xref: /XiangShan/src/main/scala/xiangshan/mem/pipeline/StoreUnit.scala (revision 1eb8dd224d63ba7d4afa63695f72d8230e150d37)
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 LsPrefetchTrainBundle())
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 StoreNukeQueryBundle)
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_isCbo = s0_use_flow_rs && LSUOpType.isCboAll(s0_stin.uop.fuOpType)
151  // only simulation
152  val cbo_assert_flag = LSUOpType.isCboAll(s0_out.uop.fuOpType)
153  XSError(!s0_use_flow_rs && cbo_assert_flag && s0_valid, "cbo instruction selection error.")
154
155  val s0_alignType = Mux(s0_use_flow_vec, s0_vecstin.alignedType(1,0), s0_uop.fuOpType(1, 0))
156  // exception check
157  val s0_addr_aligned = LookupTree(s0_alignType, List(
158    "b00".U   -> true.B,              //b
159    "b01".U   -> (s0_vaddr(0) === 0.U),   //h
160    "b10".U   -> (s0_vaddr(1,0) === 0.U), //w
161    "b11".U   -> (s0_vaddr(2,0) === 0.U)  //d
162  )) || s0_isCbo
163  // if vector store sends 128-bit requests, its address must be 128-aligned
164  XSError(s0_use_flow_vec && s0_vaddr(3, 0) =/= 0.U && s0_vecstin.alignedType(2), "unit stride 128 bit element is not aligned!")
165
166  val s0_isMisalign = Mux(s0_use_non_prf_flow, (!s0_addr_aligned || s0_vecstin.uop.exceptionVec(storeAddrMisaligned) && s0_vecActive), false.B)
167  val s0_addr_low = s0_vaddr(4, 0)
168  val s0_addr_Up_low = LookupTree(s0_alignType, List(
169    "b00".U -> 0.U,
170    "b01".U -> 1.U,
171    "b10".U -> 3.U,
172    "b11".U -> 7.U
173  )) + s0_addr_low
174  val s0_rs_corss16Bytes = s0_addr_Up_low(4) =/= s0_addr_low(4)
175  val s0_misalignWith16Byte = !s0_rs_corss16Bytes && !s0_addr_aligned && !s0_use_flow_prf
176  s0_is128bit := Mux(s0_use_flow_ma, io.misalign_stin.bits.is128bit, is128Bit(s0_vecstin.alignedType) || s0_misalignWith16Byte)
177
178  s0_fullva := Mux(
179    s0_use_flow_rs,
180    s0_stin.src(0) + SignExt(s0_stin.uop.imm(11,0), XLEN),
181    Mux(
182      s0_use_flow_vec,
183      s0_vecstin.vaddr,
184      s0_vaddr
185    )
186  )
187
188  val s0_mask = Mux(
189    s0_use_flow_ma,
190    io.misalign_stin.bits.mask,
191    Mux(
192      s0_use_flow_rs,
193      genVWmask128(s0_saddr, s0_uop.fuOpType(2,0)),
194      Mux(
195        s0_use_flow_vec,
196        s0_vecstin.mask,
197        // -1.asSInt.asUInt
198        Fill(VLEN/8, 1.U(1.W))
199      )
200    )
201  )
202
203  io.tlb.req.valid                   := s0_valid
204  io.tlb.req.bits.vaddr              := s0_vaddr
205  io.tlb.req.bits.fullva             := s0_fullva
206  io.tlb.req.bits.checkfullva        := s0_use_flow_rs || s0_use_flow_vec
207  io.tlb.req.bits.cmd                := TlbCmd.write
208  io.tlb.req.bits.isPrefetch         := s0_use_flow_prf
209  io.tlb.req.bits.size               := s0_size
210  io.tlb.req.bits.kill               := false.B
211  io.tlb.req.bits.memidx.is_ld       := false.B
212  io.tlb.req.bits.memidx.is_st       := true.B
213  io.tlb.req.bits.memidx.idx         := s0_mem_idx
214  io.tlb.req.bits.debug.robIdx       := s0_rob_idx
215  io.tlb.req.bits.no_translate       := false.B
216  io.tlb.req.bits.debug.pc           := s0_pc
217  io.tlb.req.bits.debug.isFirstIssue := s0_isFirstIssue
218  io.tlb.req_kill                    := false.B
219  io.tlb.req.bits.hyperinst          := LSUOpType.isHsv(s0_uop.fuOpType)
220  io.tlb.req.bits.hlvx               := false.B
221  io.tlb.req.bits.pmp_addr           := DontCare
222
223  // Dcache access here: not **real** dcache write
224  // just read meta and tag in dcache, to find out the store will hit or miss
225
226  // NOTE: The store request does not wait for the dcache to be ready.
227  //       If the dcache is not ready at this time, the dcache is not queried.
228  //       But, store prefetch request will always wait for dcache to be ready to make progress.
229  io.dcache.req.valid              := s0_fire
230  io.dcache.req.bits.cmd           := MemoryOpConstants.M_PFW
231  io.dcache.req.bits.vaddr         := s0_vaddr
232  io.dcache.req.bits.instrtype     := s0_instr_type
233
234  s0_out              := DontCare
235  s0_out.vaddr        := s0_vaddr
236  s0_out.fullva       := s0_fullva
237  // Now data use its own io
238  s0_out.data         := s0_stin.src(1)
239  s0_out.uop          := s0_uop
240  s0_out.miss         := false.B
241  // For unaligned, we need to generate a base-aligned mask in storeunit and then do a shift split in StoreQueue.
242  s0_out.mask         := Mux(s0_rs_corss16Bytes && !s0_addr_aligned, genBasemask(s0_saddr,s0_alignType(1,0)), s0_mask)
243  s0_out.isFirstIssue := s0_isFirstIssue
244  s0_out.isHWPrefetch := s0_use_flow_prf
245  s0_out.wlineflag    := s0_wlineflag
246  s0_out.isvec        := s0_use_flow_vec
247  s0_out.is128bit     := s0_is128bit
248  s0_out.vecActive    := s0_vecActive
249  s0_out.usSecondInv  := s0_secondInv
250  s0_out.elemIdx      := s0_elemIdx
251  s0_out.alignedType  := s0_alignedType
252  s0_out.mbIndex      := s0_mBIndex
253  s0_out.misalignWith16Byte      := s0_misalignWith16Byte
254  s0_out.isMisalign      := s0_isMisalign
255  s0_out.vecBaseVaddr := s0_vecBaseVaddr
256  when(s0_valid && s0_isFirstIssue) {
257    s0_out.uop.debugInfo.tlbFirstReqTime := GTimer()
258  }
259  s0_out.isFrmMisAlignBuf := s0_use_flow_ma
260  s0_out.isFinalSplit := s0_isFinalSplit
261//  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
262
263  io.st_mask_out.valid       := s0_use_flow_rs || s0_use_flow_vec
264  io.st_mask_out.bits.mask   := s0_out.mask
265  io.st_mask_out.bits.sqIdx  := s0_out.uop.sqIdx
266
267  io.stin.ready := s1_ready && s0_use_flow_rs
268  io.vecstin.ready := s1_ready && s0_use_flow_vec
269  io.prefetch_req.ready := s1_ready && io.dcache.req.ready && !s0_iss_valid && !s0_vec_valid && !s0_ma_st_valid
270  io.misalign_stin.ready := s1_ready && s0_use_flow_ma
271
272  // Pipeline
273  // --------------------------------------------------------------------------------
274  // stage 1
275  // --------------------------------------------------------------------------------
276  // TLB resp (send paddr to dcache)
277  val s1_valid  = RegInit(false.B)
278  val s1_in     = RegEnable(s0_out, s0_fire)
279  val s1_out    = Wire(new LsPipelineBundle)
280  val s1_kill   = Wire(Bool())
281  val s1_can_go = s2_ready
282  val s1_fire   = s1_valid && !s1_kill && s1_can_go
283  val s1_vecActive    = RegEnable(s0_out.vecActive, true.B, s0_fire)
284  val s1_frm_mabuf    = s1_in.isFrmMisAlignBuf
285  val s1_is128bit     = s1_in.is128bit
286
287  // mmio cbo decoder
288  val s1_isCbo   = RegEnable(s0_isCbo, s0_fire)
289  val s1_vaNeedExt = io.tlb.resp.bits.excp(0).vaNeedExt
290  val s1_isHyper   = io.tlb.resp.bits.excp(0).isHyper
291  val s1_paddr     = io.tlb.resp.bits.paddr(0)
292  val s1_gpaddr    = io.tlb.resp.bits.gpaddr(0)
293  val s1_fullva    = io.tlb.resp.bits.fullva
294  val s1_isForVSnonLeafPTE   = io.tlb.resp.bits.isForVSnonLeafPTE
295  val s1_tlb_miss  = io.tlb.resp.bits.miss && io.tlb.resp.valid && s1_valid
296  val s1_pbmt      = Mux(!s1_tlb_miss, io.tlb.resp.bits.pbmt.head, 0.U(Pbmt.width.W))
297  val s1_exception = ExceptionNO.selectByFu(s1_out.uop.exceptionVec, StaCfg).asUInt.orR
298  val s1_isvec     = RegEnable(s0_out.isvec, false.B, s0_fire)
299  //We don't want `StoreUnit` to have an additional effect on the Store of vector from a `misalignBuffer,`
300  //But there are places where a marker bit is needed to enable additional processing of vector instructions.
301  //For example: `StoreQueue` is exceptionBuffer
302  val s1_frm_mab_vec = RegEnable(s0_use_flow_ma && io.misalign_stin.bits.isvec, false.B, s0_fire)
303  // val s1_isLastElem = RegEnable(s0_isLastElem, false.B, s0_fire)
304  s1_kill := s1_in.uop.robIdx.needFlush(io.redirect) || (s1_tlb_miss && !s1_isvec && !s1_frm_mabuf)
305
306  s1_ready := !s1_valid || s1_kill || s2_ready
307  io.tlb.resp.ready := true.B // TODO: why dtlbResp needs a ready?
308  when (s0_fire) { s1_valid := true.B }
309  .elsewhen (s1_fire) { s1_valid := false.B }
310  .elsewhen (s1_kill) { s1_valid := false.B }
311
312  // st-ld violation dectect request.
313  io.stld_nuke_query.valid       := s1_valid && !s1_tlb_miss && !s1_in.isHWPrefetch
314  io.stld_nuke_query.bits.robIdx := s1_in.uop.robIdx
315  io.stld_nuke_query.bits.paddr  := s1_paddr
316  io.stld_nuke_query.bits.mask   := s1_in.mask
317  io.stld_nuke_query.bits.matchLine := (s1_in.isvec || s1_in.misalignWith16Byte) && s1_in.is128bit
318
319  // issue
320  io.issue.valid := s1_valid && !s1_tlb_miss && !s1_in.isHWPrefetch && !s1_isvec && !s1_frm_mabuf
321  io.issue.bits  := RegEnable(s0_stin, s0_valid)
322
323  // trigger
324  val storeTrigger = Module(new MemTrigger(MemType.STORE))
325  storeTrigger.io.fromCsrTrigger.tdataVec             := io.fromCsrTrigger.tdataVec
326  storeTrigger.io.fromCsrTrigger.tEnableVec           := io.fromCsrTrigger.tEnableVec
327  storeTrigger.io.fromCsrTrigger.triggerCanRaiseBpExp := io.fromCsrTrigger.triggerCanRaiseBpExp
328  storeTrigger.io.fromCsrTrigger.debugMode            := io.fromCsrTrigger.debugMode
329  storeTrigger.io.fromLoadStore.vaddr                 := s1_in.vaddr
330  storeTrigger.io.fromLoadStore.isVectorUnitStride    := s1_in.isvec && s1_in.is128bit
331  storeTrigger.io.fromLoadStore.mask                  := s1_in.mask
332  storeTrigger.io.isCbo.get                           := s1_isCbo
333
334  val s1_trigger_action = storeTrigger.io.toLoadStore.triggerAction
335  val s1_trigger_debug_mode = TriggerAction.isDmode(s1_trigger_action)
336  val s1_trigger_breakpoint = TriggerAction.isExp(s1_trigger_action)
337
338  // for misalign in vsMergeBuffer
339  io.s0_s1_valid := s0_valid || s1_valid
340
341  // Send TLB feedback to store issue queue
342  // Store feedback is generated in store_s1, sent to RS in store_s2
343  val s1_feedback = Wire(Valid(new RSFeedback))
344  s1_feedback.valid                 := s1_valid & !s1_in.isHWPrefetch
345  s1_feedback.bits.hit              := !s1_tlb_miss
346  s1_feedback.bits.flushState       := io.tlb.resp.bits.ptwBack
347  s1_feedback.bits.robIdx           := s1_out.uop.robIdx
348  s1_feedback.bits.sourceType       := RSFeedbackType.tlbMiss
349  s1_feedback.bits.dataInvalidSqIdx := DontCare
350  s1_feedback.bits.sqIdx            := s1_out.uop.sqIdx
351  s1_feedback.bits.lqIdx            := s1_out.uop.lqIdx
352
353  XSDebug(s1_feedback.valid,
354    "S1 Store: tlbHit: %d robIdx: %d\n",
355    s1_feedback.bits.hit,
356    s1_feedback.bits.robIdx.value
357  )
358
359  // io.feedback_slow := s1_feedback
360
361  // get paddr from dtlb, check if rollback is needed
362  // writeback store inst to lsq
363  s1_out           := s1_in
364  s1_out.paddr     := s1_paddr
365  s1_out.gpaddr    := s1_gpaddr
366  s1_out.fullva    := s1_fullva
367  s1_out.vaNeedExt := s1_vaNeedExt
368  s1_out.isHyper   := s1_isHyper
369  s1_out.miss      := false.B
370  s1_out.nc        := Pbmt.isNC(s1_pbmt)
371  s1_out.mmio      := Pbmt.isIO(s1_pbmt)
372  s1_out.tlbMiss   := s1_tlb_miss
373  s1_out.atomic    := Pbmt.isIO(s1_pbmt)
374  s1_out.isForVSnonLeafPTE := s1_isForVSnonLeafPTE
375  when (RegNext(io.tlb.req.bits.checkfullva) &&
376    (s1_out.uop.exceptionVec(storePageFault) ||
377      s1_out.uop.exceptionVec(storeAccessFault) ||
378      s1_out.uop.exceptionVec(storeGuestPageFault))) {
379    s1_out.uop.exceptionVec(storeAddrMisaligned) := false.B
380  }
381  s1_out.uop.exceptionVec(storePageFault)      := io.tlb.resp.bits.excp(0).pf.st && s1_vecActive
382  s1_out.uop.exceptionVec(storeAccessFault)    := io.tlb.resp.bits.excp(0).af.st && s1_vecActive
383  s1_out.uop.exceptionVec(storeGuestPageFault) := io.tlb.resp.bits.excp(0).gpf.st && s1_vecActive
384
385  s1_out.uop.flushPipe                := false.B
386  s1_out.uop.trigger                  := s1_trigger_action
387  s1_out.uop.exceptionVec(breakPoint) := s1_trigger_breakpoint
388  s1_out.uop.exceptionVec(storeAddrMisaligned) := s1_out.mmio && s1_in.isMisalign
389  s1_out.vecVaddrOffset := Mux(
390    s1_trigger_debug_mode || s1_trigger_breakpoint,
391    storeTrigger.io.toLoadStore.triggerVaddr - s1_in.vecBaseVaddr,
392    s1_in.vaddr + genVFirstUnmask(s1_in.mask).asUInt - s1_in.vecBaseVaddr ,
393  )
394  s1_out.vecTriggerMask := Mux(s1_trigger_debug_mode || s1_trigger_breakpoint, storeTrigger.io.toLoadStore.triggerMask, 0.U)
395
396  // scalar store and scalar load nuke check, and also other purposes
397  //A 128-bit aligned unaligned memory access requires changing the unaligned flag bit in sq
398  io.lsq.valid     := s1_valid && !s1_in.isHWPrefetch
399  io.lsq.bits      := s1_out
400  io.lsq.bits.miss := s1_tlb_miss
401  io.lsq.bits.isvec := s1_out.isvec || s1_frm_mab_vec
402  io.lsq.bits.updateAddrValid := (!s1_in.isMisalign || s1_in.misalignWith16Byte) && (!s1_frm_mabuf || s1_in.isFinalSplit) || s1_exception
403  // kill dcache write intent request when tlb miss or exception
404  io.dcache.s1_kill  := (s1_tlb_miss || s1_exception || s1_out.mmio || s1_in.uop.robIdx.needFlush(io.redirect))
405  io.dcache.s1_paddr := s1_paddr
406
407  // write below io.out.bits assign sentence to prevent overwriting values
408  val s1_tlb_memidx = io.tlb.resp.bits.memidx
409  when(s1_tlb_memidx.is_st && io.tlb.resp.valid && !s1_tlb_miss && s1_tlb_memidx.idx === s1_out.uop.sqIdx.value) {
410    // printf("Store idx = %d\n", s1_tlb_memidx.idx)
411    s1_out.uop.debugInfo.tlbRespTime := GTimer()
412  }
413  val s1_mis_align = s1_valid && !s1_tlb_miss && !s1_in.isHWPrefetch && !s1_isCbo && !s1_out.nc && !s1_out.mmio &&
414                      GatedValidRegNext(io.csrCtrl.hd_misalign_st_enable) && s1_in.isMisalign && !s1_in.misalignWith16Byte &&
415                      !s1_trigger_breakpoint && !s1_trigger_debug_mode
416
417  // Pipeline
418  // --------------------------------------------------------------------------------
419  // stage 2
420  // --------------------------------------------------------------------------------
421  // mmio check
422  val s2_valid  = RegInit(false.B)
423  val s2_in     = RegEnable(s1_out, s1_fire)
424  val s2_out    = Wire(new LsPipelineBundle)
425  val s2_kill   = Wire(Bool())
426  val s2_can_go = s3_ready
427  val s2_fire   = s2_valid && !s2_kill && s2_can_go
428  val s2_vecActive    = RegEnable(s1_out.vecActive, true.B, s1_fire)
429  val s2_frm_mabuf    = s2_in.isFrmMisAlignBuf
430  val s2_frm_mab_vec  = RegEnable(s1_frm_mab_vec, true.B, s1_fire)
431  val s2_pbmt   = RegEnable(s1_pbmt, s1_fire)
432  val s2_trigger_debug_mode = RegEnable(s1_trigger_debug_mode, false.B, s1_fire)
433
434  s2_ready := !s2_valid || s2_kill || s3_ready
435  when (s1_fire) { s2_valid := true.B }
436  .elsewhen (s2_fire) { s2_valid := false.B }
437  .elsewhen (s2_kill) { s2_valid := false.B }
438
439  val s2_pmp = WireInit(io.pmp)
440
441  val s2_exception = RegNext(s1_feedback.bits.hit) &&
442                    (s2_trigger_debug_mode || ExceptionNO.selectByFu(s2_out.uop.exceptionVec, StaCfg).asUInt.orR) && s2_vecActive
443  val s2_un_misalign_exception =  RegNext(s1_feedback.bits.hit) &&
444                    (s2_trigger_debug_mode || ExceptionNO.selectByFuAndUnSelect(s2_out.uop.exceptionVec, StaCfg, Seq(storeAddrMisaligned)).asUInt.orR)
445
446  val s2_mmio = (s2_in.mmio || (Pbmt.isPMA(s2_pbmt) && s2_pmp.mmio)) && RegNext(s1_feedback.bits.hit)
447  val s2_memBackTypeMM = !s2_pmp.mmio
448  val s2_actually_uncache = (Pbmt.isPMA(s2_pbmt) && s2_pmp.mmio || s2_in.nc || s2_in.mmio) && RegNext(s1_feedback.bits.hit)
449  val s2_uncache = !s2_exception && !s2_in.tlbMiss && s2_actually_uncache
450  val s2_isCbo  = RegEnable(s1_isCbo, s1_fire) // all cbo instr
451  val s2_isCbo_noZero = LSUOpType.isCbo(s2_in.uop.fuOpType)
452
453  s2_kill := ((s2_mmio && !s2_exception) && !s2_in.isvec && !s2_frm_mabuf) || s2_in.uop.robIdx.needFlush(io.redirect)
454
455  s2_out        := s2_in
456  s2_out.af     := s2_out.uop.exceptionVec(storeAccessFault)
457  s2_out.mmio   := s2_mmio && !s2_exception
458  s2_out.atomic := s2_in.atomic || Pbmt.isPMA(s2_pbmt) && s2_pmp.atomic
459  s2_out.memBackTypeMM := s2_memBackTypeMM
460  s2_out.uop.exceptionVec(storeAccessFault) := (s2_in.uop.exceptionVec(storeAccessFault) ||
461                                                s2_pmp.st ||
462                                                ((s2_in.isvec || s2_frm_mabuf || s2_isCbo) && s2_actually_uncache && RegNext(s1_feedback.bits.hit))
463                                                ) && s2_vecActive
464  s2_out.uop.exceptionVec(storeAddrMisaligned) := s2_mmio && s2_in.isMisalign && !s2_un_misalign_exception
465  s2_out.uop.vpu.vstart     := s2_in.vecVaddrOffset >> s2_in.uop.vpu.veew
466
467  // kill dcache write intent request when mmio or exception
468  io.dcache.s2_kill := (s2_uncache || s2_exception || s2_in.uop.robIdx.needFlush(io.redirect))
469  io.dcache.s2_pc   := s2_out.uop.pc
470  // TODO: dcache resp
471  io.dcache.resp.ready := true.B
472
473  val s2_mis_align = s2_valid && RegEnable(s1_mis_align, s1_fire) && !s2_exception
474  // goto misalignBuffer
475  val toMisalignBufferValid = s2_valid && GatedValidRegNext(s1_mis_align && !s1_frm_mabuf)
476  io.misalign_buf.valid := toMisalignBufferValid
477  io.misalign_buf.bits  := s2_in
478  io.misalign_buf.bits.hasException := s2_exception
479  val misalignBufferNack = toMisalignBufferValid && !io.misalign_buf.ready
480
481  // feedback tlb miss to RS in store_s2
482  val feedback_slow_valid = WireInit(false.B)
483
484  feedback_slow_valid := s1_feedback.valid && !s1_out.uop.robIdx.needFlush(io.redirect) && !s1_out.isvec && !s1_frm_mabuf
485  io.feedback_slow.valid := GatedValidRegNext(feedback_slow_valid)
486  io.feedback_slow.bits  := RegEnable(s1_feedback.bits, feedback_slow_valid)
487  io.feedback_slow.bits.hit  := RegEnable(s1_feedback.bits.hit, feedback_slow_valid) && !misalignBufferNack
488
489  val s2_vecFeedback = RegNext(!s1_out.uop.robIdx.needFlush(io.redirect) && s1_feedback.bits.hit && s1_feedback.valid) &&
490                       !misalignBufferNack && s2_in.isvec && !s2_frm_mabuf
491
492  val s2_misalign_stout = WireInit(0.U.asTypeOf(io.misalign_stout))
493  s2_misalign_stout.valid := s2_valid && s2_can_go && s2_frm_mabuf
494  connectSamePort(s2_misalign_stout.bits, s2_out)
495  s2_misalign_stout.bits.need_rep := RegEnable(s1_tlb_miss, s1_fire)
496  io.misalign_stout := s2_misalign_stout
497
498  // mmio and exception
499  io.lsq_replenish := s2_out
500  io.lsq_replenish.af := s2_out.af && s2_valid && !s2_kill
501  io.lsq_replenish.mmio := (s2_mmio || s2_isCbo_noZero) && !s2_exception // reuse `mmiostall` logic in sq
502
503  // prefetch related
504  io.lsq_replenish.miss := io.dcache.resp.fire && io.dcache.resp.bits.miss // miss info
505  io.lsq_replenish.updateAddrValid := !s2_mis_align && (!s2_frm_mabuf || s2_out.isFinalSplit) || s2_exception
506  io.lsq_replenish.isvec := s2_out.isvec || s2_frm_mab_vec
507
508  io.lsq_replenish.hasException := (ExceptionNO.selectByFu(s2_out.uop.exceptionVec, StaCfg).asUInt.orR ||
509    TriggerAction.isDmode(s2_out.uop.trigger) || s2_out.af) && s2_valid && !s2_kill
510
511
512  // RegNext prefetch train for better timing
513  // ** Now, prefetch train is valid at store s3 **
514  val s2_prefetch_train_valid = WireInit(false.B)
515  s2_prefetch_train_valid := s2_valid && io.dcache.resp.fire && !s2_out.mmio && !s2_out.nc && !s2_in.tlbMiss && !s2_in.isHWPrefetch
516  if(EnableStorePrefetchSMS) {
517    io.s1_prefetch_spec := s1_fire
518    io.s2_prefetch_spec := s2_prefetch_train_valid
519    io.prefetch_train.valid := RegNext(s2_prefetch_train_valid)
520    io.prefetch_train.bits.fromLsPipelineBundle(s2_in, latch = true, enable = s2_prefetch_train_valid)
521  }else {
522    io.s1_prefetch_spec := false.B
523    io.s2_prefetch_spec := false.B
524    io.prefetch_train.valid := false.B
525    io.prefetch_train.bits.fromLsPipelineBundle(s2_in, latch = true, enable = false.B)
526  }
527  // override miss bit
528  io.prefetch_train.bits.miss := RegEnable(io.dcache.resp.bits.miss, s2_prefetch_train_valid)
529  // TODO: add prefetch and access bit
530  io.prefetch_train.bits.meta_prefetch := false.B
531  io.prefetch_train.bits.meta_access := false.B
532  io.prefetch_train.bits.isFinalSplit := false.B
533  io.prefetch_train.bits.misalignWith16Byte := false.B
534  io.prefetch_train.bits.isMisalign := false.B
535  io.prefetch_train.bits.misalignNeedWakeUp := false.B
536  io.prefetch_train.bits.updateAddrValid := false.B
537  io.prefetch_train.bits.hasException := false.B
538
539  // Pipeline
540  // --------------------------------------------------------------------------------
541  // stage 3
542  // --------------------------------------------------------------------------------
543  // store write back
544  val s3_valid  = RegInit(false.B)
545  val s3_in     = RegEnable(s2_out, s2_fire)
546  val s3_out    = Wire(new MemExuOutput(isVector = true))
547  val s3_kill   = s3_in.uop.robIdx.needFlush(io.redirect)
548  val s3_can_go = s3_ready
549  val s3_fire   = s3_valid && !s3_kill && s3_can_go
550  val s3_vecFeedback = RegEnable(s2_vecFeedback, s2_fire)
551  val s3_exception     = RegEnable(s2_exception, s2_fire)
552
553  // store misalign will not writeback to rob now
554  when (s2_fire) { s3_valid := (!s2_mmio && !s2_isCbo || s2_exception) && !s2_out.isHWPrefetch && !s2_mis_align && !s2_frm_mabuf }
555  .elsewhen (s3_fire) { s3_valid := false.B }
556  .elsewhen (s3_kill) { s3_valid := false.B }
557
558  // wb: writeback
559  val SelectGroupSize   = RollbackGroupSize
560  val lgSelectGroupSize = log2Ceil(SelectGroupSize)
561  val TotalSelectCycles = scala.math.ceil(log2Ceil(LoadQueueRAWSize).toFloat / lgSelectGroupSize).toInt + 1
562
563  s3_out                 := DontCare
564  s3_out.uop             := s3_in.uop
565  s3_out.data            := DontCare
566  s3_out.debug.isMMIO    := s3_in.mmio
567  s3_out.debug.isNC      := s3_in.nc
568  s3_out.debug.paddr     := s3_in.paddr
569  s3_out.debug.vaddr     := s3_in.vaddr
570  s3_out.debug.isPerfCnt := false.B
571
572  XSError(s3_valid && s3_in.isvec && s3_in.vecActive && !s3_in.mask.orR, "In vecActive, mask complement should not be 0")
573  // Pipeline
574  // --------------------------------------------------------------------------------
575  // stage x
576  // --------------------------------------------------------------------------------
577  // delay TotalSelectCycles - 2 cycle(s)
578  val TotalDelayCycles = TotalSelectCycles - 2
579  val sx_valid = Wire(Vec(TotalDelayCycles + 1, Bool()))
580  val sx_ready = Wire(Vec(TotalDelayCycles + 1, Bool()))
581  val sx_in    = Wire(Vec(TotalDelayCycles + 1, new VecMemExuOutput(isVector = true)))
582  val sx_in_vec = Wire(Vec(TotalDelayCycles +1, Bool()))
583
584  // backward ready signal
585  s3_ready := sx_ready.head
586  for (i <- 0 until TotalDelayCycles + 1) {
587    if (i == 0) {
588      sx_valid(i)          := s3_valid
589      sx_in(i).output      := s3_out
590      sx_in(i).vecFeedback := s3_vecFeedback
591      sx_in(i).nc          := s3_in.nc
592      sx_in(i).mmio        := s3_in.mmio
593      sx_in(i).usSecondInv := s3_in.usSecondInv
594      sx_in(i).elemIdx     := s3_in.elemIdx
595      sx_in(i).alignedType := s3_in.alignedType
596      sx_in(i).mbIndex     := s3_in.mbIndex
597      sx_in(i).mask        := s3_in.mask
598      sx_in(i).vaddr       := s3_in.fullva
599      sx_in(i).vaNeedExt   := s3_in.vaNeedExt
600      sx_in(i).gpaddr      := s3_in.gpaddr
601      sx_in(i).isForVSnonLeafPTE     := s3_in.isForVSnonLeafPTE
602      sx_in(i).vecTriggerMask := s3_in.vecTriggerMask
603      sx_in(i).hasException := s3_exception
604      sx_in_vec(i)         := s3_in.isvec
605      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))
606    } else {
607      val cur_kill   = sx_in(i).output.uop.robIdx.needFlush(io.redirect)
608      val cur_can_go = (if (i == TotalDelayCycles) io.stout.ready else sx_ready(i+1))
609      val cur_fire   = sx_valid(i) && !cur_kill && cur_can_go
610      val prev_fire  = sx_valid(i-1) && !sx_in(i-1).output.uop.robIdx.needFlush(io.redirect) && sx_ready(i)
611
612      sx_ready(i) := !sx_valid(i) || cur_kill || (if (i == TotalDelayCycles) io.stout.ready else sx_ready(i+1))
613      val sx_valid_can_go = prev_fire || cur_fire || cur_kill
614      sx_valid(i) := RegEnable(Mux(prev_fire, true.B, false.B), false.B, sx_valid_can_go)
615      sx_in(i) := RegEnable(sx_in(i-1), prev_fire)
616      sx_in_vec(i) := RegEnable(sx_in_vec(i-1), prev_fire)
617    }
618  }
619  val sx_last_valid = sx_valid.takeRight(1).head
620  val sx_last_ready = sx_ready.takeRight(1).head
621  val sx_last_in    = sx_in.takeRight(1).head
622  val sx_last_in_vec = sx_in_vec.takeRight(1).head
623  sx_last_ready := !sx_last_valid || sx_last_in.output.uop.robIdx.needFlush(io.redirect) || io.stout.ready
624
625  // write back: normal store, nc store
626  io.stout.valid := sx_last_valid && !sx_last_in_vec //isStore(sx_last_in.output.uop.fuType)
627  io.stout.bits := sx_last_in.output
628  io.stout.bits.uop.exceptionVec := ExceptionNO.selectByFu(sx_last_in.output.uop.exceptionVec, StaCfg)
629
630  io.vecstout.valid := sx_last_valid && sx_last_in_vec //isVStore(sx_last_in.output.uop.fuType)
631  // TODO: implement it!
632  io.vecstout.bits.mBIndex := sx_last_in.mbIndex
633  io.vecstout.bits.hit := sx_last_in.vecFeedback
634  io.vecstout.bits.isvec := true.B
635  io.vecstout.bits.sourceType := RSFeedbackType.tlbMiss
636  io.vecstout.bits.flushState := DontCare
637  io.vecstout.bits.trigger    := sx_last_in.output.uop.trigger
638  io.vecstout.bits.nc := sx_last_in.nc
639  io.vecstout.bits.mmio := sx_last_in.mmio
640  io.vecstout.bits.exceptionVec := ExceptionNO.selectByFu(sx_last_in.output.uop.exceptionVec, VstuCfg)
641  io.vecstout.bits.hasException := sx_last_in.hasException
642  io.vecstout.bits.usSecondInv := sx_last_in.usSecondInv
643  io.vecstout.bits.vecFeedback := sx_last_in.vecFeedback
644  io.vecstout.bits.elemIdx     := sx_last_in.elemIdx
645  io.vecstout.bits.alignedType := sx_last_in.alignedType
646  io.vecstout.bits.mask        := sx_last_in.mask
647  io.vecstout.bits.vaddr       := sx_last_in.vaddr
648  io.vecstout.bits.vaNeedExt   := sx_last_in.vaNeedExt
649  io.vecstout.bits.gpaddr      := sx_last_in.gpaddr
650  io.vecstout.bits.isForVSnonLeafPTE     := sx_last_in.isForVSnonLeafPTE
651  io.vecstout.bits.vstart      := sx_last_in.output.uop.vpu.vstart
652  io.vecstout.bits.vecTriggerMask := sx_last_in.vecTriggerMask
653  // io.vecstout.bits.reg_offset.map(_ := DontCare)
654  // io.vecstout.bits.elemIdx.map(_ := sx_last_in.elemIdx)
655  // io.vecstout.bits.elemIdxInsideVd.map(_ := DontCare)
656  // io.vecstout.bits.vecdata.map(_ := DontCare)
657  // io.vecstout.bits.mask.map(_ := DontCare)
658  // io.vecstout.bits.alignedType.map(_ := sx_last_in.alignedType)
659
660  io.debug_ls := DontCare
661  io.debug_ls.s1_robIdx := s1_in.uop.robIdx.value
662  io.debug_ls.s1_isTlbFirstMiss := io.tlb.resp.valid && io.tlb.resp.bits.miss && io.tlb.resp.bits.debug.isFirstIssue && !s1_in.isHWPrefetch
663
664  private def printPipeLine(pipeline: LsPipelineBundle, cond: Bool, name: String): Unit = {
665    XSDebug(cond,
666      p"$name" + p" pc ${Hexadecimal(pipeline.uop.pc)} " +
667        p"addr ${Hexadecimal(pipeline.vaddr)} -> ${Hexadecimal(pipeline.paddr)} " +
668        p"op ${Binary(pipeline.uop.fuOpType)} " +
669        p"data ${Hexadecimal(pipeline.data)} " +
670        p"mask ${Hexadecimal(pipeline.mask)}\n"
671    )
672  }
673
674  printPipeLine(s0_out, s0_valid, "S0")
675  printPipeLine(s1_out, s1_valid, "S1")
676
677  // perf cnt
678  XSPerfAccumulate("s0_in_valid",                s0_valid)
679  XSPerfAccumulate("s0_in_fire",                 s0_fire)
680  XSPerfAccumulate("s0_vecin_fire",              s0_fire && s0_use_flow_vec)
681  XSPerfAccumulate("s0_in_fire_first_issue",     s0_fire && s0_isFirstIssue)
682  XSPerfAccumulate("s0_addr_spec_success",       s0_fire && !s0_use_flow_vec && s0_saddr(VAddrBits-1, 12) === s0_stin.src(0)(VAddrBits-1, 12))
683  XSPerfAccumulate("s0_addr_spec_failed",        s0_fire && !s0_use_flow_vec && s0_saddr(VAddrBits-1, 12) =/= s0_stin.src(0)(VAddrBits-1, 12))
684  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)
685  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)
686
687  XSPerfAccumulate("s1_in_valid",                s1_valid)
688  XSPerfAccumulate("s1_in_fire",                 s1_fire)
689  XSPerfAccumulate("s1_in_fire_first_issue",     s1_fire && s1_in.isFirstIssue)
690  XSPerfAccumulate("s1_tlb_miss",                s1_fire && s1_tlb_miss)
691  XSPerfAccumulate("s1_tlb_miss_first_issue",    s1_fire && s1_tlb_miss && s1_in.isFirstIssue)
692  // end
693}
694