xref: /XiangShan/src/main/scala/xiangshan/backend/rename/Rename.scala (revision 83ba63b34cf09b33c0a9e1b3203138e51af4491b)
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.backend.rename
18
19import org.chipsalliance.cde.config.Parameters
20import chisel3._
21import chisel3.util._
22import utility._
23import utils._
24import xiangshan._
25import xiangshan.backend.Bundles.{DecodedInst, DynInst}
26import xiangshan.backend.decode.{FusionDecodeInfo, ImmUnion, Imm_I, Imm_LUI_LOAD, Imm_U}
27import xiangshan.backend.fu.FuType
28import xiangshan.backend.rename.freelist._
29import xiangshan.backend.rob.RobPtr
30import xiangshan.mem.mdp._
31
32class Rename(implicit p: Parameters) extends XSModule with HasCircularQueuePtrHelper with HasPerfEvents {
33
34  // params alias
35  private val numRegSrc = backendParams.numRegSrc
36  private val numVecRegSrc = backendParams.numVecRegSrc
37  private val numVecRatPorts = numVecRegSrc + 1 // +1 dst
38
39  println(s"[Rename] numRegSrc: $numRegSrc")
40
41  val io = IO(new Bundle() {
42    val redirect = Flipped(ValidIO(new Redirect))
43    val robCommits = Input(new RobCommitIO)
44    // from decode
45    val in = Vec(RenameWidth, Flipped(DecoupledIO(new DecodedInst)))
46    val fusionInfo = Vec(DecodeWidth - 1, Flipped(new FusionDecodeInfo))
47    // ssit read result
48    val ssit = Flipped(Vec(RenameWidth, Output(new SSITEntry)))
49    // waittable read result
50    val waittable = Flipped(Vec(RenameWidth, Output(Bool())))
51    // to rename table
52    val intReadPorts = Vec(RenameWidth, Vec(3, Input(UInt(PhyRegIdxWidth.W))))
53    val fpReadPorts = Vec(RenameWidth, Vec(4, Input(UInt(PhyRegIdxWidth.W))))
54    val vecReadPorts = Vec(RenameWidth, Vec(numVecRatPorts, Input(UInt(PhyRegIdxWidth.W))))
55    val intRenamePorts = Vec(RenameWidth, Output(new RatWritePort))
56    val fpRenamePorts = Vec(RenameWidth, Output(new RatWritePort))
57    val vecRenamePorts = Vec(RenameWidth, Output(new RatWritePort))
58    // from rename table
59    val int_old_pdest = Vec(CommitWidth, Input(UInt(PhyRegIdxWidth.W)))
60    val fp_old_pdest = Vec(CommitWidth, Input(UInt(PhyRegIdxWidth.W)))
61    val vec_old_pdest = Vec(CommitWidth, Input(UInt(PhyRegIdxWidth.W)))
62    val int_need_free = Vec(CommitWidth, Input(Bool()))
63    // to dispatch1
64    val out = Vec(RenameWidth, DecoupledIO(new DynInst))
65    // for snapshots
66    val snpt = Input(new SnapshotPort)
67    // debug arch ports
68    val debug_int_rat = if (backendParams.debugEn) Some(Vec(32, Input(UInt(PhyRegIdxWidth.W)))) else None
69    val debug_vconfig_rat = if (backendParams.debugEn) Some(Input(UInt(PhyRegIdxWidth.W))) else None
70    val debug_fp_rat = if (backendParams.debugEn) Some(Vec(32, Input(UInt(PhyRegIdxWidth.W)))) else None
71    val debug_vec_rat = if (backendParams.debugEn) Some(Vec(32, Input(UInt(PhyRegIdxWidth.W)))) else None
72    // perf only
73    val stallReason = new Bundle {
74      val in = Flipped(new StallReasonIO(RenameWidth))
75      val out = new StallReasonIO(RenameWidth)
76    }
77  })
78
79  val compressUnit = Module(new CompressUnit())
80  // create free list and rat
81  val intFreeList = Module(new MEFreeList(IntPhyRegs))
82  val fpFreeList = Module(new StdFreeList(VfPhyRegs - FpLogicRegs - VecLogicRegs))
83
84  intFreeList.io.commit    <> io.robCommits
85  intFreeList.io.debug_rat.foreach(_ <> io.debug_int_rat.get)
86  fpFreeList.io.commit     <> io.robCommits
87  fpFreeList.io.debug_rat.foreach(_ <> io.debug_fp_rat.get)
88
89  // decide if given instruction needs allocating a new physical register (CfCtrl: from decode; RobCommitInfo: from rob)
90  // fp and vec share `fpFreeList`
91  def needDestReg[T <: DecodedInst](reg_t: RegType, x: T): Bool = reg_t match {
92    case Reg_I => x.rfWen && x.ldest =/= 0.U
93    case Reg_F => x.fpWen
94    case Reg_V => x.vecWen
95  }
96  def needDestRegCommit[T <: RobCommitInfo](reg_t: RegType, x: T): Bool = {
97    reg_t match {
98      case Reg_I => x.rfWen
99      case Reg_F => x.fpWen
100      case Reg_V => x.vecWen
101    }
102  }
103  def needDestRegWalk[T <: RobCommitInfo](reg_t: RegType, x: T): Bool = {
104    reg_t match {
105      case Reg_I => x.rfWen && x.ldest =/= 0.U
106      case Reg_F => x.fpWen
107      case Reg_V => x.vecWen
108    }
109  }
110
111  // connect [redirect + walk] ports for __float point__ & __integer__ free list
112  Seq(fpFreeList, intFreeList).foreach { case fl =>
113    fl.io.redirect := io.redirect.valid
114    fl.io.walk := io.robCommits.isWalk
115  }
116  // only when both fp and int free list and dispatch1 has enough space can we do allocation
117  // when isWalk, freelist can definitely allocate
118  intFreeList.io.doAllocate := fpFreeList.io.canAllocate && io.out(0).ready || io.robCommits.isWalk
119  fpFreeList.io.doAllocate := intFreeList.io.canAllocate && io.out(0).ready || io.robCommits.isWalk
120
121  //           dispatch1 ready ++ float point free list ready ++ int free list ready      ++ not walk
122  val canOut = io.out(0).ready && fpFreeList.io.canAllocate && intFreeList.io.canAllocate && !io.robCommits.isWalk
123
124  compressUnit.io.in.zip(io.in).foreach{ case(sink, source) =>
125    sink.valid := source.valid
126    sink.bits := source.bits
127  }
128  val needRobFlags = compressUnit.io.out.needRobFlags
129  val instrSizesVec = compressUnit.io.out.instrSizes
130  val compressMasksVec = compressUnit.io.out.masks
131
132  // speculatively assign the instruction with an robIdx
133  val validCount = PopCount(io.in.zip(needRobFlags).map{ case(in, needRobFlag) => in.valid && in.bits.lastUop && needRobFlag}) // number of instructions waiting to enter rob (from decode)
134  val robIdxHead = RegInit(0.U.asTypeOf(new RobPtr))
135  val lastCycleMisprediction = RegNext(io.redirect.valid && !io.redirect.bits.flushItself())
136  val robIdxHeadNext = Mux(io.redirect.valid, io.redirect.bits.robIdx, // redirect: move ptr to given rob index
137         Mux(lastCycleMisprediction, robIdxHead + 1.U, // mis-predict: not flush robIdx itself
138                         Mux(canOut, robIdxHead + validCount, // instructions successfully entered next stage: increase robIdx
139                      /* default */  robIdxHead))) // no instructions passed by this cycle: stick to old value
140  robIdxHead := robIdxHeadNext
141
142  /**
143    * Rename: allocate free physical register and update rename table
144    */
145  val uops = Wire(Vec(RenameWidth, new DynInst))
146  uops.foreach( uop => {
147    uop.srcState      := DontCare
148    uop.debugInfo     := DontCare
149    uop.lqIdx         := DontCare
150    uop.sqIdx         := DontCare
151    uop.waitForRobIdx := DontCare
152    uop.singleStep    := DontCare
153    uop.snapshot      := DontCare
154    uop.dataSource    := DontCare
155    uop.l1ExuOH       := DontCare
156  })
157
158  require(RenameWidth >= CommitWidth)
159  val needVecDest    = Wire(Vec(RenameWidth, Bool()))
160  val needFpDest     = Wire(Vec(RenameWidth, Bool()))
161  val needIntDest    = Wire(Vec(RenameWidth, Bool()))
162  val hasValid = Cat(io.in.map(_.valid)).orR
163  private val inHeadValid = io.in.head.valid
164
165  val isMove = Wire(Vec(RenameWidth, Bool()))
166  isMove zip io.in.map(_.bits) foreach {
167    case (move, in) => move := Mux(in.exceptionVec.asUInt.orR, false.B, in.isMove)
168  }
169
170  val walkNeedIntDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
171  val walkNeedFpDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
172  val walkNeedVecDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
173  val walkIsMove = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
174
175  val intSpecWen = Wire(Vec(RenameWidth, Bool()))
176  val fpSpecWen  = Wire(Vec(RenameWidth, Bool()))
177  val vecSpecWen = Wire(Vec(RenameWidth, Bool()))
178
179  val walkIntSpecWen = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
180
181  val walkPdest = Wire(Vec(RenameWidth, UInt(PhyRegIdxWidth.W)))
182
183  // uop calculation
184  for (i <- 0 until RenameWidth) {
185    for ((name, data) <- uops(i).elements) {
186      if (io.in(i).bits.elements.contains(name)) {
187        data := io.in(i).bits.elements(name)
188      }
189    }
190
191    // update cf according to ssit result
192    uops(i).storeSetHit := io.ssit(i).valid
193    uops(i).loadWaitStrict := io.ssit(i).strict && io.ssit(i).valid
194    uops(i).ssid := io.ssit(i).ssid
195
196    // update cf according to waittable result
197    uops(i).loadWaitBit := io.waittable(i)
198
199    uops(i).replayInst := false.B // set by IQ or MemQ
200    // alloc a new phy reg, fp and vec share the `fpFreeList`
201    needVecDest   (i) := io.in(i).valid && needDestReg(Reg_V,       io.in(i).bits)
202    needFpDest    (i) := io.in(i).valid && needDestReg(Reg_F,       io.in(i).bits)
203    needIntDest   (i) := io.in(i).valid && needDestReg(Reg_I,       io.in(i).bits)
204    if (i < CommitWidth) {
205      walkNeedIntDest(i) := io.robCommits.walkValid(i) && needDestRegWalk(Reg_I, io.robCommits.info(i))
206      walkNeedFpDest(i) := io.robCommits.walkValid(i) && needDestRegWalk(Reg_F, io.robCommits.info(i))
207      walkNeedVecDest(i) := io.robCommits.walkValid(i) && needDestRegWalk(Reg_V, io.robCommits.info(i))
208      walkIsMove(i) := io.robCommits.info(i).isMove
209    }
210    fpFreeList.io.allocateReq(i) := needFpDest(i) || needVecDest(i)
211    fpFreeList.io.walkReq(i) := walkNeedFpDest(i) || walkNeedVecDest(i)
212    intFreeList.io.allocateReq(i) := needIntDest(i) && !isMove(i)
213    intFreeList.io.walkReq(i) := walkNeedIntDest(i) && !walkIsMove(i)
214
215    // no valid instruction from decode stage || all resources (dispatch1 + both free lists) ready
216    io.in(i).ready := !hasValid || canOut
217
218    uops(i).robIdx := robIdxHead + PopCount(io.in.zip(needRobFlags).take(i).map{ case(in, needRobFlag) => in.valid && in.bits.lastUop && needRobFlag})
219    uops(i).instrSize := instrSizesVec(i)
220    when(isMove(i)) {
221      uops(i).numUops := 0.U
222    }
223    if (i > 0) {
224      when(!needRobFlags(i - 1)) {
225        uops(i).firstUop := false.B
226        uops(i).ftqPtr := uops(i - 1).ftqPtr
227        uops(i).ftqOffset := uops(i - 1).ftqOffset
228        uops(i).numUops := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
229      }
230    }
231    when(!needRobFlags(i)) {
232      uops(i).lastUop := false.B
233      uops(i).numUops := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
234    }
235    uops(i).wfflags := (compressMasksVec(i) & Cat(io.in.map(_.bits.wfflags).reverse)).orR
236    uops(i).dirtyFs := (compressMasksVec(i) & Cat(io.in.map(_.bits.fpWen).reverse)).orR
237
238    uops(i).psrc(0) := Mux1H(uops(i).srcType(0), Seq(io.intReadPorts(i)(0), io.fpReadPorts(i)(0), io.vecReadPorts(i)(0)))
239    uops(i).psrc(1) := Mux1H(uops(i).srcType(1), Seq(io.intReadPorts(i)(1), io.fpReadPorts(i)(1), io.vecReadPorts(i)(1)))
240    uops(i).psrc(2) := Mux1H(uops(i).srcType(2)(2, 1), Seq(io.fpReadPorts(i)(2), io.vecReadPorts(i)(2)))
241    uops(i).psrc(3) := io.vecReadPorts(i)(3)
242    uops(i).psrc(4) := io.vecReadPorts(i)(4) // Todo: vl read port
243
244    // int psrc2 should be bypassed from next instruction if it is fused
245    if (i < RenameWidth - 1) {
246      when (io.fusionInfo(i).rs2FromRs2 || io.fusionInfo(i).rs2FromRs1) {
247        uops(i).psrc(1) := Mux(io.fusionInfo(i).rs2FromRs2, io.intReadPorts(i + 1)(1), io.intReadPorts(i + 1)(0))
248      }.elsewhen(io.fusionInfo(i).rs2FromZero) {
249        uops(i).psrc(1) := 0.U
250      }
251    }
252    uops(i).eliminatedMove := isMove(i)
253
254    // update pdest
255    uops(i).pdest := MuxCase(0.U, Seq(
256      needIntDest(i)                    -> intFreeList.io.allocatePhyReg(i),
257      (needFpDest(i) || needVecDest(i)) -> fpFreeList.io.allocatePhyReg(i),
258    ))
259
260    // Assign performance counters
261    uops(i).debugInfo.renameTime := GTimer()
262
263    io.out(i).valid := io.in(i).valid && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && !io.robCommits.isWalk
264    io.out(i).bits := uops(i)
265    // Todo: move these shit in decode stage
266    // dirty code for fence. The lsrc is passed by imm.
267    when (io.out(i).bits.fuType === FuType.fence.U) {
268      io.out(i).bits.imm := Cat(io.in(i).bits.lsrc(1), io.in(i).bits.lsrc(0))
269    }
270
271    // dirty code for SoftPrefetch (prefetch.r/prefetch.w)
272//    when (io.in(i).bits.isSoftPrefetch) {
273//      io.out(i).bits.fuType := FuType.ldu.U
274//      io.out(i).bits.fuOpType := Mux(io.in(i).bits.lsrc(1) === 1.U, LSUOpType.prefetch_r, LSUOpType.prefetch_w)
275//      io.out(i).bits.selImm := SelImm.IMM_S
276//      io.out(i).bits.imm := Cat(io.in(i).bits.imm(io.in(i).bits.imm.getWidth - 1, 5), 0.U(5.W))
277//    }
278
279    // dirty code for lui+addi(w) fusion
280    if (i < RenameWidth - 1) {
281      val fused_lui32 = io.in(i).bits.selImm === SelImm.IMM_LUI32 && io.in(i).bits.fuType === FuType.alu.U
282      when (fused_lui32) {
283        val lui_imm = io.in(i).bits.imm(19, 0)
284        val add_imm = io.in(i + 1).bits.imm(11, 0)
285        io.out(i).bits.imm := Imm_LUI_LOAD().immFromLuiLoad(lui_imm, add_imm)
286        val lsrcWidth = uops(i).lsrc.head.getWidth
287        val lui_imm_in_imm = ImmUnion.maxLen - Imm_I().len
288        val left_lui_imm = Imm_U().len - lui_imm_in_imm
289        require(2 * lsrcWidth >= left_lui_imm, "cannot fused lui and addi(w) with lsrc")
290        io.out(i).bits.lsrc(0) := lui_imm(lui_imm_in_imm + lsrcWidth - 1, lui_imm_in_imm)
291        io.out(i).bits.lsrc(1) := lui_imm(lui_imm.getWidth - 1, lui_imm_in_imm + lsrcWidth)
292      }
293    }
294
295    // write speculative rename table
296    // we update rat later inside commit code
297    intSpecWen(i) := needIntDest(i) && intFreeList.io.canAllocate && intFreeList.io.doAllocate && !io.robCommits.isWalk && !io.redirect.valid
298    fpSpecWen(i) := needFpDest(i) && fpFreeList.io.canAllocate && fpFreeList.io.doAllocate && !io.robCommits.isWalk && !io.redirect.valid
299    vecSpecWen(i) := needVecDest(i) && fpFreeList.io.canAllocate && fpFreeList.io.doAllocate && !io.robCommits.isWalk && !io.redirect.valid
300
301    if (i < CommitWidth) {
302      walkIntSpecWen(i) := walkNeedIntDest(i) && !io.redirect.valid
303      walkPdest(i) := io.robCommits.info(i).pdest
304    } else {
305      walkPdest(i) := io.out(i).bits.pdest
306    }
307  }
308
309  /**
310    * How to set psrc:
311    * - bypass the pdest to psrc if previous instructions write to the same ldest as lsrc
312    * - default: psrc from RAT
313    * How to set pdest:
314    * - Mux(isMove, psrc, pdest_from_freelist).
315    *
316    * The critical path of rename lies here:
317    * When move elimination is enabled, we need to update the rat with psrc.
318    * However, psrc maybe comes from previous instructions' pdest, which comes from freelist.
319    *
320    * If we expand these logic for pdest(N):
321    * pdest(N) = Mux(isMove(N), psrc(N), freelist_out(N))
322    *          = Mux(isMove(N), Mux(bypass(N, N - 1), pdest(N - 1),
323    *                           Mux(bypass(N, N - 2), pdest(N - 2),
324    *                           ...
325    *                           Mux(bypass(N, 0),     pdest(0),
326    *                                                 rat_out(N))...)),
327    *                           freelist_out(N))
328    */
329  // a simple functional model for now
330  io.out(0).bits.pdest := Mux(isMove(0), uops(0).psrc.head, uops(0).pdest)
331
332  // psrc(n) + pdest(1)
333  val bypassCond: Vec[MixedVec[UInt]] = Wire(Vec(numRegSrc + 1, MixedVec(List.tabulate(RenameWidth-1)(i => UInt((i+1).W)))))
334  require(io.in(0).bits.srcType.size == io.in(0).bits.numSrc)
335  private val pdestLoc = io.in.head.bits.srcType.size // 2 vector src: v0, vl&vtype
336  println(s"[Rename] idx of pdest in bypassCond $pdestLoc")
337  for (i <- 1 until RenameWidth) {
338    val vecCond = io.in(i).bits.srcType.map(_ === SrcType.vp) :+ needVecDest(i)
339    val fpCond  = io.in(i).bits.srcType.map(_ === SrcType.fp) :+ needFpDest(i)
340    val intCond = io.in(i).bits.srcType.map(_ === SrcType.xp) :+ needIntDest(i)
341    val target = io.in(i).bits.lsrc :+ io.in(i).bits.ldest
342    for (((((cond1, cond2), cond3), t), j) <- vecCond.zip(fpCond).zip(intCond).zip(target).zipWithIndex) {
343      val destToSrc = io.in.take(i).zipWithIndex.map { case (in, j) =>
344        val indexMatch = in.bits.ldest === t
345        val writeMatch =  cond3 && needIntDest(j) || cond2 && needFpDest(j) || cond1 && needVecDest(j)
346        indexMatch && writeMatch
347      }
348      bypassCond(j)(i - 1) := VecInit(destToSrc).asUInt
349    }
350    io.out(i).bits.psrc(0) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(0)(i-1).asBools).foldLeft(uops(i).psrc(0)) {
351      (z, next) => Mux(next._2, next._1, z)
352    }
353    io.out(i).bits.psrc(1) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(1)(i-1).asBools).foldLeft(uops(i).psrc(1)) {
354      (z, next) => Mux(next._2, next._1, z)
355    }
356    io.out(i).bits.psrc(2) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(2)(i-1).asBools).foldLeft(uops(i).psrc(2)) {
357      (z, next) => Mux(next._2, next._1, z)
358    }
359    io.out(i).bits.psrc(3) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(3)(i-1).asBools).foldLeft(uops(i).psrc(3)) {
360      (z, next) => Mux(next._2, next._1, z)
361    }
362    io.out(i).bits.psrc(4) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(4)(i-1).asBools).foldLeft(uops(i).psrc(4)) {
363      (z, next) => Mux(next._2, next._1, z)
364    }
365    io.out(i).bits.pdest := Mux(isMove(i), io.out(i).bits.psrc(0), uops(i).pdest)
366
367    // Todo: better implementation for fields reuse
368    // For fused-lui-load, load.src(0) is replaced by the imm.
369    val last_is_lui = io.in(i - 1).bits.selImm === SelImm.IMM_U && io.in(i - 1).bits.srcType(0) =/= SrcType.pc
370    val this_is_load = io.in(i).bits.fuType === FuType.ldu.U
371    val lui_to_load = io.in(i - 1).valid && io.in(i - 1).bits.ldest === io.in(i).bits.lsrc(0)
372    val fused_lui_load = last_is_lui && this_is_load && lui_to_load
373    when (fused_lui_load) {
374      // The first LOAD operand (base address) is replaced by LUI-imm and stored in {psrc, imm}
375      val lui_imm = io.in(i - 1).bits.imm(19, 0)
376      val ld_imm = io.in(i).bits.imm
377      io.out(i).bits.srcType(0) := SrcType.imm
378      io.out(i).bits.imm := Imm_LUI_LOAD().immFromLuiLoad(lui_imm, ld_imm)
379      val psrcWidth = uops(i).psrc.head.getWidth
380      val lui_imm_in_imm = 20/*Todo: uops(i).imm.getWidth*/ - Imm_I().len
381      val left_lui_imm = Imm_U().len - lui_imm_in_imm
382      require(2 * psrcWidth >= left_lui_imm, "cannot fused lui and load with psrc")
383      io.out(i).bits.psrc(0) := lui_imm(lui_imm_in_imm + psrcWidth - 1, lui_imm_in_imm)
384      io.out(i).bits.psrc(1) := lui_imm(lui_imm.getWidth - 1, lui_imm_in_imm + psrcWidth)
385    }
386
387  }
388
389  val hasCFI = VecInit(io.in.map(in => (!in.bits.preDecodeInfo.notCFI || FuType.isJump(in.bits.fuType)) && in.fire)).asUInt.orR
390  val snapshotCtr = RegInit((4 * CommitWidth).U)
391  val allowSnpt = if (EnableRenameSnapshot) !snapshotCtr.orR else false.B
392  io.out.head.bits.snapshot := hasCFI && allowSnpt
393  when(io.out.head.fire && io.out.head.bits.snapshot) {
394    snapshotCtr := (4 * CommitWidth).U - PopCount(io.out.map(_.fire))
395  }.elsewhen(io.out.head.fire) {
396    snapshotCtr := Mux(snapshotCtr < PopCount(io.out.map(_.fire)), 0.U, snapshotCtr - PopCount(io.out.map(_.fire)))
397  }
398
399  intFreeList.io.snpt := io.snpt
400  fpFreeList.io.snpt := io.snpt
401  intFreeList.io.snpt.snptEnq := io.out.head.fire && io.out.head.bits.snapshot
402  fpFreeList.io.snpt.snptEnq := io.out.head.fire && io.out.head.bits.snapshot
403
404  /**
405    * Instructions commit: update freelist and rename table
406    */
407  for (i <- 0 until CommitWidth) {
408    val commitValid = io.robCommits.isCommit && io.robCommits.commitValid(i)
409    val walkValid = io.robCommits.isWalk && io.robCommits.walkValid(i)
410
411    // I. RAT Update
412    // When redirect happens (mis-prediction), don't update the rename table
413    io.intRenamePorts(i).wen  := intSpecWen(i)
414    io.intRenamePorts(i).addr := uops(i).ldest
415    io.intRenamePorts(i).data := io.out(i).bits.pdest
416
417    io.fpRenamePorts(i).wen  := fpSpecWen(i)
418    io.fpRenamePorts(i).addr := uops(i).ldest
419    io.fpRenamePorts(i).data := fpFreeList.io.allocatePhyReg(i)
420
421    io.vecRenamePorts(i).wen := vecSpecWen(i)
422    io.vecRenamePorts(i).addr := uops(i).ldest
423    io.vecRenamePorts(i).data := fpFreeList.io.allocatePhyReg(i)
424
425    // II. Free List Update
426    intFreeList.io.freeReq(i) := io.int_need_free(i)
427    intFreeList.io.freePhyReg(i) := RegNext(io.int_old_pdest(i))
428    fpFreeList.io.freeReq(i)  := RegNext(commitValid && (needDestRegCommit(Reg_F, io.robCommits.info(i)) || needDestRegCommit(Reg_V, io.robCommits.info(i))))
429    fpFreeList.io.freePhyReg(i) := Mux(RegNext(needDestRegCommit(Reg_F, io.robCommits.info(i))), io.fp_old_pdest(i), io.vec_old_pdest(i))
430  }
431
432  /*
433  Debug and performance counters
434   */
435  def printRenameInfo(in: DecoupledIO[DecodedInst], out: DecoupledIO[DynInst]) = {
436    XSInfo(out.fire, p"pc:${Hexadecimal(in.bits.pc)} in(${in.valid},${in.ready}) " +
437      p"lsrc(0):${in.bits.lsrc(0)} -> psrc(0):${out.bits.psrc(0)} " +
438      p"lsrc(1):${in.bits.lsrc(1)} -> psrc(1):${out.bits.psrc(1)} " +
439      p"lsrc(2):${in.bits.lsrc(2)} -> psrc(2):${out.bits.psrc(2)} " +
440      p"ldest:${in.bits.ldest} -> pdest:${out.bits.pdest}\n"
441    )
442  }
443
444  for ((x,y) <- io.in.zip(io.out)) {
445    printRenameInfo(x, y)
446  }
447
448  val debugRedirect = RegEnable(io.redirect.bits, io.redirect.valid)
449  // bad speculation
450  val recStall = io.redirect.valid || io.robCommits.isWalk
451  val ctrlRecStall = Mux(io.redirect.valid, io.redirect.bits.debugIsCtrl, io.robCommits.isWalk && debugRedirect.debugIsCtrl)
452  val mvioRecStall = Mux(io.redirect.valid, io.redirect.bits.debugIsMemVio, io.robCommits.isWalk && debugRedirect.debugIsMemVio)
453  val otherRecStall = recStall && !(ctrlRecStall || mvioRecStall)
454  XSPerfAccumulate("recovery_stall", recStall)
455  XSPerfAccumulate("control_recovery_stall", ctrlRecStall)
456  XSPerfAccumulate("mem_violation_recovery_stall", mvioRecStall)
457  XSPerfAccumulate("other_recovery_stall", otherRecStall)
458  // freelist stall
459  val notRecStall = !io.out.head.valid && !recStall
460  val intFlStall = notRecStall && inHeadValid && !intFreeList.io.canAllocate
461  val fpFlStall = notRecStall && inHeadValid && intFreeList.io.canAllocate && !fpFreeList.io.canAllocate
462  // other stall
463  val otherStall = notRecStall && !intFlStall && !fpFlStall
464
465  io.stallReason.in.backReason.valid := io.stallReason.out.backReason.valid || !io.in.head.ready
466  io.stallReason.in.backReason.bits := Mux(io.stallReason.out.backReason.valid, io.stallReason.out.backReason.bits,
467    MuxCase(TopDownCounters.OtherCoreStall.id.U, Seq(
468      ctrlRecStall  -> TopDownCounters.ControlRecoveryStall.id.U,
469      mvioRecStall  -> TopDownCounters.MemVioRecoveryStall.id.U,
470      otherRecStall -> TopDownCounters.OtherRecoveryStall.id.U,
471      intFlStall    -> TopDownCounters.IntFlStall.id.U,
472      fpFlStall     -> TopDownCounters.FpFlStall.id.U
473    )
474  ))
475  io.stallReason.out.reason.zip(io.stallReason.in.reason).zip(io.in.map(_.valid)).foreach { case ((out, in), valid) =>
476    out := Mux(io.stallReason.in.backReason.valid, io.stallReason.in.backReason.bits, in)
477  }
478
479  XSDebug(io.robCommits.isWalk, p"Walk Recovery Enabled\n")
480  XSDebug(io.robCommits.isWalk, p"validVec:${Binary(io.robCommits.walkValid.asUInt)}\n")
481  for (i <- 0 until CommitWidth) {
482    val info = io.robCommits.info(i)
483    XSDebug(io.robCommits.isWalk && io.robCommits.walkValid(i), p"[#$i walk info] pc:${Hexadecimal(info.pc)} " +
484      p"ldest:${info.ldest} rfWen:${info.rfWen} fpWen:${info.fpWen} vecWen:${info.vecWen}")
485  }
486
487  XSDebug(p"inValidVec: ${Binary(Cat(io.in.map(_.valid)))}\n")
488
489  XSPerfAccumulate("in_valid_count", PopCount(io.in.map(_.valid)))
490  XSPerfAccumulate("in_fire_count", PopCount(io.in.map(_.fire)))
491  XSPerfAccumulate("in_valid_not_ready_count", PopCount(io.in.map(x => x.valid && !x.ready)))
492  XSPerfAccumulate("wait_cycle", !io.in.head.valid && io.out.head.ready)
493
494  // These stall reasons could overlap each other, but we configure the priority as fellows.
495  // walk stall > dispatch stall > int freelist stall > fp freelist stall
496  private val inHeadStall = io.in.head match { case x => x.valid && !x.ready }
497  private val stallForWalk      = inHeadValid &&  io.robCommits.isWalk
498  private val stallForDispatch  = inHeadValid && !io.robCommits.isWalk && !io.out(0).ready
499  private val stallForIntFL     = inHeadValid && !io.robCommits.isWalk &&  io.out(0).ready && !intFreeList.io.canAllocate
500  private val stallForFpFL      = inHeadValid && !io.robCommits.isWalk &&  io.out(0).ready &&  intFreeList.io.canAllocate && !fpFreeList.io.canAllocate
501  XSPerfAccumulate("stall_cycle",          inHeadStall)
502  XSPerfAccumulate("stall_cycle_walk",     stallForWalk)
503  XSPerfAccumulate("stall_cycle_dispatch", stallForDispatch)
504  XSPerfAccumulate("stall_cycle_int",      stallForIntFL)
505  XSPerfAccumulate("stall_cycle_fp",       stallForFpFL)
506
507  XSPerfHistogram("in_valid_range",  PopCount(io.in.map(_.valid)),  true.B, 0, DecodeWidth + 1, 1)
508  XSPerfHistogram("in_fire_range",   PopCount(io.in.map(_.fire)),   true.B, 0, DecodeWidth + 1, 1)
509  XSPerfHistogram("out_valid_range", PopCount(io.out.map(_.valid)), true.B, 0, DecodeWidth + 1, 1)
510  XSPerfHistogram("out_fire_range",  PopCount(io.out.map(_.fire)),  true.B, 0, DecodeWidth + 1, 1)
511
512  XSPerfAccumulate("move_instr_count", PopCount(io.out.map(out => out.fire && out.bits.isMove)))
513  val is_fused_lui_load = io.out.map(o => o.fire && o.bits.fuType === FuType.ldu.U && o.bits.srcType(0) === SrcType.imm)
514  XSPerfAccumulate("fused_lui_load_instr_count", PopCount(is_fused_lui_load))
515
516  val renamePerf = Seq(
517    ("rename_in                  ", PopCount(io.in.map(_.valid & io.in(0).ready ))                                                               ),
518    ("rename_waitinstr           ", PopCount((0 until RenameWidth).map(i => io.in(i).valid && !io.in(i).ready))                                  ),
519    ("rename_stall               ", inHeadStall),
520    ("rename_stall_cycle_walk    ", inHeadValid &&  io.robCommits.isWalk),
521    ("rename_stall_cycle_dispatch", inHeadValid && !io.robCommits.isWalk && !io.out(0).ready),
522    ("rename_stall_cycle_int     ", inHeadValid && !io.robCommits.isWalk &&  io.out(0).ready && !intFreeList.io.canAllocate),
523    ("rename_stall_cycle_fp      ", inHeadValid && !io.robCommits.isWalk &&  io.out(0).ready && intFreeList.io.canAllocate && !fpFreeList.io.canAllocate),
524  )
525  val intFlPerf = intFreeList.getPerfEvents
526  val fpFlPerf = fpFreeList.getPerfEvents
527  val perfEvents = renamePerf ++ intFlPerf ++ fpFlPerf
528  generatePerfEvent()
529}
530