xref: /XiangShan/src/main/scala/xiangshan/backend/rename/Rename.scala (revision 6374b1d62bdc41478bd6fe6a4dc48e131c2769f7)
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  // io alias
82  private val dispatchCanAcc = io.out.head.ready
83
84  val compressUnit = Module(new CompressUnit())
85  // create free list and rat
86  val intFreeList = Module(new MEFreeList(IntPhyRegs))
87  val fpFreeList = Module(new StdFreeList(FpPhyRegs - FpLogicRegs, FpLogicRegs, Reg_F))
88  val vecFreeList = Module(new StdFreeList(VfPhyRegs - VecLogicRegs, VecLogicRegs, Reg_V))
89
90  intFreeList.io.commit    <> io.rabCommits
91  intFreeList.io.debug_rat.foreach(_ <> io.debug_int_rat.get)
92  fpFreeList.io.commit     <> io.rabCommits
93  fpFreeList.io.debug_rat.foreach(_ <> io.debug_fp_rat.get)
94  vecFreeList.io.commit    <> io.rabCommits
95  vecFreeList.io.debug_rat.foreach(_ <> io.debug_vec_rat.get)
96
97  // decide if given instruction needs allocating a new physical register (CfCtrl: from decode; RobCommitInfo: from rob)
98  def needDestReg[T <: DecodedInst](reg_t: RegType, x: T): Bool = reg_t match {
99    case Reg_I => x.rfWen && x.ldest =/= 0.U
100    case Reg_F => x.fpWen
101    case Reg_V => x.vecWen
102  }
103  def needDestRegCommit[T <: RabCommitInfo](reg_t: RegType, x: T): Bool = {
104    reg_t match {
105      case Reg_I => x.rfWen
106      case Reg_F => x.fpWen
107      case Reg_V => x.vecWen
108    }
109  }
110  def needDestRegWalk[T <: RabCommitInfo](reg_t: RegType, x: T): Bool = {
111    reg_t match {
112      case Reg_I => x.rfWen && x.ldest =/= 0.U
113      case Reg_F => x.fpWen
114      case Reg_V => x.vecWen
115    }
116  }
117
118  // connect [redirect + walk] ports for fp & vec & int free list
119  Seq(fpFreeList, vecFreeList, intFreeList).foreach { case fl =>
120    fl.io.redirect := io.redirect.valid
121    fl.io.walk := io.rabCommits.isWalk
122  }
123  // only when all free list and dispatch1 has enough space can we do allocation
124  // when isWalk, freelist can definitely allocate
125  intFreeList.io.doAllocate := fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && dispatchCanAcc || io.rabCommits.isWalk
126  fpFreeList.io.doAllocate := intFreeList.io.canAllocate && vecFreeList.io.canAllocate && dispatchCanAcc || io.rabCommits.isWalk
127  vecFreeList.io.doAllocate := intFreeList.io.canAllocate && fpFreeList.io.canAllocate && dispatchCanAcc || io.rabCommits.isWalk
128
129  //           dispatch1 ready ++ float point free list ready ++ int free list ready ++ vec free list ready     ++ not walk
130  val canOut = dispatchCanAcc && fpFreeList.io.canAllocate && intFreeList.io.canAllocate && vecFreeList.io.canAllocate && !io.rabCommits.isWalk
131
132  compressUnit.io.in.zip(io.in).foreach{ case(sink, source) =>
133    sink.valid := source.valid
134    sink.bits := source.bits
135  }
136  val needRobFlags = compressUnit.io.out.needRobFlags
137  val instrSizesVec = compressUnit.io.out.instrSizes
138  val compressMasksVec = compressUnit.io.out.masks
139
140  // speculatively assign the instruction with an robIdx
141  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)
142  val robIdxHead = RegInit(0.U.asTypeOf(new RobPtr))
143  val lastCycleMisprediction = GatedValidRegNext(io.redirect.valid && !io.redirect.bits.flushItself())
144  val robIdxHeadNext = Mux(io.redirect.valid, io.redirect.bits.robIdx, // redirect: move ptr to given rob index
145         Mux(lastCycleMisprediction, robIdxHead + 1.U, // mis-predict: not flush robIdx itself
146           Mux(canOut, robIdxHead + validCount, // instructions successfully entered next stage: increase robIdx
147                      /* default */  robIdxHead))) // no instructions passed by this cycle: stick to old value
148  robIdxHead := robIdxHeadNext
149
150  /**
151    * Rename: allocate free physical register and update rename table
152    */
153  val uops = Wire(Vec(RenameWidth, new DynInst))
154  uops.foreach( uop => {
155    uop.srcState      := DontCare
156    uop.debugInfo     := DontCare
157    uop.lqIdx         := DontCare
158    uop.sqIdx         := DontCare
159    uop.waitForRobIdx := DontCare
160    uop.singleStep    := DontCare
161    uop.snapshot      := DontCare
162    uop.srcLoadDependency := DontCare
163    uop.numLsElem       :=  DontCare
164  })
165
166  val needVecDest    = Wire(Vec(RenameWidth, Bool()))
167  val needFpDest     = Wire(Vec(RenameWidth, Bool()))
168  val needIntDest    = Wire(Vec(RenameWidth, Bool()))
169  val hasValid = Cat(io.in.map(_.valid)).orR
170  private val inHeadValid = io.in.head.valid
171
172  val isMove = Wire(Vec(RenameWidth, Bool()))
173  isMove zip io.in.map(_.bits) foreach {
174    case (move, in) => move := Mux(in.exceptionVec.asUInt.orR, false.B, in.isMove)
175  }
176
177  val walkNeedIntDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
178  val walkNeedFpDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
179  val walkNeedVecDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
180  val walkIsMove = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
181
182  val intSpecWen = Wire(Vec(RenameWidth, Bool()))
183  val fpSpecWen  = Wire(Vec(RenameWidth, Bool()))
184  val vecSpecWen = Wire(Vec(RenameWidth, Bool()))
185
186  val walkIntSpecWen = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
187
188  val walkPdest = Wire(Vec(RenameWidth, UInt(PhyRegIdxWidth.W)))
189
190  // uop calculation
191  for (i <- 0 until RenameWidth) {
192    (uops(i): Data).waiveAll :<= (io.in(i).bits: Data).waiveAll
193
194    // update cf according to ssit result
195    uops(i).storeSetHit := io.ssit(i).valid
196    uops(i).loadWaitStrict := io.ssit(i).strict && io.ssit(i).valid
197    uops(i).ssid := io.ssit(i).ssid
198
199    // update cf according to waittable result
200    uops(i).loadWaitBit := io.waittable(i)
201
202    uops(i).replayInst := false.B // set by IQ or MemQ
203    // alloc a new phy reg
204    needVecDest(i) := io.in(i).valid && needDestReg(Reg_V, io.in(i).bits)
205    needFpDest(i) := io.in(i).valid && needDestReg(Reg_F, io.in(i).bits)
206    needIntDest(i) := io.in(i).valid && needDestReg(Reg_I, io.in(i).bits)
207    if (i < RabCommitWidth) {
208      walkNeedIntDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_I, io.rabCommits.info(i))
209      walkNeedFpDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_F, io.rabCommits.info(i))
210      walkNeedVecDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_V, io.rabCommits.info(i))
211      walkIsMove(i) := io.rabCommits.info(i).isMove
212    }
213    fpFreeList.io.allocateReq(i) := needFpDest(i)
214    fpFreeList.io.walkReq(i) := walkNeedFpDest(i)
215    vecFreeList.io.allocateReq(i) := needVecDest(i)
216    vecFreeList.io.walkReq(i) := walkNeedVecDest(i)
217    intFreeList.io.allocateReq(i) := needIntDest(i) && !isMove(i)
218    intFreeList.io.walkReq(i) := walkNeedIntDest(i) && !walkIsMove(i)
219
220    // no valid instruction from decode stage || all resources (dispatch1 + both free lists) ready
221    io.in(i).ready := !hasValid || canOut
222
223    uops(i).robIdx := robIdxHead + PopCount(io.in.zip(needRobFlags).take(i).map{ case(in, needRobFlag) => in.valid && in.bits.lastUop && needRobFlag})
224    uops(i).instrSize := instrSizesVec(i)
225    when(isMove(i)) {
226      uops(i).numUops := 0.U
227      uops(i).numWB := 0.U
228    }
229    if (i > 0) {
230      when(!needRobFlags(i - 1)) {
231        uops(i).firstUop := false.B
232        uops(i).ftqPtr := uops(i - 1).ftqPtr
233        uops(i).ftqOffset := uops(i - 1).ftqOffset
234        uops(i).numUops := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
235        uops(i).numWB := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
236      }
237    }
238    when(!needRobFlags(i)) {
239      uops(i).lastUop := false.B
240      uops(i).numUops := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
241      uops(i).numWB := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
242    }
243    uops(i).wfflags := (compressMasksVec(i) & Cat(io.in.map(_.bits.wfflags).reverse)).orR
244    uops(i).dirtyFs := (compressMasksVec(i) & Cat(io.in.map(_.bits.fpWen).reverse)).orR
245    // vector instructions' uopSplitType cannot be UopSplitType.SCA_SIM
246    uops(i).dirtyVs := (compressMasksVec(i) & Cat(io.in.map(_.bits.uopSplitType =/= UopSplitType.SCA_SIM).reverse)).orR
247
248    uops(i).psrc(0) := Mux1H(uops(i).srcType(0), Seq(io.intReadPorts(i)(0), io.fpReadPorts(i)(0), io.vecReadPorts(i)(0)))
249    uops(i).psrc(1) := Mux1H(uops(i).srcType(1), Seq(io.intReadPorts(i)(1), io.fpReadPorts(i)(1), io.vecReadPorts(i)(1)))
250    uops(i).psrc(2) := Mux1H(uops(i).srcType(2)(2, 1), Seq(io.fpReadPorts(i)(2), io.vecReadPorts(i)(2)))
251    uops(i).psrc(3) := io.vecReadPorts(i)(3)
252    uops(i).psrc(4) := io.vecReadPorts(i)(4) // Todo: vl read port
253
254    // int psrc2 should be bypassed from next instruction if it is fused
255    if (i < RenameWidth - 1) {
256      when (io.fusionInfo(i).rs2FromRs2 || io.fusionInfo(i).rs2FromRs1) {
257        uops(i).psrc(1) := Mux(io.fusionInfo(i).rs2FromRs2, io.intReadPorts(i + 1)(1), io.intReadPorts(i + 1)(0))
258      }.elsewhen(io.fusionInfo(i).rs2FromZero) {
259        uops(i).psrc(1) := 0.U
260      }
261    }
262    uops(i).eliminatedMove := isMove(i)
263
264    // update pdest
265    uops(i).pdest := MuxCase(0.U, Seq(
266      needIntDest(i)    ->  intFreeList.io.allocatePhyReg(i),
267      needFpDest(i)     ->  fpFreeList.io.allocatePhyReg(i),
268      needVecDest(i)    ->  vecFreeList.io.allocatePhyReg(i),
269    ))
270
271    // Assign performance counters
272    uops(i).debugInfo.renameTime := GTimer()
273
274    io.out(i).valid := io.in(i).valid && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && !io.rabCommits.isWalk
275    io.out(i).bits := uops(i)
276    // Todo: move these shit in decode stage
277    // dirty code for fence. The lsrc is passed by imm.
278    when (io.out(i).bits.fuType === FuType.fence.U) {
279      io.out(i).bits.imm := Cat(io.in(i).bits.lsrc(1), io.in(i).bits.lsrc(0))
280    }
281
282    // dirty code for SoftPrefetch (prefetch.r/prefetch.w)
283//    when (io.in(i).bits.isSoftPrefetch) {
284//      io.out(i).bits.fuType := FuType.ldu.U
285//      io.out(i).bits.fuOpType := Mux(io.in(i).bits.lsrc(1) === 1.U, LSUOpType.prefetch_r, LSUOpType.prefetch_w)
286//      io.out(i).bits.selImm := SelImm.IMM_S
287//      io.out(i).bits.imm := Cat(io.in(i).bits.imm(io.in(i).bits.imm.getWidth - 1, 5), 0.U(5.W))
288//    }
289
290    // dirty code for lui+addi(w) fusion
291    if (i < RenameWidth - 1) {
292      val fused_lui32 = io.in(i).bits.selImm === SelImm.IMM_LUI32 && io.in(i).bits.fuType === FuType.alu.U
293      when (fused_lui32) {
294        val lui_imm = io.in(i).bits.imm(19, 0)
295        val add_imm = io.in(i + 1).bits.imm(11, 0)
296        require(io.out(i).bits.imm.getWidth >= lui_imm.getWidth + add_imm.getWidth)
297        io.out(i).bits.imm := Cat(lui_imm, add_imm)
298      }
299    }
300
301    // write speculative rename table
302    // we update rat later inside commit code
303    intSpecWen(i) := needIntDest(i) && intFreeList.io.canAllocate && intFreeList.io.doAllocate && !io.rabCommits.isWalk && !io.redirect.valid
304    fpSpecWen(i)  := needFpDest(i)  && fpFreeList.io.canAllocate  && fpFreeList.io.doAllocate  && !io.rabCommits.isWalk && !io.redirect.valid
305    vecSpecWen(i) := needVecDest(i) && vecFreeList.io.canAllocate && vecFreeList.io.doAllocate && !io.rabCommits.isWalk && !io.redirect.valid
306
307
308    if (i < RabCommitWidth) {
309      walkIntSpecWen(i) := walkNeedIntDest(i) && !io.redirect.valid
310      walkPdest(i) := io.rabCommits.info(i).pdest
311    } else {
312      walkPdest(i) := io.out(i).bits.pdest
313    }
314  }
315
316  /**
317    * How to set psrc:
318    * - bypass the pdest to psrc if previous instructions write to the same ldest as lsrc
319    * - default: psrc from RAT
320    * How to set pdest:
321    * - Mux(isMove, psrc, pdest_from_freelist).
322    *
323    * The critical path of rename lies here:
324    * When move elimination is enabled, we need to update the rat with psrc.
325    * However, psrc maybe comes from previous instructions' pdest, which comes from freelist.
326    *
327    * If we expand these logic for pdest(N):
328    * pdest(N) = Mux(isMove(N), psrc(N), freelist_out(N))
329    *          = Mux(isMove(N), Mux(bypass(N, N - 1), pdest(N - 1),
330    *                           Mux(bypass(N, N - 2), pdest(N - 2),
331    *                           ...
332    *                           Mux(bypass(N, 0),     pdest(0),
333    *                                                 rat_out(N))...)),
334    *                           freelist_out(N))
335    */
336  // a simple functional model for now
337  io.out(0).bits.pdest := Mux(isMove(0), uops(0).psrc.head, uops(0).pdest)
338
339  // psrc(n) + pdest(1)
340  val bypassCond: Vec[MixedVec[UInt]] = Wire(Vec(numRegSrc + 1, MixedVec(List.tabulate(RenameWidth-1)(i => UInt((i+1).W)))))
341  require(io.in(0).bits.srcType.size == io.in(0).bits.numSrc)
342  private val pdestLoc = io.in.head.bits.srcType.size // 2 vector src: v0, vl&vtype
343  println(s"[Rename] idx of pdest in bypassCond $pdestLoc")
344  for (i <- 1 until RenameWidth) {
345    val vecCond = io.in(i).bits.srcType.map(_ === SrcType.vp) :+ needVecDest(i)
346    val fpCond  = io.in(i).bits.srcType.map(_ === SrcType.fp) :+ needFpDest(i)
347    val intCond = io.in(i).bits.srcType.map(_ === SrcType.xp) :+ needIntDest(i)
348    val target = io.in(i).bits.lsrc :+ io.in(i).bits.ldest
349    for (((((cond1, cond2), cond3), t), j) <- vecCond.zip(fpCond).zip(intCond).zip(target).zipWithIndex) {
350      val destToSrc = io.in.take(i).zipWithIndex.map { case (in, j) =>
351        val indexMatch = in.bits.ldest === t
352        val writeMatch =  cond3 && needIntDest(j) || cond2 && needFpDest(j) || cond1 && needVecDest(j)
353        indexMatch && writeMatch
354      }
355      bypassCond(j)(i - 1) := VecInit(destToSrc).asUInt
356    }
357    io.out(i).bits.psrc(0) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(0)(i-1).asBools).foldLeft(uops(i).psrc(0)) {
358      (z, next) => Mux(next._2, next._1, z)
359    }
360    io.out(i).bits.psrc(1) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(1)(i-1).asBools).foldLeft(uops(i).psrc(1)) {
361      (z, next) => Mux(next._2, next._1, z)
362    }
363    io.out(i).bits.psrc(2) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(2)(i-1).asBools).foldLeft(uops(i).psrc(2)) {
364      (z, next) => Mux(next._2, next._1, z)
365    }
366    io.out(i).bits.psrc(3) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(3)(i-1).asBools).foldLeft(uops(i).psrc(3)) {
367      (z, next) => Mux(next._2, next._1, z)
368    }
369    io.out(i).bits.psrc(4) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(4)(i-1).asBools).foldLeft(uops(i).psrc(4)) {
370      (z, next) => Mux(next._2, next._1, z)
371    }
372    io.out(i).bits.pdest := Mux(isMove(i), io.out(i).bits.psrc(0), uops(i).pdest)
373
374    // Todo: better implementation for fields reuse
375    // For fused-lui-load, load.src(0) is replaced by the imm.
376    val last_is_lui = io.in(i - 1).bits.selImm === SelImm.IMM_U && io.in(i - 1).bits.srcType(0) =/= SrcType.pc
377    val this_is_load = io.in(i).bits.fuType === FuType.ldu.U
378    val lui_to_load = io.in(i - 1).valid && io.in(i - 1).bits.ldest === io.in(i).bits.lsrc(0)
379    val fused_lui_load = last_is_lui && this_is_load && lui_to_load
380    when (fused_lui_load) {
381      // The first LOAD operand (base address) is replaced by LUI-imm and stored in imm
382      val lui_imm = io.in(i - 1).bits.imm(ImmUnion.U.len - 1, 0)
383      val ld_imm = io.in(i).bits.imm(ImmUnion.I.len - 1, 0)
384      require(io.out(i).bits.imm.getWidth >= lui_imm.getWidth + ld_imm.getWidth)
385      io.out(i).bits.srcType(0) := SrcType.imm
386      io.out(i).bits.imm := Cat(lui_imm, ld_imm)
387    }
388
389  }
390
391  val genSnapshot = Cat(io.out.map(out => out.fire && out.bits.snapshot)).orR
392  val lastCycleCreateSnpt = RegInit(false.B)
393  lastCycleCreateSnpt := genSnapshot && !io.snptIsFull
394  val sameSnptDistance = (RobCommitWidth * 4).U
395  // notInSameSnpt: 1.robidxHead - snapLastEnq >= sameSnptDistance 2.no snap
396  val notInSameSnpt = GatedValidRegNext(distanceBetween(robIdxHeadNext, io.snptLastEnq.bits) >= sameSnptDistance || !io.snptLastEnq.valid)
397  val allowSnpt = if (EnableRenameSnapshot) notInSameSnpt && !lastCycleCreateSnpt && io.in.head.bits.firstUop else false.B
398  io.out.zip(io.in).foreach{ case (out, in) => out.bits.snapshot := allowSnpt && (!in.bits.preDecodeInfo.notCFI || FuType.isJump(in.bits.fuType)) && in.fire }
399  if(backendParams.debugEn){
400    dontTouch(robIdxHeadNext)
401    dontTouch(notInSameSnpt)
402    dontTouch(genSnapshot)
403  }
404  intFreeList.io.snpt := io.snpt
405  fpFreeList.io.snpt := io.snpt
406  vecFreeList.io.snpt := io.snpt
407  intFreeList.io.snpt.snptEnq := genSnapshot
408  fpFreeList.io.snpt.snptEnq := genSnapshot
409  vecFreeList.io.snpt.snptEnq := genSnapshot
410
411  /**
412    * Instructions commit: update freelist and rename table
413    */
414  for (i <- 0 until RabCommitWidth) {
415    val commitValid = io.rabCommits.isCommit && io.rabCommits.commitValid(i)
416    val walkValid = io.rabCommits.isWalk && io.rabCommits.walkValid(i)
417
418    // I. RAT Update
419    // When redirect happens (mis-prediction), don't update the rename table
420    io.intRenamePorts(i).wen  := intSpecWen(i)
421    io.intRenamePorts(i).addr := uops(i).ldest
422    io.intRenamePorts(i).data := io.out(i).bits.pdest
423
424    io.fpRenamePorts(i).wen  := fpSpecWen(i)
425    io.fpRenamePorts(i).addr := uops(i).ldest
426    io.fpRenamePorts(i).data := fpFreeList.io.allocatePhyReg(i)
427
428    io.vecRenamePorts(i).wen := vecSpecWen(i)
429    io.vecRenamePorts(i).addr := uops(i).ldest
430    io.vecRenamePorts(i).data := vecFreeList.io.allocatePhyReg(i)
431
432    // II. Free List Update
433    intFreeList.io.freeReq(i) := io.int_need_free(i)
434    intFreeList.io.freePhyReg(i) := RegNext(io.int_old_pdest(i))
435    fpFreeList.io.freeReq(i)  := GatedValidRegNext(commitValid && needDestRegCommit(Reg_F, io.rabCommits.info(i)))
436    fpFreeList.io.freePhyReg(i) := io.fp_old_pdest(i)
437    vecFreeList.io.freeReq(i)  := GatedValidRegNext(commitValid && needDestRegCommit(Reg_V, io.rabCommits.info(i)))
438    vecFreeList.io.freePhyReg(i) := io.vec_old_pdest(i)
439  }
440
441  /*
442  Debug and performance counters
443   */
444  def printRenameInfo(in: DecoupledIO[DecodedInst], out: DecoupledIO[DynInst]) = {
445    XSInfo(out.fire, p"pc:${Hexadecimal(in.bits.pc)} in(${in.valid},${in.ready}) " +
446      p"lsrc(0):${in.bits.lsrc(0)} -> psrc(0):${out.bits.psrc(0)} " +
447      p"lsrc(1):${in.bits.lsrc(1)} -> psrc(1):${out.bits.psrc(1)} " +
448      p"lsrc(2):${in.bits.lsrc(2)} -> psrc(2):${out.bits.psrc(2)} " +
449      p"ldest:${in.bits.ldest} -> pdest:${out.bits.pdest}\n"
450    )
451  }
452
453  for ((x,y) <- io.in.zip(io.out)) {
454    printRenameInfo(x, y)
455  }
456
457  io.out.map { case x =>
458    when(x.valid && x.bits.rfWen){
459      assert(x.bits.ldest =/= 0.U, "rfWen cannot be 1 when Int regfile ldest is 0")
460    }
461  }
462  val debugRedirect = RegEnable(io.redirect.bits, io.redirect.valid)
463  // bad speculation
464  val recStall = io.redirect.valid || io.rabCommits.isWalk
465  val ctrlRecStall = Mux(io.redirect.valid, io.redirect.bits.debugIsCtrl, io.rabCommits.isWalk && debugRedirect.debugIsCtrl)
466  val mvioRecStall = Mux(io.redirect.valid, io.redirect.bits.debugIsMemVio, io.rabCommits.isWalk && debugRedirect.debugIsMemVio)
467  val otherRecStall = recStall && !(ctrlRecStall || mvioRecStall)
468  XSPerfAccumulate("recovery_stall", recStall)
469  XSPerfAccumulate("control_recovery_stall", ctrlRecStall)
470  XSPerfAccumulate("mem_violation_recovery_stall", mvioRecStall)
471  XSPerfAccumulate("other_recovery_stall", otherRecStall)
472  // freelist stall
473  val notRecStall = !io.out.head.valid && !recStall
474  val intFlStall = notRecStall && inHeadValid && !intFreeList.io.canAllocate
475  val fpFlStall = notRecStall && inHeadValid && intFreeList.io.canAllocate && vecFreeList.io.canAllocate && !fpFreeList.io.canAllocate
476  val vecFlStall = notRecStall && inHeadValid && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && !vecFreeList.io.canAllocate
477  // other stall
478  val otherStall = notRecStall && !intFlStall && !fpFlStall && !vecFlStall
479
480  io.stallReason.in.backReason.valid := io.stallReason.out.backReason.valid || !io.in.head.ready
481  io.stallReason.in.backReason.bits := Mux(io.stallReason.out.backReason.valid, io.stallReason.out.backReason.bits,
482    MuxCase(TopDownCounters.OtherCoreStall.id.U, Seq(
483      ctrlRecStall  -> TopDownCounters.ControlRecoveryStall.id.U,
484      mvioRecStall  -> TopDownCounters.MemVioRecoveryStall.id.U,
485      otherRecStall -> TopDownCounters.OtherRecoveryStall.id.U,
486      intFlStall    -> TopDownCounters.IntFlStall.id.U,
487      fpFlStall     -> TopDownCounters.FpFlStall.id.U,
488      vecFlStall    -> TopDownCounters.VecFlStall.id.U,
489    )
490  ))
491  io.stallReason.out.reason.zip(io.stallReason.in.reason).zip(io.in.map(_.valid)).foreach { case ((out, in), valid) =>
492    out := Mux(io.stallReason.in.backReason.valid, io.stallReason.in.backReason.bits, in)
493  }
494
495  XSDebug(io.rabCommits.isWalk, p"Walk Recovery Enabled\n")
496  XSDebug(io.rabCommits.isWalk, p"validVec:${Binary(io.rabCommits.walkValid.asUInt)}\n")
497  for (i <- 0 until RabCommitWidth) {
498    val info = io.rabCommits.info(i)
499    XSDebug(io.rabCommits.isWalk && io.rabCommits.walkValid(i), p"[#$i walk info] " +
500      p"ldest:${info.ldest} rfWen:${info.rfWen} fpWen:${info.fpWen} vecWen:${info.vecWen}")
501  }
502
503  XSDebug(p"inValidVec: ${Binary(Cat(io.in.map(_.valid)))}\n")
504
505  XSPerfAccumulate("in_valid_count", PopCount(io.in.map(_.valid)))
506  XSPerfAccumulate("in_fire_count", PopCount(io.in.map(_.fire)))
507  XSPerfAccumulate("in_valid_not_ready_count", PopCount(io.in.map(x => x.valid && !x.ready)))
508  XSPerfAccumulate("wait_cycle", !io.in.head.valid && dispatchCanAcc)
509
510  // These stall reasons could overlap each other, but we configure the priority as fellows.
511  // walk stall > dispatch stall > int freelist stall > fp freelist stall
512  private val inHeadStall = io.in.head match { case x => x.valid && !x.ready }
513  private val stallForWalk      = inHeadValid &&  io.rabCommits.isWalk
514  private val stallForDispatch  = inHeadValid && !io.rabCommits.isWalk && !dispatchCanAcc
515  private val stallForIntFL     = inHeadValid && !io.rabCommits.isWalk &&  dispatchCanAcc && !intFreeList.io.canAllocate
516  private val stallForFpFL      = inHeadValid && !io.rabCommits.isWalk &&  dispatchCanAcc &&  intFreeList.io.canAllocate && vecFreeList.io.canAllocate && !fpFreeList.io.canAllocate
517  private val stallForVecFL     = inHeadValid && !io.rabCommits.isWalk &&  dispatchCanAcc &&  intFreeList.io.canAllocate && fpFreeList.io.canAllocate && !vecFreeList.io.canAllocate
518  XSPerfAccumulate("stall_cycle",          inHeadStall)
519  XSPerfAccumulate("stall_cycle_walk",     stallForWalk)
520  XSPerfAccumulate("stall_cycle_dispatch", stallForDispatch)
521  XSPerfAccumulate("stall_cycle_int",      stallForIntFL)
522  XSPerfAccumulate("stall_cycle_fp",       stallForFpFL)
523  XSPerfAccumulate("stall_cycle_vec",      stallForVecFL)
524
525  XSPerfHistogram("in_valid_range",  PopCount(io.in.map(_.valid)),  true.B, 0, DecodeWidth + 1, 1)
526  XSPerfHistogram("in_fire_range",   PopCount(io.in.map(_.fire)),   true.B, 0, DecodeWidth + 1, 1)
527  XSPerfHistogram("out_valid_range", PopCount(io.out.map(_.valid)), true.B, 0, DecodeWidth + 1, 1)
528  XSPerfHistogram("out_fire_range",  PopCount(io.out.map(_.fire)),  true.B, 0, DecodeWidth + 1, 1)
529
530  XSPerfAccumulate("move_instr_count", PopCount(io.out.map(out => out.fire && out.bits.isMove)))
531  val is_fused_lui_load = io.out.map(o => o.fire && o.bits.fuType === FuType.ldu.U && o.bits.srcType(0) === SrcType.imm)
532  XSPerfAccumulate("fused_lui_load_instr_count", PopCount(is_fused_lui_load))
533
534  val renamePerf = Seq(
535    ("rename_in                  ", PopCount(io.in.map(_.valid & io.in(0).ready ))                                                               ),
536    ("rename_waitinstr           ", PopCount((0 until RenameWidth).map(i => io.in(i).valid && !io.in(i).ready))                                  ),
537    ("rename_stall               ", inHeadStall),
538    ("rename_stall_cycle_walk    ", inHeadValid &&  io.rabCommits.isWalk),
539    ("rename_stall_cycle_dispatch", inHeadValid && !io.rabCommits.isWalk && !dispatchCanAcc),
540    ("rename_stall_cycle_int     ", inHeadValid && !io.rabCommits.isWalk &&  dispatchCanAcc && !intFreeList.io.canAllocate),
541    ("rename_stall_cycle_fp      ", inHeadValid && !io.rabCommits.isWalk &&  dispatchCanAcc && intFreeList.io.canAllocate && vecFreeList.io.canAllocate && !fpFreeList.io.canAllocate),
542    ("rename_stall_cycle_vec     ", inHeadValid && !io.rabCommits.isWalk &&  dispatchCanAcc && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && !vecFreeList.io.canAllocate),
543  )
544  val intFlPerf = intFreeList.getPerfEvents
545  val fpFlPerf = fpFreeList.getPerfEvents
546  val vecFlPerf = vecFreeList.getPerfEvents
547  val perfEvents = renamePerf ++ intFlPerf ++ fpFlPerf ++ vecFlPerf
548  generatePerfEvent()
549}
550