1/*************************************************************************************** 2* Copyright (c) 2020-2021 Institute of Computing Technology, Chinese Academy of Sciences 3* 4* XiangShan is licensed under Mulan PSL v2. 5* You can use this software according to the terms and conditions of the Mulan PSL v2. 6* You may obtain a copy of Mulan PSL v2 at: 7* http://license.coscl.org.cn/MulanPSL2 8* 9* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, 10* EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, 11* MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE. 12* 13* See the Mulan PSL v2 for more details. 14***************************************************************************************/ 15 16package xiangshan.backend 17 18import chipsalliance.rocketchip.config.Parameters 19import chisel3._ 20import chisel3.util._ 21import utils._ 22import xiangshan._ 23import xiangshan.backend.decode.{DecodeStage, ImmUnion} 24import xiangshan.backend.rename.{BusyTable, Rename} 25import xiangshan.backend.dispatch.Dispatch 26import xiangshan.backend.exu._ 27import xiangshan.frontend.{FtqRead, HasFtqHelper} 28import xiangshan.backend.roq.{Roq, RoqCSRIO, RoqLsqIO, RoqPtr} 29import xiangshan.mem.LsqEnqIO 30 31class CtrlToIntBlockIO(implicit p: Parameters) extends XSBundle { 32 val enqIqCtrl = Vec(exuParameters.IntExuCnt, DecoupledIO(new MicroOp)) 33 val readRf = Vec(NRIntReadPorts, Output(UInt(PhyRegIdxWidth.W))) 34 val jumpPc = Output(UInt(VAddrBits.W)) 35 val jalr_target = Output(UInt(VAddrBits.W)) 36 // int block only uses port 0~7 37 val readPortIndex = Vec(exuParameters.IntExuCnt, Output(UInt(log2Ceil(8 / 2).W))) // TODO parameterize 8 here 38 val redirect = ValidIO(new Redirect) 39 val flush = Output(Bool()) 40 val debug_rat = Vec(32, Output(UInt(PhyRegIdxWidth.W))) 41} 42 43class CtrlToFpBlockIO(implicit p: Parameters) extends XSBundle { 44 val enqIqCtrl = Vec(exuParameters.FpExuCnt, DecoupledIO(new MicroOp)) 45 val readRf = Vec(NRFpReadPorts, Output(UInt(PhyRegIdxWidth.W))) 46 // fp block uses port 0~11 47 val readPortIndex = Vec(exuParameters.FpExuCnt, Output(UInt(log2Ceil((NRFpReadPorts - exuParameters.StuCnt) / 3).W))) 48 val redirect = ValidIO(new Redirect) 49 val flush = Output(Bool()) 50 val debug_rat = Vec(32, Output(UInt(PhyRegIdxWidth.W))) 51} 52 53class CtrlToLsBlockIO(implicit p: Parameters) extends XSBundle { 54 val enqIqCtrl = Vec(exuParameters.LsExuCnt, DecoupledIO(new MicroOp)) 55 val enqLsq = Flipped(new LsqEnqIO) 56 val memPredUpdate = Vec(StorePipelineWidth, Input(new MemPredUpdateReq)) 57 val redirect = ValidIO(new Redirect) 58 val flush = Output(Bool()) 59} 60 61class RedirectGenerator(implicit p: Parameters) extends XSModule 62 with HasCircularQueuePtrHelper with HasFtqHelper { 63 val numRedirect = exuParameters.JmpCnt + exuParameters.AluCnt 64 val io = IO(new Bundle() { 65 val exuMispredict = Vec(numRedirect, Flipped(ValidIO(new ExuOutput))) 66 val loadReplay = Flipped(ValidIO(new Redirect)) 67 val flush = Input(Bool()) 68 val stage1FtqRead = Vec(numRedirect + 1, new FtqRead) 69 val stage2FtqRead = new FtqRead 70 val stage2Redirect = ValidIO(new Redirect) 71 val stage3Redirect = ValidIO(new Redirect) 72 val memPredUpdate = Output(new MemPredUpdateReq) 73 val memPredFtqRead = new FtqRead // read req send form stage 2 74 }) 75 /* 76 LoadQueue Jump ALU0 ALU1 ALU2 ALU3 exception Stage1 77 | | | | | | | 78 |============= reg & compare =====| | ======== 79 | | 80 | | 81 | | Stage2 82 | | 83 redirect (flush backend) | 84 | | 85 === reg === | ======== 86 | | 87 |----- mux (exception first) -----| Stage3 88 | 89 redirect (send to frontend) 90 */ 91 private class Wrapper(val n: Int) extends Bundle { 92 val redirect = new Redirect 93 val valid = Bool() 94 val idx = UInt(log2Up(n).W) 95 } 96 def selectOldestRedirect(xs: Seq[Valid[Redirect]]): Vec[Bool] = { 97 val compareVec = (0 until xs.length).map(i => (0 until i).map(j => isAfter(xs(j).bits.roqIdx, xs(i).bits.roqIdx))) 98 val resultOnehot = VecInit((0 until xs.length).map(i => Cat((0 until xs.length).map(j => 99 (if (j < i) !xs(j).valid || compareVec(i)(j) 100 else if (j == i) xs(i).valid 101 else !xs(j).valid || !compareVec(j)(i)) 102 )).andR)) 103 resultOnehot 104 } 105 106 for((ptr, redirect) <- io.stage1FtqRead.map(_.ptr).zip( 107 io.exuMispredict.map(_.bits.redirect) :+ io.loadReplay.bits 108 )){ ptr := redirect.ftqIdx } 109 110 def getRedirect(exuOut: Valid[ExuOutput]): ValidIO[Redirect] = { 111 val redirect = Wire(Valid(new Redirect)) 112 redirect.valid := exuOut.valid && exuOut.bits.redirect.cfiUpdate.isMisPred 113 redirect.bits := exuOut.bits.redirect 114 redirect 115 } 116 117 val jumpOut = io.exuMispredict.head 118 val allRedirect = VecInit(io.exuMispredict.map(x => getRedirect(x)) :+ io.loadReplay) 119 val oldestOneHot = selectOldestRedirect(allRedirect) 120 val needFlushVec = VecInit(allRedirect.map(_.bits.roqIdx.needFlush(io.stage2Redirect, io.flush))) 121 val oldestValid = VecInit(oldestOneHot.zip(needFlushVec).map{ case (v, f) => v && !f }).asUInt.orR 122 val oldestExuOutput = Mux1H((0 until 5).map(oldestOneHot), io.exuMispredict) 123 val oldestRedirect = Mux1H(oldestOneHot, allRedirect) 124 125 val s1_jumpTarget = RegEnable(jumpOut.bits.redirect.cfiUpdate.target, jumpOut.valid) 126 val s1_imm12_reg = RegNext(oldestExuOutput.bits.uop.ctrl.imm(11, 0)) 127 val s1_pd = RegNext(oldestExuOutput.bits.uop.cf.pd) 128 val s1_redirect_bits_reg = RegNext(oldestRedirect.bits) 129 val s1_redirect_valid_reg = RegNext(oldestValid) 130 val s1_redirect_onehot = RegNext(oldestOneHot) 131 132 // stage1 -> stage2 133 io.stage2Redirect.valid := s1_redirect_valid_reg && !io.flush 134 io.stage2Redirect.bits := s1_redirect_bits_reg 135 io.stage2Redirect.bits.cfiUpdate := DontCare 136 // at stage2, we read ftq to get pc 137 io.stage2FtqRead.ptr := s1_redirect_bits_reg.ftqIdx 138 139 val s1_isReplay = s1_redirect_onehot(5) 140 val s1_isJump = s1_redirect_onehot(0) 141 val ftqRead = Mux1H(s1_redirect_onehot, io.stage1FtqRead).entry 142 val cfiUpdate_pc = Cat( 143 ftqRead.ftqPC.head(VAddrBits - s1_redirect_bits_reg.ftqOffset.getWidth - instOffsetBits), 144 s1_redirect_bits_reg.ftqOffset, 145 0.U(instOffsetBits.W) 146 ) 147 val real_pc = GetPcByFtq( 148 ftqRead.ftqPC, s1_redirect_bits_reg.ftqOffset, 149 ftqRead.lastPacketPC.valid, 150 ftqRead.lastPacketPC.bits 151 ) 152 val brTarget = real_pc + SignExt(ImmUnion.B.toImm32(s1_imm12_reg), XLEN) 153 val snpc = real_pc + Mux(s1_pd.isRVC, 2.U, 4.U) 154 val target = Mux(s1_isReplay, 155 real_pc, // repaly from itself 156 Mux(s1_redirect_bits_reg.cfiUpdate.taken, 157 Mux(s1_isJump, s1_jumpTarget, brTarget), 158 snpc 159 ) 160 ) 161 162 // get pc from ftq 163 io.memPredFtqRead.ptr := s1_redirect_bits_reg.stFtqIdx 164 // valid only if redirect is caused by load violation 165 // store_pc is used to update store set 166 val memPredFtqRead = io.memPredFtqRead.entry 167 val store_pc = GetPcByFtq(memPredFtqRead.ftqPC, RegNext(s1_redirect_bits_reg).stFtqOffset, 168 memPredFtqRead.lastPacketPC.valid, 169 memPredFtqRead.lastPacketPC.bits 170 ) 171 172 // update load violation predictor if load violation redirect triggered 173 io.memPredUpdate.valid := RegNext(s1_isReplay && s1_redirect_valid_reg, init = false.B) 174 // update wait table 175 io.memPredUpdate.waddr := RegNext(XORFold(real_pc(VAddrBits-1, 1), MemPredPCWidth)) 176 io.memPredUpdate.wdata := true.B 177 // update store set 178 io.memPredUpdate.ldpc := RegNext(XORFold(real_pc(VAddrBits-1, 1), MemPredPCWidth)) 179 // store pc is ready 1 cycle after s1_isReplay is judged 180 io.memPredUpdate.stpc := XORFold(store_pc(VAddrBits-1, 1), MemPredPCWidth) 181 182 183 val s2_br_mask = RegEnable(ftqRead.br_mask, enable = s1_redirect_valid_reg) 184 val s2_sawNotTakenBranch = RegEnable(VecInit((0 until PredictWidth).map{ i => 185 if(i == 0) false.B else Cat(ftqRead.br_mask.take(i)).orR() 186 })(s1_redirect_bits_reg.ftqOffset), enable = s1_redirect_valid_reg) 187 val s2_hist = RegEnable(ftqRead.hist, enable = s1_redirect_valid_reg) 188 val s2_target = RegEnable(target, enable = s1_redirect_valid_reg) 189 val s2_pd = RegEnable(s1_pd, enable = s1_redirect_valid_reg) 190 val s2_cfiUpdata_pc = RegEnable(cfiUpdate_pc, enable = s1_redirect_valid_reg) 191 val s2_redirect_bits_reg = RegEnable(s1_redirect_bits_reg, enable = s1_redirect_valid_reg) 192 val s2_redirect_valid_reg = RegNext(s1_redirect_valid_reg && !io.flush, init = false.B) 193 val s2_ftqRead = io.stage2FtqRead.entry 194 195 io.stage3Redirect.valid := s2_redirect_valid_reg 196 io.stage3Redirect.bits := s2_redirect_bits_reg 197 val stage3CfiUpdate = io.stage3Redirect.bits.cfiUpdate 198 stage3CfiUpdate.pc := s2_cfiUpdata_pc 199 stage3CfiUpdate.pd := s2_pd 200 stage3CfiUpdate.rasSp := s2_ftqRead.rasSp 201 stage3CfiUpdate.rasEntry := s2_ftqRead.rasTop 202 stage3CfiUpdate.predHist := s2_ftqRead.predHist 203 stage3CfiUpdate.specCnt := s2_ftqRead.specCnt 204 stage3CfiUpdate.hist := s2_hist 205 stage3CfiUpdate.predTaken := s2_redirect_bits_reg.cfiUpdate.predTaken 206 stage3CfiUpdate.sawNotTakenBranch := s2_sawNotTakenBranch 207 stage3CfiUpdate.target := s2_target 208 stage3CfiUpdate.taken := s2_redirect_bits_reg.cfiUpdate.taken 209 stage3CfiUpdate.isMisPred := s2_redirect_bits_reg.cfiUpdate.isMisPred 210} 211 212class CtrlBlock(implicit p: Parameters) extends XSModule 213 with HasCircularQueuePtrHelper with HasFtqHelper { 214 val io = IO(new Bundle { 215 val frontend = Flipped(new FrontendToBackendIO) 216 val fromIntBlock = Flipped(new IntBlockToCtrlIO) 217 val fromFpBlock = Flipped(new FpBlockToCtrlIO) 218 val fromLsBlock = Flipped(new LsBlockToCtrlIO) 219 val toIntBlock = new CtrlToIntBlockIO 220 val toFpBlock = new CtrlToFpBlockIO 221 val toLsBlock = new CtrlToLsBlockIO 222 val roqio = new Bundle { 223 // to int block 224 val toCSR = new RoqCSRIO 225 val exception = ValidIO(new ExceptionInfo) 226 // to mem block 227 val lsq = new RoqLsqIO 228 } 229 val csrCtrl = Input(new CustomCSRCtrlIO) 230 val perfInfo = Output(new Bundle{ 231 val ctrlInfo = new Bundle { 232 val roqFull = Input(Bool()) 233 val intdqFull = Input(Bool()) 234 val fpdqFull = Input(Bool()) 235 val lsdqFull = Input(Bool()) 236 } 237 }) 238 }) 239 240 val decode = Module(new DecodeStage) 241 val rename = Module(new Rename) 242 val dispatch = Module(new Dispatch) 243 val intBusyTable = Module(new BusyTable(NRIntReadPorts, NRIntWritePorts)) 244 val fpBusyTable = Module(new BusyTable(NRFpReadPorts, NRFpWritePorts)) 245 val redirectGen = Module(new RedirectGenerator) 246 247 val roqWbSize = NRIntWritePorts + NRFpWritePorts + exuParameters.StuCnt 248 val roq = Module(new Roq(roqWbSize)) 249 250 val backendRedirect = redirectGen.io.stage2Redirect 251 val frontendRedirect = redirectGen.io.stage3Redirect 252 val flush = roq.io.flushOut.valid 253 val flushReg = RegNext(flush) 254 255 val exuRedirect = io.fromIntBlock.exuRedirect.map(x => { 256 val valid = x.valid && x.bits.redirectValid 257 val killedByOlder = x.bits.uop.roqIdx.needFlush(backendRedirect, flushReg) 258 val delayed = Wire(Valid(new ExuOutput)) 259 delayed.valid := RegNext(valid && !killedByOlder, init = false.B) 260 delayed.bits := RegEnable(x.bits, x.valid) 261 delayed 262 }) 263 val loadReplay = Wire(Valid(new Redirect)) 264 loadReplay.valid := RegNext(io.fromLsBlock.replay.valid && 265 !io.fromLsBlock.replay.bits.roqIdx.needFlush(backendRedirect, flushReg), 266 init = false.B 267 ) 268 loadReplay.bits := RegEnable(io.fromLsBlock.replay.bits, io.fromLsBlock.replay.valid) 269 VecInit(io.frontend.fromFtq.ftqRead.tail.dropRight(2)) <> redirectGen.io.stage1FtqRead 270 io.frontend.fromFtq.ftqRead.dropRight(1).last <> redirectGen.io.memPredFtqRead 271 io.frontend.fromFtq.cfiRead <> redirectGen.io.stage2FtqRead 272 redirectGen.io.exuMispredict <> exuRedirect 273 redirectGen.io.loadReplay <> loadReplay 274 redirectGen.io.flush := flushReg 275 276 for(i <- 0 until CommitWidth){ 277 io.frontend.toFtq.roq_commits(i).valid := roq.io.commits.valid(i) && !roq.io.commits.isWalk 278 io.frontend.toFtq.roq_commits(i).bits := roq.io.commits.info(i) 279 } 280 io.frontend.toFtq.redirect <> backendRedirect 281 io.frontend.toFtq.flush := flushReg 282 io.frontend.toFtq.flushIdx := RegNext(roq.io.flushOut.bits.ftqIdx) 283 io.frontend.toFtq.flushOffset := RegNext(roq.io.flushOut.bits.ftqOffset) 284 io.frontend.toFtq.frontendRedirect <> frontendRedirect 285 io.frontend.toFtq.exuWriteback <> exuRedirect 286 287 io.frontend.fromFtq.ftqRead.last.ptr := roq.io.flushOut.bits.ftqIdx 288 val flushPC = GetPcByFtq( 289 io.frontend.fromFtq.ftqRead.last.entry.ftqPC, 290 RegEnable(roq.io.flushOut.bits.ftqOffset, roq.io.flushOut.valid), 291 io.frontend.fromFtq.ftqRead.last.entry.lastPacketPC.valid, 292 io.frontend.fromFtq.ftqRead.last.entry.lastPacketPC.bits 293 ) 294 295 val flushRedirect = Wire(Valid(new Redirect)) 296 flushRedirect.valid := flushReg 297 flushRedirect.bits := DontCare 298 flushRedirect.bits.ftqIdx := RegEnable(roq.io.flushOut.bits.ftqIdx, flush) 299 flushRedirect.bits.interrupt := true.B 300 flushRedirect.bits.cfiUpdate.target := Mux(io.roqio.toCSR.isXRet || roq.io.exception.valid, 301 io.roqio.toCSR.trapTarget, 302 flushPC + 4.U // flush pipe 303 ) 304 val flushRedirectReg = Wire(Valid(new Redirect)) 305 flushRedirectReg.valid := RegNext(flushRedirect.valid, init = false.B) 306 flushRedirectReg.bits := RegEnable(flushRedirect.bits, enable = flushRedirect.valid) 307 308 io.frontend.redirect_cfiUpdate := Mux(flushRedirectReg.valid, flushRedirectReg, frontendRedirect) 309 310 decode.io.in <> io.frontend.cfVec 311 // currently, we only update wait table when isReplay 312 decode.io.memPredUpdate(0) <> RegNext(redirectGen.io.memPredUpdate) 313 decode.io.memPredUpdate(1) := DontCare 314 decode.io.memPredUpdate(1).valid := false.B 315 // decode.io.memPredUpdate <> io.toLsBlock.memPredUpdate 316 decode.io.csrCtrl := RegNext(io.csrCtrl) 317 318 319 val jumpInst = dispatch.io.enqIQCtrl(0).bits 320 val ftqOffsetReg = Reg(UInt(log2Up(PredictWidth).W)) 321 ftqOffsetReg := jumpInst.cf.ftqOffset 322 io.frontend.fromFtq.ftqRead(0).ptr := jumpInst.cf.ftqPtr // jump 323 io.toIntBlock.jumpPc := GetPcByFtq( 324 io.frontend.fromFtq.ftqRead(0).entry.ftqPC, ftqOffsetReg, 325 io.frontend.fromFtq.ftqRead(0).entry.lastPacketPC.valid, 326 io.frontend.fromFtq.ftqRead(0).entry.lastPacketPC.bits 327 ) 328 io.toIntBlock.jalr_target := io.frontend.fromFtq.ftqRead(0).entry.target 329 330 // pipeline between decode and dispatch 331 for (i <- 0 until RenameWidth) { 332 PipelineConnect(decode.io.out(i), rename.io.in(i), rename.io.in(i).ready, 333 flushReg || io.frontend.redirect_cfiUpdate.valid) 334 } 335 336 rename.io.redirect <> backendRedirect 337 rename.io.flush := flushReg 338 rename.io.roqCommits <> roq.io.commits 339 rename.io.out <> dispatch.io.fromRename 340 rename.io.renameBypass <> dispatch.io.renameBypass 341 rename.io.dispatchInfo <> dispatch.io.preDpInfo 342 rename.io.csrCtrl <> RegNext(io.csrCtrl) 343 344 dispatch.io.redirect <> backendRedirect 345 dispatch.io.flush := flushReg 346 dispatch.io.enqRoq <> roq.io.enq 347 dispatch.io.enqLsq <> io.toLsBlock.enqLsq 348 dispatch.io.readIntRf <> io.toIntBlock.readRf 349 dispatch.io.readFpRf <> io.toFpBlock.readRf 350 dispatch.io.allocPregs.zipWithIndex.foreach { case (preg, i) => 351 intBusyTable.io.allocPregs(i).valid := preg.isInt 352 fpBusyTable.io.allocPregs(i).valid := preg.isFp 353 intBusyTable.io.allocPregs(i).bits := preg.preg 354 fpBusyTable.io.allocPregs(i).bits := preg.preg 355 } 356 dispatch.io.numExist <> io.fromIntBlock.numExist ++ io.fromFpBlock.numExist ++ io.fromLsBlock.numExist 357 dispatch.io.enqIQCtrl <> io.toIntBlock.enqIqCtrl ++ io.toFpBlock.enqIqCtrl ++ io.toLsBlock.enqIqCtrl 358// dispatch.io.enqIQData <> io.toIntBlock.enqIqData ++ io.toFpBlock.enqIqData ++ io.toLsBlock.enqIqData 359 dispatch.io.csrCtrl <> io.csrCtrl 360 dispatch.io.storeIssue <> io.fromLsBlock.stIn 361 362 363 fpBusyTable.io.flush := flushReg 364 intBusyTable.io.flush := flushReg 365 for((wb, setPhyRegRdy) <- io.fromIntBlock.wbRegs.zip(intBusyTable.io.wbPregs)){ 366 setPhyRegRdy.valid := wb.valid && wb.bits.uop.ctrl.rfWen 367 setPhyRegRdy.bits := wb.bits.uop.pdest 368 } 369 for((wb, setPhyRegRdy) <- io.fromFpBlock.wbRegs.zip(fpBusyTable.io.wbPregs)){ 370 setPhyRegRdy.valid := wb.valid && wb.bits.uop.ctrl.fpWen 371 setPhyRegRdy.bits := wb.bits.uop.pdest 372 } 373 intBusyTable.io.read <> dispatch.io.readIntState 374 fpBusyTable.io.read <> dispatch.io.readFpState 375 376 roq.io.redirect <> backendRedirect 377 val exeWbResults = VecInit(io.fromIntBlock.wbRegs ++ io.fromFpBlock.wbRegs ++ io.fromLsBlock.stOut) 378 for((roq_wb, wb) <- roq.io.exeWbResults.zip(exeWbResults)) { 379 roq_wb.valid := RegNext(wb.valid && !wb.bits.uop.roqIdx.needFlush(backendRedirect, flushReg)) 380 roq_wb.bits := RegNext(wb.bits) 381 } 382 383 // TODO: is 'backendRedirect' necesscary? 384 io.toIntBlock.redirect <> backendRedirect 385 io.toIntBlock.flush <> flushReg 386 io.toIntBlock.debug_rat <> rename.io.debug_int_rat 387 io.toFpBlock.redirect <> backendRedirect 388 io.toFpBlock.flush <> flushReg 389 io.toFpBlock.debug_rat <> rename.io.debug_fp_rat 390 io.toLsBlock.redirect <> backendRedirect 391 io.toLsBlock.flush <> flushReg 392 393 dispatch.io.readPortIndex.intIndex <> io.toIntBlock.readPortIndex 394 dispatch.io.readPortIndex.fpIndex <> io.toFpBlock.readPortIndex 395 396 // roq to int block 397 io.roqio.toCSR <> roq.io.csr 398 io.roqio.toCSR.perfinfo.retiredInstr <> RegNext(roq.io.csr.perfinfo.retiredInstr) 399 io.roqio.exception := roq.io.exception 400 io.roqio.exception.bits.uop.cf.pc := flushPC 401 // roq to mem block 402 io.roqio.lsq <> roq.io.lsq 403 404 io.perfInfo.ctrlInfo.roqFull := RegNext(roq.io.roqFull) 405 io.perfInfo.ctrlInfo.intdqFull := RegNext(dispatch.io.ctrlInfo.intdqFull) 406 io.perfInfo.ctrlInfo.fpdqFull := RegNext(dispatch.io.ctrlInfo.fpdqFull) 407 io.perfInfo.ctrlInfo.lsdqFull := RegNext(dispatch.io.ctrlInfo.lsdqFull) 408} 409