xref: /XiangShan/src/main/scala/xiangshan/backend/rename/Rename.scala (revision f7063a43ab34da917ba6c670d21871314340c550)
1/***************************************************************************************
2* Copyright (c) 2020-2021 Institute of Computing Technology, Chinese Academy of Sciences
3* Copyright (c) 2020-2021 Peng Cheng Laboratory
4*
5* XiangShan is licensed under Mulan PSL v2.
6* You can use this software according to the terms and conditions of the Mulan PSL v2.
7* You may obtain a copy of Mulan PSL v2 at:
8*          http://license.coscl.org.cn/MulanPSL2
9*
10* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
11* EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
12* MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
13*
14* See the Mulan PSL v2 for more details.
15***************************************************************************************/
16
17package xiangshan.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
38
39  println(s"[Rename] numRegSrc: $numRegSrc")
40
41  val io = IO(new Bundle() {
42    val redirect = Flipped(ValidIO(new Redirect))
43    val rabCommits = Input(new RabCommitIO)
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(2, Input(UInt(PhyRegIdxWidth.W))))
53    val fpReadPorts = Vec(RenameWidth, Vec(3, 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(RabCommitWidth, Input(UInt(PhyRegIdxWidth.W)))
60    val fp_old_pdest = Vec(RabCommitWidth, Input(UInt(PhyRegIdxWidth.W)))
61    val vec_old_pdest = Vec(RabCommitWidth, Input(UInt(PhyRegIdxWidth.W)))
62    val int_need_free = Vec(RabCommitWidth, 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 snptIsFull= Input(Bool())
69    // debug arch ports
70    val debug_int_rat = if (backendParams.debugEn) Some(Vec(32, Input(UInt(PhyRegIdxWidth.W)))) else None
71    val debug_vconfig_rat = if (backendParams.debugEn) Some(Input(UInt(PhyRegIdxWidth.W))) else None
72    val debug_fp_rat = if (backendParams.debugEn) Some(Vec(32, Input(UInt(PhyRegIdxWidth.W)))) else None
73    val debug_vec_rat = if (backendParams.debugEn) Some(Vec(32, Input(UInt(PhyRegIdxWidth.W)))) else None
74    // perf only
75    val stallReason = new Bundle {
76      val in = Flipped(new StallReasonIO(RenameWidth))
77      val out = new StallReasonIO(RenameWidth)
78    }
79  })
80
81  val compressUnit = Module(new CompressUnit())
82  // create free list and rat
83  val intFreeList = Module(new MEFreeList(IntPhyRegs))
84  val fpFreeList = Module(new StdFreeList(VfPhyRegs - FpLogicRegs - VecLogicRegs))
85
86  intFreeList.io.commit    <> io.rabCommits
87  intFreeList.io.debug_rat.foreach(_ <> io.debug_int_rat.get)
88  fpFreeList.io.commit     <> io.rabCommits
89  fpFreeList.io.debug_rat.foreach(_ <> io.debug_fp_rat.get)
90
91  // decide if given instruction needs allocating a new physical register (CfCtrl: from decode; RobCommitInfo: from rob)
92  // fp and vec share `fpFreeList`
93  def needDestReg[T <: DecodedInst](reg_t: RegType, x: T): Bool = reg_t match {
94    case Reg_I => x.rfWen && x.ldest =/= 0.U
95    case Reg_F => x.fpWen
96    case Reg_V => x.vecWen
97  }
98  def needDestRegCommit[T <: RabCommitInfo](reg_t: RegType, x: T): Bool = {
99    reg_t match {
100      case Reg_I => x.rfWen
101      case Reg_F => x.fpWen
102      case Reg_V => x.vecWen
103    }
104  }
105  def needDestRegWalk[T <: RabCommitInfo](reg_t: RegType, x: T): Bool = {
106    reg_t match {
107      case Reg_I => x.rfWen && x.ldest =/= 0.U
108      case Reg_F => x.fpWen
109      case Reg_V => x.vecWen
110    }
111  }
112
113  // connect [redirect + walk] ports for __float point__ & __integer__ free list
114  Seq(fpFreeList, intFreeList).foreach { case fl =>
115    fl.io.redirect := io.redirect.valid
116    fl.io.walk := io.rabCommits.isWalk
117  }
118  // only when both fp and int free list and dispatch1 has enough space can we do allocation
119  // when isWalk, freelist can definitely allocate
120  intFreeList.io.doAllocate := fpFreeList.io.canAllocate && io.out(0).ready || io.rabCommits.isWalk
121  fpFreeList.io.doAllocate := intFreeList.io.canAllocate && io.out(0).ready || io.rabCommits.isWalk
122
123  //           dispatch1 ready ++ float point free list ready ++ int free list ready      ++ not walk
124  val canOut = io.out(0).ready && fpFreeList.io.canAllocate && intFreeList.io.canAllocate && !io.rabCommits.isWalk
125
126  compressUnit.io.in.zip(io.in).foreach{ case(sink, source) =>
127    sink.valid := source.valid
128    sink.bits := source.bits
129  }
130  val needRobFlags = compressUnit.io.out.needRobFlags
131  val instrSizesVec = compressUnit.io.out.instrSizes
132  val compressMasksVec = compressUnit.io.out.masks
133
134  // speculatively assign the instruction with an robIdx
135  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)
136  val robIdxHead = RegInit(0.U.asTypeOf(new RobPtr))
137  val lastCycleMisprediction = GatedValidRegNext(io.redirect.valid && !io.redirect.bits.flushItself())
138  val robIdxHeadNext = Mux(io.redirect.valid, io.redirect.bits.robIdx, // redirect: move ptr to given rob index
139         Mux(lastCycleMisprediction, robIdxHead + 1.U, // mis-predict: not flush robIdx itself
140           Mux(canOut, robIdxHead + validCount, // instructions successfully entered next stage: increase robIdx
141                      /* default */  robIdxHead))) // no instructions passed by this cycle: stick to old value
142  robIdxHead := robIdxHeadNext
143
144  /**
145    * Rename: allocate free physical register and update rename table
146    */
147  val uops = Wire(Vec(RenameWidth, new DynInst))
148  uops.foreach( uop => {
149    uop.srcState      := DontCare
150    uop.debugInfo     := DontCare
151    uop.lqIdx         := DontCare
152    uop.sqIdx         := DontCare
153    uop.waitForRobIdx := DontCare
154    uop.singleStep    := DontCare
155    uop.snapshot      := DontCare
156    uop.srcLoadDependency := DontCare
157  })
158
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    (uops(i): Data).waiveAll :<= (io.in(i).bits: Data).waiveAll
186
187    // update cf according to ssit result
188    uops(i).storeSetHit := io.ssit(i).valid
189    uops(i).loadWaitStrict := io.ssit(i).strict && io.ssit(i).valid
190    uops(i).ssid := io.ssit(i).ssid
191
192    // update cf according to waittable result
193    uops(i).loadWaitBit := io.waittable(i)
194
195    uops(i).replayInst := false.B // set by IQ or MemQ
196    // alloc a new phy reg, fp and vec share the `fpFreeList`
197    needVecDest(i) := io.in(i).valid && needDestReg(Reg_V, io.in(i).bits)
198    needFpDest(i) := io.in(i).valid && needDestReg(Reg_F, io.in(i).bits)
199    needIntDest(i) := io.in(i).valid && needDestReg(Reg_I, io.in(i).bits)
200    if (i < RabCommitWidth) {
201      walkNeedIntDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_I, io.rabCommits.info(i))
202      walkNeedFpDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_F, io.rabCommits.info(i))
203      walkNeedVecDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_V, io.rabCommits.info(i))
204      walkIsMove(i) := io.rabCommits.info(i).isMove
205    }
206    fpFreeList.io.allocateReq(i) := needFpDest(i) || needVecDest(i)
207    fpFreeList.io.walkReq(i) := walkNeedFpDest(i) || walkNeedVecDest(i)
208    intFreeList.io.allocateReq(i) := needIntDest(i) && !isMove(i)
209    intFreeList.io.walkReq(i) := walkNeedIntDest(i) && !walkIsMove(i)
210
211    // no valid instruction from decode stage || all resources (dispatch1 + both free lists) ready
212    io.in(i).ready := !hasValid || canOut
213
214    uops(i).robIdx := robIdxHead + PopCount(io.in.zip(needRobFlags).take(i).map{ case(in, needRobFlag) => in.valid && in.bits.lastUop && needRobFlag})
215    uops(i).instrSize := instrSizesVec(i)
216    when(isMove(i)) {
217      uops(i).numUops := 0.U
218      uops(i).numWB := 0.U
219    }
220    if (i > 0) {
221      when(!needRobFlags(i - 1)) {
222        uops(i).firstUop := false.B
223        uops(i).ftqPtr := uops(i - 1).ftqPtr
224        uops(i).ftqOffset := uops(i - 1).ftqOffset
225        uops(i).numUops := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
226        uops(i).numWB := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
227      }
228    }
229    when(!needRobFlags(i)) {
230      uops(i).lastUop := false.B
231      uops(i).numUops := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
232      uops(i).numWB := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
233    }
234    uops(i).wfflags := (compressMasksVec(i) & Cat(io.in.map(_.bits.wfflags).reverse)).orR
235    uops(i).dirtyFs := (compressMasksVec(i) & Cat(io.in.map(_.bits.fpWen).reverse)).orR
236    // vector instructions' uopSplitType cannot be UopSplitType.SCA_SIM
237    uops(i).dirtyVs := (compressMasksVec(i) & Cat(io.in.map(_.bits.uopSplitType =/= UopSplitType.SCA_SIM).reverse)).orR
238
239    uops(i).psrc(0) := Mux1H(uops(i).srcType(0), Seq(io.intReadPorts(i)(0), io.fpReadPorts(i)(0), io.vecReadPorts(i)(0)))
240    uops(i).psrc(1) := Mux1H(uops(i).srcType(1), Seq(io.intReadPorts(i)(1), io.fpReadPorts(i)(1), io.vecReadPorts(i)(1)))
241    uops(i).psrc(2) := Mux1H(uops(i).srcType(2)(2, 1), Seq(io.fpReadPorts(i)(2), io.vecReadPorts(i)(2)))
242    uops(i).psrc(3) := io.vecReadPorts(i)(3)
243    uops(i).psrc(4) := io.vecReadPorts(i)(4) // Todo: vl read port
244
245    // int psrc2 should be bypassed from next instruction if it is fused
246    if (i < RenameWidth - 1) {
247      when (io.fusionInfo(i).rs2FromRs2 || io.fusionInfo(i).rs2FromRs1) {
248        uops(i).psrc(1) := Mux(io.fusionInfo(i).rs2FromRs2, io.intReadPorts(i + 1)(1), io.intReadPorts(i + 1)(0))
249      }.elsewhen(io.fusionInfo(i).rs2FromZero) {
250        uops(i).psrc(1) := 0.U
251      }
252    }
253    uops(i).eliminatedMove := isMove(i)
254
255    // update pdest
256    uops(i).pdest := MuxCase(0.U, Seq(
257      needIntDest(i)                    -> intFreeList.io.allocatePhyReg(i),
258      (needFpDest(i) || needVecDest(i)) -> fpFreeList.io.allocatePhyReg(i),
259    ))
260
261    // Assign performance counters
262    uops(i).debugInfo.renameTime := GTimer()
263
264    io.out(i).valid := io.in(i).valid && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && !io.rabCommits.isWalk
265    io.out(i).bits := uops(i)
266    // Todo: move these shit in decode stage
267    // dirty code for fence. The lsrc is passed by imm.
268    when (io.out(i).bits.fuType === FuType.fence.U) {
269      io.out(i).bits.imm := Cat(io.in(i).bits.lsrc(1), io.in(i).bits.lsrc(0))
270    }
271
272    // dirty code for SoftPrefetch (prefetch.r/prefetch.w)
273//    when (io.in(i).bits.isSoftPrefetch) {
274//      io.out(i).bits.fuType := FuType.ldu.U
275//      io.out(i).bits.fuOpType := Mux(io.in(i).bits.lsrc(1) === 1.U, LSUOpType.prefetch_r, LSUOpType.prefetch_w)
276//      io.out(i).bits.selImm := SelImm.IMM_S
277//      io.out(i).bits.imm := Cat(io.in(i).bits.imm(io.in(i).bits.imm.getWidth - 1, 5), 0.U(5.W))
278//    }
279
280    // dirty code for lui+addi(w) fusion
281    if (i < RenameWidth - 1) {
282      val fused_lui32 = io.in(i).bits.selImm === SelImm.IMM_LUI32 && io.in(i).bits.fuType === FuType.alu.U
283      when (fused_lui32) {
284        val lui_imm = io.in(i).bits.imm(19, 0)
285        val add_imm = io.in(i + 1).bits.imm(11, 0)
286        require(io.out(i).bits.imm.getWidth >= lui_imm.getWidth + add_imm.getWidth)
287        io.out(i).bits.imm := Cat(lui_imm, add_imm)
288      }
289    }
290
291    // write speculative rename table
292    // we update rat later inside commit code
293    intSpecWen(i) := needIntDest(i) && intFreeList.io.canAllocate && intFreeList.io.doAllocate && !io.rabCommits.isWalk && !io.redirect.valid
294    fpSpecWen(i) := needFpDest(i) && fpFreeList.io.canAllocate && fpFreeList.io.doAllocate && !io.rabCommits.isWalk && !io.redirect.valid
295    vecSpecWen(i) := needVecDest(i) && fpFreeList.io.canAllocate && fpFreeList.io.doAllocate && !io.rabCommits.isWalk && !io.redirect.valid
296
297
298    if (i < RabCommitWidth) {
299      walkIntSpecWen(i) := walkNeedIntDest(i) && !io.redirect.valid
300      walkPdest(i) := io.rabCommits.info(i).pdest
301    } else {
302      walkPdest(i) := io.out(i).bits.pdest
303    }
304  }
305
306  /**
307    * How to set psrc:
308    * - bypass the pdest to psrc if previous instructions write to the same ldest as lsrc
309    * - default: psrc from RAT
310    * How to set pdest:
311    * - Mux(isMove, psrc, pdest_from_freelist).
312    *
313    * The critical path of rename lies here:
314    * When move elimination is enabled, we need to update the rat with psrc.
315    * However, psrc maybe comes from previous instructions' pdest, which comes from freelist.
316    *
317    * If we expand these logic for pdest(N):
318    * pdest(N) = Mux(isMove(N), psrc(N), freelist_out(N))
319    *          = Mux(isMove(N), Mux(bypass(N, N - 1), pdest(N - 1),
320    *                           Mux(bypass(N, N - 2), pdest(N - 2),
321    *                           ...
322    *                           Mux(bypass(N, 0),     pdest(0),
323    *                                                 rat_out(N))...)),
324    *                           freelist_out(N))
325    */
326  // a simple functional model for now
327  io.out(0).bits.pdest := Mux(isMove(0), uops(0).psrc.head, uops(0).pdest)
328
329  // psrc(n) + pdest(1)
330  val bypassCond: Vec[MixedVec[UInt]] = Wire(Vec(numRegSrc + 1, MixedVec(List.tabulate(RenameWidth-1)(i => UInt((i+1).W)))))
331  require(io.in(0).bits.srcType.size == io.in(0).bits.numSrc)
332  private val pdestLoc = io.in.head.bits.srcType.size // 2 vector src: v0, vl&vtype
333  println(s"[Rename] idx of pdest in bypassCond $pdestLoc")
334  for (i <- 1 until RenameWidth) {
335    val vecCond = io.in(i).bits.srcType.map(_ === SrcType.vp) :+ needVecDest(i)
336    val fpCond  = io.in(i).bits.srcType.map(_ === SrcType.fp) :+ needFpDest(i)
337    val intCond = io.in(i).bits.srcType.map(_ === SrcType.xp) :+ needIntDest(i)
338    val target = io.in(i).bits.lsrc :+ io.in(i).bits.ldest
339    for (((((cond1, cond2), cond3), t), j) <- vecCond.zip(fpCond).zip(intCond).zip(target).zipWithIndex) {
340      val destToSrc = io.in.take(i).zipWithIndex.map { case (in, j) =>
341        val indexMatch = in.bits.ldest === t
342        val writeMatch =  cond3 && needIntDest(j) || cond2 && needFpDest(j) || cond1 && needVecDest(j)
343        indexMatch && writeMatch
344      }
345      bypassCond(j)(i - 1) := VecInit(destToSrc).asUInt
346    }
347    io.out(i).bits.psrc(0) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(0)(i-1).asBools).foldLeft(uops(i).psrc(0)) {
348      (z, next) => Mux(next._2, next._1, z)
349    }
350    io.out(i).bits.psrc(1) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(1)(i-1).asBools).foldLeft(uops(i).psrc(1)) {
351      (z, next) => Mux(next._2, next._1, z)
352    }
353    io.out(i).bits.psrc(2) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(2)(i-1).asBools).foldLeft(uops(i).psrc(2)) {
354      (z, next) => Mux(next._2, next._1, z)
355    }
356    io.out(i).bits.psrc(3) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(3)(i-1).asBools).foldLeft(uops(i).psrc(3)) {
357      (z, next) => Mux(next._2, next._1, z)
358    }
359    io.out(i).bits.psrc(4) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(4)(i-1).asBools).foldLeft(uops(i).psrc(4)) {
360      (z, next) => Mux(next._2, next._1, z)
361    }
362    io.out(i).bits.pdest := Mux(isMove(i), io.out(i).bits.psrc(0), uops(i).pdest)
363
364    // Todo: better implementation for fields reuse
365    // For fused-lui-load, load.src(0) is replaced by the imm.
366    val last_is_lui = io.in(i - 1).bits.selImm === SelImm.IMM_U && io.in(i - 1).bits.srcType(0) =/= SrcType.pc
367    val this_is_load = io.in(i).bits.fuType === FuType.ldu.U
368    val lui_to_load = io.in(i - 1).valid && io.in(i - 1).bits.ldest === io.in(i).bits.lsrc(0)
369    val fused_lui_load = last_is_lui && this_is_load && lui_to_load
370    when (fused_lui_load) {
371      // The first LOAD operand (base address) is replaced by LUI-imm and stored in imm
372      val lui_imm = io.in(i - 1).bits.imm(ImmUnion.U.len - 1, 0)
373      val ld_imm = io.in(i).bits.imm(ImmUnion.I.len - 1, 0)
374      require(io.out(i).bits.imm.getWidth >= lui_imm.getWidth + ld_imm.getWidth)
375      io.out(i).bits.srcType(0) := SrcType.imm
376      io.out(i).bits.imm := Cat(lui_imm, ld_imm)
377    }
378
379  }
380
381  val genSnapshot = Cat(io.out.map(out => out.fire && out.bits.snapshot)).orR
382  val lastCycleCreateSnpt = RegInit(false.B)
383  lastCycleCreateSnpt := genSnapshot && !io.snptIsFull
384  val sameSnptDistance = (RobCommitWidth * 4).U
385  // notInSameSnpt: 1.robidxHead - snapLastEnq >= sameSnptDistance 2.no snap
386  val notInSameSnpt = GatedValidRegNext(distanceBetween(robIdxHeadNext, io.snptLastEnq.bits) >= sameSnptDistance || !io.snptLastEnq.valid)
387  val allowSnpt = if (EnableRenameSnapshot) notInSameSnpt && !lastCycleCreateSnpt && io.in.head.bits.firstUop else false.B
388  io.out.zip(io.in).foreach{ case (out, in) => out.bits.snapshot := allowSnpt && (!in.bits.preDecodeInfo.notCFI || FuType.isJump(in.bits.fuType)) && in.fire }
389  if(backendParams.debugEn){
390    dontTouch(robIdxHeadNext)
391    dontTouch(notInSameSnpt)
392    dontTouch(genSnapshot)
393  }
394  intFreeList.io.snpt := io.snpt
395  fpFreeList.io.snpt := io.snpt
396  intFreeList.io.snpt.snptEnq := genSnapshot
397  fpFreeList.io.snpt.snptEnq := genSnapshot
398
399  /**
400    * Instructions commit: update freelist and rename table
401    */
402  for (i <- 0 until RabCommitWidth) {
403    val commitValid = io.rabCommits.isCommit && io.rabCommits.commitValid(i)
404    val walkValid = io.rabCommits.isWalk && io.rabCommits.walkValid(i)
405
406    // I. RAT Update
407    // When redirect happens (mis-prediction), don't update the rename table
408    io.intRenamePorts(i).wen  := intSpecWen(i)
409    io.intRenamePorts(i).addr := uops(i).ldest
410    io.intRenamePorts(i).data := io.out(i).bits.pdest
411
412    io.fpRenamePorts(i).wen  := fpSpecWen(i)
413    io.fpRenamePorts(i).addr := uops(i).ldest
414    io.fpRenamePorts(i).data := fpFreeList.io.allocatePhyReg(i)
415
416    io.vecRenamePorts(i).wen := vecSpecWen(i)
417    io.vecRenamePorts(i).addr := uops(i).ldest
418    io.vecRenamePorts(i).data := fpFreeList.io.allocatePhyReg(i)
419
420    // II. Free List Update
421    intFreeList.io.freeReq(i) := io.int_need_free(i)
422    intFreeList.io.freePhyReg(i) := RegNext(io.int_old_pdest(i))
423    fpFreeList.io.freeReq(i)  := GatedValidRegNext(commitValid && (needDestRegCommit(Reg_F, io.rabCommits.info(i)) || needDestRegCommit(Reg_V, io.rabCommits.info(i))))
424    fpFreeList.io.freePhyReg(i) := Mux(RegNext(needDestRegCommit(Reg_F, io.rabCommits.info(i))), io.fp_old_pdest(i), io.vec_old_pdest(i))
425  }
426
427  /*
428  Debug and performance counters
429   */
430  def printRenameInfo(in: DecoupledIO[DecodedInst], out: DecoupledIO[DynInst]) = {
431    XSInfo(out.fire, p"pc:${Hexadecimal(in.bits.pc)} in(${in.valid},${in.ready}) " +
432      p"lsrc(0):${in.bits.lsrc(0)} -> psrc(0):${out.bits.psrc(0)} " +
433      p"lsrc(1):${in.bits.lsrc(1)} -> psrc(1):${out.bits.psrc(1)} " +
434      p"lsrc(2):${in.bits.lsrc(2)} -> psrc(2):${out.bits.psrc(2)} " +
435      p"ldest:${in.bits.ldest} -> pdest:${out.bits.pdest}\n"
436    )
437  }
438
439  for ((x,y) <- io.in.zip(io.out)) {
440    printRenameInfo(x, y)
441  }
442
443  val debugRedirect = RegEnable(io.redirect.bits, io.redirect.valid)
444  // bad speculation
445  val recStall = io.redirect.valid || io.rabCommits.isWalk
446  val ctrlRecStall = Mux(io.redirect.valid, io.redirect.bits.debugIsCtrl, io.rabCommits.isWalk && debugRedirect.debugIsCtrl)
447  val mvioRecStall = Mux(io.redirect.valid, io.redirect.bits.debugIsMemVio, io.rabCommits.isWalk && debugRedirect.debugIsMemVio)
448  val otherRecStall = recStall && !(ctrlRecStall || mvioRecStall)
449  XSPerfAccumulate("recovery_stall", recStall)
450  XSPerfAccumulate("control_recovery_stall", ctrlRecStall)
451  XSPerfAccumulate("mem_violation_recovery_stall", mvioRecStall)
452  XSPerfAccumulate("other_recovery_stall", otherRecStall)
453  // freelist stall
454  val notRecStall = !io.out.head.valid && !recStall
455  val intFlStall = notRecStall && inHeadValid && !intFreeList.io.canAllocate
456  val fpFlStall = notRecStall && inHeadValid && intFreeList.io.canAllocate && !fpFreeList.io.canAllocate
457  // other stall
458  val otherStall = notRecStall && !intFlStall && !fpFlStall
459
460  io.stallReason.in.backReason.valid := io.stallReason.out.backReason.valid || !io.in.head.ready
461  io.stallReason.in.backReason.bits := Mux(io.stallReason.out.backReason.valid, io.stallReason.out.backReason.bits,
462    MuxCase(TopDownCounters.OtherCoreStall.id.U, Seq(
463      ctrlRecStall  -> TopDownCounters.ControlRecoveryStall.id.U,
464      mvioRecStall  -> TopDownCounters.MemVioRecoveryStall.id.U,
465      otherRecStall -> TopDownCounters.OtherRecoveryStall.id.U,
466      intFlStall    -> TopDownCounters.IntFlStall.id.U,
467      fpFlStall     -> TopDownCounters.FpFlStall.id.U
468    )
469  ))
470  io.stallReason.out.reason.zip(io.stallReason.in.reason).zip(io.in.map(_.valid)).foreach { case ((out, in), valid) =>
471    out := Mux(io.stallReason.in.backReason.valid, io.stallReason.in.backReason.bits, in)
472  }
473
474  XSDebug(io.rabCommits.isWalk, p"Walk Recovery Enabled\n")
475  XSDebug(io.rabCommits.isWalk, p"validVec:${Binary(io.rabCommits.walkValid.asUInt)}\n")
476  for (i <- 0 until RabCommitWidth) {
477    val info = io.rabCommits.info(i)
478    XSDebug(io.rabCommits.isWalk && io.rabCommits.walkValid(i), p"[#$i walk info] " +
479      p"ldest:${info.ldest} rfWen:${info.rfWen} fpWen:${info.fpWen} vecWen:${info.vecWen}")
480  }
481
482  XSDebug(p"inValidVec: ${Binary(Cat(io.in.map(_.valid)))}\n")
483
484  XSPerfAccumulate("in_valid_count", PopCount(io.in.map(_.valid)))
485  XSPerfAccumulate("in_fire_count", PopCount(io.in.map(_.fire)))
486  XSPerfAccumulate("in_valid_not_ready_count", PopCount(io.in.map(x => x.valid && !x.ready)))
487  XSPerfAccumulate("wait_cycle", !io.in.head.valid && io.out.head.ready)
488
489  // These stall reasons could overlap each other, but we configure the priority as fellows.
490  // walk stall > dispatch stall > int freelist stall > fp freelist stall
491  private val inHeadStall = io.in.head match { case x => x.valid && !x.ready }
492  private val stallForWalk      = inHeadValid &&  io.rabCommits.isWalk
493  private val stallForDispatch  = inHeadValid && !io.rabCommits.isWalk && !io.out(0).ready
494  private val stallForIntFL     = inHeadValid && !io.rabCommits.isWalk &&  io.out(0).ready && !intFreeList.io.canAllocate
495  private val stallForFpFL      = inHeadValid && !io.rabCommits.isWalk &&  io.out(0).ready &&  intFreeList.io.canAllocate && !fpFreeList.io.canAllocate
496  XSPerfAccumulate("stall_cycle",          inHeadStall)
497  XSPerfAccumulate("stall_cycle_walk",     stallForWalk)
498  XSPerfAccumulate("stall_cycle_dispatch", stallForDispatch)
499  XSPerfAccumulate("stall_cycle_int",      stallForIntFL)
500  XSPerfAccumulate("stall_cycle_fp",       stallForFpFL)
501
502  XSPerfHistogram("in_valid_range",  PopCount(io.in.map(_.valid)),  true.B, 0, DecodeWidth + 1, 1)
503  XSPerfHistogram("in_fire_range",   PopCount(io.in.map(_.fire)),   true.B, 0, DecodeWidth + 1, 1)
504  XSPerfHistogram("out_valid_range", PopCount(io.out.map(_.valid)), true.B, 0, DecodeWidth + 1, 1)
505  XSPerfHistogram("out_fire_range",  PopCount(io.out.map(_.fire)),  true.B, 0, DecodeWidth + 1, 1)
506
507  XSPerfAccumulate("move_instr_count", PopCount(io.out.map(out => out.fire && out.bits.isMove)))
508  val is_fused_lui_load = io.out.map(o => o.fire && o.bits.fuType === FuType.ldu.U && o.bits.srcType(0) === SrcType.imm)
509  XSPerfAccumulate("fused_lui_load_instr_count", PopCount(is_fused_lui_load))
510
511  val renamePerf = Seq(
512    ("rename_in                  ", PopCount(io.in.map(_.valid & io.in(0).ready ))                                                               ),
513    ("rename_waitinstr           ", PopCount((0 until RenameWidth).map(i => io.in(i).valid && !io.in(i).ready))                                  ),
514    ("rename_stall               ", inHeadStall),
515    ("rename_stall_cycle_walk    ", inHeadValid &&  io.rabCommits.isWalk),
516    ("rename_stall_cycle_dispatch", inHeadValid && !io.rabCommits.isWalk && !io.out(0).ready),
517    ("rename_stall_cycle_int     ", inHeadValid && !io.rabCommits.isWalk &&  io.out(0).ready && !intFreeList.io.canAllocate),
518    ("rename_stall_cycle_fp      ", inHeadValid && !io.rabCommits.isWalk &&  io.out(0).ready && intFreeList.io.canAllocate && !fpFreeList.io.canAllocate),
519  )
520  val intFlPerf = intFreeList.getPerfEvents
521  val fpFlPerf = fpFreeList.getPerfEvents
522  val perfEvents = renamePerf ++ intFlPerf ++ fpFlPerf
523  generatePerfEvent()
524}
525