xref: /XiangShan/src/main/scala/xiangshan/backend/rename/Rename.scala (revision 00f9d184d720c2934146e1d3661fc6dc93cbdf1e)
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._
31import xiangshan.ExceptionNO._
32
33class Rename(implicit p: Parameters) extends XSModule with HasCircularQueuePtrHelper with HasPerfEvents {
34
35  // params alias
36  private val numRegSrc = backendParams.numRegSrc
37  private val numVecRegSrc = backendParams.numVecRegSrc
38  private val numVecRatPorts = numVecRegSrc
39
40  println(s"[Rename] numRegSrc: $numRegSrc")
41
42  val io = IO(new Bundle() {
43    val redirect = Flipped(ValidIO(new Redirect))
44    val rabCommits = Input(new RabCommitIO)
45    // from decode
46    val in = Vec(RenameWidth, Flipped(DecoupledIO(new DecodedInst)))
47    val fusionInfo = Vec(DecodeWidth - 1, Flipped(new FusionDecodeInfo))
48    // ssit read result
49    val ssit = Flipped(Vec(RenameWidth, Output(new SSITEntry)))
50    // waittable read result
51    val waittable = Flipped(Vec(RenameWidth, Output(Bool())))
52    // to rename table
53    val intReadPorts = Vec(RenameWidth, Vec(2, Input(UInt(PhyRegIdxWidth.W))))
54    val fpReadPorts = Vec(RenameWidth, Vec(3, Input(UInt(PhyRegIdxWidth.W))))
55    val vecReadPorts = Vec(RenameWidth, Vec(numVecRatPorts, Input(UInt(PhyRegIdxWidth.W))))
56    val intRenamePorts = Vec(RenameWidth, Output(new RatWritePort))
57    val fpRenamePorts = Vec(RenameWidth, Output(new RatWritePort))
58    val vecRenamePorts = Vec(RenameWidth, Output(new RatWritePort))
59    // from rename table
60    val int_old_pdest = Vec(RabCommitWidth, Input(UInt(PhyRegIdxWidth.W)))
61    val fp_old_pdest = Vec(RabCommitWidth, Input(UInt(PhyRegIdxWidth.W)))
62    val vec_old_pdest = Vec(RabCommitWidth, Input(UInt(PhyRegIdxWidth.W)))
63    val int_need_free = Vec(RabCommitWidth, Input(Bool()))
64    // to dispatch1
65    val out = Vec(RenameWidth, DecoupledIO(new DynInst))
66    // for snapshots
67    val snpt = Input(new SnapshotPort)
68    val snptLastEnq = Flipped(ValidIO(new RobPtr))
69    val snptIsFull= Input(Bool())
70    // debug arch ports
71    val debug_int_rat = if (backendParams.debugEn) Some(Vec(32, Input(UInt(PhyRegIdxWidth.W)))) else None
72    val debug_vconfig_rat = if (backendParams.debugEn) Some(Input(UInt(PhyRegIdxWidth.W))) else None
73    val debug_fp_rat = if (backendParams.debugEn) Some(Vec(32, Input(UInt(PhyRegIdxWidth.W)))) else None
74    val debug_vec_rat = if (backendParams.debugEn) Some(Vec(32, Input(UInt(PhyRegIdxWidth.W)))) else None
75    // perf only
76    val stallReason = new Bundle {
77      val in = Flipped(new StallReasonIO(RenameWidth))
78      val out = new StallReasonIO(RenameWidth)
79    }
80  })
81
82  // io alias
83  private val dispatchCanAcc = io.out.head.ready
84
85  val compressUnit = Module(new CompressUnit())
86  // create free list and rat
87  val intFreeList = Module(new MEFreeList(IntPhyRegs))
88  val fpFreeList = Module(new StdFreeList(FpPhyRegs - FpLogicRegs, FpLogicRegs, Reg_F))
89  val vecFreeList = Module(new StdFreeList(VfPhyRegs - VecLogicRegs, VecLogicRegs, Reg_V))
90
91  intFreeList.io.commit    <> io.rabCommits
92  intFreeList.io.debug_rat.foreach(_ <> io.debug_int_rat.get)
93  fpFreeList.io.commit     <> io.rabCommits
94  fpFreeList.io.debug_rat.foreach(_ <> io.debug_fp_rat.get)
95  vecFreeList.io.commit    <> io.rabCommits
96  vecFreeList.io.debug_rat.foreach(_ <> io.debug_vec_rat.get)
97
98  // decide if given instruction needs allocating a new physical register (CfCtrl: from decode; RobCommitInfo: from rob)
99  def needDestReg[T <: DecodedInst](reg_t: RegType, x: T): Bool = reg_t match {
100    case Reg_I => x.rfWen && x.ldest =/= 0.U
101    case Reg_F => x.fpWen
102    case Reg_V => x.vecWen
103  }
104  def needDestRegCommit[T <: RabCommitInfo](reg_t: RegType, x: T): Bool = {
105    reg_t match {
106      case Reg_I => x.rfWen
107      case Reg_F => x.fpWen
108      case Reg_V => x.vecWen
109    }
110  }
111  def needDestRegWalk[T <: RabCommitInfo](reg_t: RegType, x: T): Bool = {
112    reg_t match {
113      case Reg_I => x.rfWen && x.ldest =/= 0.U
114      case Reg_F => x.fpWen
115      case Reg_V => x.vecWen
116    }
117  }
118
119  // connect [redirect + walk] ports for fp & vec & int free list
120  Seq(fpFreeList, vecFreeList, intFreeList).foreach { case fl =>
121    fl.io.redirect := io.redirect.valid
122    fl.io.walk := io.rabCommits.isWalk
123  }
124  // only when all free list and dispatch1 has enough space can we do allocation
125  // when isWalk, freelist can definitely allocate
126  intFreeList.io.doAllocate := fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && dispatchCanAcc || io.rabCommits.isWalk
127  fpFreeList.io.doAllocate := intFreeList.io.canAllocate && vecFreeList.io.canAllocate && dispatchCanAcc || io.rabCommits.isWalk
128  vecFreeList.io.doAllocate := intFreeList.io.canAllocate && fpFreeList.io.canAllocate && dispatchCanAcc || io.rabCommits.isWalk
129
130  //           dispatch1 ready ++ float point free list ready ++ int free list ready ++ vec free list ready     ++ not walk
131  val canOut = dispatchCanAcc && fpFreeList.io.canAllocate && intFreeList.io.canAllocate && vecFreeList.io.canAllocate && !io.rabCommits.isWalk
132
133  compressUnit.io.in.zip(io.in).foreach{ case(sink, source) =>
134    sink.valid := source.valid
135    sink.bits := source.bits
136  }
137  val needRobFlags = compressUnit.io.out.needRobFlags
138  val instrSizesVec = compressUnit.io.out.instrSizes
139  val compressMasksVec = compressUnit.io.out.masks
140
141  // speculatively assign the instruction with an robIdx
142  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)
143  val robIdxHead = RegInit(0.U.asTypeOf(new RobPtr))
144  val lastCycleMisprediction = GatedValidRegNext(io.redirect.valid && !io.redirect.bits.flushItself())
145  val robIdxHeadNext = Mux(io.redirect.valid, io.redirect.bits.robIdx, // redirect: move ptr to given rob index
146         Mux(lastCycleMisprediction, robIdxHead + 1.U, // mis-predict: not flush robIdx itself
147           Mux(canOut, robIdxHead + validCount, // instructions successfully entered next stage: increase robIdx
148                      /* default */  robIdxHead))) // no instructions passed by this cycle: stick to old value
149  robIdxHead := robIdxHeadNext
150
151  /**
152    * Rename: allocate free physical register and update rename table
153    */
154  val uops = Wire(Vec(RenameWidth, new DynInst))
155  uops.foreach( uop => {
156    uop.srcState      := DontCare
157    uop.debugInfo     := DontCare
158    uop.lqIdx         := DontCare
159    uop.sqIdx         := DontCare
160    uop.waitForRobIdx := DontCare
161    uop.singleStep    := DontCare
162    uop.snapshot      := DontCare
163    uop.srcLoadDependency := DontCare
164    uop.numLsElem       :=  DontCare
165    uop.hasException  :=  DontCare
166  })
167
168  val needVecDest    = Wire(Vec(RenameWidth, Bool()))
169  val needFpDest     = Wire(Vec(RenameWidth, Bool()))
170  val needIntDest    = Wire(Vec(RenameWidth, Bool()))
171  val hasValid = Cat(io.in.map(_.valid)).orR
172  private val inHeadValid = io.in.head.valid
173
174  val isMove = Wire(Vec(RenameWidth, Bool()))
175  isMove zip io.in.map(_.bits) foreach {
176    case (move, in) => move := Mux(in.exceptionVec.asUInt.orR, false.B, in.isMove)
177  }
178
179  val walkNeedIntDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
180  val walkNeedFpDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
181  val walkNeedVecDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
182  val walkIsMove = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
183
184  val intSpecWen = Wire(Vec(RenameWidth, Bool()))
185  val fpSpecWen  = Wire(Vec(RenameWidth, Bool()))
186  val vecSpecWen = Wire(Vec(RenameWidth, Bool()))
187
188  val walkIntSpecWen = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
189
190  val walkPdest = Wire(Vec(RenameWidth, UInt(PhyRegIdxWidth.W)))
191
192  // uop calculation
193  for (i <- 0 until RenameWidth) {
194    (uops(i): Data).waiveAll :<= (io.in(i).bits: Data).waiveAll
195
196    // update cf according to ssit result
197    uops(i).storeSetHit := io.ssit(i).valid
198    uops(i).loadWaitStrict := io.ssit(i).strict && io.ssit(i).valid
199    uops(i).ssid := io.ssit(i).ssid
200
201    // update cf according to waittable result
202    uops(i).loadWaitBit := io.waittable(i)
203
204    uops(i).replayInst := false.B // set by IQ or MemQ
205    // alloc a new phy reg
206    needVecDest(i) := io.in(i).valid && needDestReg(Reg_V, io.in(i).bits)
207    needFpDest(i) := io.in(i).valid && needDestReg(Reg_F, io.in(i).bits)
208    needIntDest(i) := io.in(i).valid && needDestReg(Reg_I, io.in(i).bits)
209    if (i < RabCommitWidth) {
210      walkNeedIntDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_I, io.rabCommits.info(i))
211      walkNeedFpDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_F, io.rabCommits.info(i))
212      walkNeedVecDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_V, io.rabCommits.info(i))
213      walkIsMove(i) := io.rabCommits.info(i).isMove
214    }
215    fpFreeList.io.allocateReq(i) := needFpDest(i)
216    fpFreeList.io.walkReq(i) := walkNeedFpDest(i)
217    vecFreeList.io.allocateReq(i) := needVecDest(i)
218    vecFreeList.io.walkReq(i) := walkNeedVecDest(i)
219    intFreeList.io.allocateReq(i) := needIntDest(i) && !isMove(i)
220    intFreeList.io.walkReq(i) := walkNeedIntDest(i) && !walkIsMove(i)
221
222    // no valid instruction from decode stage || all resources (dispatch1 + both free lists) ready
223    io.in(i).ready := !hasValid || canOut
224
225    uops(i).robIdx := robIdxHead + PopCount(io.in.zip(needRobFlags).take(i).map{ case(in, needRobFlag) => in.valid && in.bits.lastUop && needRobFlag})
226    uops(i).instrSize := instrSizesVec(i)
227    when(isMove(i)) {
228      uops(i).numUops := 0.U
229      uops(i).numWB := 0.U
230    }
231    if (i > 0) {
232      when(!needRobFlags(i - 1)) {
233        uops(i).firstUop := false.B
234        uops(i).ftqPtr := uops(i - 1).ftqPtr
235        uops(i).ftqOffset := uops(i - 1).ftqOffset
236        uops(i).numUops := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
237        uops(i).numWB := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
238      }
239    }
240    when(!needRobFlags(i)) {
241      uops(i).lastUop := false.B
242      uops(i).numUops := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
243      uops(i).numWB := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
244    }
245    uops(i).wfflags := (compressMasksVec(i) & Cat(io.in.map(_.bits.wfflags).reverse)).orR
246    uops(i).dirtyFs := (compressMasksVec(i) & Cat(io.in.map(_.bits.fpWen).reverse)).orR
247    // vector instructions' uopSplitType cannot be UopSplitType.SCA_SIM
248    uops(i).dirtyVs := (compressMasksVec(i) & Cat(io.in.map(_.bits.uopSplitType =/= UopSplitType.SCA_SIM).reverse)).orR
249
250    uops(i).psrc(0) := Mux1H(uops(i).srcType(0), Seq(io.intReadPorts(i)(0), io.fpReadPorts(i)(0), io.vecReadPorts(i)(0)))
251    uops(i).psrc(1) := Mux1H(uops(i).srcType(1), Seq(io.intReadPorts(i)(1), io.fpReadPorts(i)(1), io.vecReadPorts(i)(1)))
252    uops(i).psrc(2) := Mux1H(uops(i).srcType(2)(2, 1), Seq(io.fpReadPorts(i)(2), io.vecReadPorts(i)(2)))
253    uops(i).psrc(3) := io.vecReadPorts(i)(3)
254    uops(i).psrc(4) := io.vecReadPorts(i)(4) // Todo: vl read port
255
256    // int psrc2 should be bypassed from next instruction if it is fused
257    if (i < RenameWidth - 1) {
258      when (io.fusionInfo(i).rs2FromRs2 || io.fusionInfo(i).rs2FromRs1) {
259        uops(i).psrc(1) := Mux(io.fusionInfo(i).rs2FromRs2, io.intReadPorts(i + 1)(1), io.intReadPorts(i + 1)(0))
260      }.elsewhen(io.fusionInfo(i).rs2FromZero) {
261        uops(i).psrc(1) := 0.U
262      }
263    }
264    uops(i).eliminatedMove := isMove(i)
265
266    // update pdest
267    uops(i).pdest := MuxCase(0.U, Seq(
268      needIntDest(i)    ->  intFreeList.io.allocatePhyReg(i),
269      needFpDest(i)     ->  fpFreeList.io.allocatePhyReg(i),
270      needVecDest(i)    ->  vecFreeList.io.allocatePhyReg(i),
271    ))
272
273    // Assign performance counters
274    uops(i).debugInfo.renameTime := GTimer()
275
276    io.out(i).valid := io.in(i).valid && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && !io.rabCommits.isWalk
277    io.out(i).bits := uops(i)
278    // Todo: move these shit in decode stage
279    // dirty code for fence. The lsrc is passed by imm.
280    when (io.out(i).bits.fuType === FuType.fence.U) {
281      io.out(i).bits.imm := Cat(io.in(i).bits.lsrc(1), io.in(i).bits.lsrc(0))
282    }
283
284    // dirty code for SoftPrefetch (prefetch.r/prefetch.w)
285//    when (io.in(i).bits.isSoftPrefetch) {
286//      io.out(i).bits.fuType := FuType.ldu.U
287//      io.out(i).bits.fuOpType := Mux(io.in(i).bits.lsrc(1) === 1.U, LSUOpType.prefetch_r, LSUOpType.prefetch_w)
288//      io.out(i).bits.selImm := SelImm.IMM_S
289//      io.out(i).bits.imm := Cat(io.in(i).bits.imm(io.in(i).bits.imm.getWidth - 1, 5), 0.U(5.W))
290//    }
291
292    // dirty code for lui+addi(w) fusion
293    if (i < RenameWidth - 1) {
294      val fused_lui32 = io.in(i).bits.selImm === SelImm.IMM_LUI32 && io.in(i).bits.fuType === FuType.alu.U
295      when (fused_lui32) {
296        val lui_imm = io.in(i).bits.imm(19, 0)
297        val add_imm = io.in(i + 1).bits.imm(11, 0)
298        require(io.out(i).bits.imm.getWidth >= lui_imm.getWidth + add_imm.getWidth)
299        io.out(i).bits.imm := Cat(lui_imm, add_imm)
300      }
301    }
302
303    // write speculative rename table
304    // we update rat later inside commit code
305    intSpecWen(i) := needIntDest(i) && intFreeList.io.canAllocate && intFreeList.io.doAllocate && !io.rabCommits.isWalk && !io.redirect.valid
306    fpSpecWen(i)  := needFpDest(i)  && fpFreeList.io.canAllocate  && fpFreeList.io.doAllocate  && !io.rabCommits.isWalk && !io.redirect.valid
307    vecSpecWen(i) := needVecDest(i) && vecFreeList.io.canAllocate && vecFreeList.io.doAllocate && !io.rabCommits.isWalk && !io.redirect.valid
308
309
310    if (i < RabCommitWidth) {
311      walkIntSpecWen(i) := walkNeedIntDest(i) && !io.redirect.valid
312      walkPdest(i) := io.rabCommits.info(i).pdest
313    } else {
314      walkPdest(i) := io.out(i).bits.pdest
315    }
316  }
317
318  /**
319    * How to set psrc:
320    * - bypass the pdest to psrc if previous instructions write to the same ldest as lsrc
321    * - default: psrc from RAT
322    * How to set pdest:
323    * - Mux(isMove, psrc, pdest_from_freelist).
324    *
325    * The critical path of rename lies here:
326    * When move elimination is enabled, we need to update the rat with psrc.
327    * However, psrc maybe comes from previous instructions' pdest, which comes from freelist.
328    *
329    * If we expand these logic for pdest(N):
330    * pdest(N) = Mux(isMove(N), psrc(N), freelist_out(N))
331    *          = Mux(isMove(N), Mux(bypass(N, N - 1), pdest(N - 1),
332    *                           Mux(bypass(N, N - 2), pdest(N - 2),
333    *                           ...
334    *                           Mux(bypass(N, 0),     pdest(0),
335    *                                                 rat_out(N))...)),
336    *                           freelist_out(N))
337    */
338  // a simple functional model for now
339  io.out(0).bits.pdest := Mux(isMove(0), uops(0).psrc.head, uops(0).pdest)
340
341  // psrc(n) + pdest(1)
342  val bypassCond: Vec[MixedVec[UInt]] = Wire(Vec(numRegSrc + 1, MixedVec(List.tabulate(RenameWidth-1)(i => UInt((i+1).W)))))
343  require(io.in(0).bits.srcType.size == io.in(0).bits.numSrc)
344  private val pdestLoc = io.in.head.bits.srcType.size // 2 vector src: v0, vl&vtype
345  println(s"[Rename] idx of pdest in bypassCond $pdestLoc")
346  for (i <- 1 until RenameWidth) {
347    val vecCond = io.in(i).bits.srcType.map(_ === SrcType.vp) :+ needVecDest(i)
348    val fpCond  = io.in(i).bits.srcType.map(_ === SrcType.fp) :+ needFpDest(i)
349    val intCond = io.in(i).bits.srcType.map(_ === SrcType.xp) :+ needIntDest(i)
350    val target = io.in(i).bits.lsrc :+ io.in(i).bits.ldest
351    for (((((cond1, cond2), cond3), t), j) <- vecCond.zip(fpCond).zip(intCond).zip(target).zipWithIndex) {
352      val destToSrc = io.in.take(i).zipWithIndex.map { case (in, j) =>
353        val indexMatch = in.bits.ldest === t
354        val writeMatch =  cond3 && needIntDest(j) || cond2 && needFpDest(j) || cond1 && needVecDest(j)
355        indexMatch && writeMatch
356      }
357      bypassCond(j)(i - 1) := VecInit(destToSrc).asUInt
358    }
359    io.out(i).bits.psrc(0) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(0)(i-1).asBools).foldLeft(uops(i).psrc(0)) {
360      (z, next) => Mux(next._2, next._1, z)
361    }
362    io.out(i).bits.psrc(1) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(1)(i-1).asBools).foldLeft(uops(i).psrc(1)) {
363      (z, next) => Mux(next._2, next._1, z)
364    }
365    io.out(i).bits.psrc(2) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(2)(i-1).asBools).foldLeft(uops(i).psrc(2)) {
366      (z, next) => Mux(next._2, next._1, z)
367    }
368    io.out(i).bits.psrc(3) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(3)(i-1).asBools).foldLeft(uops(i).psrc(3)) {
369      (z, next) => Mux(next._2, next._1, z)
370    }
371    io.out(i).bits.psrc(4) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(4)(i-1).asBools).foldLeft(uops(i).psrc(4)) {
372      (z, next) => Mux(next._2, next._1, z)
373    }
374    io.out(i).bits.pdest := Mux(isMove(i), io.out(i).bits.psrc(0), uops(i).pdest)
375
376    // Todo: better implementation for fields reuse
377    // For fused-lui-load, load.src(0) is replaced by the imm.
378    val last_is_lui = io.in(i - 1).bits.selImm === SelImm.IMM_U && io.in(i - 1).bits.srcType(0) =/= SrcType.pc
379    val this_is_load = io.in(i).bits.fuType === FuType.ldu.U
380    val lui_to_load = io.in(i - 1).valid && io.in(i - 1).bits.ldest === io.in(i).bits.lsrc(0)
381    val fused_lui_load = last_is_lui && this_is_load && lui_to_load
382    when (fused_lui_load) {
383      // The first LOAD operand (base address) is replaced by LUI-imm and stored in imm
384      val lui_imm = io.in(i - 1).bits.imm(ImmUnion.U.len - 1, 0)
385      val ld_imm = io.in(i).bits.imm(ImmUnion.I.len - 1, 0)
386      require(io.out(i).bits.imm.getWidth >= lui_imm.getWidth + ld_imm.getWidth)
387      io.out(i).bits.srcType(0) := SrcType.imm
388      io.out(i).bits.imm := Cat(lui_imm, ld_imm)
389    }
390
391  }
392
393  val genSnapshot = Cat(io.out.map(out => out.fire && out.bits.snapshot)).orR
394  val lastCycleCreateSnpt = RegInit(false.B)
395  lastCycleCreateSnpt := genSnapshot && !io.snptIsFull
396  val sameSnptDistance = (RobCommitWidth * 4).U
397  // notInSameSnpt: 1.robidxHead - snapLastEnq >= sameSnptDistance 2.no snap
398  val notInSameSnpt = GatedValidRegNext(distanceBetween(robIdxHeadNext, io.snptLastEnq.bits) >= sameSnptDistance || !io.snptLastEnq.valid)
399  val allowSnpt = if (EnableRenameSnapshot) notInSameSnpt && !lastCycleCreateSnpt && io.in.head.bits.firstUop else false.B
400  io.out.zip(io.in).foreach{ case (out, in) => out.bits.snapshot := allowSnpt && (!in.bits.preDecodeInfo.notCFI || FuType.isJump(in.bits.fuType)) && in.fire }
401  io.out.map{ x =>
402    x.bits.hasException := Cat(selectFrontend(x.bits.exceptionVec) :+ x.bits.exceptionVec(illegalInstr) :+ x.bits.exceptionVec(virtualInstr)).orR || x.bits.trigger.getFrontendCanFire
403  }
404  if(backendParams.debugEn){
405    dontTouch(robIdxHeadNext)
406    dontTouch(notInSameSnpt)
407    dontTouch(genSnapshot)
408  }
409  intFreeList.io.snpt := io.snpt
410  fpFreeList.io.snpt := io.snpt
411  vecFreeList.io.snpt := io.snpt
412  intFreeList.io.snpt.snptEnq := genSnapshot
413  fpFreeList.io.snpt.snptEnq := genSnapshot
414  vecFreeList.io.snpt.snptEnq := genSnapshot
415
416  /**
417    * Instructions commit: update freelist and rename table
418    */
419  for (i <- 0 until RabCommitWidth) {
420    val commitValid = io.rabCommits.isCommit && io.rabCommits.commitValid(i)
421    val walkValid = io.rabCommits.isWalk && io.rabCommits.walkValid(i)
422
423    // I. RAT Update
424    // When redirect happens (mis-prediction), don't update the rename table
425    io.intRenamePorts(i).wen  := intSpecWen(i)
426    io.intRenamePorts(i).addr := uops(i).ldest
427    io.intRenamePorts(i).data := io.out(i).bits.pdest
428
429    io.fpRenamePorts(i).wen  := fpSpecWen(i)
430    io.fpRenamePorts(i).addr := uops(i).ldest
431    io.fpRenamePorts(i).data := fpFreeList.io.allocatePhyReg(i)
432
433    io.vecRenamePorts(i).wen := vecSpecWen(i)
434    io.vecRenamePorts(i).addr := uops(i).ldest
435    io.vecRenamePorts(i).data := vecFreeList.io.allocatePhyReg(i)
436
437    // II. Free List Update
438    intFreeList.io.freeReq(i) := io.int_need_free(i)
439    intFreeList.io.freePhyReg(i) := RegNext(io.int_old_pdest(i))
440    fpFreeList.io.freeReq(i)  := GatedValidRegNext(commitValid && needDestRegCommit(Reg_F, io.rabCommits.info(i)))
441    fpFreeList.io.freePhyReg(i) := io.fp_old_pdest(i)
442    vecFreeList.io.freeReq(i)  := GatedValidRegNext(commitValid && needDestRegCommit(Reg_V, io.rabCommits.info(i)))
443    vecFreeList.io.freePhyReg(i) := io.vec_old_pdest(i)
444  }
445
446  /*
447  Debug and performance counters
448   */
449  def printRenameInfo(in: DecoupledIO[DecodedInst], out: DecoupledIO[DynInst]) = {
450    XSInfo(out.fire, p"pc:${Hexadecimal(in.bits.pc)} in(${in.valid},${in.ready}) " +
451      p"lsrc(0):${in.bits.lsrc(0)} -> psrc(0):${out.bits.psrc(0)} " +
452      p"lsrc(1):${in.bits.lsrc(1)} -> psrc(1):${out.bits.psrc(1)} " +
453      p"lsrc(2):${in.bits.lsrc(2)} -> psrc(2):${out.bits.psrc(2)} " +
454      p"ldest:${in.bits.ldest} -> pdest:${out.bits.pdest}\n"
455    )
456  }
457
458  for ((x,y) <- io.in.zip(io.out)) {
459    printRenameInfo(x, y)
460  }
461
462  io.out.map { case x =>
463    when(x.valid && x.bits.rfWen){
464      assert(x.bits.ldest =/= 0.U, "rfWen cannot be 1 when Int regfile ldest is 0")
465    }
466  }
467  val debugRedirect = RegEnable(io.redirect.bits, io.redirect.valid)
468  // bad speculation
469  val recStall = io.redirect.valid || io.rabCommits.isWalk
470  val ctrlRecStall = Mux(io.redirect.valid, io.redirect.bits.debugIsCtrl, io.rabCommits.isWalk && debugRedirect.debugIsCtrl)
471  val mvioRecStall = Mux(io.redirect.valid, io.redirect.bits.debugIsMemVio, io.rabCommits.isWalk && debugRedirect.debugIsMemVio)
472  val otherRecStall = recStall && !(ctrlRecStall || mvioRecStall)
473  XSPerfAccumulate("recovery_stall", recStall)
474  XSPerfAccumulate("control_recovery_stall", ctrlRecStall)
475  XSPerfAccumulate("mem_violation_recovery_stall", mvioRecStall)
476  XSPerfAccumulate("other_recovery_stall", otherRecStall)
477  // freelist stall
478  val notRecStall = !io.out.head.valid && !recStall
479  val intFlStall = notRecStall && inHeadValid && !intFreeList.io.canAllocate
480  val fpFlStall = notRecStall && inHeadValid && intFreeList.io.canAllocate && vecFreeList.io.canAllocate && !fpFreeList.io.canAllocate
481  val vecFlStall = notRecStall && inHeadValid && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && !vecFreeList.io.canAllocate
482  // other stall
483  val otherStall = notRecStall && !intFlStall && !fpFlStall && !vecFlStall
484
485  io.stallReason.in.backReason.valid := io.stallReason.out.backReason.valid || !io.in.head.ready
486  io.stallReason.in.backReason.bits := Mux(io.stallReason.out.backReason.valid, io.stallReason.out.backReason.bits,
487    MuxCase(TopDownCounters.OtherCoreStall.id.U, Seq(
488      ctrlRecStall  -> TopDownCounters.ControlRecoveryStall.id.U,
489      mvioRecStall  -> TopDownCounters.MemVioRecoveryStall.id.U,
490      otherRecStall -> TopDownCounters.OtherRecoveryStall.id.U,
491      intFlStall    -> TopDownCounters.IntFlStall.id.U,
492      fpFlStall     -> TopDownCounters.FpFlStall.id.U,
493      vecFlStall    -> TopDownCounters.VecFlStall.id.U,
494    )
495  ))
496  io.stallReason.out.reason.zip(io.stallReason.in.reason).zip(io.in.map(_.valid)).foreach { case ((out, in), valid) =>
497    out := Mux(io.stallReason.in.backReason.valid, io.stallReason.in.backReason.bits, in)
498  }
499
500  XSDebug(io.rabCommits.isWalk, p"Walk Recovery Enabled\n")
501  XSDebug(io.rabCommits.isWalk, p"validVec:${Binary(io.rabCommits.walkValid.asUInt)}\n")
502  for (i <- 0 until RabCommitWidth) {
503    val info = io.rabCommits.info(i)
504    XSDebug(io.rabCommits.isWalk && io.rabCommits.walkValid(i), p"[#$i walk info] " +
505      p"ldest:${info.ldest} rfWen:${info.rfWen} fpWen:${info.fpWen} vecWen:${info.vecWen}")
506  }
507
508  XSDebug(p"inValidVec: ${Binary(Cat(io.in.map(_.valid)))}\n")
509
510  XSPerfAccumulate("in_valid_count", PopCount(io.in.map(_.valid)))
511  XSPerfAccumulate("in_fire_count", PopCount(io.in.map(_.fire)))
512  XSPerfAccumulate("in_valid_not_ready_count", PopCount(io.in.map(x => x.valid && !x.ready)))
513  XSPerfAccumulate("wait_cycle", !io.in.head.valid && dispatchCanAcc)
514
515  // These stall reasons could overlap each other, but we configure the priority as fellows.
516  // walk stall > dispatch stall > int freelist stall > fp freelist stall
517  private val inHeadStall = io.in.head match { case x => x.valid && !x.ready }
518  private val stallForWalk      = inHeadValid &&  io.rabCommits.isWalk
519  private val stallForDispatch  = inHeadValid && !io.rabCommits.isWalk && !dispatchCanAcc
520  private val stallForIntFL     = inHeadValid && !io.rabCommits.isWalk &&  dispatchCanAcc && !intFreeList.io.canAllocate
521  private val stallForFpFL      = inHeadValid && !io.rabCommits.isWalk &&  dispatchCanAcc &&  intFreeList.io.canAllocate && vecFreeList.io.canAllocate && !fpFreeList.io.canAllocate
522  private val stallForVecFL     = inHeadValid && !io.rabCommits.isWalk &&  dispatchCanAcc &&  intFreeList.io.canAllocate && fpFreeList.io.canAllocate && !vecFreeList.io.canAllocate
523  XSPerfAccumulate("stall_cycle",          inHeadStall)
524  XSPerfAccumulate("stall_cycle_walk",     stallForWalk)
525  XSPerfAccumulate("stall_cycle_dispatch", stallForDispatch)
526  XSPerfAccumulate("stall_cycle_int",      stallForIntFL)
527  XSPerfAccumulate("stall_cycle_fp",       stallForFpFL)
528  XSPerfAccumulate("stall_cycle_vec",      stallForVecFL)
529
530  XSPerfHistogram("in_valid_range",  PopCount(io.in.map(_.valid)),  true.B, 0, DecodeWidth + 1, 1)
531  XSPerfHistogram("in_fire_range",   PopCount(io.in.map(_.fire)),   true.B, 0, DecodeWidth + 1, 1)
532  XSPerfHistogram("out_valid_range", PopCount(io.out.map(_.valid)), true.B, 0, DecodeWidth + 1, 1)
533  XSPerfHistogram("out_fire_range",  PopCount(io.out.map(_.fire)),  true.B, 0, DecodeWidth + 1, 1)
534
535  XSPerfAccumulate("move_instr_count", PopCount(io.out.map(out => out.fire && out.bits.isMove)))
536  val is_fused_lui_load = io.out.map(o => o.fire && o.bits.fuType === FuType.ldu.U && o.bits.srcType(0) === SrcType.imm)
537  XSPerfAccumulate("fused_lui_load_instr_count", PopCount(is_fused_lui_load))
538
539  val renamePerf = Seq(
540    ("rename_in                  ", PopCount(io.in.map(_.valid & io.in(0).ready ))                                                               ),
541    ("rename_waitinstr           ", PopCount((0 until RenameWidth).map(i => io.in(i).valid && !io.in(i).ready))                                  ),
542    ("rename_stall               ", inHeadStall),
543    ("rename_stall_cycle_walk    ", inHeadValid &&  io.rabCommits.isWalk),
544    ("rename_stall_cycle_dispatch", inHeadValid && !io.rabCommits.isWalk && !dispatchCanAcc),
545    ("rename_stall_cycle_int     ", inHeadValid && !io.rabCommits.isWalk &&  dispatchCanAcc && !intFreeList.io.canAllocate),
546    ("rename_stall_cycle_fp      ", inHeadValid && !io.rabCommits.isWalk &&  dispatchCanAcc && intFreeList.io.canAllocate && vecFreeList.io.canAllocate && !fpFreeList.io.canAllocate),
547    ("rename_stall_cycle_vec     ", inHeadValid && !io.rabCommits.isWalk &&  dispatchCanAcc && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && !vecFreeList.io.canAllocate),
548  )
549  val intFlPerf = intFreeList.getPerfEvents
550  val fpFlPerf = fpFreeList.getPerfEvents
551  val vecFlPerf = vecFreeList.getPerfEvents
552  val perfEvents = renamePerf ++ intFlPerf ++ fpFlPerf ++ vecFlPerf
553  generatePerfEvent()
554}
555