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._ 32import xiangshan.backend.fu.FuType._ 33import xiangshan.mem.{EewLog2, GenUSWholeEmul} 34import xiangshan.mem.GenRealFlowNum 35 36class Rename(implicit p: Parameters) extends XSModule with HasCircularQueuePtrHelper with HasPerfEvents { 37 38 // params alias 39 private val numRegSrc = backendParams.numRegSrc 40 private val numVecRegSrc = backendParams.numVecRegSrc 41 private val numVecRatPorts = numVecRegSrc 42 43 println(s"[Rename] numRegSrc: $numRegSrc") 44 45 val io = IO(new Bundle() { 46 val redirect = Flipped(ValidIO(new Redirect)) 47 val rabCommits = Input(new RabCommitIO) 48 // from decode 49 val in = Vec(RenameWidth, Flipped(DecoupledIO(new DecodedInst))) 50 val fusionInfo = Vec(DecodeWidth - 1, Flipped(new FusionDecodeInfo)) 51 // ssit read result 52 val ssit = Flipped(Vec(RenameWidth, Output(new SSITEntry))) 53 // waittable read result 54 val waittable = Flipped(Vec(RenameWidth, Output(Bool()))) 55 // to rename table 56 val intReadPorts = Vec(RenameWidth, Vec(2, Input(UInt(PhyRegIdxWidth.W)))) 57 val fpReadPorts = Vec(RenameWidth, Vec(3, Input(UInt(PhyRegIdxWidth.W)))) 58 val vecReadPorts = Vec(RenameWidth, Vec(numVecRatPorts, Input(UInt(PhyRegIdxWidth.W)))) 59 val v0ReadPorts = Vec(RenameWidth, Vec(1, Input(UInt(PhyRegIdxWidth.W)))) 60 val vlReadPorts = Vec(RenameWidth, Vec(1, Input(UInt(PhyRegIdxWidth.W)))) 61 val intRenamePorts = Vec(RenameWidth, Output(new RatWritePort)) 62 val fpRenamePorts = Vec(RenameWidth, Output(new RatWritePort)) 63 val vecRenamePorts = Vec(RenameWidth, Output(new RatWritePort)) 64 val v0RenamePorts = Vec(RenameWidth, Output(new RatWritePort)) 65 val vlRenamePorts = Vec(RenameWidth, Output(new RatWritePort)) 66 // from rename table 67 val int_old_pdest = Vec(RabCommitWidth, Input(UInt(PhyRegIdxWidth.W))) 68 val fp_old_pdest = Vec(RabCommitWidth, Input(UInt(PhyRegIdxWidth.W))) 69 val vec_old_pdest = Vec(RabCommitWidth, Input(UInt(PhyRegIdxWidth.W))) 70 val v0_old_pdest = Vec(RabCommitWidth, Input(UInt(PhyRegIdxWidth.W))) 71 val vl_old_pdest = Vec(RabCommitWidth, Input(UInt(PhyRegIdxWidth.W))) 72 val int_need_free = Vec(RabCommitWidth, Input(Bool())) 73 // to dispatch1 74 val out = Vec(RenameWidth, DecoupledIO(new DynInst)) 75 // for snapshots 76 val snpt = Input(new SnapshotPort) 77 val snptLastEnq = Flipped(ValidIO(new RobPtr)) 78 val snptIsFull= Input(Bool()) 79 // debug arch ports 80 val debug_int_rat = if (backendParams.debugEn) Some(Vec(32, Input(UInt(PhyRegIdxWidth.W)))) else None 81 val debug_fp_rat = if (backendParams.debugEn) Some(Vec(32, Input(UInt(PhyRegIdxWidth.W)))) else None 82 val debug_vec_rat = if (backendParams.debugEn) Some(Vec(31, Input(UInt(PhyRegIdxWidth.W)))) else None 83 val debug_v0_rat = if (backendParams.debugEn) Some(Vec(1, Input(UInt(PhyRegIdxWidth.W)))) else None 84 val debug_vl_rat = if (backendParams.debugEn) Some(Vec(1, Input(UInt(PhyRegIdxWidth.W)))) else None 85 // perf only 86 val stallReason = new Bundle { 87 val in = Flipped(new StallReasonIO(RenameWidth)) 88 val out = new StallReasonIO(RenameWidth) 89 } 90 }) 91 92 // io alias 93 private val dispatchCanAcc = io.out.head.ready 94 95 val compressUnit = Module(new CompressUnit()) 96 // create free list and rat 97 val intFreeList = Module(new MEFreeList(IntPhyRegs)) 98 val fpFreeList = Module(new StdFreeList(FpPhyRegs - FpLogicRegs, FpLogicRegs, Reg_F)) 99 val vecFreeList = Module(new StdFreeList(VfPhyRegs - VecLogicRegs, VecLogicRegs, Reg_V, 31)) 100 val v0FreeList = Module(new StdFreeList(V0PhyRegs - V0LogicRegs, V0LogicRegs, Reg_V0, 1)) 101 val vlFreeList = Module(new StdFreeList(VlPhyRegs - VlLogicRegs, VlLogicRegs, Reg_Vl, 1)) 102 103 104 intFreeList.io.commit <> io.rabCommits 105 intFreeList.io.debug_rat.foreach(_ <> io.debug_int_rat.get) 106 fpFreeList.io.commit <> io.rabCommits 107 fpFreeList.io.debug_rat.foreach(_ <> io.debug_fp_rat.get) 108 vecFreeList.io.commit <> io.rabCommits 109 vecFreeList.io.debug_rat.foreach(_ <> io.debug_vec_rat.get) 110 v0FreeList.io.commit <> io.rabCommits 111 v0FreeList.io.debug_rat.foreach(_ <> io.debug_v0_rat.get) 112 vlFreeList.io.commit <> io.rabCommits 113 vlFreeList.io.debug_rat.foreach(_ <> io.debug_vl_rat.get) 114 115 // decide if given instruction needs allocating a new physical register (CfCtrl: from decode; RobCommitInfo: from rob) 116 def needDestReg[T <: DecodedInst](reg_t: RegType, x: T): Bool = reg_t match { 117 case Reg_I => x.rfWen && x.ldest =/= 0.U 118 case Reg_F => x.fpWen 119 case Reg_V => x.vecWen 120 case Reg_V0 => x.v0Wen 121 case Reg_Vl => x.vlWen 122 } 123 def needDestRegCommit[T <: RabCommitInfo](reg_t: RegType, x: T): Bool = { 124 reg_t match { 125 case Reg_I => x.rfWen 126 case Reg_F => x.fpWen 127 case Reg_V => x.vecWen 128 case Reg_V0 => x.v0Wen 129 case Reg_Vl => x.vlWen 130 } 131 } 132 def needDestRegWalk[T <: RabCommitInfo](reg_t: RegType, x: T): Bool = { 133 reg_t match { 134 case Reg_I => x.rfWen && x.ldest =/= 0.U 135 case Reg_F => x.fpWen 136 case Reg_V => x.vecWen 137 case Reg_V0 => x.v0Wen 138 case Reg_Vl => x.vlWen 139 } 140 } 141 142 // connect [redirect + walk] ports for fp & vec & int free list 143 Seq(fpFreeList, vecFreeList, intFreeList, v0FreeList, vlFreeList).foreach { case fl => 144 fl.io.redirect := io.redirect.valid 145 fl.io.walk := io.rabCommits.isWalk 146 } 147 // only when all free list and dispatch1 has enough space can we do allocation 148 // when isWalk, freelist can definitely allocate 149 intFreeList.io.doAllocate := fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && dispatchCanAcc || io.rabCommits.isWalk 150 fpFreeList.io.doAllocate := intFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && dispatchCanAcc || io.rabCommits.isWalk 151 vecFreeList.io.doAllocate := intFreeList.io.canAllocate && fpFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && dispatchCanAcc || io.rabCommits.isWalk 152 v0FreeList.io.doAllocate := intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && vlFreeList.io.canAllocate && dispatchCanAcc || io.rabCommits.isWalk 153 vlFreeList.io.doAllocate := intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && dispatchCanAcc || io.rabCommits.isWalk 154 155 // dispatch1 ready ++ float point free list ready ++ int free list ready ++ vec free list ready ++ not walk 156 val canOut = dispatchCanAcc && fpFreeList.io.canAllocate && intFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !io.rabCommits.isWalk 157 158 compressUnit.io.in.zip(io.in).foreach{ case(sink, source) => 159 sink.valid := source.valid 160 sink.bits := source.bits 161 } 162 val needRobFlags = compressUnit.io.out.needRobFlags 163 val instrSizesVec = compressUnit.io.out.instrSizes 164 val compressMasksVec = compressUnit.io.out.masks 165 166 // speculatively assign the instruction with an robIdx 167 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) 168 val robIdxHead = RegInit(0.U.asTypeOf(new RobPtr)) 169 val lastCycleMisprediction = GatedValidRegNext(io.redirect.valid && !io.redirect.bits.flushItself()) 170 val robIdxHeadNext = Mux(io.redirect.valid, io.redirect.bits.robIdx, // redirect: move ptr to given rob index 171 Mux(lastCycleMisprediction, robIdxHead + 1.U, // mis-predict: not flush robIdx itself 172 Mux(canOut, robIdxHead + validCount, // instructions successfully entered next stage: increase robIdx 173 /* default */ robIdxHead))) // no instructions passed by this cycle: stick to old value 174 robIdxHead := robIdxHeadNext 175 176 /** 177 * Rename: allocate free physical register and update rename table 178 */ 179 val uops = Wire(Vec(RenameWidth, new DynInst)) 180 uops.foreach( uop => { 181 uop.srcState := DontCare 182 uop.debugInfo := DontCare 183 uop.lqIdx := DontCare 184 uop.sqIdx := DontCare 185 uop.waitForRobIdx := DontCare 186 uop.singleStep := DontCare 187 uop.snapshot := DontCare 188 uop.srcLoadDependency := DontCare 189 uop.numLsElem := DontCare 190 uop.hasException := DontCare 191 }) 192 private val fuType = uops.map(_.fuType) 193 private val fuOpType = uops.map(_.fuOpType) 194 private val vtype = uops.map(_.vpu.vtype) 195 private val sew = vtype.map(_.vsew) 196 private val lmul = vtype.map(_.vlmul) 197 private val eew = uops.map(_.vpu.veew) 198 private val mop = fuOpType.map(fuOpTypeItem => LSUOpType.getVecLSMop(fuOpTypeItem)) 199 private val isVlsType = fuType.map(fuTypeItem => isVls(fuTypeItem)) 200 private val isSegment = fuType.map(fuTypeItem => isVsegls(fuTypeItem)) 201 private val isUnitStride = fuOpType.map(fuOpTypeItem => LSUOpType.isAllUS(fuOpTypeItem)) 202 private val nf = fuOpType.zip(uops.map(_.vpu.nf)).map { case (fuOpTypeItem, nfItem) => Mux(LSUOpType.isWhole(fuOpTypeItem), 0.U, nfItem) } 203 private val mulBits = 3 // dirty code 204 private val emul = fuOpType.zipWithIndex.map { case (fuOpTypeItem, index) => 205 Mux( 206 LSUOpType.isWhole(fuOpTypeItem), 207 GenUSWholeEmul(nf(index)), 208 Mux( 209 LSUOpType.isMasked(fuOpTypeItem), 210 0.U(mulBits.W), 211 EewLog2(eew(index)) - sew(index) + lmul(index) 212 ) 213 ) 214 } 215 private val isVecUnitType = isVlsType.zip(isUnitStride).map { case (isVlsTypeItme, isUnitStrideItem) => 216 isVlsTypeItme && isUnitStrideItem 217 } 218 private val instType = isSegment.zip(mop).map { case (isSegementItem, mopItem) => Cat(isSegementItem, mopItem) } 219 // There is no way to calculate the 'flow' for 'unit-stride' exactly: 220 // Whether 'unit-stride' needs to be split can only be known after obtaining the address. 221 // For scalar instructions, this is not handled here, and different assignments are done later according to the situation. 222 private val numLsElem = instType.zipWithIndex.map { case (instTypeItem, index) => 223 Mux( 224 isVecUnitType(index), 225 VecMemUnitStrideMaxFlowNum.U, 226 GenRealFlowNum(instTypeItem, emul(index), lmul(index), eew(index), sew(index)) 227 ) 228 } 229 uops.zipWithIndex.map { case(u, i) => 230 u.numLsElem := Mux(io.in(i).valid & isVlsType(i), numLsElem(i), 0.U) 231 } 232 233 val needVecDest = Wire(Vec(RenameWidth, Bool())) 234 val needFpDest = Wire(Vec(RenameWidth, Bool())) 235 val needIntDest = Wire(Vec(RenameWidth, Bool())) 236 val needV0Dest = Wire(Vec(RenameWidth, Bool())) 237 val needVlDest = Wire(Vec(RenameWidth, Bool())) 238 val hasValid = Cat(io.in.map(_.valid)).orR 239 private val inHeadValid = io.in.head.valid 240 241 val isMove = Wire(Vec(RenameWidth, Bool())) 242 isMove zip io.in.map(_.bits) foreach { 243 case (move, in) => move := Mux(in.exceptionVec.asUInt.orR, false.B, in.isMove) 244 } 245 246 val walkNeedIntDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B))) 247 val walkNeedFpDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B))) 248 val walkNeedVecDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B))) 249 val walkNeedV0Dest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B))) 250 val walkNeedVlDest = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B))) 251 val walkIsMove = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B))) 252 253 val intSpecWen = Wire(Vec(RenameWidth, Bool())) 254 val fpSpecWen = Wire(Vec(RenameWidth, Bool())) 255 val vecSpecWen = Wire(Vec(RenameWidth, Bool())) 256 val v0SpecWen = Wire(Vec(RenameWidth, Bool())) 257 val vlSpecWen = Wire(Vec(RenameWidth, Bool())) 258 259 val walkIntSpecWen = WireDefault(VecInit(Seq.fill(RenameWidth)(false.B))) 260 261 val walkPdest = Wire(Vec(RenameWidth, UInt(PhyRegIdxWidth.W))) 262 263 // uop calculation 264 for (i <- 0 until RenameWidth) { 265 (uops(i): Data).waiveAll :<= (io.in(i).bits: Data).waiveAll 266 267 // update cf according to ssit result 268 uops(i).storeSetHit := io.ssit(i).valid 269 uops(i).loadWaitStrict := io.ssit(i).strict && io.ssit(i).valid 270 uops(i).ssid := io.ssit(i).ssid 271 272 // update cf according to waittable result 273 uops(i).loadWaitBit := io.waittable(i) 274 275 uops(i).replayInst := false.B // set by IQ or MemQ 276 // alloc a new phy reg 277 needV0Dest(i) := io.in(i).valid && needDestReg(Reg_V0, io.in(i).bits) 278 needVlDest(i) := io.in(i).valid && needDestReg(Reg_Vl, io.in(i).bits) 279 needVecDest(i) := io.in(i).valid && needDestReg(Reg_V, io.in(i).bits) 280 needFpDest(i) := io.in(i).valid && needDestReg(Reg_F, io.in(i).bits) 281 needIntDest(i) := io.in(i).valid && needDestReg(Reg_I, io.in(i).bits) 282 if (i < RabCommitWidth) { 283 walkNeedIntDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_I, io.rabCommits.info(i)) 284 walkNeedFpDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_F, io.rabCommits.info(i)) 285 walkNeedVecDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_V, io.rabCommits.info(i)) 286 walkNeedV0Dest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_V0, io.rabCommits.info(i)) 287 walkNeedVlDest(i) := io.rabCommits.walkValid(i) && needDestRegWalk(Reg_Vl, io.rabCommits.info(i)) 288 walkIsMove(i) := io.rabCommits.info(i).isMove 289 } 290 fpFreeList.io.allocateReq(i) := needFpDest(i) 291 fpFreeList.io.walkReq(i) := walkNeedFpDest(i) 292 vecFreeList.io.allocateReq(i) := needVecDest(i) 293 vecFreeList.io.walkReq(i) := walkNeedVecDest(i) 294 v0FreeList.io.allocateReq(i) := needV0Dest(i) 295 v0FreeList.io.walkReq(i) := walkNeedV0Dest(i) 296 vlFreeList.io.allocateReq(i) := needVlDest(i) 297 vlFreeList.io.walkReq(i) := walkNeedVlDest(i) 298 intFreeList.io.allocateReq(i) := needIntDest(i) && !isMove(i) 299 intFreeList.io.walkReq(i) := walkNeedIntDest(i) && !walkIsMove(i) 300 301 // no valid instruction from decode stage || all resources (dispatch1 + both free lists) ready 302 io.in(i).ready := !hasValid || canOut 303 304 uops(i).robIdx := robIdxHead + PopCount(io.in.zip(needRobFlags).take(i).map{ case(in, needRobFlag) => in.valid && in.bits.lastUop && needRobFlag}) 305 uops(i).instrSize := instrSizesVec(i) 306 when(isMove(i)) { 307 uops(i).numUops := 0.U 308 uops(i).numWB := 0.U 309 } 310 if (i > 0) { 311 when(!needRobFlags(i - 1)) { 312 uops(i).firstUop := false.B 313 uops(i).ftqPtr := uops(i - 1).ftqPtr 314 uops(i).ftqOffset := uops(i - 1).ftqOffset 315 uops(i).numUops := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse)) 316 uops(i).numWB := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse)) 317 } 318 } 319 when(!needRobFlags(i)) { 320 uops(i).lastUop := false.B 321 uops(i).numUops := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse)) 322 uops(i).numWB := instrSizesVec(i) - PopCount(compressMasksVec(i) & Cat(isMove.reverse)) 323 } 324 uops(i).wfflags := (compressMasksVec(i) & Cat(io.in.map(_.bits.wfflags).reverse)).orR 325 uops(i).dirtyFs := (compressMasksVec(i) & Cat(io.in.map(_.bits.fpWen).reverse)).orR 326 // vector instructions' uopSplitType cannot be UopSplitType.SCA_SIM 327 uops(i).dirtyVs := (compressMasksVec(i) & Cat(io.in.map(_.bits.uopSplitType =/= UopSplitType.SCA_SIM).reverse)).orR 328 // psrc0,psrc1,psrc2 don't require v0ReadPorts because their srcType can distinguish whether they are V0 or not 329 uops(i).psrc(0) := Mux1H(uops(i).srcType(0)(2, 0), Seq(io.intReadPorts(i)(0), io.fpReadPorts(i)(0), io.vecReadPorts(i)(0))) 330 uops(i).psrc(1) := Mux1H(uops(i).srcType(1)(2, 0), Seq(io.intReadPorts(i)(1), io.fpReadPorts(i)(1), io.vecReadPorts(i)(1))) 331 uops(i).psrc(2) := Mux1H(uops(i).srcType(2)(2, 1), Seq(io.fpReadPorts(i)(2), io.vecReadPorts(i)(2))) 332 uops(i).psrc(3) := io.v0ReadPorts(i)(0) 333 uops(i).psrc(4) := io.vlReadPorts(i)(0) 334 335 // int psrc2 should be bypassed from next instruction if it is fused 336 if (i < RenameWidth - 1) { 337 when (io.fusionInfo(i).rs2FromRs2 || io.fusionInfo(i).rs2FromRs1) { 338 uops(i).psrc(1) := Mux(io.fusionInfo(i).rs2FromRs2, io.intReadPorts(i + 1)(1), io.intReadPorts(i + 1)(0)) 339 }.elsewhen(io.fusionInfo(i).rs2FromZero) { 340 uops(i).psrc(1) := 0.U 341 } 342 } 343 uops(i).eliminatedMove := isMove(i) 344 345 // update pdest 346 uops(i).pdest := MuxCase(0.U, Seq( 347 needIntDest(i) -> intFreeList.io.allocatePhyReg(i), 348 needFpDest(i) -> fpFreeList.io.allocatePhyReg(i), 349 needVecDest(i) -> vecFreeList.io.allocatePhyReg(i), 350 needV0Dest(i) -> v0FreeList.io.allocatePhyReg(i), 351 needVlDest(i) -> vlFreeList.io.allocatePhyReg(i), 352 )) 353 354 // Assign performance counters 355 uops(i).debugInfo.renameTime := GTimer() 356 357 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 358 io.out(i).bits := uops(i) 359 // Todo: move these shit in decode stage 360 // dirty code for fence. The lsrc is passed by imm. 361 when (io.out(i).bits.fuType === FuType.fence.U) { 362 io.out(i).bits.imm := Cat(io.in(i).bits.lsrc(1), io.in(i).bits.lsrc(0)) 363 } 364 365 // dirty code for SoftPrefetch (prefetch.r/prefetch.w) 366// when (io.in(i).bits.isSoftPrefetch) { 367// io.out(i).bits.fuType := FuType.ldu.U 368// io.out(i).bits.fuOpType := Mux(io.in(i).bits.lsrc(1) === 1.U, LSUOpType.prefetch_r, LSUOpType.prefetch_w) 369// io.out(i).bits.selImm := SelImm.IMM_S 370// io.out(i).bits.imm := Cat(io.in(i).bits.imm(io.in(i).bits.imm.getWidth - 1, 5), 0.U(5.W)) 371// } 372 373 // dirty code for lui+addi(w) fusion 374 if (i < RenameWidth - 1) { 375 val fused_lui32 = io.in(i).bits.selImm === SelImm.IMM_LUI32 && io.in(i).bits.fuType === FuType.alu.U 376 when (fused_lui32) { 377 val lui_imm = io.in(i).bits.imm(19, 0) 378 val add_imm = io.in(i + 1).bits.imm(11, 0) 379 require(io.out(i).bits.imm.getWidth >= lui_imm.getWidth + add_imm.getWidth) 380 io.out(i).bits.imm := Cat(lui_imm, add_imm) 381 } 382 } 383 384 // write speculative rename table 385 // we update rat later inside commit code 386 intSpecWen(i) := needIntDest(i) && intFreeList.io.canAllocate && intFreeList.io.doAllocate && !io.rabCommits.isWalk && !io.redirect.valid 387 fpSpecWen(i) := needFpDest(i) && fpFreeList.io.canAllocate && fpFreeList.io.doAllocate && !io.rabCommits.isWalk && !io.redirect.valid 388 vecSpecWen(i) := needVecDest(i) && vecFreeList.io.canAllocate && vecFreeList.io.doAllocate && !io.rabCommits.isWalk && !io.redirect.valid 389 v0SpecWen(i) := needV0Dest(i) && v0FreeList.io.canAllocate && v0FreeList.io.doAllocate && !io.rabCommits.isWalk && !io.redirect.valid 390 vlSpecWen(i) := needVlDest(i) && vlFreeList.io.canAllocate && vlFreeList.io.doAllocate && !io.rabCommits.isWalk && !io.redirect.valid 391 392 393 if (i < RabCommitWidth) { 394 walkIntSpecWen(i) := walkNeedIntDest(i) && !io.redirect.valid 395 walkPdest(i) := io.rabCommits.info(i).pdest 396 } else { 397 walkPdest(i) := io.out(i).bits.pdest 398 } 399 } 400 401 /** 402 * How to set psrc: 403 * - bypass the pdest to psrc if previous instructions write to the same ldest as lsrc 404 * - default: psrc from RAT 405 * How to set pdest: 406 * - Mux(isMove, psrc, pdest_from_freelist). 407 * 408 * The critical path of rename lies here: 409 * When move elimination is enabled, we need to update the rat with psrc. 410 * However, psrc maybe comes from previous instructions' pdest, which comes from freelist. 411 * 412 * If we expand these logic for pdest(N): 413 * pdest(N) = Mux(isMove(N), psrc(N), freelist_out(N)) 414 * = Mux(isMove(N), Mux(bypass(N, N - 1), pdest(N - 1), 415 * Mux(bypass(N, N - 2), pdest(N - 2), 416 * ... 417 * Mux(bypass(N, 0), pdest(0), 418 * rat_out(N))...)), 419 * freelist_out(N)) 420 */ 421 // a simple functional model for now 422 io.out(0).bits.pdest := Mux(isMove(0), uops(0).psrc.head, uops(0).pdest) 423 424 // psrc(n) + pdest(1) 425 val bypassCond: Vec[MixedVec[UInt]] = Wire(Vec(numRegSrc + 1, MixedVec(List.tabulate(RenameWidth-1)(i => UInt((i+1).W))))) 426 require(io.in(0).bits.srcType.size == io.in(0).bits.numSrc) 427 private val pdestLoc = io.in.head.bits.srcType.size // 2 vector src: v0, vl&vtype 428 println(s"[Rename] idx of pdest in bypassCond $pdestLoc") 429 for (i <- 1 until RenameWidth) { 430 val v0Cond = io.in(i).bits.srcType.zipWithIndex.map{ case (s, i) => 431 if (i == 3) (s === SrcType.vp) || (s === SrcType.v0) 432 else false.B 433 } :+ needV0Dest(i) 434 val vlCond = io.in(i).bits.srcType.zipWithIndex.map{ case (s, i) => 435 if (i == 4) s === SrcType.vp 436 else false.B 437 } :+ needVlDest(i) 438 val vecCond = io.in(i).bits.srcType.map(_ === SrcType.vp) :+ needVecDest(i) 439 val fpCond = io.in(i).bits.srcType.map(_ === SrcType.fp) :+ needFpDest(i) 440 val intCond = io.in(i).bits.srcType.map(_ === SrcType.xp) :+ needIntDest(i) 441 val target = io.in(i).bits.lsrc :+ io.in(i).bits.ldest 442 for ((((((cond1, (condV0, condVl)), cond2), cond3), t), j) <- vecCond.zip(v0Cond.zip(vlCond)).zip(fpCond).zip(intCond).zip(target).zipWithIndex) { 443 val destToSrc = io.in.take(i).zipWithIndex.map { case (in, j) => 444 val indexMatch = in.bits.ldest === t 445 val writeMatch = cond3 && needIntDest(j) || cond2 && needFpDest(j) || cond1 && needVecDest(j) 446 val v0vlMatch = condV0 && needV0Dest(j) || condVl && needVlDest(j) 447 indexMatch && writeMatch || v0vlMatch 448 } 449 bypassCond(j)(i - 1) := VecInit(destToSrc).asUInt 450 } 451 io.out(i).bits.psrc(0) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(0)(i-1).asBools).foldLeft(uops(i).psrc(0)) { 452 (z, next) => Mux(next._2, next._1, z) 453 } 454 io.out(i).bits.psrc(1) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(1)(i-1).asBools).foldLeft(uops(i).psrc(1)) { 455 (z, next) => Mux(next._2, next._1, z) 456 } 457 io.out(i).bits.psrc(2) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(2)(i-1).asBools).foldLeft(uops(i).psrc(2)) { 458 (z, next) => Mux(next._2, next._1, z) 459 } 460 io.out(i).bits.psrc(3) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(3)(i-1).asBools).foldLeft(uops(i).psrc(3)) { 461 (z, next) => Mux(next._2, next._1, z) 462 } 463 io.out(i).bits.psrc(4) := io.out.take(i).map(_.bits.pdest).zip(bypassCond(4)(i-1).asBools).foldLeft(uops(i).psrc(4)) { 464 (z, next) => Mux(next._2, next._1, z) 465 } 466 io.out(i).bits.pdest := Mux(isMove(i), io.out(i).bits.psrc(0), uops(i).pdest) 467 468 // Todo: better implementation for fields reuse 469 // For fused-lui-load, load.src(0) is replaced by the imm. 470 val last_is_lui = io.in(i - 1).bits.selImm === SelImm.IMM_U && io.in(i - 1).bits.srcType(0) =/= SrcType.pc 471 val this_is_load = io.in(i).bits.fuType === FuType.ldu.U 472 val lui_to_load = io.in(i - 1).valid && io.in(i - 1).bits.ldest === io.in(i).bits.lsrc(0) 473 val fused_lui_load = last_is_lui && this_is_load && lui_to_load 474 when (fused_lui_load) { 475 // The first LOAD operand (base address) is replaced by LUI-imm and stored in imm 476 val lui_imm = io.in(i - 1).bits.imm(ImmUnion.U.len - 1, 0) 477 val ld_imm = io.in(i).bits.imm(ImmUnion.I.len - 1, 0) 478 require(io.out(i).bits.imm.getWidth >= lui_imm.getWidth + ld_imm.getWidth) 479 io.out(i).bits.srcType(0) := SrcType.imm 480 io.out(i).bits.imm := Cat(lui_imm, ld_imm) 481 } 482 483 } 484 485 val genSnapshot = Cat(io.out.map(out => out.fire && out.bits.snapshot)).orR 486 val lastCycleCreateSnpt = RegInit(false.B) 487 lastCycleCreateSnpt := genSnapshot && !io.snptIsFull 488 val sameSnptDistance = (RobCommitWidth * 4).U 489 // notInSameSnpt: 1.robidxHead - snapLastEnq >= sameSnptDistance 2.no snap 490 val notInSameSnpt = GatedValidRegNext(distanceBetween(robIdxHeadNext, io.snptLastEnq.bits) >= sameSnptDistance || !io.snptLastEnq.valid) 491 val allowSnpt = if (EnableRenameSnapshot) notInSameSnpt && !lastCycleCreateSnpt && io.in.head.bits.firstUop else false.B 492 io.out.zip(io.in).foreach{ case (out, in) => out.bits.snapshot := allowSnpt && (!in.bits.preDecodeInfo.notCFI || FuType.isJump(in.bits.fuType)) && in.fire } 493 io.out.map{ x => 494 x.bits.hasException := Cat(selectFrontend(x.bits.exceptionVec) :+ x.bits.exceptionVec(illegalInstr) :+ x.bits.exceptionVec(virtualInstr)).orR || x.bits.trigger.getFrontendCanFire 495 } 496 if(backendParams.debugEn){ 497 dontTouch(robIdxHeadNext) 498 dontTouch(notInSameSnpt) 499 dontTouch(genSnapshot) 500 } 501 intFreeList.io.snpt := io.snpt 502 fpFreeList.io.snpt := io.snpt 503 vecFreeList.io.snpt := io.snpt 504 v0FreeList.io.snpt := io.snpt 505 vlFreeList.io.snpt := io.snpt 506 intFreeList.io.snpt.snptEnq := genSnapshot 507 fpFreeList.io.snpt.snptEnq := genSnapshot 508 vecFreeList.io.snpt.snptEnq := genSnapshot 509 v0FreeList.io.snpt.snptEnq := genSnapshot 510 vlFreeList.io.snpt.snptEnq := genSnapshot 511 512 /** 513 * Instructions commit: update freelist and rename table 514 */ 515 for (i <- 0 until RabCommitWidth) { 516 val commitValid = io.rabCommits.isCommit && io.rabCommits.commitValid(i) 517 val walkValid = io.rabCommits.isWalk && io.rabCommits.walkValid(i) 518 519 // I. RAT Update 520 // When redirect happens (mis-prediction), don't update the rename table 521 io.intRenamePorts(i).wen := intSpecWen(i) 522 io.intRenamePorts(i).addr := uops(i).ldest 523 io.intRenamePorts(i).data := io.out(i).bits.pdest 524 525 io.fpRenamePorts(i).wen := fpSpecWen(i) 526 io.fpRenamePorts(i).addr := uops(i).ldest 527 io.fpRenamePorts(i).data := fpFreeList.io.allocatePhyReg(i) 528 529 io.vecRenamePorts(i).wen := vecSpecWen(i) 530 io.vecRenamePorts(i).addr := uops(i).ldest 531 io.vecRenamePorts(i).data := vecFreeList.io.allocatePhyReg(i) 532 533 io.v0RenamePorts(i).wen := v0SpecWen(i) 534 io.v0RenamePorts(i).addr := uops(i).ldest 535 io.v0RenamePorts(i).data := v0FreeList.io.allocatePhyReg(i) 536 537 io.vlRenamePorts(i).wen := vlSpecWen(i) 538 io.vlRenamePorts(i).addr := uops(i).ldest 539 io.vlRenamePorts(i).data := vlFreeList.io.allocatePhyReg(i) 540 541 // II. Free List Update 542 intFreeList.io.freeReq(i) := io.int_need_free(i) 543 intFreeList.io.freePhyReg(i) := RegNext(io.int_old_pdest(i)) 544 fpFreeList.io.freeReq(i) := GatedValidRegNext(commitValid && needDestRegCommit(Reg_F, io.rabCommits.info(i))) 545 fpFreeList.io.freePhyReg(i) := io.fp_old_pdest(i) 546 vecFreeList.io.freeReq(i) := GatedValidRegNext(commitValid && needDestRegCommit(Reg_V, io.rabCommits.info(i))) 547 vecFreeList.io.freePhyReg(i) := io.vec_old_pdest(i) 548 v0FreeList.io.freeReq(i) := GatedValidRegNext(commitValid && needDestRegCommit(Reg_V0, io.rabCommits.info(i))) 549 v0FreeList.io.freePhyReg(i) := io.v0_old_pdest(i) 550 vlFreeList.io.freeReq(i) := GatedValidRegNext(commitValid && needDestRegCommit(Reg_Vl, io.rabCommits.info(i))) 551 vlFreeList.io.freePhyReg(i) := io.vl_old_pdest(i) 552 } 553 554 /* 555 Debug and performance counters 556 */ 557 def printRenameInfo(in: DecoupledIO[DecodedInst], out: DecoupledIO[DynInst]) = { 558 XSInfo(out.fire, p"pc:${Hexadecimal(in.bits.pc)} in(${in.valid},${in.ready}) " + 559 p"lsrc(0):${in.bits.lsrc(0)} -> psrc(0):${out.bits.psrc(0)} " + 560 p"lsrc(1):${in.bits.lsrc(1)} -> psrc(1):${out.bits.psrc(1)} " + 561 p"lsrc(2):${in.bits.lsrc(2)} -> psrc(2):${out.bits.psrc(2)} " + 562 p"ldest:${in.bits.ldest} -> pdest:${out.bits.pdest}\n" 563 ) 564 } 565 566 for ((x,y) <- io.in.zip(io.out)) { 567 printRenameInfo(x, y) 568 } 569 570 io.out.map { case x => 571 when(x.valid && x.bits.rfWen){ 572 assert(x.bits.ldest =/= 0.U, "rfWen cannot be 1 when Int regfile ldest is 0") 573 } 574 } 575 val debugRedirect = RegEnable(io.redirect.bits, io.redirect.valid) 576 // bad speculation 577 val recStall = io.redirect.valid || io.rabCommits.isWalk 578 val ctrlRecStall = Mux(io.redirect.valid, io.redirect.bits.debugIsCtrl, io.rabCommits.isWalk && debugRedirect.debugIsCtrl) 579 val mvioRecStall = Mux(io.redirect.valid, io.redirect.bits.debugIsMemVio, io.rabCommits.isWalk && debugRedirect.debugIsMemVio) 580 val otherRecStall = recStall && !(ctrlRecStall || mvioRecStall) 581 XSPerfAccumulate("recovery_stall", recStall) 582 XSPerfAccumulate("control_recovery_stall", ctrlRecStall) 583 XSPerfAccumulate("mem_violation_recovery_stall", mvioRecStall) 584 XSPerfAccumulate("other_recovery_stall", otherRecStall) 585 // freelist stall 586 val notRecStall = !io.out.head.valid && !recStall 587 val intFlStall = notRecStall && inHeadValid && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !intFreeList.io.canAllocate 588 val fpFlStall = notRecStall && inHeadValid && intFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !fpFreeList.io.canAllocate 589 val vecFlStall = notRecStall && inHeadValid && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !vecFreeList.io.canAllocate 590 val v0FlStall = notRecStall && inHeadValid && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && vlFreeList.io.canAllocate && !v0FreeList.io.canAllocate 591 val vlFlStall = notRecStall && inHeadValid && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && !vlFreeList.io.canAllocate 592 val multiFlStall = notRecStall && inHeadValid && (PopCount(Cat( 593 !intFreeList.io.canAllocate, 594 !fpFreeList.io.canAllocate, 595 !vecFreeList.io.canAllocate, 596 !v0FreeList.io.canAllocate, 597 !vlFreeList.io.canAllocate, 598 )) > 1.U) 599 // other stall 600 val otherStall = notRecStall && !intFlStall && !fpFlStall && !vecFlStall && !v0FlStall && !vlFlStall && !multiFlStall 601 602 io.stallReason.in.backReason.valid := io.stallReason.out.backReason.valid || !io.in.head.ready 603 io.stallReason.in.backReason.bits := Mux(io.stallReason.out.backReason.valid, io.stallReason.out.backReason.bits, 604 MuxCase(TopDownCounters.OtherCoreStall.id.U, Seq( 605 ctrlRecStall -> TopDownCounters.ControlRecoveryStall.id.U, 606 mvioRecStall -> TopDownCounters.MemVioRecoveryStall.id.U, 607 otherRecStall -> TopDownCounters.OtherRecoveryStall.id.U, 608 intFlStall -> TopDownCounters.IntFlStall.id.U, 609 fpFlStall -> TopDownCounters.FpFlStall.id.U, 610 vecFlStall -> TopDownCounters.VecFlStall.id.U, 611 v0FlStall -> TopDownCounters.V0FlStall.id.U, 612 vlFlStall -> TopDownCounters.VlFlStall.id.U, 613 multiFlStall -> TopDownCounters.MultiFlStall.id.U, 614 ) 615 )) 616 io.stallReason.out.reason.zip(io.stallReason.in.reason).zip(io.in.map(_.valid)).foreach { case ((out, in), valid) => 617 out := Mux(io.stallReason.in.backReason.valid, io.stallReason.in.backReason.bits, in) 618 } 619 620 XSDebug(io.rabCommits.isWalk, p"Walk Recovery Enabled\n") 621 XSDebug(io.rabCommits.isWalk, p"validVec:${Binary(io.rabCommits.walkValid.asUInt)}\n") 622 for (i <- 0 until RabCommitWidth) { 623 val info = io.rabCommits.info(i) 624 XSDebug(io.rabCommits.isWalk && io.rabCommits.walkValid(i), p"[#$i walk info] " + 625 p"ldest:${info.ldest} rfWen:${info.rfWen} fpWen:${info.fpWen} vecWen:${info.vecWen} v0Wen:${info.v0Wen} vlWen:${info.vlWen}") 626 } 627 628 XSDebug(p"inValidVec: ${Binary(Cat(io.in.map(_.valid)))}\n") 629 630 XSPerfAccumulate("in_valid_count", PopCount(io.in.map(_.valid))) 631 XSPerfAccumulate("in_fire_count", PopCount(io.in.map(_.fire))) 632 XSPerfAccumulate("in_valid_not_ready_count", PopCount(io.in.map(x => x.valid && !x.ready))) 633 XSPerfAccumulate("wait_cycle", !io.in.head.valid && dispatchCanAcc) 634 635 // These stall reasons could overlap each other, but we configure the priority as fellows. 636 // walk stall > dispatch stall > int freelist stall > fp freelist stall 637 private val inHeadStall = io.in.head match { case x => x.valid && !x.ready } 638 private val stallForWalk = inHeadValid && io.rabCommits.isWalk 639 private val stallForDispatch = inHeadValid && !io.rabCommits.isWalk && !dispatchCanAcc 640 private val stallForIntFL = inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !intFreeList.io.canAllocate 641 private val stallForFpFL = inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && intFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !fpFreeList.io.canAllocate 642 private val stallForVecFL = inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !vecFreeList.io.canAllocate 643 private val stallForV0FL = inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && vlFreeList.io.canAllocate && !v0FreeList.io.canAllocate 644 private val stallForVlFL = inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && !vlFreeList.io.canAllocate 645 XSPerfAccumulate("stall_cycle", inHeadStall) 646 XSPerfAccumulate("stall_cycle_walk", stallForWalk) 647 XSPerfAccumulate("stall_cycle_dispatch", stallForDispatch) 648 XSPerfAccumulate("stall_cycle_int", stallForIntFL) 649 XSPerfAccumulate("stall_cycle_fp", stallForFpFL) 650 XSPerfAccumulate("stall_cycle_vec", stallForVecFL) 651 XSPerfAccumulate("stall_cycle_vec", stallForV0FL) 652 XSPerfAccumulate("stall_cycle_vec", stallForVlFL) 653 654 XSPerfHistogram("in_valid_range", PopCount(io.in.map(_.valid)), true.B, 0, DecodeWidth + 1, 1) 655 XSPerfHistogram("in_fire_range", PopCount(io.in.map(_.fire)), true.B, 0, DecodeWidth + 1, 1) 656 XSPerfHistogram("out_valid_range", PopCount(io.out.map(_.valid)), true.B, 0, DecodeWidth + 1, 1) 657 XSPerfHistogram("out_fire_range", PopCount(io.out.map(_.fire)), true.B, 0, DecodeWidth + 1, 1) 658 659 XSPerfAccumulate("move_instr_count", PopCount(io.out.map(out => out.fire && out.bits.isMove))) 660 val is_fused_lui_load = io.out.map(o => o.fire && o.bits.fuType === FuType.ldu.U && o.bits.srcType(0) === SrcType.imm) 661 XSPerfAccumulate("fused_lui_load_instr_count", PopCount(is_fused_lui_load)) 662 663 val renamePerf = Seq( 664 ("rename_in ", PopCount(io.in.map(_.valid & io.in(0).ready )) ), 665 ("rename_waitinstr ", PopCount((0 until RenameWidth).map(i => io.in(i).valid && !io.in(i).ready)) ), 666 ("rename_stall ", inHeadStall), 667 ("rename_stall_cycle_walk ", inHeadValid && io.rabCommits.isWalk), 668 ("rename_stall_cycle_dispatch", inHeadValid && !io.rabCommits.isWalk && !dispatchCanAcc), 669 ("rename_stall_cycle_int ", inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !intFreeList.io.canAllocate), 670 ("rename_stall_cycle_fp ", inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && intFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !fpFreeList.io.canAllocate), 671 ("rename_stall_cycle_vec ", inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && v0FreeList.io.canAllocate && vlFreeList.io.canAllocate && !vecFreeList.io.canAllocate), 672 ("rename_stall_cycle_v0 ", inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && vlFreeList.io.canAllocate && !v0FreeList.io.canAllocate), 673 ("rename_stall_cycle_vl ", inHeadValid && !io.rabCommits.isWalk && dispatchCanAcc && intFreeList.io.canAllocate && fpFreeList.io.canAllocate && vecFreeList.io.canAllocate && v0FreeList.io.canAllocate && !vlFreeList.io.canAllocate), 674 ) 675 val intFlPerf = intFreeList.getPerfEvents 676 val fpFlPerf = fpFreeList.getPerfEvents 677 val vecFlPerf = vecFreeList.getPerfEvents 678 val v0FlPerf = v0FreeList.getPerfEvents 679 val vlFlPerf = vlFreeList.getPerfEvents 680 val perfEvents = renamePerf ++ intFlPerf ++ fpFlPerf ++ vecFlPerf ++ v0FlPerf ++ vlFlPerf 681 generatePerfEvent() 682} 683