xref: /XiangShan/src/main/scala/xiangshan/backend/rename/Rename.scala (revision 3bec463ea7e6217896b6ae73531cd2ed8e8b8d92)
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 v0ReadPorts = Vec(RenameWidth, Vec(1, Input(UInt(PhyRegIdxWidth.W))))
57    val vlReadPorts = Vec(RenameWidth, Vec(1, Input(UInt(PhyRegIdxWidth.W))))
58    val intRenamePorts = Vec(RenameWidth, Output(new RatWritePort))
59    val fpRenamePorts = Vec(RenameWidth, Output(new RatWritePort))
60    val vecRenamePorts = Vec(RenameWidth, Output(new RatWritePort))
61    val v0RenamePorts = Vec(RenameWidth, Output(new RatWritePort))
62    val vlRenamePorts = Vec(RenameWidth, Output(new RatWritePort))
63    // from rename table
64    val int_old_pdest = Vec(RabCommitWidth, Input(UInt(PhyRegIdxWidth.W)))
65    val fp_old_pdest = Vec(RabCommitWidth, Input(UInt(PhyRegIdxWidth.W)))
66    val vec_old_pdest = Vec(RabCommitWidth, Input(UInt(PhyRegIdxWidth.W)))
67    val v0_old_pdest = Vec(RabCommitWidth, Input(UInt(PhyRegIdxWidth.W)))
68    val vl_old_pdest = Vec(RabCommitWidth, Input(UInt(PhyRegIdxWidth.W)))
69    val int_need_free = Vec(RabCommitWidth, Input(Bool()))
70    // to dispatch1
71    val out = Vec(RenameWidth, DecoupledIO(new DynInst))
72    // for snapshots
73    val snpt = Input(new SnapshotPort)
74    val snptLastEnq = Flipped(ValidIO(new RobPtr))
75    val snptIsFull= Input(Bool())
76    // debug arch ports
77    val debug_int_rat = if (backendParams.debugEn) Some(Vec(32, Input(UInt(PhyRegIdxWidth.W)))) else None
78    val debug_fp_rat = if (backendParams.debugEn) Some(Vec(32, Input(UInt(PhyRegIdxWidth.W)))) else None
79    val debug_vec_rat = if (backendParams.debugEn) Some(Vec(31, Input(UInt(PhyRegIdxWidth.W)))) else None
80    val debug_v0_rat = if (backendParams.debugEn) Some(Vec(1, Input(UInt(PhyRegIdxWidth.W)))) else None
81    val debug_vl_rat = if (backendParams.debugEn) Some(Vec(1, Input(UInt(PhyRegIdxWidth.W)))) else None
82    // perf only
83    val stallReason = new Bundle {
84      val in = Flipped(new StallReasonIO(RenameWidth))
85      val out = new StallReasonIO(RenameWidth)
86    }
87  })
88
89  // io alias
90  private val dispatchCanAcc = io.out.head.ready
91
92  val compressUnit = Module(new CompressUnit())
93  // create free list and rat
94  val intFreeList = Module(new MEFreeList(IntPhyRegs))
95  val fpFreeList = Module(new StdFreeList(FpPhyRegs - FpLogicRegs, FpLogicRegs, Reg_F))
96  val vecFreeList = Module(new StdFreeList(VfPhyRegs - VecLogicRegs, VecLogicRegs, Reg_V, 31))
97  val v0FreeList = Module(new StdFreeList(V0PhyRegs - V0LogicRegs, V0LogicRegs, Reg_V0, 1))
98  val vlFreeList = Module(new StdFreeList(VlPhyRegs - VlLogicRegs, VlLogicRegs, Reg_Vl, 1))
99
100
101  intFreeList.io.commit    <> io.rabCommits
102  intFreeList.io.debug_rat.foreach(_ <> io.debug_int_rat.get)
103  fpFreeList.io.commit     <> io.rabCommits
104  fpFreeList.io.debug_rat.foreach(_ <> io.debug_fp_rat.get)
105  vecFreeList.io.commit    <> io.rabCommits
106  vecFreeList.io.debug_rat.foreach(_ <> io.debug_vec_rat.get)
107  v0FreeList.io.commit <> io.rabCommits
108  v0FreeList.io.debug_rat.foreach(_ <> io.debug_v0_rat.get)
109  vlFreeList.io.commit <> io.rabCommits
110  vlFreeList.io.debug_rat.foreach(_ <> io.debug_vl_rat.get)
111
112  // decide if given instruction needs allocating a new physical register (CfCtrl: from decode; RobCommitInfo: from rob)
113  def needDestReg[T <: DecodedInst](reg_t: RegType, x: T): Bool = reg_t match {
114    case Reg_I => x.rfWen && x.ldest =/= 0.U
115    case Reg_F => x.fpWen
116    case Reg_V => x.vecWen
117    case Reg_V0 => x.v0Wen
118    case Reg_Vl => x.vlWen
119  }
120  def needDestRegCommit[T <: RabCommitInfo](reg_t: RegType, x: T): Bool = {
121    reg_t match {
122      case Reg_I => x.rfWen
123      case Reg_F => x.fpWen
124      case Reg_V => x.vecWen
125      case Reg_V0 => x.v0Wen
126      case Reg_Vl => x.vlWen
127    }
128  }
129  def needDestRegWalk[T <: RabCommitInfo](reg_t: RegType, x: T): Bool = {
130    reg_t match {
131      case Reg_I => x.rfWen && x.ldest =/= 0.U
132      case Reg_F => x.fpWen
133      case Reg_V => x.vecWen
134      case Reg_V0 => x.v0Wen
135      case Reg_Vl => x.vlWen
136    }
137  }
138
139  // connect [redirect + walk] ports for fp & vec & int free list
140  Seq(fpFreeList, vecFreeList, intFreeList, v0FreeList, vlFreeList).foreach { case fl =>
141    fl.io.redirect := io.redirect.valid
142    fl.io.walk := io.rabCommits.isWalk
143  }
144  // only when all free list and dispatch1 has enough space can we do allocation
145  // when isWalk, freelist can definitely allocate
146  intFreeList.io.doAllocate := fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && dispatchCanAcc || io.rabCommits.isWalk
147  fpFreeList.io.doAllocate := intFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && dispatchCanAcc || io.rabCommits.isWalk
148  vecFreeList.io.doAllocate := intFreeList.io.canAllocate && fpFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && dispatchCanAcc || io.rabCommits.isWalk
149  v0FreeList.io.doAllocate := intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && vlFreeList.io.canAllocate && dispatchCanAcc || io.rabCommits.isWalk
150  vlFreeList.io.doAllocate := intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && dispatchCanAcc || io.rabCommits.isWalk
151
152  //           dispatch1 ready ++ float point free list ready ++ int free list ready ++ vec free list ready     ++ not walk
153  val canOut = dispatchCanAcc && fpFreeList.io.canAllocate && intFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !io.rabCommits.isWalk
154
155  compressUnit.io.in.zip(io.in).foreach{ case(sink, source) =>
156    sink.valid := source.valid
157    sink.bits := source.bits
158  }
159  val needRobFlags = compressUnit.io.out.needRobFlags
160  val instrSizesVec = compressUnit.io.out.instrSizes
161  val compressMasksVec = compressUnit.io.out.masks
162
163  // speculatively assign the instruction with an robIdx
164  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)
165  val robIdxHead = RegInit(0.U.asTypeOf(new RobPtr))
166  val lastCycleMisprediction = GatedValidRegNext(io.redirect.valid && !io.redirect.bits.flushItself())
167  val robIdxHeadNext = Mux(io.redirect.valid, io.redirect.bits.robIdx, // redirect: move ptr to given rob index
168         Mux(lastCycleMisprediction, robIdxHead + 1.U, // mis-predict: not flush robIdx itself
169           Mux(canOut, robIdxHead + validCount, // instructions successfully entered next stage: increase robIdx
170                      /* default */  robIdxHead))) // no instructions passed by this cycle: stick to old value
171  robIdxHead := robIdxHeadNext
172
173  /**
174    * Rename: allocate free physical register and update rename table
175    */
176  val uops = Wire(Vec(RenameWidth, new DynInst))
177  uops.foreach( uop => {
178    uop.srcState      := DontCare
179    uop.debugInfo     := DontCare
180    uop.lqIdx         := DontCare
181    uop.sqIdx         := DontCare
182    uop.waitForRobIdx := DontCare
183    uop.singleStep    := DontCare
184    uop.snapshot      := DontCare
185    uop.srcLoadDependency := DontCare
186    uop.numLsElem       :=  DontCare
187    uop.hasException  :=  DontCare
188  })
189
190  val needVecDest    = Wire(Vec(RenameWidth, Bool()))
191  val needFpDest     = Wire(Vec(RenameWidth, Bool()))
192  val needIntDest    = Wire(Vec(RenameWidth, Bool()))
193  val needV0Dest     = Wire(Vec(RenameWidth, Bool()))
194  val needVlDest     = Wire(Vec(RenameWidth, Bool()))
195  val hasValid = Cat(io.in.map(_.valid)).orR
196  private val inHeadValid = io.in.head.valid
197
198  val isMove = Wire(Vec(RenameWidth, Bool()))
199  isMove zip io.in.map(_.bits) foreach {
200    case (move, in) => move := Mux(in.exceptionVec.asUInt.orR, false.B, in.isMove)
201  }
202
203  val walkNeedIntDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
204  val walkNeedFpDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
205  val walkNeedVecDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
206  val walkNeedV0Dest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
207  val walkNeedVlDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
208  val walkIsMove = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
209
210  val intSpecWen = Wire(Vec(RenameWidth, Bool()))
211  val fpSpecWen  = Wire(Vec(RenameWidth, Bool()))
212  val vecSpecWen = Wire(Vec(RenameWidth, Bool()))
213  val v0SpecWen = Wire(Vec(RenameWidth, Bool()))
214  val vlSpecWen = Wire(Vec(RenameWidth, Bool()))
215
216  val walkIntSpecWen = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B)))
217
218  val walkPdest = Wire(Vec(RenameWidth, UInt(PhyRegIdxWidth.W)))
219
220  // uop calculation
221  for (i <- 0 until RenameWidth) {
222    (uops(i): Data).waiveAll :<= (io.in(i).bits: Data).waiveAll
223
224    // update cf according to ssit result
225    uops(i).storeSetHit := io.ssit(i).valid
226    uops(i).loadWaitStrict := io.ssit(i).strict && io.ssit(i).valid
227    uops(i).ssid := io.ssit(i).ssid
228
229    // update cf according to waittable result
230    uops(i).loadWaitBit := io.waittable(i)
231
232    uops(i).replayInst := false.B // set by IQ or MemQ
233    // alloc a new phy reg
234    needV0Dest(i) := io.in(i).valid && needDestReg(Reg_V0, io.in(i).bits)
235    needVlDest(i) := io.in(i).valid && needDestReg(Reg_Vl, io.in(i).bits)
236    needVecDest(i) := io.in(i).valid && needDestReg(Reg_V, io.in(i).bits)
237    needFpDest(i) := io.in(i).valid && needDestReg(Reg_F, io.in(i).bits)
238    needIntDest(i) := io.in(i).valid && needDestReg(Reg_I, io.in(i).bits)
239    if (i < RabCommitWidth) {
240      walkNeedIntDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_I, io.rabCommits.info(i))
241      walkNeedFpDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_F, io.rabCommits.info(i))
242      walkNeedVecDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_V, io.rabCommits.info(i))
243      walkNeedV0Dest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_V0, io.rabCommits.info(i))
244      walkNeedVlDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_Vl, io.rabCommits.info(i))
245      walkIsMove(i) := io.rabCommits.info(i).isMove
246    }
247    fpFreeList.io.allocateReq(i) := needFpDest(i)
248    fpFreeList.io.walkReq(i) := walkNeedFpDest(i)
249    vecFreeList.io.allocateReq(i) := needVecDest(i)
250    vecFreeList.io.walkReq(i) := walkNeedVecDest(i)
251    v0FreeList.io.allocateReq(i) := needV0Dest(i)
252    v0FreeList.io.walkReq(i) := walkNeedV0Dest(i)
253    vlFreeList.io.allocateReq(i) := needVlDest(i)
254    vlFreeList.io.walkReq(i) := walkNeedVlDest(i)
255    intFreeList.io.allocateReq(i) := needIntDest(i) && !isMove(i)
256    intFreeList.io.walkReq(i) := walkNeedIntDest(i) && !walkIsMove(i)
257
258    // no valid instruction from decode stage || all resources (dispatch1 + both free lists) ready
259    io.in(i).ready := !hasValid || canOut
260
261    uops(i).robIdx := robIdxHead + PopCount(io.in.zip(needRobFlags).take(i).map{ case(in, needRobFlag) => in.valid && in.bits.lastUop && needRobFlag})
262    uops(i).instrSize := instrSizesVec(i)
263    when(isMove(i)) {
264      uops(i).numUops := 0.U
265      uops(i).numWB := 0.U
266    }
267    if (i > 0) {
268      when(!needRobFlags(i - 1)) {
269        uops(i).firstUop := false.B
270        uops(i).ftqPtr := uops(i - 1).ftqPtr
271        uops(i).ftqOffset := uops(i - 1).ftqOffset
272        uops(i).numUops := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
273        uops(i).numWB := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
274      }
275    }
276    when(!needRobFlags(i)) {
277      uops(i).lastUop := false.B
278      uops(i).numUops := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
279      uops(i).numWB := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse))
280    }
281    uops(i).wfflags := (compressMasksVec(i) & Cat(io.in.map(_.bits.wfflags).reverse)).orR
282    uops(i).dirtyFs := (compressMasksVec(i) & Cat(io.in.map(_.bits.fpWen).reverse)).orR
283    // vector instructions' uopSplitType cannot be UopSplitType.SCA_SIM
284    uops(i).dirtyVs := (compressMasksVec(i) & Cat(io.in.map(_.bits.uopSplitType =/= UopSplitType.SCA_SIM).reverse)).orR
285    // psrc0,psrc1,psrc2 don't require v0ReadPorts because their srcType can distinguish whether they are V0 or not
286    uops(i).psrc(0) := Mux1H(uops(i).srcType(0)(2, 0), Seq(io.intReadPorts(i)(0), io.fpReadPorts(i)(0), io.vecReadPorts(i)(0)))
287    uops(i).psrc(1) := Mux1H(uops(i).srcType(1)(2, 0), Seq(io.intReadPorts(i)(1), io.fpReadPorts(i)(1), io.vecReadPorts(i)(1)))
288    uops(i).psrc(2) := Mux1H(uops(i).srcType(2)(2, 1), Seq(io.fpReadPorts(i)(2), io.vecReadPorts(i)(2)))
289    uops(i).psrc(3) := io.v0ReadPorts(i)(0)
290    uops(i).psrc(4) := io.vlReadPorts(i)(0)
291
292    // int psrc2 should be bypassed from next instruction if it is fused
293    if (i < RenameWidth - 1) {
294      when (io.fusionInfo(i).rs2FromRs2 || io.fusionInfo(i).rs2FromRs1) {
295        uops(i).psrc(1) := Mux(io.fusionInfo(i).rs2FromRs2, io.intReadPorts(i + 1)(1), io.intReadPorts(i + 1)(0))
296      }.elsewhen(io.fusionInfo(i).rs2FromZero) {
297        uops(i).psrc(1) := 0.U
298      }
299    }
300    uops(i).eliminatedMove := isMove(i)
301
302    // update pdest
303    uops(i).pdest := MuxCase(0.U, Seq(
304      needIntDest(i)    ->  intFreeList.io.allocatePhyReg(i),
305      needFpDest(i)     ->  fpFreeList.io.allocatePhyReg(i),
306      needVecDest(i)    ->  vecFreeList.io.allocatePhyReg(i),
307      needV0Dest(i)    ->  v0FreeList.io.allocatePhyReg(i),
308      needVlDest(i)    ->  vlFreeList.io.allocatePhyReg(i),
309    ))
310
311    // Assign performance counters
312    uops(i).debugInfo.renameTime := GTimer()
313
314    io.out(i).valid := io.in(i).valid && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !io.rabCommits.isWalk
315    io.out(i).bits := uops(i)
316    // Todo: move these shit in decode stage
317    // dirty code for fence. The lsrc is passed by imm.
318    when (io.out(i).bits.fuType === FuType.fence.U) {
319      io.out(i).bits.imm := Cat(io.in(i).bits.lsrc(1), io.in(i).bits.lsrc(0))
320    }
321
322    // dirty code for SoftPrefetch (prefetch.r/prefetch.w)
323//    when (io.in(i).bits.isSoftPrefetch) {
324//      io.out(i).bits.fuType := FuType.ldu.U
325//      io.out(i).bits.fuOpType := Mux(io.in(i).bits.lsrc(1) === 1.U, LSUOpType.prefetch_r, LSUOpType.prefetch_w)
326//      io.out(i).bits.selImm := SelImm.IMM_S
327//      io.out(i).bits.imm := Cat(io.in(i).bits.imm(io.in(i).bits.imm.getWidth - 1, 5), 0.U(5.W))
328//    }
329
330    // dirty code for lui+addi(w) fusion
331    if (i < RenameWidth - 1) {
332      val fused_lui32 = io.in(i).bits.selImm === SelImm.IMM_LUI32 && io.in(i).bits.fuType === FuType.alu.U
333      when (fused_lui32) {
334        val lui_imm = io.in(i).bits.imm(19, 0)
335        val add_imm = io.in(i + 1).bits.imm(11, 0)
336        require(io.out(i).bits.imm.getWidth >= lui_imm.getWidth + add_imm.getWidth)
337        io.out(i).bits.imm := Cat(lui_imm, add_imm)
338      }
339    }
340
341    // write speculative rename table
342    // we update rat later inside commit code
343    intSpecWen(i) := needIntDest(i) && intFreeList.io.canAllocate && intFreeList.io.doAllocate && !io.rabCommits.isWalk && !io.redirect.valid
344    fpSpecWen(i)  := needFpDest(i)  && fpFreeList.io.canAllocate  && fpFreeList.io.doAllocate  && !io.rabCommits.isWalk && !io.redirect.valid
345    vecSpecWen(i) := needVecDest(i) && vecFreeList.io.canAllocate && vecFreeList.io.doAllocate && !io.rabCommits.isWalk && !io.redirect.valid
346    v0SpecWen(i) := needV0Dest(i) && v0FreeList.io.canAllocate && v0FreeList.io.doAllocate && !io.rabCommits.isWalk && !io.redirect.valid
347    vlSpecWen(i) := needVlDest(i) && vlFreeList.io.canAllocate && vlFreeList.io.doAllocate && !io.rabCommits.isWalk && !io.redirect.valid
348
349
350    if (i < RabCommitWidth) {
351      walkIntSpecWen(i) := walkNeedIntDest(i) && !io.redirect.valid
352      walkPdest(i) := io.rabCommits.info(i).pdest
353    } else {
354      walkPdest(i) := io.out(i).bits.pdest
355    }
356  }
357
358  /**
359    * How to set psrc:
360    * - bypass the pdest to psrc if previous instructions write to the same ldest as lsrc
361    * - default: psrc from RAT
362    * How to set pdest:
363    * - Mux(isMove, psrc, pdest_from_freelist).
364    *
365    * The critical path of rename lies here:
366    * When move elimination is enabled, we need to update the rat with psrc.
367    * However, psrc maybe comes from previous instructions' pdest, which comes from freelist.
368    *
369    * If we expand these logic for pdest(N):
370    * pdest(N) = Mux(isMove(N), psrc(N), freelist_out(N))
371    *          = Mux(isMove(N), Mux(bypass(N, N - 1), pdest(N - 1),
372    *                           Mux(bypass(N, N - 2), pdest(N - 2),
373    *                           ...
374    *                           Mux(bypass(N, 0),     pdest(0),
375    *                                                 rat_out(N))...)),
376    *                           freelist_out(N))
377    */
378  // a simple functional model for now
379  io.out(0).bits.pdest := Mux(isMove(0), uops(0).psrc.head, uops(0).pdest)
380
381  // psrc(n) + pdest(1)
382  val bypassCond: Vec[MixedVec[UInt]] = Wire(Vec(numRegSrc + 1, MixedVec(List.tabulate(RenameWidth-1)(i => UInt((i+1).W)))))
383  require(io.in(0).bits.srcType.size == io.in(0).bits.numSrc)
384  private val pdestLoc = io.in.head.bits.srcType.size // 2 vector src: v0, vl&vtype
385  println(s"[Rename] idx of pdest in bypassCond $pdestLoc")
386  for (i <- 1 until RenameWidth) {
387    val v0Cond = io.in(i).bits.srcType.zipWithIndex.map{ case (s, i) =>
388      if (i == 3) (s === SrcType.vp) || (s === SrcType.v0)
389      else false.B
390    } :+ needV0Dest(i)
391    val vlCond = io.in(i).bits.srcType.zipWithIndex.map{ case (s, i) =>
392      if (i == 4) s === SrcType.vp
393      else false.B
394    } :+ needVlDest(i)
395    val vecCond = io.in(i).bits.srcType.map(_ === SrcType.vp) :+ needVecDest(i)
396    val fpCond  = io.in(i).bits.srcType.map(_ === SrcType.fp) :+ needFpDest(i)
397    val intCond = io.in(i).bits.srcType.map(_ === SrcType.xp) :+ needIntDest(i)
398    val target = io.in(i).bits.lsrc :+ io.in(i).bits.ldest
399    for ((((((cond1, (condV0, condVl)), cond2), cond3), t), j) <- vecCond.zip(v0Cond.zip(vlCond)).zip(fpCond).zip(intCond).zip(target).zipWithIndex) {
400      val destToSrc = io.in.take(i).zipWithIndex.map { case (in, j) =>
401        val indexMatch = in.bits.ldest === t
402        val writeMatch =  cond3 && needIntDest(j) || cond2 && needFpDest(j) || cond1 && needVecDest(j)
403        val v0vlMatch = condV0 && needV0Dest(j) || condVl && needVlDest(j)
404        indexMatch && writeMatch || v0vlMatch
405      }
406      bypassCond(j)(i - 1) := VecInit(destToSrc).asUInt
407    }
408    io.out(i).bits.psrc(0) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(0)(i-1).asBools).foldLeft(uops(i).psrc(0)) {
409      (z, next) => Mux(next._2, next._1, z)
410    }
411    io.out(i).bits.psrc(1) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(1)(i-1).asBools).foldLeft(uops(i).psrc(1)) {
412      (z, next) => Mux(next._2, next._1, z)
413    }
414    io.out(i).bits.psrc(2) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(2)(i-1).asBools).foldLeft(uops(i).psrc(2)) {
415      (z, next) => Mux(next._2, next._1, z)
416    }
417    io.out(i).bits.psrc(3) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(3)(i-1).asBools).foldLeft(uops(i).psrc(3)) {
418      (z, next) => Mux(next._2, next._1, z)
419    }
420    io.out(i).bits.psrc(4) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(4)(i-1).asBools).foldLeft(uops(i).psrc(4)) {
421      (z, next) => Mux(next._2, next._1, z)
422    }
423    io.out(i).bits.pdest := Mux(isMove(i), io.out(i).bits.psrc(0), uops(i).pdest)
424
425    // Todo: better implementation for fields reuse
426    // For fused-lui-load, load.src(0) is replaced by the imm.
427    val last_is_lui = io.in(i - 1).bits.selImm === SelImm.IMM_U && io.in(i - 1).bits.srcType(0) =/= SrcType.pc
428    val this_is_load = io.in(i).bits.fuType === FuType.ldu.U
429    val lui_to_load = io.in(i - 1).valid && io.in(i - 1).bits.ldest === io.in(i).bits.lsrc(0)
430    val fused_lui_load = last_is_lui && this_is_load && lui_to_load
431    when (fused_lui_load) {
432      // The first LOAD operand (base address) is replaced by LUI-imm and stored in imm
433      val lui_imm = io.in(i - 1).bits.imm(ImmUnion.U.len - 1, 0)
434      val ld_imm = io.in(i).bits.imm(ImmUnion.I.len - 1, 0)
435      require(io.out(i).bits.imm.getWidth >= lui_imm.getWidth + ld_imm.getWidth)
436      io.out(i).bits.srcType(0) := SrcType.imm
437      io.out(i).bits.imm := Cat(lui_imm, ld_imm)
438    }
439
440  }
441
442  val genSnapshot = Cat(io.out.map(out => out.fire && out.bits.snapshot)).orR
443  val lastCycleCreateSnpt = RegInit(false.B)
444  lastCycleCreateSnpt := genSnapshot && !io.snptIsFull
445  val sameSnptDistance = (RobCommitWidth * 4).U
446  // notInSameSnpt: 1.robidxHead - snapLastEnq >= sameSnptDistance 2.no snap
447  val notInSameSnpt = GatedValidRegNext(distanceBetween(robIdxHeadNext, io.snptLastEnq.bits) >= sameSnptDistance || !io.snptLastEnq.valid)
448  val allowSnpt = if (EnableRenameSnapshot) notInSameSnpt && !lastCycleCreateSnpt && io.in.head.bits.firstUop else false.B
449  io.out.zip(io.in).foreach{ case (out, in) => out.bits.snapshot := allowSnpt && (!in.bits.preDecodeInfo.notCFI || FuType.isJump(in.bits.fuType)) && in.fire }
450  io.out.map{ x =>
451    x.bits.hasException := Cat(selectFrontend(x.bits.exceptionVec) :+ x.bits.exceptionVec(illegalInstr) :+ x.bits.exceptionVec(virtualInstr)).orR || x.bits.trigger.getFrontendCanFire
452  }
453  if(backendParams.debugEn){
454    dontTouch(robIdxHeadNext)
455    dontTouch(notInSameSnpt)
456    dontTouch(genSnapshot)
457  }
458  intFreeList.io.snpt := io.snpt
459  fpFreeList.io.snpt := io.snpt
460  vecFreeList.io.snpt := io.snpt
461  v0FreeList.io.snpt := io.snpt
462  vlFreeList.io.snpt := io.snpt
463  intFreeList.io.snpt.snptEnq := genSnapshot
464  fpFreeList.io.snpt.snptEnq := genSnapshot
465  vecFreeList.io.snpt.snptEnq := genSnapshot
466  v0FreeList.io.snpt.snptEnq := genSnapshot
467  vlFreeList.io.snpt.snptEnq := genSnapshot
468
469  /**
470    * Instructions commit: update freelist and rename table
471    */
472  for (i <- 0 until RabCommitWidth) {
473    val commitValid = io.rabCommits.isCommit && io.rabCommits.commitValid(i)
474    val walkValid = io.rabCommits.isWalk && io.rabCommits.walkValid(i)
475
476    // I. RAT Update
477    // When redirect happens (mis-prediction), don't update the rename table
478    io.intRenamePorts(i).wen  := intSpecWen(i)
479    io.intRenamePorts(i).addr := uops(i).ldest
480    io.intRenamePorts(i).data := io.out(i).bits.pdest
481
482    io.fpRenamePorts(i).wen  := fpSpecWen(i)
483    io.fpRenamePorts(i).addr := uops(i).ldest
484    io.fpRenamePorts(i).data := fpFreeList.io.allocatePhyReg(i)
485
486    io.vecRenamePorts(i).wen := vecSpecWen(i)
487    io.vecRenamePorts(i).addr := uops(i).ldest
488    io.vecRenamePorts(i).data := vecFreeList.io.allocatePhyReg(i)
489
490    io.v0RenamePorts(i).wen := v0SpecWen(i)
491    io.v0RenamePorts(i).addr := uops(i).ldest
492    io.v0RenamePorts(i).data := v0FreeList.io.allocatePhyReg(i)
493
494    io.vlRenamePorts(i).wen := vlSpecWen(i)
495    io.vlRenamePorts(i).addr := uops(i).ldest
496    io.vlRenamePorts(i).data := vlFreeList.io.allocatePhyReg(i)
497
498    // II. Free List Update
499    intFreeList.io.freeReq(i) := io.int_need_free(i)
500    intFreeList.io.freePhyReg(i) := RegNext(io.int_old_pdest(i))
501    fpFreeList.io.freeReq(i)  := GatedValidRegNext(commitValid && needDestRegCommit(Reg_F, io.rabCommits.info(i)))
502    fpFreeList.io.freePhyReg(i) := io.fp_old_pdest(i)
503    vecFreeList.io.freeReq(i)  := GatedValidRegNext(commitValid && needDestRegCommit(Reg_V, io.rabCommits.info(i)))
504    vecFreeList.io.freePhyReg(i) := io.vec_old_pdest(i)
505    v0FreeList.io.freeReq(i) := GatedValidRegNext(commitValid && needDestRegCommit(Reg_V0, io.rabCommits.info(i)))
506    v0FreeList.io.freePhyReg(i) := io.v0_old_pdest(i)
507    vlFreeList.io.freeReq(i) := GatedValidRegNext(commitValid && needDestRegCommit(Reg_Vl, io.rabCommits.info(i)))
508    vlFreeList.io.freePhyReg(i) := io.vl_old_pdest(i)
509  }
510
511  /*
512  Debug and performance counters
513   */
514  def printRenameInfo(in: DecoupledIO[DecodedInst], out: DecoupledIO[DynInst]) = {
515    XSInfo(out.fire, p"pc:${Hexadecimal(in.bits.pc)} in(${in.valid},${in.ready}) " +
516      p"lsrc(0):${in.bits.lsrc(0)} -> psrc(0):${out.bits.psrc(0)} " +
517      p"lsrc(1):${in.bits.lsrc(1)} -> psrc(1):${out.bits.psrc(1)} " +
518      p"lsrc(2):${in.bits.lsrc(2)} -> psrc(2):${out.bits.psrc(2)} " +
519      p"ldest:${in.bits.ldest} -> pdest:${out.bits.pdest}\n"
520    )
521  }
522
523  for ((x,y) <- io.in.zip(io.out)) {
524    printRenameInfo(x, y)
525  }
526
527  io.out.map { case x =>
528    when(x.valid && x.bits.rfWen){
529      assert(x.bits.ldest =/= 0.U, "rfWen cannot be 1 when Int regfile ldest is 0")
530    }
531  }
532  val debugRedirect = RegEnable(io.redirect.bits, io.redirect.valid)
533  // bad speculation
534  val recStall = io.redirect.valid || io.rabCommits.isWalk
535  val ctrlRecStall = Mux(io.redirect.valid, io.redirect.bits.debugIsCtrl, io.rabCommits.isWalk && debugRedirect.debugIsCtrl)
536  val mvioRecStall = Mux(io.redirect.valid, io.redirect.bits.debugIsMemVio, io.rabCommits.isWalk && debugRedirect.debugIsMemVio)
537  val otherRecStall = recStall && !(ctrlRecStall || mvioRecStall)
538  XSPerfAccumulate("recovery_stall", recStall)
539  XSPerfAccumulate("control_recovery_stall", ctrlRecStall)
540  XSPerfAccumulate("mem_violation_recovery_stall", mvioRecStall)
541  XSPerfAccumulate("other_recovery_stall", otherRecStall)
542  // freelist stall
543  val notRecStall = !io.out.head.valid && !recStall
544  val intFlStall = notRecStall && inHeadValid && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !intFreeList.io.canAllocate
545  val fpFlStall = notRecStall && inHeadValid && intFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !fpFreeList.io.canAllocate
546  val vecFlStall = notRecStall && inHeadValid && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !vecFreeList.io.canAllocate
547  val v0FlStall = notRecStall && inHeadValid && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && vlFreeList.io.canAllocate && !v0FreeList.io.canAllocate
548  val vlFlStall = notRecStall && inHeadValid && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && !vlFreeList.io.canAllocate
549  val multiFlStall = notRecStall && inHeadValid && (PopCount(Cat(
550    !intFreeList.io.canAllocate,
551    !fpFreeList.io.canAllocate,
552    !vecFreeList.io.canAllocate,
553    !v0FreeList.io.canAllocate,
554    !vlFreeList.io.canAllocate,
555  )) > 1.U)
556  // other stall
557  val otherStall = notRecStall && !intFlStall && !fpFlStall && !vecFlStall && !v0FlStall && !vlFlStall && !multiFlStall
558
559  io.stallReason.in.backReason.valid := io.stallReason.out.backReason.valid || !io.in.head.ready
560  io.stallReason.in.backReason.bits := Mux(io.stallReason.out.backReason.valid, io.stallReason.out.backReason.bits,
561    MuxCase(TopDownCounters.OtherCoreStall.id.U, Seq(
562      ctrlRecStall  -> TopDownCounters.ControlRecoveryStall.id.U,
563      mvioRecStall  -> TopDownCounters.MemVioRecoveryStall.id.U,
564      otherRecStall -> TopDownCounters.OtherRecoveryStall.id.U,
565      intFlStall    -> TopDownCounters.IntFlStall.id.U,
566      fpFlStall     -> TopDownCounters.FpFlStall.id.U,
567      vecFlStall    -> TopDownCounters.VecFlStall.id.U,
568      v0FlStall     -> TopDownCounters.V0FlStall.id.U,
569      vlFlStall     -> TopDownCounters.VlFlStall.id.U,
570      multiFlStall  -> TopDownCounters.MultiFlStall.id.U,
571    )
572  ))
573  io.stallReason.out.reason.zip(io.stallReason.in.reason).zip(io.in.map(_.valid)).foreach { case ((out, in), valid) =>
574    out := Mux(io.stallReason.in.backReason.valid, io.stallReason.in.backReason.bits, in)
575  }
576
577  XSDebug(io.rabCommits.isWalk, p"Walk Recovery Enabled\n")
578  XSDebug(io.rabCommits.isWalk, p"validVec:${Binary(io.rabCommits.walkValid.asUInt)}\n")
579  for (i <- 0 until RabCommitWidth) {
580    val info = io.rabCommits.info(i)
581    XSDebug(io.rabCommits.isWalk && io.rabCommits.walkValid(i), p"[#$i walk info] " +
582      p"ldest:${info.ldest} rfWen:${info.rfWen} fpWen:${info.fpWen} vecWen:${info.vecWen} v0Wen:${info.v0Wen} vlWen:${info.vlWen}")
583  }
584
585  XSDebug(p"inValidVec: ${Binary(Cat(io.in.map(_.valid)))}\n")
586
587  XSPerfAccumulate("in_valid_count", PopCount(io.in.map(_.valid)))
588  XSPerfAccumulate("in_fire_count", PopCount(io.in.map(_.fire)))
589  XSPerfAccumulate("in_valid_not_ready_count", PopCount(io.in.map(x => x.valid && !x.ready)))
590  XSPerfAccumulate("wait_cycle", !io.in.head.valid && dispatchCanAcc)
591
592  // These stall reasons could overlap each other, but we configure the priority as fellows.
593  // walk stall > dispatch stall > int freelist stall > fp freelist stall
594  private val inHeadStall = io.in.head match { case x => x.valid && !x.ready }
595  private val stallForWalk      = inHeadValid &&  io.rabCommits.isWalk
596  private val stallForDispatch  = inHeadValid && !io.rabCommits.isWalk && !dispatchCanAcc
597  private val stallForIntFL     = inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !intFreeList.io.canAllocate
598  private val stallForFpFL      = inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && intFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !fpFreeList.io.canAllocate
599  private val stallForVecFL     = inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !vecFreeList.io.canAllocate
600  private val stallForV0FL      = inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && vlFreeList.io.canAllocate && !v0FreeList.io.canAllocate
601  private val stallForVlFL      = inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && !vlFreeList.io.canAllocate
602  XSPerfAccumulate("stall_cycle",          inHeadStall)
603  XSPerfAccumulate("stall_cycle_walk",     stallForWalk)
604  XSPerfAccumulate("stall_cycle_dispatch", stallForDispatch)
605  XSPerfAccumulate("stall_cycle_int",      stallForIntFL)
606  XSPerfAccumulate("stall_cycle_fp",       stallForFpFL)
607  XSPerfAccumulate("stall_cycle_vec",      stallForVecFL)
608  XSPerfAccumulate("stall_cycle_vec",      stallForV0FL)
609  XSPerfAccumulate("stall_cycle_vec",      stallForVlFL)
610
611  XSPerfHistogram("in_valid_range",  PopCount(io.in.map(_.valid)),  true.B, 0, DecodeWidth + 1, 1)
612  XSPerfHistogram("in_fire_range",   PopCount(io.in.map(_.fire)),   true.B, 0, DecodeWidth + 1, 1)
613  XSPerfHistogram("out_valid_range", PopCount(io.out.map(_.valid)), true.B, 0, DecodeWidth + 1, 1)
614  XSPerfHistogram("out_fire_range",  PopCount(io.out.map(_.fire)),  true.B, 0, DecodeWidth + 1, 1)
615
616  XSPerfAccumulate("move_instr_count", PopCount(io.out.map(out => out.fire && out.bits.isMove)))
617  val is_fused_lui_load = io.out.map(o => o.fire && o.bits.fuType === FuType.ldu.U && o.bits.srcType(0) === SrcType.imm)
618  XSPerfAccumulate("fused_lui_load_instr_count", PopCount(is_fused_lui_load))
619
620  val renamePerf = Seq(
621    ("rename_in                  ", PopCount(io.in.map(_.valid & io.in(0).ready ))                                                               ),
622    ("rename_waitinstr           ", PopCount((0 until RenameWidth).map(i => io.in(i).valid && !io.in(i).ready))                                  ),
623    ("rename_stall               ", inHeadStall),
624    ("rename_stall_cycle_walk    ", inHeadValid &&  io.rabCommits.isWalk),
625    ("rename_stall_cycle_dispatch", inHeadValid && !io.rabCommits.isWalk && !dispatchCanAcc),
626    ("rename_stall_cycle_int     ", inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !intFreeList.io.canAllocate),
627    ("rename_stall_cycle_fp      ", inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && intFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !fpFreeList.io.canAllocate),
628    ("rename_stall_cycle_vec     ", inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !vecFreeList.io.canAllocate),
629    ("rename_stall_cycle_v0      ", inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && vlFreeList.io.canAllocate && !v0FreeList.io.canAllocate),
630    ("rename_stall_cycle_vl      ", inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && !vlFreeList.io.canAllocate),
631  )
632  val intFlPerf = intFreeList.getPerfEvents
633  val fpFlPerf = fpFreeList.getPerfEvents
634  val vecFlPerf = vecFreeList.getPerfEvents
635  val v0FlPerf = v0FreeList.getPerfEvents
636  val vlFlPerf = vlFreeList.getPerfEvents
637  val perfEvents = renamePerf ++ intFlPerf ++ fpFlPerf ++ vecFlPerf ++ v0FlPerf ++ vlFlPerf
638  generatePerfEvent()
639}
640