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.backend.Bundles.{DynInst, MemExuOutput} 26import xiangshan.cache._ 27import xiangshan.cache.{DCacheWordIO, DCacheLineIO, MemoryOpConstants} 28import xiangshan.cache.mmu.{TlbRequestIO, TlbHintIO} 29import xiangshan.mem._ 30import xiangshan.backend._ 31import xiangshan.backend.rob.RobLsqIO 32import coupledL2.{CMOReq, CMOResp} 33import xiangshan.backend.fu.FuType 34 35class ExceptionAddrIO(implicit p: Parameters) extends XSBundle { 36 val isStore = Input(Bool()) 37 val vaddr = Output(UInt(VAddrBits.W)) 38 val vstart = Output(UInt((log2Up(VLEN) + 1).W)) 39 val vl = Output(UInt((log2Up(VLEN) + 1).W)) 40 val gpaddr = Output(UInt(GPAddrBits.W)) 41} 42 43class FwdEntry extends Bundle { 44 val validFast = Bool() // validFast is generated the same cycle with query 45 val valid = Bool() // valid is generated 1 cycle after query request 46 val data = UInt(8.W) // data is generated 1 cycle after query request 47} 48 49// inflight miss block reqs 50class InflightBlockInfo(implicit p: Parameters) extends XSBundle { 51 val block_addr = UInt(PAddrBits.W) 52 val valid = Bool() 53} 54 55class LsqEnqIO(implicit p: Parameters) extends MemBlockBundle { 56 val canAccept = Output(Bool()) 57 val needAlloc = Vec(LSQEnqWidth, Input(UInt(2.W))) 58 val req = Vec(LSQEnqWidth, Flipped(ValidIO(new DynInst))) 59 val iqAccept = Input(Vec(LSQEnqWidth, Bool())) 60 val resp = Vec(LSQEnqWidth, Output(new LSIdx)) 61} 62 63// Load / Store Queue Wrapper for XiangShan Out of Order LSU 64class LsqWrapper(implicit p: Parameters) extends XSModule with HasDCacheParameters with HasPerfEvents { 65 val io = IO(new Bundle() { 66 val hartId = Input(UInt(hartIdLen.W)) 67 val brqRedirect = Flipped(ValidIO(new Redirect)) 68 val stvecFeedback = Vec(VecStorePipelineWidth, Flipped(ValidIO(new FeedbackToLsqIO))) 69 val ldvecFeedback = Vec(VecLoadPipelineWidth, Flipped(ValidIO(new FeedbackToLsqIO))) 70 val enq = new LsqEnqIO 71 val ldu = new Bundle() { 72 val stld_nuke_query = Vec(LoadPipelineWidth, Flipped(new LoadNukeQueryIO)) // from load_s2 73 val ldld_nuke_query = Vec(LoadPipelineWidth, Flipped(new LoadNukeQueryIO)) // from load_s2 74 val ldin = Vec(LoadPipelineWidth, Flipped(Decoupled(new LqWriteBundle))) // from load_s3 75 } 76 val sta = new Bundle() { 77 val storeMaskIn = Vec(StorePipelineWidth, Flipped(Valid(new StoreMaskBundle))) // from store_s0, store mask, send to sq from rs 78 val storeAddrIn = Vec(StorePipelineWidth, Flipped(Valid(new LsPipelineBundle))) // from store_s1 79 val storeAddrInRe = Vec(StorePipelineWidth, Input(new LsPipelineBundle())) // from store_s2 80 } 81 val std = new Bundle() { 82 val storeDataIn = Vec(StorePipelineWidth, Flipped(Valid(new MemExuOutput(isVector = true)))) // from store_s0, store data, send to sq from rs 83 } 84 val ldout = Vec(LoadPipelineWidth, DecoupledIO(new MemExuOutput)) 85 val ld_raw_data = Vec(LoadPipelineWidth, Output(new LoadDataFromLQBundle)) 86 val replay = Vec(LoadPipelineWidth, Decoupled(new LsPipelineBundle)) 87 val sbuffer = Vec(EnsbufferWidth, Decoupled(new DCacheWordReqWithVaddrAndPfFlag)) 88 val sbufferVecDifftestInfo = Vec(EnsbufferWidth, Decoupled(new DynInst)) // The vector store difftest needs is 89 val forward = Vec(LoadPipelineWidth, Flipped(new PipeLoadForwardQueryIO)) 90 val rob = Flipped(new RobLsqIO) 91 val nuke_rollback = Vec(StorePipelineWidth, Output(Valid(new Redirect))) 92 val nack_rollback = Output(Valid(new Redirect)) 93 val release = Flipped(Valid(new Release)) 94 // val refill = Flipped(Valid(new Refill)) 95 val tl_d_channel = Input(new DcacheToLduForwardIO) 96 val maControl = Flipped(new StoreMaBufToSqControlIO) 97 val uncacheOutstanding = Input(Bool()) 98 val uncache = new UncacheWordIO 99 val mmioStout = DecoupledIO(new MemExuOutput) // writeback uncached store 100 // TODO: implement vector store 101 val vecmmioStout = DecoupledIO(new MemExuOutput(isVector = true)) // vec writeback uncached store 102 val sqEmpty = Output(Bool()) 103 val lq_rep_full = Output(Bool()) 104 val sqFull = Output(Bool()) 105 val lqFull = Output(Bool()) 106 val sqCancelCnt = Output(UInt(log2Up(StoreQueueSize+1).W)) 107 val lqCancelCnt = Output(UInt(log2Up(VirtualLoadQueueSize+1).W)) 108 val lqDeq = Output(UInt(log2Up(CommitWidth + 1).W)) 109 val sqDeq = Output(UInt(log2Ceil(EnsbufferWidth + 1).W)) 110 val lqCanAccept = Output(Bool()) 111 val sqCanAccept = Output(Bool()) 112 val lqDeqPtr = Output(new LqPtr) 113 val sqDeqPtr = Output(new SqPtr) 114 val exceptionAddr = new ExceptionAddrIO 115 val flushFrmMaBuf = Input(Bool()) 116 val issuePtrExt = Output(new SqPtr) 117 val l2_hint = Input(Valid(new L2ToL1Hint())) 118 val tlb_hint = Flipped(new TlbHintIO) 119 val cmoOpReq = DecoupledIO(new CMOReq) 120 val cmoOpResp = Flipped(DecoupledIO(new CMOResp)) 121 val flushSbuffer = new SbufferFlushBundle 122 val force_write = Output(Bool()) 123 val lqEmpty = Output(Bool()) 124 125 // top-down 126 val debugTopDown = new LoadQueueTopDownIO 127 }) 128 129 val loadQueue = Module(new LoadQueue) 130 val storeQueue = Module(new StoreQueue) 131 132 storeQueue.io.hartId := io.hartId 133 storeQueue.io.uncacheOutstanding := io.uncacheOutstanding 134 135 136 dontTouch(loadQueue.io.tlbReplayDelayCycleCtrl) 137 // Todo: imm 138 val tlbReplayDelayCycleCtrl = WireInit(VecInit(Seq(14.U(ReSelectLen.W), 0.U(ReSelectLen.W), 125.U(ReSelectLen.W), 0.U(ReSelectLen.W)))) 139 loadQueue.io.tlbReplayDelayCycleCtrl := tlbReplayDelayCycleCtrl 140 141 // io.enq logic 142 // LSQ: send out canAccept when both load queue and store queue are ready 143 // Dispatch: send instructions to LSQ only when they are ready 144 io.enq.canAccept := loadQueue.io.enq.canAccept && storeQueue.io.enq.canAccept 145 io.lqCanAccept := loadQueue.io.enq.canAccept 146 io.sqCanAccept := storeQueue.io.enq.canAccept 147 loadQueue.io.enq.sqCanAccept := storeQueue.io.enq.canAccept 148 storeQueue.io.enq.lqCanAccept := loadQueue.io.enq.canAccept 149 io.lqDeqPtr := loadQueue.io.lqDeqPtr 150 io.sqDeqPtr := storeQueue.io.sqDeqPtr 151 for (i <- io.enq.req.indices) { 152 loadQueue.io.enq.needAlloc(i) := io.enq.needAlloc(i)(0) 153 loadQueue.io.enq.req(i).valid := io.enq.needAlloc(i)(0) && io.enq.req(i).valid 154 loadQueue.io.enq.req(i).bits := io.enq.req(i).bits 155 loadQueue.io.enq.req(i).bits.sqIdx := storeQueue.io.enq.resp(i) 156 157 storeQueue.io.enq.needAlloc(i) := io.enq.needAlloc(i)(1) 158 storeQueue.io.enq.req(i).valid := io.enq.needAlloc(i)(1) && io.enq.req(i).valid 159 storeQueue.io.enq.req(i).bits := io.enq.req(i).bits 160 storeQueue.io.enq.req(i).bits.lqIdx := loadQueue.io.enq.resp(i) 161 162 io.enq.resp(i).lqIdx := loadQueue.io.enq.resp(i) 163 io.enq.resp(i).sqIdx := storeQueue.io.enq.resp(i) 164 } 165 166 // store queue wiring 167 storeQueue.io.brqRedirect <> io.brqRedirect 168 storeQueue.io.vecFeedback <> io.stvecFeedback 169 storeQueue.io.storeAddrIn <> io.sta.storeAddrIn // from store_s1 170 storeQueue.io.storeAddrInRe <> io.sta.storeAddrInRe // from store_s2 171 storeQueue.io.storeDataIn <> io.std.storeDataIn // from store_s0 172 storeQueue.io.storeMaskIn <> io.sta.storeMaskIn // from store_s0 173 storeQueue.io.sbuffer <> io.sbuffer 174 storeQueue.io.sbufferVecDifftestInfo <> io.sbufferVecDifftestInfo 175 storeQueue.io.mmioStout <> io.mmioStout 176 storeQueue.io.vecmmioStout <> io.vecmmioStout 177 storeQueue.io.rob <> io.rob 178 storeQueue.io.exceptionAddr.isStore := DontCare 179 storeQueue.io.sqCancelCnt <> io.sqCancelCnt 180 storeQueue.io.sqDeq <> io.sqDeq 181 storeQueue.io.sqEmpty <> io.sqEmpty 182 storeQueue.io.sqFull <> io.sqFull 183 storeQueue.io.forward <> io.forward // overlap forwardMask & forwardData, DO NOT CHANGE SEQUENCE 184 storeQueue.io.force_write <> io.force_write 185 storeQueue.io.cmoOpReq <> io.cmoOpReq 186 storeQueue.io.cmoOpResp <> io.cmoOpResp 187 storeQueue.io.flushSbuffer <> io.flushSbuffer 188 storeQueue.io.maControl <> io.maControl 189 190 /* <------- DANGEROUS: Don't change sequence here ! -------> */ 191 192 // load queue wiring 193 loadQueue.io.redirect <> io.brqRedirect 194 loadQueue.io.vecFeedback <> io.ldvecFeedback 195 loadQueue.io.ldu <> io.ldu 196 loadQueue.io.ldout <> io.ldout 197 loadQueue.io.ld_raw_data <> io.ld_raw_data 198 loadQueue.io.rob <> io.rob 199 loadQueue.io.nuke_rollback <> io.nuke_rollback 200 loadQueue.io.nack_rollback <> io.nack_rollback 201 loadQueue.io.replay <> io.replay 202 // loadQueue.io.refill <> io.refill 203 loadQueue.io.tl_d_channel <> io.tl_d_channel 204 loadQueue.io.release <> io.release 205 loadQueue.io.exceptionAddr.isStore := DontCare 206 loadQueue.io.flushFrmMaBuf := io.flushFrmMaBuf 207 loadQueue.io.lqCancelCnt <> io.lqCancelCnt 208 loadQueue.io.sq.stAddrReadySqPtr <> storeQueue.io.stAddrReadySqPtr 209 loadQueue.io.sq.stAddrReadyVec <> storeQueue.io.stAddrReadyVec 210 loadQueue.io.sq.stDataReadySqPtr <> storeQueue.io.stDataReadySqPtr 211 loadQueue.io.sq.stDataReadyVec <> storeQueue.io.stDataReadyVec 212 loadQueue.io.sq.stIssuePtr <> storeQueue.io.stIssuePtr 213 loadQueue.io.sq.sqEmpty <> storeQueue.io.sqEmpty 214 loadQueue.io.sta.storeAddrIn <> io.sta.storeAddrIn // store_s1 215 loadQueue.io.std.storeDataIn <> io.std.storeDataIn // store_s0 216 loadQueue.io.lqFull <> io.lqFull 217 loadQueue.io.lq_rep_full <> io.lq_rep_full 218 loadQueue.io.lqDeq <> io.lqDeq 219 loadQueue.io.l2_hint <> io.l2_hint 220 loadQueue.io.tlb_hint <> io.tlb_hint 221 loadQueue.io.lqEmpty <> io.lqEmpty 222 223 // rob commits for lsq is delayed for two cycles, which causes the delayed update for deqPtr in lq/sq 224 // s0: commit 225 // s1: exception find 226 // s2: exception triggered 227 // s3: ptr updated & new address 228 // address will be used at the next cycle after exception is triggered 229 io.exceptionAddr.vaddr := Mux(RegNext(io.exceptionAddr.isStore), storeQueue.io.exceptionAddr.vaddr, loadQueue.io.exceptionAddr.vaddr) 230 io.exceptionAddr.vstart := Mux(RegNext(io.exceptionAddr.isStore), storeQueue.io.exceptionAddr.vstart, loadQueue.io.exceptionAddr.vstart) 231 io.exceptionAddr.vl := Mux(RegNext(io.exceptionAddr.isStore), storeQueue.io.exceptionAddr.vl, loadQueue.io.exceptionAddr.vl) 232 io.exceptionAddr.gpaddr := Mux(RegNext(io.exceptionAddr.isStore), storeQueue.io.exceptionAddr.gpaddr, loadQueue.io.exceptionAddr.gpaddr) 233 io.issuePtrExt := storeQueue.io.stAddrReadySqPtr 234 235 // naive uncache arbiter 236 val s_idle :: s_load :: s_store :: Nil = Enum(3) 237 val pendingstate = RegInit(s_idle) 238 239 switch(pendingstate){ 240 is(s_idle){ 241 when(io.uncache.req.fire){ 242 pendingstate := Mux(loadQueue.io.uncache.req.valid, s_load, 243 Mux(io.uncacheOutstanding, s_idle, s_store)) 244 } 245 } 246 is(s_load){ 247 when(io.uncache.resp.fire){ 248 pendingstate := s_idle 249 } 250 } 251 is(s_store){ 252 when(io.uncache.resp.fire){ 253 pendingstate := s_idle 254 } 255 } 256 } 257 258 loadQueue.io.uncache := DontCare 259 storeQueue.io.uncache := DontCare 260 loadQueue.io.uncache.req.ready := false.B 261 storeQueue.io.uncache.req.ready := false.B 262 loadQueue.io.uncache.resp.valid := false.B 263 storeQueue.io.uncache.resp.valid := false.B 264 when(loadQueue.io.uncache.req.valid){ 265 io.uncache.req <> loadQueue.io.uncache.req 266 }.otherwise{ 267 io.uncache.req <> storeQueue.io.uncache.req 268 } 269 when (io.uncacheOutstanding) { 270 io.uncache.resp <> loadQueue.io.uncache.resp 271 } .otherwise { 272 when(pendingstate === s_load){ 273 io.uncache.resp <> loadQueue.io.uncache.resp 274 }.otherwise{ 275 io.uncache.resp <> storeQueue.io.uncache.resp 276 } 277 } 278 279 loadQueue.io.debugTopDown <> io.debugTopDown 280 281 assert(!(loadQueue.io.uncache.req.valid && storeQueue.io.uncache.req.valid)) 282 assert(!(loadQueue.io.uncache.resp.valid && storeQueue.io.uncache.resp.valid)) 283 when (!io.uncacheOutstanding) { 284 assert(!((loadQueue.io.uncache.resp.valid || storeQueue.io.uncache.resp.valid) && pendingstate === s_idle)) 285 } 286 287 288 val perfEvents = Seq(loadQueue, storeQueue).flatMap(_.getPerfEvents) 289 generatePerfEvent() 290} 291 292class LsqEnqCtrl(implicit p: Parameters) extends XSModule 293 with HasVLSUParameters { 294 val io = IO(new Bundle { 295 val redirect = Flipped(ValidIO(new Redirect)) 296 // to dispatch 297 val enq = new LsqEnqIO 298 // from `memBlock.io.lqDeq 299 val lcommit = Input(UInt(log2Up(CommitWidth + 1).W)) 300 // from `memBlock.io.sqDeq` 301 val scommit = Input(UInt(log2Ceil(EnsbufferWidth + 1).W)) 302 // from/tp lsq 303 val lqCancelCnt = Input(UInt(log2Up(VirtualLoadQueueSize + 1).W)) 304 val sqCancelCnt = Input(UInt(log2Up(StoreQueueSize + 1).W)) 305 val lqFreeCount = Output(UInt(log2Up(VirtualLoadQueueSize + 1).W)) 306 val sqFreeCount = Output(UInt(log2Up(StoreQueueSize + 1).W)) 307 val enqLsq = Flipped(new LsqEnqIO) 308 }) 309 310 val lqPtr = RegInit(0.U.asTypeOf(new LqPtr)) 311 val sqPtr = RegInit(0.U.asTypeOf(new SqPtr)) 312 val lqCounter = RegInit(VirtualLoadQueueSize.U(log2Up(VirtualLoadQueueSize + 1).W)) 313 val sqCounter = RegInit(StoreQueueSize.U(log2Up(StoreQueueSize + 1).W)) 314 val canAccept = RegInit(false.B) 315 316 val blockVec = io.enq.iqAccept.map(!_) :+ true.B 317 val numLsElem = io.enq.req.map(_.bits.numLsElem) 318 val needEnqLoadQueue = VecInit(io.enq.req.map(x => FuType.isLoad(x.bits.fuType) || FuType.isVNonsegLoad(x.bits.fuType))) 319 val needEnqStoreQueue = VecInit(io.enq.req.map(x => FuType.isStore(x.bits.fuType) || FuType.isVNonsegStore(x.bits.fuType))) 320 val loadQueueElem = needEnqLoadQueue.zip(numLsElem).map(x => Mux(x._1, x._2, 0.U)) 321 val storeQueueElem = needEnqStoreQueue.zip(numLsElem).map(x => Mux(x._1, x._2, 0.U)) 322 val loadFlowPopCount = 0.U +: loadQueueElem.zipWithIndex.map{ case (l, i) => 323 loadQueueElem.take(i + 1).reduce(_ + _) 324 } 325 val storeFlowPopCount = 0.U +: storeQueueElem.zipWithIndex.map { case (s, i) => 326 storeQueueElem.take(i + 1).reduce(_ + _) 327 } 328 val lqAllocNumber = PriorityMux(blockVec.zip(loadFlowPopCount)) 329 val sqAllocNumber = PriorityMux(blockVec.zip(storeFlowPopCount)) 330 331 io.lqFreeCount := lqCounter 332 io.sqFreeCount := sqCounter 333 // How to update ptr and counter: 334 // (1) by default, updated according to enq/commit 335 // (2) when redirect and dispatch queue is empty, update according to lsq 336 val t1_redirect = RegNext(io.redirect.valid) 337 val t2_redirect = RegNext(t1_redirect) 338 val t2_update = t2_redirect && !VecInit(io.enq.needAlloc.map(_.orR)).asUInt.orR 339 val t3_update = RegNext(t2_update) 340 val t3_lqCancelCnt = GatedRegNext(io.lqCancelCnt) 341 val t3_sqCancelCnt = GatedRegNext(io.sqCancelCnt) 342 when (t3_update) { 343 lqPtr := lqPtr - t3_lqCancelCnt 344 lqCounter := lqCounter + io.lcommit + t3_lqCancelCnt 345 sqPtr := sqPtr - t3_sqCancelCnt 346 sqCounter := sqCounter + io.scommit + t3_sqCancelCnt 347 }.elsewhen (!io.redirect.valid && io.enq.canAccept) { 348 lqPtr := lqPtr + lqAllocNumber 349 lqCounter := lqCounter + io.lcommit - lqAllocNumber 350 sqPtr := sqPtr + sqAllocNumber 351 sqCounter := sqCounter + io.scommit - sqAllocNumber 352 }.otherwise { 353 lqCounter := lqCounter + io.lcommit 354 sqCounter := sqCounter + io.scommit 355 } 356 357 358 //TODO MaxAllocate and width of lqOffset/sqOffset needs to be discussed 359 val lqMaxAllocate = LSQLdEnqWidth 360 val sqMaxAllocate = LSQStEnqWidth 361 val maxAllocate = lqMaxAllocate max sqMaxAllocate 362 val ldCanAccept = lqCounter >= lqAllocNumber +& lqMaxAllocate.U 363 val sqCanAccept = sqCounter >= sqAllocNumber +& sqMaxAllocate.U 364 // It is possible that t3_update and enq are true at the same clock cycle. 365 // For example, if redirect.valid lasts more than one clock cycle, 366 // after the last redirect, new instructions may enter but previously redirect has not been resolved (updated according to the cancel count from LSQ). 367 // To solve the issue easily, we block enqueue when t3_update, which is RegNext(t2_update). 368 io.enq.canAccept := RegNext(ldCanAccept && sqCanAccept && !t2_update) 369 val lqOffset = Wire(Vec(io.enq.resp.length, UInt(lqPtr.value.getWidth.W))) 370 val sqOffset = Wire(Vec(io.enq.resp.length, UInt(sqPtr.value.getWidth.W))) 371 for ((resp, i) <- io.enq.resp.zipWithIndex) { 372 lqOffset(i) := loadFlowPopCount(i) 373 resp.lqIdx := lqPtr + lqOffset(i) 374 sqOffset(i) := storeFlowPopCount(i) 375 resp.sqIdx := sqPtr + sqOffset(i) 376 } 377 378 io.enqLsq.needAlloc := RegNext(io.enq.needAlloc) 379 io.enqLsq.iqAccept := RegNext(io.enq.iqAccept) 380 io.enqLsq.req.zip(io.enq.req).zip(io.enq.resp).foreach{ case ((toLsq, enq), resp) => 381 val do_enq = enq.valid && !io.redirect.valid && io.enq.canAccept 382 toLsq.valid := RegNext(do_enq) 383 toLsq.bits := RegEnable(enq.bits, do_enq) 384 toLsq.bits.lqIdx := RegEnable(resp.lqIdx, do_enq) 385 toLsq.bits.sqIdx := RegEnable(resp.sqIdx, do_enq) 386 } 387 388}