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