xref: /XiangShan/src/main/scala/xiangshan/mem/pipeline/StoreUnit.scala (revision 233f2ad08102497ba2a93670bfce33bf22bf645f)
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 org.chipsalliance.cde.config.Parameters
20import chisel3._
21import chisel3.util._
22import utils._
23import utility._
24import xiangshan.ExceptionNO._
25import xiangshan._
26import xiangshan.backend.Bundles.{MemExuInput, MemExuOutput}
27import xiangshan.backend.fu.PMPRespBundle
28import xiangshan.backend.fu.FuConfig._
29import xiangshan.backend.fu.FuType._
30import xiangshan.backend.ctrlblock.DebugLsInfoBundle
31import xiangshan.backend.fu.NewCSR._
32import xiangshan.cache.mmu.{TlbCmd, TlbReq, TlbRequestIO, TlbResp, Pbmt}
33import xiangshan.cache.{DcacheStoreRequestIO, DCacheStoreIO, MemoryOpConstants, HasDCacheParameters, StorePrefetchReq}
34
35class StoreUnit(implicit p: Parameters) extends XSModule
36  with HasDCacheParameters
37  with HasVLSUParameters
38  {
39  val io = IO(new Bundle() {
40    val redirect        = Flipped(ValidIO(new Redirect))
41    val csrCtrl         = Flipped(new CustomCSRCtrlIO)
42    val stin            = Flipped(Decoupled(new MemExuInput))
43    val issue           = Valid(new MemExuInput)
44    // misalignBuffer issue path
45    val misalign_stin   = Flipped(Decoupled(new LsPipelineBundle))
46    val misalign_stout  = Valid(new SqWriteBundle)
47    val tlb             = new TlbRequestIO()
48    val dcache          = new DCacheStoreIO
49    val pmp             = Flipped(new PMPRespBundle())
50    val lsq             = ValidIO(new LsPipelineBundle)
51    val lsq_replenish   = Output(new LsPipelineBundle())
52    val feedback_slow   = ValidIO(new RSFeedback)
53    val prefetch_req    = Flipped(DecoupledIO(new StorePrefetchReq))
54    // provide prefetch info to sms
55    val prefetch_train  = ValidIO(new StPrefetchTrainBundle())
56    // speculative for gated control
57    val s1_prefetch_spec = Output(Bool())
58    val s2_prefetch_spec = Output(Bool())
59    val stld_nuke_query = Valid(new StoreNukeQueryIO)
60    val stout           = DecoupledIO(new MemExuOutput) // writeback store
61    val vecstout        = DecoupledIO(new VecPipelineFeedbackIO(isVStore = true))
62    // store mask, send to sq in store_s0
63    val st_mask_out     = Valid(new StoreMaskBundle)
64    val debug_ls        = Output(new DebugLsInfoBundle)
65    // vector
66    val vecstin           = Flipped(Decoupled(new VecPipeBundle(isVStore = true)))
67    val vec_isFirstIssue  = Input(Bool())
68    // writeback to misalign buffer
69    val misalign_buf = Valid(new LsPipelineBundle)
70    // trigger
71    val fromCsrTrigger = Input(new CsrTriggerBundle)
72  })
73
74  val s1_ready, s2_ready, s3_ready = WireInit(false.B)
75
76  // Pipeline
77  // --------------------------------------------------------------------------------
78  // stage 0
79  // --------------------------------------------------------------------------------
80  // generate addr, use addr to query DCache and DTLB
81  val s0_iss_valid        = io.stin.valid
82  val s0_prf_valid        = io.prefetch_req.valid && io.dcache.req.ready
83  val s0_vec_valid        = io.vecstin.valid
84  val s0_ma_st_valid      = io.misalign_stin.valid
85  val s0_valid            = s0_iss_valid || s0_prf_valid || s0_vec_valid || s0_ma_st_valid
86  val s0_use_flow_ma      = s0_ma_st_valid
87  val s0_use_flow_vec     = s0_vec_valid && !s0_ma_st_valid
88  val s0_use_flow_rs      = s0_iss_valid && !s0_vec_valid && !s0_ma_st_valid
89  val s0_use_flow_prf     = s0_prf_valid && !s0_iss_valid && !s0_vec_valid && !s0_ma_st_valid
90  val s0_use_non_prf_flow = s0_use_flow_rs || s0_use_flow_vec || s0_use_flow_ma
91  val s0_stin             = Mux(s0_use_flow_rs, io.stin.bits, 0.U.asTypeOf(io.stin.bits))
92  val s0_vecstin          = Mux(s0_use_flow_vec, io.vecstin.bits, 0.U.asTypeOf(io.vecstin.bits))
93  val s0_uop              = Mux(
94    s0_use_flow_ma,
95    io.misalign_stin.bits.uop,
96    Mux(
97      s0_use_flow_rs,
98      s0_stin.uop,
99      s0_vecstin.uop
100    )
101  )
102  val s0_isFirstIssue = Mux(
103    s0_use_flow_ma,
104    false.B,
105    s0_use_flow_rs && io.stin.bits.isFirstIssue || s0_use_flow_vec && io.vec_isFirstIssue
106  )
107  val s0_size         = Mux(s0_use_non_prf_flow, s0_uop.fuOpType(2,0), 0.U)// may broken if use it in feature
108  val s0_mem_idx      = Mux(s0_use_non_prf_flow, s0_uop.sqIdx.value, 0.U)
109  val s0_rob_idx      = Mux(s0_use_non_prf_flow, s0_uop.robIdx, 0.U.asTypeOf(s0_uop.robIdx))
110  val s0_pc           = Mux(s0_use_non_prf_flow, s0_uop.pc, 0.U)
111  val s0_instr_type   = Mux(s0_use_non_prf_flow, STORE_SOURCE.U, DCACHE_PREFETCH_SOURCE.U)
112  val s0_wlineflag    = Mux(s0_use_flow_rs, s0_uop.fuOpType === LSUOpType.cbo_zero, false.B)
113  val s0_out          = Wire(new LsPipelineBundle)
114  val s0_kill         = s0_uop.robIdx.needFlush(io.redirect)
115  val s0_can_go       = s1_ready
116  val s0_fire         = s0_valid && !s0_kill && s0_can_go
117  val s0_is128bit     = Mux(s0_use_flow_ma, io.misalign_stin.bits.is128bit, is128Bit(s0_vecstin.alignedType))
118  // vector
119  val s0_vecActive    = !s0_use_flow_vec || s0_vecstin.vecActive
120  // val s0_flowPtr      = s0_vecstin.flowPtr
121  // val s0_isLastElem   = s0_vecstin.isLastElem
122  val s0_secondInv    = s0_vecstin.usSecondInv
123  val s0_elemIdx      = s0_vecstin.elemIdx
124  val s0_alignedType  = s0_vecstin.alignedType
125  val s0_mBIndex      = s0_vecstin.mBIndex
126
127  // generate addr
128  val s0_saddr = s0_stin.src(0) + SignExt(s0_uop.imm(11,0), VAddrBits)
129  val s0_fullva = Wire(UInt(XLEN.W))
130  val s0_vaddr = Mux(
131    s0_use_flow_ma,
132    io.misalign_stin.bits.vaddr,
133    Mux(
134      s0_use_flow_rs,
135      s0_saddr,
136      Mux(
137        s0_use_flow_vec,
138        s0_vecstin.vaddr(VAddrBits - 1, 0),
139        io.prefetch_req.bits.vaddr
140      )
141    )
142  )
143  s0_fullva := Mux(
144    s0_use_flow_rs,
145    s0_stin.src(0) + SignExt(s0_uop.imm(11,0), XLEN),
146    Mux(
147      s0_use_flow_vec,
148      s0_vecstin.vaddr,
149      s0_vaddr
150    )
151  )
152
153  val s0_mask = Mux(
154    s0_use_flow_ma,
155    io.misalign_stin.bits.mask,
156    Mux(
157      s0_use_flow_rs,
158      genVWmask128(s0_saddr, s0_uop.fuOpType(2,0)),
159      Mux(
160        s0_use_flow_vec,
161        s0_vecstin.mask,
162        // -1.asSInt.asUInt
163        Fill(VLEN/8, 1.U(1.W))
164      )
165    )
166  )
167
168  io.tlb.req.valid                   := s0_valid
169  io.tlb.req.bits.vaddr              := s0_vaddr
170  io.tlb.req.bits.fullva             := s0_fullva
171  io.tlb.req.bits.checkfullva        := s0_use_flow_rs || s0_use_flow_vec
172  io.tlb.req.bits.cmd                := TlbCmd.write
173  io.tlb.req.bits.size               := s0_size
174  io.tlb.req.bits.kill               := false.B
175  io.tlb.req.bits.memidx.is_ld       := false.B
176  io.tlb.req.bits.memidx.is_st       := true.B
177  io.tlb.req.bits.memidx.idx         := s0_mem_idx
178  io.tlb.req.bits.debug.robIdx       := s0_rob_idx
179  io.tlb.req.bits.no_translate       := false.B
180  io.tlb.req.bits.debug.pc           := s0_pc
181  io.tlb.req.bits.debug.isFirstIssue := s0_isFirstIssue
182  io.tlb.req_kill                    := false.B
183  io.tlb.req.bits.hyperinst          := LSUOpType.isHsv(s0_uop.fuOpType)
184  io.tlb.req.bits.hlvx               := false.B
185  io.tlb.req.bits.pmp_addr           := DontCare
186
187  // Dcache access here: not **real** dcache write
188  // just read meta and tag in dcache, to find out the store will hit or miss
189
190  // NOTE: The store request does not wait for the dcache to be ready.
191  //       If the dcache is not ready at this time, the dcache is not queried.
192  //       But, store prefetch request will always wait for dcache to be ready to make progress.
193  io.dcache.req.valid              := s0_fire
194  io.dcache.req.bits.cmd           := MemoryOpConstants.M_PFW
195  io.dcache.req.bits.vaddr         := s0_vaddr
196  io.dcache.req.bits.instrtype     := s0_instr_type
197
198  s0_out              := DontCare
199  s0_out.vaddr        := s0_vaddr
200  s0_out.fullva       := s0_fullva
201  // Now data use its own io
202  // s1_out.data := genWdata(s1_in.src(1), s1_in.uop.fuOpType(1,0))
203  s0_out.data         := s0_stin.src(1)
204  s0_out.uop          := s0_uop
205  s0_out.miss         := false.B
206  s0_out.mask         := s0_mask
207  s0_out.isFirstIssue := s0_isFirstIssue
208  s0_out.isHWPrefetch := s0_use_flow_prf
209  s0_out.wlineflag    := s0_wlineflag
210  s0_out.isvec        := s0_use_flow_vec
211  s0_out.is128bit     := s0_is128bit
212  s0_out.vecActive    := s0_vecActive
213  s0_out.usSecondInv  := s0_secondInv
214  s0_out.elemIdx      := s0_elemIdx
215  s0_out.alignedType  := s0_alignedType
216  s0_out.mbIndex      := s0_mBIndex
217  when(s0_valid && s0_isFirstIssue) {
218    s0_out.uop.debugInfo.tlbFirstReqTime := GTimer()
219  }
220  s0_out.isFrmMisAlignBuf := s0_use_flow_ma
221
222  // exception check
223  val s0_addr_aligned = LookupTree(Mux(s0_use_flow_vec, s0_vecstin.alignedType(1,0), s0_uop.fuOpType(1, 0)), List(
224    "b00".U   -> true.B,              //b
225    "b01".U   -> (s0_out.vaddr(0) === 0.U),   //h
226    "b10".U   -> (s0_out.vaddr(1,0) === 0.U), //w
227    "b11".U   -> (s0_out.vaddr(2,0) === 0.U)  //d
228  ))
229  // if vector store sends 128-bit requests, its address must be 128-aligned
230  XSError(s0_use_flow_vec && s0_out.vaddr(3, 0) =/= 0.U && s0_vecstin.alignedType(2), "unit stride 128 bit element is not aligned!")
231  s0_out.uop.exceptionVec(storeAddrMisaligned) := Mux(s0_use_non_prf_flow, (!s0_addr_aligned || s0_vecstin.uop.exceptionVec(storeAddrMisaligned) && s0_vecActive), false.B)
232
233  io.st_mask_out.valid       := s0_use_flow_rs || s0_use_flow_vec
234  io.st_mask_out.bits.mask   := s0_out.mask
235  io.st_mask_out.bits.sqIdx  := s0_out.uop.sqIdx
236
237  io.stin.ready := s1_ready && s0_use_flow_rs
238  io.vecstin.ready := s1_ready && s0_use_flow_vec
239  io.prefetch_req.ready := s1_ready && io.dcache.req.ready && !s0_iss_valid && !s0_vec_valid && !s0_ma_st_valid
240  io.misalign_stin.ready := s1_ready && s0_use_flow_ma
241
242  // Pipeline
243  // --------------------------------------------------------------------------------
244  // stage 1
245  // --------------------------------------------------------------------------------
246  // TLB resp (send paddr to dcache)
247  val s1_valid  = RegInit(false.B)
248  val s1_in     = RegEnable(s0_out, s0_fire)
249  val s1_out    = Wire(new LsPipelineBundle)
250  val s1_kill   = Wire(Bool())
251  val s1_can_go = s2_ready
252  val s1_fire   = s1_valid && !s1_kill && s1_can_go
253  val s1_vecActive    = RegEnable(s0_out.vecActive, true.B, s0_fire)
254  val s1_frm_mabuf    = s1_in.isFrmMisAlignBuf
255
256  // mmio cbo decoder
257  val s1_mmio_cbo  = s1_in.uop.fuOpType === LSUOpType.cbo_clean ||
258                     s1_in.uop.fuOpType === LSUOpType.cbo_flush ||
259                     s1_in.uop.fuOpType === LSUOpType.cbo_inval
260  val s1_paddr     = io.tlb.resp.bits.paddr(0)
261  val s1_gpaddr    = io.tlb.resp.bits.gpaddr(0)
262  val s1_tlb_miss  = io.tlb.resp.bits.miss
263  val s1_mmio      = s1_mmio_cbo
264  val s1_pbmt      = io.tlb.resp.bits.pbmt(0)
265  val s1_exception = ExceptionNO.selectByFu(s1_out.uop.exceptionVec, StaCfg).asUInt.orR
266  val s1_isvec     = RegEnable(s0_out.isvec, false.B, s0_fire)
267  // val s1_isLastElem = RegEnable(s0_isLastElem, false.B, s0_fire)
268  s1_kill := s1_in.uop.robIdx.needFlush(io.redirect) || (s1_tlb_miss && !s1_isvec && !s1_frm_mabuf)
269
270  s1_ready := !s1_valid || s1_kill || s2_ready
271  io.tlb.resp.ready := true.B // TODO: why dtlbResp needs a ready?
272  when (s0_fire) { s1_valid := true.B }
273  .elsewhen (s1_fire) { s1_valid := false.B }
274  .elsewhen (s1_kill) { s1_valid := false.B }
275
276  // st-ld violation dectect request.
277  io.stld_nuke_query.valid       := s1_valid && !s1_tlb_miss && !s1_in.isHWPrefetch && !s1_frm_mabuf
278  io.stld_nuke_query.bits.robIdx := s1_in.uop.robIdx
279  io.stld_nuke_query.bits.paddr  := s1_paddr
280  io.stld_nuke_query.bits.mask   := s1_in.mask
281  io.stld_nuke_query.bits.matchLine := s1_in.isvec && s1_in.is128bit
282
283  // issue
284  io.issue.valid := s1_valid && !s1_tlb_miss && !s1_in.isHWPrefetch && !s1_isvec && !s1_frm_mabuf
285  io.issue.bits  := RegEnable(s0_stin, s0_valid)
286
287
288  // Send TLB feedback to store issue queue
289  // Store feedback is generated in store_s1, sent to RS in store_s2
290  val s1_feedback = Wire(Valid(new RSFeedback))
291  s1_feedback.valid                 := s1_valid & !s1_in.isHWPrefetch
292  s1_feedback.bits.hit              := !s1_tlb_miss
293  s1_feedback.bits.flushState       := io.tlb.resp.bits.ptwBack
294  s1_feedback.bits.robIdx           := s1_out.uop.robIdx
295  s1_feedback.bits.sourceType       := RSFeedbackType.tlbMiss
296  s1_feedback.bits.dataInvalidSqIdx := DontCare
297  s1_feedback.bits.sqIdx            := s1_out.uop.sqIdx
298  s1_feedback.bits.lqIdx            := s1_out.uop.lqIdx
299
300  XSDebug(s1_feedback.valid,
301    "S1 Store: tlbHit: %d robIdx: %d\n",
302    s1_feedback.bits.hit,
303    s1_feedback.bits.robIdx.value
304  )
305
306  // io.feedback_slow := s1_feedback
307
308  // get paddr from dtlb, check if rollback is needed
309  // writeback store inst to lsq
310  s1_out         := s1_in
311  s1_out.paddr   := s1_paddr
312  s1_out.gpaddr  := s1_gpaddr
313  s1_out.miss    := false.B
314  s1_out.mmio    := s1_mmio
315  s1_out.tlbMiss := s1_tlb_miss
316  s1_out.atomic  := s1_mmio
317  when (!s1_out.isFrmMisAlignBuf && RegNext(io.tlb.req.bits.checkfullva) && (s1_out.uop.exceptionVec(storePageFault) || s1_out.uop.exceptionVec(storeAccessFault) || s1_out.uop.exceptionVec(storeGuestPageFault))) {
318    s1_out.uop.exceptionVec(storeAddrMisaligned) := false.B
319  }
320  s1_out.uop.exceptionVec(storePageFault)      := io.tlb.resp.bits.excp(0).pf.st && s1_vecActive
321  s1_out.uop.exceptionVec(storeAccessFault)    := io.tlb.resp.bits.excp(0).af.st && s1_vecActive
322  s1_out.uop.exceptionVec(storeGuestPageFault) := io.tlb.resp.bits.excp(0).gpf.st && s1_vecActive
323
324  // trigger
325  val storeTrigger = Module(new MemTrigger(MemType.STORE))
326  storeTrigger.io.fromCsrTrigger.tdataVec             := io.fromCsrTrigger.tdataVec
327  storeTrigger.io.fromCsrTrigger.tEnableVec           := io.fromCsrTrigger.tEnableVec
328  storeTrigger.io.fromCsrTrigger.triggerCanRaiseBpExp := io.fromCsrTrigger.triggerCanRaiseBpExp
329  storeTrigger.io.fromCsrTrigger.debugMode            := io.fromCsrTrigger.debugMode
330  storeTrigger.io.fromLoadStore.vaddr                 := s1_in.vaddr
331
332  val s1_trigger_action = storeTrigger.io.toLoadStore.triggerAction
333  val s1_trigger_debug_mode = TriggerAction.isDmode(s1_trigger_action)
334  val s1_trigger_breakpoint = TriggerAction.isExp(s1_trigger_action)
335
336  s1_out.uop.flushPipe                := false.B
337  s1_out.uop.trigger                  := s1_trigger_action
338  s1_out.uop.exceptionVec(breakPoint) := s1_trigger_breakpoint
339
340  // scalar store and scalar load nuke check, and also other purposes
341  io.lsq.valid     := s1_valid && !s1_in.isHWPrefetch && !s1_frm_mabuf
342  io.lsq.bits      := s1_out
343  io.lsq.bits.miss := s1_tlb_miss
344
345  // goto misalignBuffer
346  io.misalign_buf.valid := s1_valid && !s1_in.isHWPrefetch && GatedValidRegNext(io.csrCtrl.hd_misalign_st_enable) && !s1_in.isvec
347  io.misalign_buf.bits  := io.lsq.bits
348
349  // kill dcache write intent request when tlb miss or exception
350  io.dcache.s1_kill  := (s1_tlb_miss || s1_exception || s1_mmio || s1_in.uop.robIdx.needFlush(io.redirect))
351  io.dcache.s1_paddr := s1_paddr
352
353  // write below io.out.bits assign sentence to prevent overwriting values
354  val s1_tlb_memidx = io.tlb.resp.bits.memidx
355  when(s1_tlb_memidx.is_st && io.tlb.resp.valid && !s1_tlb_miss && s1_tlb_memidx.idx === s1_out.uop.sqIdx.value) {
356    // printf("Store idx = %d\n", s1_tlb_memidx.idx)
357    s1_out.uop.debugInfo.tlbRespTime := GTimer()
358  }
359
360  // Pipeline
361  // --------------------------------------------------------------------------------
362  // stage 2
363  // --------------------------------------------------------------------------------
364  // mmio check
365  val s2_valid  = RegInit(false.B)
366  val s2_in     = RegEnable(s1_out, s1_fire)
367  val s2_out    = Wire(new LsPipelineBundle)
368  val s2_kill   = Wire(Bool())
369  val s2_can_go = s3_ready
370  val s2_fire   = s2_valid && !s2_kill && s2_can_go
371  val s2_vecActive    = RegEnable(s1_out.vecActive, true.B, s1_fire)
372  val s2_frm_mabuf    = s2_in.isFrmMisAlignBuf
373  val s2_pbmt   = RegEnable(s1_pbmt, s1_fire)
374  val s2_trigger_debug_mode = RegEnable(s1_trigger_debug_mode, false.B, s1_fire)
375  val s2_mis_align = GatedValidRegNext(io.csrCtrl.hd_misalign_st_enable) && !s2_in.isvec &&
376                     s2_in.uop.exceptionVec(storeAddrMisaligned) && !s2_in.uop.exceptionVec(breakPoint) && !s2_trigger_debug_mode
377
378  s2_ready := !s2_valid || s2_kill || s3_ready
379  when (s1_fire) { s2_valid := true.B }
380  .elsewhen (s2_fire) { s2_valid := false.B }
381  .elsewhen (s2_kill) { s2_valid := false.B }
382
383  val s2_pmp = WireInit(io.pmp)
384
385  val s2_exception = RegNext(s1_feedback.bits.hit) &&
386                    (s2_trigger_debug_mode || ExceptionNO.selectByFu(s2_out.uop.exceptionVec, StaCfg).asUInt.orR)
387  val s2_mmio = (s2_in.mmio || s2_pmp.mmio || Pbmt.isUncache(s2_pbmt)) && RegNext(s1_feedback.bits.hit)
388  s2_kill := ((s2_mmio && !s2_exception) && !s2_in.isvec) || s2_in.uop.robIdx.needFlush(io.redirect)
389
390  s2_out        := s2_in
391  s2_out.af     := s2_pmp.st && !s2_in.isvec
392  s2_out.mmio   := s2_mmio && !s2_exception
393  s2_out.atomic := s2_in.atomic || s2_pmp.atomic
394  s2_out.uop.exceptionVec(storeAccessFault) := (s2_in.uop.exceptionVec(storeAccessFault) ||
395                                                s2_pmp.st ||
396                                                (s2_in.isvec && s2_pmp.mmio && RegNext(s1_feedback.bits.hit))
397                                                ) && s2_vecActive
398
399  // kill dcache write intent request when mmio or exception
400  io.dcache.s2_kill := (s2_mmio || s2_exception || s2_in.uop.robIdx.needFlush(io.redirect))
401  io.dcache.s2_pc   := s2_out.uop.pc
402  // TODO: dcache resp
403  io.dcache.resp.ready := true.B
404
405  // feedback tlb miss to RS in store_s2
406  val feedback_slow_valid = WireInit(false.B)
407  feedback_slow_valid := s1_feedback.valid && !s1_out.uop.robIdx.needFlush(io.redirect) && !s1_out.isvec && !s1_frm_mabuf
408  io.feedback_slow.valid := GatedValidRegNext(feedback_slow_valid)
409  io.feedback_slow.bits  := RegEnable(s1_feedback.bits, feedback_slow_valid)
410
411  val s2_vecFeedback = RegNext(!s1_out.uop.robIdx.needFlush(io.redirect) && s1_feedback.bits.hit) && s2_in.isvec
412
413  val s2_misalign_stout = WireInit(0.U.asTypeOf(io.misalign_stout))
414  s2_misalign_stout.valid := s2_valid && s2_can_go && s2_frm_mabuf
415  s2_misalign_stout.bits.mmio := s2_out.mmio
416  s2_misalign_stout.bits.vaddr := s2_out.vaddr
417  s2_misalign_stout.bits.paddr := s2_out.paddr
418  s2_misalign_stout.bits.gpaddr := s2_out.gpaddr
419  s2_misalign_stout.bits.need_rep := RegEnable(s1_tlb_miss, s1_fire)
420  s2_misalign_stout.bits.uop.exceptionVec := s2_out.uop.exceptionVec
421  io.misalign_stout := s2_misalign_stout
422
423  // mmio and exception
424  io.lsq_replenish := s2_out
425  io.lsq_replenish.af := s2_out.af && s2_valid && !s2_kill
426
427  // prefetch related
428  io.lsq_replenish.miss := io.dcache.resp.fire && io.dcache.resp.bits.miss // miss info
429
430  // RegNext prefetch train for better timing
431  // ** Now, prefetch train is valid at store s3 **
432  val s2_prefetch_train_valid = WireInit(false.B)
433  s2_prefetch_train_valid := s2_valid && io.dcache.resp.fire && !s2_out.mmio && !s2_in.tlbMiss && !s2_in.isHWPrefetch
434  if(EnableStorePrefetchSMS) {
435    io.s1_prefetch_spec := s1_fire
436    io.s2_prefetch_spec := s2_prefetch_train_valid
437    io.prefetch_train.valid := RegNext(s2_prefetch_train_valid)
438    io.prefetch_train.bits.fromLsPipelineBundle(s2_in, latch = true, enable = s2_prefetch_train_valid)
439  }else {
440    io.s1_prefetch_spec := false.B
441    io.s2_prefetch_spec := false.B
442    io.prefetch_train.valid := false.B
443    io.prefetch_train.bits.fromLsPipelineBundle(s2_in, latch = true, enable = false.B)
444  }
445  // override miss bit
446  io.prefetch_train.bits.miss := RegEnable(io.dcache.resp.bits.miss, s2_prefetch_train_valid)
447  // TODO: add prefetch and access bit
448  io.prefetch_train.bits.meta_prefetch := false.B
449  io.prefetch_train.bits.meta_access := false.B
450
451  // Pipeline
452  // --------------------------------------------------------------------------------
453  // stage 3
454  // --------------------------------------------------------------------------------
455  // store write back
456  val s3_valid  = RegInit(false.B)
457  val s3_in     = RegEnable(s2_out, s2_fire)
458  val s3_out    = Wire(new MemExuOutput(isVector = true))
459  val s3_kill   = s3_in.uop.robIdx.needFlush(io.redirect)
460  val s3_can_go = s3_ready
461  val s3_fire   = s3_valid && !s3_kill && s3_can_go
462  val s3_vecFeedback = RegEnable(s2_vecFeedback, s2_fire)
463
464  // store misalign will not writeback to rob now
465  when (s2_fire) { s3_valid := (!s2_mmio || s2_exception) && !s2_out.isHWPrefetch && !s2_mis_align && !s2_frm_mabuf }
466  .elsewhen (s3_fire) { s3_valid := false.B }
467  .elsewhen (s3_kill) { s3_valid := false.B }
468
469  // wb: writeback
470  val SelectGroupSize   = RollbackGroupSize
471  val lgSelectGroupSize = log2Ceil(SelectGroupSize)
472  val TotalSelectCycles = scala.math.ceil(log2Ceil(LoadQueueRAWSize).toFloat / lgSelectGroupSize).toInt + 1
473
474  s3_out                 := DontCare
475  s3_out.uop             := s3_in.uop
476  s3_out.data            := DontCare
477  s3_out.debug.isMMIO    := s3_in.mmio
478  s3_out.debug.paddr     := s3_in.paddr
479  s3_out.debug.vaddr     := s3_in.vaddr
480  s3_out.debug.isPerfCnt := false.B
481
482  // Pipeline
483  // --------------------------------------------------------------------------------
484  // stage x
485  // --------------------------------------------------------------------------------
486  // delay TotalSelectCycles - 2 cycle(s)
487  val TotalDelayCycles = TotalSelectCycles - 2
488  val sx_valid = Wire(Vec(TotalDelayCycles + 1, Bool()))
489  val sx_ready = Wire(Vec(TotalDelayCycles + 1, Bool()))
490  val sx_in    = Wire(Vec(TotalDelayCycles + 1, new VecMemExuOutput(isVector = true)))
491
492  // backward ready signal
493  s3_ready := sx_ready.head
494  for (i <- 0 until TotalDelayCycles + 1) {
495    if (i == 0) {
496      sx_valid(i)          := s3_valid
497      sx_in(i).output      := s3_out
498      sx_in(i).vecFeedback := s3_vecFeedback
499      sx_in(i).mmio        := s3_in.mmio
500      sx_in(i).usSecondInv := s3_in.usSecondInv
501      sx_in(i).elemIdx     := s3_in.elemIdx
502      sx_in(i).alignedType := s3_in.alignedType
503      sx_in(i).mbIndex     := s3_in.mbIndex
504      sx_in(i).mask        := s3_in.mask
505      sx_in(i).vaddr       := s3_in.fullva
506      sx_in(i).gpaddr      := s3_in.gpaddr
507      sx_ready(i) := !s3_valid(i) || sx_in(i).output.uop.robIdx.needFlush(io.redirect) || (if (TotalDelayCycles == 0) io.stout.ready else sx_ready(i+1))
508    } else {
509      val cur_kill   = sx_in(i).output.uop.robIdx.needFlush(io.redirect)
510      val cur_can_go = (if (i == TotalDelayCycles) io.stout.ready else sx_ready(i+1))
511      val cur_fire   = sx_valid(i) && !cur_kill && cur_can_go
512      val prev_fire  = sx_valid(i-1) && !sx_in(i-1).output.uop.robIdx.needFlush(io.redirect) && sx_ready(i)
513
514      sx_ready(i) := !sx_valid(i) || cur_kill || (if (i == TotalDelayCycles) io.stout.ready else sx_ready(i+1))
515      val sx_valid_can_go = prev_fire || cur_fire || cur_kill
516      sx_valid(i) := RegEnable(Mux(prev_fire, true.B, false.B), false.B, sx_valid_can_go)
517      sx_in(i) := RegEnable(sx_in(i-1), prev_fire)
518    }
519  }
520  val sx_last_valid = sx_valid.takeRight(1).head
521  val sx_last_ready = sx_ready.takeRight(1).head
522  val sx_last_in    = sx_in.takeRight(1).head
523  sx_last_ready := !sx_last_valid || sx_last_in.output.uop.robIdx.needFlush(io.redirect) || io.stout.ready
524
525  io.stout.valid := sx_last_valid && !sx_last_in.output.uop.robIdx.needFlush(io.redirect) && isStore(sx_last_in.output.uop.fuType)
526  io.stout.bits := sx_last_in.output
527  io.stout.bits.uop.exceptionVec := ExceptionNO.selectByFu(sx_last_in.output.uop.exceptionVec, StaCfg)
528
529  io.vecstout.valid := sx_last_valid && !sx_last_in.output.uop.robIdx.needFlush(io.redirect) && isVStore(sx_last_in.output.uop.fuType)
530  // TODO: implement it!
531  io.vecstout.bits.mBIndex := sx_last_in.mbIndex
532  io.vecstout.bits.hit := sx_last_in.vecFeedback
533  io.vecstout.bits.isvec := true.B
534  io.vecstout.bits.sourceType := RSFeedbackType.tlbMiss
535  io.vecstout.bits.flushState := DontCare
536  io.vecstout.bits.mmio := sx_last_in.mmio
537  io.vecstout.bits.exceptionVec := ExceptionNO.selectByFu(sx_last_in.output.uop.exceptionVec, VstuCfg)
538  io.vecstout.bits.usSecondInv := sx_last_in.usSecondInv
539  io.vecstout.bits.vecFeedback := sx_last_in.vecFeedback
540  io.vecstout.bits.elemIdx     := sx_last_in.elemIdx
541  io.vecstout.bits.alignedType := sx_last_in.alignedType
542  io.vecstout.bits.mask        := sx_last_in.mask
543  io.vecstout.bits.vaddr       := sx_last_in.vaddr
544  io.vecstout.bits.gpaddr      := sx_last_in.gpaddr
545  // io.vecstout.bits.reg_offset.map(_ := DontCare)
546  // io.vecstout.bits.elemIdx.map(_ := sx_last_in.elemIdx)
547  // io.vecstout.bits.elemIdxInsideVd.map(_ := DontCare)
548  // io.vecstout.bits.vecdata.map(_ := DontCare)
549  // io.vecstout.bits.mask.map(_ := DontCare)
550  // io.vecstout.bits.alignedType.map(_ := sx_last_in.alignedType)
551
552  io.debug_ls := DontCare
553  io.debug_ls.s1_robIdx := s1_in.uop.robIdx.value
554  io.debug_ls.s1_isTlbFirstMiss := io.tlb.resp.valid && io.tlb.resp.bits.miss && io.tlb.resp.bits.debug.isFirstIssue && !s1_in.isHWPrefetch
555
556  private def printPipeLine(pipeline: LsPipelineBundle, cond: Bool, name: String): Unit = {
557    XSDebug(cond,
558      p"$name" + p" pc ${Hexadecimal(pipeline.uop.pc)} " +
559        p"addr ${Hexadecimal(pipeline.vaddr)} -> ${Hexadecimal(pipeline.paddr)} " +
560        p"op ${Binary(pipeline.uop.fuOpType)} " +
561        p"data ${Hexadecimal(pipeline.data)} " +
562        p"mask ${Hexadecimal(pipeline.mask)}\n"
563    )
564  }
565
566  printPipeLine(s0_out, s0_valid, "S0")
567  printPipeLine(s1_out, s1_valid, "S1")
568
569  // perf cnt
570  XSPerfAccumulate("s0_in_valid",                s0_valid)
571  XSPerfAccumulate("s0_in_fire",                 s0_fire)
572  XSPerfAccumulate("s0_vecin_fire",              s0_fire && s0_use_flow_vec)
573  XSPerfAccumulate("s0_in_fire_first_issue",     s0_fire && s0_isFirstIssue)
574  XSPerfAccumulate("s0_addr_spec_success",       s0_fire && !s0_use_flow_vec && s0_saddr(VAddrBits-1, 12) === s0_stin.src(0)(VAddrBits-1, 12))
575  XSPerfAccumulate("s0_addr_spec_failed",        s0_fire && !s0_use_flow_vec && s0_saddr(VAddrBits-1, 12) =/= s0_stin.src(0)(VAddrBits-1, 12))
576  XSPerfAccumulate("s0_addr_spec_success_once",  s0_fire && !s0_use_flow_vec && s0_saddr(VAddrBits-1, 12) === s0_stin.src(0)(VAddrBits-1, 12) && s0_isFirstIssue)
577  XSPerfAccumulate("s0_addr_spec_failed_once",   s0_fire && !s0_use_flow_vec && s0_saddr(VAddrBits-1, 12) =/= s0_stin.src(0)(VAddrBits-1, 12) && s0_isFirstIssue)
578
579  XSPerfAccumulate("s1_in_valid",                s1_valid)
580  XSPerfAccumulate("s1_in_fire",                 s1_fire)
581  XSPerfAccumulate("s1_in_fire_first_issue",     s1_fire && s1_in.isFirstIssue)
582  XSPerfAccumulate("s1_tlb_miss",                s1_fire && s1_tlb_miss)
583  XSPerfAccumulate("s1_tlb_miss_first_issue",    s1_fire && s1_tlb_miss && s1_in.isFirstIssue)
584  // end
585}
586