fetch_impl.hh revision 4636:afc8da9f526e
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 "config/use_checker.hh" 33 34#include "arch/isa_traits.hh" 35#include "arch/utility.hh" 36#include "cpu/checker/cpu.hh" 37#include "cpu/exetrace.hh" 38#include "cpu/o3/fetch.hh" 39#include "mem/packet.hh" 40#include "mem/request.hh" 41#include "sim/byteswap.hh" 42#include "sim/host.hh" 43#include "sim/core.hh" 44 45#if FULL_SYSTEM 46#include "arch/tlb.hh" 47#include "arch/vtophys.hh" 48#include "sim/system.hh" 49#endif // FULL_SYSTEM 50 51#include <algorithm> 52 53template<class Impl> 54void 55DefaultFetch<Impl>::IcachePort::setPeer(Port *port) 56{ 57 Port::setPeer(port); 58 59 fetch->setIcache(); 60} 61 62template<class Impl> 63Tick 64DefaultFetch<Impl>::IcachePort::recvAtomic(PacketPtr pkt) 65{ 66 panic("DefaultFetch doesn't expect recvAtomic callback!"); 67 return curTick; 68} 69 70template<class Impl> 71void 72DefaultFetch<Impl>::IcachePort::recvFunctional(PacketPtr pkt) 73{ 74 DPRINTF(Fetch, "DefaultFetch doesn't update its state from a " 75 "functional call."); 76} 77 78template<class Impl> 79void 80DefaultFetch<Impl>::IcachePort::recvStatusChange(Status status) 81{ 82 if (status == RangeChange) { 83 if (!snoopRangeSent) { 84 snoopRangeSent = true; 85 sendStatusChange(Port::RangeChange); 86 } 87 return; 88 } 89 90 panic("DefaultFetch doesn't expect recvStatusChange callback!"); 91} 92 93template<class Impl> 94bool 95DefaultFetch<Impl>::IcachePort::recvTiming(PacketPtr pkt) 96{ 97 DPRINTF(Fetch, "Received timing\n"); 98 if (pkt->isResponse()) { 99 fetch->processCacheCompletion(pkt); 100 } 101 //else Snooped a coherence request, just return 102 return true; 103} 104 105template<class Impl> 106void 107DefaultFetch<Impl>::IcachePort::recvRetry() 108{ 109 fetch->recvRetry(); 110} 111 112template<class Impl> 113DefaultFetch<Impl>::DefaultFetch(O3CPU *_cpu, Params *params) 114 : cpu(_cpu), 115 branchPred(params), 116 predecoder(NULL), 117 decodeToFetchDelay(params->decodeToFetchDelay), 118 renameToFetchDelay(params->renameToFetchDelay), 119 iewToFetchDelay(params->iewToFetchDelay), 120 commitToFetchDelay(params->commitToFetchDelay), 121 fetchWidth(params->fetchWidth), 122 cacheBlocked(false), 123 retryPkt(NULL), 124 retryTid(-1), 125 numThreads(params->numberOfThreads), 126 numFetchingThreads(params->smtNumFetchingThreads), 127 interruptPending(false), 128 drainPending(false), 129 switchedOut(false) 130{ 131 if (numThreads > Impl::MaxThreads) 132 fatal("numThreads is not a valid value\n"); 133 134 // Set fetch stage's status to inactive. 135 _status = Inactive; 136 137 std::string policy = params->smtFetchPolicy; 138 139 // Convert string to lowercase 140 std::transform(policy.begin(), policy.end(), policy.begin(), 141 (int(*)(int)) tolower); 142 143 // Figure out fetch policy 144 if (policy == "singlethread") { 145 fetchPolicy = SingleThread; 146 if (numThreads > 1) 147 panic("Invalid Fetch Policy for a SMT workload."); 148 } else if (policy == "roundrobin") { 149 fetchPolicy = RoundRobin; 150 DPRINTF(Fetch, "Fetch policy set to Round Robin\n"); 151 } else if (policy == "branch") { 152 fetchPolicy = Branch; 153 DPRINTF(Fetch, "Fetch policy set to Branch Count\n"); 154 } else if (policy == "iqcount") { 155 fetchPolicy = IQ; 156 DPRINTF(Fetch, "Fetch policy set to IQ count\n"); 157 } else if (policy == "lsqcount") { 158 fetchPolicy = LSQ; 159 DPRINTF(Fetch, "Fetch policy set to LSQ count\n"); 160 } else { 161 fatal("Invalid Fetch Policy. Options Are: {SingleThread," 162 " RoundRobin,LSQcount,IQcount}\n"); 163 } 164 165 // Get the size of an instruction. 166 instSize = sizeof(TheISA::MachInst); 167 168 // Name is finally available, so create the port. 169 icachePort = new IcachePort(this); 170 171 icachePort->snoopRangeSent = false; 172 173#if USE_CHECKER 174 if (cpu->checker) { 175 cpu->checker->setIcachePort(icachePort); 176 } 177#endif 178} 179 180template <class Impl> 181std::string 182DefaultFetch<Impl>::name() const 183{ 184 return cpu->name() + ".fetch"; 185} 186 187template <class Impl> 188void 189DefaultFetch<Impl>::regStats() 190{ 191 icacheStallCycles 192 .name(name() + ".icacheStallCycles") 193 .desc("Number of cycles fetch is stalled on an Icache miss") 194 .prereq(icacheStallCycles); 195 196 fetchedInsts 197 .name(name() + ".Insts") 198 .desc("Number of instructions fetch has processed") 199 .prereq(fetchedInsts); 200 201 fetchedBranches 202 .name(name() + ".Branches") 203 .desc("Number of branches that fetch encountered") 204 .prereq(fetchedBranches); 205 206 predictedBranches 207 .name(name() + ".predictedBranches") 208 .desc("Number of branches that fetch has predicted taken") 209 .prereq(predictedBranches); 210 211 fetchCycles 212 .name(name() + ".Cycles") 213 .desc("Number of cycles fetch has run and was not squashing or" 214 " blocked") 215 .prereq(fetchCycles); 216 217 fetchSquashCycles 218 .name(name() + ".SquashCycles") 219 .desc("Number of cycles fetch has spent squashing") 220 .prereq(fetchSquashCycles); 221 222 fetchIdleCycles 223 .name(name() + ".IdleCycles") 224 .desc("Number of cycles fetch was idle") 225 .prereq(fetchIdleCycles); 226 227 fetchBlockedCycles 228 .name(name() + ".BlockedCycles") 229 .desc("Number of cycles fetch has spent blocked") 230 .prereq(fetchBlockedCycles); 231 232 fetchedCacheLines 233 .name(name() + ".CacheLines") 234 .desc("Number of cache lines fetched") 235 .prereq(fetchedCacheLines); 236 237 fetchMiscStallCycles 238 .name(name() + ".MiscStallCycles") 239 .desc("Number of cycles fetch has spent waiting on interrupts, or " 240 "bad addresses, or out of MSHRs") 241 .prereq(fetchMiscStallCycles); 242 243 fetchIcacheSquashes 244 .name(name() + ".IcacheSquashes") 245 .desc("Number of outstanding Icache misses that were squashed") 246 .prereq(fetchIcacheSquashes); 247 248 fetchNisnDist 249 .init(/* base value */ 0, 250 /* last value */ fetchWidth, 251 /* bucket size */ 1) 252 .name(name() + ".rateDist") 253 .desc("Number of instructions fetched each cycle (Total)") 254 .flags(Stats::pdf); 255 256 idleRate 257 .name(name() + ".idleRate") 258 .desc("Percent of cycles fetch was idle") 259 .prereq(idleRate); 260 idleRate = fetchIdleCycles * 100 / cpu->numCycles; 261 262 branchRate 263 .name(name() + ".branchRate") 264 .desc("Number of branch fetches per cycle") 265 .flags(Stats::total); 266 branchRate = fetchedBranches / cpu->numCycles; 267 268 fetchRate 269 .name(name() + ".rate") 270 .desc("Number of inst fetches per cycle") 271 .flags(Stats::total); 272 fetchRate = fetchedInsts / cpu->numCycles; 273 274 branchPred.regStats(); 275} 276 277template<class Impl> 278void 279DefaultFetch<Impl>::setTimeBuffer(TimeBuffer<TimeStruct> *time_buffer) 280{ 281 timeBuffer = time_buffer; 282 283 // Create wires to get information from proper places in time buffer. 284 fromDecode = timeBuffer->getWire(-decodeToFetchDelay); 285 fromRename = timeBuffer->getWire(-renameToFetchDelay); 286 fromIEW = timeBuffer->getWire(-iewToFetchDelay); 287 fromCommit = timeBuffer->getWire(-commitToFetchDelay); 288} 289 290template<class Impl> 291void 292DefaultFetch<Impl>::setActiveThreads(std::list<unsigned> *at_ptr) 293{ 294 activeThreads = at_ptr; 295} 296 297template<class Impl> 298void 299DefaultFetch<Impl>::setFetchQueue(TimeBuffer<FetchStruct> *fq_ptr) 300{ 301 fetchQueue = fq_ptr; 302 303 // Create wire to write information to proper place in fetch queue. 304 toDecode = fetchQueue->getWire(0); 305} 306 307template<class Impl> 308void 309DefaultFetch<Impl>::initStage() 310{ 311 // Setup PC and nextPC with initial state. 312 for (int tid = 0; tid < numThreads; tid++) { 313 PC[tid] = cpu->readPC(tid); 314 nextPC[tid] = cpu->readNextPC(tid); 315 microPC[tid] = cpu->readMicroPC(tid); 316 } 317 318 for (int tid=0; tid < numThreads; tid++) { 319 320 fetchStatus[tid] = Running; 321 322 priorityList.push_back(tid); 323 324 memReq[tid] = NULL; 325 326 stalls[tid].decode = false; 327 stalls[tid].rename = false; 328 stalls[tid].iew = false; 329 stalls[tid].commit = false; 330 } 331 332 // Schedule fetch to get the correct PC from the CPU 333 // scheduleFetchStartupEvent(1); 334 335 // Fetch needs to start fetching instructions at the very beginning, 336 // so it must start up in active state. 337 switchToActive(); 338} 339 340template<class Impl> 341void 342DefaultFetch<Impl>::setIcache() 343{ 344 // Size of cache block. 345 cacheBlkSize = icachePort->peerBlockSize(); 346 347 // Create mask to get rid of offset bits. 348 cacheBlkMask = (cacheBlkSize - 1); 349 350 for (int tid=0; tid < numThreads; tid++) { 351 // Create space to store a cache line. 352 cacheData[tid] = new uint8_t[cacheBlkSize]; 353 cacheDataPC[tid] = 0; 354 cacheDataValid[tid] = false; 355 } 356} 357 358template<class Impl> 359void 360DefaultFetch<Impl>::processCacheCompletion(PacketPtr pkt) 361{ 362 unsigned tid = pkt->req->getThreadNum(); 363 364 DPRINTF(Fetch, "[tid:%u] Waking up from cache miss.\n",tid); 365 366 // Only change the status if it's still waiting on the icache access 367 // to return. 368 if (fetchStatus[tid] != IcacheWaitResponse || 369 pkt->req != memReq[tid] || 370 isSwitchedOut()) { 371 ++fetchIcacheSquashes; 372 delete pkt->req; 373 delete pkt; 374 return; 375 } 376 377 memcpy(cacheData[tid], pkt->getPtr<uint8_t *>(), cacheBlkSize); 378 cacheDataValid[tid] = true; 379 380 if (!drainPending) { 381 // Wake up the CPU (if it went to sleep and was waiting on 382 // this completion event). 383 cpu->wakeCPU(); 384 385 DPRINTF(Activity, "[tid:%u] Activating fetch due to cache completion\n", 386 tid); 387 388 switchToActive(); 389 } 390 391 // Only switch to IcacheAccessComplete if we're not stalled as well. 392 if (checkStall(tid)) { 393 fetchStatus[tid] = Blocked; 394 } else { 395 fetchStatus[tid] = IcacheAccessComplete; 396 } 397 398 // Reset the mem req to NULL. 399 delete pkt->req; 400 delete pkt; 401 memReq[tid] = NULL; 402} 403 404template <class Impl> 405bool 406DefaultFetch<Impl>::drain() 407{ 408 // Fetch is ready to drain at any time. 409 cpu->signalDrained(); 410 drainPending = true; 411 return true; 412} 413 414template <class Impl> 415void 416DefaultFetch<Impl>::resume() 417{ 418 drainPending = false; 419} 420 421template <class Impl> 422void 423DefaultFetch<Impl>::switchOut() 424{ 425 switchedOut = true; 426 // Branch predictor needs to have its state cleared. 427 branchPred.switchOut(); 428} 429 430template <class Impl> 431void 432DefaultFetch<Impl>::takeOverFrom() 433{ 434 // Reset all state 435 for (int i = 0; i < Impl::MaxThreads; ++i) { 436 stalls[i].decode = 0; 437 stalls[i].rename = 0; 438 stalls[i].iew = 0; 439 stalls[i].commit = 0; 440 PC[i] = cpu->readPC(i); 441 nextPC[i] = cpu->readNextPC(i); 442 microPC[i] = cpu->readMicroPC(i); 443 fetchStatus[i] = Running; 444 } 445 numInst = 0; 446 wroteToTimeBuffer = false; 447 _status = Inactive; 448 switchedOut = false; 449 interruptPending = false; 450 branchPred.takeOverFrom(); 451} 452 453template <class Impl> 454void 455DefaultFetch<Impl>::wakeFromQuiesce() 456{ 457 DPRINTF(Fetch, "Waking up from quiesce\n"); 458 // Hopefully this is safe 459 // @todo: Allow other threads to wake from quiesce. 460 fetchStatus[0] = Running; 461} 462 463template <class Impl> 464inline void 465DefaultFetch<Impl>::switchToActive() 466{ 467 if (_status == Inactive) { 468 DPRINTF(Activity, "Activating stage.\n"); 469 470 cpu->activateStage(O3CPU::FetchIdx); 471 472 _status = Active; 473 } 474} 475 476template <class Impl> 477inline void 478DefaultFetch<Impl>::switchToInactive() 479{ 480 if (_status == Active) { 481 DPRINTF(Activity, "Deactivating stage.\n"); 482 483 cpu->deactivateStage(O3CPU::FetchIdx); 484 485 _status = Inactive; 486 } 487} 488 489template <class Impl> 490bool 491DefaultFetch<Impl>::lookupAndUpdateNextPC(DynInstPtr &inst, Addr &next_PC, 492 Addr &next_NPC, Addr &next_MicroPC) 493{ 494 // Do branch prediction check here. 495 // A bit of a misnomer...next_PC is actually the current PC until 496 // this function updates it. 497 bool predict_taken; 498 499 if (!inst->isControl()) { 500 if (inst->isMicroOp() && !inst->isLastMicroOp()) { 501 next_MicroPC++; 502 } else { 503 next_PC = next_NPC; 504 next_NPC = next_NPC + instSize; 505 next_MicroPC = 0; 506 } 507 inst->setPredTarg(next_PC, next_NPC, next_MicroPC); 508 inst->setPredTaken(false); 509 return false; 510 } 511 512 //Assume for now that all control flow is to a different macroop which 513 //would reset the micro pc to 0. 514 next_MicroPC = 0; 515 516 int tid = inst->threadNumber; 517 Addr pred_PC = next_PC; 518 predict_taken = branchPred.predict(inst, pred_PC, tid); 519 520/* if (predict_taken) { 521 DPRINTF(Fetch, "[tid:%i]: Branch predicted to be taken to %#x.\n", 522 tid, pred_PC); 523 } else { 524 DPRINTF(Fetch, "[tid:%i]: Branch predicted to be not taken.\n", tid); 525 }*/ 526 527#if ISA_HAS_DELAY_SLOT 528 next_PC = next_NPC; 529 if (predict_taken) 530 next_NPC = pred_PC; 531 else 532 next_NPC += instSize; 533#else 534 if (predict_taken) 535 next_PC = pred_PC; 536 else 537 next_PC += instSize; 538 next_NPC = next_PC + instSize; 539#endif 540/* DPRINTF(Fetch, "[tid:%i]: Branch predicted to go to %#x and then %#x.\n", 541 tid, next_PC, next_NPC);*/ 542 inst->setPredTarg(next_PC, next_NPC, next_MicroPC); 543 inst->setPredTaken(predict_taken); 544 545 ++fetchedBranches; 546 547 if (predict_taken) { 548 ++predictedBranches; 549 } 550 551 return predict_taken; 552} 553 554template <class Impl> 555bool 556DefaultFetch<Impl>::fetchCacheLine(Addr fetch_PC, Fault &ret_fault, unsigned tid) 557{ 558 Fault fault = NoFault; 559 560 //AlphaDep 561 if (cacheBlocked) { 562 DPRINTF(Fetch, "[tid:%i] Can't fetch cache line, cache blocked\n", 563 tid); 564 return false; 565 } else if (isSwitchedOut()) { 566 DPRINTF(Fetch, "[tid:%i] Can't fetch cache line, switched out\n", 567 tid); 568 return false; 569 } else if (interruptPending && !(fetch_PC & 0x3)) { 570 // Hold off fetch from getting new instructions when: 571 // Cache is blocked, or 572 // while an interrupt is pending and we're not in PAL mode, or 573 // fetch is switched out. 574 DPRINTF(Fetch, "[tid:%i] Can't fetch cache line, interrupt pending\n", 575 tid); 576 return false; 577 } 578 579 // Align the fetch PC so it's at the start of a cache block. 580 Addr block_PC = icacheBlockAlignPC(fetch_PC); 581 582 // If we've already got the block, no need to try to fetch it again. 583 if (cacheDataValid[tid] && block_PC == cacheDataPC[tid]) { 584 return true; 585 } 586 587 // Setup the memReq to do a read of the first instruction's address. 588 // Set the appropriate read size and flags as well. 589 // Build request here. 590 RequestPtr mem_req = new Request(tid, block_PC, cacheBlkSize, 0, 591 fetch_PC, cpu->readCpuId(), tid); 592 593 memReq[tid] = mem_req; 594 595 // Translate the instruction request. 596 fault = cpu->translateInstReq(mem_req, cpu->thread[tid]); 597 598 // In the case of faults, the fetch stage may need to stall and wait 599 // for the ITB miss to be handled. 600 601 // If translation was successful, attempt to read the first 602 // instruction. 603 if (fault == NoFault) { 604#if 0 605 if (cpu->system->memctrl->badaddr(memReq[tid]->paddr) || 606 memReq[tid]->isUncacheable()) { 607 DPRINTF(Fetch, "Fetch: Bad address %#x (hopefully on a " 608 "misspeculating path)!", 609 memReq[tid]->paddr); 610 ret_fault = TheISA::genMachineCheckFault(); 611 return false; 612 } 613#endif 614 615 // Build packet here. 616 PacketPtr data_pkt = new Packet(mem_req, 617 MemCmd::ReadReq, Packet::Broadcast); 618 data_pkt->dataDynamicArray(new uint8_t[cacheBlkSize]); 619 620 cacheDataPC[tid] = block_PC; 621 cacheDataValid[tid] = false; 622 623 DPRINTF(Fetch, "Fetch: Doing instruction read.\n"); 624 625 fetchedCacheLines++; 626 627 // Now do the timing access to see whether or not the instruction 628 // exists within the cache. 629 if (!icachePort->sendTiming(data_pkt)) { 630 if (data_pkt->result == Packet::BadAddress) { 631 fault = TheISA::genMachineCheckFault(); 632 delete mem_req; 633 memReq[tid] = NULL; 634 warn("Bad address!\n"); 635 } 636 assert(retryPkt == NULL); 637 assert(retryTid == -1); 638 DPRINTF(Fetch, "[tid:%i] Out of MSHRs!\n", tid); 639 fetchStatus[tid] = IcacheWaitRetry; 640 retryPkt = data_pkt; 641 retryTid = tid; 642 cacheBlocked = true; 643 return false; 644 } 645 646 DPRINTF(Fetch, "[tid:%i]: Doing cache access.\n", tid); 647 648 lastIcacheStall[tid] = curTick; 649 650 DPRINTF(Activity, "[tid:%i]: Activity: Waiting on I-cache " 651 "response.\n", tid); 652 653 fetchStatus[tid] = IcacheWaitResponse; 654 } else { 655 delete mem_req; 656 memReq[tid] = NULL; 657 } 658 659 ret_fault = fault; 660 return true; 661} 662 663template <class Impl> 664inline void 665DefaultFetch<Impl>::doSquash(const Addr &new_PC, 666 const Addr &new_NPC, const Addr &new_microPC, unsigned tid) 667{ 668 DPRINTF(Fetch, "[tid:%i]: Squashing, setting PC to: %#x, NPC to: %#x.\n", 669 tid, new_PC, new_NPC); 670 671 PC[tid] = new_PC; 672 nextPC[tid] = new_NPC; 673 microPC[tid] = new_microPC; 674 675 // Clear the icache miss if it's outstanding. 676 if (fetchStatus[tid] == IcacheWaitResponse) { 677 DPRINTF(Fetch, "[tid:%i]: Squashing outstanding Icache miss.\n", 678 tid); 679 memReq[tid] = NULL; 680 } 681 682 // Get rid of the retrying packet if it was from this thread. 683 if (retryTid == tid) { 684 assert(cacheBlocked); 685 if (retryPkt) { 686 delete retryPkt->req; 687 delete retryPkt; 688 } 689 retryPkt = NULL; 690 retryTid = -1; 691 } 692 693 fetchStatus[tid] = Squashing; 694 695 ++fetchSquashCycles; 696} 697 698template<class Impl> 699void 700DefaultFetch<Impl>::squashFromDecode(const Addr &new_PC, const Addr &new_NPC, 701 const Addr &new_MicroPC, 702 const InstSeqNum &seq_num, unsigned tid) 703{ 704 DPRINTF(Fetch, "[tid:%i]: Squashing from decode.\n",tid); 705 706 doSquash(new_PC, new_NPC, new_MicroPC, tid); 707 708 // Tell the CPU to remove any instructions that are in flight between 709 // fetch and decode. 710 cpu->removeInstsUntil(seq_num, tid); 711} 712 713template<class Impl> 714bool 715DefaultFetch<Impl>::checkStall(unsigned tid) const 716{ 717 bool ret_val = false; 718 719 if (cpu->contextSwitch) { 720 DPRINTF(Fetch,"[tid:%i]: Stalling for a context switch.\n",tid); 721 ret_val = true; 722 } else if (stalls[tid].decode) { 723 DPRINTF(Fetch,"[tid:%i]: Stall from Decode stage detected.\n",tid); 724 ret_val = true; 725 } else if (stalls[tid].rename) { 726 DPRINTF(Fetch,"[tid:%i]: Stall from Rename stage detected.\n",tid); 727 ret_val = true; 728 } else if (stalls[tid].iew) { 729 DPRINTF(Fetch,"[tid:%i]: Stall from IEW stage detected.\n",tid); 730 ret_val = true; 731 } else if (stalls[tid].commit) { 732 DPRINTF(Fetch,"[tid:%i]: Stall from Commit stage detected.\n",tid); 733 ret_val = true; 734 } 735 736 return ret_val; 737} 738 739template<class Impl> 740typename DefaultFetch<Impl>::FetchStatus 741DefaultFetch<Impl>::updateFetchStatus() 742{ 743 //Check Running 744 std::list<unsigned>::iterator threads = activeThreads->begin(); 745 std::list<unsigned>::iterator end = activeThreads->end(); 746 747 while (threads != end) { 748 unsigned tid = *threads++; 749 750 if (fetchStatus[tid] == Running || 751 fetchStatus[tid] == Squashing || 752 fetchStatus[tid] == IcacheAccessComplete) { 753 754 if (_status == Inactive) { 755 DPRINTF(Activity, "[tid:%i]: Activating stage.\n",tid); 756 757 if (fetchStatus[tid] == IcacheAccessComplete) { 758 DPRINTF(Activity, "[tid:%i]: Activating fetch due to cache" 759 "completion\n",tid); 760 } 761 762 cpu->activateStage(O3CPU::FetchIdx); 763 } 764 765 return Active; 766 } 767 } 768 769 // Stage is switching from active to inactive, notify CPU of it. 770 if (_status == Active) { 771 DPRINTF(Activity, "Deactivating stage.\n"); 772 773 cpu->deactivateStage(O3CPU::FetchIdx); 774 } 775 776 return Inactive; 777} 778 779template <class Impl> 780void 781DefaultFetch<Impl>::squash(const Addr &new_PC, const Addr &new_NPC, 782 const Addr &new_MicroPC, 783 const InstSeqNum &seq_num, unsigned tid) 784{ 785 DPRINTF(Fetch, "[tid:%u]: Squash from commit.\n",tid); 786 787 doSquash(new_PC, new_NPC, new_MicroPC, tid); 788 789 // Tell the CPU to remove any instructions that are not in the ROB. 790 cpu->removeInstsNotInROB(tid); 791} 792 793template <class Impl> 794void 795DefaultFetch<Impl>::tick() 796{ 797 std::list<unsigned>::iterator threads = activeThreads->begin(); 798 std::list<unsigned>::iterator end = activeThreads->end(); 799 bool status_change = false; 800 801 wroteToTimeBuffer = false; 802 803 while (threads != end) { 804 unsigned tid = *threads++; 805 806 // Check the signals for each thread to determine the proper status 807 // for each thread. 808 bool updated_status = checkSignalsAndUpdate(tid); 809 status_change = status_change || updated_status; 810 } 811 812 DPRINTF(Fetch, "Running stage.\n"); 813 814 // Reset the number of the instruction we're fetching. 815 numInst = 0; 816 817#if FULL_SYSTEM 818 if (fromCommit->commitInfo[0].interruptPending) { 819 interruptPending = true; 820 } 821 822 if (fromCommit->commitInfo[0].clearInterrupt) { 823 interruptPending = false; 824 } 825#endif 826 827 for (threadFetched = 0; threadFetched < numFetchingThreads; 828 threadFetched++) { 829 // Fetch each of the actively fetching threads. 830 fetch(status_change); 831 } 832 833 // Record number of instructions fetched this cycle for distribution. 834 fetchNisnDist.sample(numInst); 835 836 if (status_change) { 837 // Change the fetch stage status if there was a status change. 838 _status = updateFetchStatus(); 839 } 840 841 // If there was activity this cycle, inform the CPU of it. 842 if (wroteToTimeBuffer || cpu->contextSwitch) { 843 DPRINTF(Activity, "Activity this cycle.\n"); 844 845 cpu->activityThisCycle(); 846 } 847} 848 849template <class Impl> 850bool 851DefaultFetch<Impl>::checkSignalsAndUpdate(unsigned tid) 852{ 853 // Update the per thread stall statuses. 854 if (fromDecode->decodeBlock[tid]) { 855 stalls[tid].decode = true; 856 } 857 858 if (fromDecode->decodeUnblock[tid]) { 859 assert(stalls[tid].decode); 860 assert(!fromDecode->decodeBlock[tid]); 861 stalls[tid].decode = false; 862 } 863 864 if (fromRename->renameBlock[tid]) { 865 stalls[tid].rename = true; 866 } 867 868 if (fromRename->renameUnblock[tid]) { 869 assert(stalls[tid].rename); 870 assert(!fromRename->renameBlock[tid]); 871 stalls[tid].rename = false; 872 } 873 874 if (fromIEW->iewBlock[tid]) { 875 stalls[tid].iew = true; 876 } 877 878 if (fromIEW->iewUnblock[tid]) { 879 assert(stalls[tid].iew); 880 assert(!fromIEW->iewBlock[tid]); 881 stalls[tid].iew = false; 882 } 883 884 if (fromCommit->commitBlock[tid]) { 885 stalls[tid].commit = true; 886 } 887 888 if (fromCommit->commitUnblock[tid]) { 889 assert(stalls[tid].commit); 890 assert(!fromCommit->commitBlock[tid]); 891 stalls[tid].commit = false; 892 } 893 894 // Check squash signals from commit. 895 if (fromCommit->commitInfo[tid].squash) { 896 897 DPRINTF(Fetch, "[tid:%u]: Squashing instructions due to squash " 898 "from commit.\n",tid); 899 // In any case, squash. 900 squash(fromCommit->commitInfo[tid].nextPC, 901 fromCommit->commitInfo[tid].nextNPC, 902 fromCommit->commitInfo[tid].nextMicroPC, 903 fromCommit->commitInfo[tid].doneSeqNum, 904 tid); 905 906 // Also check if there's a mispredict that happened. 907 if (fromCommit->commitInfo[tid].branchMispredict) { 908 branchPred.squash(fromCommit->commitInfo[tid].doneSeqNum, 909 fromCommit->commitInfo[tid].nextPC, 910 fromCommit->commitInfo[tid].branchTaken, 911 tid); 912 } else { 913 branchPred.squash(fromCommit->commitInfo[tid].doneSeqNum, 914 tid); 915 } 916 917 return true; 918 } else if (fromCommit->commitInfo[tid].doneSeqNum) { 919 // Update the branch predictor if it wasn't a squashed instruction 920 // that was broadcasted. 921 branchPred.update(fromCommit->commitInfo[tid].doneSeqNum, tid); 922 } 923 924 // Check ROB squash signals from commit. 925 if (fromCommit->commitInfo[tid].robSquashing) { 926 DPRINTF(Fetch, "[tid:%u]: ROB is still squashing.\n", tid); 927 928 // Continue to squash. 929 fetchStatus[tid] = Squashing; 930 931 return true; 932 } 933 934 // Check squash signals from decode. 935 if (fromDecode->decodeInfo[tid].squash) { 936 DPRINTF(Fetch, "[tid:%u]: Squashing instructions due to squash " 937 "from decode.\n",tid); 938 939 // Update the branch predictor. 940 if (fromDecode->decodeInfo[tid].branchMispredict) { 941 branchPred.squash(fromDecode->decodeInfo[tid].doneSeqNum, 942 fromDecode->decodeInfo[tid].nextPC, 943 fromDecode->decodeInfo[tid].branchTaken, 944 tid); 945 } else { 946 branchPred.squash(fromDecode->decodeInfo[tid].doneSeqNum, 947 tid); 948 } 949 950 if (fetchStatus[tid] != Squashing) { 951 952 DPRINTF(Fetch, "Squashing from decode with PC = %#x, NPC = %#x\n", 953 fromDecode->decodeInfo[tid].nextPC, 954 fromDecode->decodeInfo[tid].nextNPC); 955 // Squash unless we're already squashing 956 squashFromDecode(fromDecode->decodeInfo[tid].nextPC, 957 fromDecode->decodeInfo[tid].nextNPC, 958 fromDecode->decodeInfo[tid].nextMicroPC, 959 fromDecode->decodeInfo[tid].doneSeqNum, 960 tid); 961 962 return true; 963 } 964 } 965 966 if (checkStall(tid) && 967 fetchStatus[tid] != IcacheWaitResponse && 968 fetchStatus[tid] != IcacheWaitRetry) { 969 DPRINTF(Fetch, "[tid:%i]: Setting to blocked\n",tid); 970 971 fetchStatus[tid] = Blocked; 972 973 return true; 974 } 975 976 if (fetchStatus[tid] == Blocked || 977 fetchStatus[tid] == Squashing) { 978 // Switch status to running if fetch isn't being told to block or 979 // squash this cycle. 980 DPRINTF(Fetch, "[tid:%i]: Done squashing, switching to running.\n", 981 tid); 982 983 fetchStatus[tid] = Running; 984 985 return true; 986 } 987 988 // If we've reached this point, we have not gotten any signals that 989 // cause fetch to change its status. Fetch remains the same as before. 990 return false; 991} 992 993template<class Impl> 994void 995DefaultFetch<Impl>::fetch(bool &status_change) 996{ 997 ////////////////////////////////////////// 998 // Start actual fetch 999 ////////////////////////////////////////// 1000 int tid = getFetchingThread(fetchPolicy); 1001 1002 if (tid == -1 || drainPending) { 1003 DPRINTF(Fetch,"There are no more threads available to fetch from.\n"); 1004 1005 // Breaks looping condition in tick() 1006 threadFetched = numFetchingThreads; 1007 return; 1008 } 1009 1010 DPRINTF(Fetch, "Attempting to fetch from [tid:%i]\n", tid); 1011 1012 // The current PC. 1013 Addr fetch_PC = PC[tid]; 1014 Addr fetch_NPC = nextPC[tid]; 1015 Addr fetch_MicroPC = microPC[tid]; 1016 1017 // Fault code for memory access. 1018 Fault fault = NoFault; 1019 1020 // If returning from the delay of a cache miss, then update the status 1021 // to running, otherwise do the cache access. Possibly move this up 1022 // to tick() function. 1023 if (fetchStatus[tid] == IcacheAccessComplete) { 1024 DPRINTF(Fetch, "[tid:%i]: Icache miss is complete.\n", 1025 tid); 1026 1027 fetchStatus[tid] = Running; 1028 status_change = true; 1029 } else if (fetchStatus[tid] == Running) { 1030 DPRINTF(Fetch, "[tid:%i]: Attempting to translate and read " 1031 "instruction, starting at PC %08p.\n", 1032 tid, fetch_PC); 1033 1034 bool fetch_success = fetchCacheLine(fetch_PC, fault, tid); 1035 if (!fetch_success) { 1036 if (cacheBlocked) { 1037 ++icacheStallCycles; 1038 } else { 1039 ++fetchMiscStallCycles; 1040 } 1041 return; 1042 } 1043 } else { 1044 if (fetchStatus[tid] == Idle) { 1045 ++fetchIdleCycles; 1046 DPRINTF(Fetch, "[tid:%i]: Fetch is idle!\n", tid); 1047 } else if (fetchStatus[tid] == Blocked) { 1048 ++fetchBlockedCycles; 1049 DPRINTF(Fetch, "[tid:%i]: Fetch is blocked!\n", tid); 1050 } else if (fetchStatus[tid] == Squashing) { 1051 ++fetchSquashCycles; 1052 DPRINTF(Fetch, "[tid:%i]: Fetch is squashing!\n", tid); 1053 } else if (fetchStatus[tid] == IcacheWaitResponse) { 1054 ++icacheStallCycles; 1055 DPRINTF(Fetch, "[tid:%i]: Fetch is waiting cache response!\n", tid); 1056 } 1057 1058 // Status is Idle, Squashing, Blocked, or IcacheWaitResponse, so 1059 // fetch should do nothing. 1060 return; 1061 } 1062 1063 ++fetchCycles; 1064 1065 // If we had a stall due to an icache miss, then return. 1066 if (fetchStatus[tid] == IcacheWaitResponse) { 1067 ++icacheStallCycles; 1068 status_change = true; 1069 return; 1070 } 1071 1072 Addr next_PC = fetch_PC; 1073 Addr next_NPC = fetch_NPC; 1074 Addr next_MicroPC = fetch_MicroPC; 1075 1076 InstSeqNum inst_seq; 1077 MachInst inst; 1078 ExtMachInst ext_inst; 1079 // @todo: Fix this hack. 1080 unsigned offset = (fetch_PC & cacheBlkMask) & ~3; 1081 1082 StaticInstPtr staticInst = NULL; 1083 StaticInstPtr macroop = NULL; 1084 1085 if (fault == NoFault) { 1086 // If the read of the first instruction was successful, then grab the 1087 // instructions from the rest of the cache line and put them into the 1088 // queue heading to decode. 1089 1090 DPRINTF(Fetch, "[tid:%i]: Adding instructions to queue to " 1091 "decode.\n",tid); 1092 1093 // Need to keep track of whether or not a predicted branch 1094 // ended this fetch block. 1095 bool predicted_branch = false; 1096 1097 for (; 1098 offset < cacheBlkSize && 1099 numInst < fetchWidth && 1100 !predicted_branch; 1101 ++numInst) { 1102 1103 // If we're branching after this instruction, quite fetching 1104 // from the same block then. 1105 predicted_branch = 1106 (fetch_PC + sizeof(TheISA::MachInst) != fetch_NPC); 1107 if (predicted_branch) { 1108 DPRINTF(Fetch, "Branch detected with PC = %#x, NPC = %#x\n", 1109 fetch_PC, fetch_NPC); 1110 } 1111 1112 1113 // Get a sequence number. 1114 inst_seq = cpu->getAndIncrementInstSeq(); 1115 1116 // Make sure this is a valid index. 1117 assert(offset <= cacheBlkSize - instSize); 1118 1119 if (!macroop) { 1120 // Get the instruction from the array of the cache line. 1121 inst = TheISA::gtoh(*reinterpret_cast<TheISA::MachInst *> 1122 (&cacheData[tid][offset])); 1123 1124 predecoder.setTC(cpu->thread[tid]->getTC()); 1125 predecoder.moreBytes(fetch_PC, 0, inst); 1126 1127 ext_inst = predecoder.getExtMachInst(); 1128 staticInst = StaticInstPtr(ext_inst); 1129 if (staticInst->isMacroOp()) 1130 macroop = staticInst; 1131 } 1132 if (macroop) { 1133 staticInst = macroop->fetchMicroOp(fetch_MicroPC); 1134 if (staticInst->isLastMicroOp()) 1135 macroop = NULL; 1136 } 1137 1138 // Create a new DynInst from the instruction fetched. 1139 DynInstPtr instruction = new DynInst(staticInst, 1140 fetch_PC, fetch_NPC, fetch_MicroPC, 1141 next_PC, next_NPC, next_MicroPC, 1142 inst_seq, cpu); 1143 instruction->setTid(tid); 1144 1145 instruction->setASID(tid); 1146 1147 instruction->setThreadState(cpu->thread[tid]); 1148 1149 DPRINTF(Fetch, "[tid:%i]: Instruction PC %#x created " 1150 "[sn:%lli]\n", 1151 tid, instruction->readPC(), inst_seq); 1152 1153 //DPRINTF(Fetch, "[tid:%i]: MachInst is %#x\n", tid, ext_inst); 1154 1155 DPRINTF(Fetch, "[tid:%i]: Instruction is: %s\n", 1156 tid, instruction->staticInst->disassemble(fetch_PC)); 1157 1158 instruction->traceData = 1159 Trace::getInstRecord(curTick, cpu->tcBase(tid), 1160 instruction->staticInst, 1161 instruction->readPC()); 1162 1163 ///FIXME This needs to be more robust in dealing with delay slots 1164 lookupAndUpdateNextPC(instruction, next_PC, next_NPC, next_MicroPC); 1165 predicted_branch |= (next_PC != fetch_NPC); 1166 1167 // Add instruction to the CPU's list of instructions. 1168 instruction->setInstListIt(cpu->addInst(instruction)); 1169 1170 // Write the instruction to the first slot in the queue 1171 // that heads to decode. 1172 toDecode->insts[numInst] = instruction; 1173 1174 toDecode->size++; 1175 1176 // Increment stat of fetched instructions. 1177 ++fetchedInsts; 1178 1179 // Move to the next instruction, unless we have a branch. 1180 fetch_PC = next_PC; 1181 fetch_NPC = next_NPC; 1182 fetch_MicroPC = next_MicroPC; 1183 1184 if (instruction->isQuiesce()) { 1185 DPRINTF(Fetch, "Quiesce instruction encountered, halting fetch!", 1186 curTick); 1187 fetchStatus[tid] = QuiescePending; 1188 ++numInst; 1189 status_change = true; 1190 break; 1191 } 1192 1193 if (!macroop) 1194 offset += instSize; 1195 } 1196 1197 if (offset >= cacheBlkSize) { 1198 DPRINTF(Fetch, "[tid:%i]: Done fetching, reached the end of cache " 1199 "block.\n", tid); 1200 } else if (numInst >= fetchWidth) { 1201 DPRINTF(Fetch, "[tid:%i]: Done fetching, reached fetch bandwidth " 1202 "for this cycle.\n", tid); 1203 } else if (predicted_branch) { 1204 DPRINTF(Fetch, "[tid:%i]: Done fetching, predicted branch " 1205 "instruction encountered.\n", tid); 1206 } 1207 } 1208 1209 if (numInst > 0) { 1210 wroteToTimeBuffer = true; 1211 } 1212 1213 // Now that fetching is completed, update the PC to signify what the next 1214 // cycle will be. 1215 if (fault == NoFault) { 1216 PC[tid] = next_PC; 1217 nextPC[tid] = next_NPC; 1218 microPC[tid] = next_MicroPC; 1219 DPRINTF(Fetch, "[tid:%i]: Setting PC to %08p.\n", tid, next_PC); 1220 } else { 1221 // We shouldn't be in an icache miss and also have a fault (an ITB 1222 // miss) 1223 if (fetchStatus[tid] == IcacheWaitResponse) { 1224 panic("Fetch should have exited prior to this!"); 1225 } 1226 1227 // Send the fault to commit. This thread will not do anything 1228 // until commit handles the fault. The only other way it can 1229 // wake up is if a squash comes along and changes the PC. 1230#if FULL_SYSTEM 1231 assert(numInst < fetchWidth); 1232 // Get a sequence number. 1233 inst_seq = cpu->getAndIncrementInstSeq(); 1234 // We will use a nop in order to carry the fault. 1235 ext_inst = TheISA::NoopMachInst; 1236 1237 StaticInstPtr staticInst = new StaticInst(ext_inst); 1238 // Create a new DynInst from the dummy nop. 1239 DynInstPtr instruction = new DynInst(staticInst, 1240 fetch_PC, fetch_NPC, 1241 next_PC, next_NPC, 1242 inst_seq, cpu); 1243 instruction->setPredTarg(next_PC, next_NPC); 1244 instruction->setTid(tid); 1245 1246 instruction->setASID(tid); 1247 1248 instruction->setThreadState(cpu->thread[tid]); 1249 1250 instruction->traceData = NULL; 1251 1252 instruction->setInstListIt(cpu->addInst(instruction)); 1253 1254 instruction->fault = fault; 1255 1256 toDecode->insts[numInst] = instruction; 1257 toDecode->size++; 1258 1259 DPRINTF(Fetch, "[tid:%i]: Blocked, need to handle the trap.\n",tid); 1260 1261 fetchStatus[tid] = TrapPending; 1262 status_change = true; 1263#else // !FULL_SYSTEM 1264 fetchStatus[tid] = TrapPending; 1265 status_change = true; 1266 1267#endif // FULL_SYSTEM 1268 DPRINTF(Fetch, "[tid:%i]: fault (%s) detected @ PC %08p", 1269 tid, fault->name(), PC[tid]); 1270 } 1271} 1272 1273template<class Impl> 1274void 1275DefaultFetch<Impl>::recvRetry() 1276{ 1277 if (retryPkt != NULL) { 1278 assert(cacheBlocked); 1279 assert(retryTid != -1); 1280 assert(fetchStatus[retryTid] == IcacheWaitRetry); 1281 1282 if (icachePort->sendTiming(retryPkt)) { 1283 fetchStatus[retryTid] = IcacheWaitResponse; 1284 retryPkt = NULL; 1285 retryTid = -1; 1286 cacheBlocked = false; 1287 } 1288 } else { 1289 assert(retryTid == -1); 1290 // Access has been squashed since it was sent out. Just clear 1291 // the cache being blocked. 1292 cacheBlocked = false; 1293 } 1294} 1295 1296/////////////////////////////////////// 1297// // 1298// SMT FETCH POLICY MAINTAINED HERE // 1299// // 1300/////////////////////////////////////// 1301template<class Impl> 1302int 1303DefaultFetch<Impl>::getFetchingThread(FetchPriority &fetch_priority) 1304{ 1305 if (numThreads > 1) { 1306 switch (fetch_priority) { 1307 1308 case SingleThread: 1309 return 0; 1310 1311 case RoundRobin: 1312 return roundRobin(); 1313 1314 case IQ: 1315 return iqCount(); 1316 1317 case LSQ: 1318 return lsqCount(); 1319 1320 case Branch: 1321 return branchCount(); 1322 1323 default: 1324 return -1; 1325 } 1326 } else { 1327 std::list<unsigned>::iterator thread = activeThreads->begin(); 1328 assert(thread != activeThreads->end()); 1329 int tid = *thread; 1330 1331 if (fetchStatus[tid] == Running || 1332 fetchStatus[tid] == IcacheAccessComplete || 1333 fetchStatus[tid] == Idle) { 1334 return tid; 1335 } else { 1336 return -1; 1337 } 1338 } 1339 1340} 1341 1342 1343template<class Impl> 1344int 1345DefaultFetch<Impl>::roundRobin() 1346{ 1347 std::list<unsigned>::iterator pri_iter = priorityList.begin(); 1348 std::list<unsigned>::iterator end = priorityList.end(); 1349 1350 int high_pri; 1351 1352 while (pri_iter != end) { 1353 high_pri = *pri_iter; 1354 1355 assert(high_pri <= numThreads); 1356 1357 if (fetchStatus[high_pri] == Running || 1358 fetchStatus[high_pri] == IcacheAccessComplete || 1359 fetchStatus[high_pri] == Idle) { 1360 1361 priorityList.erase(pri_iter); 1362 priorityList.push_back(high_pri); 1363 1364 return high_pri; 1365 } 1366 1367 pri_iter++; 1368 } 1369 1370 return -1; 1371} 1372 1373template<class Impl> 1374int 1375DefaultFetch<Impl>::iqCount() 1376{ 1377 std::priority_queue<unsigned> PQ; 1378 1379 std::list<unsigned>::iterator threads = activeThreads->begin(); 1380 std::list<unsigned>::iterator end = activeThreads->end(); 1381 1382 while (threads != end) { 1383 unsigned tid = *threads++; 1384 1385 PQ.push(fromIEW->iewInfo[tid].iqCount); 1386 } 1387 1388 while (!PQ.empty()) { 1389 1390 unsigned high_pri = PQ.top(); 1391 1392 if (fetchStatus[high_pri] == Running || 1393 fetchStatus[high_pri] == IcacheAccessComplete || 1394 fetchStatus[high_pri] == Idle) 1395 return high_pri; 1396 else 1397 PQ.pop(); 1398 1399 } 1400 1401 return -1; 1402} 1403 1404template<class Impl> 1405int 1406DefaultFetch<Impl>::lsqCount() 1407{ 1408 std::priority_queue<unsigned> PQ; 1409 1410 std::list<unsigned>::iterator threads = activeThreads->begin(); 1411 std::list<unsigned>::iterator end = activeThreads->end(); 1412 1413 while (threads != end) { 1414 unsigned tid = *threads++; 1415 1416 PQ.push(fromIEW->iewInfo[tid].ldstqCount); 1417 } 1418 1419 while (!PQ.empty()) { 1420 1421 unsigned high_pri = PQ.top(); 1422 1423 if (fetchStatus[high_pri] == Running || 1424 fetchStatus[high_pri] == IcacheAccessComplete || 1425 fetchStatus[high_pri] == Idle) 1426 return high_pri; 1427 else 1428 PQ.pop(); 1429 1430 } 1431 1432 return -1; 1433} 1434 1435template<class Impl> 1436int 1437DefaultFetch<Impl>::branchCount() 1438{ 1439 std::list<unsigned>::iterator thread = activeThreads->begin(); 1440 assert(thread != activeThreads->end()); 1441 unsigned tid = *thread; 1442 1443 panic("Branch Count Fetch policy unimplemented\n"); 1444 return 0 * tid; 1445} 1446