xref: /XiangShan/src/main/scala/xiangshan/mem/lsqueue/LoadQueueRAR.scala (revision f7063a43ab34da917ba6c670d21871314340c550)
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***************************************************************************************/
16package xiangshan.mem
17
18import chisel3._
19import chisel3.util._
20import org.chipsalliance.cde.config._
21import xiangshan._
22import xiangshan.backend.rob.RobPtr
23import xiangshan.cache._
24import utils._
25import utility._
26import xiangshan.backend.Bundles.DynInst
27
28class LoadQueueRAR(implicit p: Parameters) extends XSModule
29  with HasDCacheParameters
30  with HasCircularQueuePtrHelper
31  with HasLoadHelper
32  with HasPerfEvents
33{
34  val io = IO(new Bundle() {
35    // control
36    val redirect = Flipped(Valid(new Redirect))
37
38    // violation query
39    val query = Vec(LoadPipelineWidth, Flipped(new LoadNukeQueryIO))
40
41    // release cacheline
42    val release = Flipped(Valid(new Release))
43
44    // from VirtualLoadQueue
45    val ldWbPtr = Input(new LqPtr)
46
47    // global
48    val lqFull = Output(Bool())
49  })
50
51  println("LoadQueueRAR: size: " + LoadQueueRARSize)
52  //  LoadQueueRAR field
53  //  +-------+-------+-------+----------+
54  //  | Valid |  Uop  | PAddr | Released |
55  //  +-------+-------+-------+----------+
56  //
57  //  Field descriptions:
58  //  Allocated   : entry is valid.
59  //  MicroOp     : Micro-op
60  //  PAddr       : physical address.
61  //  Released    : DCache released.
62  //
63  val allocated = RegInit(VecInit(List.fill(LoadQueueRARSize)(false.B))) // The control signals need to explicitly indicate the initial value
64  val uop = Reg(Vec(LoadQueueRARSize, new DynInst))
65  val paddrModule = Module(new LqPAddrModule(
66    gen = UInt(PAddrBits.W),
67    numEntries = LoadQueueRARSize,
68    numRead = LoadPipelineWidth,
69    numWrite = LoadPipelineWidth,
70    numWBank = LoadQueueNWriteBanks,
71    numWDelay = 2,
72    numCamPort = LoadPipelineWidth
73  ))
74  paddrModule.io := DontCare
75  val released = RegInit(VecInit(List.fill(LoadQueueRARSize)(false.B)))
76  val bypassPAddr = Reg(Vec(LoadPipelineWidth, UInt(PAddrBits.W)))
77
78  // freeliset: store valid entries index.
79  // +---+---+--------------+-----+-----+
80  // | 0 | 1 |      ......  | n-2 | n-1 |
81  // +---+---+--------------+-----+-----+
82  val freeList = Module(new FreeList(
83    size = LoadQueueRARSize,
84    allocWidth = LoadPipelineWidth,
85    freeWidth = 4,
86    enablePreAlloc = true,
87    moduleName = "LoadQueueRAR freelist"
88  ))
89  freeList.io := DontCare
90
91  // Real-allocation: load_s2
92  // PAddr write needs 2 cycles, release signal should delay 1 cycle so that
93  // load enqueue can catch release.
94  val release1Cycle = io.release
95  val release2Cycle = RegNext(io.release)
96  val release2Cycle_dup_lsu = RegNext(io.release)
97
98  // LoadQueueRAR enqueue condition:
99  // There are still not completed load instructions before the current load instruction.
100  // (e.g. "not completed" means that load instruction get the data or exception).
101  val canEnqueue = io.query.map(_.req.valid)
102  val cancelEnqueue = io.query.map(_.req.bits.uop.robIdx.needFlush(io.redirect))
103  val hasNotWritebackedLoad = io.query.map(_.req.bits.uop.lqIdx).map(lqIdx => isAfter(lqIdx, io.ldWbPtr))
104  val needEnqueue = canEnqueue.zip(hasNotWritebackedLoad).zip(cancelEnqueue).map { case ((v, r), c) => v && r && !c }
105
106  // Allocate logic
107  val acceptedVec = Wire(Vec(LoadPipelineWidth, Bool()))
108  val enqIndexVec = Wire(Vec(LoadPipelineWidth, UInt()))
109
110  for ((enq, w) <- io.query.map(_.req).zipWithIndex) {
111    acceptedVec(w) := false.B
112    paddrModule.io.wen(w) := false.B
113    freeList.io.doAllocate(w) := false.B
114
115    freeList.io.allocateReq(w) := true.B
116
117    //  Allocate ready
118    val offset = PopCount(needEnqueue.take(w))
119    val canAccept = freeList.io.canAllocate(offset)
120    val enqIndex = freeList.io.allocateSlot(offset)
121    enq.ready := Mux(needEnqueue(w), canAccept, true.B)
122
123    enqIndexVec(w) := enqIndex
124    when (needEnqueue(w) && enq.ready) {
125      acceptedVec(w) := true.B
126
127      val debug_robIdx = enq.bits.uop.robIdx.asUInt
128      XSError(allocated(enqIndex), p"LoadQueueRAR: You can not write an valid entry! check: ldu $w, robIdx $debug_robIdx")
129
130      freeList.io.doAllocate(w) := true.B
131
132      //  Allocate new entry
133      allocated(enqIndex) := true.B
134
135      //  Write paddr
136      paddrModule.io.wen(w) := true.B
137      paddrModule.io.waddr(w) := enqIndex
138      paddrModule.io.wdata(w) := enq.bits.paddr
139      bypassPAddr(w) := enq.bits.paddr
140
141      //  Fill info
142      uop(enqIndex) := enq.bits.uop
143      released(enqIndex) :=
144        enq.bits.data_valid &&
145        (release2Cycle.valid &&
146        enq.bits.paddr(PAddrBits-1, DCacheLineOffset) === release2Cycle.bits.paddr(PAddrBits-1, DCacheLineOffset) ||
147        release1Cycle.valid &&
148        enq.bits.paddr(PAddrBits-1, DCacheLineOffset) === release1Cycle.bits.paddr(PAddrBits-1, DCacheLineOffset))
149    }
150  }
151
152  //  LoadQueueRAR deallocate
153  val freeMaskVec = Wire(Vec(LoadQueueRARSize, Bool()))
154
155  // init
156  freeMaskVec.map(e => e := false.B)
157
158  // when the loads that "older than" current load were writebacked,
159  // current load will be released.
160  for (i <- 0 until LoadQueueRARSize) {
161    val deqNotBlock = !isBefore(io.ldWbPtr, uop(i).lqIdx)
162    val needFlush = uop(i).robIdx.needFlush(io.redirect)
163
164    when (allocated(i) && (deqNotBlock || needFlush)) {
165      allocated(i) := false.B
166      freeMaskVec(i) := true.B
167    }
168  }
169
170  // if need replay revoke entry
171  val lastCanAccept = RegNext(acceptedVec)
172  val lastAllocIndex = RegNext(enqIndexVec)
173
174  for ((revoke, w) <- io.query.map(_.revoke).zipWithIndex) {
175    val revokeValid = revoke && lastCanAccept(w)
176    val revokeIndex = lastAllocIndex(w)
177
178    when (allocated(revokeIndex) && revokeValid) {
179      allocated(revokeIndex) := false.B
180      freeMaskVec(revokeIndex) := true.B
181    }
182  }
183
184  freeList.io.free := freeMaskVec.asUInt
185
186  // LoadQueueRAR Query
187  // Load-to-Load violation check condition:
188  // 1. Physical address match by CAM port.
189  // 2. release is set.
190  // 3. Younger than current load instruction.
191  val ldLdViolation = Wire(Vec(LoadPipelineWidth, Bool()))
192  val allocatedUInt = RegNext(allocated.asUInt)
193  for ((query, w) <- io.query.zipWithIndex) {
194    ldLdViolation(w) := false.B
195    paddrModule.io.releaseViolationMdata(w) := query.req.bits.paddr
196
197    query.resp.valid := RegNext(query.req.valid)
198    // Generate real violation mask
199    val robIdxMask = VecInit(uop.map(_.robIdx).map(isAfter(_, query.req.bits.uop.robIdx)))
200    val matchMask = (0 until LoadQueueRARSize).map(i => {
201                      RegNext(allocated(i) &
202                      paddrModule.io.releaseViolationMmask(w)(i) &
203                      robIdxMask(i) &&
204                      released(i))
205                    })
206    //  Load-to-Load violation check result
207    val ldLdViolationMask = VecInit(matchMask)
208    ldLdViolationMask.suggestName("ldLdViolationMask_" + w)
209    query.resp.bits.rep_frm_fetch := ParallelORR(ldLdViolationMask)
210  }
211
212
213  // When io.release.valid (release1cycle.valid), it uses the last ld-ld paddr cam port to
214  // update release flag in 1 cycle
215  val releaseVioMask = Reg(Vec(LoadQueueRARSize, Bool()))
216  when (release1Cycle.valid) {
217    paddrModule.io.releaseMdata.takeRight(1)(0) := release1Cycle.bits.paddr
218  }
219
220  val lastAllocIndexOH = lastAllocIndex.map(UIntToOH(_))
221  val lastReleasePAddrMatch = VecInit((0 until LoadPipelineWidth).map(i => {
222    (bypassPAddr(i)(PAddrBits-1, DCacheLineOffset) === release1Cycle.bits.paddr(PAddrBits-1, DCacheLineOffset))
223  }))
224  (0 until LoadQueueRARSize).map(i => {
225    val bypassMatch = VecInit((0 until LoadPipelineWidth).map(j => lastCanAccept(j) && lastAllocIndexOH(j)(i) && lastReleasePAddrMatch(j))).asUInt.orR
226    when (RegNext((paddrModule.io.releaseMmask.takeRight(1)(0)(i) || bypassMatch) && allocated(i) && release1Cycle.valid)) {
227      // Note: if a load has missed in dcache and is waiting for refill in load queue,
228      // its released flag still needs to be set as true if addr matches.
229      released(i) := true.B
230    }
231  })
232
233  io.lqFull := freeList.io.empty
234
235  // perf cnt
236  val canEnqCount = PopCount(io.query.map(_.req.fire))
237  val validCount = freeList.io.validCount
238  val allowEnqueue = validCount <= (LoadQueueRARSize - LoadPipelineWidth).U
239  val ldLdViolationCount = PopCount(io.query.map(_.resp).map(resp => resp.valid && resp.bits.rep_frm_fetch))
240
241  QueuePerf(LoadQueueRARSize, validCount, !allowEnqueue)
242  XSPerfAccumulate("enq", canEnqCount)
243  XSPerfAccumulate("ld_ld_violation", ldLdViolationCount)
244  val perfEvents: Seq[(String, UInt)] = Seq(
245    ("enq", canEnqCount),
246    ("ld_ld_violation", ldLdViolationCount)
247  )
248  generatePerfEvent()
249  // End
250}