaitStrict)) {
579      io.forward(i).addrInvalidSqIdx := RegNext(io.forward(i).uop.sqIdx - 1.U)
580    } .elsewhen (addrInvalidFlag) {
581      io.forward(i).addrInvalidSqIdx.flag := Mux(!s2_differentFlag || addrInvalidSqIdx >= s2_deqPtrExt.value, s2_deqPtrExt.flag, s2_enqPtrExt.flag)
582      io.forward(i).addrInvalidSqIdx.value := addrInvalidSqIdx
583    } .otherwise {
584      // may be store inst has been written to sbuffer already.
585      io.forward(i).addrInvalidSqIdx := RegNext(io.forward(i).uop.sqIdx)
586    }
587    io.forward(i).addrInvalid := Mux(RegNext(io.forward(i).uop.loadWaitStrict), RegNext(hasInvalidAddr), addrInvalidFlag)
588
589    // data invalid sq index
590    // make chisel happy
591    val dataInvalidMaskRegWire = Wire(UInt(StoreQueueSize.W))
592    dataInvalidMaskRegWire := dataInvalidMaskReg
593    val dataInvalidFlag = dataInvalidMaskRegWire.orR
594
595    val dataInvalidSqIdx1 = OHToUInt(Reverse(PriorityEncoderOH(Reverse(dataInvalidMask1Reg))))
596    val dataInvalidSqIdx2 = OHToUInt(Reverse(PriorityEncoderOH(Reverse(dataInvalidMask2Reg))))
597    val dataInvalidSqIdx = Mux(dataInvalidMask2Reg.orR, dataInvalidSqIdx2, dataInvalidSqIdx1)
598
599    when (dataInvalidFlag) {
600      io.forward(i).dataInvalidSqIdx.flag := Mux(!s2_differentFlag || dataInvalidSqIdx >= s2_deqPtrExt.value, s2_deqPtrExt.flag, s2_enqPtrExt.flag)
601      io.forward(i).dataInvalidSqIdx.value := dataInvalidSqIdx
602    } .otherwise {
603      // may be store inst has been written to sbuffer already.
604      io.forward(i).dataInvalidSqIdx := RegNext(io.forward(i).uop.sqIdx)
605    }
606  }
607
608  /**
609    * Memory mapped IO / other uncached operations
610    *
611    * States:
612    * (1) writeback from store units: mark as pending
613    * (2) when they reach ROB's head, they can be sent to uncache channel
614    * (3) response from uncache channel: mark as datavalidmask.wen
615    * (4) writeback to ROB (and other units): mark as writebacked
616    * (5) ROB commits the instruction: same as normal instructions
617    */
618  //(2) when they reach ROB's head, they can be sent to uncache channel
619  val s_idle :: s_req :: s_resp :: s_wb :: s_wait :: Nil = Enum(5)
620  val uncacheState = RegInit(s_idle)
621  switch(uncacheState) {
622    is(s_idle) {
623      when(RegNext(io.rob.pendingst && pending(deqPtr) && allocated(deqPtr) && datavalid(deqPtr) && addrvalid(deqPtr))) {
624        uncacheState := s_req
625      }
626    }
627    is(s_req) {
628      when (io.uncache.req.fire) {
629        when (io.uncacheOutstanding) {
630          uncacheState := s_wb
631        } .otherwise {
632          uncacheState := s_resp
633        }
634      }
635    }
636    is(s_resp) {
637      when(io.uncache.resp.fire) {
638        uncacheState := s_wb
639      }
640    }
641    is(s_wb) {
642      when (io.mmioStout.fire) {
643        uncacheState := s_wait
644      }
645    }
646    is(s_wait) {
647      when(commitCount > 0.U) {
648        uncacheState := s_idle // ready for next mmio
649      }
650    }
651  }
652  io.uncache.req.valid := uncacheState === s_req
653
654  io.uncache.req.bits := DontCare
655  io.uncache.req.bits.cmd  := MemoryOpConstants.M_XWR
656  io.uncache.req.bits.addr := paddrModule.io.rdata(0) // data(deqPtr) -> rdata(0)
657  io.uncache.req.bits.data := shiftDataToLow(paddrModule.io.rdata(0), dataModule.io.rdata(0).data)
658  io.uncache.req.bits.mask := shiftMaskToLow(paddrModule.io.rdata(0), dataModule.io.rdata(0).mask)
659
660  // CBO op type check can be delayed for 1 cycle,
661  // as uncache op will not start in s_idle
662  val cbo_mmio_addr = paddrModule.io.rdata(0) >> 2 << 2 // clear lowest 2 bits for op
663  val cbo_mmio_op = 0.U //TODO
664  val cbo_mmio_data = cbo_mmio_addr | cbo_mmio_op
665  when(RegNext(LSUOpType.isCbo(uop(deqPtr).fuOpType))){
666    io.uncache.req.bits.addr := DontCare // TODO
667    io.uncache.req.bits.data := paddrModule.io.rdata(0)
668    io.uncache.req.bits.mask := DontCare // TODO
669  }
670
671  io.uncache.req.bits.atomic := atomic(RegNext(rdataPtrExtNext(0)).value)
672
673  when(io.uncache.req.fire){
674    // mmio store should not be committed until uncache req is sent
675    pending(deqPtr) := false.B
676
677    XSDebug(
678      p"uncache req: pc ${Hexadecimal(uop(deqPtr).pc)} " +
679      p"addr ${Hexadecimal(io.uncache.req.bits.addr)} " +
680      p"data ${Hexadecimal(io.uncache.req.bits.data)} " +
681      p"op ${Hexadecimal(io.uncache.req.bits.cmd)} " +
682      p"mask ${Hexadecimal(io.uncache.req.bits.mask)}\n"
683    )
684  }
685
686  // (3) response from uncache channel: mark as datavalid
687  io.uncache.resp.ready := true.B
688
689  // (4) writeback to ROB (and other units): mark as writebacked
690  io.mmioStout.valid := uncacheState === s_wb
691  io.mmioStout.bits.uop := uop(deqPtr)
692  io.mmioStout.bits.uop.sqIdx := deqPtrExt(0)
693  io.mmioStout.bits.data := shiftDataToLow(paddrModule.io.rdata(0), dataModule.io.rdata(0).data) // dataModule.io.rdata.read(deqPtr)
694  io.mmioStout.bits.debug.isMMIO := true.B
695  io.mmioStout.bits.debug.paddr := DontCare
696  io.mmioStout.bits.debug.isPerfCnt := false.B
697  io.mmioStout.bits.debug.vaddr := DontCare
698  // Remove MMIO inst from store queue after MMIO request is being sent
699  // That inst will be traced by uncache state machine
700  when (io.mmioStout.fire) {
701    allocated(deqPtr) := false.B
702  }
703
704  /**
705    * ROB commits store instructions (mark them as committed)
706    *
707    * (1) When store commits, mark it as committed.
708    * (2) They will not be cancelled and can be sent to lower level.
709    */
710  XSError(uncacheState =/= s_idle && uncacheState =/= s_wait && commitCount > 0.U,
711   "should not commit instruction when MMIO has not been finished\n")
712  for (i <- 0 until CommitWidth) {
713    when (commitCount > i.U) { // MMIO inst is not in progress
714      if(i == 0){
715        // MMIO inst should not update committed flag
716        // Note that commit count has been delayed for 1 cycle
717        when(uncacheState === s_idle){
718          committed(cmtPtrExt(0).value) := true.B
719        }
720      } else {
721        committed(cmtPtrExt(i).value) := true.B
722      }
723    }
724  }
725  cmtPtrExt := cmtPtrExt.map(_ + commitCount)
726
727  // committed stores will not be cancelled and can be sent to lower level.
728  // remove retired insts from sq, add retired store to sbuffer
729
730  // Read data from data module
731  // As store queue grows larger and larger, time needed to read data from data
732  // module keeps growing higher. Now we give data read a whole cycle.
733
734  // Vector stores are written to sbuffer by vector store flow queue rather than sq
735  XSError(io.vecStoreRetire.valid && !vec(rdataPtrExt(0).value), "Vector store flow queue is trying to retire a scalar store")
736  XSError(io.vecStoreRetire.valid && !allocated(rdataPtrExt(0).value), "Vector store flow queue is trying to retire an invalid entry")
737  XSError(io.vecStoreRetire.valid && vec(rdataPtrExt(0).value) && !vecAddrvalid(rdataPtrExt(0).value), "Vector store is trying to retire without write last element!")
738  when (io.vecStoreRetire.valid) {
739    assert(io.vecStoreRetire.bits === rdataPtrExt(0))
740    vec(rdataPtrExt(0).value) := false.B
741    vecAddrvalid(rdataPtrExt(0).value) := false.B
742    allocated(rdataPtrExt(0).value) := false.B
743  }
744
745  val mmioStall = mmio(rdataPtrExt(0).value)
746  val vecStall = vec(rdataPtrExt(0).value)
747  for (i <- 0 until EnsbufferWidth) {
748    val ptr = rdataPtrExt(i).value
749    dataBuffer.io.enq(i).valid := allocated(ptr) && committed(ptr) && !mmioStall && !vecStall
750    // Note that store data/addr should both be valid after store's commit
751    assert(!dataBuffer.io.enq(i).valid || allvalid(ptr))
752    dataBuffer.io.enq(i).bits.addr     := paddrModule.io.rdata(i)
753    dataBuffer.io.enq(i).bits.vaddr    := vaddrModule.io.rdata(i)
754    dataBuffer.io.enq(i).bits.data     := dataModule.io.rdata(i).data
755    dataBuffer.io.enq(i).bits.mask     := dataModule.io.rdata(i).mask
756    dataBuffer.io.enq(i).bits.wline    := paddrModule.io.rlineflag(i)
757    dataBuffer.io.enq(i).bits.sqPtr    := rdataPtrExt(i)
758    dataBuffer.io.enq(i).bits.prefetch := prefetch(ptr)
759  }
760
761  // Send data stored in sbufferReqBitsReg to sbuffer
762  for (i <- 0 until EnsbufferWidth) {
763    io.sbuffer(i).valid := dataBuffer.io.deq(i).valid
764    dataBuffer.io.deq(i).ready := io.sbuffer(i).ready
765    // Write line request should have all 1 mask
766    assert(!(io.sbuffer(i).valid && io.sbuffer(i).bits.wline && !io.sbuffer(i).bits.mask.andR))
767    io.sbuffer(i).bits := DontCare
768    io.sbuffer(i).bits.cmd   := MemoryOpConstants.M_XWR
769    io.sbuffer(i).bits.addr  := dataBuffer.io.deq(i).bits.addr
770    io.sbuffer(i).bits.vaddr := dataBuffer.io.deq(i).bits.vaddr
771    io.sbuffer(i).bits.data  := dataBuffer.io.deq(i).bits.data
772    io.sbuffer(i).bits.mask  := dataBuffer.io.deq(i).bits.mask
773    io.sbuffer(i).bits.wline := dataBuffer.io.deq(i).bits.wline
774    io.sbuffer(i).bits.prefetch := dataBuffer.io.deq(i).bits.prefetch
775
776    // io.sbuffer(i).fire is RegNexted, as sbuffer data write takes 2 cycles.
777    // Before data write finish, sbuffer is unable to provide store to load
778    // forward data. As an workaround, deqPtrExt and allocated flag update
779    // is delayed so that load can get the right data from store queue.
780    val ptr = dataBuffer.io.deq(i).bits.sqPtr.value
781    when (RegNext(io.sbuffer(i).fire)) {
782      allocated(RegEnable(ptr, io.sbuffer(i).fire)) := false.B
783      XSDebug("sbuffer "+i+" fire: ptr %d\n", ptr)
784    }
785  }
786  (1 until EnsbufferWidth).foreach(i => when(io.sbuffer(i).fire) { assert(io.sbuffer(i - 1).fire) })
787  if (coreParams.dcacheParametersOpt.isEmpty) {
788    for (i <- 0 until EnsbufferWidth) {
789      val ptr = deqPtrExt(i).value
790      val ram = DifftestMem(64L * 1024 * 1024 * 1024, 8)
791      val wen = allocated(ptr) && committed(ptr) && !mmio(ptr)
792      val waddr = ((paddrModule.io.rdata(i) - "h80000000".U) >> 3).asUInt
793      val wdata = Mux(paddrModule.io.rdata(i)(3), dataModule.io.rdata(i).data(127, 64), dataModule.io.rdata(i).data(63, 0))
794      val wmask = Mux(paddrModule.io.rdata(i)(3), dataModule.io.rdata(i).mask(15, 8), dataModule.io.rdata(i).mask(7, 0))
795      when (wen) {
796        ram.write(waddr, wdata.asTypeOf(Vec(8, UInt(8.W))), wmask.asBools)
797      }
798    }
799  }
800
801  // Read vaddr for mem exception
802  io.exceptionAddr.vaddr := vaddrModule.io.rdata(EnsbufferWidth)
803  io.exceptionAddr.gpaddr := gpaddrModule.io.rdata(EnsbufferWidth)
804  // misprediction recovery / exception redirect
805  // invalidate sq term using robIdx
806  val needCancel = Wire(Vec(StoreQueueSize, Bool()))
807  for (i <- 0 until StoreQueueSize) {
808    needCancel(i) := uop(i).robIdx.needFlush(io.brqRedirect) && allocated(i) && !committed(i)
809    when (needCancel(i)) {
810      allocated(i) := false.B
811    }
812  }
813
814 /**
815* update pointers
816**/
817  val lastEnqCancel = PopCount(RegNext(VecInit(canEnqueue.zip(enqCancel).map(x => x._1 && x._2)))) // 1 cycle after redirect
818  val lastCycleCancelCount = PopCount(RegNext(needCancel)) // 1 cycle after redirect
819  val lastCycleRedirect = RegNext(io.brqRedirect.valid) // 1 cycle after redirect
820  val enqNumber = Mux(!lastCycleRedirect&&io.enq.canAccept && io.enq.lqCanAccept, PopCount(io.enq.req.map(_.valid)), 0.U) // 1 cycle after redirect
821
822  val lastlastCycleRedirect=RegNext(lastCycleRedirect)// 2 cycle after redirect
823  val redirectCancelCount = RegEnable(lastCycleCancelCount + lastEnqCancel, lastCycleRedirect) // 2 cycle after redirect
824
825  when (lastlastCycleRedirect) {
826    // we recover the pointers in 2 cycle after redirect for better timing
827    enqPtrExt := VecInit(enqPtrExt.map(_ - redirectCancelCount))
828  }.otherwise {
829    // lastCycleRedirect.valid or nornal case
830    // when lastCycleRedirect.valid, enqNumber === 0.U, enqPtrExt will not change
831    enqPtrExt := VecInit(enqPtrExt.map(_ + enqNumber))
832  }
833  assert(!(lastCycleRedirect && enqNumber =/= 0.U))
834
835  deqPtrExt := deqPtrExtNext
836  rdataPtrExt := rdataPtrExtNext
837
838  // val dequeueCount = Mux(io.sbuffer(1).fire, 2.U, Mux(io.sbuffer(0).fire || io.mmioStout.fire, 1.U, 0.U))
839
840  // If redirect at T0, sqCancelCnt is at T2
841  io.sqCancelCnt := redirectCancelCount
842  val ForceWriteUpper = Wire(UInt(log2Up(StoreQueueSize + 1).W))
843  ForceWriteUpper := Constantin.createRecord("ForceWriteUpper_"+p(XSCoreParamsKey).HartId.toString(), initValue = 60.U)
844  val ForceWriteLower = Wire(UInt(log2Up(StoreQueueSize + 1).W))
845  ForceWriteLower := Constantin.createRecord("ForceWriteLower_"+p(XSCoreParamsKey).HartId.toString(), initValue = 55.U)
846
847  val valid_cnt = PopCount(allocated)
848  io.force_write := RegNext(Mux(valid_cnt >= ForceWriteUpper, true.B, valid_cnt >= ForceWriteLower && io.force_write), init = false.B)
849
850  // io.sqempty will be used by sbuffer
851  // We delay it for 1 cycle for better timing
852  // When sbuffer need to check if it is empty, the pipeline is blocked, which means delay io.sqempty
853  // for 1 cycle will also promise that sq is empty in that cycle
854  io.sqEmpty := RegNext(
855    enqPtrExt(0).value === deqPtrExt(0).value &&
856    enqPtrExt(0).flag === deqPtrExt(0).flag
857  )
858  // perf counter
859  QueuePerf(StoreQueueSize, validCount, !allowEnqueue)
860  io.sqFull := !allowEnqueue
861  XSPerfAccumulate("mmioCycle", uncacheState =/= s_idle) // lq is busy dealing with uncache req
862  XSPerfAccumulate("mmioCnt", io.uncache.req.fire)
863  XSPerfAccumulate("mmio_wb_success", io.mmioStout.fire)
864  XSPerfAccumulate("mmio_wb_blocked", io.mmioStout.valid && !io.mmioStout.ready)
865  XSPerfAccumulate("validEntryCnt", distanceBetween(enqPtrExt(0), deqPtrExt(0)))
866  XSPerfAccumulate("cmtEntryCnt", distanceBetween(cmtPtrExt(0), deqPtrExt(0)))
867  XSPerfAccumulate("nCmtEntryCnt", distanceBetween(enqPtrExt(0), cmtPtrExt(0)))
868
869  val perfValidCount = distanceBetween(enqPtrExt(0), deqPtrExt(0))
870  val perfEvents = Seq(
871    ("mmioCycle      ", uncacheState =/= s_idle),
872    ("mmioCnt        ", io.uncache.req.fire),
873    ("mmio_wb_success", io.mmioStout.fire),
874    ("mmio_wb_blocked", io.mmioStout.valid && !io.mmioStout.ready),
875    ("stq_1_4_valid  ", (perfValidCount < (StoreQueueSize.U/4.U))),
876    ("stq_2_4_valid  ", (perfValidCount > (StoreQueueSize.U/4.U)) & (perfValidCount <= (StoreQueueSize.U/2.U))),
877    ("stq_3_4_valid  ", (perfValidCount > (StoreQueueSize.U/2.U)) & (perfValidCount <= (StoreQueueSize.U*3.U/4.U))),
878    ("stq_4_4_valid  ", (perfValidCount > (StoreQueueSize.U*3.U/4.U))),
879  )
880  generatePerfEvent()
881
882  // debug info
883  XSDebug("enqPtrExt %d:%d deqPtrExt %d:%d\n", enqPtrExt(0).flag, enqPtr, deqPtrExt(0).flag, deqPtr)
884
885  def PrintFlag(flag: Bool, name: String): Unit = {
886    when(flag) {
887      XSDebug(false, true.B, name)
888    }.otherwise {
889      XSDebug(false, true.B, " ")
890    }
891  }
892
893  for (i <- 0 until StoreQueueSize) {
894    XSDebug(i + ": pc %x va %x pa %x data %x ",
895      uop(i).pc,
896      debug_vaddr(i),
897      debug_paddr(i),
898      debug_data(i)
899    )
900    PrintFlag(allocated(i), "a")
901    PrintFlag(allocated(i) && addrvalid(i), "a")
902    PrintFlag(allocated(i) && datavalid(i), "d")
903    PrintFlag(allocated(i) && committed(i), "c")
904    PrintFlag(allocated(i) && pending(i), "p")
905    PrintFlag(allocated(i) && mmio(i), "m")
906    XSDebug(false, true.B, "\n")
907  }
908
909}
910