xref: /XiangShan/src/main/scala/xiangshan/backend/Backend.scala (revision 0ecbc6d638a24a599c96fe49b2e26d2af34051ef) 
1package xiangshan.backend
2
3import bus.simplebus.SimpleBusUC
4import chisel3._
5import chisel3.util._
6import chisel3.util.experimental.BoringUtils
7import noop.MemMMUIO
8import xiangshan._
9import xiangshan.backend.decode.{DecodeBuffer, DecodeStage}
10import xiangshan.backend.rename.Rename
11import xiangshan.backend.brq.Brq
12import xiangshan.backend.dispatch.Dispatch
13import xiangshan.backend.exu._
14import xiangshan.backend.fu.FunctionUnit
15import xiangshan.backend.issue.{IssueQueue, ReservationStationNew}
16import xiangshan.backend.regfile.{Regfile, RfWritePort}
17import xiangshan.backend.roq.Roq
18import xiangshan.mem._
19import utils.ParallelOR
20
21/** Backend Pipeline:
22  * Decode -> Rename -> Dispatch-1 -> Dispatch-2 -> Issue -> Exe
23  */
24class Backend extends XSModule
25  with NeedImpl {
26  val io = IO(new Bundle {
27    val frontend = Flipped(new FrontendToBackendIO)
28    val mem = Flipped(new MemToBackendIO)
29  })
30
31
32  val aluExeUnits =Array.tabulate(exuParameters.AluCnt)(_ => Module(new AluExeUnit))
33  val jmpExeUnit = Module(new JmpExeUnit)
34  val mulExeUnits = Array.tabulate(exuParameters.MulCnt)(_ => Module(new MulExeUnit))
35  val mduExeUnits = Array.tabulate(exuParameters.MduCnt)(_ => Module(new MulDivExeUnit))
36  // val fmacExeUnits = Array.tabulate(exuParameters.FmacCnt)(_ => Module(new Fmac))
37  // val fmiscExeUnits = Array.tabulate(exuParameters.FmiscCnt)(_ => Module(new Fmisc))
38  // val fmiscDivSqrtExeUnits = Array.tabulate(exuParameters.FmiscDivSqrtCnt)(_ => Module(new FmiscDivSqrt))
39  val exeUnits = jmpExeUnit +: (aluExeUnits ++ mulExeUnits ++ mduExeUnits)
40  exeUnits.foreach(_.io.exception := DontCare)
41  exeUnits.foreach(_.io.dmem := DontCare)
42  exeUnits.foreach(_.io.mcommit := DontCare)
43
44  val decode = Module(new DecodeStage)
45  val brq = Module(new Brq)
46  val decBuf = Module(new DecodeBuffer)
47  val rename = Module(new Rename)
48  val dispatch = Module(new Dispatch)
49  val roq = Module(new Roq)
50  val intRf = Module(new Regfile(
51    numReadPorts = NRIntReadPorts,
52    numWirtePorts = NRIntWritePorts,
53    hasZero = true
54  ))
55  val fpRf = Module(new Regfile(
56    numReadPorts = NRFpReadPorts,
57    numWirtePorts = NRFpWritePorts,
58    hasZero = false
59  ))
60  val memRf = Module(new Regfile(
61    numReadPorts = 2*exuParameters.StuCnt + exuParameters.LduCnt,
62    numWirtePorts = NRIntWritePorts,
63    hasZero = true,
64    isMemRf = true
65  ))
66
67  // backend redirect, flush pipeline
68  val redirect = Mux(
69    roq.io.redirect.valid,
70    roq.io.redirect,
71    Mux(
72      brq.io.redirect.valid,
73      brq.io.redirect,
74      io.mem.replayAll
75    )
76  )
77
78  io.frontend.redirect := redirect
79  io.frontend.redirect.valid := redirect.valid && !redirect.bits.isReplay
80
81  val memConfigs =
82    Seq.fill(exuParameters.LduCnt)(Exu.ldExeUnitCfg) ++
83    Seq.fill(exuParameters.StuCnt)(Exu.stExeUnitCfg)
84
85  val exuConfigs = exeUnits.map(_.config) ++ memConfigs
86
87  val exeWbReqs = exeUnits.map(_.io.out) ++ io.mem.ldout ++ io.mem.stout
88
89  def needWakeup(cfg: ExuConfig): Boolean =
90    (cfg.readIntRf && cfg.writeIntRf) || (cfg.readFpRf && cfg.writeFpRf)
91
92  def needData(a: ExuConfig, b: ExuConfig): Boolean =
93    (a.readIntRf && b.writeIntRf) || (a.readFpRf && b.writeFpRf)
94
95
96  val reservedStations  = exeUnits.zipWithIndex.map({ case (exu, i) =>
97
98    val cfg = exu.config
99
100    val writeBackedData = exuConfigs.zip(exeWbReqs).filter(x => x._1.hasCertainLatency && needData(x._1, cfg)).map(_._2.bits.data)
101    val wakeupCnt = writeBackedData.length
102
103    val extraListenPorts = exuConfigs
104      .zip(exeWbReqs)
105      .filter(x => x._1.hasUncertainlatency && needData(x._1, cfg))
106      .map(_._2)
107    val extraListenPortsCnt = extraListenPorts.length
108
109
110    val rs = Module(new ReservationStationNew(cfg, wakeupCnt, extraListenPortsCnt))
111
112    println(s"exu:${cfg.name} wakeupCnt: ${wakeupCnt} extraListenPorts: ${extraListenPortsCnt}")
113
114    rs.io.redirect <> redirect
115    rs.io.numExist <> dispatch.io.numExist(i)
116    rs.io.enqCtrl <> dispatch.io.enqIQCtrl(i)
117    rs.io.enqData <> dispatch.io.enqIQData(i)
118
119
120    rs.io.writeBackedData <> writeBackedData
121    for((x, y) <- rs.io.extraListenPorts.zip(extraListenPorts)){
122      x.valid := y.fire()
123      x.bits := y.bits
124    }
125
126    exu.io.in <> rs.io.deq
127    exu.io.redirect <> redirect
128
129    rs
130  })
131
132  for( rs <- reservedStations){
133
134    rs.io.broadcastedUops <> reservedStations.
135        filter(x => x.exuCfg.hasCertainLatency && needData(rs.exuCfg, x.exuCfg)).
136        map(_.io.selectedUop)
137
138  }
139
140
141  val issueQueues = exuConfigs.
142    zipWithIndex.
143    takeRight(exuParameters.LduCnt + exuParameters.StuCnt).
144    map({case (cfg, i) =>
145      val wakeUpDateVec = exuConfigs.zip(exeWbReqs).filter(x => needData(cfg, x._1)).map(_._2)
146      val bypassUopVec = reservedStations.
147        filter(r => r.exuCfg.enableBypass && needData(cfg, r.exuCfg)).map(_.io.selectedUop)
148      val bypassDataVec = exuConfigs.zip(exeWbReqs).
149        filter(x => x._1.enableBypass && needData(cfg, x._1)).map(_._2)
150
151      val iq = Module(new IssueQueue(
152        cfg, wakeUpDateVec.length, bypassUopVec.length
153      ))
154      println(s"exu:${cfg.name} wakeupCnt:${wakeUpDateVec.length} bypassCnt:${bypassUopVec.length}")
155      iq.io.redirect <> redirect
156      iq.io.tlbFeedback := io.mem.tlbFeedback(i - exuParameters.ExuCnt + exuParameters.LduCnt + exuParameters.StuCnt)
157      iq.io.enq <> dispatch.io.enqIQCtrl(i)
158      dispatch.io.numExist(i) := iq.io.numExist
159      for(
160        (wakeUpPort, exuOut) <-
161        iq.io.wakeUpPorts.zip(wakeUpDateVec)
162      ){
163        wakeUpPort.bits := exuOut.bits
164        wakeUpPort.valid := exuOut.fire() // data after arbit
165      }
166      iq.io.bypassUops <> bypassUopVec
167      for(i <- bypassDataVec.indices){
168        iq.io.bypassData(i).valid := bypassDataVec(i).valid
169        iq.io.bypassData(i).bits := bypassDataVec(i).bits
170      }
171      iq
172    })
173
174  io.mem.commits <> roq.io.commits
175  io.mem.roqDeqPtr := roq.io.roqDeqPtr
176  io.mem.ldin <> issueQueues.filter(_.exuCfg == Exu.ldExeUnitCfg).map(_.io.deq)
177  io.mem.stin <> issueQueues.filter(_.exuCfg == Exu.stExeUnitCfg).map(_.io.deq)
178  jmpExeUnit.io.exception.valid := roq.io.redirect.valid && roq.io.redirect.bits.isException
179  jmpExeUnit.io.exception.bits := roq.io.exception
180
181  io.frontend.outOfOrderBrInfo <> brq.io.outOfOrderBrInfo
182  io.frontend.inOrderBrInfo <> brq.io.inOrderBrInfo
183
184  decode.io.in <> io.frontend.cfVec
185  brq.io.roqRedirect <> roq.io.redirect
186  brq.io.memRedirect <> io.mem.replayAll
187  brq.io.bcommit := roq.io.bcommit
188  brq.io.enqReqs <> decode.io.toBrq
189  for ((x, y) <- brq.io.exuRedirect.zip(exeUnits.filter(_.config.hasRedirect))) {
190    x.bits := y.io.out.bits
191    x.valid := y.io.out.fire() && y.io.out.bits.redirectValid
192  }
193  decode.io.brTags <> brq.io.brTags
194  decBuf.io.isWalking := ParallelOR(roq.io.commits.map(c => c.valid && c.bits.isWalk)) // TODO: opt this
195  decBuf.io.redirect <> redirect
196  decBuf.io.in <> decode.io.out
197
198  rename.io.redirect <> redirect
199  rename.io.roqCommits <> roq.io.commits
200  rename.io.in <> decBuf.io.out
201  rename.io.intRfReadAddr <> dispatch.io.readIntRf.map(_.addr) ++ dispatch.io.intMemRegAddr
202  rename.io.intPregRdy <> dispatch.io.intPregRdy ++ dispatch.io.intMemRegRdy
203  rename.io.fpRfReadAddr <> dispatch.io.readFpRf.map(_.addr) ++ dispatch.io.fpMemRegAddr
204  rename.io.fpPregRdy <> dispatch.io.fpPregRdy ++ dispatch.io.fpMemRegRdy
205  rename.io.replayPregReq <> dispatch.io.replayPregReq
206  dispatch.io.redirect <> redirect
207  dispatch.io.fromRename <> rename.io.out
208
209  roq.io.memRedirect <> io.mem.replayAll
210  roq.io.brqRedirect <> brq.io.redirect
211  roq.io.dp1Req <> dispatch.io.toRoq
212  dispatch.io.roqIdxs <> roq.io.roqIdxs
213  io.mem.dp1Req <> dispatch.io.toLsroq
214  dispatch.io.lsIdxs <> io.mem.lsIdxs
215  dispatch.io.dequeueRoqIndex.valid := roq.io.commitRoqIndex.valid || io.mem.oldestStore.valid
216  // store writeback must be after commit roqIdx
217  dispatch.io.dequeueRoqIndex.bits := Mux(io.mem.oldestStore.valid, io.mem.oldestStore.bits, roq.io.commitRoqIndex.bits)
218
219
220  intRf.io.readPorts <> dispatch.io.readIntRf
221  fpRf.io.readPorts <> dispatch.io.readFpRf ++ issueQueues.flatMap(_.io.readFpRf)
222  memRf.io.readPorts <> issueQueues.flatMap(_.io.readIntRf)
223
224  io.mem.redirect <> redirect
225
226  val wbu = Module(new Wbu(exuConfigs))
227  wbu.io.in <> exeWbReqs
228
229  val wbIntResults = wbu.io.toIntRf
230  val wbFpResults = wbu.io.toFpRf
231
232  def exuOutToRfWrite(x: Valid[ExuOutput]): RfWritePort = {
233    val rfWrite = Wire(new RfWritePort)
234    rfWrite.wen := x.valid
235    rfWrite.addr := x.bits.uop.pdest
236    rfWrite.data := x.bits.data
237    rfWrite
238  }
239  val intRfWrite = wbIntResults.map(exuOutToRfWrite)
240  intRf.io.writePorts <> intRfWrite
241  memRf.io.writePorts <> intRfWrite
242  fpRf.io.writePorts <> wbFpResults.map(exuOutToRfWrite)
243
244  rename.io.wbIntResults <> wbIntResults
245  rename.io.wbFpResults <> wbFpResults
246
247  roq.io.exeWbResults.take(exeWbReqs.length).zip(wbu.io.toRoq).foreach(x => x._1 := x._2)
248  roq.io.exeWbResults.last := brq.io.out
249
250
251  // TODO: Remove sink and source
252  val tmp = WireInit(0.U)
253  val sinks = Array[String](
254    "DTLBFINISH",
255    "DTLBPF",
256    "DTLBENABLE",
257    "perfCntCondMdcacheLoss",
258    "perfCntCondMl2cacheLoss",
259    "perfCntCondMdcacheHit",
260    "lsuMMIO",
261    "perfCntCondMl2cacheHit",
262    "perfCntCondMl2cacheReq",
263    "mtip",
264    "perfCntCondMdcacheReq",
265    "meip"
266  )
267  for (s <- sinks) {
268    BoringUtils.addSink(tmp, s)
269  }
270
271  val debugIntReg, debugFpReg = WireInit(VecInit(Seq.fill(32)(0.U(XLEN.W))))
272  BoringUtils.addSink(debugIntReg, "DEBUG_INT_ARCH_REG")
273  BoringUtils.addSink(debugFpReg, "DEBUG_FP_ARCH_REG")
274  val debugArchReg = WireInit(VecInit(debugIntReg ++ debugFpReg))
275  if (!env.FPGAPlatform) {
276    BoringUtils.addSource(debugArchReg, "difftestRegs")
277  }
278
279}
280