xref: /XiangShan/src/main/scala/xiangshan/backend/rename/Rename.scala (revision c3f1642540a337cd2431926bddabf6267b2c2774)
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.{RobEnqIO, 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    val snptLastEnq = Flipped(ValidIO(new RobPtr))
68    val robIsEmpty = Input(Bool())
69    val toDispatchIsFp = Output(Vec(RenameWidth,Bool()))
70    val toDispatchIsInt = Output(Vec(RenameWidth,Bool()))
71    // debug arch ports
72    val debug_int_rat = if (backendParams.debugEn) Some(Vec(32, Input(UInt(PhyRegIdxWidth.W)))) else None
73    val debug_vconfig_rat = if (backendParams.debugEn) Some(Input(UInt(PhyRegIdxWidth.W))) else None
74    val debug_fp_rat = if (backendParams.debugEn) Some(Vec(32, Input(UInt(PhyRegIdxWidth.W)))) else None
75    val debug_vec_rat = if (backendParams.debugEn) Some(Vec(32, Input(UInt(PhyRegIdxWidth.W)))) else None
76    // perf only
77    val stallReason = new Bundle {
78      val in = Flipped(new StallReasonIO(RenameWidth))
79      val out = new StallReasonIO(RenameWidth)
80    }
81  })
82
83  val compressUnit = Module(new CompressUnit())
84  // create free list and rat
85  val intFreeList = Module(new MEFreeList(IntPhyRegs))
86  val fpFreeList = Module(new StdFreeList(VfPhyRegs - FpLogicRegs - VecLogicRegs))
87
88  intFreeList.io.commit    <> io.robCommits
89  intFreeList.io.debug_rat.foreach(_ <> io.debug_int_rat.get)
90  fpFreeList.io.commit     <> io.robCommits
91  fpFreeList.io.debug_rat.foreach(_ <> io.debug_fp_rat.get)
92
93  // decide if given instruction needs allocating a new physical register (CfCtrl: from decode; RobCommitInfo: from rob)
94  // fp and vec share `fpFreeList`
95  def needDestReg[T <: DecodedInst](reg_t: RegType, x: T): Bool = reg_t match {
96    case Reg_I => x.rfWen && x.ldest =/= 0.U
97    case Reg_F => x.fpWen
98    case Reg_V => x.vecWen
99  }
100  def needDestRegCommit[T <: RobCommitInfo](reg_t: RegType, x: T): Bool = {
101    reg_t match {
102      case Reg_I => x.rfWen
103      case Reg_F => x.fpWen
104      case Reg_V => x.vecWen
105    }
106  }
107  def needDestRegWalk[T <: RobCommitInfo](reg_t: RegType, x: T): Bool = {
108    reg_t match {
109      case Reg_I => x.rfWen && x.ldest =/= 0.U
110      case Reg_F => x.fpWen
111      case Reg_V => x.vecWen
112    }
113  }
114
115  // connect [redirect + walk] ports for __float point__ & __integer__ free list
116  Seq(fpFreeList, intFreeList).foreach { case fl =>
117    fl.io.redirect := io.redirect.valid
118    fl.io.walk := io.robCommits.isWalk
119  }
120  // only when both fp and int free list and dispatch1 has enough space can we do allocation
121  // when isWalk, freelist can definitely allocate
122  intFreeList.io.doAllocate := fpFreeList.io.canAllocate && io.out.map(_.ready).reduce(_ || _) || io.robCommits.isWalk
123  fpFreeList.io.doAllocate := intFreeList.io.canAllocate && io.out.map(_.ready).reduce(_ || _) || io.robCommits.isWalk
124
125  //           dispatch1 ready ++ float point free list ready ++ int free list ready      ++ not walk
126  val canOut = io.out(0).ready && fpFreeList.io.canAllocate && intFreeList.io.canAllocate && !io.robCommits.isWalk
127
128  compressUnit.io.in.zip(io.in).foreach{ case(sink, source) =>
129    sink.valid := source.valid
130    sink.bits := source.bits
131  }
132  val needRobFlags = compressUnit.io.out.needRobFlags
133  val instrSizesVec = compressUnit.io.out.instrSizes
134  val compressMasksVec = compressUnit.io.out.masks
135
136  // speculatively assign the instruction with an robIdx
137  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)
138  val robIdxHead = RegInit(0.U.asTypeOf(new RobPtr))
139  val lastCycleMisprediction = RegNext(io.redirect.valid && !io.redirect.bits.flushItself())
140  val robIdxHeadNext = Mux(io.redirect.valid, io.redirect.bits.robIdx, // redirect: move ptr to given rob index
141         Mux(lastCycleMisprediction, robIdxHead + 1.U, // mis-predict: not flush robIdx itself
142                         Mux(canOut && io.in(0).fire, robIdxHead + validCount, // instructions successfully entered next stage: increase robIdx
143                      /* default */  robIdxHead))) // no instructions passed by this cycle: stick to old value
144  robIdxHead := robIdxHeadNext
145
146  /**
147    * Rename: allocate free physical register and update rename table
148    */
149  val uops = Wire(Vec(RenameWidth, new DynInst))
150  uops.foreach( uop => {
151    uop.srcState      := DontCare
152    uop.debugInfo     := DontCare
153    uop.lqIdx         := DontCare
154    uop.sqIdx         := DontCare
155    uop.waitForRobIdx := DontCare
156    uop.singleStep    := DontCare
157    uop.snapshot      := DontCare
158    uop.dataSource    := DontCare
159    uop.l1ExuOH       := DontCare
160  })
161
162  require(RenameWidth >= CommitWidth)
163  val needVecDest    = Wire(Vec(RenameWidth, Bool()))
164  val needFpDest     = Wire(Vec(RenameWidth, Bool()))
165  val needIntDest    = Wire(Vec(RenameWidth, Bool()))
166  val hasValid = Cat(io.in.map(_.valid)).orR
167  private val inHeadValid = io.in.head.valid
168
169  val isMove = Wire(Vec(RenameWidth, Bool()))
170  isMove zip io.in.map(_.bits) foreach {
171    case (move, in) => move := Mux(in.exceptionVec.asUInt.orR, false.B, in.isMove)
172  }
173
174  val walkNeedIntDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
175  val walkNeedFpDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
176  val walkNeedVecDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
177  val walkIsMove = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
178
179  val intSpecWen = Wire(Vec(RenameWidth, Bool()))
180  val fpSpecWen  = Wire(Vec(RenameWidth, Bool()))
181  val vecSpecWen = Wire(Vec(RenameWidth, Bool()))
182
183  val walkIntSpecWen = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
184
185  val walkPdest = Wire(Vec(RenameWidth, UInt(PhyRegIdxWidth.W)))
186
187  val hasInstr = RegInit(false.B)
188  hasInstr := io.in.head.valid && !io.in.head.ready
189  val outFireNum = RegInit(0.U(RenameWidth.U.getWidth.W))
190  val outFireNumNext = Mux(io.in.head.fire || io.redirect.valid, 0.U, outFireNum + PopCount(io.out.map(_.fire)))
191  outFireNum := outFireNumNext
192  val inValidNum = PopCount(io.in.map(_.valid))
193  val allOut = inValidNum === outFireNum + PopCount(io.out.map(_.fire))
194  val outValidMask = Wire(Vec(RenameWidth, Bool()))
195  outValidMask.zipWithIndex.map{ case(m,i) =>
196    m := Mux(hasInstr, Mux(outFireNum > PopCount(io.in.map(_.valid).take(i)), false.B, true.B), true.B)
197  }
198  val validWaitForward = io.in.map(_.bits.waitForward).zip(outValidMask).map(x => x._1 && x._2)
199  val isWaitForward = VecInit((0 until RenameWidth).map(i => validWaitForward.take(i).fold(false.B)(_ || _)))
200  // uop calculation
201  for (i <- 0 until RenameWidth) {
202    for ((name, data) <- uops(i).elements) {
203      if (io.in(i).bits.elements.contains(name)) {
204        data := io.in(i).bits.elements(name)
205      }
206    }
207
208    // update cf according to ssit result
209    uops(i).storeSetHit := io.ssit(i).valid
210    uops(i).loadWaitStrict := io.ssit(i).strict && io.ssit(i).valid
211    uops(i).ssid := io.ssit(i).ssid
212
213    // update cf according to waittable result
214    uops(i).loadWaitBit := io.waittable(i)
215
216    uops(i).replayInst := false.B // set by IQ or MemQ
217    // alloc a new phy reg, fp and vec share the `fpFreeList`
218    needVecDest   (i) := io.in(i).valid && io.out(i).fire && needDestReg(Reg_V,io.in(i).bits) && outValidMask(i) && fpFreeList.io.canAllocate && fpFreeList.io.doAllocate && !io.robCommits.isWalk && !io.redirect.valid
219    needFpDest    (i) := io.in(i).valid && io.out(i).fire && needDestReg(Reg_F,io.in(i).bits) && outValidMask(i) && fpFreeList.io.canAllocate && fpFreeList.io.doAllocate && !io.robCommits.isWalk && !io.redirect.valid
220    needIntDest   (i) := io.in(i).valid && io.out(i).fire && needDestReg(Reg_I,io.in(i).bits) && outValidMask(i) && intFreeList.io.canAllocate && intFreeList.io.doAllocate && !io.robCommits.isWalk && !io.redirect.valid
221    if (i < CommitWidth) {
222      walkNeedIntDest(i) := io.robCommits.walkValid(i) && needDestRegWalk(Reg_I, io.robCommits.info(i))
223      walkNeedFpDest(i) := io.robCommits.walkValid(i) && needDestRegWalk(Reg_F, io.robCommits.info(i))
224      walkNeedVecDest(i) := io.robCommits.walkValid(i) && needDestRegWalk(Reg_V, io.robCommits.info(i))
225      walkIsMove(i) := io.robCommits.info(i).isMove
226    }
227    fpFreeList.io.allocateReq(i) := needFpDest(i) || needVecDest(i)
228    fpFreeList.io.walkReq(i) := walkNeedFpDest(i) || walkNeedVecDest(i)
229    intFreeList.io.allocateReq(i) := needIntDest(i) && !isMove(i)
230    intFreeList.io.walkReq(i) := walkNeedIntDest(i) && !walkIsMove(i)
231
232    // no valid instruction from decode stage || all resources (dispatch1 + both free lists) ready
233    io.in(i).ready := !hasValid || (canOut && allOut)
234
235    uops(i).robIdx := robIdxHead + PopCount(io.in.zip(needRobFlags).take(i).map{ case(in, needRobFlag) => in.valid && in.bits.lastUop && needRobFlag})
236    uops(i).instrSize := instrSizesVec(i)
237    when(isMove(i)) {
238      uops(i).numUops := 0.U
239      uops(i).numWB := 0.U
240    }
241    if (i > 0) {
242      when(!needRobFlags(i - 1)) {
243        uops(i).firstUop := false.B
244        uops(i).ftqPtr := uops(i - 1).ftqPtr
245        uops(i).ftqOffset := uops(i - 1).ftqOffset
246        uops(i).numUops := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
247        uops(i).numWB := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
248      }
249    }
250    when(!needRobFlags(i)) {
251      uops(i).lastUop := false.B
252      uops(i).numUops := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
253      uops(i).numWB := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
254    }
255    uops(i).wfflags := (compressMasksVec(i) & Cat(io.in.map(_.bits.wfflags).reverse)).orR
256    uops(i).dirtyFs := (compressMasksVec(i) & Cat(io.in.map(_.bits.fpWen).reverse)).orR
257
258    uops(i).psrc(0) := Mux1H(uops(i).srcType(0), Seq(io.intReadPorts(i)(0), io.fpReadPorts(i)(0), io.vecReadPorts(i)(0)))
259    uops(i).psrc(1) := Mux1H(uops(i).srcType(1), Seq(io.intReadPorts(i)(1), io.fpReadPorts(i)(1), io.vecReadPorts(i)(1)))
260    uops(i).psrc(2) := Mux1H(uops(i).srcType(2)(2, 1), Seq(io.fpReadPorts(i)(2), io.vecReadPorts(i)(2)))
261    uops(i).psrc(3) := io.vecReadPorts(i)(3)
262    uops(i).psrc(4) := io.vecReadPorts(i)(4) // Todo: vl read port
263
264    // int psrc2 should be bypassed from next instruction if it is fused
265    if (i < RenameWidth - 1) {
266      when (io.fusionInfo(i).rs2FromRs2 || io.fusionInfo(i).rs2FromRs1) {
267        uops(i).psrc(1) := Mux(io.fusionInfo(i).rs2FromRs2, io.intReadPorts(i + 1)(1), io.intReadPorts(i + 1)(0))
268      }.elsewhen(io.fusionInfo(i).rs2FromZero) {
269        uops(i).psrc(1) := 0.U
270      }
271    }
272    uops(i).eliminatedMove := isMove(i)
273
274    // update pdest
275    uops(i).pdest := MuxCase(0.U, Seq(
276      (needIntDest(i) && !isMove(i))    -> intFreeList.io.allocatePhyReg(i),
277      (needFpDest(i) || needVecDest(i)) -> fpFreeList.io.allocatePhyReg(i),
278    ))
279
280    // Assign performance counters
281    uops(i).debugInfo.renameTime := GTimer()
282
283    dontTouch(isWaitForward)
284    io.out(i).valid := !isWaitForward(i) && (!io.in(i).bits.waitForward || (io.in(i).bits.waitForward && io.robIsEmpty) ) && outValidMask(i) && io.in(i).valid && io.out(i).ready && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && !io.robCommits.isWalk
285    io.out(i).bits := uops(i)
286    // Todo: move these shit in decode stage
287    // dirty code for fence. The lsrc is passed by imm.
288    when (io.out(i).bits.fuType === FuType.fence.U) {
289      io.out(i).bits.imm := Cat(io.in(i).bits.lsrc(1), io.in(i).bits.lsrc(0))
290    }
291
292    // dirty code for SoftPrefetch (prefetch.r/prefetch.w)
293//    when (io.in(i).bits.isSoftPrefetch) {
294//      io.out(i).bits.fuType := FuType.ldu.U
295//      io.out(i).bits.fuOpType := Mux(io.in(i).bits.lsrc(1) === 1.U, LSUOpType.prefetch_r, LSUOpType.prefetch_w)
296//      io.out(i).bits.selImm := SelImm.IMM_S
297//      io.out(i).bits.imm := Cat(io.in(i).bits.imm(io.in(i).bits.imm.getWidth - 1, 5), 0.U(5.W))
298//    }
299
300    // dirty code for lui+addi(w) fusion
301    if (i < RenameWidth - 1) {
302      val fused_lui32 = io.in(i).bits.selImm === SelImm.IMM_LUI32 && io.in(i).bits.fuType === FuType.alu.U
303      when (fused_lui32) {
304        val lui_imm = io.in(i).bits.imm(19, 0)
305        val add_imm = io.in(i + 1).bits.imm(11, 0)
306        io.out(i).bits.imm := Imm_LUI_LOAD().immFromLuiLoad(lui_imm, add_imm)
307        val lsrcWidth = uops(i).lsrc.head.getWidth
308        val lui_imm_in_imm = ImmUnion.maxLen - Imm_I().len
309        val left_lui_imm = Imm_U().len - lui_imm_in_imm
310        require(2 * lsrcWidth >= left_lui_imm, "cannot fused lui and addi(w) with lsrc")
311        io.out(i).bits.lsrc(0) := lui_imm(lui_imm_in_imm + lsrcWidth - 1, lui_imm_in_imm)
312        io.out(i).bits.lsrc(1) := lui_imm(lui_imm.getWidth - 1, lui_imm_in_imm + lsrcWidth)
313      }
314    }
315
316    // write speculative rename table
317    // we update rat later inside commit code
318    intSpecWen(i) := needIntDest(i) && intFreeList.io.canAllocate && intFreeList.io.doAllocate && !io.robCommits.isWalk && !io.redirect.valid
319    fpSpecWen(i) := needFpDest(i) && fpFreeList.io.canAllocate && fpFreeList.io.doAllocate && !io.robCommits.isWalk && !io.redirect.valid
320    vecSpecWen(i) := needVecDest(i) && fpFreeList.io.canAllocate && fpFreeList.io.doAllocate && !io.robCommits.isWalk && !io.redirect.valid
321    io.toDispatchIsFp := fpSpecWen
322    io.toDispatchIsInt := intSpecWen
323
324    if (i < CommitWidth) {
325      walkIntSpecWen(i) := walkNeedIntDest(i) && !io.redirect.valid
326      walkPdest(i) := io.robCommits.info(i).pdest
327    } else {
328      walkPdest(i) := io.out(i).bits.pdest
329    }
330  }
331
332  /**
333    * How to set psrc:
334    * - bypass the pdest to psrc if previous instructions write to the same ldest as lsrc
335    * - default: psrc from RAT
336    * How to set pdest:
337    * - Mux(isMove, psrc, pdest_from_freelist).
338    *
339    * The critical path of rename lies here:
340    * When move elimination is enabled, we need to update the rat with psrc.
341    * However, psrc maybe comes from previous instructions' pdest, which comes from freelist.
342    *
343    * If we expand these logic for pdest(N):
344    * pdest(N) = Mux(isMove(N), psrc(N), freelist_out(N))
345    *          = Mux(isMove(N), Mux(bypass(N, N - 1), pdest(N - 1),
346    *                           Mux(bypass(N, N - 2), pdest(N - 2),
347    *                           ...
348    *                           Mux(bypass(N, 0),     pdest(0),
349    *                                                 rat_out(N))...)),
350    *                           freelist_out(N))
351    */
352  // a simple functional model for now
353  io.out(0).bits.pdest := Mux(isMove(0), uops(0).psrc.head, uops(0).pdest)
354
355  // psrc(n) + pdest(1)
356  val bypassCond: Vec[MixedVec[UInt]] = Wire(Vec(numRegSrc + 1, MixedVec(List.tabulate(RenameWidth-1)(i => UInt((i+1).W)))))
357  require(io.in(0).bits.srcType.size == io.in(0).bits.numSrc)
358  private val pdestLoc = io.in.head.bits.srcType.size // 2 vector src: v0, vl&vtype
359  println(s"[Rename] idx of pdest in bypassCond $pdestLoc")
360  for (i <- 1 until RenameWidth) {
361    val vecCond = io.in(i).bits.srcType.map(_ === SrcType.vp) :+ needVecDest(i)
362    val fpCond  = io.in(i).bits.srcType.map(_ === SrcType.fp) :+ needFpDest(i)
363    val intCond = io.in(i).bits.srcType.map(_ === SrcType.xp) :+ needIntDest(i)
364    val target = io.in(i).bits.lsrc :+ io.in(i).bits.ldest
365    for (((((cond1, cond2), cond3), t), j) <- vecCond.zip(fpCond).zip(intCond).zip(target).zipWithIndex) {
366      val destToSrc = io.in.take(i).zipWithIndex.map { case (in, j) =>
367        val indexMatch = in.bits.ldest === t
368        val writeMatch =  cond3 && needIntDest(j) || cond2 && needFpDest(j) || cond1 && needVecDest(j)
369        indexMatch && writeMatch
370      }
371      bypassCond(j)(i - 1) := VecInit(destToSrc).asUInt
372    }
373    io.out(i).bits.psrc(0) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(0)(i-1).asBools).foldLeft(uops(i).psrc(0)) {
374      (z, next) => Mux(next._2, next._1, z)
375    }
376    io.out(i).bits.psrc(1) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(1)(i-1).asBools).foldLeft(uops(i).psrc(1)) {
377      (z, next) => Mux(next._2, next._1, z)
378    }
379    io.out(i).bits.psrc(2) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(2)(i-1).asBools).foldLeft(uops(i).psrc(2)) {
380      (z, next) => Mux(next._2, next._1, z)
381    }
382    io.out(i).bits.psrc(3) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(3)(i-1).asBools).foldLeft(uops(i).psrc(3)) {
383      (z, next) => Mux(next._2, next._1, z)
384    }
385    io.out(i).bits.psrc(4) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(4)(i-1).asBools).foldLeft(uops(i).psrc(4)) {
386      (z, next) => Mux(next._2, next._1, z)
387    }
388    io.out(i).bits.pdest := Mux(isMove(i), io.out(i).bits.psrc(0), uops(i).pdest)
389
390    // Todo: better implementation for fields reuse
391    // For fused-lui-load, load.src(0) is replaced by the imm.
392    val last_is_lui = io.in(i - 1).bits.selImm === SelImm.IMM_U && io.in(i - 1).bits.srcType(0) =/= SrcType.pc
393    val this_is_load = io.in(i).bits.fuType === FuType.ldu.U
394    val lui_to_load = io.in(i - 1).valid && io.in(i - 1).bits.ldest === io.in(i).bits.lsrc(0)
395    val fused_lui_load = last_is_lui && this_is_load && lui_to_load
396    when (fused_lui_load) {
397      // The first LOAD operand (base address) is replaced by LUI-imm and stored in {psrc, imm}
398      val lui_imm = io.in(i - 1).bits.imm(19, 0)
399      val ld_imm = io.in(i).bits.imm
400      io.out(i).bits.srcType(0) := SrcType.imm
401      io.out(i).bits.imm := Imm_LUI_LOAD().immFromLuiLoad(lui_imm, ld_imm)
402      val psrcWidth = uops(i).psrc.head.getWidth
403      val lui_imm_in_imm = 20/*Todo: uops(i).imm.getWidth*/ - Imm_I().len
404      val left_lui_imm = Imm_U().len - lui_imm_in_imm
405      require(2 * psrcWidth >= left_lui_imm, "cannot fused lui and load with psrc")
406      io.out(i).bits.psrc(0) := lui_imm(lui_imm_in_imm + psrcWidth - 1, lui_imm_in_imm)
407      io.out(i).bits.psrc(1) := lui_imm(lui_imm.getWidth - 1, lui_imm_in_imm + psrcWidth)
408    }
409
410  }
411
412  val genSnapshot = Cat(io.out.map(out => out.fire && out.bits.snapshot)).orR
413  val snapshotCtr = RegInit((4 * CommitWidth).U)
414  val notInSameSnpt = RegNext(distanceBetween(robIdxHeadNext, io.snptLastEnq.bits) >= CommitWidth.U || !io.snptLastEnq.valid)
415  val allowSnpt = if (EnableRenameSnapshot) !snapshotCtr.orR && notInSameSnpt else false.B
416  io.out.zip(io.in).foreach{ case (out, in) => out.bits.snapshot := allowSnpt && (!in.bits.preDecodeInfo.notCFI || FuType.isJump(in.bits.fuType)) && in.fire }
417  when(genSnapshot) {
418    snapshotCtr := (4 * CommitWidth).U - PopCount(io.out.map(_.fire))
419  }.elsewhen(io.out.head.fire) {
420    snapshotCtr := Mux(snapshotCtr < PopCount(io.out.map(_.fire)), 0.U, snapshotCtr - PopCount(io.out.map(_.fire)))
421  }
422
423  intFreeList.io.snpt := io.snpt
424  fpFreeList.io.snpt := io.snpt
425  intFreeList.io.snpt.snptEnq := genSnapshot
426  fpFreeList.io.snpt.snptEnq := genSnapshot
427
428  /**
429    * Instructions commit: update freelist and rename table
430    */
431  for (i <- 0 until CommitWidth) {
432    val commitValid = io.robCommits.isCommit && io.robCommits.commitValid(i)
433    val walkValid = io.robCommits.isWalk && io.robCommits.walkValid(i)
434
435    // I. RAT Update
436    // When redirect happens (mis-prediction), don't update the rename table
437    io.intRenamePorts(i).wen  := intSpecWen(i)
438    io.intRenamePorts(i).addr := uops(i).ldest
439    io.intRenamePorts(i).data := io.out(i).bits.pdest
440
441    io.fpRenamePorts(i).wen  := fpSpecWen(i)
442    io.fpRenamePorts(i).addr := uops(i).ldest
443    io.fpRenamePorts(i).data := fpFreeList.io.allocatePhyReg(i)
444
445    io.vecRenamePorts(i).wen := vecSpecWen(i)
446    io.vecRenamePorts(i).addr := uops(i).ldest
447    io.vecRenamePorts(i).data := fpFreeList.io.allocatePhyReg(i)
448
449    // II. Free List Update
450    intFreeList.io.freeReq(i) := io.int_need_free(i)
451    intFreeList.io.freePhyReg(i) := RegNext(io.int_old_pdest(i))
452    fpFreeList.io.freeReq(i)  := RegNext(commitValid && (needDestRegCommit(Reg_F, io.robCommits.info(i)) || needDestRegCommit(Reg_V, io.robCommits.info(i))))
453    fpFreeList.io.freePhyReg(i) := Mux(RegNext(needDestRegCommit(Reg_F, io.robCommits.info(i))), io.fp_old_pdest(i), io.vec_old_pdest(i))
454  }
455
456  /*
457  Debug and performance counters
458   */
459  def printRenameInfo(in: DecoupledIO[DecodedInst], out: DecoupledIO[DynInst]) = {
460    XSInfo(out.fire, p"pc:${Hexadecimal(in.bits.pc)} in(${in.valid},${in.ready}) " +
461      p"lsrc(0):${in.bits.lsrc(0)} -> psrc(0):${out.bits.psrc(0)} " +
462      p"lsrc(1):${in.bits.lsrc(1)} -> psrc(1):${out.bits.psrc(1)} " +
463      p"lsrc(2):${in.bits.lsrc(2)} -> psrc(2):${out.bits.psrc(2)} " +
464      p"ldest:${in.bits.ldest} -> pdest:${out.bits.pdest}\n"
465    )
466  }
467
468  for ((x,y) <- io.in.zip(io.out)) {
469    printRenameInfo(x, y)
470  }
471
472  val debugRedirect = RegEnable(io.redirect.bits, io.redirect.valid)
473  // bad speculation
474  val recStall = io.redirect.valid || io.robCommits.isWalk
475  val ctrlRecStall = Mux(io.redirect.valid, io.redirect.bits.debugIsCtrl, io.robCommits.isWalk && debugRedirect.debugIsCtrl)
476  val mvioRecStall = Mux(io.redirect.valid, io.redirect.bits.debugIsMemVio, io.robCommits.isWalk && debugRedirect.debugIsMemVio)
477  val otherRecStall = recStall && !(ctrlRecStall || mvioRecStall)
478  XSPerfAccumulate("recovery_stall", recStall)
479  XSPerfAccumulate("control_recovery_stall", ctrlRecStall)
480  XSPerfAccumulate("mem_violation_recovery_stall", mvioRecStall)
481  XSPerfAccumulate("other_recovery_stall", otherRecStall)
482  // freelist stall
483  val notRecStall = !io.out.head.valid && !recStall
484  val intFlStall = notRecStall && inHeadValid && !intFreeList.io.canAllocate
485  val fpFlStall = notRecStall && inHeadValid && intFreeList.io.canAllocate && !fpFreeList.io.canAllocate
486  // other stall
487  val otherStall = notRecStall && !intFlStall && !fpFlStall
488
489  io.stallReason.in.backReason.valid := io.stallReason.out.backReason.valid || !io.in.head.ready
490  io.stallReason.in.backReason.bits := Mux(io.stallReason.out.backReason.valid, io.stallReason.out.backReason.bits,
491    MuxCase(TopDownCounters.OtherCoreStall.id.U, Seq(
492      ctrlRecStall  -> TopDownCounters.ControlRecoveryStall.id.U,
493      mvioRecStall  -> TopDownCounters.MemVioRecoveryStall.id.U,
494      otherRecStall -> TopDownCounters.OtherRecoveryStall.id.U,
495      intFlStall    -> TopDownCounters.IntFlStall.id.U,
496      fpFlStall     -> TopDownCounters.FpFlStall.id.U
497    )
498  ))
499  io.stallReason.out.reason.zip(io.stallReason.in.reason).zip(io.in.map(_.valid)).foreach { case ((out, in), valid) =>
500    out := Mux(io.stallReason.in.backReason.valid, io.stallReason.in.backReason.bits, in)
501  }
502
503  XSDebug(io.robCommits.isWalk, p"Walk Recovery Enabled\n")
504  XSDebug(io.robCommits.isWalk, p"validVec:${Binary(io.robCommits.walkValid.asUInt)}\n")
505  for (i <- 0 until CommitWidth) {
506    val info = io.robCommits.info(i)
507    XSDebug(io.robCommits.isWalk && io.robCommits.walkValid(i), p"[#$i walk info] pc:${Hexadecimal(info.pc)} " +
508      p"ldest:${info.ldest} rfWen:${info.rfWen} fpWen:${info.fpWen} vecWen:${info.vecWen}")
509  }
510
511  XSDebug(p"inValidVec: ${Binary(Cat(io.in.map(_.valid)))}\n")
512
513  XSPerfAccumulate("in_valid_count", PopCount(io.in.map(_.valid)))
514  XSPerfAccumulate("in_fire_count", PopCount(io.in.map(_.fire)))
515  XSPerfAccumulate("in_valid_not_ready_count", PopCount(io.in.map(x => x.valid && !x.ready)))
516  XSPerfAccumulate("wait_cycle", !io.in.head.valid && io.out.head.ready)
517
518  // These stall reasons could overlap each other, but we configure the priority as fellows.
519  // walk stall > dispatch stall > int freelist stall > fp freelist stall
520  private val inHeadStall = io.in.head match { case x => x.valid && !x.ready }
521  private val stallForWalk      = inHeadValid &&  io.robCommits.isWalk
522  private val stallForDispatch  = inHeadValid && !io.robCommits.isWalk && !io.out(0).ready
523  private val stallForIntFL     = inHeadValid && !io.robCommits.isWalk &&  io.out(0).ready && !intFreeList.io.canAllocate
524  private val stallForFpFL      = inHeadValid && !io.robCommits.isWalk &&  io.out(0).ready &&  intFreeList.io.canAllocate && !fpFreeList.io.canAllocate
525  XSPerfAccumulate("stall_cycle",          inHeadStall)
526  XSPerfAccumulate("stall_cycle_walk",     stallForWalk)
527  XSPerfAccumulate("stall_cycle_dispatch", stallForDispatch)
528  XSPerfAccumulate("stall_cycle_int",      stallForIntFL)
529  XSPerfAccumulate("stall_cycle_fp",       stallForFpFL)
530
531  XSPerfHistogram("in_valid_range",  PopCount(io.in.map(_.valid)),  true.B, 0, DecodeWidth + 1, 1)
532  XSPerfHistogram("in_fire_range",   PopCount(io.in.map(_.fire)),   true.B, 0, DecodeWidth + 1, 1)
533  XSPerfHistogram("out_valid_range", PopCount(io.out.map(_.valid)), true.B, 0, DecodeWidth + 1, 1)
534  XSPerfHistogram("out_fire_range",  PopCount(io.out.map(_.fire)),  true.B, 0, DecodeWidth + 1, 1)
535
536  XSPerfAccumulate("move_instr_count", PopCount(io.out.map(out => out.fire && out.bits.isMove)))
537  val is_fused_lui_load = io.out.map(o => o.fire && o.bits.fuType === FuType.ldu.U && o.bits.srcType(0) === SrcType.imm)
538  XSPerfAccumulate("fused_lui_load_instr_count", PopCount(is_fused_lui_load))
539
540  val renamePerf = Seq(
541    ("rename_in                  ", PopCount(io.in.map(_.valid & io.in(0).ready ))                                                               ),
542    ("rename_waitinstr           ", PopCount((0 until RenameWidth).map(i => io.in(i).valid && !io.in(i).ready))                                  ),
543    ("rename_stall               ", inHeadStall),
544    ("rename_stall_cycle_walk    ", inHeadValid &&  io.robCommits.isWalk),
545    ("rename_stall_cycle_dispatch", inHeadValid && !io.robCommits.isWalk && !io.out(0).ready),
546    ("rename_stall_cycle_int     ", inHeadValid && !io.robCommits.isWalk &&  io.out(0).ready && !intFreeList.io.canAllocate),
547    ("rename_stall_cycle_fp      ", inHeadValid && !io.robCommits.isWalk &&  io.out(0).ready && intFreeList.io.canAllocate && !fpFreeList.io.canAllocate),
548  )
549  val intFlPerf = intFreeList.getPerfEvents
550  val fpFlPerf = fpFreeList.getPerfEvents
551  val perfEvents = renamePerf ++ intFlPerf ++ fpFlPerf
552  generatePerfEvent()
553}
554