xref: /XiangShan/src/main/scala/xiangshan/mem/pipeline/StoreUnit.scala (revision 5299b43275f62c4fe2e4fde91bc30c97f5be89a2)
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._
25import xiangshan.ExceptionNO._
26import xiangshan.backend.Bundles.{MemExuInput, MemExuOutput, connectSamePort}
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.mem.Bundles._
33import xiangshan.cache.mmu.{Pbmt, TlbCmd, TlbReq, TlbRequestIO, TlbResp}
34import xiangshan.cache.{DCacheStoreIO, DcacheStoreRequestIO, HasDCacheParameters, MemoryOpConstants, StorePrefetchReq}
35
36class StoreUnit(implicit p: Parameters) extends XSModule
37  with HasDCacheParameters
38  with HasVLSUParameters
39  {
40  val io = IO(new Bundle() {
41    val redirect        = Flipped(ValidIO(new Redirect))
42    val csrCtrl         = Flipped(new CustomCSRCtrlIO)
43    val stin            = Flipped(Decoupled(new MemExuInput))
44    val issue           = Valid(new MemExuInput)
45    // misalignBuffer issue path
46    val misalign_stin   = Flipped(Decoupled(new LsPipelineBundle))
47    val misalign_stout  = Valid(new SqWriteBundle)
48    val tlb             = new TlbRequestIO()
49    val dcache          = new DCacheStoreIO
50    val pmp             = Flipped(new PMPRespBundle())
51    val lsq             = ValidIO(new LsPipelineBundle)
52    val lsq_replenish   = Output(new LsPipelineBundle())
53    val feedback_slow   = ValidIO(new RSFeedback)
54    val prefetch_req    = Flipped(DecoupledIO(new StorePrefetchReq))
55    // provide prefetch info to sms
56    val prefetch_train  = ValidIO(new StPrefetchTrainBundle())
57    // speculative for gated control
58    val s1_prefetch_spec = Output(Bool())
59    val s2_prefetch_spec = Output(Bool())
60    val stld_nuke_query = Valid(new StoreNukeQueryIO)
61    val stout           = DecoupledIO(new MemExuOutput) // writeback store
62    val vecstout        = DecoupledIO(new VecPipelineFeedbackIO(isVStore = true))
63    // store mask, send to sq in store_s0
64    val st_mask_out     = Valid(new StoreMaskBundle)
65    val debug_ls        = Output(new DebugLsInfoBundle)
66    // vector
67    val vecstin           = Flipped(Decoupled(new VecPipeBundle(isVStore = true)))
68    val vec_isFirstIssue  = Input(Bool())
69    // writeback to misalign buffer
70    val misalign_buf = Decoupled(new LsPipelineBundle)
71    // trigger
72    val fromCsrTrigger = Input(new CsrTriggerBundle)
73
74    val s0_s1_valid = Output(Bool())
75  })
76
77  val s1_ready, s2_ready, s3_ready = WireInit(false.B)
78
79  // Pipeline
80  // --------------------------------------------------------------------------------
81  // stage 0
82  // --------------------------------------------------------------------------------
83  // generate addr, use addr to query DCache and DTLB
84  val s0_iss_valid        = io.stin.valid
85  val s0_prf_valid        = io.prefetch_req.valid && io.dcache.req.ready
86  val s0_vec_valid        = io.vecstin.valid
87  val s0_ma_st_valid      = io.misalign_stin.valid
88  val s0_valid            = s0_iss_valid || s0_prf_valid || s0_vec_valid || s0_ma_st_valid
89  val s0_use_flow_ma      = s0_ma_st_valid
90  val s0_use_flow_vec     = s0_vec_valid && !s0_ma_st_valid
91  val s0_use_flow_rs      = s0_iss_valid && !s0_vec_valid && !s0_ma_st_valid
92  val s0_use_flow_prf     = s0_prf_valid && !s0_iss_valid && !s0_vec_valid && !s0_ma_st_valid
93  val s0_use_non_prf_flow = s0_use_flow_rs || s0_use_flow_vec || s0_use_flow_ma
94  val s0_stin             = Mux(s0_use_flow_rs, io.stin.bits, 0.U.asTypeOf(io.stin.bits))
95  val s0_vecstin          = Mux(s0_use_flow_vec, io.vecstin.bits, 0.U.asTypeOf(io.vecstin.bits))
96  val s0_uop              = Mux(
97    s0_use_flow_ma,
98    io.misalign_stin.bits.uop,
99    Mux(
100      s0_use_flow_rs,
101      s0_stin.uop,
102      s0_vecstin.uop
103    )
104  )
105  val s0_isFirstIssue = Mux(
106    s0_use_flow_ma,
107    false.B,
108    s0_use_flow_rs && io.stin.bits.isFirstIssue || s0_use_flow_vec && io.vec_isFirstIssue
109  )
110  val s0_size         = Mux(s0_use_non_prf_flow, s0_uop.fuOpType(2,0), 0.U)// may broken if use it in feature
111  val s0_mem_idx      = Mux(s0_use_non_prf_flow, s0_uop.sqIdx.value, 0.U)
112  val s0_rob_idx      = Mux(s0_use_non_prf_flow, s0_uop.robIdx, 0.U.asTypeOf(s0_uop.robIdx))
113  val s0_pc           = Mux(s0_use_non_prf_flow, s0_uop.pc, 0.U)
114  val s0_instr_type   = Mux(s0_use_non_prf_flow, STORE_SOURCE.U, DCACHE_PREFETCH_SOURCE.U)
115  val s0_wlineflag    = Mux(s0_use_flow_rs, s0_uop.fuOpType === LSUOpType.cbo_zero, false.B)
116  val s0_out          = Wire(new LsPipelineBundle)
117  val s0_kill         = s0_uop.robIdx.needFlush(io.redirect)
118  val s0_can_go       = s1_ready
119  val s0_fire         = s0_valid && !s0_kill && s0_can_go
120  val s0_is128bit     = Wire(Bool())
121  // vector
122  val s0_vecActive    = !s0_use_flow_vec || s0_vecstin.vecActive
123  // val s0_flowPtr      = s0_vecstin.flowPtr
124  // val s0_isLastElem   = s0_vecstin.isLastElem
125  val s0_secondInv    = s0_vecstin.usSecondInv
126  val s0_elemIdx      = s0_vecstin.elemIdx
127  val s0_alignedType  = s0_vecstin.alignedType
128  val s0_mBIndex      = s0_vecstin.mBIndex
129  val s0_vecBaseVaddr = s0_vecstin.basevaddr
130  val s0_isFinalSplit = io.misalign_stin.valid && io.misalign_stin.bits.isFinalSplit
131
132  // generate addr
133  val s0_saddr = s0_stin.src(0) + SignExt(s0_stin.uop.imm(11,0), VAddrBits)
134  val s0_fullva = Wire(UInt(XLEN.W))
135
136  val s0_vaddr = Mux(
137    s0_use_flow_ma,
138    io.misalign_stin.bits.vaddr,
139    Mux(
140      s0_use_flow_rs,
141      s0_saddr,
142      Mux(
143        s0_use_flow_vec,
144        s0_vecstin.vaddr(VAddrBits - 1, 0),
145        io.prefetch_req.bits.vaddr
146      )
147    )
148  )
149
150    val s0_alignTpye = Mux(s0_use_flow_vec, s0_vecstin.alignedType(1,0), s0_uop.fuOpType(1, 0))
151  // exception check
152  val s0_addr_aligned = LookupTree(s0_alignTpye, List(
153    "b00".U   -> true.B,              //b
154    "b01".U   -> (s0_vaddr(0) === 0.U),   //h
155    "b10".U   -> (s0_vaddr(1,0) === 0.U), //w
156    "b11".U   -> (s0_vaddr(2,0) === 0.U)  //d
157  ))
158  // if vector store sends 128-bit requests, its address must be 128-aligned
159  XSError(s0_use_flow_vec && s0_vaddr(3, 0) =/= 0.U && s0_vecstin.alignedType(2), "unit stride 128 bit element is not aligned!")
160
161  val s0_isMisalign = Mux(s0_use_non_prf_flow, (!s0_addr_aligned || s0_vecstin.uop.exceptionVec(storeAddrMisaligned) && s0_vecActive), false.B)
162  val s0_addr_low = s0_vaddr(4, 0)
163  val s0_addr_Up_low = LookupTree(s0_alignTpye, List(
164    "b00".U -> 0.U,
165    "b01".U -> 1.U,
166    "b10".U -> 3.U,
167    "b11".U -> 7.U
168  )) + s0_addr_low
169  val s0_rs_corss16Bytes = s0_addr_Up_low(4) =/= s0_addr_low(4)
170  val s0_misalignWith16Byte = !s0_rs_corss16Bytes && !s0_addr_aligned && !s0_use_flow_prf
171  s0_is128bit := Mux(s0_use_flow_ma, io.misalign_stin.bits.is128bit, is128Bit(s0_vecstin.alignedType) || s0_misalignWith16Byte)
172
173  s0_fullva := Mux(
174    s0_use_flow_rs,
175    s0_stin.src(0) + SignExt(s0_stin.uop.imm(11,0), XLEN),
176    Mux(
177      s0_use_flow_vec,
178      s0_vecstin.vaddr,
179      s0_vaddr
180    )
181  )
182
183  val s0_mask = Mux(
184    s0_use_flow_ma,
185    io.misalign_stin.bits.mask,
186    Mux(
187      s0_use_flow_rs,
188      genVWmask128(s0_saddr, s0_uop.fuOpType(2,0)),
189      Mux(
190        s0_use_flow_vec,
191        s0_vecstin.mask,
192        // -1.asSInt.asUInt
193        Fill(VLEN/8, 1.U(1.W))
194      )
195    )
196  )
197
198  io.tlb.req.valid                   := s0_valid
199  io.tlb.req.bits.vaddr              := s0_vaddr
200  io.tlb.req.bits.fullva             := s0_fullva
201  io.tlb.req.bits.checkfullva        := s0_use_flow_rs || s0_use_flow_vec
202  io.tlb.req.bits.cmd                := TlbCmd.write
203  io.tlb.req.bits.isPrefetch         := s0_use_flow_prf
204  io.tlb.req.bits.size               := s0_size
205  io.tlb.req.bits.kill               := false.B
206  io.tlb.req.bits.memidx.is_ld       := false.B
207  io.tlb.req.bits.memidx.is_st       := true.B
208  io.tlb.req.bits.memidx.idx         := s0_mem_idx
209  io.tlb.req.bits.debug.robIdx       := s0_rob_idx
210  io.tlb.req.bits.no_translate       := false.B
211  io.tlb.req.bits.debug.pc           := s0_pc
212  io.tlb.req.bits.debug.isFirstIssue := s0_isFirstIssue
213  io.tlb.req_kill                    := false.B
214  io.tlb.req.bits.hyperinst          := LSUOpType.isHsv(s0_uop.fuOpType)
215  io.tlb.req.bits.hlvx               := false.B
216  io.tlb.req.bits.pmp_addr           := DontCare
217
218  // Dcache access here: not **real** dcache write
219  // just read meta and tag in dcache, to find out the store will hit or miss
220
221  // NOTE: The store request does not wait for the dcache to be ready.
222  //       If the dcache is not ready at this time, the dcache is not queried.
223  //       But, store prefetch request will always wait for dcache to be ready to make progress.
224  io.dcache.req.valid              := s0_fire
225  io.dcache.req.bits.cmd           := MemoryOpConstants.M_PFW
226  io.dcache.req.bits.vaddr         := s0_vaddr
227  io.dcache.req.bits.instrtype     := s0_instr_type
228
229  s0_out              := DontCare
230  s0_out.vaddr        := s0_vaddr
231  s0_out.fullva       := s0_fullva
232  // Now data use its own io
233  s0_out.data         := s0_stin.src(1)
234  s0_out.uop          := s0_uop
235  s0_out.miss         := false.B
236  // For unaligned, we need to generate a base-aligned mask in storeunit and then do a shift split in StoreQueue.
237  s0_out.mask         := Mux(s0_rs_corss16Bytes && !s0_addr_aligned, genBasemask(s0_saddr,s0_alignTpye(1,0)), s0_mask)
238  s0_out.isFirstIssue := s0_isFirstIssue
239  s0_out.isHWPrefetch := s0_use_flow_prf
240  s0_out.wlineflag    := s0_wlineflag
241  s0_out.isvec        := s0_use_flow_vec
242  s0_out.is128bit     := s0_is128bit
243  s0_out.vecActive    := s0_vecActive
244  s0_out.usSecondInv  := s0_secondInv
245  s0_out.elemIdx      := s0_elemIdx
246  s0_out.alignedType  := s0_alignedType
247  s0_out.mbIndex      := s0_mBIndex
248  s0_out.misalignWith16Byte      := s0_misalignWith16Byte
249  s0_out.isMisalign      := s0_isMisalign
250  s0_out.vecBaseVaddr := s0_vecBaseVaddr
251  when(s0_valid && s0_isFirstIssue) {
252    s0_out.uop.debugInfo.tlbFirstReqTime := GTimer()
253  }
254  s0_out.isFrmMisAlignBuf := s0_use_flow_ma
255  s0_out.isFinalSplit := s0_isFinalSplit
256//  s0_out.uop.exceptionVec(storeAddrMisaligned) := Mux(s0_use_non_prf_flow, (!s0_addr_aligned || s0_vecstin.uop.exceptionVec(storeAddrMisaligned) && s0_vecActive), false.B) && !s0_misalignWith16Byte
257
258  io.st_mask_out.valid       := s0_use_flow_rs || s0_use_flow_vec
259  io.st_mask_out.bits.mask   := s0_out.mask
260  io.st_mask_out.bits.sqIdx  := s0_out.uop.sqIdx
261
262  io.stin.ready := s1_ready && s0_use_flow_rs
263  io.vecstin.ready := s1_ready && s0_use_flow_vec
264  io.prefetch_req.ready := s1_ready && io.dcache.req.ready && !s0_iss_valid && !s0_vec_valid && !s0_ma_st_valid
265  io.misalign_stin.ready := s1_ready && s0_use_flow_ma
266
267  // Pipeline
268  // --------------------------------------------------------------------------------
269  // stage 1
270  // --------------------------------------------------------------------------------
271  // TLB resp (send paddr to dcache)
272  val s1_valid  = RegInit(false.B)
273  val s1_in     = RegEnable(s0_out, s0_fire)
274  val s1_out    = Wire(new LsPipelineBundle)
275  val s1_kill   = Wire(Bool())
276  val s1_can_go = s2_ready
277  val s1_fire   = s1_valid && !s1_kill && s1_can_go
278  val s1_vecActive    = RegEnable(s0_out.vecActive, true.B, s0_fire)
279  val s1_frm_mabuf    = s1_in.isFrmMisAlignBuf
280  val s1_is128bit     = s1_in.is128bit
281
282  // mmio cbo decoder
283  val s1_isCbo   = s1_in.uop.fuOpType === LSUOpType.cbo_clean ||
284                   s1_in.uop.fuOpType === LSUOpType.cbo_flush ||
285                   s1_in.uop.fuOpType === LSUOpType.cbo_inval ||
286                   s1_in.uop.fuOpType === LSUOpType.cbo_zero
287  val s1_vaNeedExt = io.tlb.resp.bits.excp(0).vaNeedExt
288  val s1_isHyper   = io.tlb.resp.bits.excp(0).isHyper
289  val s1_paddr     = io.tlb.resp.bits.paddr(0)
290  val s1_gpaddr    = io.tlb.resp.bits.gpaddr(0)
291  val s1_fullva    = io.tlb.resp.bits.fullva
292  val s1_isForVSnonLeafPTE   = io.tlb.resp.bits.isForVSnonLeafPTE
293  val s1_tlb_miss  = io.tlb.resp.bits.miss && io.tlb.resp.valid && s1_valid
294  val s1_pbmt      = Mux(!s1_tlb_miss, io.tlb.resp.bits.pbmt.head, 0.U(Pbmt.width.W))
295  val s1_exception = ExceptionNO.selectByFu(s1_out.uop.exceptionVec, StaCfg).asUInt.orR
296  val s1_isvec     = RegEnable(s0_out.isvec, false.B, s0_fire)
297  //We don't want `StoreUnit` to have an additional effect on the Store of vector from a `misalignBuffer,`
298  //But there are places where a marker bit is needed to enable additional processing of vector instructions.
299  //For example: `StoreQueue` is exceptionBuffer
300  val s1_frm_mab_vec = RegEnable(s0_use_flow_ma && io.misalign_stin.bits.isvec, false.B, s0_fire)
301  // val s1_isLastElem = RegEnable(s0_isLastElem, false.B, s0_fire)
302  s1_kill := s1_in.uop.robIdx.needFlush(io.redirect) || (s1_tlb_miss && !s1_isvec && !s1_frm_mabuf)
303
304  s1_ready := !s1_valid || s1_kill || s2_ready
305  io.tlb.resp.ready := true.B // TODO: why dtlbResp needs a ready?
306  when (s0_fire) { s1_valid := true.B }
307  .elsewhen (s1_fire) { s1_valid := false.B }
308  .elsewhen (s1_kill) { s1_valid := false.B }
309
310  // st-ld violation dectect request.
311  io.stld_nuke_query.valid       := s1_valid && !s1_tlb_miss && !s1_in.isHWPrefetch
312  io.stld_nuke_query.bits.robIdx := s1_in.uop.robIdx
313  io.stld_nuke_query.bits.paddr  := s1_paddr
314  io.stld_nuke_query.bits.mask   := s1_in.mask
315  io.stld_nuke_query.bits.matchLine := (s1_in.isvec || s1_in.misalignWith16Byte) && s1_in.is128bit
316
317  // issue
318  io.issue.valid := s1_valid && !s1_tlb_miss && !s1_in.isHWPrefetch && !s1_isvec && !s1_frm_mabuf
319  io.issue.bits  := RegEnable(s0_stin, s0_valid)
320
321  // trigger
322  val storeTrigger = Module(new MemTrigger(MemType.STORE))
323  storeTrigger.io.fromCsrTrigger.tdataVec             := io.fromCsrTrigger.tdataVec
324  storeTrigger.io.fromCsrTrigger.tEnableVec           := io.fromCsrTrigger.tEnableVec
325  storeTrigger.io.fromCsrTrigger.triggerCanRaiseBpExp := io.fromCsrTrigger.triggerCanRaiseBpExp
326  storeTrigger.io.fromCsrTrigger.debugMode            := io.fromCsrTrigger.debugMode
327  storeTrigger.io.fromLoadStore.vaddr                 := s1_in.vaddr
328  storeTrigger.io.fromLoadStore.isVectorUnitStride    := s1_in.isvec && s1_in.is128bit
329  storeTrigger.io.fromLoadStore.mask                  := s1_in.mask
330
331  val s1_trigger_action = storeTrigger.io.toLoadStore.triggerAction
332  val s1_trigger_debug_mode = TriggerAction.isDmode(s1_trigger_action)
333  val s1_trigger_breakpoint = TriggerAction.isExp(s1_trigger_action)
334
335  // for misalign in vsMergeBuffer
336  io.s0_s1_valid := s0_valid || s1_valid
337
338  // Send TLB feedback to store issue queue
339  // Store feedback is generated in store_s1, sent to RS in store_s2
340  val s1_feedback = Wire(Valid(new RSFeedback))
341  s1_feedback.valid                 := s1_valid & !s1_in.isHWPrefetch
342  s1_feedback.bits.hit              := !s1_tlb_miss
343  s1_feedback.bits.flushState       := io.tlb.resp.bits.ptwBack
344  s1_feedback.bits.robIdx           := s1_out.uop.robIdx
345  s1_feedback.bits.sourceType       := RSFeedbackType.tlbMiss
346  s1_feedback.bits.dataInvalidSqIdx := DontCare
347  s1_feedback.bits.sqIdx            := s1_out.uop.sqIdx
348  s1_feedback.bits.lqIdx            := s1_out.uop.lqIdx
349
350  XSDebug(s1_feedback.valid,
351    "S1 Store: tlbHit: %d robIdx: %d\n",
352    s1_feedback.bits.hit,
353    s1_feedback.bits.robIdx.value
354  )
355
356  // io.feedback_slow := s1_feedback
357
358  // get paddr from dtlb, check if rollback is needed
359  // writeback store inst to lsq
360  s1_out           := s1_in
361  s1_out.paddr     := s1_paddr
362  s1_out.gpaddr    := s1_gpaddr
363  s1_out.fullva    := s1_fullva
364  s1_out.vaNeedExt := s1_vaNeedExt
365  s1_out.isHyper   := s1_isHyper
366  s1_out.miss      := false.B
367  s1_out.nc        := Pbmt.isNC(s1_pbmt)
368  s1_out.mmio      := Pbmt.isIO(s1_pbmt)
369  s1_out.tlbMiss   := s1_tlb_miss
370  s1_out.atomic    := Pbmt.isIO(s1_pbmt)
371  s1_out.isForVSnonLeafPTE := s1_isForVSnonLeafPTE
372  when (RegNext(io.tlb.req.bits.checkfullva) &&
373    (s1_out.uop.exceptionVec(storePageFault) ||
374      s1_out.uop.exceptionVec(storeAccessFault) ||
375      s1_out.uop.exceptionVec(storeGuestPageFault))) {
376    s1_out.uop.exceptionVec(storeAddrMisaligned) := false.B
377  }
378  s1_out.uop.exceptionVec(storePageFault)      := io.tlb.resp.bits.excp(0).pf.st && s1_vecActive
379  s1_out.uop.exceptionVec(storeAccessFault)    := io.tlb.resp.bits.excp(0).af.st && s1_vecActive
380  s1_out.uop.exceptionVec(storeGuestPageFault) := io.tlb.resp.bits.excp(0).gpf.st && s1_vecActive
381
382  s1_out.uop.flushPipe                := false.B
383  s1_out.uop.trigger                  := s1_trigger_action
384  s1_out.uop.exceptionVec(breakPoint) := s1_trigger_breakpoint
385  s1_out.uop.exceptionVec(storeAddrMisaligned) := s1_out.mmio && s1_in.isMisalign
386  s1_out.vecVaddrOffset := Mux(
387    s1_trigger_debug_mode || s1_trigger_breakpoint,
388    storeTrigger.io.toLoadStore.triggerVaddr - s1_in.vecBaseVaddr,
389    s1_in.vaddr + genVFirstUnmask(s1_in.mask).asUInt - s1_in.vecBaseVaddr ,
390  )
391  s1_out.vecTriggerMask := Mux(s1_trigger_debug_mode || s1_trigger_breakpoint, storeTrigger.io.toLoadStore.triggerMask, 0.U)
392
393  // scalar store and scalar load nuke check, and also other purposes
394  //A 128-bit aligned unaligned memory access requires changing the unaligned flag bit in sq
395  io.lsq.valid     := s1_valid && !s1_in.isHWPrefetch
396  io.lsq.bits      := s1_out
397  io.lsq.bits.miss := s1_tlb_miss
398  io.lsq.bits.isvec := s1_out.isvec || s1_frm_mab_vec
399  io.lsq.bits.updateAddrValid := (!s1_in.isMisalign || s1_in.misalignWith16Byte) && (!s1_frm_mabuf || s1_in.isFinalSplit) || s1_exception
400  // kill dcache write intent request when tlb miss or exception
401  io.dcache.s1_kill  := (s1_tlb_miss || s1_exception || s1_out.mmio || s1_in.uop.robIdx.needFlush(io.redirect))
402  io.dcache.s1_paddr := s1_paddr
403
404  // write below io.out.bits assign sentence to prevent overwriting values
405  val s1_tlb_memidx = io.tlb.resp.bits.memidx
406  when(s1_tlb_memidx.is_st && io.tlb.resp.valid && !s1_tlb_miss && s1_tlb_memidx.idx === s1_out.uop.sqIdx.value) {
407    // printf("Store idx = %d\n", s1_tlb_memidx.idx)
408    s1_out.uop.debugInfo.tlbRespTime := GTimer()
409  }
410  val s1_mis_align = s1_valid && !s1_tlb_miss && !s1_in.isHWPrefetch && !s1_isCbo && !s1_out.nc && !s1_out.mmio &&
411                      GatedValidRegNext(io.csrCtrl.hd_misalign_st_enable) && s1_in.isMisalign && !s1_in.misalignWith16Byte &&
412                      !s1_trigger_breakpoint && !s1_trigger_debug_mode
413
414  // Pipeline
415  // --------------------------------------------------------------------------------
416  // stage 2
417  // --------------------------------------------------------------------------------
418  // mmio check
419  val s2_valid  = RegInit(false.B)
420  val s2_in     = RegEnable(s1_out, s1_fire)
421  val s2_out    = Wire(new LsPipelineBundle)
422  val s2_kill   = Wire(Bool())
423  val s2_can_go = s3_ready
424  val s2_fire   = s2_valid && !s2_kill && s2_can_go
425  val s2_vecActive    = RegEnable(s1_out.vecActive, true.B, s1_fire)
426  val s2_frm_mabuf    = s2_in.isFrmMisAlignBuf
427  val s2_frm_mab_vec  = RegEnable(s1_frm_mab_vec, true.B, s1_fire)
428  val s2_pbmt   = RegEnable(s1_pbmt, s1_fire)
429  val s2_trigger_debug_mode = RegEnable(s1_trigger_debug_mode, false.B, s1_fire)
430
431  s2_ready := !s2_valid || s2_kill || s3_ready
432  when (s1_fire) { s2_valid := true.B }
433  .elsewhen (s2_fire) { s2_valid := false.B }
434  .elsewhen (s2_kill) { s2_valid := false.B }
435
436  val s2_pmp = WireInit(io.pmp)
437
438  val s2_exception = RegNext(s1_feedback.bits.hit) &&
439                    (s2_trigger_debug_mode || ExceptionNO.selectByFu(s2_out.uop.exceptionVec, StaCfg).asUInt.orR) && s2_vecActive
440  val s2_un_misalign_exception =  RegNext(s1_feedback.bits.hit) &&
441                    (s2_trigger_debug_mode || ExceptionNO.selectByFuAndUnSelect(s2_out.uop.exceptionVec, StaCfg, Seq(storeAddrMisaligned)).asUInt.orR)
442
443  val s2_mmio = (s2_in.mmio || (Pbmt.isPMA(s2_pbmt) && s2_pmp.mmio)) && RegNext(s1_feedback.bits.hit)
444  val s2_memBackTypeMM = !s2_pmp.mmio
445  val s2_actually_uncache = (Pbmt.isPMA(s2_pbmt) && s2_pmp.mmio || s2_in.nc || s2_in.mmio) && RegNext(s1_feedback.bits.hit)
446  val s2_uncache = !s2_exception && !s2_in.tlbMiss && s2_actually_uncache
447  val s2_isCbo  = RegEnable(s1_isCbo, s1_fire) // all cbo instr
448  val s2_isCbo_noZero = LSUOpType.isCbo(s2_in.uop.fuOpType)
449
450  s2_kill := ((s2_mmio && !s2_exception) && !s2_in.isvec && !s2_frm_mabuf) || s2_in.uop.robIdx.needFlush(io.redirect)
451
452  s2_out        := s2_in
453  s2_out.af     := s2_out.uop.exceptionVec(storeAccessFault)
454  s2_out.mmio   := s2_mmio && !s2_exception
455  s2_out.atomic := s2_in.atomic || Pbmt.isPMA(s2_pbmt) && s2_pmp.atomic
456  s2_out.memBackTypeMM := s2_memBackTypeMM
457  s2_out.uop.exceptionVec(storeAccessFault) := (s2_in.uop.exceptionVec(storeAccessFault) ||
458                                                s2_pmp.st ||
459                                                ((s2_in.isvec || s2_frm_mabuf || s2_isCbo) && s2_actually_uncache && RegNext(s1_feedback.bits.hit))
460                                                ) && s2_vecActive
461  s2_out.uop.exceptionVec(storeAddrMisaligned) := s2_mmio && s2_in.isMisalign && !s2_un_misalign_exception
462  s2_out.uop.vpu.vstart     := s2_in.vecVaddrOffset >> s2_in.uop.vpu.veew
463
464  // kill dcache write intent request when mmio or exception
465  io.dcache.s2_kill := (s2_uncache || s2_exception || s2_in.uop.robIdx.needFlush(io.redirect))
466  io.dcache.s2_pc   := s2_out.uop.pc
467  // TODO: dcache resp
468  io.dcache.resp.ready := true.B
469
470  val s2_mis_align = s2_valid && RegEnable(s1_mis_align, s1_fire) && !s2_exception
471  // goto misalignBuffer
472  val toMisalignBufferValid =  s2_mis_align && !s2_frm_mabuf
473  io.misalign_buf.valid := toMisalignBufferValid
474  io.misalign_buf.bits  := s2_in
475  val misalignBufferNack = toMisalignBufferValid && !io.misalign_buf.ready
476
477  // feedback tlb miss to RS in store_s2
478  val feedback_slow_valid = WireInit(false.B)
479
480  feedback_slow_valid := s1_feedback.valid && !s1_out.uop.robIdx.needFlush(io.redirect) && !s1_out.isvec && !s1_frm_mabuf
481  io.feedback_slow.valid := GatedValidRegNext(feedback_slow_valid)
482  io.feedback_slow.bits  := RegEnable(s1_feedback.bits, feedback_slow_valid)
483  io.feedback_slow.bits.hit  := RegEnable(s1_feedback.bits.hit, feedback_slow_valid) && !misalignBufferNack
484
485  val s2_vecFeedback = RegNext(!s1_out.uop.robIdx.needFlush(io.redirect) && s1_feedback.bits.hit && s1_feedback.valid) &&
486                       !misalignBufferNack && s2_in.isvec && !s2_frm_mabuf
487
488  val s2_misalign_stout = WireInit(0.U.asTypeOf(io.misalign_stout))
489  s2_misalign_stout.valid := s2_valid && s2_can_go && s2_frm_mabuf
490  connectSamePort(s2_misalign_stout.bits, s2_out)
491  s2_misalign_stout.bits.need_rep := RegEnable(s1_tlb_miss, s1_fire)
492  io.misalign_stout := s2_misalign_stout
493
494  // mmio and exception
495  io.lsq_replenish := s2_out
496  io.lsq_replenish.af := s2_out.af && s2_valid && !s2_kill
497  io.lsq_replenish.mmio := (s2_mmio || s2_isCbo_noZero) && !s2_exception // reuse `mmiostall` logic in sq
498
499  // prefetch related
500  io.lsq_replenish.miss := io.dcache.resp.fire && io.dcache.resp.bits.miss // miss info
501  io.lsq_replenish.updateAddrValid := !s2_mis_align && (!s2_frm_mabuf || s2_out.isFinalSplit) || s2_exception
502  io.lsq_replenish.isvec := s2_out.isvec || s2_frm_mab_vec
503
504  io.lsq_replenish.hasException := (ExceptionNO.selectByFu(s2_out.uop.exceptionVec, StaCfg).asUInt.orR ||
505    TriggerAction.isDmode(s2_out.uop.trigger) || s2_out.af) && s2_valid && !s2_kill
506
507
508  // RegNext prefetch train for better timing
509  // ** Now, prefetch train is valid at store s3 **
510  val s2_prefetch_train_valid = WireInit(false.B)
511  s2_prefetch_train_valid := s2_valid && io.dcache.resp.fire && !s2_out.mmio && !s2_out.nc && !s2_in.tlbMiss && !s2_in.isHWPrefetch
512  if(EnableStorePrefetchSMS) {
513    io.s1_prefetch_spec := s1_fire
514    io.s2_prefetch_spec := s2_prefetch_train_valid
515    io.prefetch_train.valid := RegNext(s2_prefetch_train_valid)
516    io.prefetch_train.bits.fromLsPipelineBundle(s2_in, latch = true, enable = s2_prefetch_train_valid)
517  }else {
518    io.s1_prefetch_spec := false.B
519    io.s2_prefetch_spec := false.B
520    io.prefetch_train.valid := false.B
521    io.prefetch_train.bits.fromLsPipelineBundle(s2_in, latch = true, enable = false.B)
522  }
523  // override miss bit
524  io.prefetch_train.bits.miss := RegEnable(io.dcache.resp.bits.miss, s2_prefetch_train_valid)
525  // TODO: add prefetch and access bit
526  io.prefetch_train.bits.meta_prefetch := false.B
527  io.prefetch_train.bits.meta_access := false.B
528  io.prefetch_train.bits.isFinalSplit := false.B
529  io.prefetch_train.bits.misalignWith16Byte := false.B
530  io.prefetch_train.bits.isMisalign := false.B
531  io.prefetch_train.bits.misalignNeedWakeUp := false.B
532  io.prefetch_train.bits.updateAddrValid := false.B
533  io.prefetch_train.bits.hasException := false.B
534
535  // Pipeline
536  // --------------------------------------------------------------------------------
537  // stage 3
538  // --------------------------------------------------------------------------------
539  // store write back
540  val s3_valid  = RegInit(false.B)
541  val s3_in     = RegEnable(s2_out, s2_fire)
542  val s3_out    = Wire(new MemExuOutput(isVector = true))
543  val s3_kill   = s3_in.uop.robIdx.needFlush(io.redirect)
544  val s3_can_go = s3_ready
545  val s3_fire   = s3_valid && !s3_kill && s3_can_go
546  val s3_vecFeedback = RegEnable(s2_vecFeedback, s2_fire)
547  val s3_exception     = RegEnable(s2_exception, s2_fire)
548
549  // store misalign will not writeback to rob now
550  when (s2_fire) { s3_valid := (!s2_mmio && !s2_isCbo_noZero || s2_exception) && !s2_out.isHWPrefetch && !s2_mis_align && !s2_frm_mabuf }
551  .elsewhen (s3_fire) { s3_valid := false.B }
552  .elsewhen (s3_kill) { s3_valid := false.B }
553
554  // wb: writeback
555  val SelectGroupSize   = RollbackGroupSize
556  val lgSelectGroupSize = log2Ceil(SelectGroupSize)
557  val TotalSelectCycles = scala.math.ceil(log2Ceil(LoadQueueRAWSize).toFloat / lgSelectGroupSize).toInt + 1
558
559  s3_out                 := DontCare
560  s3_out.uop             := s3_in.uop
561  s3_out.data            := DontCare
562  s3_out.debug.isMMIO    := s3_in.mmio
563  s3_out.debug.isNC      := s3_in.nc
564  s3_out.debug.paddr     := s3_in.paddr
565  s3_out.debug.vaddr     := s3_in.vaddr
566  s3_out.debug.isPerfCnt := false.B
567
568  XSError(s3_valid && s3_in.isvec && s3_in.vecActive && !s3_in.mask.orR, "In vecActive, mask complement should not be 0")
569  // Pipeline
570  // --------------------------------------------------------------------------------
571  // stage x
572  // --------------------------------------------------------------------------------
573  // delay TotalSelectCycles - 2 cycle(s)
574  val TotalDelayCycles = TotalSelectCycles - 2
575  val sx_valid = Wire(Vec(TotalDelayCycles + 1, Bool()))
576  val sx_ready = Wire(Vec(TotalDelayCycles + 1, Bool()))
577  val sx_in    = Wire(Vec(TotalDelayCycles + 1, new VecMemExuOutput(isVector = true)))
578  val sx_in_vec = Wire(Vec(TotalDelayCycles +1, Bool()))
579
580  // backward ready signal
581  s3_ready := sx_ready.head
582  for (i <- 0 until TotalDelayCycles + 1) {
583    if (i == 0) {
584      sx_valid(i)          := s3_valid
585      sx_in(i).output      := s3_out
586      sx_in(i).vecFeedback := s3_vecFeedback
587      sx_in(i).nc          := s3_in.nc
588      sx_in(i).mmio        := s3_in.mmio
589      sx_in(i).usSecondInv := s3_in.usSecondInv
590      sx_in(i).elemIdx     := s3_in.elemIdx
591      sx_in(i).alignedType := s3_in.alignedType
592      sx_in(i).mbIndex     := s3_in.mbIndex
593      sx_in(i).mask        := s3_in.mask
594      sx_in(i).vaddr       := s3_in.fullva
595      sx_in(i).vaNeedExt   := s3_in.vaNeedExt
596      sx_in(i).gpaddr      := s3_in.gpaddr
597      sx_in(i).isForVSnonLeafPTE     := s3_in.isForVSnonLeafPTE
598      sx_in(i).vecTriggerMask := s3_in.vecTriggerMask
599      sx_in(i).hasException := s3_exception
600      sx_in_vec(i)         := s3_in.isvec
601      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))
602    } else {
603      val cur_kill   = sx_in(i).output.uop.robIdx.needFlush(io.redirect)
604      val cur_can_go = (if (i == TotalDelayCycles) io.stout.ready else sx_ready(i+1))
605      val cur_fire   = sx_valid(i) && !cur_kill && cur_can_go
606      val prev_fire  = sx_valid(i-1) && !sx_in(i-1).output.uop.robIdx.needFlush(io.redirect) && sx_ready(i)
607
608      sx_ready(i) := !sx_valid(i) || cur_kill || (if (i == TotalDelayCycles) io.stout.ready else sx_ready(i+1))
609      val sx_valid_can_go = prev_fire || cur_fire || cur_kill
610      sx_valid(i) := RegEnable(Mux(prev_fire, true.B, false.B), false.B, sx_valid_can_go)
611      sx_in(i) := RegEnable(sx_in(i-1), prev_fire)
612      sx_in_vec(i) := RegEnable(sx_in_vec(i-1), prev_fire)
613    }
614  }
615  val sx_last_valid = sx_valid.takeRight(1).head
616  val sx_last_ready = sx_ready.takeRight(1).head
617  val sx_last_in    = sx_in.takeRight(1).head
618  val sx_last_in_vec = sx_in_vec.takeRight(1).head
619  sx_last_ready := !sx_last_valid || sx_last_in.output.uop.robIdx.needFlush(io.redirect) || io.stout.ready
620
621  // write back: normal store, nc store
622  io.stout.valid := sx_last_valid && !sx_last_in_vec //isStore(sx_last_in.output.uop.fuType)
623  io.stout.bits := sx_last_in.output
624  io.stout.bits.uop.exceptionVec := ExceptionNO.selectByFu(sx_last_in.output.uop.exceptionVec, StaCfg)
625
626  io.vecstout.valid := sx_last_valid && sx_last_in_vec //isVStore(sx_last_in.output.uop.fuType)
627  // TODO: implement it!
628  io.vecstout.bits.mBIndex := sx_last_in.mbIndex
629  io.vecstout.bits.hit := sx_last_in.vecFeedback
630  io.vecstout.bits.isvec := true.B
631  io.vecstout.bits.sourceType := RSFeedbackType.tlbMiss
632  io.vecstout.bits.flushState := DontCare
633  io.vecstout.bits.trigger    := sx_last_in.output.uop.trigger
634  io.vecstout.bits.nc := sx_last_in.nc
635  io.vecstout.bits.mmio := sx_last_in.mmio
636  io.vecstout.bits.exceptionVec := ExceptionNO.selectByFu(sx_last_in.output.uop.exceptionVec, VstuCfg)
637  io.vecstout.bits.hasException := sx_last_in.hasException
638  io.vecstout.bits.usSecondInv := sx_last_in.usSecondInv
639  io.vecstout.bits.vecFeedback := sx_last_in.vecFeedback
640  io.vecstout.bits.elemIdx     := sx_last_in.elemIdx
641  io.vecstout.bits.alignedType := sx_last_in.alignedType
642  io.vecstout.bits.mask        := sx_last_in.mask
643  io.vecstout.bits.vaddr       := sx_last_in.vaddr
644  io.vecstout.bits.vaNeedExt   := sx_last_in.vaNeedExt
645  io.vecstout.bits.gpaddr      := sx_last_in.gpaddr
646  io.vecstout.bits.isForVSnonLeafPTE     := sx_last_in.isForVSnonLeafPTE
647  io.vecstout.bits.vstart      := sx_last_in.output.uop.vpu.vstart
648  io.vecstout.bits.vecTriggerMask := sx_last_in.vecTriggerMask
649  // io.vecstout.bits.reg_offset.map(_ := DontCare)
650  // io.vecstout.bits.elemIdx.map(_ := sx_last_in.elemIdx)
651  // io.vecstout.bits.elemIdxInsideVd.map(_ := DontCare)
652  // io.vecstout.bits.vecdata.map(_ := DontCare)
653  // io.vecstout.bits.mask.map(_ := DontCare)
654  // io.vecstout.bits.alignedType.map(_ := sx_last_in.alignedType)
655
656  io.debug_ls := DontCare
657  io.debug_ls.s1_robIdx := s1_in.uop.robIdx.value
658  io.debug_ls.s1_isTlbFirstMiss := io.tlb.resp.valid && io.tlb.resp.bits.miss && io.tlb.resp.bits.debug.isFirstIssue && !s1_in.isHWPrefetch
659
660  private def printPipeLine(pipeline: LsPipelineBundle, cond: Bool, name: String): Unit = {
661    XSDebug(cond,
662      p"$name" + p" pc ${Hexadecimal(pipeline.uop.pc)} " +
663        p"addr ${Hexadecimal(pipeline.vaddr)} -> ${Hexadecimal(pipeline.paddr)} " +
664        p"op ${Binary(pipeline.uop.fuOpType)} " +
665        p"data ${Hexadecimal(pipeline.data)} " +
666        p"mask ${Hexadecimal(pipeline.mask)}\n"
667    )
668  }
669
670  printPipeLine(s0_out, s0_valid, "S0")
671  printPipeLine(s1_out, s1_valid, "S1")
672
673  // perf cnt
674  XSPerfAccumulate("s0_in_valid",                s0_valid)
675  XSPerfAccumulate("s0_in_fire",                 s0_fire)
676  XSPerfAccumulate("s0_vecin_fire",              s0_fire && s0_use_flow_vec)
677  XSPerfAccumulate("s0_in_fire_first_issue",     s0_fire && s0_isFirstIssue)
678  XSPerfAccumulate("s0_addr_spec_success",       s0_fire && !s0_use_flow_vec && s0_saddr(VAddrBits-1, 12) === s0_stin.src(0)(VAddrBits-1, 12))
679  XSPerfAccumulate("s0_addr_spec_failed",        s0_fire && !s0_use_flow_vec && s0_saddr(VAddrBits-1, 12) =/= s0_stin.src(0)(VAddrBits-1, 12))
680  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)
681  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)
682
683  XSPerfAccumulate("s1_in_valid",                s1_valid)
684  XSPerfAccumulate("s1_in_fire",                 s1_fire)
685  XSPerfAccumulate("s1_in_fire_first_issue",     s1_fire && s1_in.isFirstIssue)
686  XSPerfAccumulate("s1_tlb_miss",                s1_fire && s1_tlb_miss)
687  XSPerfAccumulate("s1_tlb_miss_first_issue",    s1_fire && s1_tlb_miss && s1_in.isFirstIssue)
688  // end
689}
690