xref: /XiangShan/src/main/scala/xiangshan/mem/lsqueue/LoadQueue.scala (revision a4e57ea3a91431261d57a58df4810c0d9f0366ef)
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 chipsalliance.rocketchip.config.Parameters
20import chisel3._
21import chisel3.util._
22import utils._
23import xiangshan._
24import xiangshan.backend.fu.fpu.FPU
25import xiangshan.backend.rob.RobLsqIO
26import xiangshan.cache._
27import xiangshan.frontend.FtqPtr
28
29
30class LqPtr(implicit p: Parameters) extends CircularQueuePtr[LqPtr](
31  p => p(XSCoreParamsKey).LoadQueueSize
32){
33  override def cloneType = (new LqPtr).asInstanceOf[this.type]
34}
35
36object LqPtr {
37  def apply(f: Bool, v: UInt)(implicit p: Parameters): LqPtr = {
38    val ptr = Wire(new LqPtr)
39    ptr.flag := f
40    ptr.value := v
41    ptr
42  }
43}
44
45trait HasLoadHelper { this: XSModule =>
46  def rdataHelper(uop: MicroOp, rdata: UInt): UInt = {
47    val fpWen = uop.ctrl.fpWen
48    LookupTree(uop.ctrl.fuOpType, List(
49      LSUOpType.lb   -> SignExt(rdata(7, 0) , XLEN),
50      LSUOpType.lh   -> SignExt(rdata(15, 0), XLEN),
51      /*
52          riscv-spec-20191213: 12.2 NaN Boxing of Narrower Values
53          Any operation that writes a narrower result to an f register must write
54          all 1s to the uppermost FLEN−n bits to yield a legal NaN-boxed value.
55      */
56      LSUOpType.lw   -> Mux(fpWen, FPU.box(rdata, FPU.S), SignExt(rdata(31, 0), XLEN)),
57      LSUOpType.ld   -> Mux(fpWen, FPU.box(rdata, FPU.D), SignExt(rdata(63, 0), XLEN)),
58      LSUOpType.lbu  -> ZeroExt(rdata(7, 0) , XLEN),
59      LSUOpType.lhu  -> ZeroExt(rdata(15, 0), XLEN),
60      LSUOpType.lwu  -> ZeroExt(rdata(31, 0), XLEN),
61    ))
62  }
63}
64
65class LqEnqIO(implicit p: Parameters) extends XSBundle {
66  val canAccept = Output(Bool())
67  val sqCanAccept = Input(Bool())
68  val needAlloc = Vec(exuParameters.LsExuCnt, Input(Bool()))
69  val req = Vec(exuParameters.LsExuCnt, Flipped(ValidIO(new MicroOp)))
70  val resp = Vec(exuParameters.LsExuCnt, Output(new LqPtr))
71}
72
73class LqTriggerIO(implicit p: Parameters) extends XSBundle {
74  val hitLoadAddrTriggerHitVec = Input(Vec(3, Bool()))
75  val lqLoadAddrTriggerHitVec = Output(Vec(3, Bool()))
76}
77
78// Load Queue
79class LoadQueue(implicit p: Parameters) extends XSModule
80  with HasDCacheParameters
81  with HasCircularQueuePtrHelper
82  with HasLoadHelper
83  with HasPerfEvents
84{
85  val io = IO(new Bundle() {
86    val enq = new LqEnqIO
87    val brqRedirect = Flipped(ValidIO(new Redirect))
88    val loadIn = Vec(LoadPipelineWidth, Flipped(Valid(new LsPipelineBundle)))
89    val storeIn = Vec(StorePipelineWidth, Flipped(Valid(new LsPipelineBundle)))
90    val loadDataForwarded = Vec(LoadPipelineWidth, Input(Bool()))
91    val needReplayFromRS = Vec(LoadPipelineWidth, Input(Bool()))
92    val ldout = Vec(2, DecoupledIO(new ExuOutput)) // writeback int load
93    val load_s1 = Vec(LoadPipelineWidth, Flipped(new PipeLoadForwardQueryIO)) // TODO: to be renamed
94    val loadViolationQuery = Vec(LoadPipelineWidth, Flipped(new LoadViolationQueryIO))
95    val rob = Flipped(new RobLsqIO)
96    val rollback = Output(Valid(new Redirect)) // replay now starts from load instead of store
97    val dcache = Flipped(ValidIO(new Refill)) // TODO: to be renamed
98    val release = Flipped(ValidIO(new Release))
99    val uncache = new DCacheWordIO
100    val exceptionAddr = new ExceptionAddrIO
101    val lqFull = Output(Bool())
102    val trigger = Vec(LoadPipelineWidth, new LqTriggerIO)
103  })
104
105  println("LoadQueue: size:" + LoadQueueSize)
106
107  val uop = Reg(Vec(LoadQueueSize, new MicroOp))
108  // val data = Reg(Vec(LoadQueueSize, new LsRobEntry))
109  val dataModule = Module(new LoadQueueDataWrapper(LoadQueueSize, wbNumRead = LoadPipelineWidth, wbNumWrite = LoadPipelineWidth))
110  dataModule.io := DontCare
111  val vaddrModule = Module(new SyncDataModuleTemplate(UInt(VAddrBits.W), LoadQueueSize, numRead = 3, numWrite = LoadPipelineWidth))
112  vaddrModule.io := DontCare
113  val vaddrTriggerResultModule = Module(new SyncDataModuleTemplate(Vec(3, Bool()), LoadQueueSize, numRead = LoadPipelineWidth, numWrite = LoadPipelineWidth))
114  vaddrTriggerResultModule.io := DontCare
115  val allocated = RegInit(VecInit(List.fill(LoadQueueSize)(false.B))) // lq entry has been allocated
116  val datavalid = RegInit(VecInit(List.fill(LoadQueueSize)(false.B))) // data is valid
117  val writebacked = RegInit(VecInit(List.fill(LoadQueueSize)(false.B))) // inst has been writebacked to CDB
118  val released = RegInit(VecInit(List.fill(LoadQueueSize)(false.B))) // load data has been released by dcache
119  val error = RegInit(VecInit(List.fill(LoadQueueSize)(false.B))) // load data has been corrupted
120  val miss = Reg(Vec(LoadQueueSize, Bool())) // load inst missed, waiting for miss queue to accept miss request
121  // val listening = Reg(Vec(LoadQueueSize, Bool())) // waiting for refill result
122  val pending = Reg(Vec(LoadQueueSize, Bool())) // mmio pending: inst is an mmio inst, it will not be executed until it reachs the end of rob
123  val refilling = WireInit(VecInit(List.fill(LoadQueueSize)(false.B))) // inst has been writebacked to CDB
124
125  val debug_mmio = Reg(Vec(LoadQueueSize, Bool())) // mmio: inst is an mmio inst
126  val debug_paddr = Reg(Vec(LoadQueueSize, UInt(PAddrBits.W))) // mmio: inst is an mmio inst
127
128  val enqPtrExt = RegInit(VecInit((0 until io.enq.req.length).map(_.U.asTypeOf(new LqPtr))))
129  val deqPtrExt = RegInit(0.U.asTypeOf(new LqPtr))
130  val deqPtrExtNext = Wire(new LqPtr)
131  val allowEnqueue = RegInit(true.B)
132
133  val enqPtr = enqPtrExt(0).value
134  val deqPtr = deqPtrExt.value
135
136  val deqMask = UIntToMask(deqPtr, LoadQueueSize)
137  val enqMask = UIntToMask(enqPtr, LoadQueueSize)
138
139  val commitCount = RegNext(io.rob.lcommit)
140
141  /**
142    * Enqueue at dispatch
143    *
144    * Currently, LoadQueue only allows enqueue when #emptyEntries > EnqWidth
145    */
146  io.enq.canAccept := allowEnqueue
147
148  for (i <- 0 until io.enq.req.length) {
149    val offset = if (i == 0) 0.U else PopCount(io.enq.needAlloc.take(i))
150    val lqIdx = enqPtrExt(offset)
151    val index = lqIdx.value
152    when (io.enq.req(i).valid && io.enq.canAccept && io.enq.sqCanAccept && !io.brqRedirect.valid) {
153      uop(index) := io.enq.req(i).bits
154      allocated(index) := true.B
155      datavalid(index) := false.B
156      writebacked(index) := false.B
157      released(index) := false.B
158      miss(index) := false.B
159      pending(index) := false.B
160      error(index) := false.B
161    }
162    io.enq.resp(i) := lqIdx
163  }
164  XSDebug(p"(ready, valid): ${io.enq.canAccept}, ${Binary(Cat(io.enq.req.map(_.valid)))}\n")
165
166  /**
167    * Writeback load from load units
168    *
169    * Most load instructions writeback to regfile at the same time.
170    * However,
171    *   (1) For an mmio instruction with exceptions, it writes back to ROB immediately.
172    *   (2) For an mmio instruction without exceptions, it does not write back.
173    * The mmio instruction will be sent to lower level when it reaches ROB's head.
174    * After uncache response, it will write back through arbiter with loadUnit.
175    *   (3) For cache misses, it is marked miss and sent to dcache later.
176    * After cache refills, it will write back through arbiter with loadUnit.
177    */
178  for (i <- 0 until LoadPipelineWidth) {
179    dataModule.io.wb.wen(i) := false.B
180    vaddrTriggerResultModule.io.wen(i) := false.B
181    val loadWbIndex = io.loadIn(i).bits.uop.lqIdx.value
182    when(io.loadIn(i).fire()) {
183      when(io.loadIn(i).bits.miss) {
184        XSInfo(io.loadIn(i).valid, "load miss write to lq idx %d pc 0x%x vaddr %x paddr %x data %x mask %x forwardData %x forwardMask: %x mmio %x\n",
185          io.loadIn(i).bits.uop.lqIdx.asUInt,
186          io.loadIn(i).bits.uop.cf.pc,
187          io.loadIn(i).bits.vaddr,
188          io.loadIn(i).bits.paddr,
189          io.loadIn(i).bits.data,
190          io.loadIn(i).bits.mask,
191          io.loadIn(i).bits.forwardData.asUInt,
192          io.loadIn(i).bits.forwardMask.asUInt,
193          io.loadIn(i).bits.mmio
194        )
195      }.otherwise {
196        XSInfo(io.loadIn(i).valid, "load hit write to cbd lqidx %d pc 0x%x vaddr %x paddr %x data %x mask %x forwardData %x forwardMask: %x mmio %x\n",
197        io.loadIn(i).bits.uop.lqIdx.asUInt,
198        io.loadIn(i).bits.uop.cf.pc,
199        io.loadIn(i).bits.vaddr,
200        io.loadIn(i).bits.paddr,
201        io.loadIn(i).bits.data,
202        io.loadIn(i).bits.mask,
203        io.loadIn(i).bits.forwardData.asUInt,
204        io.loadIn(i).bits.forwardMask.asUInt,
205        io.loadIn(i).bits.mmio
206      )}
207      datavalid(loadWbIndex) := (!io.loadIn(i).bits.miss || io.loadDataForwarded(i)) &&
208        !io.loadIn(i).bits.mmio && // mmio data is not valid until we finished uncache access
209        !io.needReplayFromRS(i) // do not writeback if that inst will be resend from rs
210      writebacked(loadWbIndex) := !io.loadIn(i).bits.miss && !io.loadIn(i).bits.mmio
211
212      val loadWbData = Wire(new LQDataEntry)
213      loadWbData.paddr := io.loadIn(i).bits.paddr
214      loadWbData.mask := io.loadIn(i).bits.mask
215      loadWbData.data := io.loadIn(i).bits.forwardData.asUInt // fwd data
216      loadWbData.fwdMask := io.loadIn(i).bits.forwardMask
217      dataModule.io.wbWrite(i, loadWbIndex, loadWbData)
218      dataModule.io.wb.wen(i) := true.B
219
220      vaddrTriggerResultModule.io.waddr(i) := loadWbIndex
221      vaddrTriggerResultModule.io.wdata(i) := io.trigger(i).hitLoadAddrTriggerHitVec
222      vaddrTriggerResultModule.io.wen(i) := true.B
223
224      debug_mmio(loadWbIndex) := io.loadIn(i).bits.mmio
225      debug_paddr(loadWbIndex) := io.loadIn(i).bits.paddr
226
227      val dcacheMissed = io.loadIn(i).bits.miss && !io.loadIn(i).bits.mmio
228      miss(loadWbIndex) := dcacheMissed && !io.loadDataForwarded(i) && !io.needReplayFromRS(i)
229      pending(loadWbIndex) := io.loadIn(i).bits.mmio
230      uop(loadWbIndex).debugInfo := io.loadIn(i).bits.uop.debugInfo
231      // update replayInst (replay from fetch) bit,
232      // for replayInst may be set to true in load pipeline
233      uop(loadWbIndex).ctrl.replayInst := io.loadIn(i).bits.uop.ctrl.replayInst
234    }
235    // vaddrModule write is delayed, as vaddrModule will not be read right after write
236    vaddrModule.io.waddr(i) := RegNext(loadWbIndex)
237    vaddrModule.io.wdata(i) := RegNext(io.loadIn(i).bits.vaddr)
238    vaddrModule.io.wen(i) := RegNext(io.loadIn(i).fire())
239  }
240
241  when(io.dcache.valid) {
242    XSDebug("miss resp: paddr:0x%x data %x\n", io.dcache.bits.addr, io.dcache.bits.data)
243  }
244
245  // Refill 64 bit in a cycle
246  // Refill data comes back from io.dcache.resp
247  dataModule.io.refill.valid := io.dcache.valid
248  dataModule.io.refill.paddr := io.dcache.bits.addr
249  dataModule.io.refill.data := io.dcache.bits.data
250
251  (0 until LoadQueueSize).map(i => {
252    dataModule.io.refill.refillMask(i) := allocated(i) && miss(i)
253    when(dataModule.io.refill.valid && dataModule.io.refill.refillMask(i) && dataModule.io.refill.matchMask(i)) {
254      datavalid(i) := true.B
255      miss(i) := false.B
256      refilling(i) := true.B
257      when(io.dcache.bits.error) {
258        error(i) := true.B
259      }
260    }
261  })
262
263  // Writeback up to 2 missed load insts to CDB
264  //
265  // Pick 2 missed load (data refilled), write them back to cdb
266  // 2 refilled load will be selected from even/odd entry, separately
267
268  // Stage 0
269  // Generate writeback indexes
270
271  def getEvenBits(input: UInt): UInt = {
272    VecInit((0 until LoadQueueSize/2).map(i => {input(2*i)})).asUInt
273  }
274  def getOddBits(input: UInt): UInt = {
275    VecInit((0 until LoadQueueSize/2).map(i => {input(2*i+1)})).asUInt
276  }
277
278  val loadWbSel = Wire(Vec(LoadPipelineWidth, UInt(log2Up(LoadQueueSize).W))) // index selected last cycle
279  val loadWbSelV = Wire(Vec(LoadPipelineWidth, Bool())) // index selected in last cycle is valid
280
281  val loadWbSelVec = VecInit((0 until LoadQueueSize).map(i => {
282    allocated(i) && !writebacked(i) && (datavalid(i) || refilling(i))
283  })).asUInt() // use uint instead vec to reduce verilog lines
284  val evenDeqMask = getEvenBits(deqMask)
285  val oddDeqMask = getOddBits(deqMask)
286  // generate lastCycleSelect mask
287  val evenFireMask = getEvenBits(UIntToOH(loadWbSel(0)))
288  val oddFireMask = getOddBits(UIntToOH(loadWbSel(1)))
289  // generate real select vec
290  def toVec(a: UInt): Vec[Bool] = {
291    VecInit(a.asBools)
292  }
293  val loadEvenSelVecFire = getEvenBits(loadWbSelVec) & ~evenFireMask
294  val loadOddSelVecFire = getOddBits(loadWbSelVec) & ~oddFireMask
295  val loadEvenSelVecNotFire = getEvenBits(loadWbSelVec)
296  val loadOddSelVecNotFire = getOddBits(loadWbSelVec)
297  val loadEvenSel = Mux(
298    io.ldout(0).fire(),
299    getFirstOne(toVec(loadEvenSelVecFire), evenDeqMask),
300    getFirstOne(toVec(loadEvenSelVecNotFire), evenDeqMask)
301  )
302  val loadOddSel= Mux(
303    io.ldout(1).fire(),
304    getFirstOne(toVec(loadOddSelVecFire), oddDeqMask),
305    getFirstOne(toVec(loadOddSelVecNotFire), oddDeqMask)
306  )
307
308
309  val loadWbSelGen = Wire(Vec(LoadPipelineWidth, UInt(log2Up(LoadQueueSize).W)))
310  val loadWbSelVGen = Wire(Vec(LoadPipelineWidth, Bool()))
311  loadWbSelGen(0) := Cat(loadEvenSel, 0.U(1.W))
312  loadWbSelVGen(0):= Mux(io.ldout(0).fire(), loadEvenSelVecFire.asUInt.orR, loadEvenSelVecNotFire.asUInt.orR)
313  loadWbSelGen(1) := Cat(loadOddSel, 1.U(1.W))
314  loadWbSelVGen(1) := Mux(io.ldout(1).fire(), loadOddSelVecFire.asUInt.orR, loadOddSelVecNotFire.asUInt.orR)
315
316  (0 until LoadPipelineWidth).map(i => {
317    loadWbSel(i) := RegNext(loadWbSelGen(i))
318    loadWbSelV(i) := RegNext(loadWbSelVGen(i), init = false.B)
319    when(io.ldout(i).fire()){
320      // Mark them as writebacked, so they will not be selected in the next cycle
321      writebacked(loadWbSel(i)) := true.B
322    }
323  })
324
325  // Stage 1
326  // Use indexes generated in cycle 0 to read data
327  // writeback data to cdb
328  (0 until LoadPipelineWidth).map(i => {
329    // data select
330    dataModule.io.wb.raddr(i) := loadWbSelGen(i)
331    val rdata = dataModule.io.wb.rdata(i).data
332    val seluop = uop(loadWbSel(i))
333    val func = seluop.ctrl.fuOpType
334    val raddr = dataModule.io.wb.rdata(i).paddr
335    val rdataSel = LookupTree(raddr(2, 0), List(
336      "b000".U -> rdata(63, 0),
337      "b001".U -> rdata(63, 8),
338      "b010".U -> rdata(63, 16),
339      "b011".U -> rdata(63, 24),
340      "b100".U -> rdata(63, 32),
341      "b101".U -> rdata(63, 40),
342      "b110".U -> rdata(63, 48),
343      "b111".U -> rdata(63, 56)
344    ))
345    val rdataPartialLoad = rdataHelper(seluop, rdataSel)
346
347    // writeback missed int/fp load
348    //
349    // Int load writeback will finish (if not blocked) in one cycle
350    io.ldout(i).bits.uop := seluop
351    io.ldout(i).bits.uop.lqIdx := loadWbSel(i).asTypeOf(new LqPtr)
352    io.ldout(i).bits.data := rdataPartialLoad
353    io.ldout(i).bits.redirectValid := false.B
354    io.ldout(i).bits.redirect := DontCare
355    io.ldout(i).bits.debug.isMMIO := debug_mmio(loadWbSel(i))
356    io.ldout(i).bits.debug.isPerfCnt := false.B
357    io.ldout(i).bits.debug.paddr := debug_paddr(loadWbSel(i))
358    io.ldout(i).bits.debug.vaddr := vaddrModule.io.rdata(i+1)
359    io.ldout(i).bits.fflags := DontCare
360    io.ldout(i).valid := loadWbSelV(i)
361
362    when(io.ldout(i).fire()) {
363      XSInfo("int load miss write to cbd robidx %d lqidx %d pc 0x%x mmio %x\n",
364        io.ldout(i).bits.uop.robIdx.asUInt,
365        io.ldout(i).bits.uop.lqIdx.asUInt,
366        io.ldout(i).bits.uop.cf.pc,
367        debug_mmio(loadWbSel(i))
368      )
369    }
370
371  })
372
373  /**
374    * Load commits
375    *
376    * When load commited, mark it as !allocated and move deqPtrExt forward.
377    */
378  (0 until CommitWidth).map(i => {
379    when(commitCount > i.U){
380      allocated((deqPtrExt+i.U).value) := false.B
381    }
382  })
383
384  def getFirstOne(mask: Vec[Bool], startMask: UInt) = {
385    val length = mask.length
386    val highBits = (0 until length).map(i => mask(i) & ~startMask(i))
387    val highBitsUint = Cat(highBits.reverse)
388    PriorityEncoder(Mux(highBitsUint.orR(), highBitsUint, mask.asUInt))
389  }
390
391  def getOldestInTwo(valid: Seq[Bool], uop: Seq[MicroOp]) = {
392    assert(valid.length == uop.length)
393    assert(valid.length == 2)
394    Mux(valid(0) && valid(1),
395      Mux(isAfter(uop(0).robIdx, uop(1).robIdx), uop(1), uop(0)),
396      Mux(valid(0) && !valid(1), uop(0), uop(1)))
397  }
398
399  def getAfterMask(valid: Seq[Bool], uop: Seq[MicroOp]) = {
400    assert(valid.length == uop.length)
401    val length = valid.length
402    (0 until length).map(i => {
403      (0 until length).map(j => {
404        Mux(valid(i) && valid(j),
405          isAfter(uop(i).robIdx, uop(j).robIdx),
406          Mux(!valid(i), true.B, false.B))
407      })
408    })
409  }
410
411  /**
412    * Store-Load Memory violation detection
413    *
414    * When store writes back, it searches LoadQueue for younger load instructions
415    * with the same load physical address. They loaded wrong data and need re-execution.
416    *
417    * Cycle 0: Store Writeback
418    *   Generate match vector for store address with rangeMask(stPtr, enqPtr).
419    *   Besides, load instructions in LoadUnit_S1 and S2 are also checked.
420    * Cycle 1: Redirect Generation
421    *   There're three possible types of violations, up to 6 possible redirect requests.
422    *   Choose the oldest load (part 1). (4 + 2) -> (1 + 2)
423    * Cycle 2: Redirect Fire
424    *   Choose the oldest load (part 2). (3 -> 1)
425    *   Prepare redirect request according to the detected violation.
426    *   Fire redirect request (if valid)
427    */
428
429  // stage 0:        lq l1 wb     l1 wb lq
430  //                 |  |  |      |  |  |  (paddr match)
431  // stage 1:        lq l1 wb     l1 wb lq
432  //                 |  |  |      |  |  |
433  //                 |  |------------|  |
434  //                 |        |         |
435  // stage 2:        lq      l1wb       lq
436  //                 |        |         |
437  //                 --------------------
438  //                          |
439  //                      rollback req
440  io.load_s1 := DontCare
441  def detectRollback(i: Int) = {
442    val startIndex = io.storeIn(i).bits.uop.lqIdx.value
443    val lqIdxMask = UIntToMask(startIndex, LoadQueueSize)
444    val xorMask = lqIdxMask ^ enqMask
445    val sameFlag = io.storeIn(i).bits.uop.lqIdx.flag === enqPtrExt(0).flag
446    val toEnqPtrMask = Mux(sameFlag, xorMask, ~xorMask)
447
448    // check if load already in lq needs to be rolledback
449    dataModule.io.violation(i).paddr := io.storeIn(i).bits.paddr
450    dataModule.io.violation(i).mask := io.storeIn(i).bits.mask
451    val addrMaskMatch = RegNext(dataModule.io.violation(i).violationMask)
452    val entryNeedCheck = RegNext(VecInit((0 until LoadQueueSize).map(j => {
453      allocated(j) && toEnqPtrMask(j) && (datavalid(j) || miss(j))
454    })))
455    val lqViolationVec = VecInit((0 until LoadQueueSize).map(j => {
456      addrMaskMatch(j) && entryNeedCheck(j)
457    }))
458    val lqViolation = lqViolationVec.asUInt().orR()
459    val lqViolationIndex = getFirstOne(lqViolationVec, RegNext(lqIdxMask))
460    val lqViolationUop = uop(lqViolationIndex)
461    // lqViolationUop.lqIdx.flag := deqMask(lqViolationIndex) ^ deqPtrExt.flag
462    // lqViolationUop.lqIdx.value := lqViolationIndex
463    XSDebug(lqViolation, p"${Binary(Cat(lqViolationVec))}, $startIndex, $lqViolationIndex\n")
464
465    // when l/s writeback to rob together, check if rollback is needed
466    val wbViolationVec = RegNext(VecInit((0 until LoadPipelineWidth).map(j => {
467      io.loadIn(j).valid &&
468        isAfter(io.loadIn(j).bits.uop.robIdx, io.storeIn(i).bits.uop.robIdx) &&
469        io.storeIn(i).bits.paddr(PAddrBits - 1, 3) === io.loadIn(j).bits.paddr(PAddrBits - 1, 3) &&
470        (io.storeIn(i).bits.mask & io.loadIn(j).bits.mask).orR
471    })))
472    val wbViolation = wbViolationVec.asUInt().orR()
473    val wbViolationUop = getOldestInTwo(wbViolationVec, RegNext(VecInit(io.loadIn.map(_.bits.uop))))
474    XSDebug(wbViolation, p"${Binary(Cat(wbViolationVec))}, $wbViolationUop\n")
475
476    // check if rollback is needed for load in l1
477    val l1ViolationVec = RegNext(VecInit((0 until LoadPipelineWidth).map(j => {
478      io.load_s1(j).valid && // L1 valid
479        isAfter(io.load_s1(j).uop.robIdx, io.storeIn(i).bits.uop.robIdx) &&
480        io.storeIn(i).bits.paddr(PAddrBits - 1, 3) === io.load_s1(j).paddr(PAddrBits - 1, 3) &&
481        (io.storeIn(i).bits.mask & io.load_s1(j).mask).orR
482    })))
483    val l1Violation = l1ViolationVec.asUInt().orR()
484    val l1ViolationUop = getOldestInTwo(l1ViolationVec, RegNext(VecInit(io.load_s1.map(_.uop))))
485    XSDebug(l1Violation, p"${Binary(Cat(l1ViolationVec))}, $l1ViolationUop\n")
486
487    XSDebug(
488      l1Violation,
489      "need rollback (l1 load) pc %x robidx %d target %x\n",
490      io.storeIn(i).bits.uop.cf.pc, io.storeIn(i).bits.uop.robIdx.asUInt, l1ViolationUop.robIdx.asUInt
491    )
492    XSDebug(
493      lqViolation,
494      "need rollback (ld wb before store) pc %x robidx %d target %x\n",
495      io.storeIn(i).bits.uop.cf.pc, io.storeIn(i).bits.uop.robIdx.asUInt, lqViolationUop.robIdx.asUInt
496    )
497    XSDebug(
498      wbViolation,
499      "need rollback (ld/st wb together) pc %x robidx %d target %x\n",
500      io.storeIn(i).bits.uop.cf.pc, io.storeIn(i).bits.uop.robIdx.asUInt, wbViolationUop.robIdx.asUInt
501    )
502
503    ((lqViolation, lqViolationUop), (wbViolation, wbViolationUop), (l1Violation, l1ViolationUop))
504  }
505
506  def rollbackSel(a: Valid[MicroOpRbExt], b: Valid[MicroOpRbExt]): ValidIO[MicroOpRbExt] = {
507    Mux(
508      a.valid,
509      Mux(
510        b.valid,
511        Mux(isAfter(a.bits.uop.robIdx, b.bits.uop.robIdx), b, a), // a,b both valid, sel oldest
512        a // sel a
513      ),
514      b // sel b
515    )
516  }
517  val lastCycleRedirect = RegNext(io.brqRedirect)
518  val lastlastCycleRedirect = RegNext(lastCycleRedirect)
519
520  // S2: select rollback (part1) and generate rollback request
521  // rollback check
522  // Wb/L1 rollback seq check is done in s2
523  val rollbackWb = Wire(Vec(StorePipelineWidth, Valid(new MicroOpRbExt)))
524  val rollbackL1 = Wire(Vec(StorePipelineWidth, Valid(new MicroOpRbExt)))
525  val rollbackL1Wb = Wire(Vec(StorePipelineWidth*2, Valid(new MicroOpRbExt)))
526  // Lq rollback seq check is done in s3 (next stage), as getting rollbackLq MicroOp is slow
527  val rollbackLq = Wire(Vec(StorePipelineWidth, Valid(new MicroOpRbExt)))
528  // store ftq index for store set update
529  val stFtqIdxS2 = Wire(Vec(StorePipelineWidth, new FtqPtr))
530  val stFtqOffsetS2 = Wire(Vec(StorePipelineWidth, UInt(log2Up(PredictWidth).W)))
531  for (i <- 0 until StorePipelineWidth) {
532    val detectedRollback = detectRollback(i)
533    rollbackLq(i).valid := detectedRollback._1._1 && RegNext(io.storeIn(i).valid)
534    rollbackLq(i).bits.uop := detectedRollback._1._2
535    rollbackLq(i).bits.flag := i.U
536    rollbackWb(i).valid := detectedRollback._2._1 && RegNext(io.storeIn(i).valid)
537    rollbackWb(i).bits.uop := detectedRollback._2._2
538    rollbackWb(i).bits.flag := i.U
539    rollbackL1(i).valid := detectedRollback._3._1 && RegNext(io.storeIn(i).valid)
540    rollbackL1(i).bits.uop := detectedRollback._3._2
541    rollbackL1(i).bits.flag := i.U
542    rollbackL1Wb(2*i) := rollbackL1(i)
543    rollbackL1Wb(2*i+1) := rollbackWb(i)
544    stFtqIdxS2(i) := RegNext(io.storeIn(i).bits.uop.cf.ftqPtr)
545    stFtqOffsetS2(i) := RegNext(io.storeIn(i).bits.uop.cf.ftqOffset)
546  }
547
548  val rollbackL1WbSelected = ParallelOperation(rollbackL1Wb, rollbackSel)
549  val rollbackL1WbVReg = RegNext(rollbackL1WbSelected.valid)
550  val rollbackL1WbReg = RegEnable(rollbackL1WbSelected.bits, rollbackL1WbSelected.valid)
551  val rollbackLq0VReg = RegNext(rollbackLq(0).valid)
552  val rollbackLq0Reg = RegEnable(rollbackLq(0).bits, rollbackLq(0).valid)
553  val rollbackLq1VReg = RegNext(rollbackLq(1).valid)
554  val rollbackLq1Reg = RegEnable(rollbackLq(1).bits, rollbackLq(1).valid)
555
556  // S3: select rollback (part2), generate rollback request, then fire rollback request
557  // Note that we use robIdx - 1.U to flush the load instruction itself.
558  // Thus, here if last cycle's robIdx equals to this cycle's robIdx, it still triggers the redirect.
559
560  // FIXME: this is ugly
561  val rollbackValidVec = Seq(rollbackL1WbVReg, rollbackLq0VReg, rollbackLq1VReg)
562  val rollbackUopExtVec = Seq(rollbackL1WbReg, rollbackLq0Reg, rollbackLq1Reg)
563
564  // select uop in parallel
565  val mask = getAfterMask(rollbackValidVec, rollbackUopExtVec.map(i => i.uop))
566  val oneAfterZero = mask(1)(0)
567  val rollbackUopExt = Mux(oneAfterZero && mask(2)(0),
568    rollbackUopExtVec(0),
569    Mux(!oneAfterZero && mask(2)(1), rollbackUopExtVec(1), rollbackUopExtVec(2)))
570  val stFtqIdxS3 = RegNext(stFtqIdxS2)
571  val stFtqOffsetS3 = RegNext(stFtqOffsetS2)
572  val rollbackUop = rollbackUopExt.uop
573  val rollbackStFtqIdx = stFtqIdxS3(rollbackUopExt.flag)
574  val rollbackStFtqOffset = stFtqOffsetS3(rollbackUopExt.flag)
575
576  // check if rollback request is still valid in parallel
577  val rollbackValidVecChecked = Wire(Vec(3, Bool()))
578  for(((v, uop), idx) <- rollbackValidVec.zip(rollbackUopExtVec.map(i => i.uop)).zipWithIndex) {
579    rollbackValidVecChecked(idx) := v &&
580      (!lastCycleRedirect.valid || isBefore(uop.robIdx, lastCycleRedirect.bits.robIdx)) &&
581      (!lastlastCycleRedirect.valid || isBefore(uop.robIdx, lastlastCycleRedirect.bits.robIdx))
582  }
583
584  io.rollback.bits.robIdx := rollbackUop.robIdx
585  io.rollback.bits.ftqIdx := rollbackUop.cf.ftqPtr
586  io.rollback.bits.stFtqIdx := rollbackStFtqIdx
587  io.rollback.bits.ftqOffset := rollbackUop.cf.ftqOffset
588  io.rollback.bits.stFtqOffset := rollbackStFtqOffset
589  io.rollback.bits.level := RedirectLevel.flush
590  io.rollback.bits.interrupt := DontCare
591  io.rollback.bits.cfiUpdate := DontCare
592  io.rollback.bits.cfiUpdate.target := rollbackUop.cf.pc
593  io.rollback.bits.debug_runahead_checkpoint_id := rollbackUop.debugInfo.runahead_checkpoint_id
594  // io.rollback.bits.pc := DontCare
595
596  io.rollback.valid := rollbackValidVecChecked.asUInt.orR
597
598  when(io.rollback.valid) {
599    // XSDebug("Mem rollback: pc %x robidx %d\n", io.rollback.bits.cfi, io.rollback.bits.robIdx.asUInt)
600  }
601
602  /**
603  * Load-Load Memory violation detection
604  *
605  * When load arrives load_s1, it searches LoadQueue for younger load instructions
606  * with the same load physical address. If younger load has been released (or observed),
607  * the younger load needs to be re-execed.
608  *
609  * For now, if re-exec it found to be needed in load_s1, we mark the older load as replayInst,
610  * the two loads will be replayed if the older load becomes the head of rob.
611  *
612  * When dcache releases a line, mark all writebacked entrys in load queue with
613  * the same line paddr as released.
614  */
615
616  // Load-Load Memory violation query
617  val deqRightMask = UIntToMask.rightmask(deqPtr, LoadQueueSize)
618  (0 until LoadPipelineWidth).map(i => {
619    dataModule.io.release_violation(i).paddr := io.loadViolationQuery(i).req.bits.paddr
620    io.loadViolationQuery(i).req.ready := true.B
621    io.loadViolationQuery(i).resp.valid := RegNext(io.loadViolationQuery(i).req.fire())
622    // Generate real violation mask
623    // Note that we use UIntToMask.rightmask here
624    val startIndex = io.loadViolationQuery(i).req.bits.uop.lqIdx.value
625    val lqIdxMask = UIntToMask.rightmask(startIndex, LoadQueueSize)
626    val xorMask = lqIdxMask ^ deqRightMask
627    val sameFlag = io.loadViolationQuery(i).req.bits.uop.lqIdx.flag === deqPtrExt.flag
628    val toDeqPtrMask = Mux(sameFlag, xorMask, ~xorMask)
629    val ldld_violation_mask = WireInit(VecInit((0 until LoadQueueSize).map(j => {
630      dataModule.io.release_violation(i).match_mask(j) && // addr match
631      toDeqPtrMask(j) && // the load is younger than current load
632      allocated(j) && // entry is valid
633      released(j) && // cacheline is released
634      (datavalid(j) || miss(j)) // paddr is valid
635    })))
636    dontTouch(ldld_violation_mask)
637    ldld_violation_mask.suggestName("ldldViolationMask_" + i)
638    io.loadViolationQuery(i).resp.bits.have_violation := RegNext(ldld_violation_mask.asUInt.orR)
639  })
640
641  // "released" flag update
642  //
643  // When io.release.valid, it uses the last ld-ld paddr cam port to
644  // update release flag in 1 cycle
645  when(io.release.valid){
646    // Take over ld-ld paddr cam port
647    dataModule.io.release_violation.takeRight(1)(0).paddr := io.release.bits.paddr
648    io.loadViolationQuery.takeRight(1)(0).req.ready := false.B
649    // If a load needs that cam port, replay it from rs
650  }
651
652  (0 until LoadQueueSize).map(i => {
653    when(RegNext(dataModule.io.release_violation.takeRight(1)(0).match_mask(i) &&
654      allocated(i) &&
655      writebacked(i) &&
656      io.release.valid
657    )){
658      // Note: if a load has missed in dcache and is waiting for refill in load queue,
659      // its released flag still needs to be set as true if addr matches.
660      released(i) := true.B
661    }
662  })
663
664  /**
665    * Memory mapped IO / other uncached operations
666    *
667    * States:
668    * (1) writeback from store units: mark as pending
669    * (2) when they reach ROB's head, they can be sent to uncache channel
670    * (3) response from uncache channel: mark as datavalid
671    * (4) writeback to ROB (and other units): mark as writebacked
672    * (5) ROB commits the instruction: same as normal instructions
673    */
674  //(2) when they reach ROB's head, they can be sent to uncache channel
675  val lqTailMmioPending = WireInit(pending(deqPtr))
676  val lqTailAllocated = WireInit(allocated(deqPtr))
677  val s_idle :: s_req :: s_resp :: s_wait :: Nil = Enum(4)
678  val uncacheState = RegInit(s_idle)
679  switch(uncacheState) {
680    is(s_idle) {
681      when(RegNext(io.rob.pendingld && lqTailMmioPending && lqTailAllocated)) {
682        uncacheState := s_req
683      }
684    }
685    is(s_req) {
686      when(io.uncache.req.fire()) {
687        uncacheState := s_resp
688      }
689    }
690    is(s_resp) {
691      when(io.uncache.resp.fire()) {
692        uncacheState := s_wait
693      }
694    }
695    is(s_wait) {
696      when(RegNext(io.rob.commit)) {
697        uncacheState := s_idle // ready for next mmio
698      }
699    }
700  }
701  io.uncache.req.valid := uncacheState === s_req
702
703  dataModule.io.uncache.raddr := deqPtrExtNext.value
704
705  io.uncache.req.bits.cmd  := MemoryOpConstants.M_XRD
706  io.uncache.req.bits.addr := dataModule.io.uncache.rdata.paddr
707  io.uncache.req.bits.data := dataModule.io.uncache.rdata.data
708  io.uncache.req.bits.mask := dataModule.io.uncache.rdata.mask
709
710  io.uncache.req.bits.id   := DontCare
711  io.uncache.req.bits.instrtype := DontCare
712
713  io.uncache.resp.ready := true.B
714
715  when (io.uncache.req.fire()) {
716    pending(deqPtr) := false.B
717
718    XSDebug("uncache req: pc %x addr %x data %x op %x mask %x\n",
719      uop(deqPtr).cf.pc,
720      io.uncache.req.bits.addr,
721      io.uncache.req.bits.data,
722      io.uncache.req.bits.cmd,
723      io.uncache.req.bits.mask
724    )
725  }
726
727  // (3) response from uncache channel: mark as datavalid
728  dataModule.io.uncache.wen := false.B
729  when(io.uncache.resp.fire()){
730    datavalid(deqPtr) := true.B
731    dataModule.io.uncacheWrite(deqPtr, io.uncache.resp.bits.data(XLEN-1, 0))
732    dataModule.io.uncache.wen := true.B
733
734    XSDebug("uncache resp: data %x\n", io.dcache.bits.data)
735  }
736
737  // Read vaddr for mem exception
738  // no inst will be commited 1 cycle before tval update
739  vaddrModule.io.raddr(0) := (deqPtrExt + commitCount).value
740  io.exceptionAddr.vaddr := vaddrModule.io.rdata(0)
741
742  // Read vaddr for debug
743  (0 until LoadPipelineWidth).map(i => {
744    vaddrModule.io.raddr(i+1) := loadWbSel(i)
745  })
746
747  (0 until LoadPipelineWidth).map(i => {
748    vaddrTriggerResultModule.io.raddr(i) := loadWbSelGen(i)
749    io.trigger(i).lqLoadAddrTriggerHitVec := vaddrTriggerResultModule.io.rdata(i)
750  })
751
752  // misprediction recovery / exception redirect
753  // invalidate lq term using robIdx
754  val needCancel = Wire(Vec(LoadQueueSize, Bool()))
755  for (i <- 0 until LoadQueueSize) {
756    needCancel(i) := uop(i).robIdx.needFlush(io.brqRedirect) && allocated(i)
757    when (needCancel(i)) {
758        allocated(i) := false.B
759    }
760  }
761
762  /**
763    * update pointers
764    */
765  val lastCycleCancelCount = PopCount(RegNext(needCancel))
766  // when io.brqRedirect.valid, we don't allow eneuque even though it may fire.
767  val enqNumber = Mux(io.enq.canAccept && io.enq.sqCanAccept && !io.brqRedirect.valid, PopCount(io.enq.req.map(_.valid)), 0.U)
768  when (lastCycleRedirect.valid) {
769    // we recover the pointers in the next cycle after redirect
770    enqPtrExt := VecInit(enqPtrExt.map(_ - lastCycleCancelCount))
771  }.otherwise {
772    enqPtrExt := VecInit(enqPtrExt.map(_ + enqNumber))
773  }
774
775  deqPtrExtNext := deqPtrExt + commitCount
776  deqPtrExt := deqPtrExtNext
777
778  val validCount = distanceBetween(enqPtrExt(0), deqPtrExt)
779
780  allowEnqueue := validCount + enqNumber <= (LoadQueueSize - io.enq.req.length).U
781
782  /**
783    * misc
784    */
785  // perf counter
786  QueuePerf(LoadQueueSize, validCount, !allowEnqueue)
787  io.lqFull := !allowEnqueue
788  XSPerfAccumulate("rollback", io.rollback.valid) // rollback redirect generated
789  XSPerfAccumulate("mmioCycle", uncacheState =/= s_idle) // lq is busy dealing with uncache req
790  XSPerfAccumulate("mmioCnt", io.uncache.req.fire())
791  XSPerfAccumulate("refill", io.dcache.valid)
792  XSPerfAccumulate("writeback_success", PopCount(VecInit(io.ldout.map(i => i.fire()))))
793  XSPerfAccumulate("writeback_blocked", PopCount(VecInit(io.ldout.map(i => i.valid && !i.ready))))
794  XSPerfAccumulate("utilization_miss", PopCount((0 until LoadQueueSize).map(i => allocated(i) && miss(i))))
795
796  val perfEvents = Seq(
797    ("rollback         ", io.rollback.valid                                                               ),
798    ("mmioCycle        ", uncacheState =/= s_idle                                                         ),
799    ("mmio_Cnt         ", io.uncache.req.fire()                                                           ),
800    ("refill           ", io.dcache.valid                                                                 ),
801    ("writeback_success", PopCount(VecInit(io.ldout.map(i => i.fire())))                                  ),
802    ("writeback_blocked", PopCount(VecInit(io.ldout.map(i => i.valid && !i.ready)))                       ),
803    ("ltq_1_4_valid    ", (validCount < (LoadQueueSize.U/4.U))                                            ),
804    ("ltq_2_4_valid    ", (validCount > (LoadQueueSize.U/4.U)) & (validCount <= (LoadQueueSize.U/2.U))    ),
805    ("ltq_3_4_valid    ", (validCount > (LoadQueueSize.U/2.U)) & (validCount <= (LoadQueueSize.U*3.U/4.U))),
806    ("ltq_4_4_valid    ", (validCount > (LoadQueueSize.U*3.U/4.U))                                        )
807  )
808  generatePerfEvent()
809
810  // debug info
811  XSDebug("enqPtrExt %d:%d deqPtrExt %d:%d\n", enqPtrExt(0).flag, enqPtr, deqPtrExt.flag, deqPtr)
812
813  def PrintFlag(flag: Bool, name: String): Unit = {
814    when(flag) {
815      XSDebug(false, true.B, name)
816    }.otherwise {
817      XSDebug(false, true.B, " ")
818    }
819  }
820
821  for (i <- 0 until LoadQueueSize) {
822    XSDebug(i + " pc %x pa %x ", uop(i).cf.pc, debug_paddr(i))
823    PrintFlag(allocated(i), "a")
824    PrintFlag(allocated(i) && datavalid(i), "v")
825    PrintFlag(allocated(i) && writebacked(i), "w")
826    PrintFlag(allocated(i) && miss(i), "m")
827    PrintFlag(allocated(i) && pending(i), "p")
828    XSDebug(false, true.B, "\n")
829  }
830
831}
832