xref: /XiangShan/src/main/scala/xiangshan/backend/rename/Rename.scala (revision 4daa5bf3c3f27e7fd090866d52405b21e107eb8d)
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(FpPhyRegs - FpLogicRegs, FpLogicRegs, Reg_F))
85  val vecFreeList = Module(new StdFreeList(VfPhyRegs - VecLogicRegs, VecLogicRegs, Reg_V))
86
87  intFreeList.io.commit    <> io.rabCommits
88  intFreeList.io.debug_rat.foreach(_ <> io.debug_int_rat.get)
89  fpFreeList.io.commit     <> io.rabCommits
90  fpFreeList.io.debug_rat.foreach(_ <> io.debug_fp_rat.get)
91  vecFreeList.io.commit    <> io.rabCommits
92  vecFreeList.io.debug_rat.foreach(_ <> io.debug_vec_rat.get)
93
94  // decide if given instruction needs allocating a new physical register (CfCtrl: from decode; RobCommitInfo: from rob)
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 <: RabCommitInfo](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 <: RabCommitInfo](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 fp & vec & int free list
116  Seq(fpFreeList, vecFreeList, intFreeList).foreach { case fl =>
117    fl.io.redirect := io.redirect.valid
118    fl.io.walk := io.rabCommits.isWalk
119  }
120  // only when all 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 && vecFreeList.io.canAllocate && io.out(0).ready || io.rabCommits.isWalk
123  fpFreeList.io.doAllocate := intFreeList.io.canAllocate && vecFreeList.io.canAllocate && io.out(0).ready || io.rabCommits.isWalk
124  vecFreeList.io.doAllocate := intFreeList.io.canAllocate && fpFreeList.io.canAllocate && io.out(0).ready || io.rabCommits.isWalk
125
126  //           dispatch1 ready ++ float point free list ready ++ int free list ready ++ vec free list ready     ++ not walk
127  val canOut = io.out(0).ready && fpFreeList.io.canAllocate && intFreeList.io.canAllocate && vecFreeList.io.canAllocate && !io.rabCommits.isWalk
128
129  compressUnit.io.in.zip(io.in).foreach{ case(sink, source) =>
130    sink.valid := source.valid
131    sink.bits := source.bits
132  }
133  val needRobFlags = compressUnit.io.out.needRobFlags
134  val instrSizesVec = compressUnit.io.out.instrSizes
135  val compressMasksVec = compressUnit.io.out.masks
136
137  // speculatively assign the instruction with an robIdx
138  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)
139  val robIdxHead = RegInit(0.U.asTypeOf(new RobPtr))
140  val lastCycleMisprediction = GatedValidRegNext(io.redirect.valid && !io.redirect.bits.flushItself())
141  val robIdxHeadNext = Mux(io.redirect.valid, io.redirect.bits.robIdx, // redirect: move ptr to given rob index
142         Mux(lastCycleMisprediction, robIdxHead + 1.U, // mis-predict: not flush robIdx itself
143           Mux(canOut, robIdxHead + validCount, // instructions successfully entered next stage: increase robIdx
144                      /* default */  robIdxHead))) // no instructions passed by this cycle: stick to old value
145  robIdxHead := robIdxHeadNext
146
147  /**
148    * Rename: allocate free physical register and update rename table
149    */
150  val uops = Wire(Vec(RenameWidth, new DynInst))
151  uops.foreach( uop => {
152    uop.srcState      := DontCare
153    uop.debugInfo     := DontCare
154    uop.lqIdx         := DontCare
155    uop.sqIdx         := DontCare
156    uop.waitForRobIdx := DontCare
157    uop.singleStep    := DontCare
158    uop.snapshot      := DontCare
159    uop.srcLoadDependency := DontCare
160  })
161
162  val needVecDest    = Wire(Vec(RenameWidth, Bool()))
163  val needFpDest     = Wire(Vec(RenameWidth, Bool()))
164  val needIntDest    = Wire(Vec(RenameWidth, Bool()))
165  val hasValid = Cat(io.in.map(_.valid)).orR
166  private val inHeadValid = io.in.head.valid
167
168  val isMove = Wire(Vec(RenameWidth, Bool()))
169  isMove zip io.in.map(_.bits) foreach {
170    case (move, in) => move := Mux(in.exceptionVec.asUInt.orR, false.B, in.isMove)
171  }
172
173  val walkNeedIntDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
174  val walkNeedFpDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
175  val walkNeedVecDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
176  val walkIsMove = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
177
178  val intSpecWen = Wire(Vec(RenameWidth, Bool()))
179  val fpSpecWen  = Wire(Vec(RenameWidth, Bool()))
180  val vecSpecWen = Wire(Vec(RenameWidth, Bool()))
181
182  val walkIntSpecWen = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
183
184  val walkPdest = Wire(Vec(RenameWidth, UInt(PhyRegIdxWidth.W)))
185
186  // uop calculation
187  for (i <- 0 until RenameWidth) {
188    (uops(i): Data).waiveAll :<= (io.in(i).bits: Data).waiveAll
189
190    // update cf according to ssit result
191    uops(i).storeSetHit := io.ssit(i).valid
192    uops(i).loadWaitStrict := io.ssit(i).strict && io.ssit(i).valid
193    uops(i).ssid := io.ssit(i).ssid
194
195    // update cf according to waittable result
196    uops(i).loadWaitBit := io.waittable(i)
197
198    uops(i).replayInst := false.B // set by IQ or MemQ
199    // alloc a new phy reg
200    needVecDest(i) := io.in(i).valid && needDestReg(Reg_V, io.in(i).bits)
201    needFpDest(i) := io.in(i).valid && needDestReg(Reg_F, io.in(i).bits)
202    needIntDest(i) := io.in(i).valid && needDestReg(Reg_I, io.in(i).bits)
203    if (i < RabCommitWidth) {
204      walkNeedIntDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_I, io.rabCommits.info(i))
205      walkNeedFpDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_F, io.rabCommits.info(i))
206      walkNeedVecDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_V, io.rabCommits.info(i))
207      walkIsMove(i) := io.rabCommits.info(i).isMove
208    }
209    fpFreeList.io.allocateReq(i) := needFpDest(i)
210    fpFreeList.io.walkReq(i) := walkNeedFpDest(i)
211    vecFreeList.io.allocateReq(i) := needVecDest(i)
212    vecFreeList.io.walkReq(i) := walkNeedVecDest(i)
213    intFreeList.io.allocateReq(i) := needIntDest(i) && !isMove(i)
214    intFreeList.io.walkReq(i) := walkNeedIntDest(i) && !walkIsMove(i)
215
216    // no valid instruction from decode stage || all resources (dispatch1 + both free lists) ready
217    io.in(i).ready := !hasValid || canOut
218
219    uops(i).robIdx := robIdxHead + PopCount(io.in.zip(needRobFlags).take(i).map{ case(in, needRobFlag) => in.valid && in.bits.lastUop && needRobFlag})
220    uops(i).instrSize := instrSizesVec(i)
221    when(isMove(i)) {
222      uops(i).numUops := 0.U
223      uops(i).numWB := 0.U
224    }
225    if (i > 0) {
226      when(!needRobFlags(i - 1)) {
227        uops(i).firstUop := false.B
228        uops(i).ftqPtr := uops(i - 1).ftqPtr
229        uops(i).ftqOffset := uops(i - 1).ftqOffset
230        uops(i).numUops := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
231        uops(i).numWB := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
232      }
233    }
234    when(!needRobFlags(i)) {
235      uops(i).lastUop := false.B
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    uops(i).wfflags := (compressMasksVec(i) & Cat(io.in.map(_.bits.wfflags).reverse)).orR
240    uops(i).dirtyFs := (compressMasksVec(i) & Cat(io.in.map(_.bits.fpWen).reverse)).orR
241    // vector instructions' uopSplitType cannot be UopSplitType.SCA_SIM
242    uops(i).dirtyVs := (compressMasksVec(i) & Cat(io.in.map(_.bits.uopSplitType =/= UopSplitType.SCA_SIM).reverse)).orR
243
244    uops(i).psrc(0) := Mux1H(uops(i).srcType(0), Seq(io.intReadPorts(i)(0), io.fpReadPorts(i)(0), io.vecReadPorts(i)(0)))
245    uops(i).psrc(1) := Mux1H(uops(i).srcType(1), Seq(io.intReadPorts(i)(1), io.fpReadPorts(i)(1), io.vecReadPorts(i)(1)))
246    uops(i).psrc(2) := Mux1H(uops(i).srcType(2)(2, 1), Seq(io.fpReadPorts(i)(2), io.vecReadPorts(i)(2)))
247    uops(i).psrc(3) := io.vecReadPorts(i)(3)
248    uops(i).psrc(4) := io.vecReadPorts(i)(4) // Todo: vl read port
249
250    // int psrc2 should be bypassed from next instruction if it is fused
251    if (i < RenameWidth - 1) {
252      when (io.fusionInfo(i).rs2FromRs2 || io.fusionInfo(i).rs2FromRs1) {
253        uops(i).psrc(1) := Mux(io.fusionInfo(i).rs2FromRs2, io.intReadPorts(i + 1)(1), io.intReadPorts(i + 1)(0))
254      }.elsewhen(io.fusionInfo(i).rs2FromZero) {
255        uops(i).psrc(1) := 0.U
256      }
257    }
258    uops(i).eliminatedMove := isMove(i)
259
260    // update pdest
261    uops(i).pdest := MuxCase(0.U, Seq(
262      needIntDest(i)    ->  intFreeList.io.allocatePhyReg(i),
263      needFpDest(i)     ->  fpFreeList.io.allocatePhyReg(i),
264      needVecDest(i)    ->  vecFreeList.io.allocatePhyReg(i),
265    ))
266
267    // Assign performance counters
268    uops(i).debugInfo.renameTime := GTimer()
269
270    io.out(i).valid := io.in(i).valid && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && !io.rabCommits.isWalk
271    io.out(i).bits := uops(i)
272    // Todo: move these shit in decode stage
273    // dirty code for fence. The lsrc is passed by imm.
274    when (io.out(i).bits.fuType === FuType.fence.U) {
275      io.out(i).bits.imm := Cat(io.in(i).bits.lsrc(1), io.in(i).bits.lsrc(0))
276    }
277
278    // dirty code for SoftPrefetch (prefetch.r/prefetch.w)
279//    when (io.in(i).bits.isSoftPrefetch) {
280//      io.out(i).bits.fuType := FuType.ldu.U
281//      io.out(i).bits.fuOpType := Mux(io.in(i).bits.lsrc(1) === 1.U, LSUOpType.prefetch_r, LSUOpType.prefetch_w)
282//      io.out(i).bits.selImm := SelImm.IMM_S
283//      io.out(i).bits.imm := Cat(io.in(i).bits.imm(io.in(i).bits.imm.getWidth - 1, 5), 0.U(5.W))
284//    }
285
286    // dirty code for lui+addi(w) fusion
287    if (i < RenameWidth - 1) {
288      val fused_lui32 = io.in(i).bits.selImm === SelImm.IMM_LUI32 && io.in(i).bits.fuType === FuType.alu.U
289      when (fused_lui32) {
290        val lui_imm = io.in(i).bits.imm(19, 0)
291        val add_imm = io.in(i + 1).bits.imm(11, 0)
292        require(io.out(i).bits.imm.getWidth >= lui_imm.getWidth + add_imm.getWidth)
293        io.out(i).bits.imm := Cat(lui_imm, add_imm)
294      }
295    }
296
297    // write speculative rename table
298    // we update rat later inside commit code
299    intSpecWen(i) := needIntDest(i) && intFreeList.io.canAllocate && intFreeList.io.doAllocate && !io.rabCommits.isWalk && !io.redirect.valid
300    fpSpecWen(i)  := needFpDest(i)  && fpFreeList.io.canAllocate  && fpFreeList.io.doAllocate  && !io.rabCommits.isWalk && !io.redirect.valid
301    vecSpecWen(i) := needVecDest(i) && vecFreeList.io.canAllocate && vecFreeList.io.doAllocate && !io.rabCommits.isWalk && !io.redirect.valid
302
303
304    if (i < RabCommitWidth) {
305      walkIntSpecWen(i) := walkNeedIntDest(i) && !io.redirect.valid
306      walkPdest(i) := io.rabCommits.info(i).pdest
307    } else {
308      walkPdest(i) := io.out(i).bits.pdest
309    }
310  }
311
312  /**
313    * How to set psrc:
314    * - bypass the pdest to psrc if previous instructions write to the same ldest as lsrc
315    * - default: psrc from RAT
316    * How to set pdest:
317    * - Mux(isMove, psrc, pdest_from_freelist).
318    *
319    * The critical path of rename lies here:
320    * When move elimination is enabled, we need to update the rat with psrc.
321    * However, psrc maybe comes from previous instructions' pdest, which comes from freelist.
322    *
323    * If we expand these logic for pdest(N):
324    * pdest(N) = Mux(isMove(N), psrc(N), freelist_out(N))
325    *          = Mux(isMove(N), Mux(bypass(N, N - 1), pdest(N - 1),
326    *                           Mux(bypass(N, N - 2), pdest(N - 2),
327    *                           ...
328    *                           Mux(bypass(N, 0),     pdest(0),
329    *                                                 rat_out(N))...)),
330    *                           freelist_out(N))
331    */
332  // a simple functional model for now
333  io.out(0).bits.pdest := Mux(isMove(0), uops(0).psrc.head, uops(0).pdest)
334
335  // psrc(n) + pdest(1)
336  val bypassCond: Vec[MixedVec[UInt]] = Wire(Vec(numRegSrc + 1, MixedVec(List.tabulate(RenameWidth-1)(i => UInt((i+1).W)))))
337  require(io.in(0).bits.srcType.size == io.in(0).bits.numSrc)
338  private val pdestLoc = io.in.head.bits.srcType.size // 2 vector src: v0, vl&vtype
339  println(s"[Rename] idx of pdest in bypassCond $pdestLoc")
340  for (i <- 1 until RenameWidth) {
341    val vecCond = io.in(i).bits.srcType.map(_ === SrcType.vp) :+ needVecDest(i)
342    val fpCond  = io.in(i).bits.srcType.map(_ === SrcType.fp) :+ needFpDest(i)
343    val intCond = io.in(i).bits.srcType.map(_ === SrcType.xp) :+ needIntDest(i)
344    val target = io.in(i).bits.lsrc :+ io.in(i).bits.ldest
345    for (((((cond1, cond2), cond3), t), j) <- vecCond.zip(fpCond).zip(intCond).zip(target).zipWithIndex) {
346      val destToSrc = io.in.take(i).zipWithIndex.map { case (in, j) =>
347        val indexMatch = in.bits.ldest === t
348        val writeMatch =  cond3 && needIntDest(j) || cond2 && needFpDest(j) || cond1 && needVecDest(j)
349        indexMatch && writeMatch
350      }
351      bypassCond(j)(i - 1) := VecInit(destToSrc).asUInt
352    }
353    io.out(i).bits.psrc(0) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(0)(i-1).asBools).foldLeft(uops(i).psrc(0)) {
354      (z, next) => Mux(next._2, next._1, z)
355    }
356    io.out(i).bits.psrc(1) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(1)(i-1).asBools).foldLeft(uops(i).psrc(1)) {
357      (z, next) => Mux(next._2, next._1, z)
358    }
359    io.out(i).bits.psrc(2) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(2)(i-1).asBools).foldLeft(uops(i).psrc(2)) {
360      (z, next) => Mux(next._2, next._1, z)
361    }
362    io.out(i).bits.psrc(3) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(3)(i-1).asBools).foldLeft(uops(i).psrc(3)) {
363      (z, next) => Mux(next._2, next._1, z)
364    }
365    io.out(i).bits.psrc(4) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(4)(i-1).asBools).foldLeft(uops(i).psrc(4)) {
366      (z, next) => Mux(next._2, next._1, z)
367    }
368    io.out(i).bits.pdest := Mux(isMove(i), io.out(i).bits.psrc(0), uops(i).pdest)
369
370    // Todo: better implementation for fields reuse
371    // For fused-lui-load, load.src(0) is replaced by the imm.
372    val last_is_lui = io.in(i - 1).bits.selImm === SelImm.IMM_U && io.in(i - 1).bits.srcType(0) =/= SrcType.pc
373    val this_is_load = io.in(i).bits.fuType === FuType.ldu.U
374    val lui_to_load = io.in(i - 1).valid && io.in(i - 1).bits.ldest === io.in(i).bits.lsrc(0)
375    val fused_lui_load = last_is_lui && this_is_load && lui_to_load
376    when (fused_lui_load) {
377      // The first LOAD operand (base address) is replaced by LUI-imm and stored in imm
378      val lui_imm = io.in(i - 1).bits.imm(ImmUnion.U.len - 1, 0)
379      val ld_imm = io.in(i).bits.imm(ImmUnion.I.len - 1, 0)
380      require(io.out(i).bits.imm.getWidth >= lui_imm.getWidth + ld_imm.getWidth)
381      io.out(i).bits.srcType(0) := SrcType.imm
382      io.out(i).bits.imm := Cat(lui_imm, ld_imm)
383    }
384
385  }
386
387  val genSnapshot = Cat(io.out.map(out => out.fire && out.bits.snapshot)).orR
388  val lastCycleCreateSnpt = RegInit(false.B)
389  lastCycleCreateSnpt := genSnapshot && !io.snptIsFull
390  val sameSnptDistance = (RobCommitWidth * 4).U
391  // notInSameSnpt: 1.robidxHead - snapLastEnq >= sameSnptDistance 2.no snap
392  val notInSameSnpt = GatedValidRegNext(distanceBetween(robIdxHeadNext, io.snptLastEnq.bits) >= sameSnptDistance || !io.snptLastEnq.valid)
393  val allowSnpt = if (EnableRenameSnapshot) notInSameSnpt && !lastCycleCreateSnpt && io.in.head.bits.firstUop else false.B
394  io.out.zip(io.in).foreach{ case (out, in) => out.bits.snapshot := allowSnpt && (!in.bits.preDecodeInfo.notCFI || FuType.isJump(in.bits.fuType)) && in.fire }
395  if(backendParams.debugEn){
396    dontTouch(robIdxHeadNext)
397    dontTouch(notInSameSnpt)
398    dontTouch(genSnapshot)
399  }
400  intFreeList.io.snpt := io.snpt
401  fpFreeList.io.snpt := io.snpt
402  vecFreeList.io.snpt := io.snpt
403  intFreeList.io.snpt.snptEnq := genSnapshot
404  fpFreeList.io.snpt.snptEnq := genSnapshot
405  vecFreeList.io.snpt.snptEnq := genSnapshot
406
407  /**
408    * Instructions commit: update freelist and rename table
409    */
410  for (i <- 0 until RabCommitWidth) {
411    val commitValid = io.rabCommits.isCommit && io.rabCommits.commitValid(i)
412    val walkValid = io.rabCommits.isWalk && io.rabCommits.walkValid(i)
413
414    // I. RAT Update
415    // When redirect happens (mis-prediction), don't update the rename table
416    io.intRenamePorts(i).wen  := intSpecWen(i)
417    io.intRenamePorts(i).addr := uops(i).ldest
418    io.intRenamePorts(i).data := io.out(i).bits.pdest
419
420    io.fpRenamePorts(i).wen  := fpSpecWen(i)
421    io.fpRenamePorts(i).addr := uops(i).ldest
422    io.fpRenamePorts(i).data := fpFreeList.io.allocatePhyReg(i)
423
424    io.vecRenamePorts(i).wen := vecSpecWen(i)
425    io.vecRenamePorts(i).addr := uops(i).ldest
426    io.vecRenamePorts(i).data := vecFreeList.io.allocatePhyReg(i)
427
428    // II. Free List Update
429    intFreeList.io.freeReq(i) := io.int_need_free(i)
430    intFreeList.io.freePhyReg(i) := RegNext(io.int_old_pdest(i))
431    fpFreeList.io.freeReq(i)  := GatedValidRegNext(commitValid && needDestRegCommit(Reg_F, io.rabCommits.info(i)))
432    fpFreeList.io.freePhyReg(i) := io.fp_old_pdest(i)
433    vecFreeList.io.freeReq(i)  := GatedValidRegNext(commitValid && needDestRegCommit(Reg_V, io.rabCommits.info(i)))
434    vecFreeList.io.freePhyReg(i) := io.vec_old_pdest(i)
435  }
436
437  /*
438  Debug and performance counters
439   */
440  def printRenameInfo(in: DecoupledIO[DecodedInst], out: DecoupledIO[DynInst]) = {
441    XSInfo(out.fire, p"pc:${Hexadecimal(in.bits.pc)} in(${in.valid},${in.ready}) " +
442      p"lsrc(0):${in.bits.lsrc(0)} -> psrc(0):${out.bits.psrc(0)} " +
443      p"lsrc(1):${in.bits.lsrc(1)} -> psrc(1):${out.bits.psrc(1)} " +
444      p"lsrc(2):${in.bits.lsrc(2)} -> psrc(2):${out.bits.psrc(2)} " +
445      p"ldest:${in.bits.ldest} -> pdest:${out.bits.pdest}\n"
446    )
447  }
448
449  for ((x,y) <- io.in.zip(io.out)) {
450    printRenameInfo(x, y)
451  }
452
453  val debugRedirect = RegEnable(io.redirect.bits, io.redirect.valid)
454  // bad speculation
455  val recStall = io.redirect.valid || io.rabCommits.isWalk
456  val ctrlRecStall = Mux(io.redirect.valid, io.redirect.bits.debugIsCtrl, io.rabCommits.isWalk && debugRedirect.debugIsCtrl)
457  val mvioRecStall = Mux(io.redirect.valid, io.redirect.bits.debugIsMemVio, io.rabCommits.isWalk && debugRedirect.debugIsMemVio)
458  val otherRecStall = recStall && !(ctrlRecStall || mvioRecStall)
459  XSPerfAccumulate("recovery_stall", recStall)
460  XSPerfAccumulate("control_recovery_stall", ctrlRecStall)
461  XSPerfAccumulate("mem_violation_recovery_stall", mvioRecStall)
462  XSPerfAccumulate("other_recovery_stall", otherRecStall)
463  // freelist stall
464  val notRecStall = !io.out.head.valid && !recStall
465  val intFlStall = notRecStall && inHeadValid && !intFreeList.io.canAllocate
466  val fpFlStall = notRecStall && inHeadValid && intFreeList.io.canAllocate && vecFreeList.io.canAllocate && !fpFreeList.io.canAllocate
467  val vecFlStall = notRecStall && inHeadValid && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && !vecFreeList.io.canAllocate
468  // other stall
469  val otherStall = notRecStall && !intFlStall && !fpFlStall && !vecFlStall
470
471  io.stallReason.in.backReason.valid := io.stallReason.out.backReason.valid || !io.in.head.ready
472  io.stallReason.in.backReason.bits := Mux(io.stallReason.out.backReason.valid, io.stallReason.out.backReason.bits,
473    MuxCase(TopDownCounters.OtherCoreStall.id.U, Seq(
474      ctrlRecStall  -> TopDownCounters.ControlRecoveryStall.id.U,
475      mvioRecStall  -> TopDownCounters.MemVioRecoveryStall.id.U,
476      otherRecStall -> TopDownCounters.OtherRecoveryStall.id.U,
477      intFlStall    -> TopDownCounters.IntFlStall.id.U,
478      fpFlStall     -> TopDownCounters.FpFlStall.id.U,
479      vecFlStall    -> TopDownCounters.VecFlStall.id.U,
480    )
481  ))
482  io.stallReason.out.reason.zip(io.stallReason.in.reason).zip(io.in.map(_.valid)).foreach { case ((out, in), valid) =>
483    out := Mux(io.stallReason.in.backReason.valid, io.stallReason.in.backReason.bits, in)
484  }
485
486  XSDebug(io.rabCommits.isWalk, p"Walk Recovery Enabled\n")
487  XSDebug(io.rabCommits.isWalk, p"validVec:${Binary(io.rabCommits.walkValid.asUInt)}\n")
488  for (i <- 0 until RabCommitWidth) {
489    val info = io.rabCommits.info(i)
490    XSDebug(io.rabCommits.isWalk && io.rabCommits.walkValid(i), p"[#$i walk info] " +
491      p"ldest:${info.ldest} rfWen:${info.rfWen} fpWen:${info.fpWen} vecWen:${info.vecWen}")
492  }
493
494  XSDebug(p"inValidVec: ${Binary(Cat(io.in.map(_.valid)))}\n")
495
496  XSPerfAccumulate("in_valid_count", PopCount(io.in.map(_.valid)))
497  XSPerfAccumulate("in_fire_count", PopCount(io.in.map(_.fire)))
498  XSPerfAccumulate("in_valid_not_ready_count", PopCount(io.in.map(x => x.valid && !x.ready)))
499  XSPerfAccumulate("wait_cycle", !io.in.head.valid && io.out.head.ready)
500
501  // These stall reasons could overlap each other, but we configure the priority as fellows.
502  // walk stall > dispatch stall > int freelist stall > fp freelist stall
503  private val inHeadStall = io.in.head match { case x => x.valid && !x.ready }
504  private val stallForWalk      = inHeadValid &&  io.rabCommits.isWalk
505  private val stallForDispatch  = inHeadValid && !io.rabCommits.isWalk && !io.out(0).ready
506  private val stallForIntFL     = inHeadValid && !io.rabCommits.isWalk &&  io.out(0).ready && !intFreeList.io.canAllocate
507  private val stallForFpFL      = inHeadValid && !io.rabCommits.isWalk &&  io.out(0).ready &&  intFreeList.io.canAllocate && vecFreeList.io.canAllocate && !fpFreeList.io.canAllocate
508  private val stallForVecFL     = inHeadValid && !io.rabCommits.isWalk &&  io.out(0).ready &&  intFreeList.io.canAllocate && fpFreeList.io.canAllocate && !vecFreeList.io.canAllocate
509  XSPerfAccumulate("stall_cycle",          inHeadStall)
510  XSPerfAccumulate("stall_cycle_walk",     stallForWalk)
511  XSPerfAccumulate("stall_cycle_dispatch", stallForDispatch)
512  XSPerfAccumulate("stall_cycle_int",      stallForIntFL)
513  XSPerfAccumulate("stall_cycle_fp",       stallForFpFL)
514  XSPerfAccumulate("stall_cycle_vec",      stallForVecFL)
515
516  XSPerfHistogram("in_valid_range",  PopCount(io.in.map(_.valid)),  true.B, 0, DecodeWidth + 1, 1)
517  XSPerfHistogram("in_fire_range",   PopCount(io.in.map(_.fire)),   true.B, 0, DecodeWidth + 1, 1)
518  XSPerfHistogram("out_valid_range", PopCount(io.out.map(_.valid)), true.B, 0, DecodeWidth + 1, 1)
519  XSPerfHistogram("out_fire_range",  PopCount(io.out.map(_.fire)),  true.B, 0, DecodeWidth + 1, 1)
520
521  XSPerfAccumulate("move_instr_count", PopCount(io.out.map(out => out.fire && out.bits.isMove)))
522  val is_fused_lui_load = io.out.map(o => o.fire && o.bits.fuType === FuType.ldu.U && o.bits.srcType(0) === SrcType.imm)
523  XSPerfAccumulate("fused_lui_load_instr_count", PopCount(is_fused_lui_load))
524
525  val renamePerf = Seq(
526    ("rename_in                  ", PopCount(io.in.map(_.valid & io.in(0).ready ))                                                               ),
527    ("rename_waitinstr           ", PopCount((0 until RenameWidth).map(i => io.in(i).valid && !io.in(i).ready))                                  ),
528    ("rename_stall               ", inHeadStall),
529    ("rename_stall_cycle_walk    ", inHeadValid &&  io.rabCommits.isWalk),
530    ("rename_stall_cycle_dispatch", inHeadValid && !io.rabCommits.isWalk && !io.out(0).ready),
531    ("rename_stall_cycle_int     ", inHeadValid && !io.rabCommits.isWalk &&  io.out(0).ready && !intFreeList.io.canAllocate),
532    ("rename_stall_cycle_fp      ", inHeadValid && !io.rabCommits.isWalk &&  io.out(0).ready && intFreeList.io.canAllocate && vecFreeList.io.canAllocate && !fpFreeList.io.canAllocate),
533    ("rename_stall_cycle_vec     ", inHeadValid && !io.rabCommits.isWalk &&  io.out(0).ready && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && !vecFreeList.io.canAllocate),
534  )
535  val intFlPerf = intFreeList.getPerfEvents
536  val fpFlPerf = fpFreeList.getPerfEvents
537  val vecFlPerf = vecFreeList.getPerfEvents
538  val perfEvents = renamePerf ++ intFlPerf ++ fpFlPerf ++ vecFlPerf
539  generatePerfEvent()
540}
541