1/* 2 * Copyright (c) 2010-2014 ARM Limited 3 * Copyright (c) 2012-2013 AMD 4 * All rights reserved. 5 * 6 * The license below extends only to copyright in the software and shall 7 * not be construed as granting a license to any other intellectual 8 * property including but not limited to intellectual property relating 9 * to a hardware implementation of the functionality of the software 10 * licensed hereunder. You may use the software subject to the license 11 * terms below provided that you ensure that this notice is replicated 12 * unmodified and in its entirety in all distributions of the software, 13 * modified or unmodified, in source code or in binary form. 14 * 15 * Copyright (c) 2004-2006 The Regents of The University of Michigan 16 * All rights reserved. 17 * 18 * Redistribution and use in source and binary forms, with or without 19 * modification, are permitted provided that the following conditions are 20 * met: redistributions of source code must retain the above copyright 21 * notice, this list of conditions and the following disclaimer; 22 * redistributions in binary form must reproduce the above copyright 23 * notice, this list of conditions and the following disclaimer in the 24 * documentation and/or other materials provided with the distribution; 25 * neither the name of the copyright holders nor the names of its 26 * contributors may be used to endorse or promote products derived from 27 * this software without specific prior written permission. 28 * 29 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 30 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 31 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 32 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 33 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 34 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 35 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 36 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 37 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 38 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 39 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 40 * 41 * Authors: Kevin Lim 42 * Korey Sewell 43 */ 44 45#ifndef __CPU_O3_FETCH_IMPL_HH__ 46#define __CPU_O3_FETCH_IMPL_HH__ 47 48#include <algorithm> 49#include <cstring> 50#include <list> 51#include <map> 52#include <queue> 53 54#include "arch/generic/tlb.hh" 55#include "arch/isa_traits.hh" 56#include "arch/utility.hh" 57#include "arch/vtophys.hh" 58#include "base/random.hh" 59#include "base/types.hh" 60#include "config/the_isa.hh" 61#include "cpu/base.hh" 62//#include "cpu/checker/cpu.hh" 63#include "cpu/o3/cpu.hh" 64#include "cpu/o3/fetch.hh" 65#include "cpu/exetrace.hh" 66#include "debug/Activity.hh" 67#include "debug/Drain.hh" 68#include "debug/Fetch.hh" 69#include "debug/O3CPU.hh" 70#include "debug/O3PipeView.hh" 71#include "mem/packet.hh" 72#include "params/DerivO3CPU.hh" 73#include "sim/byteswap.hh" 74#include "sim/core.hh" 75#include "sim/eventq.hh" 76#include "sim/full_system.hh" 77#include "sim/system.hh" 78#include "cpu/o3/isa_specific.hh" 79 80using namespace std; 81 82template<class Impl> 83DefaultFetch<Impl>::DefaultFetch(O3CPU *_cpu, DerivO3CPUParams *params) 84 : fetchPolicy(params->smtFetchPolicy), 85 cpu(_cpu), 86 branchPred(nullptr), 87 decodeToFetchDelay(params->decodeToFetchDelay), 88 renameToFetchDelay(params->renameToFetchDelay), 89 iewToFetchDelay(params->iewToFetchDelay), 90 commitToFetchDelay(params->commitToFetchDelay), 91 fetchWidth(params->fetchWidth), 92 decodeWidth(params->decodeWidth), 93 retryPkt(NULL), 94 retryTid(InvalidThreadID), 95 cacheBlkSize(cpu->cacheLineSize()), 96 fetchBufferSize(params->fetchBufferSize), 97 fetchBufferMask(fetchBufferSize - 1), 98 fetchQueueSize(params->fetchQueueSize), 99 numThreads(params->numThreads), 100 numFetchingThreads(params->smtNumFetchingThreads), 101 icachePort(this, _cpu), 102 finishTranslationEvent(this) 103{ 104 if (numThreads > Impl::MaxThreads) 105 fatal("numThreads (%d) is larger than compiled limit (%d),\n" 106 "\tincrease MaxThreads in src/cpu/o3/impl.hh\n", 107 numThreads, static_cast<int>(Impl::MaxThreads)); 108 if (fetchWidth > Impl::MaxWidth) 109 fatal("fetchWidth (%d) is larger than compiled limit (%d),\n" 110 "\tincrease MaxWidth in src/cpu/o3/impl.hh\n", 111 fetchWidth, static_cast<int>(Impl::MaxWidth)); 112 if (fetchBufferSize > cacheBlkSize) 113 fatal("fetch buffer size (%u bytes) is greater than the cache " 114 "block size (%u bytes)\n", fetchBufferSize, cacheBlkSize); 115 if (cacheBlkSize % fetchBufferSize) 116 fatal("cache block (%u bytes) is not a multiple of the " 117 "fetch buffer (%u bytes)\n", cacheBlkSize, fetchBufferSize); 118 119 // Figure out fetch policy 120 panic_if(fetchPolicy == FetchPolicy::SingleThread && numThreads > 1, 121 "Invalid Fetch Policy for a SMT workload."); 122 123 // Get the size of an instruction. 124 instSize = sizeof(TheISA::MachInst); 125 126 for (int i = 0; i < Impl::MaxThreads; i++) { 127 fetchStatus[i] = Idle; 128 decoder[i] = nullptr; 129 pc[i] = 0; 130 fetchOffset[i] = 0; 131 macroop[i] = nullptr; 132 delayedCommit[i] = false; 133 memReq[i] = nullptr; 134 stalls[i] = {false, false}; 135 fetchBuffer[i] = NULL; 136 fetchBufferPC[i] = 0; 137 fetchBufferValid[i] = false; 138 lastIcacheStall[i] = 0; 139 issuePipelinedIfetch[i] = false; 140 } 141 142 branchPred = params->branchPred; 143 144 for (ThreadID tid = 0; tid < numThreads; tid++) { 145 decoder[tid] = new TheISA::Decoder(params->isa[tid]); 146 // Create space to buffer the cache line data, 147 // which may not hold the entire cache line. 148 fetchBuffer[tid] = new uint8_t[fetchBufferSize]; 149 } 150} 151 152template <class Impl> 153std::string 154DefaultFetch<Impl>::name() const 155{ 156 return cpu->name() + ".fetch"; 157} 158 159template <class Impl> 160void 161DefaultFetch<Impl>::regProbePoints() 162{ 163 ppFetch = new ProbePointArg<DynInstPtr>(cpu->getProbeManager(), "Fetch"); 164 ppFetchRequestSent = new ProbePointArg<RequestPtr>(cpu->getProbeManager(), 165 "FetchRequest"); 166 167} 168 169template <class Impl> 170void 171DefaultFetch<Impl>::regStats() 172{ 173 icacheStallCycles 174 .name(name() + ".icacheStallCycles") 175 .desc("Number of cycles fetch is stalled on an Icache miss") 176 .prereq(icacheStallCycles); 177 178 fetchedInsts 179 .name(name() + ".Insts") 180 .desc("Number of instructions fetch has processed") 181 .prereq(fetchedInsts); 182 183 fetchedBranches 184 .name(name() + ".Branches") 185 .desc("Number of branches that fetch encountered") 186 .prereq(fetchedBranches); 187 188 predictedBranches 189 .name(name() + ".predictedBranches") 190 .desc("Number of branches that fetch has predicted taken") 191 .prereq(predictedBranches); 192 193 fetchCycles 194 .name(name() + ".Cycles") 195 .desc("Number of cycles fetch has run and was not squashing or" 196 " blocked") 197 .prereq(fetchCycles); 198 199 fetchSquashCycles 200 .name(name() + ".SquashCycles") 201 .desc("Number of cycles fetch has spent squashing") 202 .prereq(fetchSquashCycles); 203 204 fetchTlbCycles 205 .name(name() + ".TlbCycles") 206 .desc("Number of cycles fetch has spent waiting for tlb") 207 .prereq(fetchTlbCycles); 208 209 fetchIdleCycles 210 .name(name() + ".IdleCycles") 211 .desc("Number of cycles fetch was idle") 212 .prereq(fetchIdleCycles); 213 214 fetchBlockedCycles 215 .name(name() + ".BlockedCycles") 216 .desc("Number of cycles fetch has spent blocked") 217 .prereq(fetchBlockedCycles); 218 219 fetchedCacheLines 220 .name(name() + ".CacheLines") 221 .desc("Number of cache lines fetched") 222 .prereq(fetchedCacheLines); 223 224 fetchMiscStallCycles 225 .name(name() + ".MiscStallCycles") 226 .desc("Number of cycles fetch has spent waiting on interrupts, or " 227 "bad addresses, or out of MSHRs") 228 .prereq(fetchMiscStallCycles); 229 230 fetchPendingDrainCycles 231 .name(name() + ".PendingDrainCycles") 232 .desc("Number of cycles fetch has spent waiting on pipes to drain") 233 .prereq(fetchPendingDrainCycles); 234 235 fetchNoActiveThreadStallCycles 236 .name(name() + ".NoActiveThreadStallCycles") 237 .desc("Number of stall cycles due to no active thread to fetch from") 238 .prereq(fetchNoActiveThreadStallCycles); 239 240 fetchPendingTrapStallCycles 241 .name(name() + ".PendingTrapStallCycles") 242 .desc("Number of stall cycles due to pending traps") 243 .prereq(fetchPendingTrapStallCycles); 244 245 fetchPendingQuiesceStallCycles 246 .name(name() + ".PendingQuiesceStallCycles") 247 .desc("Number of stall cycles due to pending quiesce instructions") 248 .prereq(fetchPendingQuiesceStallCycles); 249 250 fetchIcacheWaitRetryStallCycles 251 .name(name() + ".IcacheWaitRetryStallCycles") 252 .desc("Number of stall cycles due to full MSHR") 253 .prereq(fetchIcacheWaitRetryStallCycles); 254 255 fetchIcacheSquashes 256 .name(name() + ".IcacheSquashes") 257 .desc("Number of outstanding Icache misses that were squashed") 258 .prereq(fetchIcacheSquashes); 259 260 fetchTlbSquashes 261 .name(name() + ".ItlbSquashes") 262 .desc("Number of outstanding ITLB misses that were squashed") 263 .prereq(fetchTlbSquashes); 264 265 fetchNisnDist 266 .init(/* base value */ 0, 267 /* last value */ fetchWidth, 268 /* bucket size */ 1) 269 .name(name() + ".rateDist") 270 .desc("Number of instructions fetched each cycle (Total)") 271 .flags(Stats::pdf); 272 273 idleRate 274 .name(name() + ".idleRate") 275 .desc("Percent of cycles fetch was idle") 276 .prereq(idleRate); 277 idleRate = fetchIdleCycles * 100 / cpu->numCycles; 278 279 branchRate 280 .name(name() + ".branchRate") 281 .desc("Number of branch fetches per cycle") 282 .flags(Stats::total); 283 branchRate = fetchedBranches / cpu->numCycles; 284 285 fetchRate 286 .name(name() + ".rate") 287 .desc("Number of inst fetches per cycle") 288 .flags(Stats::total); 289 fetchRate = fetchedInsts / cpu->numCycles; 290} 291 292template<class Impl> 293void 294DefaultFetch<Impl>::setTimeBuffer(TimeBuffer<TimeStruct> *time_buffer) 295{ 296 timeBuffer = time_buffer; 297 298 // Create wires to get information from proper places in time buffer. 299 fromDecode = timeBuffer->getWire(-decodeToFetchDelay); 300 fromRename = timeBuffer->getWire(-renameToFetchDelay); 301 fromIEW = timeBuffer->getWire(-iewToFetchDelay); 302 fromCommit = timeBuffer->getWire(-commitToFetchDelay); 303} 304 305template<class Impl> 306void 307DefaultFetch<Impl>::setActiveThreads(std::list<ThreadID> *at_ptr) 308{ 309 activeThreads = at_ptr; 310} 311 312template<class Impl> 313void 314DefaultFetch<Impl>::setFetchQueue(TimeBuffer<FetchStruct> *ftb_ptr) 315{ 316 // Create wire to write information to proper place in fetch time buf. 317 toDecode = ftb_ptr->getWire(0); 318} 319 320template<class Impl> 321void 322DefaultFetch<Impl>::startupStage() 323{ 324 assert(priorityList.empty()); 325 resetStage(); 326 327 // Fetch needs to start fetching instructions at the very beginning, 328 // so it must start up in active state. 329 switchToActive(); 330} 331 332template<class Impl> 333void 334DefaultFetch<Impl>::clearStates(ThreadID tid) 335{ 336 fetchStatus[tid] = Running; 337 pc[tid] = cpu->pcState(tid); 338 fetchOffset[tid] = 0; 339 macroop[tid] = NULL; 340 delayedCommit[tid] = false; 341 memReq[tid] = NULL; 342 stalls[tid].decode = false; 343 stalls[tid].drain = false; 344 fetchBufferPC[tid] = 0; 345 fetchBufferValid[tid] = false; 346 fetchQueue[tid].clear(); 347 348 // TODO not sure what to do with priorityList for now 349 // priorityList.push_back(tid); 350} 351 352template<class Impl> 353void 354DefaultFetch<Impl>::resetStage() 355{ 356 numInst = 0; 357 interruptPending = false; 358 cacheBlocked = false; 359 360 priorityList.clear(); 361 362 // Setup PC and nextPC with initial state. 363 for (ThreadID tid = 0; tid < numThreads; ++tid) { 364 fetchStatus[tid] = Running; 365 pc[tid] = cpu->pcState(tid); 366 fetchOffset[tid] = 0; 367 macroop[tid] = NULL; 368 369 delayedCommit[tid] = false; 370 memReq[tid] = NULL; 371 372 stalls[tid].decode = false; 373 stalls[tid].drain = false; 374 375 fetchBufferPC[tid] = 0; 376 fetchBufferValid[tid] = false; 377 378 fetchQueue[tid].clear(); 379 380 priorityList.push_back(tid); 381 } 382 383 wroteToTimeBuffer = false; 384 _status = Inactive; 385} 386 387template<class Impl> 388void 389DefaultFetch<Impl>::processCacheCompletion(PacketPtr pkt) 390{ 391 ThreadID tid = cpu->contextToThread(pkt->req->contextId()); 392 393 DPRINTF(Fetch, "[tid:%i] Waking up from cache miss.\n", tid); 394 assert(!cpu->switchedOut()); 395 396 // Only change the status if it's still waiting on the icache access 397 // to return. 398 if (fetchStatus[tid] != IcacheWaitResponse || 399 pkt->req != memReq[tid]) { 400 ++fetchIcacheSquashes; 401 delete pkt; 402 return; 403 } 404 405 memcpy(fetchBuffer[tid], pkt->getConstPtr<uint8_t>(), fetchBufferSize); 406 fetchBufferValid[tid] = true; 407 408 // Wake up the CPU (if it went to sleep and was waiting on 409 // this completion event). 410 cpu->wakeCPU(); 411 412 DPRINTF(Activity, "[tid:%i] Activating fetch due to cache completion\n", 413 tid); 414 415 switchToActive(); 416 417 // Only switch to IcacheAccessComplete if we're not stalled as well. 418 if (checkStall(tid)) { 419 fetchStatus[tid] = Blocked; 420 } else { 421 fetchStatus[tid] = IcacheAccessComplete; 422 } 423 424 pkt->req->setAccessLatency(); 425 cpu->ppInstAccessComplete->notify(pkt); 426 // Reset the mem req to NULL. 427 delete pkt; 428 memReq[tid] = NULL; 429} 430 431template <class Impl> 432void 433DefaultFetch<Impl>::drainResume() 434{ 435 for (ThreadID i = 0; i < numThreads; ++i) { 436 stalls[i].decode = false; 437 stalls[i].drain = false; 438 } 439} 440 441template <class Impl> 442void 443DefaultFetch<Impl>::drainSanityCheck() const 444{ 445 assert(isDrained()); 446 assert(retryPkt == NULL); 447 assert(retryTid == InvalidThreadID); 448 assert(!cacheBlocked); 449 assert(!interruptPending); 450 451 for (ThreadID i = 0; i < numThreads; ++i) { 452 assert(!memReq[i]); 453 assert(fetchStatus[i] == Idle || stalls[i].drain); 454 } 455 456 branchPred->drainSanityCheck(); 457} 458 459template <class Impl> 460bool 461DefaultFetch<Impl>::isDrained() const 462{ 463 /* Make sure that threads are either idle of that the commit stage 464 * has signaled that draining has completed by setting the drain 465 * stall flag. This effectively forces the pipeline to be disabled 466 * until the whole system is drained (simulation may continue to 467 * drain other components). 468 */ 469 for (ThreadID i = 0; i < numThreads; ++i) { 470 // Verify fetch queues are drained 471 if (!fetchQueue[i].empty()) 472 return false; 473 474 // Return false if not idle or drain stalled 475 if (fetchStatus[i] != Idle) { 476 if (fetchStatus[i] == Blocked && stalls[i].drain) 477 continue; 478 else 479 return false; 480 } 481 } 482 483 /* The pipeline might start up again in the middle of the drain 484 * cycle if the finish translation event is scheduled, so make 485 * sure that's not the case. 486 */ 487 return !finishTranslationEvent.scheduled(); 488} 489 490template <class Impl> 491void 492DefaultFetch<Impl>::takeOverFrom() 493{ 494 assert(cpu->getInstPort().isConnected()); 495 resetStage(); 496 497} 498 499template <class Impl> 500void 501DefaultFetch<Impl>::drainStall(ThreadID tid) 502{ 503 assert(cpu->isDraining()); 504 assert(!stalls[tid].drain); 505 DPRINTF(Drain, "%i: Thread drained.\n", tid); 506 stalls[tid].drain = true; 507} 508 509template <class Impl> 510void 511DefaultFetch<Impl>::wakeFromQuiesce() 512{ 513 DPRINTF(Fetch, "Waking up from quiesce\n"); 514 // Hopefully this is safe 515 // @todo: Allow other threads to wake from quiesce. 516 fetchStatus[0] = Running; 517} 518 519template <class Impl> 520inline void 521DefaultFetch<Impl>::switchToActive() 522{ 523 if (_status == Inactive) { 524 DPRINTF(Activity, "Activating stage.\n"); 525 526 cpu->activateStage(O3CPU::FetchIdx); 527 528 _status = Active; 529 } 530} 531 532template <class Impl> 533inline void 534DefaultFetch<Impl>::switchToInactive() 535{ 536 if (_status == Active) { 537 DPRINTF(Activity, "Deactivating stage.\n"); 538 539 cpu->deactivateStage(O3CPU::FetchIdx); 540 541 _status = Inactive; 542 } 543} 544 545template <class Impl> 546void 547DefaultFetch<Impl>::deactivateThread(ThreadID tid) 548{ 549 // Update priority list 550 auto thread_it = std::find(priorityList.begin(), priorityList.end(), tid); 551 if (thread_it != priorityList.end()) { 552 priorityList.erase(thread_it); 553 } 554} 555 556template <class Impl> 557bool 558DefaultFetch<Impl>::lookupAndUpdateNextPC( 559 const DynInstPtr &inst, TheISA::PCState &nextPC) 560{ 561 // Do branch prediction check here. 562 // A bit of a misnomer...next_PC is actually the current PC until 563 // this function updates it. 564 bool predict_taken; 565 566 if (!inst->isControl()) { 567 TheISA::advancePC(nextPC, inst->staticInst); 568 inst->setPredTarg(nextPC); 569 inst->setPredTaken(false); 570 return false; 571 } 572 573 ThreadID tid = inst->threadNumber; 574 predict_taken = branchPred->predict(inst->staticInst, inst->seqNum, 575 nextPC, tid); 576 577 if (predict_taken) { 578 DPRINTF(Fetch, "[tid:%i] [sn:%llu] Branch at PC %#x " 579 "predicted to be taken to %s\n", 580 tid, inst->seqNum, inst->pcState().instAddr(), nextPC); 581 } else { 582 DPRINTF(Fetch, "[tid:%i] [sn:%llu] Branch at PC %#x " 583 "predicted to be not taken\n", 584 tid, inst->seqNum, inst->pcState().instAddr()); 585 } 586 587 DPRINTF(Fetch, "[tid:%i] [sn:%llu] Branch at PC %#x " 588 "predicted to go to %s\n", 589 tid, inst->seqNum, inst->pcState().instAddr(), nextPC); 590 inst->setPredTarg(nextPC); 591 inst->setPredTaken(predict_taken); 592 593 ++fetchedBranches; 594 595 if (predict_taken) { 596 ++predictedBranches; 597 } 598 599 return predict_taken; 600} 601 602template <class Impl> 603bool 604DefaultFetch<Impl>::fetchCacheLine(Addr vaddr, ThreadID tid, Addr pc) 605{ 606 Fault fault = NoFault; 607 608 assert(!cpu->switchedOut()); 609 610 // @todo: not sure if these should block translation. 611 //AlphaDep 612 if (cacheBlocked) { 613 DPRINTF(Fetch, "[tid:%i] Can't fetch cache line, cache blocked\n", 614 tid); 615 return false; 616 } else if (checkInterrupt(pc) && !delayedCommit[tid]) { 617 // Hold off fetch from getting new instructions when: 618 // Cache is blocked, or 619 // while an interrupt is pending and we're not in PAL mode, or 620 // fetch is switched out. 621 DPRINTF(Fetch, "[tid:%i] Can't fetch cache line, interrupt pending\n", 622 tid); 623 return false; 624 } 625 626 // Align the fetch address to the start of a fetch buffer segment. 627 Addr fetchBufferBlockPC = fetchBufferAlignPC(vaddr); 628 629 DPRINTF(Fetch, "[tid:%i] Fetching cache line %#x for addr %#x\n", 630 tid, fetchBufferBlockPC, vaddr); 631 632 // Setup the memReq to do a read of the first instruction's address. 633 // Set the appropriate read size and flags as well. 634 // Build request here. 635 RequestPtr mem_req = std::make_shared<Request>( 636 tid, fetchBufferBlockPC, fetchBufferSize, 637 Request::INST_FETCH, cpu->instMasterId(), pc, 638 cpu->thread[tid]->contextId()); 639 640 mem_req->taskId(cpu->taskId()); 641 642 memReq[tid] = mem_req; 643 644 // Initiate translation of the icache block 645 fetchStatus[tid] = ItlbWait; 646 FetchTranslation *trans = new FetchTranslation(this); 647 cpu->itb->translateTiming(mem_req, cpu->thread[tid]->getTC(), 648 trans, BaseTLB::Execute); 649 return true; 650} 651 652template <class Impl> 653void 654DefaultFetch<Impl>::finishTranslation(const Fault &fault, 655 const RequestPtr &mem_req) 656{ 657 ThreadID tid = cpu->contextToThread(mem_req->contextId()); 658 Addr fetchBufferBlockPC = mem_req->getVaddr(); 659 660 assert(!cpu->switchedOut()); 661 662 // Wake up CPU if it was idle 663 cpu->wakeCPU(); 664 665 if (fetchStatus[tid] != ItlbWait || mem_req != memReq[tid] || 666 mem_req->getVaddr() != memReq[tid]->getVaddr()) { 667 DPRINTF(Fetch, "[tid:%i] Ignoring itlb completed after squash\n", 668 tid); 669 ++fetchTlbSquashes; 670 return; 671 } 672 673 674 // If translation was successful, attempt to read the icache block. 675 if (fault == NoFault) { 676 // Check that we're not going off into random memory 677 // If we have, just wait around for commit to squash something and put 678 // us on the right track 679 if (!cpu->system->isMemAddr(mem_req->getPaddr())) { 680 warn("Address %#x is outside of physical memory, stopping fetch\n", 681 mem_req->getPaddr()); 682 fetchStatus[tid] = NoGoodAddr; 683 memReq[tid] = NULL; 684 return; 685 } 686 687 // Build packet here. 688 PacketPtr data_pkt = new Packet(mem_req, MemCmd::ReadReq); 689 data_pkt->dataDynamic(new uint8_t[fetchBufferSize]); 690 691 fetchBufferPC[tid] = fetchBufferBlockPC; 692 fetchBufferValid[tid] = false; 693 DPRINTF(Fetch, "Fetch: Doing instruction read.\n"); 694 695 fetchedCacheLines++; 696 697 // Access the cache. 698 if (!icachePort.sendTimingReq(data_pkt)) { 699 assert(retryPkt == NULL); 700 assert(retryTid == InvalidThreadID); 701 DPRINTF(Fetch, "[tid:%i] Out of MSHRs!\n", tid); 702 703 fetchStatus[tid] = IcacheWaitRetry; 704 retryPkt = data_pkt; 705 retryTid = tid; 706 cacheBlocked = true; 707 } else { 708 DPRINTF(Fetch, "[tid:%i] Doing Icache access.\n", tid); 709 DPRINTF(Activity, "[tid:%i] Activity: Waiting on I-cache " 710 "response.\n", tid); 711 lastIcacheStall[tid] = curTick(); 712 fetchStatus[tid] = IcacheWaitResponse; 713 // Notify Fetch Request probe when a packet containing a fetch 714 // request is successfully sent 715 ppFetchRequestSent->notify(mem_req); 716 } 717 } else { 718 // Don't send an instruction to decode if we can't handle it. 719 if (!(numInst < fetchWidth) || !(fetchQueue[tid].size() < fetchQueueSize)) { 720 assert(!finishTranslationEvent.scheduled()); 721 finishTranslationEvent.setFault(fault); 722 finishTranslationEvent.setReq(mem_req); 723 cpu->schedule(finishTranslationEvent, 724 cpu->clockEdge(Cycles(1))); 725 return; 726 } 727 DPRINTF(Fetch, "[tid:%i] Got back req with addr %#x but expected %#x\n", 728 tid, mem_req->getVaddr(), memReq[tid]->getVaddr()); 729 // Translation faulted, icache request won't be sent. 730 memReq[tid] = NULL; 731 732 // Send the fault to commit. This thread will not do anything 733 // until commit handles the fault. The only other way it can 734 // wake up is if a squash comes along and changes the PC. 735 TheISA::PCState fetchPC = pc[tid]; 736 737 DPRINTF(Fetch, "[tid:%i] Translation faulted, building noop.\n", tid); 738 // We will use a nop in ordier to carry the fault. 739 DynInstPtr instruction = buildInst(tid, StaticInst::nopStaticInstPtr, 740 NULL, fetchPC, fetchPC, false); 741 instruction->setNotAnInst(); 742 743 instruction->setPredTarg(fetchPC); 744 instruction->fault = fault; 745 wroteToTimeBuffer = true; 746 747 DPRINTF(Activity, "Activity this cycle.\n"); 748 cpu->activityThisCycle(); 749 750 fetchStatus[tid] = TrapPending; 751 752 DPRINTF(Fetch, "[tid:%i] Blocked, need to handle the trap.\n", tid); 753 DPRINTF(Fetch, "[tid:%i] fault (%s) detected @ PC %s.\n", 754 tid, fault->name(), pc[tid]); 755 } 756 _status = updateFetchStatus(); 757} 758 759template <class Impl> 760inline void 761DefaultFetch<Impl>::doSquash(const TheISA::PCState &newPC, 762 const DynInstPtr squashInst, ThreadID tid) 763{ 764 DPRINTF(Fetch, "[tid:%i] Squashing, setting PC to: %s.\n", 765 tid, newPC); 766 767 pc[tid] = newPC; 768 fetchOffset[tid] = 0; 769 if (squashInst && squashInst->pcState().instAddr() == newPC.instAddr()) 770 macroop[tid] = squashInst->macroop; 771 else 772 macroop[tid] = NULL; 773 decoder[tid]->reset(); 774 775 // Clear the icache miss if it's outstanding. 776 if (fetchStatus[tid] == IcacheWaitResponse) { 777 DPRINTF(Fetch, "[tid:%i] Squashing outstanding Icache miss.\n", 778 tid); 779 memReq[tid] = NULL; 780 } else if (fetchStatus[tid] == ItlbWait) { 781 DPRINTF(Fetch, "[tid:%i] Squashing outstanding ITLB miss.\n", 782 tid); 783 memReq[tid] = NULL; 784 } 785 786 // Get rid of the retrying packet if it was from this thread. 787 if (retryTid == tid) { 788 assert(cacheBlocked); 789 if (retryPkt) { 790 delete retryPkt; 791 } 792 retryPkt = NULL; 793 retryTid = InvalidThreadID; 794 } 795 796 fetchStatus[tid] = Squashing; 797 798 // Empty fetch queue 799 fetchQueue[tid].clear(); 800 801 // microops are being squashed, it is not known wheather the 802 // youngest non-squashed microop was marked delayed commit 803 // or not. Setting the flag to true ensures that the 804 // interrupts are not handled when they cannot be, though 805 // some opportunities to handle interrupts may be missed. 806 delayedCommit[tid] = true; 807 808 ++fetchSquashCycles; 809} 810 811template<class Impl> 812void 813DefaultFetch<Impl>::squashFromDecode(const TheISA::PCState &newPC, 814 const DynInstPtr squashInst, 815 const InstSeqNum seq_num, ThreadID tid) 816{ 817 DPRINTF(Fetch, "[tid:%i] Squashing from decode.\n", tid); 818 819 doSquash(newPC, squashInst, tid); 820 821 // Tell the CPU to remove any instructions that are in flight between 822 // fetch and decode. 823 cpu->removeInstsUntil(seq_num, tid); 824} 825 826template<class Impl> 827bool 828DefaultFetch<Impl>::checkStall(ThreadID tid) const 829{ 830 bool ret_val = false; 831 832 if (stalls[tid].drain) { 833 assert(cpu->isDraining()); 834 DPRINTF(Fetch,"[tid:%i] Drain stall detected.\n",tid); 835 ret_val = true; 836 } 837 838 return ret_val; 839} 840 841template<class Impl> 842typename DefaultFetch<Impl>::FetchStatus 843DefaultFetch<Impl>::updateFetchStatus() 844{ 845 //Check Running 846 list<ThreadID>::iterator threads = activeThreads->begin(); 847 list<ThreadID>::iterator end = activeThreads->end(); 848 849 while (threads != end) { 850 ThreadID tid = *threads++; 851 852 if (fetchStatus[tid] == Running || 853 fetchStatus[tid] == Squashing || 854 fetchStatus[tid] == IcacheAccessComplete) { 855 856 if (_status == Inactive) { 857 DPRINTF(Activity, "[tid:%i] Activating stage.\n",tid); 858 859 if (fetchStatus[tid] == IcacheAccessComplete) { 860 DPRINTF(Activity, "[tid:%i] Activating fetch due to cache" 861 "completion\n",tid); 862 } 863 864 cpu->activateStage(O3CPU::FetchIdx); 865 } 866 867 return Active; 868 } 869 } 870 871 // Stage is switching from active to inactive, notify CPU of it. 872 if (_status == Active) { 873 DPRINTF(Activity, "Deactivating stage.\n"); 874 875 cpu->deactivateStage(O3CPU::FetchIdx); 876 } 877 878 return Inactive; 879} 880 881template <class Impl> 882void 883DefaultFetch<Impl>::squash(const TheISA::PCState &newPC, 884 const InstSeqNum seq_num, DynInstPtr squashInst, 885 ThreadID tid) 886{ 887 DPRINTF(Fetch, "[tid:%i] Squash from commit.\n", tid); 888 889 doSquash(newPC, squashInst, tid); 890 891 // Tell the CPU to remove any instructions that are not in the ROB. 892 cpu->removeInstsNotInROB(tid); 893} 894 895template <class Impl> 896void 897DefaultFetch<Impl>::tick() 898{ 899 list<ThreadID>::iterator threads = activeThreads->begin(); 900 list<ThreadID>::iterator end = activeThreads->end(); 901 bool status_change = false; 902 903 wroteToTimeBuffer = false; 904 905 for (ThreadID i = 0; i < numThreads; ++i) { 906 issuePipelinedIfetch[i] = false; 907 } 908 909 while (threads != end) { 910 ThreadID tid = *threads++; 911 912 // Check the signals for each thread to determine the proper status 913 // for each thread. 914 bool updated_status = checkSignalsAndUpdate(tid); 915 status_change = status_change || updated_status; 916 } 917 918 DPRINTF(Fetch, "Running stage.\n"); 919 920 if (FullSystem) { 921 if (fromCommit->commitInfo[0].interruptPending) { 922 interruptPending = true; 923 } 924 925 if (fromCommit->commitInfo[0].clearInterrupt) { 926 interruptPending = false; 927 } 928 } 929 930 for (threadFetched = 0; threadFetched < numFetchingThreads; 931 threadFetched++) { 932 // Fetch each of the actively fetching threads. 933 fetch(status_change); 934 } 935 936 // Record number of instructions fetched this cycle for distribution. 937 fetchNisnDist.sample(numInst); 938 939 if (status_change) { 940 // Change the fetch stage status if there was a status change. 941 _status = updateFetchStatus(); 942 } 943 944 // Issue the next I-cache request if possible. 945 for (ThreadID i = 0; i < numThreads; ++i) { 946 if (issuePipelinedIfetch[i]) { 947 pipelineIcacheAccesses(i); 948 } 949 } 950 951 // Send instructions enqueued into the fetch queue to decode. 952 // Limit rate by fetchWidth. Stall if decode is stalled. 953 unsigned insts_to_decode = 0; 954 unsigned available_insts = 0; 955 956 for (auto tid : *activeThreads) { 957 if (!stalls[tid].decode) { 958 available_insts += fetchQueue[tid].size(); 959 } 960 } 961 962 // Pick a random thread to start trying to grab instructions from 963 auto tid_itr = activeThreads->begin(); 964 std::advance(tid_itr, random_mt.random<uint8_t>(0, activeThreads->size() - 1)); 965 966 while (available_insts != 0 && insts_to_decode < decodeWidth) { 967 ThreadID tid = *tid_itr; 968 if (!stalls[tid].decode && !fetchQueue[tid].empty()) { 969 const auto& inst = fetchQueue[tid].front(); 970 toDecode->insts[toDecode->size++] = inst; 971 DPRINTF(Fetch, "[tid:%i] [sn:%llu] Sending instruction to decode " 972 "from fetch queue. Fetch queue size: %i.\n", 973 tid, inst->seqNum, fetchQueue[tid].size()); 974 975 wroteToTimeBuffer = true; 976 fetchQueue[tid].pop_front(); 977 insts_to_decode++; 978 available_insts--; 979 } 980 981 tid_itr++; 982 // Wrap around if at end of active threads list 983 if (tid_itr == activeThreads->end()) 984 tid_itr = activeThreads->begin(); 985 } 986 987 // If there was activity this cycle, inform the CPU of it. 988 if (wroteToTimeBuffer) { 989 DPRINTF(Activity, "Activity this cycle.\n"); 990 cpu->activityThisCycle(); 991 } 992 993 // Reset the number of the instruction we've fetched. 994 numInst = 0; 995} 996 997template <class Impl> 998bool 999DefaultFetch<Impl>::checkSignalsAndUpdate(ThreadID tid) 1000{ 1001 // Update the per thread stall statuses. 1002 if (fromDecode->decodeBlock[tid]) { 1003 stalls[tid].decode = true; 1004 } 1005 1006 if (fromDecode->decodeUnblock[tid]) { 1007 assert(stalls[tid].decode); 1008 assert(!fromDecode->decodeBlock[tid]); 1009 stalls[tid].decode = false; 1010 } 1011 1012 // Check squash signals from commit. 1013 if (fromCommit->commitInfo[tid].squash) { 1014 1015 DPRINTF(Fetch, "[tid:%i] Squashing instructions due to squash " 1016 "from commit.\n",tid); 1017 // In any case, squash. 1018 squash(fromCommit->commitInfo[tid].pc, 1019 fromCommit->commitInfo[tid].doneSeqNum, 1020 fromCommit->commitInfo[tid].squashInst, tid); 1021 1022 // If it was a branch mispredict on a control instruction, update the 1023 // branch predictor with that instruction, otherwise just kill the 1024 // invalid state we generated in after sequence number 1025 if (fromCommit->commitInfo[tid].mispredictInst && 1026 fromCommit->commitInfo[tid].mispredictInst->isControl()) { 1027 branchPred->squash(fromCommit->commitInfo[tid].doneSeqNum, 1028 fromCommit->commitInfo[tid].pc, 1029 fromCommit->commitInfo[tid].branchTaken, 1030 tid); 1031 } else { 1032 branchPred->squash(fromCommit->commitInfo[tid].doneSeqNum, 1033 tid); 1034 } 1035 1036 return true; 1037 } else if (fromCommit->commitInfo[tid].doneSeqNum) { 1038 // Update the branch predictor if it wasn't a squashed instruction 1039 // that was broadcasted. 1040 branchPred->update(fromCommit->commitInfo[tid].doneSeqNum, tid); 1041 } 1042 1043 // Check squash signals from decode. 1044 if (fromDecode->decodeInfo[tid].squash) { 1045 DPRINTF(Fetch, "[tid:%i] Squashing instructions due to squash " 1046 "from decode.\n",tid); 1047 1048 // Update the branch predictor. 1049 if (fromDecode->decodeInfo[tid].branchMispredict) { 1050 branchPred->squash(fromDecode->decodeInfo[tid].doneSeqNum, 1051 fromDecode->decodeInfo[tid].nextPC, 1052 fromDecode->decodeInfo[tid].branchTaken, 1053 tid); 1054 } else { 1055 branchPred->squash(fromDecode->decodeInfo[tid].doneSeqNum, 1056 tid); 1057 } 1058 1059 if (fetchStatus[tid] != Squashing) { 1060 1061 DPRINTF(Fetch, "Squashing from decode with PC = %s\n", 1062 fromDecode->decodeInfo[tid].nextPC); 1063 // Squash unless we're already squashing 1064 squashFromDecode(fromDecode->decodeInfo[tid].nextPC, 1065 fromDecode->decodeInfo[tid].squashInst, 1066 fromDecode->decodeInfo[tid].doneSeqNum, 1067 tid); 1068 1069 return true; 1070 } 1071 } 1072 1073 if (checkStall(tid) && 1074 fetchStatus[tid] != IcacheWaitResponse && 1075 fetchStatus[tid] != IcacheWaitRetry && 1076 fetchStatus[tid] != ItlbWait && 1077 fetchStatus[tid] != QuiescePending) { 1078 DPRINTF(Fetch, "[tid:%i] Setting to blocked\n",tid); 1079 1080 fetchStatus[tid] = Blocked; 1081 1082 return true; 1083 } 1084 1085 if (fetchStatus[tid] == Blocked || 1086 fetchStatus[tid] == Squashing) { 1087 // Switch status to running if fetch isn't being told to block or 1088 // squash this cycle. 1089 DPRINTF(Fetch, "[tid:%i] Done squashing, switching to running.\n", 1090 tid); 1091 1092 fetchStatus[tid] = Running; 1093 1094 return true; 1095 } 1096 1097 // If we've reached this point, we have not gotten any signals that 1098 // cause fetch to change its status. Fetch remains the same as before. 1099 return false; 1100} 1101 1102template<class Impl> 1103typename Impl::DynInstPtr 1104DefaultFetch<Impl>::buildInst(ThreadID tid, StaticInstPtr staticInst, 1105 StaticInstPtr curMacroop, TheISA::PCState thisPC, 1106 TheISA::PCState nextPC, bool trace) 1107{ 1108 // Get a sequence number. 1109 InstSeqNum seq = cpu->getAndIncrementInstSeq(); 1110 1111 // Create a new DynInst from the instruction fetched. 1112 DynInstPtr instruction = 1113 new DynInst(staticInst, curMacroop, thisPC, nextPC, seq, cpu); 1114 instruction->setTid(tid); 1115 1116 instruction->setASID(tid); 1117 1118 instruction->setThreadState(cpu->thread[tid]); 1119 1120 DPRINTF(Fetch, "[tid:%i] Instruction PC %#x (%d) created " 1121 "[sn:%lli].\n", tid, thisPC.instAddr(), 1122 thisPC.microPC(), seq); 1123 1124 DPRINTF(Fetch, "[tid:%i] Instruction is: %s\n", tid, 1125 instruction->staticInst-> 1126 disassemble(thisPC.instAddr())); 1127 1128#if TRACING_ON 1129 if (trace) { 1130 instruction->traceData = 1131 cpu->getTracer()->getInstRecord(curTick(), cpu->tcBase(tid), 1132 instruction->staticInst, thisPC, curMacroop); 1133 } 1134#else 1135 instruction->traceData = NULL; 1136#endif 1137 1138 // Add instruction to the CPU's list of instructions. 1139 instruction->setInstListIt(cpu->addInst(instruction)); 1140 1141 // Write the instruction to the first slot in the queue 1142 // that heads to decode. 1143 assert(numInst < fetchWidth); 1144 fetchQueue[tid].push_back(instruction); 1145 assert(fetchQueue[tid].size() <= fetchQueueSize); 1146 DPRINTF(Fetch, "[tid:%i] Fetch queue entry created (%i/%i).\n", 1147 tid, fetchQueue[tid].size(), fetchQueueSize); 1148 //toDecode->insts[toDecode->size++] = instruction; 1149 1150 // Keep track of if we can take an interrupt at this boundary 1151 delayedCommit[tid] = instruction->isDelayedCommit(); 1152 1153 return instruction; 1154} 1155 1156template<class Impl> 1157void 1158DefaultFetch<Impl>::fetch(bool &status_change) 1159{ 1160 ////////////////////////////////////////// 1161 // Start actual fetch 1162 ////////////////////////////////////////// 1163 ThreadID tid = getFetchingThread(); 1164 1165 assert(!cpu->switchedOut()); 1166 1167 if (tid == InvalidThreadID) { 1168 // Breaks looping condition in tick() 1169 threadFetched = numFetchingThreads; 1170 1171 if (numThreads == 1) { // @todo Per-thread stats 1172 profileStall(0); 1173 } 1174 1175 return; 1176 } 1177 1178 DPRINTF(Fetch, "Attempting to fetch from [tid:%i]\n", tid); 1179 1180 // The current PC. 1181 TheISA::PCState thisPC = pc[tid]; 1182 1183 Addr pcOffset = fetchOffset[tid]; 1184 Addr fetchAddr = (thisPC.instAddr() + pcOffset) & BaseCPU::PCMask; 1185 1186 bool inRom = isRomMicroPC(thisPC.microPC()); 1187 1188 // If returning from the delay of a cache miss, then update the status 1189 // to running, otherwise do the cache access. Possibly move this up 1190 // to tick() function. 1191 if (fetchStatus[tid] == IcacheAccessComplete) { 1192 DPRINTF(Fetch, "[tid:%i] Icache miss is complete.\n", tid); 1193 1194 fetchStatus[tid] = Running; 1195 status_change = true; 1196 } else if (fetchStatus[tid] == Running) { 1197 // Align the fetch PC so its at the start of a fetch buffer segment. 1198 Addr fetchBufferBlockPC = fetchBufferAlignPC(fetchAddr); 1199 1200 // If buffer is no longer valid or fetchAddr has moved to point 1201 // to the next cache block, AND we have no remaining ucode 1202 // from a macro-op, then start fetch from icache. 1203 if (!(fetchBufferValid[tid] && fetchBufferBlockPC == fetchBufferPC[tid]) 1204 && !inRom && !macroop[tid]) { 1205 DPRINTF(Fetch, "[tid:%i] Attempting to translate and read " 1206 "instruction, starting at PC %s.\n", tid, thisPC); 1207 1208 fetchCacheLine(fetchAddr, tid, thisPC.instAddr()); 1209 1210 if (fetchStatus[tid] == IcacheWaitResponse) 1211 ++icacheStallCycles; 1212 else if (fetchStatus[tid] == ItlbWait) 1213 ++fetchTlbCycles; 1214 else 1215 ++fetchMiscStallCycles; 1216 return; 1217 } else if ((checkInterrupt(thisPC.instAddr()) && !delayedCommit[tid])) { 1218 // Stall CPU if an interrupt is posted and we're not issuing 1219 // an delayed commit micro-op currently (delayed commit instructions 1220 // are not interruptable by interrupts, only faults) 1221 ++fetchMiscStallCycles; 1222 DPRINTF(Fetch, "[tid:%i] Fetch is stalled!\n", tid); 1223 return; 1224 } 1225 } else { 1226 if (fetchStatus[tid] == Idle) { 1227 ++fetchIdleCycles; 1228 DPRINTF(Fetch, "[tid:%i] Fetch is idle!\n", tid); 1229 } 1230 1231 // Status is Idle, so fetch should do nothing. 1232 return; 1233 } 1234 1235 ++fetchCycles; 1236 1237 TheISA::PCState nextPC = thisPC; 1238 1239 StaticInstPtr staticInst = NULL; 1240 StaticInstPtr curMacroop = macroop[tid]; 1241 1242 // If the read of the first instruction was successful, then grab the 1243 // instructions from the rest of the cache line and put them into the 1244 // queue heading to decode. 1245 1246 DPRINTF(Fetch, "[tid:%i] Adding instructions to queue to " 1247 "decode.\n", tid); 1248 1249 // Need to keep track of whether or not a predicted branch 1250 // ended this fetch block. 1251 bool predictedBranch = false; 1252 1253 // Need to halt fetch if quiesce instruction detected 1254 bool quiesce = false; 1255 1256 TheISA::MachInst *cacheInsts = 1257 reinterpret_cast<TheISA::MachInst *>(fetchBuffer[tid]); 1258 1259 const unsigned numInsts = fetchBufferSize / instSize; 1260 unsigned blkOffset = (fetchAddr - fetchBufferPC[tid]) / instSize; 1261 1262 // Loop through instruction memory from the cache. 1263 // Keep issuing while fetchWidth is available and branch is not 1264 // predicted taken 1265 while (numInst < fetchWidth && fetchQueue[tid].size() < fetchQueueSize 1266 && !predictedBranch && !quiesce) { 1267 // We need to process more memory if we aren't going to get a 1268 // StaticInst from the rom, the current macroop, or what's already 1269 // in the decoder. 1270 bool needMem = !inRom && !curMacroop && 1271 !decoder[tid]->instReady(); 1272 fetchAddr = (thisPC.instAddr() + pcOffset) & BaseCPU::PCMask; 1273 Addr fetchBufferBlockPC = fetchBufferAlignPC(fetchAddr); 1274 1275 if (needMem) { 1276 // If buffer is no longer valid or fetchAddr has moved to point 1277 // to the next cache block then start fetch from icache. 1278 if (!fetchBufferValid[tid] || 1279 fetchBufferBlockPC != fetchBufferPC[tid]) 1280 break; 1281 1282 if (blkOffset >= numInsts) { 1283 // We need to process more memory, but we've run out of the 1284 // current block. 1285 break; 1286 } 1287 1288 MachInst inst = TheISA::gtoh(cacheInsts[blkOffset]); 1289 decoder[tid]->moreBytes(thisPC, fetchAddr, inst); 1290 1291 if (decoder[tid]->needMoreBytes()) { 1292 blkOffset++; 1293 fetchAddr += instSize; 1294 pcOffset += instSize; 1295 } 1296 } 1297 1298 // Extract as many instructions and/or microops as we can from 1299 // the memory we've processed so far. 1300 do { 1301 if (!(curMacroop || inRom)) { 1302 if (decoder[tid]->instReady()) { 1303 staticInst = decoder[tid]->decode(thisPC); 1304 1305 // Increment stat of fetched instructions. 1306 ++fetchedInsts; 1307 1308 if (staticInst->isMacroop()) { 1309 curMacroop = staticInst; 1310 } else { 1311 pcOffset = 0; 1312 } 1313 } else { 1314 // We need more bytes for this instruction so blkOffset and 1315 // pcOffset will be updated 1316 break; 1317 } 1318 } 1319 // Whether we're moving to a new macroop because we're at the 1320 // end of the current one, or the branch predictor incorrectly 1321 // thinks we are... 1322 bool newMacro = false; 1323 if (curMacroop || inRom) { 1324 if (inRom) { 1325 staticInst = cpu->microcodeRom.fetchMicroop( 1326 thisPC.microPC(), curMacroop); 1327 } else { 1328 staticInst = curMacroop->fetchMicroop(thisPC.microPC()); 1329 } 1330 newMacro |= staticInst->isLastMicroop(); 1331 } 1332 1333 DynInstPtr instruction = 1334 buildInst(tid, staticInst, curMacroop, 1335 thisPC, nextPC, true); 1336 1337 ppFetch->notify(instruction); 1338 numInst++; 1339 1340#if TRACING_ON 1341 if (DTRACE(O3PipeView)) { 1342 instruction->fetchTick = curTick(); 1343 } 1344#endif 1345 1346 nextPC = thisPC; 1347 1348 // If we're branching after this instruction, quit fetching 1349 // from the same block. 1350 predictedBranch |= thisPC.branching(); 1351 predictedBranch |= 1352 lookupAndUpdateNextPC(instruction, nextPC); 1353 if (predictedBranch) { 1354 DPRINTF(Fetch, "Branch detected with PC = %s\n", thisPC); 1355 } 1356 1357 newMacro |= thisPC.instAddr() != nextPC.instAddr(); 1358 1359 // Move to the next instruction, unless we have a branch. 1360 thisPC = nextPC; 1361 inRom = isRomMicroPC(thisPC.microPC()); 1362 1363 if (newMacro) { 1364 fetchAddr = thisPC.instAddr() & BaseCPU::PCMask; 1365 blkOffset = (fetchAddr - fetchBufferPC[tid]) / instSize; 1366 pcOffset = 0; 1367 curMacroop = NULL; 1368 } 1369 1370 if (instruction->isQuiesce()) { 1371 DPRINTF(Fetch, 1372 "Quiesce instruction encountered, halting fetch!\n"); 1373 fetchStatus[tid] = QuiescePending; 1374 status_change = true; 1375 quiesce = true; 1376 break; 1377 } 1378 } while ((curMacroop || decoder[tid]->instReady()) && 1379 numInst < fetchWidth && 1380 fetchQueue[tid].size() < fetchQueueSize); 1381 1382 // Re-evaluate whether the next instruction to fetch is in micro-op ROM 1383 // or not. 1384 inRom = isRomMicroPC(thisPC.microPC()); 1385 } 1386 1387 if (predictedBranch) { 1388 DPRINTF(Fetch, "[tid:%i] Done fetching, predicted branch " 1389 "instruction encountered.\n", tid); 1390 } else if (numInst >= fetchWidth) { 1391 DPRINTF(Fetch, "[tid:%i] Done fetching, reached fetch bandwidth " 1392 "for this cycle.\n", tid); 1393 } else if (blkOffset >= fetchBufferSize) { 1394 DPRINTF(Fetch, "[tid:%i] Done fetching, reached the end of the" 1395 "fetch buffer.\n", tid); 1396 } 1397 1398 macroop[tid] = curMacroop; 1399 fetchOffset[tid] = pcOffset; 1400 1401 if (numInst > 0) { 1402 wroteToTimeBuffer = true; 1403 } 1404 1405 pc[tid] = thisPC; 1406 1407 // pipeline a fetch if we're crossing a fetch buffer boundary and not in 1408 // a state that would preclude fetching 1409 fetchAddr = (thisPC.instAddr() + pcOffset) & BaseCPU::PCMask; 1410 Addr fetchBufferBlockPC = fetchBufferAlignPC(fetchAddr); 1411 issuePipelinedIfetch[tid] = fetchBufferBlockPC != fetchBufferPC[tid] && 1412 fetchStatus[tid] != IcacheWaitResponse && 1413 fetchStatus[tid] != ItlbWait && 1414 fetchStatus[tid] != IcacheWaitRetry && 1415 fetchStatus[tid] != QuiescePending && 1416 !curMacroop; 1417} 1418 1419template<class Impl> 1420void 1421DefaultFetch<Impl>::recvReqRetry() 1422{ 1423 if (retryPkt != NULL) { 1424 assert(cacheBlocked); 1425 assert(retryTid != InvalidThreadID); 1426 assert(fetchStatus[retryTid] == IcacheWaitRetry); 1427 1428 if (icachePort.sendTimingReq(retryPkt)) { 1429 fetchStatus[retryTid] = IcacheWaitResponse; 1430 // Notify Fetch Request probe when a retryPkt is successfully sent. 1431 // Note that notify must be called before retryPkt is set to NULL. 1432 ppFetchRequestSent->notify(retryPkt->req); 1433 retryPkt = NULL; 1434 retryTid = InvalidThreadID; 1435 cacheBlocked = false; 1436 } 1437 } else { 1438 assert(retryTid == InvalidThreadID); 1439 // Access has been squashed since it was sent out. Just clear 1440 // the cache being blocked. 1441 cacheBlocked = false; 1442 } 1443} 1444 1445/////////////////////////////////////// 1446// // 1447// SMT FETCH POLICY MAINTAINED HERE // 1448// // 1449/////////////////////////////////////// 1450template<class Impl> 1451ThreadID 1452DefaultFetch<Impl>::getFetchingThread() 1453{ 1454 if (numThreads > 1) { 1455 switch (fetchPolicy) { 1456 case FetchPolicy::RoundRobin: 1457 return roundRobin(); 1458 case FetchPolicy::IQCount: 1459 return iqCount(); 1460 case FetchPolicy::LSQCount: 1461 return lsqCount(); 1462 case FetchPolicy::Branch: 1463 return branchCount(); 1464 default: 1465 return InvalidThreadID; 1466 } 1467 } else { 1468 list<ThreadID>::iterator thread = activeThreads->begin(); 1469 if (thread == activeThreads->end()) { 1470 return InvalidThreadID; 1471 } 1472 1473 ThreadID tid = *thread; 1474 1475 if (fetchStatus[tid] == Running || 1476 fetchStatus[tid] == IcacheAccessComplete || 1477 fetchStatus[tid] == Idle) { 1478 return tid; 1479 } else { 1480 return InvalidThreadID; 1481 } 1482 } 1483} 1484 1485 1486template<class Impl> 1487ThreadID 1488DefaultFetch<Impl>::roundRobin() 1489{ 1490 list<ThreadID>::iterator pri_iter = priorityList.begin(); 1491 list<ThreadID>::iterator end = priorityList.end(); 1492 1493 ThreadID high_pri; 1494 1495 while (pri_iter != end) { 1496 high_pri = *pri_iter; 1497 1498 assert(high_pri <= numThreads); 1499 1500 if (fetchStatus[high_pri] == Running || 1501 fetchStatus[high_pri] == IcacheAccessComplete || 1502 fetchStatus[high_pri] == Idle) { 1503 1504 priorityList.erase(pri_iter); 1505 priorityList.push_back(high_pri); 1506 1507 return high_pri; 1508 } 1509 1510 pri_iter++; 1511 } 1512 1513 return InvalidThreadID; 1514} 1515 1516template<class Impl> 1517ThreadID 1518DefaultFetch<Impl>::iqCount() 1519{ 1520 //sorted from lowest->highest 1521 std::priority_queue<unsigned,vector<unsigned>, 1522 std::greater<unsigned> > PQ; 1523 std::map<unsigned, ThreadID> threadMap; 1524 1525 list<ThreadID>::iterator threads = activeThreads->begin(); 1526 list<ThreadID>::iterator end = activeThreads->end(); 1527 1528 while (threads != end) { 1529 ThreadID tid = *threads++; 1530 unsigned iqCount = fromIEW->iewInfo[tid].iqCount; 1531 1532 //we can potentially get tid collisions if two threads 1533 //have the same iqCount, but this should be rare. 1534 PQ.push(iqCount); 1535 threadMap[iqCount] = tid; 1536 } 1537 1538 while (!PQ.empty()) { 1539 ThreadID high_pri = threadMap[PQ.top()]; 1540 1541 if (fetchStatus[high_pri] == Running || 1542 fetchStatus[high_pri] == IcacheAccessComplete || 1543 fetchStatus[high_pri] == Idle) 1544 return high_pri; 1545 else 1546 PQ.pop(); 1547 1548 } 1549 1550 return InvalidThreadID; 1551} 1552 1553template<class Impl> 1554ThreadID 1555DefaultFetch<Impl>::lsqCount() 1556{ 1557 //sorted from lowest->highest 1558 std::priority_queue<unsigned,vector<unsigned>, 1559 std::greater<unsigned> > PQ; 1560 std::map<unsigned, ThreadID> threadMap; 1561 1562 list<ThreadID>::iterator threads = activeThreads->begin(); 1563 list<ThreadID>::iterator end = activeThreads->end(); 1564 1565 while (threads != end) { 1566 ThreadID tid = *threads++; 1567 unsigned ldstqCount = fromIEW->iewInfo[tid].ldstqCount; 1568 1569 //we can potentially get tid collisions if two threads 1570 //have the same iqCount, but this should be rare. 1571 PQ.push(ldstqCount); 1572 threadMap[ldstqCount] = tid; 1573 } 1574 1575 while (!PQ.empty()) { 1576 ThreadID high_pri = threadMap[PQ.top()]; 1577 1578 if (fetchStatus[high_pri] == Running || 1579 fetchStatus[high_pri] == IcacheAccessComplete || 1580 fetchStatus[high_pri] == Idle) 1581 return high_pri; 1582 else 1583 PQ.pop(); 1584 } 1585 1586 return InvalidThreadID; 1587} 1588 1589template<class Impl> 1590ThreadID 1591DefaultFetch<Impl>::branchCount() 1592{ 1593 panic("Branch Count Fetch policy unimplemented\n"); 1594 return InvalidThreadID; 1595} 1596 1597template<class Impl> 1598void 1599DefaultFetch<Impl>::pipelineIcacheAccesses(ThreadID tid) 1600{ 1601 if (!issuePipelinedIfetch[tid]) { 1602 return; 1603 } 1604 1605 // The next PC to access. 1606 TheISA::PCState thisPC = pc[tid]; 1607 1608 if (isRomMicroPC(thisPC.microPC())) { 1609 return; 1610 } 1611 1612 Addr pcOffset = fetchOffset[tid]; 1613 Addr fetchAddr = (thisPC.instAddr() + pcOffset) & BaseCPU::PCMask; 1614 1615 // Align the fetch PC so its at the start of a fetch buffer segment. 1616 Addr fetchBufferBlockPC = fetchBufferAlignPC(fetchAddr); 1617 1618 // Unless buffer already got the block, fetch it from icache. 1619 if (!(fetchBufferValid[tid] && fetchBufferBlockPC == fetchBufferPC[tid])) { 1620 DPRINTF(Fetch, "[tid:%i] Issuing a pipelined I-cache access, " 1621 "starting at PC %s.\n", tid, thisPC); 1622 1623 fetchCacheLine(fetchAddr, tid, thisPC.instAddr()); 1624 } 1625} 1626 1627template<class Impl> 1628void 1629DefaultFetch<Impl>::profileStall(ThreadID tid) { 1630 DPRINTF(Fetch,"There are no more threads available to fetch from.\n"); 1631 1632 // @todo Per-thread stats 1633 1634 if (stalls[tid].drain) { 1635 ++fetchPendingDrainCycles; 1636 DPRINTF(Fetch, "Fetch is waiting for a drain!\n"); 1637 } else if (activeThreads->empty()) { 1638 ++fetchNoActiveThreadStallCycles; 1639 DPRINTF(Fetch, "Fetch has no active thread!\n"); 1640 } else if (fetchStatus[tid] == Blocked) { 1641 ++fetchBlockedCycles; 1642 DPRINTF(Fetch, "[tid:%i] Fetch is blocked!\n", tid); 1643 } else if (fetchStatus[tid] == Squashing) { 1644 ++fetchSquashCycles; 1645 DPRINTF(Fetch, "[tid:%i] Fetch is squashing!\n", tid); 1646 } else if (fetchStatus[tid] == IcacheWaitResponse) { 1647 ++icacheStallCycles; 1648 DPRINTF(Fetch, "[tid:%i] Fetch is waiting cache response!\n", 1649 tid); 1650 } else if (fetchStatus[tid] == ItlbWait) { 1651 ++fetchTlbCycles; 1652 DPRINTF(Fetch, "[tid:%i] Fetch is waiting ITLB walk to " 1653 "finish!\n", tid); 1654 } else if (fetchStatus[tid] == TrapPending) { 1655 ++fetchPendingTrapStallCycles; 1656 DPRINTF(Fetch, "[tid:%i] Fetch is waiting for a pending trap!\n", 1657 tid); 1658 } else if (fetchStatus[tid] == QuiescePending) { 1659 ++fetchPendingQuiesceStallCycles; 1660 DPRINTF(Fetch, "[tid:%i] Fetch is waiting for a pending quiesce " 1661 "instruction!\n", tid); 1662 } else if (fetchStatus[tid] == IcacheWaitRetry) { 1663 ++fetchIcacheWaitRetryStallCycles; 1664 DPRINTF(Fetch, "[tid:%i] Fetch is waiting for an I-cache retry!\n", 1665 tid); 1666 } else if (fetchStatus[tid] == NoGoodAddr) { 1667 DPRINTF(Fetch, "[tid:%i] Fetch predicted non-executable address\n", 1668 tid); 1669 } else { 1670 DPRINTF(Fetch, "[tid:%i] Unexpected fetch stall reason " 1671 "(Status: %i)\n", 1672 tid, fetchStatus[tid]); 1673 } 1674} 1675 1676template<class Impl> 1677bool 1678DefaultFetch<Impl>::IcachePort::recvTimingResp(PacketPtr pkt) 1679{ 1680 DPRINTF(O3CPU, "Fetch unit received timing\n"); 1681 // We shouldn't ever get a cacheable block in Modified state 1682 assert(pkt->req->isUncacheable() || 1683 !(pkt->cacheResponding() && !pkt->hasSharers())); 1684 fetch->processCacheCompletion(pkt); 1685 1686 return true; 1687} 1688 1689template<class Impl> 1690void 1691DefaultFetch<Impl>::IcachePort::recvReqRetry() 1692{ 1693 fetch->recvReqRetry(); 1694} 1695 1696#endif//__CPU_O3_FETCH_IMPL_HH__ 1697