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.{DynInst, MemExuInput, MemExuOutput} 27import xiangshan.backend.fu.PMPRespBundle 28import xiangshan.backend.fu.FuConfig._ 29import xiangshan.backend.ctrlblock.{DebugLsInfoBundle, LsTopdownInfo} 30import xiangshan.backend.rob.RobPtr 31import xiangshan.backend.fu._ 32import xiangshan.backend.fu.util.SdtrigExt 33import xiangshan.cache._ 34import xiangshan.cache.wpu.ReplayCarry 35import xiangshan.cache.mmu.{TlbCmd, TlbHintReq, TlbReq, TlbRequestIO, TlbResp} 36import xiangshan.mem.mdp._ 37 38class HybridUnit(implicit p: Parameters) extends XSModule 39 with HasLoadHelper 40 with HasPerfEvents 41 with HasDCacheParameters 42 with HasCircularQueuePtrHelper 43 with HasVLSUParameters 44 with SdtrigExt 45{ 46 val io = IO(new Bundle() { 47 // control 48 val redirect = Flipped(ValidIO(new Redirect)) 49 val csrCtrl = Flipped(new CustomCSRCtrlIO) 50 51 // flow in 52 val lsin = Flipped(Decoupled(new MemExuInput)) 53 54 // flow out 55 val ldout = DecoupledIO(new MemExuOutput) 56 val stout = DecoupledIO(new MemExuOutput) 57 58 val ldu_io = new Bundle() { 59 // dcache 60 val dcache = new DCacheLoadIO 61 62 // data path 63 val sbuffer = new LoadForwardQueryIO 64 val vec_forward = new LoadForwardQueryIO 65 val lsq = new LoadToLsqIO 66 val tl_d_channel = Input(new DcacheToLduForwardIO) 67 val forward_mshr = Flipped(new LduToMissqueueForwardIO) 68 val tlb_hint = Flipped(new TlbHintReq) 69 val l2_hint = Input(Valid(new L2ToL1Hint)) 70 71 // fast wakeup 72 val fast_uop = ValidIO(new DynInst) // early wakeup signal generated in load_s1, send to RS in load_s2 73 74 // trigger 75 val trigger = Vec(TriggerNum, new LoadUnitTriggerIO) 76 77 // load to load fast path 78 val l2l_fwd_in = Input(new LoadToLoadIO) 79 val l2l_fwd_out = Output(new LoadToLoadIO) 80 81 val ld_fast_match = Input(Bool()) 82 val ld_fast_fuOpType = Input(UInt()) 83 val ld_fast_imm = Input(UInt(12.W)) 84 85 // hardware prefetch to l1 cache req 86 val prefetch_req = Flipped(ValidIO(new L1PrefetchReq)) 87 88 // iq cancel 89 val ldCancel = Output(new LoadCancelIO()) // use to cancel the uops waked by this load, and cancel load 90 91 // iq wakeup, use to wakeup consumer uop at load s2 92 val wakeup = ValidIO(new DynInst) 93 94 // load ecc error 95 val s3_dly_ld_err = Output(Bool()) // Note that io.s3_dly_ld_err and io.lsq.s3_dly_ld_err is different 96 97 // schedule error query 98 val stld_nuke_query = Flipped(Vec(StorePipelineWidth, Valid(new StoreNukeQueryIO))) 99 100 // queue-based replay 101 val replay = Flipped(Decoupled(new LsPipelineBundle)) 102 val lq_rep_full = Input(Bool()) 103 104 // misc 105 val s2_ptr_chasing = Output(Bool()) // provide right pc for hw prefetch 106 107 // Load fast replay path 108 val fast_rep_in = Flipped(Decoupled(new LqWriteBundle)) 109 val fast_rep_out = Decoupled(new LqWriteBundle) 110 111 // Load RAR rollback 112 val rollback = Valid(new Redirect) 113 114 // perf 115 val debug_ls = Output(new DebugLsInfoBundle) 116 val lsTopdownInfo = Output(new LsTopdownInfo) 117 } 118 119 val stu_io = new Bundle() { 120 val dcache = new DCacheStoreIO 121 val prefetch_req = Flipped(DecoupledIO(new StorePrefetchReq)) 122 val issue = Valid(new MemExuInput) 123 val lsq = ValidIO(new LsPipelineBundle) 124 val lsq_replenish = Output(new LsPipelineBundle()) 125 val stld_nuke_query = Valid(new StoreNukeQueryIO) 126 val st_mask_out = Valid(new StoreMaskBundle) 127 val debug_ls = Output(new DebugLsInfoBundle) 128 } 129 130 val vec_stu_io = new Bundle() { 131 val in = Flipped(DecoupledIO(new VecPipeBundle())) 132 val isFirstIssue = Input(Bool()) 133 val lsq = ValidIO(new LsPipelineBundle()) 134 val feedbackSlow = ValidIO(new VSFQFeedback) 135 } 136 137 // speculative for gated control 138 val s0_prefetch_spec = Output(Bool()) 139 val s1_prefetch_spec = Output(Bool()) 140 // prefetch 141 val prefetch_train = ValidIO(new LdPrefetchTrainBundle()) // provide prefetch info to sms 142 val prefetch_train_l1 = ValidIO(new LdPrefetchTrainBundle()) // provide prefetch info to stream & stride 143 val canAcceptLowConfPrefetch = Output(Bool()) 144 val canAcceptHighConfPrefetch = Output(Bool()) 145 val correctMissTrain = Input(Bool()) 146 147 // data path 148 val tlb = new TlbRequestIO(2) 149 val pmp = Flipped(new PMPRespBundle()) // arrive same to tlb now 150 151 // rs feedback 152 val feedback_fast = ValidIO(new RSFeedback) // stage 2 153 val feedback_slow = ValidIO(new RSFeedback) // stage 3 154 }) 155 156 val StorePrefetchL1Enabled = EnableStorePrefetchAtCommit || EnableStorePrefetchAtIssue || EnableStorePrefetchSPB 157 val s1_ready, s2_ready, s3_ready, sx_can_go = WireInit(false.B) 158 159 // Pipeline 160 // -------------------------------------------------------------------------------- 161 // stage 0 162 // -------------------------------------------------------------------------------- 163 // generate addr, use addr to query DCache and DTLB 164 val s0_valid = Wire(Bool()) 165 val s0_dcache_ready = Wire(Bool()) 166 val s0_kill = Wire(Bool()) 167 val s0_vaddr = Wire(UInt(VAddrBits.W)) 168 val s0_mask = Wire(UInt((VLEN/8).W)) 169 val s0_uop = Wire(new DynInst) 170 val s0_has_rob_entry = Wire(Bool()) 171 val s0_rsIdx = Wire(UInt(log2Up(MemIQSizeMax).W)) 172 val s0_mshrid = Wire(UInt()) 173 val s0_try_l2l = Wire(Bool()) 174 val s0_rep_carry = Wire(new ReplayCarry(nWays)) 175 val s0_isFirstIssue = Wire(Bool()) 176 val s0_fast_rep = Wire(Bool()) 177 val s0_ld_rep = Wire(Bool()) 178 val s0_l2l_fwd = Wire(Bool()) 179 val s0_sched_idx = Wire(UInt()) 180 val s0_can_go = s1_ready 181 val s0_fire = s0_valid && s0_dcache_ready && s0_can_go 182 val s0_out = Wire(new LqWriteBundle) 183 // vector 184 val s0_isvec = WireInit(false.B) 185 val s0_vecActive = WireInit(true.B) 186 // val s0_flowPtr = WireInit(0.U.asTypeOf(new VsFlowPtr)) 187 val s0_isLastElem = WireInit(false.B) 188 189 // load flow select/gen 190 // src0: super load replayed by LSQ (cache miss replay) (io.ldu_io.replay) 191 // src1: fast load replay (io.ldu_io.fast_rep_in) 192 // src2: load replayed by LSQ (io.ldu_io.replay) 193 // src3: hardware prefetch from prefetchor (high confidence) (io.prefetch) 194 // src4: int read / software prefetch first issue from RS (io.in) 195 // src5: vec read first issue from RS (TODO) 196 // src6: load try pointchaising when no issued or replayed load (io.fastpath) 197 // src7: hardware prefetch from prefetchor (high confidence) (io.prefetch) 198 // priority: high to low 199 val s0_ld_flow = FuType.isLoad(s0_uop.fuType) || FuType.isVLoad(s0_uop.fuType) 200 val s0_rep_stall = io.lsin.valid && isAfter(io.ldu_io.replay.bits.uop.robIdx, io.lsin.bits.uop.robIdx) 201 val s0_super_ld_rep_valid = io.ldu_io.replay.valid && io.ldu_io.replay.bits.forward_tlDchannel 202 val s0_ld_fast_rep_valid = io.ldu_io.fast_rep_in.valid 203 val s0_ld_rep_valid = io.ldu_io.replay.valid && !io.ldu_io.replay.bits.forward_tlDchannel && !s0_rep_stall 204 val s0_high_conf_prf_valid = io.ldu_io.prefetch_req.valid && io.ldu_io.prefetch_req.bits.confidence > 0.U 205 val s0_int_iss_valid = io.lsin.valid // int flow first issue or software prefetch 206 val s0_vec_iss_valid = io.vec_stu_io.in.valid 207 val s0_l2l_fwd_valid = io.ldu_io.l2l_fwd_in.valid && io.ldu_io.ld_fast_match 208 val s0_low_conf_prf_valid = io.ldu_io.prefetch_req.valid && io.ldu_io.prefetch_req.bits.confidence === 0.U 209 dontTouch(s0_super_ld_rep_valid) 210 dontTouch(s0_ld_fast_rep_valid) 211 dontTouch(s0_ld_rep_valid) 212 dontTouch(s0_high_conf_prf_valid) 213 dontTouch(s0_int_iss_valid) 214 dontTouch(s0_vec_iss_valid) 215 dontTouch(s0_l2l_fwd_valid) 216 dontTouch(s0_low_conf_prf_valid) 217 218 // load flow source ready 219 val s0_super_ld_rep_ready = WireInit(true.B) 220 val s0_ld_fast_rep_ready = !s0_super_ld_rep_valid 221 val s0_ld_rep_ready = !s0_super_ld_rep_valid && 222 !s0_ld_fast_rep_valid 223 val s0_high_conf_prf_ready = !s0_super_ld_rep_valid && 224 !s0_ld_fast_rep_valid && 225 !s0_ld_rep_valid 226 227 val s0_int_iss_ready = !s0_super_ld_rep_valid && 228 !s0_ld_fast_rep_valid && 229 !s0_ld_rep_valid && 230 !s0_high_conf_prf_valid 231 232 val s0_vec_iss_ready = !s0_super_ld_rep_valid && 233 !s0_ld_fast_rep_valid && 234 !s0_ld_rep_valid && 235 !s0_high_conf_prf_valid && 236 !s0_int_iss_valid 237 238 val s0_l2l_fwd_ready = !s0_super_ld_rep_valid && 239 !s0_ld_fast_rep_valid && 240 !s0_ld_rep_valid && 241 !s0_high_conf_prf_valid && 242 !s0_int_iss_valid && 243 !s0_vec_iss_valid 244 245 val s0_low_conf_prf_ready = !s0_super_ld_rep_valid && 246 !s0_ld_fast_rep_valid && 247 !s0_ld_rep_valid && 248 !s0_high_conf_prf_valid && 249 !s0_int_iss_valid && 250 !s0_vec_iss_valid && 251 !s0_l2l_fwd_valid 252 dontTouch(s0_super_ld_rep_ready) 253 dontTouch(s0_ld_fast_rep_ready) 254 dontTouch(s0_ld_rep_ready) 255 dontTouch(s0_high_conf_prf_ready) 256 dontTouch(s0_int_iss_ready) 257 dontTouch(s0_vec_iss_ready) 258 dontTouch(s0_l2l_fwd_ready) 259 dontTouch(s0_low_conf_prf_ready) 260 261 // load flow source select (OH) 262 val s0_super_ld_rep_select = s0_super_ld_rep_valid && s0_super_ld_rep_ready 263 val s0_ld_fast_rep_select = s0_ld_fast_rep_valid && s0_ld_fast_rep_ready 264 val s0_ld_rep_select = s0_ld_rep_valid && s0_ld_rep_ready 265 val s0_hw_prf_select = s0_high_conf_prf_ready && s0_high_conf_prf_valid || 266 s0_low_conf_prf_ready && s0_low_conf_prf_valid 267 val s0_int_iss_select = s0_int_iss_ready && s0_int_iss_valid 268 val s0_vec_iss_select = s0_vec_iss_ready && s0_vec_iss_valid 269 val s0_l2l_fwd_select = s0_l2l_fwd_ready && s0_l2l_fwd_valid 270 dontTouch(s0_super_ld_rep_select) 271 dontTouch(s0_ld_fast_rep_select) 272 dontTouch(s0_ld_rep_select) 273 dontTouch(s0_hw_prf_select) 274 dontTouch(s0_int_iss_select) 275 dontTouch(s0_vec_iss_select) 276 dontTouch(s0_l2l_fwd_select) 277 278 s0_valid := (s0_super_ld_rep_valid || 279 s0_ld_fast_rep_valid || 280 s0_ld_rep_valid || 281 s0_high_conf_prf_valid || 282 s0_int_iss_valid || 283 s0_vec_iss_valid || 284 s0_l2l_fwd_valid || 285 s0_low_conf_prf_valid) && !s0_kill 286 287 // which is S0's out is ready and dcache is ready 288 val s0_try_ptr_chasing = s0_l2l_fwd_select 289 val s0_do_try_ptr_chasing = s0_try_ptr_chasing && s0_can_go && io.ldu_io.dcache.req.ready 290 val s0_ptr_chasing_vaddr = io.ldu_io.l2l_fwd_in.data(5, 0) +& io.ldu_io.ld_fast_imm(5, 0) 291 val s0_ptr_chasing_canceled = WireInit(false.B) 292 s0_kill := s0_ptr_chasing_canceled || (s0_out.uop.robIdx.needFlush(io.redirect) && !s0_try_ptr_chasing) 293 294 // prefetch related ctrl signal 295 val s0_prf = Wire(Bool()) 296 val s0_prf_rd = Wire(Bool()) 297 val s0_prf_wr = Wire(Bool()) 298 val s0_hw_prf = s0_hw_prf_select 299 300 io.canAcceptLowConfPrefetch := s0_low_conf_prf_ready && io.ldu_io.dcache.req.ready 301 io.canAcceptHighConfPrefetch := s0_high_conf_prf_ready && io.ldu_io.dcache.req.ready 302 303 if (StorePrefetchL1Enabled) { 304 s0_dcache_ready := Mux(s0_ld_flow, io.ldu_io.dcache.req.ready, io.stu_io.dcache.req.ready) 305 } else { 306 s0_dcache_ready := Mux(s0_ld_flow, io.ldu_io.dcache.req.ready, true.B) 307 } 308 309 // query DTLB 310 io.tlb.req.valid := s0_valid && s0_dcache_ready 311 io.tlb.req.bits.cmd := Mux(s0_prf, 312 Mux(s0_prf_wr, TlbCmd.write, TlbCmd.read), 313 Mux(s0_ld_flow, TlbCmd.read, TlbCmd.write) 314 ) 315 io.tlb.req.bits.vaddr := Mux(s0_hw_prf_select, io.ldu_io.prefetch_req.bits.paddr, s0_vaddr) 316 io.tlb.req.bits.size := Mux(s0_isvec, io.vec_stu_io.in.bits.alignedType(1, 0), LSUOpType.size(s0_uop.fuOpType)) // may broken if use it in feature 317 io.tlb.req.bits.kill := s0_kill 318 io.tlb.req.bits.memidx.is_ld := s0_ld_flow 319 io.tlb.req.bits.memidx.is_st := !s0_ld_flow 320 io.tlb.req.bits.memidx.idx := s0_uop.lqIdx.value 321 io.tlb.req.bits.debug.robIdx := s0_uop.robIdx 322 io.tlb.req.bits.no_translate := s0_hw_prf_select // hw b.reqetch addr does not need to be translated 323 io.tlb.req.bits.debug.pc := s0_uop.pc 324 io.tlb.req.bits.debug.isFirstIssue := s0_isFirstIssue 325 326 // query DCache 327 // for load 328 io.ldu_io.dcache.req.valid := s0_valid && s0_dcache_ready && s0_ld_flow 329 io.ldu_io.dcache.req.bits.cmd := Mux(s0_prf_rd, MemoryOpConstants.M_PFR, 330 Mux(s0_prf_wr, MemoryOpConstants.M_PFW, MemoryOpConstants.M_XRD)) 331 io.ldu_io.dcache.req.bits.vaddr := s0_vaddr 332 io.ldu_io.dcache.req.bits.mask := s0_mask 333 io.ldu_io.dcache.req.bits.data := DontCare 334 io.ldu_io.dcache.req.bits.isFirstIssue := s0_isFirstIssue 335 io.ldu_io.dcache.req.bits.instrtype := Mux(s0_prf, DCACHE_PREFETCH_SOURCE.U, LOAD_SOURCE.U) 336 io.ldu_io.dcache.req.bits.debug_robIdx := s0_uop.robIdx.value 337 io.ldu_io.dcache.req.bits.replayCarry := s0_rep_carry 338 io.ldu_io.dcache.req.bits.id := DontCare // TODO: update cache meta 339 io.ldu_io.dcache.pf_source := Mux(s0_hw_prf_select, io.ldu_io.prefetch_req.bits.pf_source.value, L1_HW_PREFETCH_NULL) 340 io.ldu_io.dcache.is128Req := is128Bit(io.vec_stu_io.in.bits.alignedType) && io.vec_stu_io.in.valid && s0_vec_iss_select 341 342 // for store 343 io.stu_io.dcache.req.valid := s0_valid && s0_dcache_ready && !s0_ld_flow && !s0_prf 344 io.stu_io.dcache.req.bits.cmd := MemoryOpConstants.M_PFW 345 io.stu_io.dcache.req.bits.vaddr := s0_vaddr 346 io.stu_io.dcache.req.bits.instrtype := Mux(s0_prf, DCACHE_PREFETCH_SOURCE.U, STORE_SOURCE.U) 347 348 // load flow priority mux 349 def fromNullSource() = { 350 s0_vaddr := 0.U 351 s0_mask := 0.U 352 s0_uop := 0.U.asTypeOf(new DynInst) 353 s0_try_l2l := false.B 354 s0_has_rob_entry := false.B 355 s0_rsIdx := 0.U 356 s0_rep_carry := 0.U.asTypeOf(s0_rep_carry.cloneType) 357 s0_mshrid := 0.U 358 s0_isFirstIssue := false.B 359 s0_fast_rep := false.B 360 s0_ld_rep := false.B 361 s0_l2l_fwd := false.B 362 s0_prf := false.B 363 s0_prf_rd := false.B 364 s0_prf_wr := false.B 365 s0_sched_idx := 0.U 366 } 367 368 def fromFastReplaySource(src: LqWriteBundle) = { 369 s0_vaddr := src.vaddr 370 s0_mask := src.mask 371 s0_uop := src.uop 372 s0_try_l2l := false.B 373 s0_has_rob_entry := src.hasROBEntry 374 s0_rep_carry := src.rep_info.rep_carry 375 s0_mshrid := src.rep_info.mshr_id 376 s0_rsIdx := src.rsIdx 377 s0_isFirstIssue := false.B 378 s0_fast_rep := true.B 379 s0_ld_rep := src.isLoadReplay 380 s0_l2l_fwd := false.B 381 s0_prf := LSUOpType.isPrefetch(src.uop.fuOpType) 382 s0_prf_rd := src.uop.fuOpType === LSUOpType.prefetch_r 383 s0_prf_wr := src.uop.fuOpType === LSUOpType.prefetch_w 384 s0_sched_idx := src.schedIndex 385 } 386 387 def fromNormalReplaySource(src: LsPipelineBundle) = { 388 s0_vaddr := src.vaddr 389 s0_mask := genVWmask(src.vaddr, src.uop.fuOpType(1, 0)) 390 s0_uop := src.uop 391 s0_try_l2l := false.B 392 s0_has_rob_entry := true.B 393 s0_rsIdx := src.rsIdx 394 s0_rep_carry := src.replayCarry 395 s0_mshrid := src.mshrid 396 s0_isFirstIssue := false.B 397 s0_fast_rep := false.B 398 s0_ld_rep := true.B 399 s0_l2l_fwd := false.B 400 s0_prf := LSUOpType.isPrefetch(src.uop.fuOpType) 401 s0_prf_rd := src.uop.fuOpType === LSUOpType.prefetch_r 402 s0_prf_wr := src.uop.fuOpType === LSUOpType.prefetch_w 403 s0_sched_idx := src.schedIndex 404 } 405 406 def fromPrefetchSource(src: L1PrefetchReq) = { 407 s0_vaddr := src.getVaddr() 408 s0_mask := 0.U 409 s0_uop := DontCare 410 s0_try_l2l := false.B 411 s0_has_rob_entry := false.B 412 s0_rsIdx := 0.U 413 s0_rep_carry := 0.U.asTypeOf(s0_rep_carry.cloneType) 414 s0_mshrid := 0.U 415 s0_isFirstIssue := false.B 416 s0_fast_rep := false.B 417 s0_ld_rep := false.B 418 s0_l2l_fwd := false.B 419 s0_prf := true.B 420 s0_prf_rd := !src.is_store 421 s0_prf_wr := src.is_store 422 s0_sched_idx := 0.U 423 } 424 425 def fromIntIssueSource(src: MemExuInput) = { 426 s0_vaddr := src.src(0) + SignExt(src.uop.imm(11, 0), VAddrBits) 427 s0_mask := genVWmask(s0_vaddr, src.uop.fuOpType(1,0)) 428 s0_uop := src.uop 429 s0_try_l2l := false.B 430 s0_has_rob_entry := true.B 431 s0_rsIdx := src.iqIdx 432 s0_rep_carry := 0.U.asTypeOf(s0_rep_carry.cloneType) 433 s0_mshrid := 0.U 434 s0_isFirstIssue := true.B 435 s0_fast_rep := false.B 436 s0_ld_rep := false.B 437 s0_l2l_fwd := false.B 438 s0_prf := LSUOpType.isPrefetch(src.uop.fuOpType) 439 s0_prf_rd := src.uop.fuOpType === LSUOpType.prefetch_r 440 s0_prf_wr := src.uop.fuOpType === LSUOpType.prefetch_w 441 s0_sched_idx := 0.U 442 } 443 444 def fromVecIssueSource(src: VecPipeBundle) = { 445 // For now, vector port handles only vector store flows 446 s0_vaddr := src.vaddr 447 s0_mask := src.mask 448 s0_uop := src.uop 449 s0_try_l2l := false.B 450 s0_has_rob_entry := true.B 451 s0_rsIdx := 0.U 452 s0_rep_carry := 0.U.asTypeOf(s0_rep_carry.cloneType) 453 s0_mshrid := 0.U 454 // s0_isFirstIssue := src.isFirstIssue 455 s0_fast_rep := false.B 456 s0_ld_rep := false.B 457 s0_l2l_fwd := false.B 458 s0_prf := false.B 459 s0_prf_rd := false.B 460 s0_prf_wr := false.B 461 s0_sched_idx := 0.U 462 463 s0_isvec := true.B 464 s0_vecActive := io.vec_stu_io.in.bits.vecActive 465 // s0_flowPtr := io.vec_stu_io.in.bits.flowPtr 466 // s0_isLastElem := io.vec_stu_io.in.bits.isLastElem 467 } 468 469 def fromLoadToLoadSource(src: LoadToLoadIO) = { 470 s0_vaddr := Cat(src.data(XLEN-1, 6), s0_ptr_chasing_vaddr(5,0)) 471 s0_mask := genVWmask(s0_vaddr, io.ldu_io.ld_fast_fuOpType(1, 0)) 472 // When there's no valid instruction from RS and LSQ, we try the load-to-load forwarding. 473 // Assume the pointer chasing is always ld. 474 s0_uop.fuOpType := io.ldu_io.ld_fast_fuOpType 475 s0_try_l2l := true.B 476 // we dont care s0_isFirstIssue and s0_rsIdx and s0_sqIdx in S0 when trying pointchasing 477 // because these signals will be updated in S1 478 s0_has_rob_entry := false.B 479 s0_rsIdx := 0.U 480 s0_mshrid := 0.U 481 s0_rep_carry := 0.U.asTypeOf(s0_rep_carry.cloneType) 482 s0_isFirstIssue := true.B 483 s0_fast_rep := false.B 484 s0_ld_rep := false.B 485 s0_l2l_fwd := true.B 486 s0_prf := false.B 487 s0_prf_rd := false.B 488 s0_prf_wr := false.B 489 s0_sched_idx := 0.U 490 } 491 492 // set default 493 s0_uop := DontCare 494 when (s0_super_ld_rep_select) { fromNormalReplaySource(io.ldu_io.replay.bits) } 495 .elsewhen (s0_ld_fast_rep_select) { fromFastReplaySource(io.ldu_io.fast_rep_in.bits) } 496 .elsewhen (s0_ld_rep_select) { fromNormalReplaySource(io.ldu_io.replay.bits) } 497 .elsewhen (s0_hw_prf_select) { fromPrefetchSource(io.ldu_io.prefetch_req.bits) } 498 .elsewhen (s0_int_iss_select) { fromIntIssueSource(io.lsin.bits) } 499 .elsewhen (s0_vec_iss_select) { fromVecIssueSource(io.vec_stu_io.in.bits) } 500 .otherwise { 501 if (EnableLoadToLoadForward) { 502 fromLoadToLoadSource(io.ldu_io.l2l_fwd_in) 503 } else { 504 fromNullSource() 505 } 506 } 507 508 // address align check 509 val s0_addr_aligned = LookupTree(Mux(s0_isvec, io.vec_stu_io.in.bits.alignedType(1,0), s0_uop.fuOpType(1, 0)), List( 510 "b00".U -> true.B, //b 511 "b01".U -> (s0_vaddr(0) === 0.U), //h 512 "b10".U -> (s0_vaddr(1, 0) === 0.U), //w 513 "b11".U -> (s0_vaddr(2, 0) === 0.U) //d 514 ))// may broken if use it in feature 515 516 // accept load flow if dcache ready (tlb is always ready) 517 // TODO: prefetch need writeback to loadQueueFlag 518 s0_out := DontCare 519 s0_out.rsIdx := s0_rsIdx 520 s0_out.vaddr := s0_vaddr 521 s0_out.mask := s0_mask 522 s0_out.uop := s0_uop 523 s0_out.isFirstIssue := s0_isFirstIssue 524 s0_out.hasROBEntry := s0_has_rob_entry 525 s0_out.isPrefetch := s0_prf 526 s0_out.isHWPrefetch := s0_hw_prf 527 s0_out.isFastReplay := s0_fast_rep 528 s0_out.isLoadReplay := s0_ld_rep 529 s0_out.isFastPath := s0_l2l_fwd 530 s0_out.mshrid := s0_mshrid 531 s0_out.isvec := s0_isvec 532 s0_out.isLastElem := s0_isLastElem 533 s0_out.vecActive := s0_vecActive 534 // s0_out.sflowPtr := s0_flowPtr 535 s0_out.uop.exceptionVec(loadAddrMisaligned) := !s0_addr_aligned && s0_ld_flow 536 s0_out.uop.exceptionVec(storeAddrMisaligned) := !s0_addr_aligned && !s0_ld_flow 537 s0_out.forward_tlDchannel := s0_super_ld_rep_select 538 when(io.tlb.req.valid && s0_isFirstIssue) { 539 s0_out.uop.debugInfo.tlbFirstReqTime := GTimer() 540 }.otherwise{ 541 s0_out.uop.debugInfo.tlbFirstReqTime := s0_uop.debugInfo.tlbFirstReqTime 542 } 543 s0_out.schedIndex := s0_sched_idx 544 545 // load fast replay 546 io.ldu_io.fast_rep_in.ready := (s0_can_go && io.ldu_io.dcache.req.ready && s0_ld_fast_rep_ready) 547 548 // load flow source ready 549 // cache missed load has highest priority 550 // always accept cache missed load flow from load replay queue 551 io.ldu_io.replay.ready := (s0_can_go && io.ldu_io.dcache.req.ready && (s0_ld_rep_ready && !s0_rep_stall || s0_super_ld_rep_select)) 552 553 // accept load flow from rs when: 554 // 1) there is no lsq-replayed load 555 // 2) there is no fast replayed load 556 // 3) there is no high confidence prefetch request 557 io.lsin.ready := (s0_can_go && 558 Mux(FuType.isLoad(io.lsin.bits.uop.fuType), io.ldu_io.dcache.req.ready, 559 (if (StorePrefetchL1Enabled) io.stu_io.dcache.req.ready else true.B)) && s0_int_iss_ready) 560 io.vec_stu_io.in.ready := s0_can_go && io.ldu_io.dcache.req.ready && s0_vec_iss_ready 561 562 563 // for hw prefetch load flow feedback, to be added later 564 // io.prefetch_in.ready := s0_hw_prf_select 565 566 // dcache replacement extra info 567 // TODO: should prefetch load update replacement? 568 io.ldu_io.dcache.replacementUpdated := Mux(s0_ld_rep_select || s0_super_ld_rep_select, io.ldu_io.replay.bits.replacementUpdated, false.B) 569 570 io.stu_io.prefetch_req.ready := s1_ready && io.stu_io.dcache.req.ready && !io.lsin.valid 571 572 // load debug 573 XSDebug(io.ldu_io.dcache.req.fire && s0_ld_flow, 574 p"[DCACHE LOAD REQ] pc ${Hexadecimal(s0_uop.pc)}, vaddr ${Hexadecimal(s0_vaddr)}\n" 575 ) 576 XSDebug(s0_valid && s0_ld_flow, 577 p"S0: pc ${Hexadecimal(s0_out.uop.pc)}, lqIdx ${Hexadecimal(s0_out.uop.lqIdx.asUInt)}, " + 578 p"vaddr ${Hexadecimal(s0_out.vaddr)}, mask ${Hexadecimal(s0_out.mask)}\n") 579 580 // store debug 581 XSDebug(io.stu_io.dcache.req.fire && !s0_ld_flow, 582 p"[DCACHE STORE REQ] pc ${Hexadecimal(s0_uop.pc)}, vaddr ${Hexadecimal(s0_vaddr)}\n" 583 ) 584 XSDebug(s0_valid && !s0_ld_flow, 585 p"S0: pc ${Hexadecimal(s0_out.uop.pc)}, sqIdx ${Hexadecimal(s0_out.uop.sqIdx.asUInt)}, " + 586 p"vaddr ${Hexadecimal(s0_out.vaddr)}, mask ${Hexadecimal(s0_out.mask)}\n") 587 588 589 // Pipeline 590 // -------------------------------------------------------------------------------- 591 // stage 1 592 // -------------------------------------------------------------------------------- 593 // TLB resp (send paddr to dcache) 594 val s1_valid = RegInit(false.B) 595 val s1_in = Wire(new LqWriteBundle) 596 val s1_out = Wire(new LqWriteBundle) 597 val s1_kill = Wire(Bool()) 598 val s1_can_go = s2_ready 599 val s1_fire = s1_valid && !s1_kill && s1_can_go 600 val s1_ld_flow = RegNext(s0_ld_flow) 601 val s1_isvec = RegEnable(s0_out.isvec, false.B, s0_fire) 602 val s1_isLastElem = RegEnable(s0_out.isLastElem, false.B, s0_fire) 603 604 s1_ready := !s1_valid || s1_kill || s2_ready 605 when (s0_fire) { s1_valid := true.B } 606 .elsewhen (s1_fire) { s1_valid := false.B } 607 .elsewhen (s1_kill) { s1_valid := false.B } 608 s1_in := RegEnable(s0_out, s0_fire) 609 610 val s1_fast_rep_dly_err = RegNext(io.ldu_io.fast_rep_in.bits.delayedLoadError) 611 val s1_fast_rep_kill = s1_fast_rep_dly_err && s1_in.isFastReplay 612 val s1_l2l_fwd_dly_err = RegNext(io.ldu_io.l2l_fwd_in.dly_ld_err) 613 val s1_l2l_fwd_kill = s1_l2l_fwd_dly_err && s1_in.isFastPath 614 val s1_late_kill = s1_fast_rep_kill || s1_l2l_fwd_kill 615 val s1_vaddr_hi = Wire(UInt()) 616 val s1_vaddr_lo = Wire(UInt()) 617 val s1_vaddr = Wire(UInt()) 618 val s1_paddr_dup_lsu = Wire(UInt()) 619 val s1_paddr_dup_dcache = Wire(UInt()) 620 val s1_ld_exception = ExceptionNO.selectByFu(s1_out.uop.exceptionVec, LduCfg).asUInt.orR // af & pf exception were modified below. 621 val s1_st_exception = ExceptionNO.selectByFu(s1_out.uop.exceptionVec, StaCfg).asUInt.orR // af & pf exception were modified below. 622 val s1_exception = (s1_ld_flow && s1_ld_exception) || (!s1_ld_flow && s1_st_exception) 623 val s1_tlb_miss = io.tlb.resp.bits.miss 624 val s1_prf = s1_in.isPrefetch 625 val s1_hw_prf = s1_in.isHWPrefetch 626 val s1_sw_prf = s1_prf && !s1_hw_prf 627 val s1_tlb_memidx = io.tlb.resp.bits.memidx 628 629 // mmio cbo decoder 630 val s1_mmio_cbo = (s1_in.uop.fuOpType === LSUOpType.cbo_clean || 631 s1_in.uop.fuOpType === LSUOpType.cbo_flush || 632 s1_in.uop.fuOpType === LSUOpType.cbo_inval) && !s1_ld_flow && !s1_prf 633 val s1_mmio = s1_mmio_cbo 634 635 s1_vaddr_hi := s1_in.vaddr(VAddrBits - 1, 6) 636 s1_vaddr_lo := s1_in.vaddr(5, 0) 637 s1_vaddr := Cat(s1_vaddr_hi, s1_vaddr_lo) 638 s1_paddr_dup_lsu := io.tlb.resp.bits.paddr(0) 639 s1_paddr_dup_dcache := io.tlb.resp.bits.paddr(1) 640 641 when (s1_tlb_memidx.is_ld && io.tlb.resp.valid && !s1_tlb_miss && 642 s1_tlb_memidx.idx === s1_in.uop.lqIdx.value && s1_ld_flow) { 643 // printf("Load idx = %d\n", s1_tlb_memidx.idx) 644 s1_out.uop.debugInfo.tlbRespTime := GTimer() 645 } .elsewhen(s1_tlb_memidx.is_st && io.tlb.resp.valid && !s1_tlb_miss && 646 s1_tlb_memidx.idx === s1_out.uop.sqIdx.value && !s1_ld_flow) { 647 // printf("Store idx = %d\n", s1_tlb_memidx.idx) 648 s1_out.uop.debugInfo.tlbRespTime := GTimer() 649 } 650 651 io.tlb.req_kill := s1_kill 652 io.tlb.resp.ready := true.B 653 654 io.ldu_io.dcache.s1_paddr_dup_lsu <> s1_paddr_dup_lsu 655 io.ldu_io.dcache.s1_paddr_dup_dcache <> s1_paddr_dup_dcache 656 io.ldu_io.dcache.s1_kill := s1_kill || s1_tlb_miss || s1_exception 657 658 // store to load forwarding 659 io.ldu_io.sbuffer.valid := s1_valid && !(s1_exception || s1_tlb_miss || s1_kill || s1_fast_rep_kill || s1_prf || !s1_ld_flow) 660 io.ldu_io.sbuffer.vaddr := s1_vaddr 661 io.ldu_io.sbuffer.paddr := s1_paddr_dup_lsu 662 io.ldu_io.sbuffer.uop := s1_in.uop 663 io.ldu_io.sbuffer.sqIdx := s1_in.uop.sqIdx 664 io.ldu_io.sbuffer.mask := s1_in.mask 665 io.ldu_io.sbuffer.pc := s1_in.uop.pc // FIXME: remove it 666 667 io.ldu_io.vec_forward.valid := s1_valid && !(s1_exception || s1_tlb_miss || s1_kill || s1_fast_rep_kill || s1_prf || !s1_ld_flow) 668 io.ldu_io.vec_forward.vaddr := s1_vaddr 669 io.ldu_io.vec_forward.paddr := s1_paddr_dup_lsu 670 io.ldu_io.vec_forward.uop := s1_in.uop 671 io.ldu_io.vec_forward.sqIdx := s1_in.uop.sqIdx 672 io.ldu_io.vec_forward.mask := s1_in.mask 673 io.ldu_io.vec_forward.pc := s1_in.uop.pc // FIXME: remove it 674 675 io.ldu_io.lsq.forward.valid := s1_valid && !(s1_exception || s1_tlb_miss || s1_kill || s1_fast_rep_kill || s1_prf || !s1_ld_flow) 676 io.ldu_io.lsq.forward.vaddr := s1_vaddr 677 io.ldu_io.lsq.forward.paddr := s1_paddr_dup_lsu 678 io.ldu_io.lsq.forward.uop := s1_in.uop 679 io.ldu_io.lsq.forward.sqIdx := s1_in.uop.sqIdx 680 io.ldu_io.lsq.forward.sqIdxMask := 0.U 681 io.ldu_io.lsq.forward.mask := s1_in.mask 682 io.ldu_io.lsq.forward.pc := s1_in.uop.pc // FIXME: remove it 683 684 // st-ld violation query 685 val s1_nuke = VecInit((0 until StorePipelineWidth).map(w => { 686 io.ldu_io.stld_nuke_query(w).valid && // query valid 687 isAfter(s1_in.uop.robIdx, io.ldu_io.stld_nuke_query(w).bits.robIdx) && // older store 688 // TODO: Fix me when vector instruction 689 (s1_paddr_dup_lsu(PAddrBits-1, 3) === io.ldu_io.stld_nuke_query(w).bits.paddr(PAddrBits-1, 3)) && // paddr match 690 (s1_in.mask & io.ldu_io.stld_nuke_query(w).bits.mask).orR // data mask contain 691 })).asUInt.orR && !s1_tlb_miss && s1_ld_flow 692 693 s1_out := s1_in 694 s1_out.vaddr := s1_vaddr 695 s1_out.paddr := s1_paddr_dup_lsu 696 s1_out.tlbMiss := s1_tlb_miss 697 s1_out.ptwBack := io.tlb.resp.bits.ptwBack 698 s1_out.rsIdx := s1_in.rsIdx 699 s1_out.rep_info.debug := s1_in.uop.debugInfo 700 s1_out.rep_info.nuke := s1_nuke && !s1_sw_prf 701 s1_out.lateKill := s1_late_kill 702 703 when (s1_ld_flow) { 704 when (!s1_late_kill) { 705 // current ori test will cause the case of ldest == 0, below will be modifeid in the future. 706 // af & pf exception were modified 707 s1_out.uop.exceptionVec(loadPageFault) := io.tlb.resp.bits.excp(0).pf.ld 708 s1_out.uop.exceptionVec(loadGuestPageFault) := io.tlb.resp.bits.excp(0).gpf.ld 709 s1_out.uop.exceptionVec(loadAccessFault) := io.tlb.resp.bits.excp(0).af.ld 710 } .otherwise { 711 s1_out.uop.exceptionVec(loadAddrMisaligned) := false.B 712 s1_out.uop.exceptionVec(loadAccessFault) := s1_late_kill 713 } 714 } .otherwise { 715 s1_out.uop.exceptionVec(storePageFault) := io.tlb.resp.bits.excp(0).pf.st 716 s1_out.uop.exceptionVec(storeGuestPageFault) := io.tlb.resp.bits.excp(0).gpf.st 717 s1_out.uop.exceptionVec(storeAccessFault) := io.tlb.resp.bits.excp(0).af.st 718 } 719 720 // pointer chasing 721 val s1_try_ptr_chasing = RegNext(s0_do_try_ptr_chasing, false.B) 722 val s1_ptr_chasing_vaddr = RegEnable(s0_ptr_chasing_vaddr, s0_do_try_ptr_chasing) 723 val s1_fu_op_type_not_ld = WireInit(false.B) 724 val s1_not_fast_match = WireInit(false.B) 725 val s1_addr_mismatch = WireInit(false.B) 726 val s1_addr_misaligned = WireInit(false.B) 727 val s1_ptr_chasing_canceled = WireInit(false.B) 728 val s1_cancel_ptr_chasing = WireInit(false.B) 729 730 s1_kill := s1_late_kill || 731 s1_cancel_ptr_chasing || 732 s1_in.uop.robIdx.needFlush(io.redirect) || 733 RegEnable(s0_kill, false.B, io.lsin.valid || io.ldu_io.replay.valid || io.ldu_io.l2l_fwd_in.valid || io.ldu_io.fast_rep_in.valid || io.vec_stu_io.in.valid) 734 735 if (EnableLoadToLoadForward) { 736 // Sometimes, we need to cancel the load-load forwarding. 737 // These can be put at S0 if timing is bad at S1. 738 // Case 0: CACHE_SET(base + offset) != CACHE_SET(base) (lowest 6-bit addition has an overflow) 739 s1_addr_mismatch := s1_ptr_chasing_vaddr(6) || RegEnable(io.ldu_io.ld_fast_imm(11, 6).orR, s0_do_try_ptr_chasing) 740 // Case 1: the address is misaligned, kill s1 741 s1_addr_misaligned := LookupTree(s1_in.uop.fuOpType(1, 0), List( 742 "b00".U -> false.B, //b 743 "b01".U -> (s1_vaddr(0) =/= 0.U), //h 744 "b10".U -> (s1_vaddr(1, 0) =/= 0.U), //w 745 "b11".U -> (s1_vaddr(2, 0) =/= 0.U) //d 746 )) 747 // Case 2: this load-load uop is cancelled 748 s1_ptr_chasing_canceled := !io.lsin.valid || FuType.isStore(io.lsin.bits.uop.fuType) 749 750 when (s1_try_ptr_chasing) { 751 s1_cancel_ptr_chasing := s1_addr_mismatch || s1_addr_misaligned || s1_ptr_chasing_canceled 752 753 s1_in.uop := io.lsin.bits.uop 754 s1_in.rsIdx := io.lsin.bits.iqIdx 755 s1_in.isFirstIssue := io.lsin.bits.isFirstIssue 756 s1_vaddr_lo := s1_ptr_chasing_vaddr(5, 0) 757 s1_paddr_dup_lsu := Cat(io.tlb.resp.bits.paddr(0)(PAddrBits - 1, 6), s1_vaddr_lo) 758 s1_paddr_dup_dcache := Cat(io.tlb.resp.bits.paddr(0)(PAddrBits - 1, 6), s1_vaddr_lo) 759 760 // recored tlb time when get the data to ensure the correctness of the latency calculation (although it should not record in here, because it does not use tlb) 761 s1_in.uop.debugInfo.tlbFirstReqTime := GTimer() 762 s1_in.uop.debugInfo.tlbRespTime := GTimer() 763 } 764 when (!s1_cancel_ptr_chasing) { 765 s0_ptr_chasing_canceled := s1_try_ptr_chasing && !io.ldu_io.replay.fire && !io.ldu_io.fast_rep_in.fire && !(s0_high_conf_prf_valid && io.canAcceptHighConfPrefetch) 766 when (s1_try_ptr_chasing) { 767 io.lsin.ready := true.B 768 } 769 } 770 } 771 772 // pre-calcuate sqIdx mask in s0, then send it to lsq in s1 for forwarding 773 val s1_sqIdx_mask = RegNext(UIntToMask(s0_out.uop.sqIdx.value, StoreQueueSize)) 774 // to enable load-load, sqIdxMask must be calculated based on lsin.uop 775 // If the timing here is not OK, load-load forwarding has to be disabled. 776 // Or we calculate sqIdxMask at RS?? 777 io.ldu_io.lsq.forward.sqIdxMask := s1_sqIdx_mask 778 if (EnableLoadToLoadForward) { 779 when (s1_try_ptr_chasing) { 780 io.ldu_io.lsq.forward.sqIdxMask := UIntToMask(io.lsin.bits.uop.sqIdx.value, StoreQueueSize) 781 } 782 } 783 784 io.ldu_io.forward_mshr.valid := s1_valid && s1_out.forward_tlDchannel && s1_ld_flow 785 io.ldu_io.forward_mshr.mshrid := s1_out.mshrid 786 io.ldu_io.forward_mshr.paddr := s1_out.paddr 787 788 io.ldu_io.wakeup.valid := s0_fire && s0_ld_flow && (s0_super_ld_rep_select || s0_ld_fast_rep_select || s0_ld_rep_select || s0_int_iss_select) 789 io.ldu_io.wakeup.bits := s0_uop 790 791 io.stu_io.dcache.s1_kill := s1_tlb_miss || s1_exception || s1_mmio || s1_in.uop.robIdx.needFlush(io.redirect) 792 io.stu_io.dcache.s1_paddr := s1_paddr_dup_dcache 793 794 795 // load debug 796 XSDebug(s1_valid && s1_ld_flow, 797 p"S1: pc ${Hexadecimal(s1_out.uop.pc)}, lId ${Hexadecimal(s1_out.uop.lqIdx.asUInt)}, tlb_miss ${io.tlb.resp.bits.miss}, " + 798 p"paddr ${Hexadecimal(s1_out.paddr)}, mmio ${s1_out.mmio}\n") 799 800 // store debug 801 XSDebug(s1_valid && !s1_ld_flow, 802 p"S1: pc ${Hexadecimal(s1_out.uop.pc)}, lId ${Hexadecimal(s1_out.uop.sqIdx.asUInt)}, tlb_miss ${io.tlb.resp.bits.miss}, " + 803 p"paddr ${Hexadecimal(s1_out.paddr)}, mmio ${s1_out.mmio}\n") 804 805 // store out 806 io.stu_io.lsq.valid := s1_valid && !s1_ld_flow && !s1_prf && !s1_isvec 807 io.stu_io.lsq.bits := s1_out 808 io.stu_io.lsq.bits.miss := s1_tlb_miss 809 810 io.vec_stu_io.lsq.valid := s1_valid && !s1_ld_flow && !s1_prf && s1_isvec 811 io.vec_stu_io.lsq.bits := s1_out 812 io.vec_stu_io.lsq.bits.miss := s1_tlb_miss 813 io.vec_stu_io.lsq.bits.isLastElem := s1_isLastElem 814 815 io.stu_io.st_mask_out.valid := s1_valid && !s1_ld_flow && !s1_prf 816 io.stu_io.st_mask_out.bits.mask := s1_out.mask 817 io.stu_io.st_mask_out.bits.sqIdx := s1_out.uop.sqIdx 818 819 io.stu_io.issue.valid := s1_valid && !s1_tlb_miss && !s1_ld_flow && !s1_prf && !s1_isvec 820 io.stu_io.issue.bits := RegEnable(io.lsin.bits, io.lsin.fire) 821 822 // st-ld violation dectect request 823 io.stu_io.stld_nuke_query.valid := s1_valid && !s1_tlb_miss && !s1_ld_flow && !s1_prf 824 io.stu_io.stld_nuke_query.bits.robIdx := s1_in.uop.robIdx 825 io.stu_io.stld_nuke_query.bits.paddr := s1_paddr_dup_lsu 826 io.stu_io.stld_nuke_query.bits.mask := s1_in.mask 827 828 // Pipeline 829 // -------------------------------------------------------------------------------- 830 // stage 2 831 // -------------------------------------------------------------------------------- 832 // s2: DCache resp 833 val s2_valid = RegInit(false.B) 834 val s2_in = Wire(new LqWriteBundle) 835 val s2_out = Wire(new LqWriteBundle) 836 val s2_kill = Wire(Bool()) 837 val s2_can_go = s3_ready 838 val s2_fire = s2_valid && !s2_kill && s2_can_go 839 val s2_isvec = RegEnable(s1_isvec, false.B, s1_fire) 840 val s2_vecActive = RegEnable(s1_out.vecActive, true.B, s1_fire) 841 val s2_paddr = RegEnable(s1_paddr_dup_lsu, s1_fire) 842 843 s2_kill := s2_in.uop.robIdx.needFlush(io.redirect) 844 s2_ready := !s2_valid || s2_kill || s3_ready 845 when (s1_fire) { s2_valid := true.B } 846 .elsewhen (s2_fire) { s2_valid := false.B } 847 .elsewhen (s2_kill) { s2_valid := false.B } 848 s2_in := RegEnable(s1_out, s1_fire) 849 850 val s2_pmp = WireInit(io.pmp) 851 852 val s2_prf = s2_in.isPrefetch 853 val s2_hw_prf = s2_in.isHWPrefetch 854 val s2_ld_flow = RegEnable(s1_ld_flow, s1_fire) 855 856 // exception that may cause load addr to be invalid / illegal 857 // if such exception happen, that inst and its exception info 858 // will be force writebacked to rob 859 val s2_exception_vec = WireInit(s2_in.uop.exceptionVec) 860 when (s2_ld_flow) { 861 when (!s2_in.lateKill) { 862 s2_exception_vec(loadAccessFault) := (s2_in.uop.exceptionVec(loadAccessFault) || s2_pmp.ld) && s2_vecActive 863 // soft prefetch will not trigger any exception (but ecc error interrupt may be triggered) 864 when (s2_prf || s2_in.tlbMiss) { 865 s2_exception_vec := 0.U.asTypeOf(s2_exception_vec.cloneType) 866 } 867 } 868 } .otherwise { 869 s2_exception_vec(storeAccessFault) := s2_in.uop.exceptionVec(storeAccessFault) || s2_pmp.st 870 when (s2_prf || s2_in.tlbMiss) { 871 s2_exception_vec := 0.U.asTypeOf(s2_exception_vec.cloneType) 872 } 873 } 874 val s2_ld_exception = ExceptionNO.selectByFu(s2_exception_vec, LduCfg).asUInt.orR && s2_ld_flow 875 val s2_st_exception = ExceptionNO.selectByFu(s2_exception_vec, StaCfg).asUInt.orR && !s2_ld_flow 876 val s2_exception = s2_ld_exception || s2_st_exception 877 878 val (s2_fwd_frm_d_chan, s2_fwd_data_frm_d_chan) = io.ldu_io.tl_d_channel.forward(s1_valid && s1_out.forward_tlDchannel, s1_out.mshrid, s1_out.paddr) 879 val (s2_fwd_data_valid, s2_fwd_frm_mshr, s2_fwd_data_frm_mshr) = io.ldu_io.forward_mshr.forward() 880 val s2_fwd_frm_d_chan_or_mshr = s2_fwd_data_valid && (s2_fwd_frm_d_chan || s2_fwd_frm_mshr) 881 882 // writeback access fault caused by ecc error / bus error 883 // * ecc data error is slow to generate, so we will not use it until load stage 3 884 // * in load stage 3, an extra signal io.load_error will be used to 885 val s2_actually_mmio = s2_pmp.mmio 886 val s2_ld_mmio = !s2_prf && 887 s2_actually_mmio && 888 !s2_exception && 889 !s2_in.tlbMiss && 890 s2_ld_flow 891 val s2_st_mmio = !s2_prf && 892 (RegNext(s1_mmio) || s2_pmp.mmio) && 893 !s2_exception && 894 !s2_in.tlbMiss && 895 !s2_ld_flow 896 val s2_st_atomic = !s2_prf && 897 (RegNext(s1_mmio) || s2_pmp.atomic) && 898 !s2_exception && 899 !s2_in.tlbMiss && 900 !s2_ld_flow 901 val s2_full_fwd = Wire(Bool()) 902 val s2_mem_amb = s2_in.uop.storeSetHit && 903 io.ldu_io.lsq.forward.addrInvalid 904 905 val s2_tlb_miss = s2_in.tlbMiss 906 val s2_fwd_fail = io.ldu_io.lsq.forward.dataInvalid || io.ldu_io.vec_forward.dataInvalid 907 val s2_dcache_miss = io.ldu_io.dcache.resp.bits.miss && 908 !s2_fwd_frm_d_chan_or_mshr && 909 !s2_full_fwd 910 911 val s2_mq_nack = io.ldu_io.dcache.s2_mq_nack && 912 !s2_fwd_frm_d_chan_or_mshr && 913 !s2_full_fwd 914 915 val s2_bank_conflict = io.ldu_io.dcache.s2_bank_conflict && 916 !s2_fwd_frm_d_chan_or_mshr && 917 !s2_full_fwd 918 919 val s2_wpu_pred_fail = io.ldu_io.dcache.s2_wpu_pred_fail && 920 !s2_fwd_frm_d_chan_or_mshr && 921 !s2_full_fwd 922 923 val s2_rar_nack = io.ldu_io.lsq.ldld_nuke_query.req.valid && 924 !io.ldu_io.lsq.ldld_nuke_query.req.ready 925 926 val s2_raw_nack = io.ldu_io.lsq.stld_nuke_query.req.valid && 927 !io.ldu_io.lsq.stld_nuke_query.req.ready 928 929 // st-ld violation query 930 // NeedFastRecovery Valid when 931 // 1. Fast recovery query request Valid. 932 // 2. Load instruction is younger than requestors(store instructions). 933 // 3. Physical address match. 934 // 4. Data contains. 935 val s2_nuke = VecInit((0 until StorePipelineWidth).map(w => { 936 io.ldu_io.stld_nuke_query(w).valid && // query valid 937 isAfter(s2_in.uop.robIdx, io.ldu_io.stld_nuke_query(w).bits.robIdx) && // older store 938 // TODO: Fix me when vector instruction 939 (s2_in.paddr(PAddrBits-1, 3) === io.ldu_io.stld_nuke_query(w).bits.paddr(PAddrBits-1, 3)) && // paddr match 940 (s2_in.mask & io.ldu_io.stld_nuke_query(w).bits.mask).orR // data mask contain 941 })).asUInt.orR && s2_ld_flow || s2_in.rep_info.nuke 942 943 val s2_cache_handled = io.ldu_io.dcache.resp.bits.handled 944 val s2_cache_tag_error = RegNext(io.csrCtrl.cache_error_enable) && 945 io.ldu_io.dcache.resp.bits.tag_error 946 947 val s2_troublem = !s2_exception && 948 !s2_ld_mmio && 949 !s2_prf && 950 !s2_in.lateKill && 951 s2_ld_flow 952 953 io.ldu_io.dcache.resp.ready := true.B 954 io.stu_io.dcache.resp.ready := true.B 955 val s2_dcache_should_resp = !(s2_in.tlbMiss || s2_exception || s2_ld_mmio || s2_prf || s2_in.lateKill) && s2_ld_flow 956 assert(!(s2_valid && (s2_dcache_should_resp && !io.ldu_io.dcache.resp.valid)), "DCache response got lost") 957 958 // fast replay require 959 val s2_dcache_fast_rep = (s2_mq_nack || !s2_dcache_miss && (s2_bank_conflict || s2_wpu_pred_fail)) 960 val s2_nuke_fast_rep = !s2_mq_nack && 961 !s2_dcache_miss && 962 !s2_bank_conflict && 963 !s2_wpu_pred_fail && 964 !s2_rar_nack && 965 !s2_raw_nack && 966 s2_nuke 967 968 val s2_fast_rep = !s2_mem_amb && 969 !s2_tlb_miss && 970 !s2_fwd_fail && 971 (s2_dcache_fast_rep || s2_nuke_fast_rep) && 972 s2_troublem 973 974 // need allocate new entry 975 val s2_can_query = !s2_mem_amb && 976 !s2_tlb_miss && 977 !s2_fwd_fail && 978 !s2_dcache_fast_rep && 979 s2_troublem 980 981 val s2_data_fwded = s2_dcache_miss && (s2_full_fwd || s2_cache_tag_error) 982 983 // ld-ld violation require 984 io.ldu_io.lsq.ldld_nuke_query.req.valid := s2_valid && s2_can_query 985 io.ldu_io.lsq.ldld_nuke_query.req.bits.uop := s2_in.uop 986 io.ldu_io.lsq.ldld_nuke_query.req.bits.mask := s2_in.mask 987 io.ldu_io.lsq.ldld_nuke_query.req.bits.paddr := s2_in.paddr 988 io.ldu_io.lsq.ldld_nuke_query.req.bits.data_valid := Mux(s2_full_fwd || s2_fwd_data_valid, true.B, !s2_dcache_miss) 989 990 // st-ld violation require 991 io.ldu_io.lsq.stld_nuke_query.req.valid := s2_valid && s2_can_query 992 io.ldu_io.lsq.stld_nuke_query.req.bits.uop := s2_in.uop 993 io.ldu_io.lsq.stld_nuke_query.req.bits.mask := s2_in.mask 994 io.ldu_io.lsq.stld_nuke_query.req.bits.paddr := s2_in.paddr 995 io.ldu_io.lsq.stld_nuke_query.req.bits.data_valid := Mux(s2_full_fwd || s2_fwd_data_valid, true.B, !s2_dcache_miss) 996 997 // merge forward result 998 // lsq has higher priority than sbuffer 999 val s2_fwd_mask = Wire(Vec((VLEN/8), Bool())) 1000 val s2_fwd_data = Wire(Vec((VLEN/8), UInt(8.W))) 1001 s2_full_fwd := ((~s2_fwd_mask.asUInt).asUInt & s2_in.mask) === 0.U && !io.ldu_io.lsq.forward.dataInvalid && !io.ldu_io.vec_forward.dataInvalid 1002 // generate XLEN/8 Muxs 1003 for (i <- 0 until VLEN / 8) { 1004 s2_fwd_mask(i) := io.ldu_io.lsq.forward.forwardMask(i) || io.ldu_io.sbuffer.forwardMask(i) || io.ldu_io.vec_forward.forwardMask(i) 1005 s2_fwd_data(i) := Mux( 1006 io.ldu_io.lsq.forward.forwardMask(i), 1007 io.ldu_io.lsq.forward.forwardData(i), 1008 Mux( 1009 io.ldu_io.vec_forward.forwardMask(i), 1010 io.ldu_io.vec_forward.forwardData(i), 1011 io.ldu_io.sbuffer.forwardData(i) 1012 ) 1013 ) 1014 } 1015 1016 XSDebug(s2_fire && s2_ld_flow, "[FWD LOAD RESP] pc %x fwd %x(%b) + %x(%b)\n", 1017 s2_in.uop.pc, 1018 io.ldu_io.lsq.forward.forwardData.asUInt, io.ldu_io.lsq.forward.forwardMask.asUInt, 1019 s2_in.forwardData.asUInt, s2_in.forwardMask.asUInt 1020 ) 1021 1022 // 1023 s2_out := s2_in 1024 s2_out.data := 0.U // data will be generated in load s3 1025 s2_out.uop.fpWen := s2_in.uop.fpWen && !s2_exception && s2_ld_flow 1026 s2_out.mmio := s2_ld_mmio || s2_st_mmio 1027 s2_out.atomic := s2_st_atomic 1028 s2_out.uop.flushPipe := false.B 1029 s2_out.uop.exceptionVec := s2_exception_vec 1030 s2_out.forwardMask := s2_fwd_mask 1031 s2_out.forwardData := s2_fwd_data 1032 s2_out.handledByMSHR := s2_cache_handled 1033 s2_out.miss := s2_dcache_miss && s2_troublem 1034 s2_out.feedbacked := io.feedback_fast.valid && !io.feedback_fast.bits.hit 1035 1036 // Generate replay signal caused by: 1037 // * st-ld violation check 1038 // * tlb miss 1039 // * dcache replay 1040 // * forward data invalid 1041 // * dcache miss 1042 s2_out.rep_info.mem_amb := s2_mem_amb && s2_troublem 1043 s2_out.rep_info.tlb_miss := s2_tlb_miss && s2_troublem 1044 s2_out.rep_info.fwd_fail := s2_fwd_fail && s2_troublem 1045 s2_out.rep_info.dcache_rep := s2_mq_nack && s2_troublem 1046 s2_out.rep_info.dcache_miss := s2_dcache_miss && s2_troublem 1047 s2_out.rep_info.bank_conflict := s2_bank_conflict && s2_troublem 1048 s2_out.rep_info.wpu_fail := s2_wpu_pred_fail && s2_troublem 1049 s2_out.rep_info.rar_nack := s2_rar_nack && s2_troublem 1050 s2_out.rep_info.raw_nack := s2_raw_nack && s2_troublem 1051 s2_out.rep_info.nuke := s2_nuke && s2_troublem 1052 s2_out.rep_info.full_fwd := s2_data_fwded 1053 s2_out.rep_info.data_inv_sq_idx := Mux(io.ldu_io.vec_forward.dataInvalid, s2_out.uop.sqIdx, io.ldu_io.lsq.forward.dataInvalidSqIdx) 1054 s2_out.rep_info.addr_inv_sq_idx := Mux(io.ldu_io.vec_forward.addrInvalid, s2_out.uop.sqIdx, io.ldu_io.lsq.forward.addrInvalidSqIdx) 1055 s2_out.rep_info.rep_carry := io.ldu_io.dcache.resp.bits.replayCarry 1056 s2_out.rep_info.mshr_id := io.ldu_io.dcache.resp.bits.mshr_id 1057 s2_out.rep_info.last_beat := s2_in.paddr(log2Up(refillBytes)) 1058 s2_out.rep_info.debug := s2_in.uop.debugInfo 1059 s2_out.rep_info.tlb_id := io.ldu_io.tlb_hint.id 1060 s2_out.rep_info.tlb_full := io.ldu_io.tlb_hint.full 1061 1062 // if forward fail, replay this inst from fetch 1063 val debug_fwd_fail_rep = s2_fwd_fail && !s2_troublem && !s2_in.tlbMiss 1064 // if ld-ld violation is detected, replay from this inst from fetch 1065 val debug_ldld_nuke_rep = false.B // s2_ldld_violation && !s2_ld_mmio && !s2_is_prefetch && !s2_in.tlbMiss 1066 // io.out.bits.uop.replayInst := false.B 1067 1068 // to be removed 1069 val s2_ld_need_fb = !s2_in.isLoadReplay && // already feedbacked 1070 io.ldu_io.lq_rep_full && // LoadQueueReplay is full 1071 s2_out.rep_info.need_rep && // need replay 1072 !s2_exception && // no exception is triggered 1073 !s2_hw_prf && // not hardware prefetch 1074 !s2_isvec 1075 val s2_st_need_fb = !s2_ld_flow && !s2_hw_prf && !s2_isvec 1076 io.feedback_fast.valid := s2_valid && (s2_ld_need_fb || s2_st_need_fb) 1077 io.feedback_fast.bits.hit := Mux(s2_ld_flow, false.B, !s2_tlb_miss) 1078 io.feedback_fast.bits.flushState := s2_in.ptwBack 1079 io.feedback_fast.bits.robIdx := s2_in.uop.robIdx 1080 io.feedback_fast.bits.sourceType := Mux(s2_ld_flow, RSFeedbackType.lrqFull, RSFeedbackType.tlbMiss) 1081 io.feedback_fast.bits.dataInvalidSqIdx := DontCare 1082 1083 val s2_vec_feedback = Wire(Valid(new VSFQFeedback)) 1084 s2_vec_feedback.valid := s2_valid && !s2_ld_flow && !s2_hw_prf && s2_isvec 1085 // s2_vec_feedback.bits.flowPtr := s2_out.sflowPtr 1086 s2_vec_feedback.bits.hit := !s2_tlb_miss 1087 s2_vec_feedback.bits.sourceType := RSFeedbackType.tlbMiss 1088 s2_vec_feedback.bits.paddr := s2_paddr 1089 s2_vec_feedback.bits.mmio := s2_st_mmio 1090 s2_vec_feedback.bits.atomic := s2_st_mmio 1091 s2_vec_feedback.bits.exceptionVec := s2_exception_vec 1092 1093 io.stu_io.lsq_replenish := s2_out 1094 io.stu_io.lsq_replenish.miss := io.ldu_io.dcache.resp.fire && io.ldu_io.dcache.resp.bits.miss 1095 1096 io.ldu_io.ldCancel.ld1Cancel := false.B 1097 1098 // fast wakeup 1099 io.ldu_io.fast_uop.valid := RegNext( 1100 !io.ldu_io.dcache.s1_disable_fast_wakeup && 1101 s1_valid && 1102 !s1_kill && 1103 !io.tlb.resp.bits.miss && 1104 !io.ldu_io.lsq.forward.dataInvalidFast 1105 ) && (s2_valid && !s2_out.rep_info.need_rep && !s2_ld_mmio && s2_ld_flow) && !s2_isvec 1106 io.ldu_io.fast_uop.bits := RegNext(s1_out.uop) 1107 1108 // 1109 io.ldu_io.s2_ptr_chasing := RegEnable(s1_try_ptr_chasing && !s1_cancel_ptr_chasing, false.B, s1_fire) 1110 1111 // prefetch train 1112 io.s0_prefetch_spec := s0_fire 1113 io.s1_prefetch_spec := s1_fire 1114 io.prefetch_train.valid := s2_valid && !s2_actually_mmio && !s2_in.tlbMiss 1115 io.prefetch_train.bits.fromLsPipelineBundle(s2_in) 1116 io.prefetch_train.bits.miss := Mux(s2_ld_flow, io.ldu_io.dcache.resp.bits.miss, io.stu_io.dcache.resp.bits.miss) // TODO: use trace with bank conflict? 1117 io.prefetch_train.bits.meta_prefetch := Mux(s2_ld_flow, io.ldu_io.dcache.resp.bits.meta_prefetch, false.B) 1118 io.prefetch_train.bits.meta_access := Mux(s2_ld_flow, io.ldu_io.dcache.resp.bits.meta_access, false.B) 1119 1120 io.prefetch_train_l1.valid := s2_valid && !s2_actually_mmio && s2_ld_flow 1121 io.prefetch_train_l1.bits.fromLsPipelineBundle(s2_in) 1122 io.prefetch_train_l1.bits.miss := io.ldu_io.dcache.resp.bits.miss 1123 io.prefetch_train_l1.bits.meta_prefetch := io.ldu_io.dcache.resp.bits.meta_prefetch 1124 io.prefetch_train_l1.bits.meta_access := io.ldu_io.dcache.resp.bits.meta_access 1125 if (env.FPGAPlatform){ 1126 io.ldu_io.dcache.s0_pc := DontCare 1127 io.ldu_io.dcache.s1_pc := DontCare 1128 io.ldu_io.dcache.s2_pc := DontCare 1129 }else{ 1130 io.ldu_io.dcache.s0_pc := s0_out.uop.pc 1131 io.ldu_io.dcache.s1_pc := s1_out.uop.pc 1132 io.ldu_io.dcache.s2_pc := s2_out.uop.pc 1133 } 1134 io.ldu_io.dcache.s2_kill := s2_pmp.ld || s2_actually_mmio || s2_kill 1135 io.stu_io.dcache.s2_kill := s2_pmp.st || s2_actually_mmio || s2_kill 1136 io.stu_io.dcache.s2_pc := s2_out.uop.pc 1137 1138 val s1_ld_left_fire = s1_valid && !s1_kill && s2_ready && s1_ld_flow 1139 val s2_ld_valid_dup = RegInit(0.U(6.W)) 1140 s2_ld_valid_dup := 0x0.U(6.W) 1141 when (s1_ld_left_fire && !s1_out.isHWPrefetch && s1_ld_flow) { s2_ld_valid_dup := 0x3f.U(6.W) } 1142 when (s1_kill || s1_out.isHWPrefetch || !s1_ld_flow) { s2_ld_valid_dup := 0x0.U(6.W) } 1143 assert(RegNext((s2_valid === s2_ld_valid_dup(0)) || RegNext(s1_out.isHWPrefetch) || RegNext(!s1_ld_flow))) 1144 1145 // Pipeline 1146 // -------------------------------------------------------------------------------- 1147 // stage 3 1148 // -------------------------------------------------------------------------------- 1149 // writeback and update load queue 1150 val s3_valid = RegNext(s2_valid && !s2_out.isHWPrefetch && !s2_out.uop.robIdx.needFlush(io.redirect)) 1151 val s3_in = RegEnable(s2_out, s2_fire) 1152 val s3_out = Wire(Valid(new MemExuOutput)) 1153 val s3_dcache_rep = RegEnable(s2_dcache_fast_rep && s2_troublem, false.B, s2_fire) 1154 val s3_ld_valid_dup = RegEnable(s2_ld_valid_dup, s2_fire) 1155 val s3_fast_rep = Wire(Bool()) 1156 val s3_ld_flow = RegNext(s2_ld_flow) 1157 val s3_troublem = RegNext(s2_troublem) 1158 val s3_kill = s3_in.uop.robIdx.needFlush(io.redirect) 1159 val s3_isvec = RegNext(s2_isvec) 1160 s3_ready := !s3_valid || s3_kill || sx_can_go 1161 1162 // forwrad last beat 1163 val (s3_fwd_frm_d_chan, s3_fwd_data_frm_d_chan) = io.ldu_io.tl_d_channel.forward(s2_valid && s2_out.forward_tlDchannel, s2_out.mshrid, s2_out.paddr) 1164 val s3_fwd_data_valid = RegEnable(s2_fwd_data_valid, false.B, s2_valid) 1165 val s3_fwd_frm_d_chan_valid = (s3_fwd_frm_d_chan && s3_fwd_data_valid) && s3_ld_flow 1166 1167 1168 // s3 load fast replay 1169 io.ldu_io.fast_rep_out.valid := s3_valid && 1170 s3_fast_rep && 1171 !s3_in.uop.robIdx.needFlush(io.redirect) && 1172 s3_ld_flow && 1173 !s3_isvec 1174 io.ldu_io.fast_rep_out.bits := s3_in 1175 1176 io.ldu_io.lsq.ldin.valid := s3_valid && 1177 (!s3_fast_rep || !io.ldu_io.fast_rep_out.ready) && 1178 !s3_in.feedbacked && 1179 !s3_in.lateKill && 1180 s3_ld_flow 1181 io.ldu_io.lsq.ldin.bits := s3_in 1182 io.ldu_io.lsq.ldin.bits.miss := s3_in.miss && !s3_fwd_frm_d_chan_valid 1183 1184 /* <------- DANGEROUS: Don't change sequence here ! -------> */ 1185 io.ldu_io.lsq.ldin.bits.data_wen_dup := s3_ld_valid_dup.asBools 1186 io.ldu_io.lsq.ldin.bits.replacementUpdated := io.ldu_io.dcache.resp.bits.replacementUpdated 1187 io.ldu_io.lsq.ldin.bits.missDbUpdated := RegNext(s2_fire && s2_in.hasROBEntry && !s2_in.tlbMiss && !s2_in.missDbUpdated) 1188 1189 val s3_dly_ld_err = 1190 if (EnableAccurateLoadError) { 1191 (s3_in.lateKill || io.ldu_io.dcache.resp.bits.error_delayed) && RegNext(io.csrCtrl.cache_error_enable) 1192 } else { 1193 WireInit(false.B) 1194 } 1195 io.ldu_io.s3_dly_ld_err := false.B // s3_dly_ld_err && s3_valid 1196 io.ldu_io.fast_rep_out.bits.delayedLoadError := s3_dly_ld_err 1197 io.ldu_io.lsq.ldin.bits.dcacheRequireReplay := s3_dcache_rep 1198 1199 val s3_vp_match_fail = RegNext(io.ldu_io.lsq.forward.matchInvalid || io.ldu_io.sbuffer.matchInvalid) && s3_troublem 1200 val s3_ldld_rep_inst = 1201 io.ldu_io.lsq.ldld_nuke_query.resp.valid && 1202 io.ldu_io.lsq.ldld_nuke_query.resp.bits.rep_frm_fetch && 1203 RegNext(io.csrCtrl.ldld_vio_check_enable) 1204 1205 val s3_rep_info = WireInit(s3_in.rep_info) 1206 s3_rep_info.dcache_miss := s3_in.rep_info.dcache_miss && !s3_fwd_frm_d_chan_valid && s3_troublem 1207 val s3_rep_frm_fetch = s3_vp_match_fail 1208 val s3_flushPipe = s3_ldld_rep_inst 1209 val s3_sel_rep_cause = PriorityEncoderOH(s3_rep_info.cause.asUInt) 1210 val s3_force_rep = s3_sel_rep_cause(LoadReplayCauses.C_TM) && 1211 !s3_in.uop.exceptionVec(loadAddrMisaligned) && 1212 s3_troublem 1213 1214 val s3_ld_exception = ExceptionNO.selectByFu(s3_in.uop.exceptionVec, LduCfg).asUInt.orR && s3_ld_flow 1215 val s3_st_exception = ExceptionNO.selectByFu(s3_in.uop.exceptionVec, StaCfg).asUInt.orR && !s3_ld_flow 1216 val s3_exception = s3_ld_exception || s3_st_exception 1217 when ((s3_ld_exception || s3_dly_ld_err || s3_rep_frm_fetch) && !s3_force_rep) { 1218 io.ldu_io.lsq.ldin.bits.rep_info.cause := 0.U.asTypeOf(s3_rep_info.cause.cloneType) 1219 } .otherwise { 1220 io.ldu_io.lsq.ldin.bits.rep_info.cause := VecInit(s3_sel_rep_cause.asBools) 1221 } 1222 1223 // Int flow, if hit, will be writebacked at s3 1224 s3_out.valid := s3_valid && 1225 (!s3_ld_flow && !s3_in.feedbacked || !io.ldu_io.lsq.ldin.bits.rep_info.need_rep) && !s3_in.mmio 1226 s3_out.bits.uop := s3_in.uop 1227 s3_out.bits.uop.exceptionVec(loadAccessFault) := (s3_dly_ld_err || s3_in.uop.exceptionVec(loadAccessFault)) && s3_ld_flow 1228 s3_out.bits.uop.replayInst := s3_rep_frm_fetch 1229 s3_out.bits.data := s3_in.data 1230 s3_out.bits.debug.isMMIO := s3_in.mmio 1231 s3_out.bits.debug.isPerfCnt := false.B 1232 s3_out.bits.debug.paddr := s3_in.paddr 1233 s3_out.bits.debug.vaddr := s3_in.vaddr 1234 1235 when (s3_force_rep) { 1236 s3_out.bits.uop.exceptionVec := 0.U.asTypeOf(s3_in.uop.exceptionVec.cloneType) 1237 } 1238 1239 io.ldu_io.rollback.valid := s3_valid && (s3_rep_frm_fetch || s3_flushPipe) && !s3_exception && s3_ld_flow 1240 io.ldu_io.rollback.bits := DontCare 1241 io.ldu_io.rollback.bits.isRVC := s3_out.bits.uop.preDecodeInfo.isRVC 1242 io.ldu_io.rollback.bits.robIdx := s3_out.bits.uop.robIdx 1243 io.ldu_io.rollback.bits.ftqIdx := s3_out.bits.uop.ftqPtr 1244 io.ldu_io.rollback.bits.ftqOffset := s3_out.bits.uop.ftqOffset 1245 io.ldu_io.rollback.bits.level := Mux(s3_rep_frm_fetch, RedirectLevel.flush, RedirectLevel.flushAfter) 1246 io.ldu_io.rollback.bits.cfiUpdate.target := s3_out.bits.uop.pc 1247 io.ldu_io.rollback.bits.debug_runahead_checkpoint_id := s3_out.bits.uop.debugInfo.runahead_checkpoint_id 1248 /* <------- DANGEROUS: Don't change sequence here ! -------> */ 1249 io.ldu_io.lsq.ldin.bits.uop := s3_out.bits.uop 1250 1251 val s3_revoke = s3_exception || io.ldu_io.lsq.ldin.bits.rep_info.need_rep 1252 io.ldu_io.lsq.ldld_nuke_query.revoke := s3_revoke 1253 io.ldu_io.lsq.stld_nuke_query.revoke := s3_revoke 1254 1255 // feedback slow 1256 s3_fast_rep := RegNext(s2_fast_rep) && 1257 !s3_in.feedbacked && 1258 !s3_in.lateKill && 1259 !s3_rep_frm_fetch && 1260 !s3_exception 1261 1262 val s3_fb_no_waiting = !s3_in.isLoadReplay && !(s3_fast_rep && io.ldu_io.fast_rep_out.ready) && !s3_in.feedbacked 1263 1264 // 1265 io.feedback_slow.valid := s3_valid && !s3_in.uop.robIdx.needFlush(io.redirect) && s3_fb_no_waiting && s3_ld_flow 1266 io.feedback_slow.bits.hit := !io.ldu_io.lsq.ldin.bits.rep_info.need_rep || io.ldu_io.lsq.ldin.ready 1267 io.feedback_slow.bits.flushState := s3_in.ptwBack 1268 io.feedback_slow.bits.robIdx := s3_in.uop.robIdx 1269 io.feedback_slow.bits.sourceType := RSFeedbackType.lrqFull 1270 io.feedback_slow.bits.dataInvalidSqIdx := DontCare 1271 1272 io.vec_stu_io.feedbackSlow.valid := RegNext(s2_vec_feedback.valid && !s2_out.uop.robIdx.needFlush(io.redirect)) 1273 io.vec_stu_io.feedbackSlow.bits := RegNext(s2_vec_feedback.bits) 1274 1275 io.ldu_io.ldCancel.ld2Cancel := s3_valid && s3_ld_flow && ( // is load 1276 io.ldu_io.lsq.ldin.bits.rep_info.need_rep || s3_in.mmio // exe fail or is mmio 1277 ) 1278 1279 // data from dcache hit 1280 val s3_ld_raw_data_frm_cache = Wire(new LoadDataFromDcacheBundle) 1281 s3_ld_raw_data_frm_cache.respDcacheData := io.ldu_io.dcache.resp.bits.data_delayed 1282 s3_ld_raw_data_frm_cache.forwardMask := RegEnable(s2_fwd_mask, s2_valid) 1283 s3_ld_raw_data_frm_cache.forwardData := RegEnable(s2_fwd_data, s2_valid) 1284 s3_ld_raw_data_frm_cache.uop := RegEnable(s2_out.uop, s2_valid) 1285 s3_ld_raw_data_frm_cache.addrOffset := RegEnable(s2_out.paddr(3, 0), s2_valid) 1286 s3_ld_raw_data_frm_cache.forward_D := RegEnable(s2_fwd_frm_d_chan, false.B, s2_valid) || s3_fwd_frm_d_chan_valid 1287 s3_ld_raw_data_frm_cache.forwardData_D := Mux(s3_fwd_frm_d_chan_valid, s3_fwd_data_frm_d_chan, RegEnable(s2_fwd_data_frm_d_chan, s2_valid)) 1288 s3_ld_raw_data_frm_cache.forward_mshr := RegEnable(s2_fwd_frm_mshr, false.B, s2_valid) 1289 s3_ld_raw_data_frm_cache.forwardData_mshr := RegEnable(s2_fwd_data_frm_mshr, s2_valid) 1290 s3_ld_raw_data_frm_cache.forward_result_valid := RegEnable(s2_fwd_data_valid, false.B, s2_valid) 1291 1292 val s3_merged_data_frm_cache = s3_ld_raw_data_frm_cache.mergedData() 1293 val s3_picked_data_frm_cache = LookupTree(s3_ld_raw_data_frm_cache.addrOffset, List( 1294 "b0000".U -> s3_merged_data_frm_cache(63, 0), 1295 "b0001".U -> s3_merged_data_frm_cache(63, 8), 1296 "b0010".U -> s3_merged_data_frm_cache(63, 16), 1297 "b0011".U -> s3_merged_data_frm_cache(63, 24), 1298 "b0100".U -> s3_merged_data_frm_cache(63, 32), 1299 "b0101".U -> s3_merged_data_frm_cache(63, 40), 1300 "b0110".U -> s3_merged_data_frm_cache(63, 48), 1301 "b0111".U -> s3_merged_data_frm_cache(63, 56), 1302 "b1000".U -> s3_merged_data_frm_cache(127, 64), 1303 "b1001".U -> s3_merged_data_frm_cache(127, 72), 1304 "b1010".U -> s3_merged_data_frm_cache(127, 80), 1305 "b1011".U -> s3_merged_data_frm_cache(127, 88), 1306 "b1100".U -> s3_merged_data_frm_cache(127, 96), 1307 "b1101".U -> s3_merged_data_frm_cache(127, 104), 1308 "b1110".U -> s3_merged_data_frm_cache(127, 112), 1309 "b1111".U -> s3_merged_data_frm_cache(127, 120) 1310 )) 1311 val s3_ld_data_frm_cache = rdataHelper(s3_ld_raw_data_frm_cache.uop, s3_picked_data_frm_cache) 1312 1313 // FIXME: add 1 cycle delay ? 1314 io.ldout.bits := s3_out.bits 1315 io.ldout.bits.data := s3_ld_data_frm_cache 1316 io.ldout.valid := s3_out.valid && !s3_out.bits.uop.robIdx.needFlush(io.redirect) && s3_ld_flow && !s3_isvec 1317 1318 // for uncache 1319 io.ldu_io.lsq.uncache.ready := true.B 1320 1321 // fast load to load forward 1322 if (EnableLoadToLoadForward) { 1323 io.ldu_io.l2l_fwd_out.valid := s3_out.valid && !s3_in.lateKill && s3_ld_flow 1324 io.ldu_io.l2l_fwd_out.data := s3_ld_data_frm_cache 1325 io.ldu_io.l2l_fwd_out.dly_ld_err := s3_dly_ld_err // ecc delayed error 1326 } else { 1327 io.ldu_io.l2l_fwd_out.valid := false.B 1328 io.ldu_io.l2l_fwd_out.data := DontCare 1329 io.ldu_io.l2l_fwd_out.dly_ld_err := DontCare 1330 } 1331 1332 // hybrid unit writeback to rob 1333 // delay params 1334 val SelectGroupSize = RollbackGroupSize 1335 val lgSelectGroupSize = log2Ceil(SelectGroupSize) 1336 val TotalSelectCycles = scala.math.ceil(log2Ceil(LoadQueueRAWSize).toFloat / lgSelectGroupSize).toInt + 1 1337 val TotalDelayCycles = TotalSelectCycles - 2 1338 1339 // writeback 1340 val sx_valid = Wire(Vec(TotalDelayCycles + 1, Bool())) 1341 val sx_ready = Wire(Vec(TotalDelayCycles + 1, Bool())) 1342 val sx_in = Wire(Vec(TotalDelayCycles + 1, new MemExuOutput)) 1343 1344 sx_can_go := sx_ready.head 1345 for (i <- 0 until TotalDelayCycles + 1) { 1346 if (i == 0) { 1347 sx_valid(i) := s3_valid && 1348 !s3_ld_flow && 1349 !s3_in.feedbacked && 1350 !s3_in.mmio 1351 sx_in(i) := s3_out.bits 1352 sx_ready(i) := !s3_valid(i) || sx_in(i).uop.robIdx.needFlush(io.redirect) || (if (TotalDelayCycles == 0) io.stout.ready else sx_ready(i+1)) 1353 } else { 1354 val cur_kill = sx_in(i).uop.robIdx.needFlush(io.redirect) 1355 val cur_can_go = (if (i == TotalDelayCycles) io.stout.ready else sx_ready(i+1)) 1356 val cur_fire = sx_valid(i) && !cur_kill && cur_can_go 1357 val prev_fire = sx_valid(i-1) && !sx_in(i-1).uop.robIdx.needFlush(io.redirect) && sx_ready(i) 1358 1359 sx_ready(i) := !sx_valid(i) || cur_kill || (if (i == TotalDelayCycles) io.stout.ready else sx_ready(i+1)) 1360 val sx_valid_can_go = prev_fire || cur_fire || cur_kill 1361 sx_valid(i) := RegEnable(Mux(prev_fire, true.B, false.B), sx_valid_can_go) 1362 sx_in(i) := RegEnable(sx_in(i-1), prev_fire) 1363 } 1364 } 1365 1366 val sx_last_valid = sx_valid.takeRight(1).head 1367 val sx_last_ready = sx_ready.takeRight(1).head 1368 val sx_last_in = sx_in.takeRight(1).head 1369 1370 sx_last_ready := !sx_last_valid || sx_last_in.uop.robIdx.needFlush(io.redirect) || io.stout.ready 1371 io.stout.valid := sx_last_valid && !sx_last_in.uop.robIdx.needFlush(io.redirect) && FuType.isStore(sx_last_in.uop.fuType) 1372 io.stout.bits := sx_last_in 1373 1374 // trigger 1375 val ld_trigger = FuType.isLoad(io.ldout.bits.uop.fuType) 1376 val last_valid_data = RegEnable(io.ldout.bits.data, io.stout.fire) 1377 val hit_ld_addr_trig_hit_vec = Wire(Vec(TriggerNum, Bool())) 1378 val lq_ld_addr_trig_hit_vec = RegNext(io.ldu_io.lsq.trigger.lqLoadAddrTriggerHitVec) 1379 (0 until TriggerNum).map{i => { 1380 val tdata2 = RegNext(RegNext(io.ldu_io.trigger(i).tdata2)) 1381 val matchType = RegNext(RegNext(io.ldu_io.trigger(i).matchType)) 1382 val tEnable = RegNext(RegNext(io.ldu_io.trigger(i).tEnable)) 1383 1384 hit_ld_addr_trig_hit_vec(i) := TriggerCmp(RegNext(s3_in.vaddr), tdata2, matchType, tEnable) 1385 io.ldu_io.trigger(i).addrHit := Mux(io.ldout.valid && ld_trigger, hit_ld_addr_trig_hit_vec(i), lq_ld_addr_trig_hit_vec(i)) 1386 }} 1387 io.ldu_io.lsq.trigger.hitLoadAddrTriggerHitVec := hit_ld_addr_trig_hit_vec 1388 1389 // FIXME: please move this part to LoadQueueReplay 1390 io.ldu_io.debug_ls := DontCare 1391 io.stu_io.debug_ls := DontCare 1392 io.stu_io.debug_ls.s1_isTlbFirstMiss := io.tlb.resp.valid && io.tlb.resp.bits.miss && io.tlb.resp.bits.debug.isFirstIssue && !s1_in.isHWPrefetch && !s1_ld_flow 1393 io.stu_io.debug_ls.s1_robIdx := s1_in.uop.robIdx.value 1394 1395 // Topdown 1396 io.ldu_io.lsTopdownInfo.s1.robIdx := s1_in.uop.robIdx.value 1397 io.ldu_io.lsTopdownInfo.s1.vaddr_valid := s1_valid && s1_in.hasROBEntry 1398 io.ldu_io.lsTopdownInfo.s1.vaddr_bits := s1_vaddr 1399 io.ldu_io.lsTopdownInfo.s2.robIdx := s2_in.uop.robIdx.value 1400 io.ldu_io.lsTopdownInfo.s2.paddr_valid := s2_fire && s2_in.hasROBEntry && !s2_in.tlbMiss 1401 io.ldu_io.lsTopdownInfo.s2.paddr_bits := s2_in.paddr 1402 io.ldu_io.lsTopdownInfo.s2.first_real_miss := io.ldu_io.dcache.resp.bits.real_miss 1403 io.ldu_io.lsTopdownInfo.s2.cache_miss_en := s2_fire && s2_in.hasROBEntry && !s2_in.tlbMiss && !s2_in.missDbUpdated 1404 1405 // perf cnt 1406 XSPerfAccumulate("s0_in_valid", io.lsin.valid) 1407 XSPerfAccumulate("s0_in_block", io.lsin.valid && !io.lsin.fire) 1408 XSPerfAccumulate("s0_in_fire_first_issue", s0_valid && s0_isFirstIssue) 1409 XSPerfAccumulate("s0_lsq_fire_first_issue", io.ldu_io.replay.fire) 1410 XSPerfAccumulate("s0_ldu_fire_first_issue", io.lsin.fire && s0_isFirstIssue) 1411 XSPerfAccumulate("s0_fast_replay_issue", io.ldu_io.fast_rep_in.fire) 1412 XSPerfAccumulate("s0_stall_out", s0_valid && !s0_can_go) 1413 XSPerfAccumulate("s0_stall_ld_dcache", s0_valid && !io.ldu_io.dcache.req.ready) 1414 XSPerfAccumulate("s0_stall_st_dcache", s0_valid && !io.stu_io.dcache.req.ready) 1415 XSPerfAccumulate("s0_addr_spec_success", s0_fire && s0_vaddr(VAddrBits-1, 12) === io.lsin.bits.src(0)(VAddrBits-1, 12)) 1416 XSPerfAccumulate("s0_addr_spec_failed", s0_fire && s0_vaddr(VAddrBits-1, 12) =/= io.lsin.bits.src(0)(VAddrBits-1, 12)) 1417 XSPerfAccumulate("s0_addr_spec_success_once", s0_fire && s0_vaddr(VAddrBits-1, 12) === io.lsin.bits.src(0)(VAddrBits-1, 12) && s0_isFirstIssue) 1418 XSPerfAccumulate("s0_addr_spec_failed_once", s0_fire && s0_vaddr(VAddrBits-1, 12) =/= io.lsin.bits.src(0)(VAddrBits-1, 12) && s0_isFirstIssue) 1419 XSPerfAccumulate("s0_forward_tl_d_channel", s0_out.forward_tlDchannel) 1420 XSPerfAccumulate("s0_hardware_prefetch_fire", s0_fire && s0_hw_prf_select) 1421 XSPerfAccumulate("s0_software_prefetch_fire", s0_fire && s0_prf && s0_int_iss_select) 1422 XSPerfAccumulate("s0_hardware_prefetch_blocked", io.ldu_io.prefetch_req.valid && !s0_hw_prf_select) 1423 XSPerfAccumulate("s0_hardware_prefetch_total", io.ldu_io.prefetch_req.valid) 1424 1425 XSPerfAccumulate("s1_in_valid", s1_valid) 1426 XSPerfAccumulate("s1_in_fire", s1_fire) 1427 XSPerfAccumulate("s1_in_fire_first_issue", s1_fire && s1_in.isFirstIssue) 1428 XSPerfAccumulate("s1_tlb_miss", s1_fire && s1_tlb_miss) 1429 XSPerfAccumulate("s1_tlb_miss_first_issue", s1_fire && s1_tlb_miss && s1_in.isFirstIssue) 1430 XSPerfAccumulate("s1_stall_out", s1_valid && !s1_can_go) 1431 XSPerfAccumulate("s1_late_kill", s1_valid && s1_fast_rep_kill) 1432 1433 XSPerfAccumulate("s2_in_valid", s2_valid) 1434 XSPerfAccumulate("s2_in_fire", s2_fire) 1435 XSPerfAccumulate("s2_in_fire_first_issue", s2_fire && s2_in.isFirstIssue) 1436 XSPerfAccumulate("s2_dcache_miss", s2_fire && io.ldu_io.dcache.resp.bits.miss) 1437 XSPerfAccumulate("s2_dcache_miss_first_issue", s2_fire && io.ldu_io.dcache.resp.bits.miss && s2_in.isFirstIssue) 1438 XSPerfAccumulate("s2_dcache_real_miss_first_issue", s2_fire && io.ldu_io.dcache.resp.bits.miss && s2_in.isFirstIssue) 1439 XSPerfAccumulate("s2_full_forward", s2_fire && s2_full_fwd) 1440 XSPerfAccumulate("s2_dcache_miss_full_forward", s2_fire && s2_dcache_miss) 1441 XSPerfAccumulate("s2_fwd_frm_d_can", s2_valid && s2_fwd_frm_d_chan) 1442 XSPerfAccumulate("s2_fwd_frm_d_chan_or_mshr", s2_valid && s2_fwd_frm_d_chan_or_mshr) 1443 XSPerfAccumulate("s2_stall_out", s2_fire && !s2_can_go) 1444 XSPerfAccumulate("s2_prefetch", s2_fire && s2_prf) 1445 XSPerfAccumulate("s2_prefetch_ignored", s2_fire && s2_prf && io.ldu_io.dcache.s2_mq_nack) // ignore prefetch for mshr full / miss req port conflict 1446 XSPerfAccumulate("s2_prefetch_miss", s2_fire && s2_prf && io.ldu_io.dcache.resp.bits.miss) // prefetch req miss in l1 1447 XSPerfAccumulate("s2_prefetch_hit", s2_fire && s2_prf && !io.ldu_io.dcache.resp.bits.miss) // prefetch req hit in l1 1448 XSPerfAccumulate("s2_prefetch_accept", s2_fire && s2_prf && io.ldu_io.dcache.resp.bits.miss && !io.ldu_io.dcache.s2_mq_nack) // prefetch a missed line in l1, and l1 accepted it 1449 XSPerfAccumulate("s2_forward_req", s2_fire && s2_in.forward_tlDchannel) 1450 XSPerfAccumulate("s2_successfully_forward_channel_D", s2_fire && s2_fwd_frm_d_chan && s2_fwd_data_valid) 1451 XSPerfAccumulate("s2_successfully_forward_mshr", s2_fire && s2_fwd_frm_mshr && s2_fwd_data_valid) 1452 1453 XSPerfAccumulate("s3_fwd_frm_d_chan", s3_valid && s3_fwd_frm_d_chan_valid) 1454 1455 XSPerfAccumulate("load_to_load_forward", s1_try_ptr_chasing && !s1_ptr_chasing_canceled) 1456 XSPerfAccumulate("load_to_load_forward_try", s1_try_ptr_chasing) 1457 XSPerfAccumulate("load_to_load_forward_fail", s1_cancel_ptr_chasing) 1458 XSPerfAccumulate("load_to_load_forward_fail_cancelled", s1_cancel_ptr_chasing && s1_ptr_chasing_canceled) 1459 XSPerfAccumulate("load_to_load_forward_fail_wakeup_mismatch", s1_cancel_ptr_chasing && !s1_ptr_chasing_canceled && s1_not_fast_match) 1460 XSPerfAccumulate("load_to_load_forward_fail_op_not_ld", s1_cancel_ptr_chasing && !s1_ptr_chasing_canceled && !s1_not_fast_match && s1_fu_op_type_not_ld) 1461 XSPerfAccumulate("load_to_load_forward_fail_addr_align", s1_cancel_ptr_chasing && !s1_ptr_chasing_canceled && !s1_not_fast_match && !s1_fu_op_type_not_ld && s1_addr_misaligned) 1462 XSPerfAccumulate("load_to_load_forward_fail_set_mismatch", s1_cancel_ptr_chasing && !s1_ptr_chasing_canceled && !s1_not_fast_match && !s1_fu_op_type_not_ld && !s1_addr_misaligned && s1_addr_mismatch) 1463 1464 // bug lyq: some signals in perfEvents are no longer suitable for the current MemBlock design 1465 // hardware performance counter 1466 val perfEvents = Seq( 1467 ("load_s0_in_fire ", s0_fire ), 1468 ("load_to_load_forward ", s1_fire && s1_try_ptr_chasing && !s1_ptr_chasing_canceled ), 1469 ("stall_dcache ", s0_valid && s0_can_go && !io.ldu_io.dcache.req.ready ), 1470 ("load_s1_in_fire ", s0_fire ), 1471 ("load_s1_tlb_miss ", s1_fire && io.tlb.resp.bits.miss ), 1472 ("load_s2_in_fire ", s1_fire ), 1473 ("load_s2_dcache_miss ", s2_fire && io.ldu_io.dcache.resp.bits.miss ), 1474 ) 1475 generatePerfEvent() 1476}