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.rob.RobPtr 26import xiangshan.backend.Bundles._ 27import xiangshan.mem._ 28import xiangshan.backend.fu.FuType 29import freechips.rocketchip.diplomacy.BufferParams 30import xiangshan.cache.mmu._ 31import xiangshan.cache._ 32import xiangshan.cache.wpu.ReplayCarry 33import xiangshan.backend.fu.util.SdtrigExt 34import xiangshan.ExceptionNO._ 35import xiangshan.backend.fu.vector.Bundles.VConfig 36import xiangshan.backend.datapath.NewPipelineConnect 37import xiangshan.backend.fu.vector.Utils.VecDataToMaskDataVec 38 39class VSegmentBundle(implicit p: Parameters) extends VLSUBundle 40{ 41 val baseVaddr = UInt(XLEN.W) 42 val uop = new DynInst 43 val paddr = UInt(PAddrBits.W) 44 val mask = UInt(VLEN.W) 45 val alignedType = UInt(alignTypeBits.W) 46 val vl = UInt(elemIdxBits.W) 47 val uopFlowNum = UInt(elemIdxBits.W) 48 val uopFlowNumMask = UInt(elemIdxBits.W) 49 // for exception 50 val vstart = UInt(elemIdxBits.W) 51 val exceptionVaddr = UInt(XLEN.W) 52 val exceptionGpaddr = UInt(XLEN.W) 53 val exceptionIsForVSnonLeafPTE = Bool() 54 val exception_va = Bool() 55 val exception_gpa = Bool() 56 val exception_pa = Bool() 57 val exceptionVstart = UInt(elemIdxBits.W) 58 val exceptionVl = UInt(elemIdxBits.W) 59 val isFof = Bool() 60} 61 62// latch each uop's VecWen, pdest, v0Wen, uopIdx 63class VSegmentUop(implicit p: Parameters) extends VLSUBundle{ 64 val uop = new DynInst 65} 66 67class VSegmentUnit (implicit p: Parameters) extends VLSUModule 68 with HasDCacheParameters 69 with MemoryOpConstants 70 with SdtrigExt 71 with HasLoadHelper 72{ 73 val io = IO(new VSegmentUnitIO) 74 75 val maxSize = VSegmentBufferSize 76 77 class VSegUPtr(implicit p: Parameters) extends CircularQueuePtr[VSegUPtr](maxSize){ 78 } 79 80 object VSegUPtr { 81 def apply(f: Bool, v: UInt)(implicit p: Parameters): VSegUPtr = { 82 val ptr = Wire(new VSegUPtr) 83 ptr.flag := f 84 ptr.value := v 85 ptr 86 } 87 } 88 89 90 /** 91 ******************************************************************************************************** 92 * Use an example to illustrate the working logic of a segmentunit: * 93 * For: * 94 * lmul=2 sew=32 emul=2 eew=32 vl=16 * 95 * Then: * 96 * Access memory in the order: * 97 * (V2,S0),(V4,S0),(V6,S0),(V8,S0), * 98 * (V2,S1),(V4,S1),(V6,S1),(V8,S1), * 99 * (V2,S2),(V4,S2),(V6,S2),(V8,S2), * 100 * (V2,S3),(V4,S3),(V6,S3),(V8,S3), * 101 * (V3,S4),(V5,S4),(V7,S4),(V9,S4), * 102 * (V3,S5),(V5,S5),(V7,S5),(V9,S5), * 103 * (V3,S6),(V5,S6),(V7,S6),(V9,S6), * 104 * (V3,S7),(V5,S7),(V7,S7),(V9,S7), * 105 * * 106 * * 107 * [[data]] saves the data generated by the access and corresponds to the register. * 108 * [[splitPtr]] controls the destination register written to. * 109 * * 110 * splitptr offset can be seen in [[splitPtrNext]] is assignment logic, * 111 * which is mainly calculated in terms of [[fieldIdx]] and [[segmentIdx]] * 112 * First access different fields of the same segment, and then visit different segments. * 113 * For the case of 'emul' greater than 1, such as the following example, * 114 * although 'v2' and 'v3' are different vd and the same field, they are still different segments, * 115 * so they should be accessed sequentially.Just like the 'Access memory in the order' above. * 116 * * 117 * [[segmentIdx]] * 118 * | * 119 * | * 120 * V * 121 * * 122 * S0 S1 S2 S3 * 123 * ---------------------------------------------------------------------------- * 124 * [[splitPtr]]--> v2 | field0 | field0 | field0 | field0 | * 125 * ---------------------------------------------------------------------------- * 126 * S4 S5 S6 S7 * 127 * ---------------------------------------------------------------------------- * 128 * v3 | field0 | field0 | field0 | field0 | * 129 * ---------------------------------------------------------------------------- * 130 * S0 S1 S2 S3 * 131 * ---------------------------------------------------------------------------- * 132 * v4 | field1 | field1 | field1 | field1 | * 133 * ---------------------------------------------------------------------------- * 134 * S4 S5 S6 S7 * 135 * ---------------------------------------------------------------------------- * 136 * v5 | field1 | field1 | field1 | field1 | * 137 * ---------------------------------------------------------------------------- * 138 * S0 S1 S2 S3 * 139 * ---------------------------------------------------------------------------- * 140 * v6 | field2 | field2 | field2 | field2 | * 141 * ---------------------------------------------------------------------------- * 142 * S4 S5 S6 S7 * 143 * ---------------------------------------------------------------------------- * 144 * v7 | field2 | field2 | field2 | field2 | * 145 * ---------------------------------------------------------------------------- * 146 * S0 S1 S2 S3 * 147 * ---------------------------------------------------------------------------- * 148 * v8 | field3 | field3 | field3 | field3 | * 149 * ---------------------------------------------------------------------------- * 150 * S4 S5 S6 S7 * 151 * ---------------------------------------------------------------------------- * 152 * v9 | field3 | field3 | field3 | field3 | * 153 * ---------------------------------------------------------------------------- * * 154 * * * 155 * * * 156 ******************************************************************************************************** 157 **/ 158 159 160 // buffer uop 161 val instMicroOp = Reg(new VSegmentBundle) 162 val instMicroOpValid = RegInit(false.B) 163 val data = Reg(Vec(maxSize, UInt(VLEN.W))) 164 val uopq = Reg(Vec(maxSize, new VSegmentUop)) 165 val stride = Reg(Vec(maxSize, UInt(VLEN.W))) 166 val allocated = RegInit(VecInit(Seq.fill(maxSize)(false.B))) 167 val enqPtr = RegInit(0.U.asTypeOf(new VSegUPtr)) 168 val deqPtr = RegInit(0.U.asTypeOf(new VSegUPtr)) 169 val stridePtr = WireInit(0.U.asTypeOf(new VSegUPtr)) // for select stride/index 170 171 val segmentIdx = RegInit(0.U(elemIdxBits.W)) 172 val fieldIdx = RegInit(0.U(fieldBits.W)) 173 val segmentOffset = RegInit(0.U(XLEN.W)) 174 val splitPtr = RegInit(0.U.asTypeOf(new VSegUPtr)) // for select load/store data 175 val splitPtrNext = WireInit(0.U.asTypeOf(new VSegUPtr)) 176 177 val exception_va = WireInit(false.B) 178 val exception_gpa = WireInit(false.B) 179 val exception_pa = WireInit(false.B) 180 181 val maxSegIdx = instMicroOp.vl - 1.U 182 val maxNfields = instMicroOp.uop.vpu.nf 183 val latchVaddr = RegInit(0.U(VAddrBits.W)) 184 185 XSError((segmentIdx > maxSegIdx) && instMicroOpValid, s"segmentIdx > vl, something error!\n") 186 XSError((fieldIdx > maxNfields) && instMicroOpValid, s"fieldIdx > nfields, something error!\n") 187 188 // MicroOp 189 val baseVaddr = instMicroOp.baseVaddr 190 val alignedType = instMicroOp.alignedType 191 val fuType = instMicroOp.uop.fuType 192 val mask = instMicroOp.mask 193 val exceptionVec = instMicroOp.uop.exceptionVec 194 val issueEew = instMicroOp.uop.vpu.veew 195 val issueLmul = instMicroOp.uop.vpu.vtype.vlmul 196 val issueSew = instMicroOp.uop.vpu.vtype.vsew 197 val issueEmul = EewLog2(issueEew) - issueSew + issueLmul 198 val elemIdxInVd = segmentIdx & instMicroOp.uopFlowNumMask 199 val issueInstType = Cat(true.B, instMicroOp.uop.fuOpType(6, 5)) // always segment instruction 200 val issueUopFlowNumLog2 = GenRealFlowLog2(issueInstType, issueEmul, issueLmul, issueEew, issueSew, true) // max element number log2 in vd 201 val issueVlMax = instMicroOp.uopFlowNum // max elementIdx in vd 202 val issueMaxIdxInIndex = GenVLMAX(Mux(issueEmul.asSInt > 0.S, 0.U, issueEmul), issueEew(1, 0)) // index element index in index register 203 val issueMaxIdxInIndexMask = GenVlMaxMask(issueMaxIdxInIndex, elemIdxBits) 204 val issueMaxIdxInIndexLog2 = GenVLMAXLog2(Mux(issueEmul.asSInt > 0.S, 0.U, issueEmul), issueEew(1, 0)) 205 val issueIndexIdx = segmentIdx & issueMaxIdxInIndexMask 206 val segmentActive = (mask & UIntToOH(segmentIdx)).orR 207 208 // sbuffer write interface 209 val sbufferOut = Wire(Decoupled(new DCacheWordReqWithVaddrAndPfFlag)) 210 211 // Segment instruction's FSM 212 /* 213 * s_idle: wait request 214 * s_flush_sbuffer_req: flush sbuffer 215 * s_wait_flush_sbuffer_resp: wait sbuffer empty 216 * s_tlb_req: request tlb 217 * s_wait_tlb_resp: wait tlb resp 218 * s_pm: check pmp 219 * s_cache_req: request cache 220 * s_cache_resp: wait cache resp 221 * s_latch_and_merge_data: for read data 222 * s_send_data: for send write data 223 * s_wait_to_sbuffer: Wait for data from the sbufferOut pipelayer to be sent to the sbuffer 224 * s_finish: 225 * */ 226 val s_idle :: s_flush_sbuffer_req :: s_wait_flush_sbuffer_resp :: s_tlb_req :: s_wait_tlb_resp :: s_pm ::s_cache_req :: s_cache_resp :: s_latch_and_merge_data :: s_send_data :: s_wait_to_sbuffer :: s_finish :: Nil = Enum(12) 227 val state = RegInit(s_idle) 228 val stateNext = WireInit(s_idle) 229 val sbufferEmpty = io.flush_sbuffer.empty 230 231 /** 232 * state update 233 */ 234 state := stateNext 235 236 /** 237 * state transfer 238 */ 239 when(state === s_idle){ 240 stateNext := Mux(isAfter(enqPtr, deqPtr), s_flush_sbuffer_req, s_idle) 241 }.elsewhen(state === s_flush_sbuffer_req){ 242 stateNext := Mux(sbufferEmpty, s_tlb_req, s_wait_flush_sbuffer_resp) // if sbuffer is empty, go to query tlb 243 244 }.elsewhen(state === s_wait_flush_sbuffer_resp){ 245 stateNext := Mux(sbufferEmpty, s_tlb_req, s_wait_flush_sbuffer_resp) 246 247 }.elsewhen(state === s_tlb_req){ 248 stateNext := Mux(segmentActive, s_wait_tlb_resp, Mux(FuType.isVLoad(instMicroOp.uop.fuType), s_latch_and_merge_data, s_send_data)) 249 250 }.elsewhen(state === s_wait_tlb_resp){ 251 stateNext := Mux(io.dtlb.resp.fire, 252 Mux(!io.dtlb.resp.bits.miss, 253 s_pm, 254 s_tlb_req), 255 s_wait_tlb_resp) 256 257 }.elsewhen(state === s_pm){ 258 /* if is vStore, send data to sbuffer, so don't need query dcache */ 259 stateNext := Mux(exception_pa || exception_va || exception_gpa, 260 s_finish, 261 Mux(FuType.isVLoad(instMicroOp.uop.fuType), s_cache_req, s_send_data)) 262 263 }.elsewhen(state === s_cache_req){ 264 stateNext := Mux(io.rdcache.req.fire, s_cache_resp, s_cache_req) 265 266 }.elsewhen(state === s_cache_resp){ 267 when(io.rdcache.resp.fire) { 268 when(io.rdcache.resp.bits.miss || io.rdcache.s2_bank_conflict) { 269 stateNext := s_cache_req 270 }.otherwise { 271 stateNext := Mux(FuType.isVLoad(instMicroOp.uop.fuType), s_latch_and_merge_data, s_send_data) 272 } 273 }.otherwise{ 274 stateNext := s_cache_resp 275 } 276 /* if segment is inactive, don't need to wait access all of the field */ 277 }.elsewhen(state === s_latch_and_merge_data) { 278 when((segmentIdx === maxSegIdx) && (fieldIdx === maxNfields) || 279 ((segmentIdx === maxSegIdx) && !segmentActive)) { 280 281 stateNext := s_finish // segment instruction finish 282 }.otherwise { 283 stateNext := s_tlb_req // need continue 284 } 285 /* if segment is inactive, don't need to wait access all of the field */ 286 }.elsewhen(state === s_send_data) { // when sbuffer accept data 287 when(!sbufferOut.fire && segmentActive) { 288 stateNext := s_send_data 289 }.elsewhen(segmentIdx === maxSegIdx && (fieldIdx === maxNfields && sbufferOut.fire || !segmentActive && io.sbuffer.valid && !io.sbuffer.ready)) { 290 stateNext := s_wait_to_sbuffer 291 }.elsewhen(segmentIdx === maxSegIdx && !segmentActive){ 292 stateNext := s_finish // segment instruction finish 293 }.otherwise { 294 stateNext := s_tlb_req // need continue 295 } 296 297 }.elsewhen(state === s_wait_to_sbuffer){ 298 stateNext := Mux(io.sbuffer.fire, s_finish, s_wait_to_sbuffer) 299 300 }.elsewhen(state === s_finish){ // writeback uop 301 stateNext := Mux(distanceBetween(enqPtr, deqPtr) === 0.U, s_idle, s_finish) 302 303 }.otherwise{ 304 stateNext := s_idle 305 XSError(true.B, s"Unknown state!\n") 306 } 307 308 /************************************************************************* 309 * enqueue logic 310 *************************************************************************/ 311 io.in.ready := true.B 312 val fuOpType = io.in.bits.uop.fuOpType 313 val vtype = io.in.bits.uop.vpu.vtype 314 val mop = fuOpType(6, 5) 315 val instType = Cat(true.B, mop) 316 val eew = io.in.bits.uop.vpu.veew 317 val sew = vtype.vsew 318 val lmul = vtype.vlmul 319 val emul = EewLog2(eew) - sew + lmul 320 val vl = instMicroOp.vl 321 val vm = instMicroOp.uop.vpu.vm 322 val vstart = instMicroOp.uop.vpu.vstart 323 val srcMask = GenFlowMask(Mux(vm, Fill(VLEN, 1.U(1.W)), io.in.bits.src_mask), vstart, vl, true) 324 // first uop enqueue, we need to latch microOp of segment instruction 325 when(io.in.fire && !instMicroOpValid){ 326 // element number in a vd 327 // TODO Rewrite it in a more elegant way. 328 val uopFlowNum = ZeroExt(GenRealFlowNum(instType, emul, lmul, eew, sew, true), elemIdxBits) 329 instMicroOp.baseVaddr := io.in.bits.src_rs1 330 instMicroOpValid := true.B // if is first uop 331 instMicroOp.alignedType := Mux(isIndexed(instType), sew(1, 0), eew) 332 instMicroOp.uop := io.in.bits.uop 333 instMicroOp.mask := srcMask 334 instMicroOp.vstart := 0.U 335 instMicroOp.uopFlowNum := uopFlowNum 336 instMicroOp.uopFlowNumMask := GenVlMaxMask(uopFlowNum, elemIdxBits) // for merge data 337 instMicroOp.vl := io.in.bits.src_vl.asTypeOf(VConfig()).vl 338 segmentOffset := 0.U 339 instMicroOp.isFof := (fuOpType === VlduType.vleff) && FuType.isVLoad(fuType) 340 } 341 // latch data 342 when(io.in.fire){ 343 data(enqPtr.value) := io.in.bits.src_vs3 344 stride(enqPtr.value) := io.in.bits.src_stride 345 uopq(enqPtr.value).uop := io.in.bits.uop 346 } 347 348 // update enqptr, only 1 port 349 when(io.in.fire){ 350 enqPtr := enqPtr + 1.U 351 } 352 353 /************************************************************************* 354 * output logic 355 *************************************************************************/ 356 357 val indexStride = IndexAddr( // index for indexed instruction 358 index = stride(stridePtr.value), 359 flow_inner_idx = issueIndexIdx, 360 eew = issueEew 361 ) 362 val realSegmentOffset = Mux(isIndexed(issueInstType), 363 indexStride, 364 segmentOffset) 365 val vaddr = baseVaddr + (fieldIdx << alignedType).asUInt + realSegmentOffset 366 367 //latch vaddr 368 when(state === s_tlb_req){ 369 latchVaddr := vaddr(VAddrBits - 1, 0) 370 } 371 /** 372 * tlb req and tlb resq 373 */ 374 375 // query DTLB IO Assign 376 io.dtlb.req := DontCare 377 io.dtlb.resp.ready := true.B 378 io.dtlb.req.valid := state === s_tlb_req && segmentActive 379 io.dtlb.req.bits.cmd := Mux(FuType.isVLoad(fuType), TlbCmd.read, TlbCmd.write) 380 io.dtlb.req.bits.vaddr := vaddr(VAddrBits - 1, 0) 381 io.dtlb.req.bits.fullva := vaddr 382 io.dtlb.req.bits.checkfullva := true.B 383 io.dtlb.req.bits.size := instMicroOp.alignedType(2,0) 384 io.dtlb.req.bits.memidx.is_ld := FuType.isVLoad(fuType) 385 io.dtlb.req.bits.memidx.is_st := FuType.isVStore(fuType) 386 io.dtlb.req.bits.debug.robIdx := instMicroOp.uop.robIdx 387 io.dtlb.req.bits.no_translate := false.B 388 io.dtlb.req.bits.debug.pc := instMicroOp.uop.pc 389 io.dtlb.req.bits.debug.isFirstIssue := DontCare 390 io.dtlb.req_kill := false.B 391 392 val canTriggerException = segmentIdx === 0.U || !instMicroOp.isFof // only elementIdx = 0 or is not fof can trigger 393 // tlb resp 394 when(io.dtlb.resp.fire && state === s_wait_tlb_resp){ 395 exceptionVec(storePageFault) := io.dtlb.resp.bits.excp(0).pf.st && canTriggerException 396 exceptionVec(loadPageFault) := io.dtlb.resp.bits.excp(0).pf.ld && canTriggerException 397 exceptionVec(storeGuestPageFault) := io.dtlb.resp.bits.excp(0).gpf.st && canTriggerException 398 exceptionVec(loadGuestPageFault) := io.dtlb.resp.bits.excp(0).gpf.ld && canTriggerException 399 exceptionVec(storeAccessFault) := io.dtlb.resp.bits.excp(0).af.st && canTriggerException 400 exceptionVec(loadAccessFault) := io.dtlb.resp.bits.excp(0).af.ld && canTriggerException 401 when(!io.dtlb.resp.bits.miss){ 402 instMicroOp.paddr := io.dtlb.resp.bits.paddr(0) 403 instMicroOp.exceptionGpaddr := io.dtlb.resp.bits.gpaddr(0) 404 instMicroOp.exceptionIsForVSnonLeafPTE := io.dtlb.resp.bits.isForVSnonLeafPTE 405 } 406 } 407 // pmp 408 // NOTE: only handle load/store exception here, if other exception happens, don't send here 409 val pmp = WireInit(io.pmpResp) 410 when(state === s_pm) { 411 val addr_aligned = LookupTree(Mux(isIndexed(issueInstType), issueSew(1, 0), issueEew(1, 0)), List( 412 "b00".U -> true.B, //b 413 "b01".U -> (vaddr(0) === 0.U), //h 414 "b10".U -> (vaddr(1, 0) === 0.U), //w 415 "b11".U -> (vaddr(2, 0) === 0.U) //d 416 )) 417 val missAligned = !addr_aligned 418 exceptionVec(loadAddrMisaligned) := missAligned && FuType.isVLoad(fuType) && canTriggerException 419 exceptionVec(storeAddrMisaligned) := missAligned && FuType.isVStore(fuType) && canTriggerException 420 421 exception_va := exceptionVec(storePageFault) || exceptionVec(loadPageFault) || 422 exceptionVec(storeAccessFault) || exceptionVec(loadAccessFault) || (missAligned && canTriggerException) 423 exception_gpa := exceptionVec(storeGuestPageFault) || exceptionVec(loadGuestPageFault) 424 exception_pa := (pmp.st || pmp.ld || pmp.mmio) && canTriggerException 425 426 instMicroOp.exception_pa := exception_pa 427 instMicroOp.exception_va := exception_va 428 instMicroOp.exception_gpa := exception_gpa 429 // update storeAccessFault bit. Currently, we don't support vector MMIO 430 exceptionVec(loadAccessFault) := (exceptionVec(loadAccessFault) || pmp.ld || pmp.mmio) && canTriggerException 431 exceptionVec(storeAccessFault) := (exceptionVec(storeAccessFault) || pmp.st || pmp.mmio) && canTriggerException 432 433 when(exception_va || exception_gpa || exception_pa) { 434 when(canTriggerException) { 435 instMicroOp.exceptionVaddr := vaddr 436 instMicroOp.exceptionVl := segmentIdx // for exception 437 instMicroOp.exceptionVstart := segmentIdx // for exception 438 }.otherwise { 439 instMicroOp.exceptionVl := segmentIdx 440 } 441 } 442 } 443 444 /** 445 * flush sbuffer IO Assign 446 */ 447 io.flush_sbuffer.valid := !sbufferEmpty && (state === s_flush_sbuffer_req) 448 449 450 /** 451 * merge data for load 452 */ 453 val cacheData = LookupTree(latchVaddr(3,0), List( 454 "b0000".U -> io.rdcache.resp.bits.data_delayed(63, 0), 455 "b0001".U -> io.rdcache.resp.bits.data_delayed(63, 8), 456 "b0010".U -> io.rdcache.resp.bits.data_delayed(63, 16), 457 "b0011".U -> io.rdcache.resp.bits.data_delayed(63, 24), 458 "b0100".U -> io.rdcache.resp.bits.data_delayed(63, 32), 459 "b0101".U -> io.rdcache.resp.bits.data_delayed(63, 40), 460 "b0110".U -> io.rdcache.resp.bits.data_delayed(63, 48), 461 "b0111".U -> io.rdcache.resp.bits.data_delayed(63, 56), 462 "b1000".U -> io.rdcache.resp.bits.data_delayed(127, 64), 463 "b1001".U -> io.rdcache.resp.bits.data_delayed(127, 72), 464 "b1010".U -> io.rdcache.resp.bits.data_delayed(127, 80), 465 "b1011".U -> io.rdcache.resp.bits.data_delayed(127, 88), 466 "b1100".U -> io.rdcache.resp.bits.data_delayed(127, 96), 467 "b1101".U -> io.rdcache.resp.bits.data_delayed(127, 104), 468 "b1110".U -> io.rdcache.resp.bits.data_delayed(127, 112), 469 "b1111".U -> io.rdcache.resp.bits.data_delayed(127, 120) 470 )) 471 val pickData = rdataVecHelper(alignedType(1,0), cacheData) 472 val mergedData = mergeDataWithElemIdx( 473 oldData = data(splitPtr.value), 474 newData = Seq(pickData), 475 alignedType = alignedType(1,0), 476 elemIdx = Seq(elemIdxInVd), 477 valids = Seq(true.B) 478 ) 479 when(state === s_latch_and_merge_data && segmentActive){ 480 data(splitPtr.value) := mergedData 481 } 482 /** 483 * split data for store 484 * */ 485 val splitData = genVSData( 486 data = data(splitPtr.value), 487 elemIdx = elemIdxInVd, 488 alignedType = alignedType 489 ) 490 val flowData = genVWdata(splitData, alignedType) // TODO: connect vstd, pass vector data 491 val wmask = genVWmask(latchVaddr, alignedType(1, 0)) & Fill(VLENB, segmentActive) 492 493 /** 494 * rdcache req, write request don't need to query dcache, because we write element to sbuffer 495 */ 496 io.rdcache.req := DontCare 497 io.rdcache.req.valid := state === s_cache_req && FuType.isVLoad(fuType) 498 io.rdcache.req.bits.cmd := MemoryOpConstants.M_XRD 499 io.rdcache.req.bits.vaddr := latchVaddr 500 io.rdcache.req.bits.mask := mask 501 io.rdcache.req.bits.data := flowData 502 io.rdcache.pf_source := LOAD_SOURCE.U 503 io.rdcache.req.bits.id := DontCare 504 io.rdcache.resp.ready := true.B 505 io.rdcache.s1_paddr_dup_lsu := instMicroOp.paddr 506 io.rdcache.s1_paddr_dup_dcache := instMicroOp.paddr 507 io.rdcache.s1_kill := false.B 508 io.rdcache.s1_kill_data_read := false.B 509 io.rdcache.s2_kill := false.B 510 if (env.FPGAPlatform){ 511 io.rdcache.s0_pc := DontCare 512 io.rdcache.s1_pc := DontCare 513 io.rdcache.s2_pc := DontCare 514 }else{ 515 io.rdcache.s0_pc := instMicroOp.uop.pc 516 io.rdcache.s1_pc := instMicroOp.uop.pc 517 io.rdcache.s2_pc := instMicroOp.uop.pc 518 } 519 io.rdcache.replacementUpdated := false.B 520 io.rdcache.is128Req := false.B 521 522 523 /** 524 * write data to sbuffer 525 * */ 526 sbufferOut.bits := DontCare 527 sbufferOut.valid := state === s_send_data && segmentActive 528 sbufferOut.bits.vecValid := state === s_send_data && segmentActive 529 sbufferOut.bits.mask := wmask 530 sbufferOut.bits.data := flowData 531 sbufferOut.bits.vaddr := latchVaddr 532 sbufferOut.bits.cmd := MemoryOpConstants.M_XWR 533 sbufferOut.bits.id := DontCare 534 sbufferOut.bits.addr := instMicroOp.paddr 535 536 NewPipelineConnect( 537 sbufferOut, io.sbuffer, io.sbuffer.fire, 538 false.B, 539 Option(s"VSegmentUnitPipelineConnect") 540 ) 541 542 io.vecDifftestInfo.valid := io.sbuffer.valid 543 io.vecDifftestInfo.bits := uopq(deqPtr.value).uop 544 545 /** 546 * update ptr 547 * */ 548 private val fieldActiveWirteFinish = sbufferOut.fire && segmentActive // writedata finish and is a active segment 549 XSError(sbufferOut.fire && !segmentActive, "Attempt write inactive segment to sbuffer, something wrong!\n") 550 551 private val segmentInactiveFinish = ((state === s_latch_and_merge_data) || (state === s_send_data)) && !segmentActive 552 553 val splitPtrOffset = Mux( 554 isIndexed(instType), 555 Mux(lmul.asSInt < 0.S, 1.U, (1.U << lmul).asUInt), 556 Mux(emul.asSInt < 0.S, 1.U, (1.U << emul).asUInt) 557 ) 558 splitPtrNext := 559 Mux(fieldIdx === maxNfields || !segmentActive, // if segment is active, need to complete this segment, otherwise jump to next segment 560 // segment finish, By shifting 'issueUopFlowNumLog2' to the right to ensure that emul != 1 can correctly generate lateral offset. 561 (deqPtr + ((segmentIdx +& 1.U) >> issueUopFlowNumLog2).asUInt), 562 // next field. 563 (splitPtr + splitPtrOffset) 564 ) 565 566 dontTouch(issueUopFlowNumLog2) 567 dontTouch(issueEmul) 568 dontTouch(splitPtrNext) 569 dontTouch(stridePtr) 570 dontTouch(segmentActive) 571 572 // update splitPtr 573 when(state === s_latch_and_merge_data || (state === s_send_data && (fieldActiveWirteFinish || !segmentActive))){ 574 splitPtr := splitPtrNext 575 }.elsewhen(io.in.fire && !instMicroOpValid){ 576 splitPtr := deqPtr // initial splitPtr 577 } 578 579 // update stridePtr, only use in index 580 val strideOffset = Mux(isIndexed(issueInstType), segmentIdx >> issueMaxIdxInIndexLog2, 0.U) 581 stridePtr := deqPtr + strideOffset 582 583 // update fieldIdx 584 when(io.in.fire && !instMicroOpValid){ // init 585 fieldIdx := 0.U 586 }.elsewhen(state === s_latch_and_merge_data && segmentActive || 587 (state === s_send_data && fieldActiveWirteFinish)){ // only if segment is active 588 589 /* next segment, only if segment complete */ 590 fieldIdx := Mux(fieldIdx === maxNfields, 0.U, fieldIdx + 1.U) 591 }.elsewhen(segmentInactiveFinish){ // segment is inactive, go to next segment 592 fieldIdx := 0.U 593 } 594 //update segmentIdx 595 when(io.in.fire && !instMicroOpValid){ 596 segmentIdx := 0.U 597 }.elsewhen(fieldIdx === maxNfields && (state === s_latch_and_merge_data || (state === s_send_data && fieldActiveWirteFinish)) && 598 segmentIdx =/= maxSegIdx){ // next segment, only if segment is active 599 600 segmentIdx := segmentIdx + 1.U 601 }.elsewhen(segmentInactiveFinish && segmentIdx =/= maxSegIdx){ // if segment is inactive, go to next segment 602 segmentIdx := segmentIdx + 1.U 603 } 604 605 //update segmentOffset 606 /* when segment is active or segment is inactive, increase segmentOffset */ 607 when((fieldIdx === maxNfields && (state === s_latch_and_merge_data || (state === s_send_data && fieldActiveWirteFinish))) || 608 segmentInactiveFinish){ 609 610 segmentOffset := segmentOffset + Mux(isUnitStride(issueInstType), (maxNfields +& 1.U) << issueEew(1, 0), stride(stridePtr.value)) 611 } 612 613 //update deqPtr 614 when((state === s_finish) && !isEmpty(enqPtr, deqPtr)){ 615 deqPtr := deqPtr + 1.U 616 } 617 618 /************************************************************************* 619 * dequeue logic 620 *************************************************************************/ 621 val vdIdxInField = GenUopIdxInField(Mux(isIndexed(instType), issueLmul, issueEmul), uopq(deqPtr.value).uop.vpu.vuopIdx) 622 /*select mask of vd, maybe remove in feature*/ 623 val realEw = Mux(isIndexed(issueInstType), issueSew(1, 0), issueEew(1, 0)) 624 val maskDataVec: Vec[UInt] = VecDataToMaskDataVec(instMicroOp.mask, realEw) 625 val maskUsed = maskDataVec(vdIdxInField) 626 627 when(stateNext === s_idle){ 628 instMicroOpValid := false.B 629 } 630 // writeback to backend 631 val writebackOut = WireInit(io.uopwriteback.bits) 632 val writebackValid = (state === s_finish) && !isEmpty(enqPtr, deqPtr) 633 writebackOut.uop := uopq(deqPtr.value).uop 634 writebackOut.uop.vpu := instMicroOp.uop.vpu 635 writebackOut.uop.exceptionVec := instMicroOp.uop.exceptionVec 636 writebackOut.mask.get := instMicroOp.mask 637 writebackOut.data := data(deqPtr.value) 638 writebackOut.vdIdx.get := vdIdxInField 639 writebackOut.uop.vpu.vl := instMicroOp.vl 640 writebackOut.uop.vpu.vstart := instMicroOp.vstart 641 writebackOut.uop.vpu.vmask := maskUsed 642 writebackOut.uop.vpu.vuopIdx := uopq(deqPtr.value).uop.vpu.vuopIdx 643 writebackOut.debug := DontCare 644 writebackOut.vdIdxInField.get := vdIdxInField 645 writebackOut.uop.robIdx := instMicroOp.uop.robIdx 646 writebackOut.uop.fuOpType := instMicroOp.uop.fuOpType 647 648 io.uopwriteback.valid := RegNext(writebackValid) 649 io.uopwriteback.bits := RegEnable(writebackOut, writebackValid) 650 651 dontTouch(writebackValid) 652 653 //to RS 654 val feedbackOut = WireInit(0.U.asTypeOf(io.feedback.bits)) 655 val feedbackValid = state === s_finish && !isEmpty(enqPtr, deqPtr) 656 feedbackOut.hit := true.B 657 feedbackOut.robIdx := instMicroOp.uop.robIdx 658 feedbackOut.sourceType := DontCare 659 feedbackOut.flushState := DontCare 660 feedbackOut.dataInvalidSqIdx := DontCare 661 feedbackOut.sqIdx := uopq(deqPtr.value).uop.sqIdx 662 feedbackOut.lqIdx := uopq(deqPtr.value).uop.lqIdx 663 664 io.feedback.valid := RegNext(feedbackValid) 665 io.feedback.bits := RegEnable(feedbackOut, feedbackValid) 666 667 dontTouch(feedbackValid) 668 669 // exception 670 io.exceptionInfo := DontCare 671 io.exceptionInfo.bits.robidx := instMicroOp.uop.robIdx 672 io.exceptionInfo.bits.uopidx := uopq(deqPtr.value).uop.vpu.vuopIdx 673 io.exceptionInfo.bits.vstart := instMicroOp.exceptionVstart 674 io.exceptionInfo.bits.vaddr := instMicroOp.exceptionVaddr 675 io.exceptionInfo.bits.gpaddr := instMicroOp.exceptionGpaddr 676 io.exceptionInfo.bits.isForVSnonLeafPTE := instMicroOp.exceptionIsForVSnonLeafPTE 677 io.exceptionInfo.bits.vl := instMicroOp.exceptionVl 678 io.exceptionInfo.valid := (state === s_finish) && instMicroOp.uop.exceptionVec.asUInt.orR && !isEmpty(enqPtr, deqPtr) 679} 680 681