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.frontend 18 19import chipsalliance.rocketchip.config.Parameters 20import chisel3._ 21import chisel3.util._ 22import freechips.rocketchip.rocket.RVCDecoder 23import xiangshan._ 24import xiangshan.cache.mmu._ 25import xiangshan.frontend.icache._ 26import utils._ 27import xiangshan.backend.fu.{PMPReqBundle, PMPRespBundle} 28 29trait HasInstrMMIOConst extends HasXSParameter with HasIFUConst{ 30 def mmioBusWidth = 64 31 def mmioBusBytes = mmioBusWidth / 8 32 def maxInstrLen = 32 33} 34 35trait HasIFUConst extends HasXSParameter{ 36 def addrAlign(addr: UInt, bytes: Int, highest: Int): UInt = Cat(addr(highest-1, log2Ceil(bytes)), 0.U(log2Ceil(bytes).W)) 37 def fetchQueueSize = 2 38 39 def getBasicBlockIdx( pc: UInt, start: UInt ): UInt = { 40 val byteOffset = pc - start 41 (byteOffset - instBytes.U)(log2Ceil(PredictWidth),instOffsetBits) 42 } 43} 44 45class IfuToFtqIO(implicit p:Parameters) extends XSBundle { 46 val pdWb = Valid(new PredecodeWritebackBundle) 47} 48 49class FtqInterface(implicit p: Parameters) extends XSBundle { 50 val fromFtq = Flipped(new FtqToIfuIO) 51 val toFtq = new IfuToFtqIO 52} 53 54class UncacheInterface(implicit p: Parameters) extends XSBundle { 55 val fromUncache = Flipped(DecoupledIO(new InsUncacheResp)) 56 val toUncache = DecoupledIO( new InsUncacheReq ) 57} 58class NewIFUIO(implicit p: Parameters) extends XSBundle { 59 val ftqInter = new FtqInterface 60 val icacheInter = Vec(2, Flipped(new ICacheMainPipeBundle)) 61 val icacheStop = Output(Bool()) 62 val icachePerfInfo = Input(new ICachePerfInfo) 63 val toIbuffer = Decoupled(new FetchToIBuffer) 64 val uncacheInter = new UncacheInterface 65 val frontendTrigger = Flipped(new FrontendTdataDistributeIO) 66 val csrTriggerEnable = Input(Vec(4, Bool())) 67 val rob_commits = Flipped(Vec(CommitWidth, Valid(new RobCommitInfo))) 68 val iTLBInter = new TlbRequestIO 69 val pmp = new ICachePMPBundle 70} 71 72// record the situation in which fallThruAddr falls into 73// the middle of an RVI inst 74class LastHalfInfo(implicit p: Parameters) extends XSBundle { 75 val valid = Bool() 76 val middlePC = UInt(VAddrBits.W) 77 def matchThisBlock(startAddr: UInt) = valid && middlePC === startAddr 78} 79 80class IfuToPreDecode(implicit p: Parameters) extends XSBundle { 81 val data = if(HasCExtension) Vec(PredictWidth + 1, UInt(16.W)) else Vec(PredictWidth, UInt(32.W)) 82 val frontendTrigger = new FrontendTdataDistributeIO 83 val csrTriggerEnable = Vec(4, Bool()) 84 val pc = Vec(PredictWidth, UInt(VAddrBits.W)) 85} 86 87 88class IfuToPredChecker(implicit p: Parameters) extends XSBundle { 89 val ftqOffset = Valid(UInt(log2Ceil(PredictWidth).W)) 90 val jumpOffset = Vec(PredictWidth, UInt(XLEN.W)) 91 val target = UInt(VAddrBits.W) 92 val instrRange = Vec(PredictWidth, Bool()) 93 val instrValid = Vec(PredictWidth, Bool()) 94 val pds = Vec(PredictWidth, new PreDecodeInfo) 95 val pc = Vec(PredictWidth, UInt(VAddrBits.W)) 96} 97 98class NewIFU(implicit p: Parameters) extends XSModule 99 with HasICacheParameters 100 with HasIFUConst 101 with HasPdConst 102 with HasCircularQueuePtrHelper 103 with HasPerfEvents 104{ 105 val io = IO(new NewIFUIO) 106 val (toFtq, fromFtq) = (io.ftqInter.toFtq, io.ftqInter.fromFtq) 107 val (toICache, fromICache) = (VecInit(io.icacheInter.map(_.req)), VecInit(io.icacheInter.map(_.resp))) 108 val (toUncache, fromUncache) = (io.uncacheInter.toUncache , io.uncacheInter.fromUncache) 109 110 def isCrossLineReq(start: UInt, end: UInt): Bool = start(blockOffBits) ^ end(blockOffBits) 111 112 def isLastInCacheline(addr: UInt): Bool = addr(blockOffBits - 1, 1) === 0.U 113 114 class TlbExept(implicit p: Parameters) extends XSBundle{ 115 val pageFault = Bool() 116 val accessFault = Bool() 117 val mmio = Bool() 118 } 119 120 val preDecoder = Module(new PreDecode) 121 val predChecker = Module(new PredChecker) 122 val frontendTrigger = Module(new FrontendTrigger) 123 val (preDecoderIn, preDecoderOut) = (preDecoder.io.in, preDecoder.io.out) 124 val (checkerIn, checkerOutStage1, checkerOutStage2) = (predChecker.io.in, predChecker.io.out.stage1Out,predChecker.io.out.stage2Out) 125 126 io.iTLBInter.req_kill := false.B 127 io.iTLBInter.resp.ready := true.B 128 129 /** 130 ****************************************************************************** 131 * IFU Stage 0 132 * - send cacheline fetch request to ICacheMainPipe 133 ****************************************************************************** 134 */ 135 136 val f0_valid = fromFtq.req.valid 137 val f0_ftq_req = fromFtq.req.bits 138 val f0_doubleLine = fromFtq.req.bits.crossCacheline 139 val f0_vSetIdx = VecInit(get_idx((f0_ftq_req.startAddr)), get_idx(f0_ftq_req.nextlineStart)) 140 val f0_fire = fromFtq.req.fire() 141 142 val f0_flush, f1_flush, f2_flush, f3_flush = WireInit(false.B) 143 val from_bpu_f0_flush, from_bpu_f1_flush, from_bpu_f2_flush, from_bpu_f3_flush = WireInit(false.B) 144 145 from_bpu_f0_flush := fromFtq.flushFromBpu.shouldFlushByStage2(f0_ftq_req.ftqIdx) || 146 fromFtq.flushFromBpu.shouldFlushByStage3(f0_ftq_req.ftqIdx) 147 148 val wb_redirect , mmio_redirect, backend_redirect= WireInit(false.B) 149 val f3_wb_not_flush = WireInit(false.B) 150 151 backend_redirect := fromFtq.redirect.valid 152 f3_flush := backend_redirect || (wb_redirect && !f3_wb_not_flush) 153 f2_flush := backend_redirect || mmio_redirect || wb_redirect 154 f1_flush := f2_flush || from_bpu_f1_flush 155 f0_flush := f1_flush || from_bpu_f0_flush 156 157 val f1_ready, f2_ready, f3_ready = WireInit(false.B) 158 159 fromFtq.req.ready := toICache(0).ready && toICache(1).ready && f1_ready //&& GTimer() > 500.U 160 161 toICache(0).valid := fromFtq.req.valid //&& !f0_flush 162 toICache(0).bits.vaddr := fromFtq.req.bits.startAddr 163 toICache(1).valid := fromFtq.req.valid && f0_doubleLine //&& !f0_flush 164 toICache(1).bits.vaddr := fromFtq.req.bits.nextlineStart//fromFtq.req.bits.startAddr + (PredictWidth * 2).U //TODO: timing critical 165 166 /** <PERF> f0 fetch bubble */ 167 168 XSPerfAccumulate("fetch_bubble_ftq_not_valid", !fromFtq.req.valid && fromFtq.req.ready ) 169 XSPerfAccumulate("fetch_bubble_pipe_stall", f0_valid && toICache(0).ready && toICache(1).ready && !f1_ready ) 170 XSPerfAccumulate("fetch_bubble_icache_0_busy", f0_valid && !toICache(0).ready ) 171 XSPerfAccumulate("fetch_bubble_icache_1_busy", f0_valid && !toICache(1).ready ) 172 XSPerfAccumulate("fetch_flush_backend_redirect", backend_redirect ) 173 XSPerfAccumulate("fetch_flush_wb_redirect", wb_redirect ) 174 XSPerfAccumulate("fetch_flush_bpu_f1_flush", from_bpu_f1_flush ) 175 XSPerfAccumulate("fetch_flush_bpu_f0_flush", from_bpu_f0_flush ) 176 177 178 /** 179 ****************************************************************************** 180 * IFU Stage 1 181 * - calculate pc/half_pc/cut_ptr for every instruction 182 ****************************************************************************** 183 */ 184 185 val f1_valid = RegInit(false.B) 186 val f1_ftq_req = RegEnable(f0_ftq_req, f0_fire) 187 // val f1_situation = RegEnable(f0_situation, f0_fire) 188 val f1_doubleLine = RegEnable(f0_doubleLine, f0_fire) 189 val f1_vSetIdx = RegEnable(f0_vSetIdx, f0_fire) 190 val f1_fire = f1_valid && f2_ready 191 192 f1_ready := f1_fire || !f1_valid 193 194 from_bpu_f1_flush := fromFtq.flushFromBpu.shouldFlushByStage3(f1_ftq_req.ftqIdx) && f1_valid 195 // from_bpu_f1_flush := false.B 196 197 when(f1_flush) {f1_valid := false.B} 198 .elsewhen(f0_fire && !f0_flush) {f1_valid := true.B} 199 .elsewhen(f1_fire) {f1_valid := false.B} 200 201 val f1_pc = VecInit((0 until PredictWidth).map(i => f1_ftq_req.startAddr + (i * 2).U)) 202 val f1_half_snpc = VecInit((0 until PredictWidth).map(i => f1_ftq_req.startAddr + ((i+2) * 2).U)) 203 val f1_cut_ptr = if(HasCExtension) VecInit((0 until PredictWidth + 1).map(i => Cat(0.U(1.W), f1_ftq_req.startAddr(blockOffBits-1, 1)) + i.U )) 204 else VecInit((0 until PredictWidth).map(i => Cat(0.U(1.W), f1_ftq_req.startAddr(blockOffBits-1, 2)) + i.U )) 205 206 /** 207 ****************************************************************************** 208 * IFU Stage 2 209 * - icache response data (latched for pipeline stop) 210 * - generate exceprion bits for every instruciton (page fault/access fault/mmio) 211 * - generate predicted instruction range (1 means this instruciton is in this fetch packet) 212 * - cut data from cachlines to packet instruction code 213 * - instruction predecode and RVC expand 214 ****************************************************************************** 215 */ 216 217 val icacheRespAllValid = WireInit(false.B) 218 219 val f2_valid = RegInit(false.B) 220 val f2_ftq_req = RegEnable(f1_ftq_req, f1_fire) 221 // val f2_situation = RegEnable(f1_situation, f1_fire) 222 val f2_doubleLine = RegEnable(f1_doubleLine, f1_fire) 223 val f2_vSetIdx = RegEnable(f1_vSetIdx, f1_fire) 224 val f2_fire = f2_valid && f3_ready && icacheRespAllValid 225 226 f2_ready := f2_fire || !f2_valid 227 //TODO: addr compare may be timing critical 228 val f2_icache_all_resp_wire = fromICache(0).valid && (fromICache(0).bits.vaddr === f2_ftq_req.startAddr) && ((fromICache(1).valid && (fromICache(1).bits.vaddr === f2_ftq_req.nextlineStart)) || !f2_doubleLine) 229 val f2_icache_all_resp_reg = RegInit(false.B) 230 231 icacheRespAllValid := f2_icache_all_resp_reg || f2_icache_all_resp_wire 232 233 io.icacheStop := !f3_ready 234 235 when(f2_flush) {f2_icache_all_resp_reg := false.B} 236 .elsewhen(f2_valid && f2_icache_all_resp_wire && !f3_ready) {f2_icache_all_resp_reg := true.B} 237 .elsewhen(f2_fire && f2_icache_all_resp_reg) {f2_icache_all_resp_reg := false.B} 238 239 when(f2_flush) {f2_valid := false.B} 240 .elsewhen(f1_fire && !f1_flush) {f2_valid := true.B } 241 .elsewhen(f2_fire) {f2_valid := false.B} 242 243 // val f2_cache_response_data = ResultHoldBypass(valid = f2_icache_all_resp_wire, data = VecInit(fromICache.map(_.bits.readData))) 244 val f2_cache_response_data = VecInit(fromICache.map(_.bits.readData)) 245 246 247 val f2_except_pf = VecInit((0 until PortNumber).map(i => fromICache(i).bits.tlbExcp.pageFault)) 248 val f2_except_af = VecInit((0 until PortNumber).map(i => fromICache(i).bits.tlbExcp.accessFault)) 249 val f2_mmio = fromICache(0).bits.tlbExcp.mmio && !fromICache(0).bits.tlbExcp.accessFault && 250 !fromICache(0).bits.tlbExcp.pageFault 251 252 val f2_pc = RegEnable(f1_pc, f1_fire) 253 val f2_half_snpc = RegEnable(f1_half_snpc, f1_fire) 254 val f2_cut_ptr = RegEnable(f1_cut_ptr, f1_fire) 255 256 val f2_resend_vaddr = RegEnable(f1_ftq_req.startAddr + 2.U, f1_fire) 257 258 def isNextLine(pc: UInt, startAddr: UInt) = { 259 startAddr(blockOffBits) ^ pc(blockOffBits) 260 } 261 262 def isLastInLine(pc: UInt) = { 263 pc(blockOffBits - 1, 0) === "b111110".U 264 } 265 266 val f2_foldpc = VecInit(f2_pc.map(i => XORFold(i(VAddrBits-1,1), MemPredPCWidth))) 267 val f2_jump_range = Fill(PredictWidth, !f2_ftq_req.ftqOffset.valid) | Fill(PredictWidth, 1.U(1.W)) >> ~f2_ftq_req.ftqOffset.bits 268 val f2_ftr_range = Fill(PredictWidth, f2_ftq_req.ftqOffset.valid) | Fill(PredictWidth, 1.U(1.W)) >> ~getBasicBlockIdx(f2_ftq_req.nextStartAddr, f2_ftq_req.startAddr) 269 val f2_instr_range = f2_jump_range & f2_ftr_range 270 val f2_pf_vec = VecInit((0 until PredictWidth).map(i => (!isNextLine(f2_pc(i), f2_ftq_req.startAddr) && f2_except_pf(0) || isNextLine(f2_pc(i), f2_ftq_req.startAddr) && f2_doubleLine && f2_except_pf(1)))) 271 val f2_af_vec = VecInit((0 until PredictWidth).map(i => (!isNextLine(f2_pc(i), f2_ftq_req.startAddr) && f2_except_af(0) || isNextLine(f2_pc(i), f2_ftq_req.startAddr) && f2_doubleLine && f2_except_af(1)))) 272 273 val f2_paddrs = VecInit((0 until PortNumber).map(i => fromICache(i).bits.paddr)) 274 val f2_perf_info = io.icachePerfInfo 275 276 def cut(cacheline: UInt, cutPtr: Vec[UInt]) : Vec[UInt] ={ 277 require(HasCExtension) 278 // if(HasCExtension){ 279 val partCacheline = cacheline((blockBytes * 8 * 2 * 3) / 4 - 1, 0) 280 val result = Wire(Vec(PredictWidth + 1, UInt(16.W))) 281 val dataVec = cacheline.asTypeOf(Vec(blockBytes * 3 /4, UInt(16.W))) //47 16-bit data vector 282 (0 until PredictWidth + 1).foreach( i => 283 result(i) := dataVec(cutPtr(i)) //the max ptr is 3*blockBytes/4-1 284 ) 285 result 286 // } else { 287 // val result = Wire(Vec(PredictWidth, UInt(32.W)) ) 288 // val dataVec = cacheline.asTypeOf(Vec(blockBytes * 2/ 4, UInt(32.W))) 289 // (0 until PredictWidth).foreach( i => 290 // result(i) := dataVec(cutPtr(i)) 291 // ) 292 // result 293 // } 294 } 295 296 val f2_datas = VecInit((0 until PortNumber).map(i => f2_cache_response_data(i))) 297 val f2_cut_data = cut( Cat(f2_datas.map(cacheline => cacheline.asUInt ).reverse).asUInt, f2_cut_ptr ) 298 299 /** predecode (include RVC expander) */ 300 preDecoderIn.data := f2_cut_data 301 preDecoderIn.frontendTrigger := io.frontendTrigger 302 preDecoderIn.csrTriggerEnable := io.csrTriggerEnable 303 preDecoderIn.pc := f2_pc 304 305 val f2_expd_instr = preDecoderOut.expInstr 306 val f2_pd = preDecoderOut.pd 307 val f2_jump_offset = preDecoderOut.jumpOffset 308 val f2_hasHalfValid = preDecoderOut.hasHalfValid 309 val f2_crossPageFault = VecInit((0 until PredictWidth).map(i => isLastInLine(f2_pc(i)) && !f2_except_pf(0) && f2_doubleLine && f2_except_pf(1) && !f2_pd(i).isRVC )) 310 311 val predecodeOutValid = WireInit(false.B) 312 313 XSPerfAccumulate("fetch_bubble_icache_not_resp", f2_valid && !icacheRespAllValid ) 314 315 316 /** 317 ****************************************************************************** 318 * IFU Stage 3 319 * - handle MMIO instruciton 320 * -send request to Uncache fetch Unit 321 * -every packet include 1 MMIO instruction 322 * -MMIO instructions will stop fetch pipeline until commiting from RoB 323 * -flush to snpc (send ifu_redirect to Ftq) 324 * - Ibuffer enqueue 325 * - check predict result in Frontend (jalFault/retFault/notCFIFault/invalidTakenFault/targetFault) 326 * - handle last half RVI instruction 327 ****************************************************************************** 328 */ 329 330 val f3_valid = RegInit(false.B) 331 val f3_ftq_req = RegEnable(f2_ftq_req, f2_fire) 332 // val f3_situation = RegEnable(f2_situation, f2_fire) 333 val f3_doubleLine = RegEnable(f2_doubleLine, f2_fire) 334 val f3_fire = io.toIbuffer.fire() 335 336 f3_ready := f3_fire || !f3_valid 337 338 val f3_cut_data = RegEnable(f2_cut_data, f2_fire) 339 340 val f3_except_pf = RegEnable(f2_except_pf, f2_fire) 341 val f3_except_af = RegEnable(f2_except_af, f2_fire) 342 val f3_mmio = RegEnable(f2_mmio , f2_fire) 343 344 val f3_expd_instr = RegEnable(f2_expd_instr, f2_fire) 345 val f3_pd = RegEnable(f2_pd, f2_fire) 346 val f3_jump_offset = RegEnable(f2_jump_offset, f2_fire) 347 val f3_af_vec = RegEnable(f2_af_vec, f2_fire) 348 val f3_pf_vec = RegEnable(f2_pf_vec , f2_fire) 349 val f3_pc = RegEnable(f2_pc, f2_fire) 350 val f3_half_snpc = RegEnable(f2_half_snpc, f2_fire) 351 val f3_instr_range = RegEnable(f2_instr_range, f2_fire) 352 val f3_foldpc = RegEnable(f2_foldpc, f2_fire) 353 val f3_crossPageFault = RegEnable(f2_crossPageFault, f2_fire) 354 val f3_hasHalfValid = RegEnable(f2_hasHalfValid, f2_fire) 355 val f3_except = VecInit((0 until 2).map{i => f3_except_pf(i) || f3_except_af(i)}) 356 val f3_has_except = f3_valid && (f3_except_af.reduce(_||_) || f3_except_pf.reduce(_||_)) 357 val f3_pAddrs = RegEnable(f2_paddrs, f2_fire) 358 val f3_resend_vaddr = RegEnable(f2_resend_vaddr, f2_fire) 359 360 when(f3_valid && !f3_ftq_req.ftqOffset.valid){ 361 assert(f3_ftq_req.startAddr + 32.U >= f3_ftq_req.nextStartAddr , "More tha 32 Bytes fetch is not allowed!") 362 } 363 364 /*** MMIO State Machine***/ 365 val f3_mmio_data = Reg(Vec(2, UInt(16.W))) 366 val mmio_is_RVC = RegInit(false.B) 367 val mmio_resend_addr =RegInit(0.U(PAddrBits.W)) 368 val mmio_resend_af = RegInit(false.B) 369 val mmio_resend_pf = RegInit(false.B) 370 371 372 val m_idle :: m_sendReq :: m_waitResp :: m_sendTLB :: m_tlbResp :: m_sendPMP :: m_resendReq :: m_waitResendResp :: m_waitCommit :: m_commited :: Nil = Enum(10) 373 val mmio_state = RegInit(m_idle) 374 375 val f3_req_is_mmio = f3_mmio && f3_valid 376 val mmio_commit = VecInit(io.rob_commits.map{commit => commit.valid && commit.bits.ftqIdx === f3_ftq_req.ftqIdx && commit.bits.ftqOffset === 0.U}).asUInt.orR 377 val f3_mmio_req_commit = f3_req_is_mmio && mmio_state === m_commited 378 379 val f3_mmio_to_commit = f3_req_is_mmio && mmio_state === m_waitCommit 380 val f3_mmio_to_commit_next = RegNext(f3_mmio_to_commit) 381 val f3_mmio_can_go = f3_mmio_to_commit && !f3_mmio_to_commit_next 382 383 val fromFtqRedirectReg = RegNext(fromFtq.redirect) 384 val f3_ftq_flush_self = fromFtqRedirectReg.valid && RedirectLevel.flushItself(fromFtqRedirectReg.bits.level) 385 val f3_ftq_flush_by_older = fromFtqRedirectReg.valid && isBefore(fromFtqRedirectReg.bits.ftqIdx, f3_ftq_req.ftqIdx) 386 387 val f3_need_not_flush = f3_req_is_mmio && fromFtqRedirectReg.valid && !f3_ftq_flush_self && !f3_ftq_flush_by_older 388 389 when(f3_flush && !f3_need_not_flush) {f3_valid := false.B} 390 .elsewhen(f2_fire && !f2_flush ) {f3_valid := true.B } 391 .elsewhen(io.toIbuffer.fire() && !f3_req_is_mmio) {f3_valid := false.B} 392 .elsewhen{f3_req_is_mmio && f3_mmio_req_commit} {f3_valid := false.B} 393 394 val f3_mmio_use_seq_pc = RegInit(false.B) 395 396 val (redirect_ftqIdx, redirect_ftqOffset) = (fromFtqRedirectReg.bits.ftqIdx,fromFtqRedirectReg.bits.ftqOffset) 397 val redirect_mmio_req = fromFtqRedirectReg.valid && redirect_ftqIdx === f3_ftq_req.ftqIdx && redirect_ftqOffset === 0.U 398 399 when(RegNext(f2_fire && !f2_flush) && f3_req_is_mmio) { f3_mmio_use_seq_pc := true.B } 400 .elsewhen(redirect_mmio_req) { f3_mmio_use_seq_pc := false.B } 401 402 f3_ready := Mux(f3_req_is_mmio, io.toIbuffer.ready && f3_mmio_req_commit || !f3_valid , io.toIbuffer.ready || !f3_valid) 403 404 // when(fromUncache.fire()) {f3_mmio_data := fromUncache.bits.data} 405 406 407 switch(mmio_state){ 408 is(m_idle){ 409 when(f3_req_is_mmio){ 410 mmio_state := m_sendReq 411 } 412 } 413 414 is(m_sendReq){ 415 mmio_state := Mux(toUncache.fire(), m_waitResp, m_sendReq ) 416 } 417 418 is(m_waitResp){ 419 when(fromUncache.fire()){ 420 val isRVC = fromUncache.bits.data(1,0) =/= 3.U 421 val needResend = !isRVC && f3_pAddrs(0)(2,1) === 3.U 422 mmio_state := Mux(needResend, m_sendTLB , m_waitCommit) 423 424 mmio_is_RVC := isRVC 425 f3_mmio_data(0) := fromUncache.bits.data(15,0) 426 f3_mmio_data(1) := fromUncache.bits.data(31,16) 427 } 428 } 429 430 is(m_sendTLB){ 431 when( io.iTLBInter.req.valid && !io.iTLBInter.resp.bits.miss ){ 432 mmio_state := m_tlbResp 433 } 434 } 435 436 is(m_tlbResp){ 437 val tlbExept = io.iTLBInter.resp.bits.excp(0).pf.instr || 438 io.iTLBInter.resp.bits.excp(0).af.instr 439 mmio_state := Mux(tlbExept,m_waitCommit,m_sendPMP) 440 mmio_resend_addr := io.iTLBInter.resp.bits.paddr(0) 441 mmio_resend_af := mmio_resend_af || io.iTLBInter.resp.bits.excp(0).af.instr 442 mmio_resend_pf := mmio_resend_pf || io.iTLBInter.resp.bits.excp(0).pf.instr 443 } 444 445 is(m_sendPMP){ 446 val pmpExcpAF = io.pmp.resp.instr || !io.pmp.resp.mmio 447 mmio_state := Mux(pmpExcpAF, m_waitCommit , m_resendReq) 448 mmio_resend_af := pmpExcpAF 449 } 450 451 is(m_resendReq){ 452 mmio_state := Mux(toUncache.fire(), m_waitResendResp, m_resendReq ) 453 } 454 455 is(m_waitResendResp){ 456 when(fromUncache.fire()){ 457 mmio_state := m_waitCommit 458 f3_mmio_data(1) := fromUncache.bits.data(15,0) 459 } 460 } 461 462 is(m_waitCommit){ 463 when(mmio_commit){ 464 mmio_state := m_commited 465 } 466 } 467 468 //normal mmio instruction 469 is(m_commited){ 470 mmio_state := m_idle 471 mmio_is_RVC := false.B 472 mmio_resend_addr := 0.U 473 } 474 } 475 476 //exception or flush by older branch prediction 477 when(f3_ftq_flush_self || f3_ftq_flush_by_older) { 478 mmio_state := m_idle 479 mmio_is_RVC := false.B 480 mmio_resend_addr := 0.U 481 mmio_resend_af := false.B 482 f3_mmio_data.map(_ := 0.U) 483 } 484 485 toUncache.valid := ((mmio_state === m_sendReq) || (mmio_state === m_resendReq)) && f3_req_is_mmio 486 toUncache.bits.addr := Mux((mmio_state === m_resendReq), mmio_resend_addr, f3_pAddrs(0)) 487 fromUncache.ready := true.B 488 489 io.iTLBInter.req.valid := (mmio_state === m_sendTLB) && f3_req_is_mmio 490 io.iTLBInter.req.bits.size := 3.U 491 io.iTLBInter.req.bits.vaddr := f3_resend_vaddr 492 io.iTLBInter.req.bits.debug.pc := f3_resend_vaddr 493 494 io.iTLBInter.req.bits.kill := false.B // IFU use itlb for mmio, doesn't need sync, set it to false 495 io.iTLBInter.req.bits.cmd := TlbCmd.exec 496 io.iTLBInter.req.bits.debug.robIdx := DontCare 497 io.iTLBInter.req.bits.debug.isFirstIssue := DontCare 498 499 io.pmp.req.valid := (mmio_state === m_sendPMP) && f3_req_is_mmio 500 io.pmp.req.bits.addr := mmio_resend_addr 501 io.pmp.req.bits.size := 3.U 502 io.pmp.req.bits.cmd := TlbCmd.exec 503 504 val f3_lastHalf = RegInit(0.U.asTypeOf(new LastHalfInfo)) 505 506 val f3_predecode_range = VecInit(preDecoderOut.pd.map(inst => inst.valid)).asUInt 507 val f3_mmio_range = VecInit((0 until PredictWidth).map(i => if(i ==0) true.B else false.B)) 508 val f3_instr_valid = Wire(Vec(PredictWidth, Bool())) 509 510 /*** prediction result check ***/ 511 checkerIn.ftqOffset := f3_ftq_req.ftqOffset 512 checkerIn.jumpOffset := f3_jump_offset 513 checkerIn.target := f3_ftq_req.nextStartAddr 514 checkerIn.instrRange := f3_instr_range.asTypeOf(Vec(PredictWidth, Bool())) 515 checkerIn.instrValid := f3_instr_valid.asTypeOf(Vec(PredictWidth, Bool())) 516 checkerIn.pds := f3_pd 517 checkerIn.pc := f3_pc 518 519 /*** handle half RVI in the last 2 Bytes ***/ 520 521 def hasLastHalf(idx: UInt) = { 522 //!f3_pd(idx).isRVC && checkerOutStage1.fixedRange(idx) && f3_instr_valid(idx) && !checkerOutStage1.fixedTaken(idx) && !checkerOutStage2.fixedMissPred(idx) && ! f3_req_is_mmio 523 !f3_pd(idx).isRVC && checkerOutStage1.fixedRange(idx) && f3_instr_valid(idx) && !checkerOutStage1.fixedTaken(idx) && ! f3_req_is_mmio 524 } 525 526 val f3_last_validIdx = ~ParallelPriorityEncoder(checkerOutStage1.fixedRange.reverse) 527 528 val f3_hasLastHalf = hasLastHalf((PredictWidth - 1).U) 529 val f3_false_lastHalf = hasLastHalf(f3_last_validIdx) 530 val f3_false_snpc = f3_half_snpc(f3_last_validIdx) 531 532 val f3_lastHalf_mask = VecInit((0 until PredictWidth).map( i => if(i ==0) false.B else true.B )).asUInt() 533 534 when (f3_flush) { 535 f3_lastHalf.valid := false.B 536 }.elsewhen (f3_fire) { 537 f3_lastHalf.valid := f3_hasLastHalf 538 f3_lastHalf.middlePC := f3_ftq_req.nextStartAddr 539 } 540 541 f3_instr_valid := Mux(f3_lastHalf.valid,f3_hasHalfValid ,VecInit(f3_pd.map(inst => inst.valid))) 542 543 /*** frontend Trigger ***/ 544 frontendTrigger.io.pds := f3_pd 545 frontendTrigger.io.pc := f3_pc 546 frontendTrigger.io.data := f3_cut_data 547 548 frontendTrigger.io.frontendTrigger := io.frontendTrigger 549 frontendTrigger.io.csrTriggerEnable := io.csrTriggerEnable 550 551 val f3_triggered = frontendTrigger.io.triggered 552 553 /*** send to Ibuffer ***/ 554 555 io.toIbuffer.valid := f3_valid && (!f3_req_is_mmio || f3_mmio_can_go) && !f3_flush 556 io.toIbuffer.bits.instrs := f3_expd_instr 557 io.toIbuffer.bits.valid := f3_instr_valid.asUInt 558 io.toIbuffer.bits.enqEnable := checkerOutStage1.fixedRange.asUInt & f3_instr_valid.asUInt 559 io.toIbuffer.bits.pd := f3_pd 560 io.toIbuffer.bits.ftqPtr := f3_ftq_req.ftqIdx 561 io.toIbuffer.bits.pc := f3_pc 562 io.toIbuffer.bits.ftqOffset.zipWithIndex.map{case(a, i) => a.bits := i.U; a.valid := checkerOutStage1.fixedTaken(i) && !f3_req_is_mmio} 563 io.toIbuffer.bits.foldpc := f3_foldpc 564 io.toIbuffer.bits.ipf := VecInit(f3_pf_vec.zip(f3_crossPageFault).map{case (pf, crossPF) => pf || crossPF}) 565 io.toIbuffer.bits.acf := f3_af_vec 566 io.toIbuffer.bits.crossPageIPFFix := f3_crossPageFault 567 io.toIbuffer.bits.triggered := f3_triggered 568 569 when(f3_lastHalf.valid){ 570 io.toIbuffer.bits.enqEnable := checkerOutStage1.fixedRange.asUInt & f3_instr_valid.asUInt & f3_lastHalf_mask 571 io.toIbuffer.bits.valid := f3_lastHalf_mask & f3_instr_valid.asUInt 572 } 573 574 /** external predecode for MMIO instruction */ 575 when(f3_req_is_mmio){ 576 val inst = Cat(f3_mmio_data(1), f3_mmio_data(0)) 577 val currentIsRVC = isRVC(inst) 578 579 val brType::isCall::isRet::Nil = brInfo(inst) 580 val jalOffset = jal_offset(inst, currentIsRVC) 581 val brOffset = br_offset(inst, currentIsRVC) 582 583 io.toIbuffer.bits.instrs (0) := new RVCDecoder(inst, XLEN).decode.bits 584 585 io.toIbuffer.bits.pd(0).valid := true.B 586 io.toIbuffer.bits.pd(0).isRVC := currentIsRVC 587 io.toIbuffer.bits.pd(0).brType := brType 588 io.toIbuffer.bits.pd(0).isCall := isCall 589 io.toIbuffer.bits.pd(0).isRet := isRet 590 591 io.toIbuffer.bits.acf(0) := mmio_resend_af 592 io.toIbuffer.bits.ipf(0) := mmio_resend_pf 593 io.toIbuffer.bits.crossPageIPFFix(0) := mmio_resend_pf 594 595 io.toIbuffer.bits.enqEnable := f3_mmio_range.asUInt 596 } 597 598 599 //Write back to Ftq 600 val f3_cache_fetch = f3_valid && !(f2_fire && !f2_flush) 601 val finishFetchMaskReg = RegNext(f3_cache_fetch) 602 603 val mmioFlushWb = Wire(Valid(new PredecodeWritebackBundle)) 604 val f3_mmio_missOffset = Wire(ValidUndirectioned(UInt(log2Ceil(PredictWidth).W))) 605 f3_mmio_missOffset.valid := f3_req_is_mmio 606 f3_mmio_missOffset.bits := 0.U 607 608 mmioFlushWb.valid := (f3_req_is_mmio && mmio_state === m_waitCommit && RegNext(fromUncache.fire()) && f3_mmio_use_seq_pc) 609 mmioFlushWb.bits.pc := f3_pc 610 mmioFlushWb.bits.pd := f3_pd 611 mmioFlushWb.bits.pd.zipWithIndex.map{case(instr,i) => instr.valid := f3_mmio_range(i)} 612 mmioFlushWb.bits.ftqIdx := f3_ftq_req.ftqIdx 613 mmioFlushWb.bits.ftqOffset := f3_ftq_req.ftqOffset.bits 614 mmioFlushWb.bits.misOffset := f3_mmio_missOffset 615 mmioFlushWb.bits.cfiOffset := DontCare 616 mmioFlushWb.bits.target := Mux(mmio_is_RVC, f3_ftq_req.startAddr + 2.U , f3_ftq_req.startAddr + 4.U) 617 mmioFlushWb.bits.jalTarget := DontCare 618 mmioFlushWb.bits.instrRange := f3_mmio_range 619 620 mmio_redirect := (f3_req_is_mmio && mmio_state === m_waitCommit && RegNext(fromUncache.fire()) && f3_mmio_use_seq_pc) 621 622 XSPerfAccumulate("fetch_bubble_ibuffer_not_ready", io.toIbuffer.valid && !io.toIbuffer.ready ) 623 624 625 /** 626 ****************************************************************************** 627 * IFU Write Back Stage 628 * - write back predecode information to Ftq to update 629 * - redirect if found fault prediction 630 * - redirect if has false hit last half (last PC is not start + 32 Bytes, but in the midle of an notCFI RVI instruction) 631 ****************************************************************************** 632 */ 633 634 val wb_valid = RegNext(RegNext(f2_fire && !f2_flush) && !f3_req_is_mmio && !f3_flush) 635 val wb_ftq_req = RegNext(f3_ftq_req) 636 637 val wb_check_result_stage1 = RegNext(checkerOutStage1) 638 val wb_check_result_stage2 = checkerOutStage2 639 val wb_instr_range = RegNext(io.toIbuffer.bits.enqEnable) 640 val wb_pc = RegNext(f3_pc) 641 val wb_pd = RegNext(f3_pd) 642 val wb_instr_valid = RegNext(f3_instr_valid) 643 644 /* false hit lastHalf */ 645 val wb_lastIdx = RegNext(f3_last_validIdx) 646 val wb_false_lastHalf = RegNext(f3_false_lastHalf) && wb_lastIdx =/= (PredictWidth - 1).U 647 val wb_false_target = RegNext(f3_false_snpc) 648 649 val wb_half_flush = wb_false_lastHalf 650 val wb_half_target = wb_false_target 651 652 /* false oversize */ 653 val lastIsRVC = wb_instr_range.asTypeOf(Vec(PredictWidth,Bool())).last && wb_pd.last.isRVC 654 val lastIsRVI = wb_instr_range.asTypeOf(Vec(PredictWidth,Bool()))(PredictWidth - 2) && !wb_pd(PredictWidth - 2).isRVC 655 val lastTaken = wb_check_result_stage1.fixedTaken.last 656 657 f3_wb_not_flush := wb_ftq_req.ftqIdx === f3_ftq_req.ftqIdx && f3_valid && wb_valid 658 659 val checkFlushWb = Wire(Valid(new PredecodeWritebackBundle)) 660 checkFlushWb.valid := wb_valid 661 checkFlushWb.bits.pc := wb_pc 662 checkFlushWb.bits.pd := wb_pd 663 checkFlushWb.bits.pd.zipWithIndex.map{case(instr,i) => instr.valid := wb_instr_valid(i)} 664 checkFlushWb.bits.ftqIdx := wb_ftq_req.ftqIdx 665 checkFlushWb.bits.ftqOffset := wb_ftq_req.ftqOffset.bits 666 checkFlushWb.bits.misOffset.valid := ParallelOR(wb_check_result_stage2.fixedMissPred) || wb_half_flush 667 checkFlushWb.bits.misOffset.bits := Mux(wb_half_flush, wb_lastIdx, ParallelPriorityEncoder(wb_check_result_stage2.fixedMissPred)) 668 checkFlushWb.bits.cfiOffset.valid := ParallelOR(wb_check_result_stage1.fixedTaken) 669 checkFlushWb.bits.cfiOffset.bits := ParallelPriorityEncoder(wb_check_result_stage1.fixedTaken) 670 checkFlushWb.bits.target := Mux(wb_half_flush, wb_half_target, wb_check_result_stage2.fixedTarget(ParallelPriorityEncoder(wb_check_result_stage2.fixedMissPred))) 671 checkFlushWb.bits.jalTarget := wb_check_result_stage2.fixedTarget(ParallelPriorityEncoder(VecInit(wb_pd.zip(wb_instr_valid).map{case (pd, v) => v && pd.isJal }))) 672 checkFlushWb.bits.instrRange := wb_instr_range.asTypeOf(Vec(PredictWidth, Bool())) 673 674 toFtq.pdWb := Mux(wb_valid, checkFlushWb, mmioFlushWb) 675 676 wb_redirect := checkFlushWb.bits.misOffset.valid && wb_valid 677 678 /*write back flush type*/ 679 val checkFaultType = wb_check_result_stage2.faultType 680 val checkJalFault = wb_valid && checkFaultType.map(_.isjalFault).reduce(_||_) 681 val checkRetFault = wb_valid && checkFaultType.map(_.isRetFault).reduce(_||_) 682 val checkTargetFault = wb_valid && checkFaultType.map(_.istargetFault).reduce(_||_) 683 val checkNotCFIFault = wb_valid && checkFaultType.map(_.notCFIFault).reduce(_||_) 684 val checkInvalidTaken = wb_valid && checkFaultType.map(_.invalidTakenFault).reduce(_||_) 685 686 687 XSPerfAccumulate("predecode_flush_jalFault", checkJalFault ) 688 XSPerfAccumulate("predecode_flush_retFault", checkRetFault ) 689 XSPerfAccumulate("predecode_flush_targetFault", checkTargetFault ) 690 XSPerfAccumulate("predecode_flush_notCFIFault", checkNotCFIFault ) 691 XSPerfAccumulate("predecode_flush_incalidTakenFault", checkInvalidTaken ) 692 693 when(checkRetFault){ 694 XSDebug("startAddr:%x nextstartAddr:%x taken:%d takenIdx:%d\n", 695 wb_ftq_req.startAddr, wb_ftq_req.nextStartAddr, wb_ftq_req.ftqOffset.valid, wb_ftq_req.ftqOffset.bits) 696 } 697 698 /** performance counter */ 699 val f3_perf_info = RegEnable(f2_perf_info, f2_fire) 700 val f3_req_0 = io.toIbuffer.fire() 701 val f3_req_1 = io.toIbuffer.fire() && f3_doubleLine 702 val f3_hit_0 = io.toIbuffer.fire() && f3_perf_info.bank_hit(0) 703 val f3_hit_1 = io.toIbuffer.fire() && f3_doubleLine & f3_perf_info.bank_hit(1) 704 val f3_hit = f3_perf_info.hit 705 val perfEvents = Seq( 706 ("frontendFlush ", wb_redirect ), 707 ("ifu_req ", io.toIbuffer.fire() ), 708 ("ifu_miss ", io.toIbuffer.fire() && !f3_perf_info.hit ), 709 ("ifu_req_cacheline_0 ", f3_req_0 ), 710 ("ifu_req_cacheline_1 ", f3_req_1 ), 711 ("ifu_req_cacheline_0_hit ", f3_hit_1 ), 712 ("ifu_req_cacheline_1_hit ", f3_hit_1 ), 713 ("only_0_hit ", f3_perf_info.only_0_hit && io.toIbuffer.fire() ), 714 ("only_0_miss ", f3_perf_info.only_0_miss && io.toIbuffer.fire() ), 715 ("hit_0_hit_1 ", f3_perf_info.hit_0_hit_1 && io.toIbuffer.fire() ), 716 ("hit_0_miss_1 ", f3_perf_info.hit_0_miss_1 && io.toIbuffer.fire() ), 717 ("miss_0_hit_1 ", f3_perf_info.miss_0_hit_1 && io.toIbuffer.fire() ), 718 ("miss_0_miss_1 ", f3_perf_info.miss_0_miss_1 && io.toIbuffer.fire() ), 719 ) 720 generatePerfEvent() 721 722 XSPerfAccumulate("ifu_req", io.toIbuffer.fire() ) 723 XSPerfAccumulate("ifu_miss", io.toIbuffer.fire() && !f3_hit ) 724 XSPerfAccumulate("ifu_req_cacheline_0", f3_req_0 ) 725 XSPerfAccumulate("ifu_req_cacheline_1", f3_req_1 ) 726 XSPerfAccumulate("ifu_req_cacheline_0_hit", f3_hit_0 ) 727 XSPerfAccumulate("ifu_req_cacheline_1_hit", f3_hit_1 ) 728 XSPerfAccumulate("frontendFlush", wb_redirect ) 729 XSPerfAccumulate("only_0_hit", f3_perf_info.only_0_hit && io.toIbuffer.fire() ) 730 XSPerfAccumulate("only_0_miss", f3_perf_info.only_0_miss && io.toIbuffer.fire() ) 731 XSPerfAccumulate("hit_0_hit_1", f3_perf_info.hit_0_hit_1 && io.toIbuffer.fire() ) 732 XSPerfAccumulate("hit_0_miss_1", f3_perf_info.hit_0_miss_1 && io.toIbuffer.fire() ) 733 XSPerfAccumulate("miss_0_hit_1", f3_perf_info.miss_0_hit_1 && io.toIbuffer.fire() ) 734 XSPerfAccumulate("miss_0_miss_1", f3_perf_info.miss_0_miss_1 && io.toIbuffer.fire() ) 735 XSPerfAccumulate("hit_0_except_1", f3_perf_info.hit_0_except_1 && io.toIbuffer.fire() ) 736 XSPerfAccumulate("miss_0_except_1", f3_perf_info.miss_0_except_1 && io.toIbuffer.fire() ) 737 XSPerfAccumulate("except_0", f3_perf_info.except_0 && io.toIbuffer.fire() ) 738} 739 740