fetch_impl.hh revision 7720:65d338a8dba4
1/* 2 * Copyright (c) 2004-2006 The Regents of The University of Michigan 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions are 7 * met: redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer; 9 * redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution; 12 * neither the name of the copyright holders nor the names of its 13 * contributors may be used to endorse or promote products derived from 14 * this software without specific prior written permission. 15 * 16 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 17 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 18 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 19 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 20 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 21 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 22 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 26 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 27 * 28 * Authors: Kevin Lim 29 * Korey Sewell 30 */ 31 32#include <algorithm> 33#include <cstring> 34 35#include "arch/isa_traits.hh" 36#include "arch/utility.hh" 37#include "base/types.hh" 38#include "config/the_isa.hh" 39#include "config/use_checker.hh" 40#include "cpu/checker/cpu.hh" 41#include "cpu/exetrace.hh" 42#include "cpu/o3/fetch.hh" 43#include "mem/packet.hh" 44#include "mem/request.hh" 45#include "params/DerivO3CPU.hh" 46#include "sim/byteswap.hh" 47#include "sim/core.hh" 48 49#if FULL_SYSTEM 50#include "arch/tlb.hh" 51#include "arch/vtophys.hh" 52#include "sim/system.hh" 53#endif // FULL_SYSTEM 54 55using namespace std; 56 57template<class Impl> 58void 59DefaultFetch<Impl>::IcachePort::setPeer(Port *port) 60{ 61 Port::setPeer(port); 62 63 fetch->setIcache(); 64} 65 66template<class Impl> 67Tick 68DefaultFetch<Impl>::IcachePort::recvAtomic(PacketPtr pkt) 69{ 70 panic("DefaultFetch doesn't expect recvAtomic callback!"); 71 return curTick; 72} 73 74template<class Impl> 75void 76DefaultFetch<Impl>::IcachePort::recvFunctional(PacketPtr pkt) 77{ 78 DPRINTF(Fetch, "DefaultFetch doesn't update its state from a " 79 "functional call."); 80} 81 82template<class Impl> 83void 84DefaultFetch<Impl>::IcachePort::recvStatusChange(Status status) 85{ 86 if (status == RangeChange) { 87 if (!snoopRangeSent) { 88 snoopRangeSent = true; 89 sendStatusChange(Port::RangeChange); 90 } 91 return; 92 } 93 94 panic("DefaultFetch doesn't expect recvStatusChange callback!"); 95} 96 97template<class Impl> 98bool 99DefaultFetch<Impl>::IcachePort::recvTiming(PacketPtr pkt) 100{ 101 DPRINTF(Fetch, "Received timing\n"); 102 if (pkt->isResponse()) { 103 fetch->processCacheCompletion(pkt); 104 } 105 //else Snooped a coherence request, just return 106 return true; 107} 108 109template<class Impl> 110void 111DefaultFetch<Impl>::IcachePort::recvRetry() 112{ 113 fetch->recvRetry(); 114} 115 116template<class Impl> 117DefaultFetch<Impl>::DefaultFetch(O3CPU *_cpu, DerivO3CPUParams *params) 118 : cpu(_cpu), 119 branchPred(params), 120 predecoder(NULL), 121 decodeToFetchDelay(params->decodeToFetchDelay), 122 renameToFetchDelay(params->renameToFetchDelay), 123 iewToFetchDelay(params->iewToFetchDelay), 124 commitToFetchDelay(params->commitToFetchDelay), 125 fetchWidth(params->fetchWidth), 126 cacheBlocked(false), 127 retryPkt(NULL), 128 retryTid(InvalidThreadID), 129 numThreads(params->numThreads), 130 numFetchingThreads(params->smtNumFetchingThreads), 131 interruptPending(false), 132 drainPending(false), 133 switchedOut(false) 134{ 135 if (numThreads > Impl::MaxThreads) 136 fatal("numThreads (%d) is larger than compiled limit (%d),\n" 137 "\tincrease MaxThreads in src/cpu/o3/impl.hh\n", 138 numThreads, static_cast<int>(Impl::MaxThreads)); 139 140 // Set fetch stage's status to inactive. 141 _status = Inactive; 142 143 std::string policy = params->smtFetchPolicy; 144 145 // Convert string to lowercase 146 std::transform(policy.begin(), policy.end(), policy.begin(), 147 (int(*)(int)) tolower); 148 149 // Figure out fetch policy 150 if (policy == "singlethread") { 151 fetchPolicy = SingleThread; 152 if (numThreads > 1) 153 panic("Invalid Fetch Policy for a SMT workload."); 154 } else if (policy == "roundrobin") { 155 fetchPolicy = RoundRobin; 156 DPRINTF(Fetch, "Fetch policy set to Round Robin\n"); 157 } else if (policy == "branch") { 158 fetchPolicy = Branch; 159 DPRINTF(Fetch, "Fetch policy set to Branch Count\n"); 160 } else if (policy == "iqcount") { 161 fetchPolicy = IQ; 162 DPRINTF(Fetch, "Fetch policy set to IQ count\n"); 163 } else if (policy == "lsqcount") { 164 fetchPolicy = LSQ; 165 DPRINTF(Fetch, "Fetch policy set to LSQ count\n"); 166 } else { 167 fatal("Invalid Fetch Policy. Options Are: {SingleThread," 168 " RoundRobin,LSQcount,IQcount}\n"); 169 } 170 171 // Get the size of an instruction. 172 instSize = sizeof(TheISA::MachInst); 173 174 // Name is finally available, so create the port. 175 icachePort = new IcachePort(this); 176 177 icachePort->snoopRangeSent = false; 178 179#if USE_CHECKER 180 if (cpu->checker) { 181 cpu->checker->setIcachePort(icachePort); 182 } 183#endif 184} 185 186template <class Impl> 187std::string 188DefaultFetch<Impl>::name() const 189{ 190 return cpu->name() + ".fetch"; 191} 192 193template <class Impl> 194void 195DefaultFetch<Impl>::regStats() 196{ 197 icacheStallCycles 198 .name(name() + ".icacheStallCycles") 199 .desc("Number of cycles fetch is stalled on an Icache miss") 200 .prereq(icacheStallCycles); 201 202 fetchedInsts 203 .name(name() + ".Insts") 204 .desc("Number of instructions fetch has processed") 205 .prereq(fetchedInsts); 206 207 fetchedBranches 208 .name(name() + ".Branches") 209 .desc("Number of branches that fetch encountered") 210 .prereq(fetchedBranches); 211 212 predictedBranches 213 .name(name() + ".predictedBranches") 214 .desc("Number of branches that fetch has predicted taken") 215 .prereq(predictedBranches); 216 217 fetchCycles 218 .name(name() + ".Cycles") 219 .desc("Number of cycles fetch has run and was not squashing or" 220 " blocked") 221 .prereq(fetchCycles); 222 223 fetchSquashCycles 224 .name(name() + ".SquashCycles") 225 .desc("Number of cycles fetch has spent squashing") 226 .prereq(fetchSquashCycles); 227 228 fetchIdleCycles 229 .name(name() + ".IdleCycles") 230 .desc("Number of cycles fetch was idle") 231 .prereq(fetchIdleCycles); 232 233 fetchBlockedCycles 234 .name(name() + ".BlockedCycles") 235 .desc("Number of cycles fetch has spent blocked") 236 .prereq(fetchBlockedCycles); 237 238 fetchedCacheLines 239 .name(name() + ".CacheLines") 240 .desc("Number of cache lines fetched") 241 .prereq(fetchedCacheLines); 242 243 fetchMiscStallCycles 244 .name(name() + ".MiscStallCycles") 245 .desc("Number of cycles fetch has spent waiting on interrupts, or " 246 "bad addresses, or out of MSHRs") 247 .prereq(fetchMiscStallCycles); 248 249 fetchIcacheSquashes 250 .name(name() + ".IcacheSquashes") 251 .desc("Number of outstanding Icache misses that were squashed") 252 .prereq(fetchIcacheSquashes); 253 254 fetchNisnDist 255 .init(/* base value */ 0, 256 /* last value */ fetchWidth, 257 /* bucket size */ 1) 258 .name(name() + ".rateDist") 259 .desc("Number of instructions fetched each cycle (Total)") 260 .flags(Stats::pdf); 261 262 idleRate 263 .name(name() + ".idleRate") 264 .desc("Percent of cycles fetch was idle") 265 .prereq(idleRate); 266 idleRate = fetchIdleCycles * 100 / cpu->numCycles; 267 268 branchRate 269 .name(name() + ".branchRate") 270 .desc("Number of branch fetches per cycle") 271 .flags(Stats::total); 272 branchRate = fetchedBranches / cpu->numCycles; 273 274 fetchRate 275 .name(name() + ".rate") 276 .desc("Number of inst fetches per cycle") 277 .flags(Stats::total); 278 fetchRate = fetchedInsts / cpu->numCycles; 279 280 branchPred.regStats(); 281} 282 283template<class Impl> 284void 285DefaultFetch<Impl>::setTimeBuffer(TimeBuffer<TimeStruct> *time_buffer) 286{ 287 timeBuffer = time_buffer; 288 289 // Create wires to get information from proper places in time buffer. 290 fromDecode = timeBuffer->getWire(-decodeToFetchDelay); 291 fromRename = timeBuffer->getWire(-renameToFetchDelay); 292 fromIEW = timeBuffer->getWire(-iewToFetchDelay); 293 fromCommit = timeBuffer->getWire(-commitToFetchDelay); 294} 295 296template<class Impl> 297void 298DefaultFetch<Impl>::setActiveThreads(std::list<ThreadID> *at_ptr) 299{ 300 activeThreads = at_ptr; 301} 302 303template<class Impl> 304void 305DefaultFetch<Impl>::setFetchQueue(TimeBuffer<FetchStruct> *fq_ptr) 306{ 307 fetchQueue = fq_ptr; 308 309 // Create wire to write information to proper place in fetch queue. 310 toDecode = fetchQueue->getWire(0); 311} 312 313template<class Impl> 314void 315DefaultFetch<Impl>::initStage() 316{ 317 // Setup PC and nextPC with initial state. 318 for (ThreadID tid = 0; tid < numThreads; tid++) { 319 pc[tid] = cpu->pcState(tid); 320 } 321 322 for (ThreadID tid = 0; tid < numThreads; tid++) { 323 324 fetchStatus[tid] = Running; 325 326 priorityList.push_back(tid); 327 328 memReq[tid] = NULL; 329 330 stalls[tid].decode = false; 331 stalls[tid].rename = false; 332 stalls[tid].iew = false; 333 stalls[tid].commit = false; 334 } 335 336 // Schedule fetch to get the correct PC from the CPU 337 // scheduleFetchStartupEvent(1); 338 339 // Fetch needs to start fetching instructions at the very beginning, 340 // so it must start up in active state. 341 switchToActive(); 342} 343 344template<class Impl> 345void 346DefaultFetch<Impl>::setIcache() 347{ 348 // Size of cache block. 349 cacheBlkSize = icachePort->peerBlockSize(); 350 351 // Create mask to get rid of offset bits. 352 cacheBlkMask = (cacheBlkSize - 1); 353 354 for (ThreadID tid = 0; tid < numThreads; tid++) { 355 // Create space to store a cache line. 356 cacheData[tid] = new uint8_t[cacheBlkSize]; 357 cacheDataPC[tid] = 0; 358 cacheDataValid[tid] = false; 359 } 360} 361 362template<class Impl> 363void 364DefaultFetch<Impl>::processCacheCompletion(PacketPtr pkt) 365{ 366 ThreadID tid = pkt->req->threadId(); 367 368 DPRINTF(Fetch, "[tid:%u] Waking up from cache miss.\n",tid); 369 370 assert(!pkt->wasNacked()); 371 372 // Only change the status if it's still waiting on the icache access 373 // to return. 374 if (fetchStatus[tid] != IcacheWaitResponse || 375 pkt->req != memReq[tid] || 376 isSwitchedOut()) { 377 ++fetchIcacheSquashes; 378 delete pkt->req; 379 delete pkt; 380 return; 381 } 382 383 memcpy(cacheData[tid], pkt->getPtr<uint8_t>(), cacheBlkSize); 384 cacheDataValid[tid] = true; 385 386 if (!drainPending) { 387 // Wake up the CPU (if it went to sleep and was waiting on 388 // this completion event). 389 cpu->wakeCPU(); 390 391 DPRINTF(Activity, "[tid:%u] Activating fetch due to cache completion\n", 392 tid); 393 394 switchToActive(); 395 } 396 397 // Only switch to IcacheAccessComplete if we're not stalled as well. 398 if (checkStall(tid)) { 399 fetchStatus[tid] = Blocked; 400 } else { 401 fetchStatus[tid] = IcacheAccessComplete; 402 } 403 404 // Reset the mem req to NULL. 405 delete pkt->req; 406 delete pkt; 407 memReq[tid] = NULL; 408} 409 410template <class Impl> 411bool 412DefaultFetch<Impl>::drain() 413{ 414 // Fetch is ready to drain at any time. 415 cpu->signalDrained(); 416 drainPending = true; 417 return true; 418} 419 420template <class Impl> 421void 422DefaultFetch<Impl>::resume() 423{ 424 drainPending = false; 425} 426 427template <class Impl> 428void 429DefaultFetch<Impl>::switchOut() 430{ 431 switchedOut = true; 432 // Branch predictor needs to have its state cleared. 433 branchPred.switchOut(); 434} 435 436template <class Impl> 437void 438DefaultFetch<Impl>::takeOverFrom() 439{ 440 // Reset all state 441 for (ThreadID i = 0; i < Impl::MaxThreads; ++i) { 442 stalls[i].decode = 0; 443 stalls[i].rename = 0; 444 stalls[i].iew = 0; 445 stalls[i].commit = 0; 446 pc[i] = cpu->pcState(i); 447 fetchStatus[i] = Running; 448 } 449 numInst = 0; 450 wroteToTimeBuffer = false; 451 _status = Inactive; 452 switchedOut = false; 453 interruptPending = false; 454 branchPred.takeOverFrom(); 455} 456 457template <class Impl> 458void 459DefaultFetch<Impl>::wakeFromQuiesce() 460{ 461 DPRINTF(Fetch, "Waking up from quiesce\n"); 462 // Hopefully this is safe 463 // @todo: Allow other threads to wake from quiesce. 464 fetchStatus[0] = Running; 465} 466 467template <class Impl> 468inline void 469DefaultFetch<Impl>::switchToActive() 470{ 471 if (_status == Inactive) { 472 DPRINTF(Activity, "Activating stage.\n"); 473 474 cpu->activateStage(O3CPU::FetchIdx); 475 476 _status = Active; 477 } 478} 479 480template <class Impl> 481inline void 482DefaultFetch<Impl>::switchToInactive() 483{ 484 if (_status == Active) { 485 DPRINTF(Activity, "Deactivating stage.\n"); 486 487 cpu->deactivateStage(O3CPU::FetchIdx); 488 489 _status = Inactive; 490 } 491} 492 493template <class Impl> 494bool 495DefaultFetch<Impl>::lookupAndUpdateNextPC( 496 DynInstPtr &inst, TheISA::PCState &nextPC) 497{ 498 // Do branch prediction check here. 499 // A bit of a misnomer...next_PC is actually the current PC until 500 // this function updates it. 501 bool predict_taken; 502 503 if (!inst->isControl()) { 504 TheISA::advancePC(nextPC, inst->staticInst); 505 inst->setPredTarg(nextPC); 506 inst->setPredTaken(false); 507 return false; 508 } 509 510 ThreadID tid = inst->threadNumber; 511 predict_taken = branchPred.predict(inst, nextPC, tid); 512 513 if (predict_taken) { 514 DPRINTF(Fetch, "[tid:%i]: [sn:%i]: Branch predicted to be taken to %s.\n", 515 tid, inst->seqNum, nextPC); 516 } else { 517 DPRINTF(Fetch, "[tid:%i]: [sn:%i]:Branch predicted to be not taken.\n", 518 tid, inst->seqNum); 519 } 520 521 DPRINTF(Fetch, "[tid:%i]: [sn:%i] Branch predicted to go to %s.\n", 522 tid, inst->seqNum, nextPC); 523 inst->setPredTarg(nextPC); 524 inst->setPredTaken(predict_taken); 525 526 ++fetchedBranches; 527 528 if (predict_taken) { 529 ++predictedBranches; 530 } 531 532 return predict_taken; 533} 534 535template <class Impl> 536bool 537DefaultFetch<Impl>::fetchCacheLine(Addr fetch_PC, Fault &ret_fault, ThreadID tid) 538{ 539 Fault fault = NoFault; 540 541 //AlphaDep 542 if (cacheBlocked) { 543 DPRINTF(Fetch, "[tid:%i] Can't fetch cache line, cache blocked\n", 544 tid); 545 return false; 546 } else if (isSwitchedOut()) { 547 DPRINTF(Fetch, "[tid:%i] Can't fetch cache line, switched out\n", 548 tid); 549 return false; 550 } else if (interruptPending && !(fetch_PC & 0x3)) { 551 // Hold off fetch from getting new instructions when: 552 // Cache is blocked, or 553 // while an interrupt is pending and we're not in PAL mode, or 554 // fetch is switched out. 555 DPRINTF(Fetch, "[tid:%i] Can't fetch cache line, interrupt pending\n", 556 tid); 557 return false; 558 } 559 560 // Align the fetch PC so it's at the start of a cache block. 561 Addr block_PC = icacheBlockAlignPC(fetch_PC); 562 563 // If we've already got the block, no need to try to fetch it again. 564 if (cacheDataValid[tid] && block_PC == cacheDataPC[tid]) { 565 return true; 566 } 567 568 // Setup the memReq to do a read of the first instruction's address. 569 // Set the appropriate read size and flags as well. 570 // Build request here. 571 RequestPtr mem_req = 572 new Request(tid, block_PC, cacheBlkSize, Request::INST_FETCH, 573 fetch_PC, cpu->thread[tid]->contextId(), tid); 574 575 memReq[tid] = mem_req; 576 577 // Translate the instruction request. 578 fault = cpu->itb->translateAtomic(mem_req, cpu->thread[tid]->getTC(), 579 BaseTLB::Execute); 580 581 // In the case of faults, the fetch stage may need to stall and wait 582 // for the ITB miss to be handled. 583 584 // If translation was successful, attempt to read the first 585 // instruction. 586 if (fault == NoFault) { 587#if 0 588 if (cpu->system->memctrl->badaddr(memReq[tid]->paddr) || 589 memReq[tid]->isUncacheable()) { 590 DPRINTF(Fetch, "Fetch: Bad address %#x (hopefully on a " 591 "misspeculating path)!", 592 memReq[tid]->paddr); 593 ret_fault = TheISA::genMachineCheckFault(); 594 return false; 595 } 596#endif 597 598 // Build packet here. 599 PacketPtr data_pkt = new Packet(mem_req, 600 MemCmd::ReadReq, Packet::Broadcast); 601 data_pkt->dataDynamicArray(new uint8_t[cacheBlkSize]); 602 603 cacheDataPC[tid] = block_PC; 604 cacheDataValid[tid] = false; 605 606 DPRINTF(Fetch, "Fetch: Doing instruction read.\n"); 607 608 fetchedCacheLines++; 609 610 // Now do the timing access to see whether or not the instruction 611 // exists within the cache. 612 if (!icachePort->sendTiming(data_pkt)) { 613 assert(retryPkt == NULL); 614 assert(retryTid == InvalidThreadID); 615 DPRINTF(Fetch, "[tid:%i] Out of MSHRs!\n", tid); 616 fetchStatus[tid] = IcacheWaitRetry; 617 retryPkt = data_pkt; 618 retryTid = tid; 619 cacheBlocked = true; 620 return false; 621 } 622 623 DPRINTF(Fetch, "[tid:%i]: Doing cache access.\n", tid); 624 625 lastIcacheStall[tid] = curTick; 626 627 DPRINTF(Activity, "[tid:%i]: Activity: Waiting on I-cache " 628 "response.\n", tid); 629 630 fetchStatus[tid] = IcacheWaitResponse; 631 } else { 632 delete mem_req; 633 memReq[tid] = NULL; 634 } 635 636 ret_fault = fault; 637 return true; 638} 639 640template <class Impl> 641inline void 642DefaultFetch<Impl>::doSquash(const TheISA::PCState &newPC, ThreadID tid) 643{ 644 DPRINTF(Fetch, "[tid:%i]: Squashing, setting PC to: %s.\n", 645 tid, newPC); 646 647 pc[tid] = newPC; 648 649 // Clear the icache miss if it's outstanding. 650 if (fetchStatus[tid] == IcacheWaitResponse) { 651 DPRINTF(Fetch, "[tid:%i]: Squashing outstanding Icache miss.\n", 652 tid); 653 memReq[tid] = NULL; 654 } 655 656 // Get rid of the retrying packet if it was from this thread. 657 if (retryTid == tid) { 658 assert(cacheBlocked); 659 if (retryPkt) { 660 delete retryPkt->req; 661 delete retryPkt; 662 } 663 retryPkt = NULL; 664 retryTid = InvalidThreadID; 665 } 666 667 fetchStatus[tid] = Squashing; 668 669 ++fetchSquashCycles; 670} 671 672template<class Impl> 673void 674DefaultFetch<Impl>::squashFromDecode(const TheISA::PCState &newPC, 675 const InstSeqNum &seq_num, ThreadID tid) 676{ 677 DPRINTF(Fetch, "[tid:%i]: Squashing from decode.\n", tid); 678 679 doSquash(newPC, tid); 680 681 // Tell the CPU to remove any instructions that are in flight between 682 // fetch and decode. 683 cpu->removeInstsUntil(seq_num, tid); 684} 685 686template<class Impl> 687bool 688DefaultFetch<Impl>::checkStall(ThreadID tid) const 689{ 690 bool ret_val = false; 691 692 if (cpu->contextSwitch) { 693 DPRINTF(Fetch,"[tid:%i]: Stalling for a context switch.\n",tid); 694 ret_val = true; 695 } else if (stalls[tid].decode) { 696 DPRINTF(Fetch,"[tid:%i]: Stall from Decode stage detected.\n",tid); 697 ret_val = true; 698 } else if (stalls[tid].rename) { 699 DPRINTF(Fetch,"[tid:%i]: Stall from Rename stage detected.\n",tid); 700 ret_val = true; 701 } else if (stalls[tid].iew) { 702 DPRINTF(Fetch,"[tid:%i]: Stall from IEW stage detected.\n",tid); 703 ret_val = true; 704 } else if (stalls[tid].commit) { 705 DPRINTF(Fetch,"[tid:%i]: Stall from Commit stage detected.\n",tid); 706 ret_val = true; 707 } 708 709 return ret_val; 710} 711 712template<class Impl> 713typename DefaultFetch<Impl>::FetchStatus 714DefaultFetch<Impl>::updateFetchStatus() 715{ 716 //Check Running 717 list<ThreadID>::iterator threads = activeThreads->begin(); 718 list<ThreadID>::iterator end = activeThreads->end(); 719 720 while (threads != end) { 721 ThreadID tid = *threads++; 722 723 if (fetchStatus[tid] == Running || 724 fetchStatus[tid] == Squashing || 725 fetchStatus[tid] == IcacheAccessComplete) { 726 727 if (_status == Inactive) { 728 DPRINTF(Activity, "[tid:%i]: Activating stage.\n",tid); 729 730 if (fetchStatus[tid] == IcacheAccessComplete) { 731 DPRINTF(Activity, "[tid:%i]: Activating fetch due to cache" 732 "completion\n",tid); 733 } 734 735 cpu->activateStage(O3CPU::FetchIdx); 736 } 737 738 return Active; 739 } 740 } 741 742 // Stage is switching from active to inactive, notify CPU of it. 743 if (_status == Active) { 744 DPRINTF(Activity, "Deactivating stage.\n"); 745 746 cpu->deactivateStage(O3CPU::FetchIdx); 747 } 748 749 return Inactive; 750} 751 752template <class Impl> 753void 754DefaultFetch<Impl>::squash(const TheISA::PCState &newPC, 755 const InstSeqNum &seq_num, ThreadID tid) 756{ 757 DPRINTF(Fetch, "[tid:%u]: Squash from commit.\n", tid); 758 759 doSquash(newPC, tid); 760 761 // Tell the CPU to remove any instructions that are not in the ROB. 762 cpu->removeInstsNotInROB(tid); 763} 764 765template <class Impl> 766void 767DefaultFetch<Impl>::tick() 768{ 769 list<ThreadID>::iterator threads = activeThreads->begin(); 770 list<ThreadID>::iterator end = activeThreads->end(); 771 bool status_change = false; 772 773 wroteToTimeBuffer = false; 774 775 while (threads != end) { 776 ThreadID tid = *threads++; 777 778 // Check the signals for each thread to determine the proper status 779 // for each thread. 780 bool updated_status = checkSignalsAndUpdate(tid); 781 status_change = status_change || updated_status; 782 } 783 784 DPRINTF(Fetch, "Running stage.\n"); 785 786 // Reset the number of the instruction we're fetching. 787 numInst = 0; 788 789#if FULL_SYSTEM 790 if (fromCommit->commitInfo[0].interruptPending) { 791 interruptPending = true; 792 } 793 794 if (fromCommit->commitInfo[0].clearInterrupt) { 795 interruptPending = false; 796 } 797#endif 798 799 for (threadFetched = 0; threadFetched < numFetchingThreads; 800 threadFetched++) { 801 // Fetch each of the actively fetching threads. 802 fetch(status_change); 803 } 804 805 // Record number of instructions fetched this cycle for distribution. 806 fetchNisnDist.sample(numInst); 807 808 if (status_change) { 809 // Change the fetch stage status if there was a status change. 810 _status = updateFetchStatus(); 811 } 812 813 // If there was activity this cycle, inform the CPU of it. 814 if (wroteToTimeBuffer || cpu->contextSwitch) { 815 DPRINTF(Activity, "Activity this cycle.\n"); 816 817 cpu->activityThisCycle(); 818 } 819} 820 821template <class Impl> 822bool 823DefaultFetch<Impl>::checkSignalsAndUpdate(ThreadID tid) 824{ 825 // Update the per thread stall statuses. 826 if (fromDecode->decodeBlock[tid]) { 827 stalls[tid].decode = true; 828 } 829 830 if (fromDecode->decodeUnblock[tid]) { 831 assert(stalls[tid].decode); 832 assert(!fromDecode->decodeBlock[tid]); 833 stalls[tid].decode = false; 834 } 835 836 if (fromRename->renameBlock[tid]) { 837 stalls[tid].rename = true; 838 } 839 840 if (fromRename->renameUnblock[tid]) { 841 assert(stalls[tid].rename); 842 assert(!fromRename->renameBlock[tid]); 843 stalls[tid].rename = false; 844 } 845 846 if (fromIEW->iewBlock[tid]) { 847 stalls[tid].iew = true; 848 } 849 850 if (fromIEW->iewUnblock[tid]) { 851 assert(stalls[tid].iew); 852 assert(!fromIEW->iewBlock[tid]); 853 stalls[tid].iew = false; 854 } 855 856 if (fromCommit->commitBlock[tid]) { 857 stalls[tid].commit = true; 858 } 859 860 if (fromCommit->commitUnblock[tid]) { 861 assert(stalls[tid].commit); 862 assert(!fromCommit->commitBlock[tid]); 863 stalls[tid].commit = false; 864 } 865 866 // Check squash signals from commit. 867 if (fromCommit->commitInfo[tid].squash) { 868 869 DPRINTF(Fetch, "[tid:%u]: Squashing instructions due to squash " 870 "from commit.\n",tid); 871 // In any case, squash. 872 squash(fromCommit->commitInfo[tid].pc, 873 fromCommit->commitInfo[tid].doneSeqNum, 874 tid); 875 876 // Also check if there's a mispredict that happened. 877 if (fromCommit->commitInfo[tid].branchMispredict) { 878 branchPred.squash(fromCommit->commitInfo[tid].doneSeqNum, 879 fromCommit->commitInfo[tid].pc, 880 fromCommit->commitInfo[tid].branchTaken, 881 tid); 882 } else { 883 branchPred.squash(fromCommit->commitInfo[tid].doneSeqNum, 884 tid); 885 } 886 887 return true; 888 } else if (fromCommit->commitInfo[tid].doneSeqNum) { 889 // Update the branch predictor if it wasn't a squashed instruction 890 // that was broadcasted. 891 branchPred.update(fromCommit->commitInfo[tid].doneSeqNum, tid); 892 } 893 894 // Check ROB squash signals from commit. 895 if (fromCommit->commitInfo[tid].robSquashing) { 896 DPRINTF(Fetch, "[tid:%u]: ROB is still squashing.\n", tid); 897 898 // Continue to squash. 899 fetchStatus[tid] = Squashing; 900 901 return true; 902 } 903 904 // Check squash signals from decode. 905 if (fromDecode->decodeInfo[tid].squash) { 906 DPRINTF(Fetch, "[tid:%u]: Squashing instructions due to squash " 907 "from decode.\n",tid); 908 909 // Update the branch predictor. 910 if (fromDecode->decodeInfo[tid].branchMispredict) { 911 branchPred.squash(fromDecode->decodeInfo[tid].doneSeqNum, 912 fromDecode->decodeInfo[tid].nextPC, 913 fromDecode->decodeInfo[tid].branchTaken, 914 tid); 915 } else { 916 branchPred.squash(fromDecode->decodeInfo[tid].doneSeqNum, 917 tid); 918 } 919 920 if (fetchStatus[tid] != Squashing) { 921 922 TheISA::PCState nextPC = fromDecode->decodeInfo[tid].nextPC; 923 DPRINTF(Fetch, "Squashing from decode with PC = %s\n", nextPC); 924 // Squash unless we're already squashing 925 squashFromDecode(fromDecode->decodeInfo[tid].nextPC, 926 fromDecode->decodeInfo[tid].doneSeqNum, 927 tid); 928 929 return true; 930 } 931 } 932 933 if (checkStall(tid) && 934 fetchStatus[tid] != IcacheWaitResponse && 935 fetchStatus[tid] != IcacheWaitRetry) { 936 DPRINTF(Fetch, "[tid:%i]: Setting to blocked\n",tid); 937 938 fetchStatus[tid] = Blocked; 939 940 return true; 941 } 942 943 if (fetchStatus[tid] == Blocked || 944 fetchStatus[tid] == Squashing) { 945 // Switch status to running if fetch isn't being told to block or 946 // squash this cycle. 947 DPRINTF(Fetch, "[tid:%i]: Done squashing, switching to running.\n", 948 tid); 949 950 fetchStatus[tid] = Running; 951 952 return true; 953 } 954 955 // If we've reached this point, we have not gotten any signals that 956 // cause fetch to change its status. Fetch remains the same as before. 957 return false; 958} 959 960template<class Impl> 961void 962DefaultFetch<Impl>::fetch(bool &status_change) 963{ 964 ////////////////////////////////////////// 965 // Start actual fetch 966 ////////////////////////////////////////// 967 ThreadID tid = getFetchingThread(fetchPolicy); 968 969 if (tid == InvalidThreadID || drainPending) { 970 DPRINTF(Fetch,"There are no more threads available to fetch from.\n"); 971 972 // Breaks looping condition in tick() 973 threadFetched = numFetchingThreads; 974 return; 975 } 976 977 DPRINTF(Fetch, "Attempting to fetch from [tid:%i]\n", tid); 978 979 // The current PC. 980 TheISA::PCState fetchPC = pc[tid]; 981 982 // Fault code for memory access. 983 Fault fault = NoFault; 984 985 // If returning from the delay of a cache miss, then update the status 986 // to running, otherwise do the cache access. Possibly move this up 987 // to tick() function. 988 if (fetchStatus[tid] == IcacheAccessComplete) { 989 DPRINTF(Fetch, "[tid:%i]: Icache miss is complete.\n", 990 tid); 991 992 fetchStatus[tid] = Running; 993 status_change = true; 994 } else if (fetchStatus[tid] == Running) { 995 DPRINTF(Fetch, "[tid:%i]: Attempting to translate and read " 996 "instruction, starting at PC %s.\n", tid, fetchPC); 997 998 bool fetch_success = fetchCacheLine(fetchPC.instAddr(), fault, tid); 999 if (!fetch_success) { 1000 if (cacheBlocked) { 1001 ++icacheStallCycles; 1002 } else { 1003 ++fetchMiscStallCycles; 1004 } 1005 return; 1006 } 1007 } else { 1008 if (fetchStatus[tid] == Idle) { 1009 ++fetchIdleCycles; 1010 DPRINTF(Fetch, "[tid:%i]: Fetch is idle!\n", tid); 1011 } else if (fetchStatus[tid] == Blocked) { 1012 ++fetchBlockedCycles; 1013 DPRINTF(Fetch, "[tid:%i]: Fetch is blocked!\n", tid); 1014 } else if (fetchStatus[tid] == Squashing) { 1015 ++fetchSquashCycles; 1016 DPRINTF(Fetch, "[tid:%i]: Fetch is squashing!\n", tid); 1017 } else if (fetchStatus[tid] == IcacheWaitResponse) { 1018 ++icacheStallCycles; 1019 DPRINTF(Fetch, "[tid:%i]: Fetch is waiting cache response!\n", tid); 1020 } 1021 1022 // Status is Idle, Squashing, Blocked, or IcacheWaitResponse, so 1023 // fetch should do nothing. 1024 return; 1025 } 1026 1027 ++fetchCycles; 1028 1029 // If we had a stall due to an icache miss, then return. 1030 if (fetchStatus[tid] == IcacheWaitResponse) { 1031 ++icacheStallCycles; 1032 status_change = true; 1033 return; 1034 } 1035 1036 TheISA::PCState nextPC = fetchPC; 1037 1038 InstSeqNum inst_seq; 1039 MachInst inst; 1040 ExtMachInst ext_inst; 1041 1042 StaticInstPtr staticInst = NULL; 1043 StaticInstPtr macroop = NULL; 1044 1045 if (fault == NoFault) { 1046 //XXX Masking out pal mode bit. This will break x86. Alpha needs 1047 //to pull the pal mode bit ouf ot the instruction address. 1048 unsigned offset = (fetchPC.instAddr() & ~1) - cacheDataPC[tid]; 1049 assert(offset < cacheBlkSize); 1050 1051 // If the read of the first instruction was successful, then grab the 1052 // instructions from the rest of the cache line and put them into the 1053 // queue heading to decode. 1054 1055 DPRINTF(Fetch, "[tid:%i]: Adding instructions to queue to " 1056 "decode.\n",tid); 1057 1058 // Need to keep track of whether or not a predicted branch 1059 // ended this fetch block. 1060 bool predicted_branch = false; 1061 1062 while (offset < cacheBlkSize && 1063 numInst < fetchWidth && 1064 !predicted_branch) { 1065 1066 // Make sure this is a valid index. 1067 assert(offset <= cacheBlkSize - instSize); 1068 1069 if (!macroop) { 1070 // Get the instruction from the array of the cache line. 1071 inst = TheISA::gtoh(*reinterpret_cast<TheISA::MachInst *> 1072 (&cacheData[tid][offset])); 1073 1074 predecoder.setTC(cpu->thread[tid]->getTC()); 1075 predecoder.moreBytes(fetchPC, fetchPC.instAddr(), inst); 1076 1077 ext_inst = predecoder.getExtMachInst(fetchPC); 1078 staticInst = StaticInstPtr(ext_inst, fetchPC.instAddr()); 1079 if (staticInst->isMacroop()) 1080 macroop = staticInst; 1081 } 1082 do { 1083 if (macroop) { 1084 staticInst = macroop->fetchMicroop(fetchPC.microPC()); 1085 if (staticInst->isLastMicroop()) 1086 macroop = NULL; 1087 } 1088 1089 // Get a sequence number. 1090 inst_seq = cpu->getAndIncrementInstSeq(); 1091 1092 // Create a new DynInst from the instruction fetched. 1093 DynInstPtr instruction = new DynInst(staticInst, 1094 fetchPC, nextPC, 1095 inst_seq, cpu); 1096 instruction->setTid(tid); 1097 1098 instruction->setASID(tid); 1099 1100 instruction->setThreadState(cpu->thread[tid]); 1101 1102 DPRINTF(Fetch, "[tid:%i]: Instruction PC %s (%d) created " 1103 "[sn:%lli]\n", tid, instruction->pcState(), 1104 instruction->microPC(), inst_seq); 1105 1106 //DPRINTF(Fetch, "[tid:%i]: MachInst is %#x\n", tid, ext_inst); 1107 1108 DPRINTF(Fetch, "[tid:%i]: Instruction is: %s\n", tid, 1109 instruction->staticInst-> 1110 disassemble(fetchPC.instAddr())); 1111 1112#if TRACING_ON 1113 instruction->traceData = 1114 cpu->getTracer()->getInstRecord(curTick, cpu->tcBase(tid), 1115 instruction->staticInst, fetchPC, macroop); 1116#else 1117 instruction->traceData = NULL; 1118#endif 1119 1120 // If we're branching after this instruction, quite fetching 1121 // from the same block then. 1122 predicted_branch = fetchPC.branching(); 1123 predicted_branch |= 1124 lookupAndUpdateNextPC(instruction, nextPC); 1125 if (predicted_branch) { 1126 DPRINTF(Fetch, "Branch detected with PC = %s\n", fetchPC); 1127 } 1128 1129 // Add instruction to the CPU's list of instructions. 1130 instruction->setInstListIt(cpu->addInst(instruction)); 1131 1132 // Write the instruction to the first slot in the queue 1133 // that heads to decode. 1134 toDecode->insts[numInst] = instruction; 1135 1136 toDecode->size++; 1137 1138 // Increment stat of fetched instructions. 1139 ++fetchedInsts; 1140 1141 // Move to the next instruction, unless we have a branch. 1142 fetchPC = nextPC; 1143 1144 if (instruction->isQuiesce()) { 1145 DPRINTF(Fetch, "Quiesce instruction encountered, halting fetch!", 1146 curTick); 1147 fetchStatus[tid] = QuiescePending; 1148 ++numInst; 1149 status_change = true; 1150 break; 1151 } 1152 1153 ++numInst; 1154 } while (staticInst->isMicroop() && 1155 !staticInst->isLastMicroop() && 1156 numInst < fetchWidth); 1157 //XXX Masking out pal mode bit. 1158 offset = (fetchPC.instAddr() & ~1) - cacheDataPC[tid]; 1159 } 1160 1161 if (predicted_branch) { 1162 DPRINTF(Fetch, "[tid:%i]: Done fetching, predicted branch " 1163 "instruction encountered.\n", tid); 1164 } else if (numInst >= fetchWidth) { 1165 DPRINTF(Fetch, "[tid:%i]: Done fetching, reached fetch bandwidth " 1166 "for this cycle.\n", tid); 1167 } else if (offset >= cacheBlkSize) { 1168 DPRINTF(Fetch, "[tid:%i]: Done fetching, reached the end of cache " 1169 "block.\n", tid); 1170 } 1171 } 1172 1173 if (numInst > 0) { 1174 wroteToTimeBuffer = true; 1175 } 1176 1177 // Now that fetching is completed, update the PC to signify what the next 1178 // cycle will be. 1179 if (fault == NoFault) { 1180 pc[tid] = nextPC; 1181 DPRINTF(Fetch, "[tid:%i]: Setting PC to %s.\n", tid, nextPC); 1182 } else { 1183 // We shouldn't be in an icache miss and also have a fault (an ITB 1184 // miss) 1185 if (fetchStatus[tid] == IcacheWaitResponse) { 1186 panic("Fetch should have exited prior to this!"); 1187 } 1188 1189 // Send the fault to commit. This thread will not do anything 1190 // until commit handles the fault. The only other way it can 1191 // wake up is if a squash comes along and changes the PC. 1192 assert(numInst < fetchWidth); 1193 // Get a sequence number. 1194 inst_seq = cpu->getAndIncrementInstSeq(); 1195 // We will use a nop in order to carry the fault. 1196 ext_inst = TheISA::NoopMachInst; 1197 1198 // Create a new DynInst from the dummy nop. 1199 DynInstPtr instruction = new DynInst(ext_inst, fetchPC, nextPC, 1200 inst_seq, cpu); 1201 TheISA::advancePC(nextPC, instruction->staticInst); 1202 instruction->setPredTarg(nextPC); 1203 instruction->setTid(tid); 1204 1205 instruction->setASID(tid); 1206 1207 instruction->setThreadState(cpu->thread[tid]); 1208 1209 instruction->traceData = NULL; 1210 1211 instruction->setInstListIt(cpu->addInst(instruction)); 1212 1213 instruction->fault = fault; 1214 1215 toDecode->insts[numInst] = instruction; 1216 toDecode->size++; 1217 1218 wroteToTimeBuffer = true; 1219 1220 DPRINTF(Fetch, "[tid:%i]: Blocked, need to handle the trap.\n",tid); 1221 1222 fetchStatus[tid] = TrapPending; 1223 status_change = true; 1224 1225 DPRINTF(Fetch, "[tid:%i]: fault (%s) detected @ PC %s", 1226 tid, fault->name(), pc[tid]); 1227 } 1228} 1229 1230template<class Impl> 1231void 1232DefaultFetch<Impl>::recvRetry() 1233{ 1234 if (retryPkt != NULL) { 1235 assert(cacheBlocked); 1236 assert(retryTid != InvalidThreadID); 1237 assert(fetchStatus[retryTid] == IcacheWaitRetry); 1238 1239 if (icachePort->sendTiming(retryPkt)) { 1240 fetchStatus[retryTid] = IcacheWaitResponse; 1241 retryPkt = NULL; 1242 retryTid = InvalidThreadID; 1243 cacheBlocked = false; 1244 } 1245 } else { 1246 assert(retryTid == InvalidThreadID); 1247 // Access has been squashed since it was sent out. Just clear 1248 // the cache being blocked. 1249 cacheBlocked = false; 1250 } 1251} 1252 1253/////////////////////////////////////// 1254// // 1255// SMT FETCH POLICY MAINTAINED HERE // 1256// // 1257/////////////////////////////////////// 1258template<class Impl> 1259ThreadID 1260DefaultFetch<Impl>::getFetchingThread(FetchPriority &fetch_priority) 1261{ 1262 if (numThreads > 1) { 1263 switch (fetch_priority) { 1264 1265 case SingleThread: 1266 return 0; 1267 1268 case RoundRobin: 1269 return roundRobin(); 1270 1271 case IQ: 1272 return iqCount(); 1273 1274 case LSQ: 1275 return lsqCount(); 1276 1277 case Branch: 1278 return branchCount(); 1279 1280 default: 1281 return InvalidThreadID; 1282 } 1283 } else { 1284 list<ThreadID>::iterator thread = activeThreads->begin(); 1285 if (thread == activeThreads->end()) { 1286 return InvalidThreadID; 1287 } 1288 1289 ThreadID tid = *thread; 1290 1291 if (fetchStatus[tid] == Running || 1292 fetchStatus[tid] == IcacheAccessComplete || 1293 fetchStatus[tid] == Idle) { 1294 return tid; 1295 } else { 1296 return InvalidThreadID; 1297 } 1298 } 1299} 1300 1301 1302template<class Impl> 1303ThreadID 1304DefaultFetch<Impl>::roundRobin() 1305{ 1306 list<ThreadID>::iterator pri_iter = priorityList.begin(); 1307 list<ThreadID>::iterator end = priorityList.end(); 1308 1309 ThreadID high_pri; 1310 1311 while (pri_iter != end) { 1312 high_pri = *pri_iter; 1313 1314 assert(high_pri <= numThreads); 1315 1316 if (fetchStatus[high_pri] == Running || 1317 fetchStatus[high_pri] == IcacheAccessComplete || 1318 fetchStatus[high_pri] == Idle) { 1319 1320 priorityList.erase(pri_iter); 1321 priorityList.push_back(high_pri); 1322 1323 return high_pri; 1324 } 1325 1326 pri_iter++; 1327 } 1328 1329 return InvalidThreadID; 1330} 1331 1332template<class Impl> 1333ThreadID 1334DefaultFetch<Impl>::iqCount() 1335{ 1336 std::priority_queue<ThreadID> PQ; 1337 1338 list<ThreadID>::iterator threads = activeThreads->begin(); 1339 list<ThreadID>::iterator end = activeThreads->end(); 1340 1341 while (threads != end) { 1342 ThreadID tid = *threads++; 1343 1344 PQ.push(fromIEW->iewInfo[tid].iqCount); 1345 } 1346 1347 while (!PQ.empty()) { 1348 ThreadID high_pri = PQ.top(); 1349 1350 if (fetchStatus[high_pri] == Running || 1351 fetchStatus[high_pri] == IcacheAccessComplete || 1352 fetchStatus[high_pri] == Idle) 1353 return high_pri; 1354 else 1355 PQ.pop(); 1356 1357 } 1358 1359 return InvalidThreadID; 1360} 1361 1362template<class Impl> 1363ThreadID 1364DefaultFetch<Impl>::lsqCount() 1365{ 1366 std::priority_queue<ThreadID> PQ; 1367 1368 list<ThreadID>::iterator threads = activeThreads->begin(); 1369 list<ThreadID>::iterator end = activeThreads->end(); 1370 1371 while (threads != end) { 1372 ThreadID tid = *threads++; 1373 1374 PQ.push(fromIEW->iewInfo[tid].ldstqCount); 1375 } 1376 1377 while (!PQ.empty()) { 1378 ThreadID high_pri = PQ.top(); 1379 1380 if (fetchStatus[high_pri] == Running || 1381 fetchStatus[high_pri] == IcacheAccessComplete || 1382 fetchStatus[high_pri] == Idle) 1383 return high_pri; 1384 else 1385 PQ.pop(); 1386 } 1387 1388 return InvalidThreadID; 1389} 1390 1391template<class Impl> 1392ThreadID 1393DefaultFetch<Impl>::branchCount() 1394{ 1395#if 0 1396 list<ThreadID>::iterator thread = activeThreads->begin(); 1397 assert(thread != activeThreads->end()); 1398 ThreadID tid = *thread; 1399#endif 1400 1401 panic("Branch Count Fetch policy unimplemented\n"); 1402 return InvalidThreadID; 1403} 1404