xref: /XiangShan/src/main/scala/xiangshan/mem/lsqueue/LoadQueue.scala (revision deb6421e9ab9b7980dc6c429456fc7bd2161357b)
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.mem
18
19import chipsalliance.rocketchip.config.Parameters
20import chisel3._
21import chisel3.util._
22import utils._
23import xiangshan._
24import xiangshan.backend.fu.fpu.FPU
25import xiangshan.backend.rob.RobLsqIO
26import xiangshan.cache._
27import xiangshan.frontend.FtqPtr
28import xiangshan.ExceptionNO._
29import chisel3.ExcitingUtils
30
31class LqPtr(implicit p: Parameters) extends CircularQueuePtr[LqPtr](
32  p => p(XSCoreParamsKey).LoadQueueSize
33){
34}
35
36object LqPtr {
37  def apply(f: Bool, v: UInt)(implicit p: Parameters): LqPtr = {
38    val ptr = Wire(new LqPtr)
39    ptr.flag := f
40    ptr.value := v
41    ptr
42  }
43}
44
45trait HasLoadHelper { this: XSModule =>
46  def rdataHelper(uop: MicroOp, rdata: UInt): UInt = {
47    val fpWen = uop.ctrl.fpWen
48    LookupTree(uop.ctrl.fuOpType, List(
49      LSUOpType.lb   -> SignExt(rdata(7, 0) , XLEN),
50      LSUOpType.lh   -> SignExt(rdata(15, 0), XLEN),
51      /*
52          riscv-spec-20191213: 12.2 NaN Boxing of Narrower Values
53          Any operation that writes a narrower result to an f register must write
54          all 1s to the uppermost FLEN−n bits to yield a legal NaN-boxed value.
55      */
56      LSUOpType.lw   -> Mux(fpWen, FPU.box(rdata, FPU.S), SignExt(rdata(31, 0), XLEN)),
57      LSUOpType.ld   -> Mux(fpWen, FPU.box(rdata, FPU.D), SignExt(rdata(63, 0), XLEN)),
58      LSUOpType.lbu  -> ZeroExt(rdata(7, 0) , XLEN),
59      LSUOpType.lhu  -> ZeroExt(rdata(15, 0), XLEN),
60      LSUOpType.lwu  -> ZeroExt(rdata(31, 0), XLEN),
61    ))
62  }
63}
64
65class LqEnqIO(implicit p: Parameters) extends XSBundle {
66  val canAccept = Output(Bool())
67  val sqCanAccept = Input(Bool())
68  val needAlloc = Vec(exuParameters.LsExuCnt, Input(Bool()))
69  val req = Vec(exuParameters.LsExuCnt, Flipped(ValidIO(new MicroOp)))
70  val resp = Vec(exuParameters.LsExuCnt, Output(new LqPtr))
71}
72
73class LqPaddrWriteBundle(implicit p: Parameters) extends XSBundle {
74  val paddr = Output(UInt(PAddrBits.W))
75  val lqIdx = Output(new LqPtr)
76}
77
78class LqVaddrWriteBundle(implicit p: Parameters) extends XSBundle {
79  val vaddr = Output(UInt(VAddrBits.W))
80  val lqIdx = Output(new LqPtr)
81}
82
83class LqTriggerIO(implicit p: Parameters) extends XSBundle {
84  val hitLoadAddrTriggerHitVec = Input(Vec(3, Bool()))
85  val lqLoadAddrTriggerHitVec = Output(Vec(3, Bool()))
86}
87
88class LoadQueueIOBundle(implicit p: Parameters) extends XSBundle {
89  val enq = new LqEnqIO
90  val brqRedirect = Flipped(ValidIO(new Redirect))
91  val loadOut = Vec(LoadPipelineWidth, Decoupled(new LsPipelineBundle)) // select load from lq to load pipeline
92  val loadPaddrIn = Vec(LoadPipelineWidth, Flipped(Valid(new LqPaddrWriteBundle)))
93  val loadVaddrIn = Vec(LoadPipelineWidth, Flipped(Valid(new LqVaddrWriteBundle)))
94  val loadIn = Vec(LoadPipelineWidth, Flipped(Valid(new LqWriteBundle)))
95  val storeIn = Vec(StorePipelineWidth, Flipped(Valid(new LsPipelineBundle)))
96  val s2_load_data_forwarded = Vec(LoadPipelineWidth, Input(Bool()))
97  val s3_delayed_load_error = Vec(LoadPipelineWidth, Input(Bool()))
98  val s2_dcache_require_replay = Vec(LoadPipelineWidth, Input(Bool()))
99  val s3_replay_from_fetch = Vec(LoadPipelineWidth, Input(Bool()))
100  val ldout = Vec(2, DecoupledIO(new ExuOutput)) // writeback int load
101  val ldRawDataOut = Vec(2, Output(new LoadDataFromLQBundle))
102  val load_s1 = Vec(LoadPipelineWidth, Flipped(new PipeLoadForwardQueryIO)) // TODO: to be renamed
103  val loadViolationQuery = Vec(LoadPipelineWidth, Flipped(new LoadViolationQueryIO))
104  val rob = Flipped(new RobLsqIO)
105  val rollback = Output(Valid(new Redirect)) // replay now starts from load instead of store
106  val refill = Flipped(ValidIO(new Refill)) // TODO: to be renamed
107  val release = Flipped(ValidIO(new Release))
108  val uncache = new UncacheWordIO
109  val exceptionAddr = new ExceptionAddrIO
110  val lqFull = Output(Bool())
111  val lqCancelCnt = Output(UInt(log2Up(LoadQueueSize + 1).W))
112  val trigger = Vec(LoadPipelineWidth, new LqTriggerIO)
113
114  // for load replay (recieve feedback from load pipe line)
115  val replayFast = Vec(LoadPipelineWidth, Flipped(new LoadToLsqFastIO))
116  val replaySlow = Vec(LoadPipelineWidth, Flipped(new LoadToLsqSlowIO))
117
118  val storeDataValidVec = Vec(StoreQueueSize, Input(Bool()))
119
120  val tlbReplayDelayCycleCtrl = Vec(4, Input(UInt(ReSelectLen.W)))
121}
122
123// Load Queue
124class LoadQueue(implicit p: Parameters) extends XSModule
125  with HasDCacheParameters
126  with HasCircularQueuePtrHelper
127  with HasLoadHelper
128  with HasPerfEvents
129{
130  val io = IO(new LoadQueueIOBundle())
131
132  // dontTouch(io)
133
134  println("LoadQueue: size:" + LoadQueueSize)
135
136  val uop = Reg(Vec(LoadQueueSize, new MicroOp))
137  // val data = Reg(Vec(LoadQueueSize, new LsRobEntry))
138  val dataModule = Module(new LoadQueueDataWrapper(LoadQueueSize, wbNumRead = LoadPipelineWidth, wbNumWrite = LoadPipelineWidth))
139  dataModule.io := DontCare
140  // vaddrModule's read port 0 for exception addr, port {1, 2} for debug module, port {3, 4} for load replay
141  val vaddrModule = Module(new SyncDataModuleTemplate(UInt(VAddrBits.W), LoadQueueSize, numRead = LoadPipelineWidth + 1 + 2, numWrite = LoadPipelineWidth))
142  vaddrModule.io := DontCare
143  val vaddrTriggerResultModule = Module(new SyncDataModuleTemplate(Vec(3, Bool()), LoadQueueSize, numRead = LoadPipelineWidth, numWrite = LoadPipelineWidth))
144  vaddrTriggerResultModule.io := DontCare
145  val allocated = RegInit(VecInit(List.fill(LoadQueueSize)(false.B))) // lq entry has been allocated
146  val datavalid = RegInit(VecInit(List.fill(LoadQueueSize)(false.B))) // data is valid
147  val writebacked = RegInit(VecInit(List.fill(LoadQueueSize)(false.B))) // inst has been writebacked to CDB
148  val released = RegInit(VecInit(List.fill(LoadQueueSize)(false.B))) // load data has been released by dcache
149  val error = RegInit(VecInit(List.fill(LoadQueueSize)(false.B))) // load data has been corrupted
150  val miss = Reg(Vec(LoadQueueSize, Bool())) // load inst missed, waiting for miss queue to accept miss request
151  // val listening = Reg(Vec(LoadQueueSize, Bool())) // waiting for refill result
152  val pending = Reg(Vec(LoadQueueSize, Bool())) // mmio pending: inst is an mmio inst, it will not be executed until it reachs the end of rob
153  val refilling = WireInit(VecInit(List.fill(LoadQueueSize)(false.B))) // inst has been writebacked to CDB
154
155  /**
156    * used for load replay control
157    */
158
159  val tlb_hited = RegInit(VecInit(List.fill(LoadQueueSize)(true.B)))
160  val ld_ld_check_ok = RegInit(VecInit(List.fill(LoadQueueSize)(true.B)))
161  val st_ld_check_ok = RegInit(VecInit(List.fill(LoadQueueSize)(true.B)))
162  val cache_bank_no_conflict = RegInit(VecInit(List.fill(LoadQueueSize)(true.B)))
163  val cache_no_replay = RegInit(VecInit(List.fill(LoadQueueSize)(true.B)))
164  val forward_data_valid = RegInit(VecInit(List.fill(LoadQueueSize)(true.B)))
165
166
167  /**
168    * used for re-select control
169    */
170
171  val credit = RegInit(VecInit(List.fill(LoadQueueSize)(0.U(ReSelectLen.W))))
172
173  // ptrs to control which cycle to choose
174  val block_ptr_tlb = RegInit(VecInit(List.fill(LoadQueueSize)(0.U(2.W))))
175  val block_ptr_others = RegInit(VecInit(List.fill(LoadQueueSize)(0.U(2.W))))
176
177  // specific cycles to block
178  val block_cycles_tlb = Reg(Vec(4, UInt(ReSelectLen.W)))
179  block_cycles_tlb := io.tlbReplayDelayCycleCtrl
180  val block_cycles_others = RegInit(VecInit(Seq(0.U(ReSelectLen.W), 0.U(ReSelectLen.W), 0.U(ReSelectLen.W), 0.U(ReSelectLen.W))))
181
182  val sel_blocked = RegInit(VecInit(List.fill(LoadQueueSize)(false.B)))
183
184  // data forward block
185  val block_sq_idx = RegInit(VecInit(List.fill(LoadQueueSize)(0.U((log2Ceil(StoreQueueSize).W)))))
186  val block_by_data_forward_fail = RegInit(VecInit(List.fill(LoadQueueSize)(false.B)))
187
188  val creditUpdate = WireInit(VecInit(List.fill(LoadQueueSize)(0.U(ReSelectLen.W))))
189
190  credit := creditUpdate
191
192  (0 until LoadQueueSize).map(i => {
193    creditUpdate(i) := Mux(credit(i) > 0.U(ReSelectLen.W), credit(i) - 1.U(ReSelectLen.W), credit(i))
194    sel_blocked(i) := creditUpdate(i) =/= 0.U(ReSelectLen.W) || credit(i) =/= 0.U(ReSelectLen.W)
195  })
196
197  (0 until LoadQueueSize).map(i => {
198    block_by_data_forward_fail(i) := Mux(block_by_data_forward_fail(i) === true.B && io.storeDataValidVec(block_sq_idx(i)) === true.B , false.B, block_by_data_forward_fail(i))
199  })
200
201  val debug_mmio = Reg(Vec(LoadQueueSize, Bool())) // mmio: inst is an mmio inst
202  val debug_paddr = Reg(Vec(LoadQueueSize, UInt(PAddrBits.W))) // mmio: inst is an mmio inst
203
204  val enqPtrExt = RegInit(VecInit((0 until io.enq.req.length).map(_.U.asTypeOf(new LqPtr))))
205  val deqPtrExt = RegInit(0.U.asTypeOf(new LqPtr))
206  val deqPtrExtNext = Wire(new LqPtr)
207
208  val enqPtr = enqPtrExt(0).value
209  val deqPtr = deqPtrExt.value
210
211  val validCount = distanceBetween(enqPtrExt(0), deqPtrExt)
212  val allowEnqueue = validCount <= (LoadQueueSize - LoadPipelineWidth).U
213
214  val deqMask = UIntToMask(deqPtr, LoadQueueSize)
215  val enqMask = UIntToMask(enqPtr, LoadQueueSize)
216
217  val commitCount = RegNext(io.rob.lcommit)
218
219  val release1cycle = io.release
220  val release2cycle = RegNext(io.release)
221  val release2cycle_dup_lsu = RegNext(io.release)
222
223  /**
224    * Enqueue at dispatch
225    *
226    * Currently, LoadQueue only allows enqueue when #emptyEntries > EnqWidth
227    */
228  io.enq.canAccept := allowEnqueue
229
230  val canEnqueue = io.enq.req.map(_.valid)
231  val enqCancel = io.enq.req.map(_.bits.robIdx.needFlush(io.brqRedirect))
232  for (i <- 0 until io.enq.req.length) {
233    val offset = if (i == 0) 0.U else PopCount(io.enq.needAlloc.take(i))
234    val lqIdx = enqPtrExt(offset)
235    val index = io.enq.req(i).bits.lqIdx.value
236    when (canEnqueue(i) && !enqCancel(i)) {
237      uop(index) := io.enq.req(i).bits
238      // NOTE: the index will be used when replay
239      uop(index).lqIdx := lqIdx
240      allocated(index) := true.B
241      datavalid(index) := false.B
242      writebacked(index) := false.B
243      released(index) := false.B
244      miss(index) := false.B
245      pending(index) := false.B
246      error(index) := false.B
247
248      /**
249        * used for load replay control
250        */
251      tlb_hited(index) := true.B
252      ld_ld_check_ok(index) := true.B
253      st_ld_check_ok(index) := true.B
254      cache_bank_no_conflict(index) := true.B
255      cache_no_replay(index) := true.B
256      forward_data_valid(index) := true.B
257
258      /**
259        * used for delaying load(block-ptr to control how many cycles to block)
260        */
261      credit(index) := 0.U(ReSelectLen.W)
262      block_ptr_tlb(index) := 0.U(2.W)
263      block_ptr_others(index) := 0.U(2.W)
264
265      block_by_data_forward_fail(index) := false.B
266
267      XSError(!io.enq.canAccept || !io.enq.sqCanAccept, s"must accept $i\n")
268      XSError(index =/= lqIdx.value, s"must be the same entry $i\n")
269    }
270    io.enq.resp(i) := lqIdx
271  }
272  XSDebug(p"(ready, valid): ${io.enq.canAccept}, ${Binary(Cat(io.enq.req.map(_.valid)))}\n")
273
274  val lastCycleRedirect = RegNext(io.brqRedirect)
275  val lastlastCycleRedirect = RegNext(lastCycleRedirect)
276
277  // replay logic
278  // replay is splited into 2 stages
279
280  // stage1: select 2 entries and read their vaddr
281  val s0_block_load_mask = WireInit(VecInit((0 until LoadQueueSize).map(x=>false.B)))
282  val s1_block_load_mask = RegNext(s0_block_load_mask)
283  val s2_block_load_mask = RegNext(s1_block_load_mask)
284
285  val loadReplaySel = Wire(Vec(LoadPipelineWidth, UInt(log2Up(LoadQueueSize).W))) // index selected last cycle
286  val loadReplaySelV = Wire(Vec(LoadPipelineWidth, Bool())) // index selected in last cycle is valid
287
288  val loadReplaySelVec = VecInit((0 until LoadQueueSize).map(i => {
289    val blocked = s1_block_load_mask(i) || s2_block_load_mask(i) || sel_blocked(i) || block_by_data_forward_fail(i)
290    allocated(i) && (!tlb_hited(i) || !ld_ld_check_ok(i) || !st_ld_check_ok(i) || !cache_bank_no_conflict(i) || !cache_no_replay(i) || !forward_data_valid(i)) && !blocked
291  })).asUInt() // use uint instead vec to reduce verilog lines
292
293  val remReplayDeqMask = Seq.tabulate(LoadPipelineWidth)(getRemBits(deqMask)(_))
294
295  // generate lastCycleSelect mask
296  val remReplayFireMask = Seq.tabulate(LoadPipelineWidth)(rem => getRemBits(UIntToOH(loadReplaySel(rem)))(rem))
297
298  val loadReplayRemSelVecFire = Seq.tabulate(LoadPipelineWidth)(rem => getRemBits(loadReplaySelVec)(rem) & ~remReplayFireMask(rem))
299  val loadReplayRemSelVecNotFire = Seq.tabulate(LoadPipelineWidth)(getRemBits(loadReplaySelVec)(_))
300
301  val replayRemFire = Seq.tabulate(LoadPipelineWidth)(rem => WireInit(false.B))
302
303  val loadReplayRemSel = Seq.tabulate(LoadPipelineWidth)(rem => Mux(
304    replayRemFire(rem),
305    getFirstOne(toVec(loadReplayRemSelVecFire(rem)), remReplayDeqMask(rem)),
306    getFirstOne(toVec(loadReplayRemSelVecNotFire(rem)), remReplayDeqMask(rem))
307  ))
308
309  val loadReplaySelGen = Wire(Vec(LoadPipelineWidth, UInt(log2Up(LoadQueueSize).W)))
310  val loadReplaySelVGen = Wire(Vec(LoadPipelineWidth, Bool()))
311
312  (0 until LoadPipelineWidth).foreach(index => {
313    loadReplaySelGen(index) := (
314      if (LoadPipelineWidth > 1) Cat(loadReplayRemSel(index), index.U(log2Ceil(LoadPipelineWidth).W))
315      else loadReplayRemSel(index)
316    )
317    loadReplaySelVGen(index) := Mux(replayRemFire(index), loadReplayRemSelVecFire(index).asUInt.orR, loadReplayRemSelVecNotFire(index).asUInt.orR)
318  })
319
320  (0 until LoadPipelineWidth).map(i => {
321    vaddrModule.io.raddr(LoadPipelineWidth + 1 + i) := loadReplaySelGen(i)
322  })
323
324  (0 until LoadPipelineWidth).map(i => {
325    loadReplaySel(i) := RegNext(loadReplaySelGen(i))
326    loadReplaySelV(i) := RegNext(loadReplaySelVGen(i), init = false.B)
327  })
328
329  // stage2: replay to load pipeline (if no load in S0)
330  (0 until LoadPipelineWidth).map(i => {
331    when(replayRemFire(i)) {
332      s0_block_load_mask(loadReplaySel(i)) := true.B
333    }
334  })
335
336  // init
337  (0 until LoadPipelineWidth).map(i => {
338    replayRemFire(i) := false.B
339  })
340
341  for(i <- 0 until LoadPipelineWidth) {
342    val replayIdx = loadReplaySel(i)
343    val notRedirectLastCycle = !uop(replayIdx).robIdx.needFlush(RegNext(io.brqRedirect))
344
345    io.loadOut(i).valid := loadReplaySelV(i) && notRedirectLastCycle
346
347    io.loadOut(i).bits := DontCare
348    io.loadOut(i).bits.uop := uop(replayIdx)
349    io.loadOut(i).bits.vaddr := vaddrModule.io.rdata(LoadPipelineWidth + 1 + i)
350    io.loadOut(i).bits.mask := genWmask(vaddrModule.io.rdata(LoadPipelineWidth + 1 + i), uop(replayIdx).ctrl.fuOpType(1,0))
351    io.loadOut(i).bits.isFirstIssue := false.B
352    io.loadOut(i).bits.isLoadReplay := true.B
353
354    when(io.loadOut(i).fire) {
355      replayRemFire(i) := true.B
356    }
357
358  }
359  /**
360    * Writeback load from load units
361    *
362    * Most load instructions writeback to regfile at the same time.
363    * However,
364    *   (1) For an mmio instruction with exceptions, it writes back to ROB immediately.
365    *   (2) For an mmio instruction without exceptions, it does not write back.
366    * The mmio instruction will be sent to lower level when it reaches ROB's head.
367    * After uncache response, it will write back through arbiter with loadUnit.
368    *   (3) For cache misses, it is marked miss and sent to dcache later.
369    * After cache refills, it will write back through arbiter with loadUnit.
370    */
371  for (i <- 0 until LoadPipelineWidth) {
372    dataModule.io.wb.wen(i) := false.B
373    dataModule.io.paddr.wen(i) := false.B
374    vaddrModule.io.wen(i) := false.B
375    vaddrTriggerResultModule.io.wen(i) := false.B
376    val loadWbIndex = io.loadIn(i).bits.uop.lqIdx.value
377
378    // most lq status need to be updated immediately after load writeback to lq
379    // flag bits in lq needs to be updated accurately
380    when(io.loadIn(i).fire()) {
381      when(io.loadIn(i).bits.miss) {
382        XSInfo(io.loadIn(i).valid, "load miss write to lq idx %d pc 0x%x vaddr %x paddr %x mask %x forwardData %x forwardMask: %x mmio %x\n",
383          io.loadIn(i).bits.uop.lqIdx.asUInt,
384          io.loadIn(i).bits.uop.cf.pc,
385          io.loadIn(i).bits.vaddr,
386          io.loadIn(i).bits.paddr,
387          io.loadIn(i).bits.mask,
388          io.loadIn(i).bits.forwardData.asUInt,
389          io.loadIn(i).bits.forwardMask.asUInt,
390          io.loadIn(i).bits.mmio
391        )
392      }.otherwise {
393        XSInfo(io.loadIn(i).valid, "load hit write to cbd lqidx %d pc 0x%x vaddr %x paddr %x mask %x forwardData %x forwardMask: %x mmio %x\n",
394        io.loadIn(i).bits.uop.lqIdx.asUInt,
395        io.loadIn(i).bits.uop.cf.pc,
396        io.loadIn(i).bits.vaddr,
397        io.loadIn(i).bits.paddr,
398        io.loadIn(i).bits.mask,
399        io.loadIn(i).bits.forwardData.asUInt,
400        io.loadIn(i).bits.forwardMask.asUInt,
401        io.loadIn(i).bits.mmio
402      )}
403      if(EnableFastForward){
404        datavalid(loadWbIndex) := (!io.loadIn(i).bits.miss || io.s2_load_data_forwarded(i)) &&
405          !io.loadIn(i).bits.mmio && // mmio data is not valid until we finished uncache access
406          !io.s2_dcache_require_replay(i) // do not writeback if that inst will be resend from rs
407      } else {
408        datavalid(loadWbIndex) := (!io.loadIn(i).bits.miss || io.s2_load_data_forwarded(i)) &&
409          !io.loadIn(i).bits.mmio // mmio data is not valid until we finished uncache access
410      }
411      writebacked(loadWbIndex) := !io.loadIn(i).bits.miss && !io.loadIn(i).bits.mmio
412
413      debug_mmio(loadWbIndex) := io.loadIn(i).bits.mmio
414      debug_paddr(loadWbIndex) := io.loadIn(i).bits.paddr
415
416      val dcacheMissed = io.loadIn(i).bits.miss && !io.loadIn(i).bits.mmio
417      if(EnableFastForward){
418        miss(loadWbIndex) := dcacheMissed && !io.s2_load_data_forwarded(i) && !io.s2_dcache_require_replay(i)
419      } else {
420        miss(loadWbIndex) := dcacheMissed && !io.s2_load_data_forwarded(i)
421      }
422      pending(loadWbIndex) := io.loadIn(i).bits.mmio
423      released(loadWbIndex) := release2cycle.valid &&
424        io.loadIn(i).bits.paddr(PAddrBits-1, DCacheLineOffset) === release2cycle.bits.paddr(PAddrBits-1, DCacheLineOffset) ||
425        release1cycle.valid &&
426        io.loadIn(i).bits.paddr(PAddrBits-1, DCacheLineOffset) === release1cycle.bits.paddr(PAddrBits-1, DCacheLineOffset)
427    }
428
429    // data bit in lq can be updated when load_s2 valid
430    // when(io.loadIn(i).bits.lq_data_wen){
431    //   val loadWbData = Wire(new LQDataEntry)
432    //   loadWbData.paddr := io.loadIn(i).bits.paddr
433    //   loadWbData.mask := io.loadIn(i).bits.mask
434    //   loadWbData.data := io.loadIn(i).bits.forwardData.asUInt // fwd data
435    //   loadWbData.fwdMask := io.loadIn(i).bits.forwardMask
436    //   dataModule.io.wbWrite(i, loadWbIndex, loadWbData)
437    //   dataModule.io.wb.wen(i) := true.B
438
439    //   // dirty code for load instr
440    //   uop(loadWbIndex).pdest := io.loadIn(i).bits.uop.pdest
441    //   uop(loadWbIndex).cf := io.loadIn(i).bits.uop.cf
442    //   uop(loadWbIndex).ctrl := io.loadIn(i).bits.uop.ctrl
443    //   uop(loadWbIndex).debugInfo := io.loadIn(i).bits.uop.debugInfo
444
445    //   vaddrTriggerResultModule.io.waddr(i) := loadWbIndex
446    //   vaddrTriggerResultModule.io.wdata(i) := io.trigger(i).hitLoadAddrTriggerHitVec
447
448    //   vaddrTriggerResultModule.io.wen(i) := true.B
449    // }
450
451    // dirty code to reduce load_s2.valid fanout
452    when(io.loadIn(i).bits.lq_data_wen_dup(0)){
453      val loadWbData = Wire(new LQDataEntry)
454      loadWbData.paddr := io.loadIn(i).bits.paddr
455      loadWbData.mask := io.loadIn(i).bits.mask
456      loadWbData.data := io.loadIn(i).bits.forwardData.asUInt // fwd data
457      loadWbData.fwdMask := io.loadIn(i).bits.forwardMask
458      dataModule.io.wbWrite(i, loadWbIndex, loadWbData)
459      dataModule.io.wb.wen(i) := true.B
460    }
461    // dirty code for load instr
462    when(io.loadIn(i).bits.lq_data_wen_dup(1)){
463      uop(loadWbIndex).pdest := io.loadIn(i).bits.uop.pdest
464    }
465    when(io.loadIn(i).bits.lq_data_wen_dup(2)){
466      uop(loadWbIndex).cf := io.loadIn(i).bits.uop.cf
467    }
468    when(io.loadIn(i).bits.lq_data_wen_dup(3)){
469      uop(loadWbIndex).ctrl := io.loadIn(i).bits.uop.ctrl
470    }
471    when(io.loadIn(i).bits.lq_data_wen_dup(4)){
472      uop(loadWbIndex).debugInfo := io.loadIn(i).bits.uop.debugInfo
473    }
474    when(io.loadIn(i).bits.lq_data_wen_dup(5)){
475      vaddrTriggerResultModule.io.waddr(i) := loadWbIndex
476      vaddrTriggerResultModule.io.wdata(i) := io.trigger(i).hitLoadAddrTriggerHitVec
477      vaddrTriggerResultModule.io.wen(i) := true.B
478    }
479
480    when(io.loadPaddrIn(i).valid) {
481      dataModule.io.paddr.wen(i) := true.B
482      dataModule.io.paddr.waddr(i) := io.loadPaddrIn(i).bits.lqIdx.value
483      dataModule.io.paddr.wdata(i) := io.loadPaddrIn(i).bits.paddr
484    }
485
486    // update vaddr in load S1
487    when(io.loadVaddrIn(i).valid) {
488      vaddrModule.io.wen(i) := true.B
489      vaddrModule.io.waddr(i) := io.loadVaddrIn(i).bits.lqIdx.value
490      vaddrModule.io.wdata(i) := io.loadVaddrIn(i).bits.vaddr
491    }
492
493    /**
494      * used for feedback and replay
495      */
496    when(io.replayFast(i).valid){
497      val idx = io.replayFast(i).ld_idx
498      val needreplay = !io.replayFast(i).ld_ld_check_ok || !io.replayFast(i).st_ld_check_ok || !io.replayFast(i).cache_bank_no_conflict
499
500      ld_ld_check_ok(idx) := io.replayFast(i).ld_ld_check_ok
501      st_ld_check_ok(idx) := io.replayFast(i).st_ld_check_ok
502      cache_bank_no_conflict(idx) := io.replayFast(i).cache_bank_no_conflict
503
504      when(needreplay) {
505        creditUpdate(idx) := block_cycles_others(block_ptr_others(idx))
506        block_ptr_others(idx) := Mux(block_ptr_others(idx) === 3.U(2.W), block_ptr_others(idx), block_ptr_others(idx) + 1.U(2.W))
507        // try to replay this load in next cycle
508        s1_block_load_mask(idx) := false.B
509        s2_block_load_mask(idx) := false.B
510
511        // replay this load in next cycle
512        loadReplaySelGen(idx(log2Ceil(LoadPipelineWidth) - 1, 0)) := idx
513        loadReplaySelVGen(idx(log2Ceil(LoadPipelineWidth) - 1, 0)) := true.B
514      }
515    }
516
517    when(io.replaySlow(i).valid){
518      val idx = io.replaySlow(i).ld_idx
519      val needreplay = !io.replaySlow(i).tlb_hited || !io.replaySlow(i).st_ld_check_ok || !io.replaySlow(i).cache_no_replay || !io.replaySlow(i).forward_data_valid
520
521      tlb_hited(idx) := io.replaySlow(i).tlb_hited
522      st_ld_check_ok(idx) := io.replaySlow(i).st_ld_check_ok
523      cache_no_replay(idx) := io.replaySlow(i).cache_no_replay
524      forward_data_valid(idx) := io.replaySlow(i).forward_data_valid
525
526      val invalid_sq_idx = io.replaySlow(i).data_invalid_sq_idx
527
528      when(needreplay) {
529        creditUpdate(idx) := Mux( !io.replaySlow(i).tlb_hited, block_cycles_tlb(block_ptr_tlb(idx)),
530                              Mux(!io.replaySlow(i).cache_no_replay || !io.replaySlow(i).st_ld_check_ok, block_cycles_others(block_ptr_others(idx)), 0.U))
531        when(!io.replaySlow(i).tlb_hited) {
532          block_ptr_tlb(idx) := Mux(block_ptr_tlb(idx) === 3.U(2.W), block_ptr_tlb(idx), block_ptr_tlb(idx) + 1.U(2.W))
533        }.elsewhen(!io.replaySlow(i).cache_no_replay || !io.replaySlow(i).st_ld_check_ok) {
534          block_ptr_others(idx) := Mux(block_ptr_others(idx) === 3.U(2.W), block_ptr_others(idx), block_ptr_others(idx) + 1.U(2.W))
535        }
536      }
537
538      block_by_data_forward_fail(idx) := false.B
539
540      when(!io.replaySlow(i).forward_data_valid) {
541        when(!io.storeDataValidVec(invalid_sq_idx)) {
542          block_by_data_forward_fail(idx) := true.B
543          block_sq_idx(idx) := invalid_sq_idx
544        }
545      }
546    }
547
548  }
549
550  when(io.refill.valid) {
551    XSDebug("miss resp: paddr:0x%x data %x\n", io.refill.bits.addr, io.refill.bits.data)
552  }
553
554  // Refill 64 bit in a cycle
555  // Refill data comes back from io.dcache.resp
556  dataModule.io.refill.valid := io.refill.valid
557  dataModule.io.refill.paddr := io.refill.bits.addr
558  dataModule.io.refill.data := io.refill.bits.data
559
560  val s2_dcache_require_replay = WireInit(VecInit((0 until LoadPipelineWidth).map(i =>{
561    RegNext(io.loadIn(i).fire()) && RegNext(io.s2_dcache_require_replay(i))
562  })))
563  dontTouch(s2_dcache_require_replay)
564
565  (0 until LoadQueueSize).map(i => {
566    dataModule.io.refill.refillMask(i) := allocated(i) && miss(i)
567    when(dataModule.io.refill.valid && dataModule.io.refill.refillMask(i) && dataModule.io.refill.matchMask(i)) {
568      datavalid(i) := true.B
569      miss(i) := false.B
570      when(!s2_dcache_require_replay.asUInt.orR){
571        refilling(i) := true.B
572      }
573      when(io.refill.bits.error) {
574        error(i) := true.B
575      }
576    }
577  })
578
579  for (i <- 0 until LoadPipelineWidth) {
580    val loadWbIndex = io.loadIn(i).bits.uop.lqIdx.value
581    val lastCycleLoadWbIndex = RegNext(loadWbIndex)
582    // update miss state in load s3
583    if(!EnableFastForward){
584      // s2_dcache_require_replay will be used to update lq flag 1 cycle after for better timing
585      //
586      // io.s2_dcache_require_replay comes from dcache miss req reject, which is quite slow to generate
587      when(s2_dcache_require_replay(i)) {
588        // do not writeback if that inst will be resend from rs
589        // rob writeback will not be triggered by a refill before inst replay
590        miss(lastCycleLoadWbIndex) := false.B // disable refill listening
591        datavalid(lastCycleLoadWbIndex) := false.B // disable refill listening
592        assert(!datavalid(lastCycleLoadWbIndex))
593      }
594    }
595    // update load error state in load s3
596    when(RegNext(io.loadIn(i).fire()) && io.s3_delayed_load_error(i)){
597      uop(lastCycleLoadWbIndex).cf.exceptionVec(loadAccessFault) := true.B
598    }
599    // update inst replay from fetch flag in s3
600    when(RegNext(io.loadIn(i).fire()) && io.s3_replay_from_fetch(i)){
601      uop(lastCycleLoadWbIndex).ctrl.replayInst := true.B
602    }
603  }
604
605
606  // Writeback up to 2 missed load insts to CDB
607  //
608  // Pick 2 missed load (data refilled), write them back to cdb
609  // 2 refilled load will be selected from even/odd entry, separately
610
611  // Stage 0
612  // Generate writeback indexes
613
614  def getRemBits(input: UInt)(rem: Int): UInt = {
615    VecInit((0 until LoadQueueSize / LoadPipelineWidth).map(i => { input(LoadPipelineWidth * i + rem) })).asUInt
616  }
617
618  val loadWbSel = Wire(Vec(LoadPipelineWidth, UInt(log2Up(LoadQueueSize).W))) // index selected last cycle
619  val loadWbSelV = Wire(Vec(LoadPipelineWidth, Bool())) // index selected in last cycle is valid
620
621  val loadWbSelVec = VecInit((0 until LoadQueueSize).map(i => {
622    // allocated(i) && !writebacked(i) && (datavalid(i) || refilling(i))
623    allocated(i) && !writebacked(i) && datavalid(i) // query refilling will cause bad timing
624  })).asUInt() // use uint instead vec to reduce verilog lines
625
626  val remDeqMask = Seq.tabulate(LoadPipelineWidth)(getRemBits(deqMask)(_))
627
628  // generate lastCycleSelect mask
629  val remFireMask = Seq.tabulate(LoadPipelineWidth)(rem => getRemBits(UIntToOH(loadWbSel(rem)))(rem))
630  // generate real select vec
631  def toVec(a: UInt): Vec[Bool] = {
632    VecInit(a.asBools)
633  }
634  val loadRemSelVecFire = Seq.tabulate(LoadPipelineWidth)(rem => getRemBits(loadWbSelVec)(rem) & ~remFireMask(rem))
635  val loadRemSelVecNotFire = Seq.tabulate(LoadPipelineWidth)(getRemBits(loadWbSelVec)(_))
636  val loadRemSel = Seq.tabulate(LoadPipelineWidth)(rem => Mux(
637    io.ldout(rem).fire(),
638    getFirstOne(toVec(loadRemSelVecFire(rem)), remDeqMask(rem)),
639    getFirstOne(toVec(loadRemSelVecNotFire(rem)), remDeqMask(rem))
640  ))
641
642
643  val loadWbSelGen = Wire(Vec(LoadPipelineWidth, UInt(log2Up(LoadQueueSize).W)))
644  val loadWbSelVGen = Wire(Vec(LoadPipelineWidth, Bool()))
645  (0 until LoadPipelineWidth).foreach(index => {
646    loadWbSelGen(index) := (
647      if (LoadPipelineWidth > 1) Cat(loadRemSel(index), index.U(log2Ceil(LoadPipelineWidth).W))
648      else loadRemSel(index)
649    )
650    loadWbSelVGen(index) := Mux(io.ldout(index).fire, loadRemSelVecFire(index).asUInt.orR, loadRemSelVecNotFire(index).asUInt.orR)
651  })
652
653  (0 until LoadPipelineWidth).map(i => {
654    loadWbSel(i) := RegNext(loadWbSelGen(i))
655    loadWbSelV(i) := RegNext(loadWbSelVGen(i), init = false.B)
656    when(io.ldout(i).fire()){
657      // Mark them as writebacked, so they will not be selected in the next cycle
658      writebacked(loadWbSel(i)) := true.B
659    }
660  })
661
662  // Stage 1
663  // Use indexes generated in cycle 0 to read data
664  // writeback data to cdb
665  (0 until LoadPipelineWidth).map(i => {
666    // data select
667    dataModule.io.wb.raddr(i) := loadWbSelGen(i)
668    val rdata = dataModule.io.wb.rdata(i).data
669    val seluop = uop(loadWbSel(i))
670    val func = seluop.ctrl.fuOpType
671    val raddr = dataModule.io.wb.rdata(i).paddr
672    val rdataSel = LookupTree(raddr(2, 0), List(
673      "b000".U -> rdata(63, 0),
674      "b001".U -> rdata(63, 8),
675      "b010".U -> rdata(63, 16),
676      "b011".U -> rdata(63, 24),
677      "b100".U -> rdata(63, 32),
678      "b101".U -> rdata(63, 40),
679      "b110".U -> rdata(63, 48),
680      "b111".U -> rdata(63, 56)
681    ))
682    val rdataPartialLoad = rdataHelper(seluop, rdataSel)
683
684    // writeback missed int/fp load
685    //
686    // Int load writeback will finish (if not blocked) in one cycle
687    io.ldout(i).bits.uop := seluop
688    io.ldout(i).bits.uop.lqIdx := loadWbSel(i).asTypeOf(new LqPtr)
689    io.ldout(i).bits.data := rdataPartialLoad // not used
690    io.ldout(i).bits.redirectValid := false.B
691    io.ldout(i).bits.redirect := DontCare
692    io.ldout(i).bits.debug.isMMIO := debug_mmio(loadWbSel(i))
693    io.ldout(i).bits.debug.isPerfCnt := false.B
694    io.ldout(i).bits.debug.paddr := debug_paddr(loadWbSel(i))
695    io.ldout(i).bits.debug.vaddr := vaddrModule.io.rdata(i+1)
696    io.ldout(i).bits.fflags := DontCare
697    io.ldout(i).valid := loadWbSelV(i)
698
699    // merged data, uop and offset for data sel in load_s3
700    io.ldRawDataOut(i).lqData := dataModule.io.wb.rdata(i).data
701    io.ldRawDataOut(i).uop := io.ldout(i).bits.uop
702    io.ldRawDataOut(i).addrOffset := dataModule.io.wb.rdata(i).paddr
703
704    when(io.ldout(i).fire()) {
705      XSInfo("int load miss write to cbd robidx %d lqidx %d pc 0x%x mmio %x\n",
706        io.ldout(i).bits.uop.robIdx.asUInt,
707        io.ldout(i).bits.uop.lqIdx.asUInt,
708        io.ldout(i).bits.uop.cf.pc,
709        debug_mmio(loadWbSel(i))
710      )
711    }
712
713  })
714
715  /**
716    * Load commits
717    *
718    * When load commited, mark it as !allocated and move deqPtrExt forward.
719    */
720  (0 until CommitWidth).map(i => {
721    when(commitCount > i.U){
722      allocated((deqPtrExt+i.U).value) := false.B
723      XSError(!allocated((deqPtrExt+i.U).value), s"why commit invalid entry $i?\n")
724    }
725  })
726
727  def getFirstOne(mask: Vec[Bool], startMask: UInt) = {
728    val length = mask.length
729    val highBits = (0 until length).map(i => mask(i) & ~startMask(i))
730    val highBitsUint = Cat(highBits.reverse)
731    PriorityEncoder(Mux(highBitsUint.orR(), highBitsUint, mask.asUInt))
732  }
733
734  def getOldest[T <: XSBundleWithMicroOp](valid: Seq[Bool], bits: Seq[T]): (Seq[Bool], Seq[T]) = {
735    assert(valid.length == bits.length)
736    assert(isPow2(valid.length))
737    if (valid.length == 1) {
738      (valid, bits)
739    } else if (valid.length == 2) {
740      val res = Seq.fill(2)(Wire(ValidIO(chiselTypeOf(bits(0)))))
741      for (i <- res.indices) {
742        res(i).valid := valid(i)
743        res(i).bits := bits(i)
744      }
745      val oldest = Mux(valid(0) && valid(1), Mux(isAfter(bits(0).uop.robIdx, bits(1).uop.robIdx), res(1), res(0)), Mux(valid(0) && !valid(1), res(0), res(1)))
746      (Seq(oldest.valid), Seq(oldest.bits))
747    } else {
748      val left = getOldest(valid.take(valid.length / 2), bits.take(valid.length / 2))
749      val right = getOldest(valid.takeRight(valid.length / 2), bits.takeRight(valid.length / 2))
750      getOldest(left._1 ++ right._1, left._2 ++ right._2)
751    }
752  }
753
754  def getAfterMask(valid: Seq[Bool], uop: Seq[MicroOp]) = {
755    assert(valid.length == uop.length)
756    val length = valid.length
757    (0 until length).map(i => {
758      (0 until length).map(j => {
759        Mux(valid(i) && valid(j),
760          isAfter(uop(i).robIdx, uop(j).robIdx),
761          Mux(!valid(i), true.B, false.B))
762      })
763    })
764  }
765
766
767  /**
768    * Store-Load Memory violation detection
769    *
770    * When store writes back, it searches LoadQueue for younger load instructions
771    * with the same load physical address. They loaded wrong data and need re-execution.
772    *
773    * Cycle 0: Store Writeback
774    *   Generate match vector for store address with rangeMask(stPtr, enqPtr).
775    * Cycle 1: Redirect Generation
776    *   There're up to 2 possible redirect requests.
777    *   Choose the oldest load (part 1).
778    * Cycle 2: Redirect Fire
779    *   Choose the oldest load (part 2).
780    *   Prepare redirect request according to the detected violation.
781    *   Fire redirect request (if valid)
782    */
783
784  // stage 0:        lq                 lq
785  //                 |                  |  (paddr match)
786  // stage 1:        lq                 lq
787  //                 |                  |
788  //                 |                  |
789  //                 |                  |
790  // stage 2:        lq                 lq
791  //                 |                  |
792  //                 --------------------
793  //                          |
794  //                      rollback req
795  io.load_s1 := DontCare
796def detectRollback(i: Int) = {
797    val startIndex = io.storeIn(i).bits.uop.lqIdx.value
798    val lqIdxMask = UIntToMask(startIndex, LoadQueueSize)
799    val xorMask = lqIdxMask ^ enqMask
800    val sameFlag = io.storeIn(i).bits.uop.lqIdx.flag === enqPtrExt(0).flag
801    val stToEnqPtrMask = Mux(sameFlag, xorMask, ~xorMask)
802
803    // check if load already in lq needs to be rolledback
804    dataModule.io.violation(i).paddr := io.storeIn(i).bits.paddr
805    dataModule.io.violation(i).mask := io.storeIn(i).bits.mask
806    val addrMaskMatch = RegNext(dataModule.io.violation(i).violationMask)
807    val entryNeedCheck = RegNext(VecInit((0 until LoadQueueSize).map(j => {
808      allocated(j) && stToEnqPtrMask(j) && (datavalid(j) || miss(j))
809    })))
810    val lqViolationVec = VecInit((0 until LoadQueueSize).map(j => {
811      addrMaskMatch(j) && entryNeedCheck(j)
812    }))
813    val lqViolation = lqViolationVec.asUInt().orR() && RegNext(!io.storeIn(i).bits.miss)
814    val lqViolationIndex = getFirstOne(lqViolationVec, RegNext(lqIdxMask))
815    val lqViolationUop = uop(lqViolationIndex)
816    // lqViolationUop.lqIdx.flag := deqMask(lqViolationIndex) ^ deqPtrExt.flag
817    // lqViolationUop.lqIdx.value := lqViolationIndex
818    XSDebug(lqViolation, p"${Binary(Cat(lqViolationVec))}, $startIndex, $lqViolationIndex\n")
819
820    XSDebug(
821      lqViolation,
822      "need rollback (ld wb before store) pc %x robidx %d target %x\n",
823      io.storeIn(i).bits.uop.cf.pc, io.storeIn(i).bits.uop.robIdx.asUInt, lqViolationUop.robIdx.asUInt
824    )
825
826    (lqViolation, lqViolationUop)
827  }
828
829  def rollbackSel(a: Valid[MicroOpRbExt], b: Valid[MicroOpRbExt]): ValidIO[MicroOpRbExt] = {
830    Mux(
831      a.valid,
832      Mux(
833        b.valid,
834        Mux(isAfter(a.bits.uop.robIdx, b.bits.uop.robIdx), b, a), // a,b both valid, sel oldest
835        a // sel a
836      ),
837      b // sel b
838    )
839  }
840
841  // S2: select rollback (part1) and generate rollback request
842  // rollback check
843  // Lq rollback seq check is done in s3 (next stage), as getting rollbackLq MicroOp is slow
844  val rollbackLq = Wire(Vec(StorePipelineWidth, Valid(new MicroOpRbExt)))
845  // store ftq index for store set update
846  val stFtqIdxS2 = Wire(Vec(StorePipelineWidth, new FtqPtr))
847  val stFtqOffsetS2 = Wire(Vec(StorePipelineWidth, UInt(log2Up(PredictWidth).W)))
848  for (i <- 0 until StorePipelineWidth) {
849    val detectedRollback = detectRollback(i)
850    rollbackLq(i).valid := detectedRollback._1 && RegNext(io.storeIn(i).valid)
851    rollbackLq(i).bits.uop := detectedRollback._2
852    rollbackLq(i).bits.flag := i.U
853    stFtqIdxS2(i) := RegNext(io.storeIn(i).bits.uop.cf.ftqPtr)
854    stFtqOffsetS2(i) := RegNext(io.storeIn(i).bits.uop.cf.ftqOffset)
855  }
856
857  val rollbackLqVReg = rollbackLq.map(x => RegNext(x.valid))
858  val rollbackLqReg = rollbackLq.map(x => RegEnable(x.bits, x.valid))
859
860  // S3: select rollback (part2), generate rollback request, then fire rollback request
861  // Note that we use robIdx - 1.U to flush the load instruction itself.
862  // Thus, here if last cycle's robIdx equals to this cycle's robIdx, it still triggers the redirect.
863
864  // select uop in parallel
865  val lqs = getOldest(rollbackLqVReg, rollbackLqReg)
866  val rollbackUopExt = lqs._2(0)
867  val stFtqIdxS3 = RegNext(stFtqIdxS2)
868  val stFtqOffsetS3 = RegNext(stFtqOffsetS2)
869  val rollbackUop = rollbackUopExt.uop
870  val rollbackStFtqIdx = stFtqIdxS3(rollbackUopExt.flag)
871  val rollbackStFtqOffset = stFtqOffsetS3(rollbackUopExt.flag)
872
873  // check if rollback request is still valid in parallel
874  io.rollback.bits.robIdx := rollbackUop.robIdx
875  io.rollback.bits.ftqIdx := rollbackUop.cf.ftqPtr
876  io.rollback.bits.stFtqIdx := rollbackStFtqIdx
877  io.rollback.bits.ftqOffset := rollbackUop.cf.ftqOffset
878  io.rollback.bits.stFtqOffset := rollbackStFtqOffset
879  io.rollback.bits.level := RedirectLevel.flush
880  io.rollback.bits.interrupt := DontCare
881  io.rollback.bits.cfiUpdate := DontCare
882  io.rollback.bits.cfiUpdate.target := rollbackUop.cf.pc
883  io.rollback.bits.debug_runahead_checkpoint_id := rollbackUop.debugInfo.runahead_checkpoint_id
884  // io.rollback.bits.pc := DontCare
885
886  io.rollback.valid := rollbackLqVReg.reduce(_|_) &&
887                        (!lastCycleRedirect.valid || isBefore(rollbackUop.robIdx, lastCycleRedirect.bits.robIdx)) &&
888                        (!lastlastCycleRedirect.valid || isBefore(rollbackUop.robIdx, lastlastCycleRedirect.bits.robIdx))
889
890  when(io.rollback.valid) {
891    // XSDebug("Mem rollback: pc %x robidx %d\n", io.rollback.bits.cfi, io.rollback.bits.robIdx.asUInt)
892  }
893
894  /**
895  * Load-Load Memory violation detection
896  *
897  * When load arrives load_s1, it searches LoadQueue for younger load instructions
898  * with the same load physical address. If younger load has been released (or observed),
899  * the younger load needs to be re-execed.
900  *
901  * For now, if re-exec it found to be needed in load_s1, we mark the older load as replayInst,
902  * the two loads will be replayed if the older load becomes the head of rob.
903  *
904  * When dcache releases a line, mark all writebacked entrys in load queue with
905  * the same line paddr as released.
906  */
907
908  // Load-Load Memory violation query
909  val deqRightMask = UIntToMask.rightmask(deqPtr, LoadQueueSize)
910  (0 until LoadPipelineWidth).map(i => {
911    dataModule.io.release_violation(i).paddr := io.loadViolationQuery(i).req.bits.paddr
912    io.loadViolationQuery(i).req.ready := true.B
913    io.loadViolationQuery(i).resp.valid := RegNext(io.loadViolationQuery(i).req.fire())
914    // Generate real violation mask
915    // Note that we use UIntToMask.rightmask here
916    val startIndex = io.loadViolationQuery(i).req.bits.uop.lqIdx.value
917    val lqIdxMask = UIntToMask(startIndex, LoadQueueSize)
918    val xorMask = lqIdxMask ^ enqMask
919    val sameFlag = io.loadViolationQuery(i).req.bits.uop.lqIdx.flag === enqPtrExt(0).flag
920    val ldToEnqPtrMask = Mux(sameFlag, xorMask, ~xorMask)
921    val ldld_violation_mask_gen_1 = WireInit(VecInit((0 until LoadQueueSize).map(j => {
922      ldToEnqPtrMask(j) && // the load is younger than current load
923      allocated(j) && // entry is valid
924      released(j) && // cacheline is released
925      (datavalid(j) || miss(j)) // paddr is valid
926    })))
927    val ldld_violation_mask_gen_2 = WireInit(VecInit((0 until LoadQueueSize).map(j => {
928      dataModule.io.release_violation(i).match_mask(j)// addr match
929      // addr match result is slow to generate, we RegNext() it
930    })))
931    val ldld_violation_mask = RegNext(ldld_violation_mask_gen_1).asUInt & RegNext(ldld_violation_mask_gen_2).asUInt
932    dontTouch(ldld_violation_mask)
933    ldld_violation_mask.suggestName("ldldViolationMask_" + i)
934    io.loadViolationQuery(i).resp.bits.have_violation := ldld_violation_mask.orR
935  })
936
937  // "released" flag update
938  //
939  // When io.release.valid (release1cycle.valid), it uses the last ld-ld paddr cam port to
940  // update release flag in 1 cycle
941
942  when(release1cycle.valid){
943    // Take over ld-ld paddr cam port
944    dataModule.io.release_violation.takeRight(1)(0).paddr := release1cycle.bits.paddr
945    io.loadViolationQuery.takeRight(1)(0).req.ready := false.B
946  }
947
948  when(release2cycle.valid){
949    // If a load comes in that cycle, we can not judge if it has ld-ld violation
950    // We replay that load inst from RS
951    io.loadViolationQuery.map(i => i.req.ready :=
952      // use lsu side release2cycle_dup_lsu paddr for better timing
953      !i.req.bits.paddr(PAddrBits-1, DCacheLineOffset) === release2cycle_dup_lsu.bits.paddr(PAddrBits-1, DCacheLineOffset)
954    )
955    // io.loadViolationQuery.map(i => i.req.ready := false.B) // For better timing
956  }
957
958  (0 until LoadQueueSize).map(i => {
959    when(RegNext(dataModule.io.release_violation.takeRight(1)(0).match_mask(i) &&
960      allocated(i) &&
961      datavalid(i) &&
962      release1cycle.valid
963    )){
964      // Note: if a load has missed in dcache and is waiting for refill in load queue,
965      // its released flag still needs to be set as true if addr matches.
966      released(i) := true.B
967    }
968  })
969
970  /**
971    * Memory mapped IO / other uncached operations
972    *
973    * States:
974    * (1) writeback from store units: mark as pending
975    * (2) when they reach ROB's head, they can be sent to uncache channel
976    * (3) response from uncache channel: mark as datavalid
977    * (4) writeback to ROB (and other units): mark as writebacked
978    * (5) ROB commits the instruction: same as normal instructions
979    */
980  //(2) when they reach ROB's head, they can be sent to uncache channel
981  val lqTailMmioPending = WireInit(pending(deqPtr))
982  val lqTailAllocated = WireInit(allocated(deqPtr))
983  val s_idle :: s_req :: s_resp :: s_wait :: Nil = Enum(4)
984  val uncacheState = RegInit(s_idle)
985  switch(uncacheState) {
986    is(s_idle) {
987      when(RegNext(io.rob.pendingld && lqTailMmioPending && lqTailAllocated)) {
988        uncacheState := s_req
989      }
990    }
991    is(s_req) {
992      when(io.uncache.req.fire()) {
993        uncacheState := s_resp
994      }
995    }
996    is(s_resp) {
997      when(io.uncache.resp.fire()) {
998        uncacheState := s_wait
999      }
1000    }
1001    is(s_wait) {
1002      when(RegNext(io.rob.commit)) {
1003        uncacheState := s_idle // ready for next mmio
1004      }
1005    }
1006  }
1007  io.uncache.req.valid := uncacheState === s_req
1008
1009  dataModule.io.uncache.raddr := deqPtrExtNext.value
1010
1011  io.uncache.req.bits.cmd  := MemoryOpConstants.M_XRD
1012  io.uncache.req.bits.addr := dataModule.io.uncache.rdata.paddr
1013  io.uncache.req.bits.data := dataModule.io.uncache.rdata.data
1014  io.uncache.req.bits.mask := dataModule.io.uncache.rdata.mask
1015  io.uncache.req.bits.id   := RegNext(deqPtrExtNext.value)
1016  io.uncache.req.bits.instrtype := DontCare
1017  io.uncache.req.bits.atomic := true.B
1018
1019  io.uncache.resp.ready := true.B
1020
1021  when (io.uncache.req.fire()) {
1022    pending(deqPtr) := false.B
1023
1024    XSDebug("uncache req: pc %x addr %x data %x op %x mask %x\n",
1025      uop(deqPtr).cf.pc,
1026      io.uncache.req.bits.addr,
1027      io.uncache.req.bits.data,
1028      io.uncache.req.bits.cmd,
1029      io.uncache.req.bits.mask
1030    )
1031  }
1032
1033  // (3) response from uncache channel: mark as datavalid
1034  dataModule.io.uncache.wen := false.B
1035  when(io.uncache.resp.fire()){
1036    datavalid(deqPtr) := true.B
1037    dataModule.io.uncacheWrite(deqPtr, io.uncache.resp.bits.data(XLEN-1, 0))
1038    dataModule.io.uncache.wen := true.B
1039
1040    XSDebug("uncache resp: data %x\n", io.refill.bits.data)
1041  }
1042
1043  // Read vaddr for mem exception
1044  // no inst will be commited 1 cycle before tval update
1045  vaddrModule.io.raddr(0) := (deqPtrExt + commitCount).value
1046  io.exceptionAddr.vaddr := vaddrModule.io.rdata(0)
1047
1048  // Read vaddr for debug
1049  (0 until LoadPipelineWidth).map(i => {
1050    vaddrModule.io.raddr(i+1) := loadWbSel(i)
1051  })
1052
1053  (0 until LoadPipelineWidth).map(i => {
1054    vaddrTriggerResultModule.io.raddr(i) := loadWbSelGen(i)
1055    io.trigger(i).lqLoadAddrTriggerHitVec := Mux(
1056      loadWbSelV(i),
1057      vaddrTriggerResultModule.io.rdata(i),
1058      VecInit(Seq.fill(3)(false.B))
1059    )
1060  })
1061
1062  // misprediction recovery / exception redirect
1063  // invalidate lq term using robIdx
1064  val needCancel = Wire(Vec(LoadQueueSize, Bool()))
1065  for (i <- 0 until LoadQueueSize) {
1066    needCancel(i) := uop(i).robIdx.needFlush(io.brqRedirect) && allocated(i)
1067    when (needCancel(i)) {
1068      allocated(i) := false.B
1069    }
1070  }
1071
1072  /**
1073    * update pointers
1074    */
1075  val lastEnqCancel = PopCount(RegNext(VecInit(canEnqueue.zip(enqCancel).map(x => x._1 && x._2))))
1076  val lastCycleCancelCount = PopCount(RegNext(needCancel))
1077  val enqNumber = Mux(io.enq.canAccept && io.enq.sqCanAccept, PopCount(io.enq.req.map(_.valid)), 0.U)
1078  when (lastCycleRedirect.valid) {
1079    // we recover the pointers in the next cycle after redirect
1080    enqPtrExt := VecInit(enqPtrExt.map(_ - (lastCycleCancelCount + lastEnqCancel)))
1081  }.otherwise {
1082    enqPtrExt := VecInit(enqPtrExt.map(_ + enqNumber))
1083  }
1084
1085  deqPtrExtNext := deqPtrExt + commitCount
1086  deqPtrExt := deqPtrExtNext
1087
1088  io.lqCancelCnt := RegNext(lastCycleCancelCount + lastEnqCancel)
1089
1090  /**
1091    * misc
1092    */
1093  // perf counter
1094  QueuePerf(LoadQueueSize, validCount, !allowEnqueue)
1095  io.lqFull := !allowEnqueue
1096  XSPerfAccumulate("rollback", io.rollback.valid) // rollback redirect generated
1097  XSPerfAccumulate("mmioCycle", uncacheState =/= s_idle) // lq is busy dealing with uncache req
1098  XSPerfAccumulate("mmioCnt", io.uncache.req.fire())
1099  XSPerfAccumulate("refill", io.refill.valid)
1100  XSPerfAccumulate("writeback_success", PopCount(VecInit(io.ldout.map(i => i.fire()))))
1101  XSPerfAccumulate("writeback_blocked", PopCount(VecInit(io.ldout.map(i => i.valid && !i.ready))))
1102  XSPerfAccumulate("utilization_miss", PopCount((0 until LoadQueueSize).map(i => allocated(i) && miss(i))))
1103
1104  if (env.EnableTopDown) {
1105    val stall_loads_bound = WireDefault(0.B)
1106    ExcitingUtils.addSink(stall_loads_bound, "stall_loads_bound", ExcitingUtils.Perf)
1107    val have_miss_entry = (allocated zip miss).map(x => x._1 && x._2).reduce(_ || _)
1108    val l1d_loads_bound = stall_loads_bound && !have_miss_entry
1109    ExcitingUtils.addSource(l1d_loads_bound, "l1d_loads_bound", ExcitingUtils.Perf)
1110    XSPerfAccumulate("l1d_loads_bound", l1d_loads_bound)
1111    val stall_l1d_load_miss = stall_loads_bound && have_miss_entry
1112    ExcitingUtils.addSource(stall_l1d_load_miss, "stall_l1d_load_miss", ExcitingUtils.Perf)
1113    ExcitingUtils.addSink(WireInit(0.U), "stall_l1d_load_miss", ExcitingUtils.Perf)
1114  }
1115
1116  val perfValidCount = RegNext(validCount)
1117
1118  val perfEvents = Seq(
1119    ("rollback         ", io.rollback.valid),
1120    ("mmioCycle        ", uncacheState =/= s_idle),
1121    ("mmio_Cnt         ", io.uncache.req.fire()),
1122    ("refill           ", io.refill.valid),
1123    ("writeback_success", PopCount(VecInit(io.ldout.map(i => i.fire())))),
1124    ("writeback_blocked", PopCount(VecInit(io.ldout.map(i => i.valid && !i.ready)))),
1125    ("ltq_1_4_valid    ", (perfValidCount < (LoadQueueSize.U/4.U))),
1126    ("ltq_2_4_valid    ", (perfValidCount > (LoadQueueSize.U/4.U)) & (perfValidCount <= (LoadQueueSize.U/2.U))),
1127    ("ltq_3_4_valid    ", (perfValidCount > (LoadQueueSize.U/2.U)) & (perfValidCount <= (LoadQueueSize.U*3.U/4.U))),
1128    ("ltq_4_4_valid    ", (perfValidCount > (LoadQueueSize.U*3.U/4.U)))
1129  )
1130  generatePerfEvent()
1131
1132  // debug info
1133  XSDebug("enqPtrExt %d:%d deqPtrExt %d:%d\n", enqPtrExt(0).flag, enqPtr, deqPtrExt.flag, deqPtr)
1134
1135  def PrintFlag(flag: Bool, name: String): Unit = {
1136    when(flag) {
1137      XSDebug(false, true.B, name)
1138    }.otherwise {
1139      XSDebug(false, true.B, " ")
1140    }
1141  }
1142
1143  for (i <- 0 until LoadQueueSize) {
1144    XSDebug(i + " pc %x pa %x ", uop(i).cf.pc, debug_paddr(i))
1145    PrintFlag(allocated(i), "a")
1146    PrintFlag(allocated(i) && datavalid(i), "v")
1147    PrintFlag(allocated(i) && writebacked(i), "w")
1148    PrintFlag(allocated(i) && miss(i), "m")
1149    PrintFlag(allocated(i) && pending(i), "p")
1150    XSDebug(false, true.B, "\n")
1151  }
1152
1153}
1154