fetch_impl.hh revision 11165:d90aec9435bd
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/isa_traits.hh" 55#include "arch/tlb.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/fetch.hh" 64#include "cpu/exetrace.hh" 65#include "debug/Activity.hh" 66#include "debug/Drain.hh" 67#include "debug/Fetch.hh" 68#include "debug/O3PipeView.hh" 69#include "mem/packet.hh" 70#include "params/DerivO3CPU.hh" 71#include "sim/byteswap.hh" 72#include "sim/core.hh" 73#include "sim/eventq.hh" 74#include "sim/full_system.hh" 75#include "sim/system.hh" 76#include "cpu/o3/isa_specific.hh" 77 78using namespace std; 79 80template<class Impl> 81DefaultFetch<Impl>::DefaultFetch(O3CPU *_cpu, DerivO3CPUParams *params) 82 : cpu(_cpu), 83 decodeToFetchDelay(params->decodeToFetchDelay), 84 renameToFetchDelay(params->renameToFetchDelay), 85 iewToFetchDelay(params->iewToFetchDelay), 86 commitToFetchDelay(params->commitToFetchDelay), 87 fetchWidth(params->fetchWidth), 88 decodeWidth(params->decodeWidth), 89 retryPkt(NULL), 90 retryTid(InvalidThreadID), 91 cacheBlkSize(cpu->cacheLineSize()), 92 fetchBufferSize(params->fetchBufferSize), 93 fetchBufferMask(fetchBufferSize - 1), 94 fetchQueueSize(params->fetchQueueSize), 95 numThreads(params->numThreads), 96 numFetchingThreads(params->smtNumFetchingThreads), 97 finishTranslationEvent(this) 98{ 99 if (numThreads > Impl::MaxThreads) 100 fatal("numThreads (%d) is larger than compiled limit (%d),\n" 101 "\tincrease MaxThreads in src/cpu/o3/impl.hh\n", 102 numThreads, static_cast<int>(Impl::MaxThreads)); 103 if (fetchWidth > Impl::MaxWidth) 104 fatal("fetchWidth (%d) is larger than compiled limit (%d),\n" 105 "\tincrease MaxWidth in src/cpu/o3/impl.hh\n", 106 fetchWidth, static_cast<int>(Impl::MaxWidth)); 107 if (fetchBufferSize > cacheBlkSize) 108 fatal("fetch buffer size (%u bytes) is greater than the cache " 109 "block size (%u bytes)\n", fetchBufferSize, cacheBlkSize); 110 if (cacheBlkSize % fetchBufferSize) 111 fatal("cache block (%u bytes) is not a multiple of the " 112 "fetch buffer (%u bytes)\n", cacheBlkSize, fetchBufferSize); 113 114 std::string policy = params->smtFetchPolicy; 115 116 // Convert string to lowercase 117 std::transform(policy.begin(), policy.end(), policy.begin(), 118 (int(*)(int)) tolower); 119 120 // Figure out fetch policy 121 if (policy == "singlethread") { 122 fetchPolicy = SingleThread; 123 if (numThreads > 1) 124 panic("Invalid Fetch Policy for a SMT workload."); 125 } else if (policy == "roundrobin") { 126 fetchPolicy = RoundRobin; 127 DPRINTF(Fetch, "Fetch policy set to Round Robin\n"); 128 } else if (policy == "branch") { 129 fetchPolicy = Branch; 130 DPRINTF(Fetch, "Fetch policy set to Branch Count\n"); 131 } else if (policy == "iqcount") { 132 fetchPolicy = IQ; 133 DPRINTF(Fetch, "Fetch policy set to IQ count\n"); 134 } else if (policy == "lsqcount") { 135 fetchPolicy = LSQ; 136 DPRINTF(Fetch, "Fetch policy set to LSQ count\n"); 137 } else { 138 fatal("Invalid Fetch Policy. Options Are: {SingleThread," 139 " RoundRobin,LSQcount,IQcount}\n"); 140 } 141 142 // Get the size of an instruction. 143 instSize = sizeof(TheISA::MachInst); 144 145 for (int i = 0; i < Impl::MaxThreads; i++) { 146 decoder[i] = NULL; 147 fetchBuffer[i] = NULL; 148 fetchBufferPC[i] = 0; 149 fetchBufferValid[i] = false; 150 } 151 152 branchPred = params->branchPred; 153 154 for (ThreadID tid = 0; tid < numThreads; tid++) { 155 decoder[tid] = new TheISA::Decoder(params->isa[tid]); 156 // Create space to buffer the cache line data, 157 // which may not hold the entire cache line. 158 fetchBuffer[tid] = new uint8_t[fetchBufferSize]; 159 } 160} 161 162template <class Impl> 163std::string 164DefaultFetch<Impl>::name() const 165{ 166 return cpu->name() + ".fetch"; 167} 168 169template <class Impl> 170void 171DefaultFetch<Impl>::regProbePoints() 172{ 173 ppFetch = new ProbePointArg<DynInstPtr>(cpu->getProbeManager(), "Fetch"); 174} 175 176template <class Impl> 177void 178DefaultFetch<Impl>::regStats() 179{ 180 icacheStallCycles 181 .name(name() + ".icacheStallCycles") 182 .desc("Number of cycles fetch is stalled on an Icache miss") 183 .prereq(icacheStallCycles); 184 185 fetchedInsts 186 .name(name() + ".Insts") 187 .desc("Number of instructions fetch has processed") 188 .prereq(fetchedInsts); 189 190 fetchedBranches 191 .name(name() + ".Branches") 192 .desc("Number of branches that fetch encountered") 193 .prereq(fetchedBranches); 194 195 predictedBranches 196 .name(name() + ".predictedBranches") 197 .desc("Number of branches that fetch has predicted taken") 198 .prereq(predictedBranches); 199 200 fetchCycles 201 .name(name() + ".Cycles") 202 .desc("Number of cycles fetch has run and was not squashing or" 203 " blocked") 204 .prereq(fetchCycles); 205 206 fetchSquashCycles 207 .name(name() + ".SquashCycles") 208 .desc("Number of cycles fetch has spent squashing") 209 .prereq(fetchSquashCycles); 210 211 fetchTlbCycles 212 .name(name() + ".TlbCycles") 213 .desc("Number of cycles fetch has spent waiting for tlb") 214 .prereq(fetchTlbCycles); 215 216 fetchIdleCycles 217 .name(name() + ".IdleCycles") 218 .desc("Number of cycles fetch was idle") 219 .prereq(fetchIdleCycles); 220 221 fetchBlockedCycles 222 .name(name() + ".BlockedCycles") 223 .desc("Number of cycles fetch has spent blocked") 224 .prereq(fetchBlockedCycles); 225 226 fetchedCacheLines 227 .name(name() + ".CacheLines") 228 .desc("Number of cache lines fetched") 229 .prereq(fetchedCacheLines); 230 231 fetchMiscStallCycles 232 .name(name() + ".MiscStallCycles") 233 .desc("Number of cycles fetch has spent waiting on interrupts, or " 234 "bad addresses, or out of MSHRs") 235 .prereq(fetchMiscStallCycles); 236 237 fetchPendingDrainCycles 238 .name(name() + ".PendingDrainCycles") 239 .desc("Number of cycles fetch has spent waiting on pipes to drain") 240 .prereq(fetchPendingDrainCycles); 241 242 fetchNoActiveThreadStallCycles 243 .name(name() + ".NoActiveThreadStallCycles") 244 .desc("Number of stall cycles due to no active thread to fetch from") 245 .prereq(fetchNoActiveThreadStallCycles); 246 247 fetchPendingTrapStallCycles 248 .name(name() + ".PendingTrapStallCycles") 249 .desc("Number of stall cycles due to pending traps") 250 .prereq(fetchPendingTrapStallCycles); 251 252 fetchPendingQuiesceStallCycles 253 .name(name() + ".PendingQuiesceStallCycles") 254 .desc("Number of stall cycles due to pending quiesce instructions") 255 .prereq(fetchPendingQuiesceStallCycles); 256 257 fetchIcacheWaitRetryStallCycles 258 .name(name() + ".IcacheWaitRetryStallCycles") 259 .desc("Number of stall cycles due to full MSHR") 260 .prereq(fetchIcacheWaitRetryStallCycles); 261 262 fetchIcacheSquashes 263 .name(name() + ".IcacheSquashes") 264 .desc("Number of outstanding Icache misses that were squashed") 265 .prereq(fetchIcacheSquashes); 266 267 fetchTlbSquashes 268 .name(name() + ".ItlbSquashes") 269 .desc("Number of outstanding ITLB misses that were squashed") 270 .prereq(fetchTlbSquashes); 271 272 fetchNisnDist 273 .init(/* base value */ 0, 274 /* last value */ fetchWidth, 275 /* bucket size */ 1) 276 .name(name() + ".rateDist") 277 .desc("Number of instructions fetched each cycle (Total)") 278 .flags(Stats::pdf); 279 280 idleRate 281 .name(name() + ".idleRate") 282 .desc("Percent of cycles fetch was idle") 283 .prereq(idleRate); 284 idleRate = fetchIdleCycles * 100 / cpu->numCycles; 285 286 branchRate 287 .name(name() + ".branchRate") 288 .desc("Number of branch fetches per cycle") 289 .flags(Stats::total); 290 branchRate = fetchedBranches / cpu->numCycles; 291 292 fetchRate 293 .name(name() + ".rate") 294 .desc("Number of inst fetches per cycle") 295 .flags(Stats::total); 296 fetchRate = fetchedInsts / cpu->numCycles; 297} 298 299template<class Impl> 300void 301DefaultFetch<Impl>::setTimeBuffer(TimeBuffer<TimeStruct> *time_buffer) 302{ 303 timeBuffer = time_buffer; 304 305 // Create wires to get information from proper places in time buffer. 306 fromDecode = timeBuffer->getWire(-decodeToFetchDelay); 307 fromRename = timeBuffer->getWire(-renameToFetchDelay); 308 fromIEW = timeBuffer->getWire(-iewToFetchDelay); 309 fromCommit = timeBuffer->getWire(-commitToFetchDelay); 310} 311 312template<class Impl> 313void 314DefaultFetch<Impl>::setActiveThreads(std::list<ThreadID> *at_ptr) 315{ 316 activeThreads = at_ptr; 317} 318 319template<class Impl> 320void 321DefaultFetch<Impl>::setFetchQueue(TimeBuffer<FetchStruct> *ftb_ptr) 322{ 323 // Create wire to write information to proper place in fetch time buf. 324 toDecode = ftb_ptr->getWire(0); 325} 326 327template<class Impl> 328void 329DefaultFetch<Impl>::startupStage() 330{ 331 assert(priorityList.empty()); 332 resetStage(); 333 334 // Fetch needs to start fetching instructions at the very beginning, 335 // so it must start up in active state. 336 switchToActive(); 337} 338 339template<class Impl> 340void 341DefaultFetch<Impl>::resetStage() 342{ 343 numInst = 0; 344 interruptPending = false; 345 cacheBlocked = false; 346 347 priorityList.clear(); 348 349 // Setup PC and nextPC with initial state. 350 for (ThreadID tid = 0; tid < numThreads; ++tid) { 351 fetchStatus[tid] = Running; 352 pc[tid] = cpu->pcState(tid); 353 fetchOffset[tid] = 0; 354 macroop[tid] = NULL; 355 356 delayedCommit[tid] = false; 357 memReq[tid] = NULL; 358 359 stalls[tid].decode = false; 360 stalls[tid].drain = false; 361 362 fetchBufferPC[tid] = 0; 363 fetchBufferValid[tid] = false; 364 365 fetchQueue[tid].clear(); 366 367 priorityList.push_back(tid); 368 } 369 370 wroteToTimeBuffer = false; 371 _status = Inactive; 372} 373 374template<class Impl> 375void 376DefaultFetch<Impl>::processCacheCompletion(PacketPtr pkt) 377{ 378 ThreadID tid = pkt->req->threadId(); 379 380 DPRINTF(Fetch, "[tid:%u] Waking up from cache miss.\n", tid); 381 assert(!cpu->switchedOut()); 382 383 // Only change the status if it's still waiting on the icache access 384 // to return. 385 if (fetchStatus[tid] != IcacheWaitResponse || 386 pkt->req != memReq[tid]) { 387 ++fetchIcacheSquashes; 388 delete pkt->req; 389 delete pkt; 390 return; 391 } 392 393 memcpy(fetchBuffer[tid], pkt->getConstPtr<uint8_t>(), fetchBufferSize); 394 fetchBufferValid[tid] = true; 395 396 // Wake up the CPU (if it went to sleep and was waiting on 397 // this completion event). 398 cpu->wakeCPU(); 399 400 DPRINTF(Activity, "[tid:%u] Activating fetch due to cache completion\n", 401 tid); 402 403 switchToActive(); 404 405 // Only switch to IcacheAccessComplete if we're not stalled as well. 406 if (checkStall(tid)) { 407 fetchStatus[tid] = Blocked; 408 } else { 409 fetchStatus[tid] = IcacheAccessComplete; 410 } 411 412 pkt->req->setAccessLatency(); 413 cpu->ppInstAccessComplete->notify(pkt); 414 // Reset the mem req to NULL. 415 delete pkt->req; 416 delete pkt; 417 memReq[tid] = NULL; 418} 419 420template <class Impl> 421void 422DefaultFetch<Impl>::drainResume() 423{ 424 for (ThreadID i = 0; i < numThreads; ++i) 425 stalls[i].drain = false; 426} 427 428template <class Impl> 429void 430DefaultFetch<Impl>::drainSanityCheck() const 431{ 432 assert(isDrained()); 433 assert(retryPkt == NULL); 434 assert(retryTid == InvalidThreadID); 435 assert(!cacheBlocked); 436 assert(!interruptPending); 437 438 for (ThreadID i = 0; i < numThreads; ++i) { 439 assert(!memReq[i]); 440 assert(fetchStatus[i] == Idle || stalls[i].drain); 441 } 442 443 branchPred->drainSanityCheck(); 444} 445 446template <class Impl> 447bool 448DefaultFetch<Impl>::isDrained() const 449{ 450 /* Make sure that threads are either idle of that the commit stage 451 * has signaled that draining has completed by setting the drain 452 * stall flag. This effectively forces the pipeline to be disabled 453 * until the whole system is drained (simulation may continue to 454 * drain other components). 455 */ 456 for (ThreadID i = 0; i < numThreads; ++i) { 457 // Verify fetch queues are drained 458 if (!fetchQueue[i].empty()) 459 return false; 460 461 // Return false if not idle or drain stalled 462 if (fetchStatus[i] != Idle) { 463 if (fetchStatus[i] == Blocked && stalls[i].drain) 464 continue; 465 else 466 return false; 467 } 468 } 469 470 /* The pipeline might start up again in the middle of the drain 471 * cycle if the finish translation event is scheduled, so make 472 * sure that's not the case. 473 */ 474 return !finishTranslationEvent.scheduled(); 475} 476 477template <class Impl> 478void 479DefaultFetch<Impl>::takeOverFrom() 480{ 481 assert(cpu->getInstPort().isConnected()); 482 resetStage(); 483 484} 485 486template <class Impl> 487void 488DefaultFetch<Impl>::drainStall(ThreadID tid) 489{ 490 assert(cpu->isDraining()); 491 assert(!stalls[tid].drain); 492 DPRINTF(Drain, "%i: Thread drained.\n", tid); 493 stalls[tid].drain = true; 494} 495 496template <class Impl> 497void 498DefaultFetch<Impl>::wakeFromQuiesce() 499{ 500 DPRINTF(Fetch, "Waking up from quiesce\n"); 501 // Hopefully this is safe 502 // @todo: Allow other threads to wake from quiesce. 503 fetchStatus[0] = Running; 504} 505 506template <class Impl> 507inline void 508DefaultFetch<Impl>::switchToActive() 509{ 510 if (_status == Inactive) { 511 DPRINTF(Activity, "Activating stage.\n"); 512 513 cpu->activateStage(O3CPU::FetchIdx); 514 515 _status = Active; 516 } 517} 518 519template <class Impl> 520inline void 521DefaultFetch<Impl>::switchToInactive() 522{ 523 if (_status == Active) { 524 DPRINTF(Activity, "Deactivating stage.\n"); 525 526 cpu->deactivateStage(O3CPU::FetchIdx); 527 528 _status = Inactive; 529 } 530} 531 532template <class Impl> 533void 534DefaultFetch<Impl>::deactivateThread(ThreadID tid) 535{ 536 // Update priority list 537 auto thread_it = std::find(priorityList.begin(), priorityList.end(), tid); 538 if (thread_it != priorityList.end()) { 539 priorityList.erase(thread_it); 540 } 541} 542 543template <class Impl> 544bool 545DefaultFetch<Impl>::lookupAndUpdateNextPC( 546 DynInstPtr &inst, TheISA::PCState &nextPC) 547{ 548 // Do branch prediction check here. 549 // A bit of a misnomer...next_PC is actually the current PC until 550 // this function updates it. 551 bool predict_taken; 552 553 if (!inst->isControl()) { 554 TheISA::advancePC(nextPC, inst->staticInst); 555 inst->setPredTarg(nextPC); 556 inst->setPredTaken(false); 557 return false; 558 } 559 560 ThreadID tid = inst->threadNumber; 561 predict_taken = branchPred->predict(inst->staticInst, inst->seqNum, 562 nextPC, tid); 563 564 if (predict_taken) { 565 DPRINTF(Fetch, "[tid:%i]: [sn:%i]: Branch predicted to be taken to %s.\n", 566 tid, inst->seqNum, nextPC); 567 } else { 568 DPRINTF(Fetch, "[tid:%i]: [sn:%i]:Branch predicted to be not taken.\n", 569 tid, inst->seqNum); 570 } 571 572 DPRINTF(Fetch, "[tid:%i]: [sn:%i] Branch predicted to go to %s.\n", 573 tid, inst->seqNum, nextPC); 574 inst->setPredTarg(nextPC); 575 inst->setPredTaken(predict_taken); 576 577 ++fetchedBranches; 578 579 if (predict_taken) { 580 ++predictedBranches; 581 } 582 583 return predict_taken; 584} 585 586template <class Impl> 587bool 588DefaultFetch<Impl>::fetchCacheLine(Addr vaddr, ThreadID tid, Addr pc) 589{ 590 Fault fault = NoFault; 591 592 assert(!cpu->switchedOut()); 593 594 // @todo: not sure if these should block translation. 595 //AlphaDep 596 if (cacheBlocked) { 597 DPRINTF(Fetch, "[tid:%i] Can't fetch cache line, cache blocked\n", 598 tid); 599 return false; 600 } else if (checkInterrupt(pc) && !delayedCommit[tid]) { 601 // Hold off fetch from getting new instructions when: 602 // Cache is blocked, or 603 // while an interrupt is pending and we're not in PAL mode, or 604 // fetch is switched out. 605 DPRINTF(Fetch, "[tid:%i] Can't fetch cache line, interrupt pending\n", 606 tid); 607 return false; 608 } 609 610 // Align the fetch address to the start of a fetch buffer segment. 611 Addr fetchBufferBlockPC = fetchBufferAlignPC(vaddr); 612 613 DPRINTF(Fetch, "[tid:%i] Fetching cache line %#x for addr %#x\n", 614 tid, fetchBufferBlockPC, vaddr); 615 616 // Setup the memReq to do a read of the first instruction's address. 617 // Set the appropriate read size and flags as well. 618 // Build request here. 619 RequestPtr mem_req = 620 new Request(tid, fetchBufferBlockPC, fetchBufferSize, 621 Request::INST_FETCH, cpu->instMasterId(), pc, 622 cpu->thread[tid]->contextId(), tid); 623 624 mem_req->taskId(cpu->taskId()); 625 626 memReq[tid] = mem_req; 627 628 // Initiate translation of the icache block 629 fetchStatus[tid] = ItlbWait; 630 FetchTranslation *trans = new FetchTranslation(this); 631 cpu->itb->translateTiming(mem_req, cpu->thread[tid]->getTC(), 632 trans, BaseTLB::Execute); 633 return true; 634} 635 636template <class Impl> 637void 638DefaultFetch<Impl>::finishTranslation(const Fault &fault, RequestPtr mem_req) 639{ 640 ThreadID tid = mem_req->threadId(); 641 Addr fetchBufferBlockPC = mem_req->getVaddr(); 642 643 assert(!cpu->switchedOut()); 644 645 // Wake up CPU if it was idle 646 cpu->wakeCPU(); 647 648 if (fetchStatus[tid] != ItlbWait || mem_req != memReq[tid] || 649 mem_req->getVaddr() != memReq[tid]->getVaddr()) { 650 DPRINTF(Fetch, "[tid:%i] Ignoring itlb completed after squash\n", 651 tid); 652 ++fetchTlbSquashes; 653 delete mem_req; 654 return; 655 } 656 657 658 // If translation was successful, attempt to read the icache block. 659 if (fault == NoFault) { 660 // Check that we're not going off into random memory 661 // If we have, just wait around for commit to squash something and put 662 // us on the right track 663 if (!cpu->system->isMemAddr(mem_req->getPaddr())) { 664 warn("Address %#x is outside of physical memory, stopping fetch\n", 665 mem_req->getPaddr()); 666 fetchStatus[tid] = NoGoodAddr; 667 delete mem_req; 668 memReq[tid] = NULL; 669 return; 670 } 671 672 // Build packet here. 673 PacketPtr data_pkt = new Packet(mem_req, MemCmd::ReadReq); 674 data_pkt->dataDynamic(new uint8_t[fetchBufferSize]); 675 676 fetchBufferPC[tid] = fetchBufferBlockPC; 677 fetchBufferValid[tid] = false; 678 DPRINTF(Fetch, "Fetch: Doing instruction read.\n"); 679 680 fetchedCacheLines++; 681 682 // Access the cache. 683 if (!cpu->getInstPort().sendTimingReq(data_pkt)) { 684 assert(retryPkt == NULL); 685 assert(retryTid == InvalidThreadID); 686 DPRINTF(Fetch, "[tid:%i] Out of MSHRs!\n", tid); 687 688 fetchStatus[tid] = IcacheWaitRetry; 689 retryPkt = data_pkt; 690 retryTid = tid; 691 cacheBlocked = true; 692 } else { 693 DPRINTF(Fetch, "[tid:%i]: Doing Icache access.\n", tid); 694 DPRINTF(Activity, "[tid:%i]: Activity: Waiting on I-cache " 695 "response.\n", tid); 696 lastIcacheStall[tid] = curTick(); 697 fetchStatus[tid] = IcacheWaitResponse; 698 } 699 } else { 700 // Don't send an instruction to decode if we can't handle it. 701 if (!(numInst < fetchWidth) || !(fetchQueue[tid].size() < fetchQueueSize)) { 702 assert(!finishTranslationEvent.scheduled()); 703 finishTranslationEvent.setFault(fault); 704 finishTranslationEvent.setReq(mem_req); 705 cpu->schedule(finishTranslationEvent, 706 cpu->clockEdge(Cycles(1))); 707 return; 708 } 709 DPRINTF(Fetch, "[tid:%i] Got back req with addr %#x but expected %#x\n", 710 tid, mem_req->getVaddr(), memReq[tid]->getVaddr()); 711 // Translation faulted, icache request won't be sent. 712 delete mem_req; 713 memReq[tid] = NULL; 714 715 // Send the fault to commit. This thread will not do anything 716 // until commit handles the fault. The only other way it can 717 // wake up is if a squash comes along and changes the PC. 718 TheISA::PCState fetchPC = pc[tid]; 719 720 DPRINTF(Fetch, "[tid:%i]: Translation faulted, building noop.\n", tid); 721 // We will use a nop in ordier to carry the fault. 722 DynInstPtr instruction = buildInst(tid, 723 decoder[tid]->decode(TheISA::NoopMachInst, fetchPC.instAddr()), 724 NULL, fetchPC, fetchPC, false); 725 726 instruction->setPredTarg(fetchPC); 727 instruction->fault = fault; 728 wroteToTimeBuffer = true; 729 730 DPRINTF(Activity, "Activity this cycle.\n"); 731 cpu->activityThisCycle(); 732 733 fetchStatus[tid] = TrapPending; 734 735 DPRINTF(Fetch, "[tid:%i]: Blocked, need to handle the trap.\n", tid); 736 DPRINTF(Fetch, "[tid:%i]: fault (%s) detected @ PC %s.\n", 737 tid, fault->name(), pc[tid]); 738 } 739 _status = updateFetchStatus(); 740} 741 742template <class Impl> 743inline void 744DefaultFetch<Impl>::doSquash(const TheISA::PCState &newPC, 745 const DynInstPtr squashInst, ThreadID tid) 746{ 747 DPRINTF(Fetch, "[tid:%i]: Squashing, setting PC to: %s.\n", 748 tid, newPC); 749 750 pc[tid] = newPC; 751 fetchOffset[tid] = 0; 752 if (squashInst && squashInst->pcState().instAddr() == newPC.instAddr()) 753 macroop[tid] = squashInst->macroop; 754 else 755 macroop[tid] = NULL; 756 decoder[tid]->reset(); 757 758 // Clear the icache miss if it's outstanding. 759 if (fetchStatus[tid] == IcacheWaitResponse) { 760 DPRINTF(Fetch, "[tid:%i]: Squashing outstanding Icache miss.\n", 761 tid); 762 memReq[tid] = NULL; 763 } else if (fetchStatus[tid] == ItlbWait) { 764 DPRINTF(Fetch, "[tid:%i]: Squashing outstanding ITLB miss.\n", 765 tid); 766 memReq[tid] = NULL; 767 } 768 769 // Get rid of the retrying packet if it was from this thread. 770 if (retryTid == tid) { 771 assert(cacheBlocked); 772 if (retryPkt) { 773 delete retryPkt->req; 774 delete retryPkt; 775 } 776 retryPkt = NULL; 777 retryTid = InvalidThreadID; 778 } 779 780 fetchStatus[tid] = Squashing; 781 782 // Empty fetch queue 783 fetchQueue[tid].clear(); 784 785 // microops are being squashed, it is not known wheather the 786 // youngest non-squashed microop was marked delayed commit 787 // or not. Setting the flag to true ensures that the 788 // interrupts are not handled when they cannot be, though 789 // some opportunities to handle interrupts may be missed. 790 delayedCommit[tid] = true; 791 792 ++fetchSquashCycles; 793} 794 795template<class Impl> 796void 797DefaultFetch<Impl>::squashFromDecode(const TheISA::PCState &newPC, 798 const DynInstPtr squashInst, 799 const InstSeqNum seq_num, ThreadID tid) 800{ 801 DPRINTF(Fetch, "[tid:%i]: Squashing from decode.\n", tid); 802 803 doSquash(newPC, squashInst, tid); 804 805 // Tell the CPU to remove any instructions that are in flight between 806 // fetch and decode. 807 cpu->removeInstsUntil(seq_num, tid); 808} 809 810template<class Impl> 811bool 812DefaultFetch<Impl>::checkStall(ThreadID tid) const 813{ 814 bool ret_val = false; 815 816 if (stalls[tid].drain) { 817 assert(cpu->isDraining()); 818 DPRINTF(Fetch,"[tid:%i]: Drain stall detected.\n",tid); 819 ret_val = true; 820 } 821 822 return ret_val; 823} 824 825template<class Impl> 826typename DefaultFetch<Impl>::FetchStatus 827DefaultFetch<Impl>::updateFetchStatus() 828{ 829 //Check Running 830 list<ThreadID>::iterator threads = activeThreads->begin(); 831 list<ThreadID>::iterator end = activeThreads->end(); 832 833 while (threads != end) { 834 ThreadID tid = *threads++; 835 836 if (fetchStatus[tid] == Running || 837 fetchStatus[tid] == Squashing || 838 fetchStatus[tid] == IcacheAccessComplete) { 839 840 if (_status == Inactive) { 841 DPRINTF(Activity, "[tid:%i]: Activating stage.\n",tid); 842 843 if (fetchStatus[tid] == IcacheAccessComplete) { 844 DPRINTF(Activity, "[tid:%i]: Activating fetch due to cache" 845 "completion\n",tid); 846 } 847 848 cpu->activateStage(O3CPU::FetchIdx); 849 } 850 851 return Active; 852 } 853 } 854 855 // Stage is switching from active to inactive, notify CPU of it. 856 if (_status == Active) { 857 DPRINTF(Activity, "Deactivating stage.\n"); 858 859 cpu->deactivateStage(O3CPU::FetchIdx); 860 } 861 862 return Inactive; 863} 864 865template <class Impl> 866void 867DefaultFetch<Impl>::squash(const TheISA::PCState &newPC, 868 const InstSeqNum seq_num, DynInstPtr squashInst, 869 ThreadID tid) 870{ 871 DPRINTF(Fetch, "[tid:%u]: Squash from commit.\n", tid); 872 873 doSquash(newPC, squashInst, tid); 874 875 // Tell the CPU to remove any instructions that are not in the ROB. 876 cpu->removeInstsNotInROB(tid); 877} 878 879template <class Impl> 880void 881DefaultFetch<Impl>::tick() 882{ 883 list<ThreadID>::iterator threads = activeThreads->begin(); 884 list<ThreadID>::iterator end = activeThreads->end(); 885 bool status_change = false; 886 887 wroteToTimeBuffer = false; 888 889 for (ThreadID i = 0; i < numThreads; ++i) { 890 issuePipelinedIfetch[i] = false; 891 } 892 893 while (threads != end) { 894 ThreadID tid = *threads++; 895 896 // Check the signals for each thread to determine the proper status 897 // for each thread. 898 bool updated_status = checkSignalsAndUpdate(tid); 899 status_change = status_change || updated_status; 900 } 901 902 DPRINTF(Fetch, "Running stage.\n"); 903 904 if (FullSystem) { 905 if (fromCommit->commitInfo[0].interruptPending) { 906 interruptPending = true; 907 } 908 909 if (fromCommit->commitInfo[0].clearInterrupt) { 910 interruptPending = false; 911 } 912 } 913 914 for (threadFetched = 0; threadFetched < numFetchingThreads; 915 threadFetched++) { 916 // Fetch each of the actively fetching threads. 917 fetch(status_change); 918 } 919 920 // Record number of instructions fetched this cycle for distribution. 921 fetchNisnDist.sample(numInst); 922 923 if (status_change) { 924 // Change the fetch stage status if there was a status change. 925 _status = updateFetchStatus(); 926 } 927 928 // Issue the next I-cache request if possible. 929 for (ThreadID i = 0; i < numThreads; ++i) { 930 if (issuePipelinedIfetch[i]) { 931 pipelineIcacheAccesses(i); 932 } 933 } 934 935 // Send instructions enqueued into the fetch queue to decode. 936 // Limit rate by fetchWidth. Stall if decode is stalled. 937 unsigned insts_to_decode = 0; 938 unsigned available_insts = 0; 939 940 for (auto tid : *activeThreads) { 941 if (!stalls[tid].decode) { 942 available_insts += fetchQueue[tid].size(); 943 } 944 } 945 946 // Pick a random thread to start trying to grab instructions from 947 auto tid_itr = activeThreads->begin(); 948 std::advance(tid_itr, random_mt.random<uint8_t>(0, activeThreads->size() - 1)); 949 950 while (available_insts != 0 && insts_to_decode < decodeWidth) { 951 ThreadID tid = *tid_itr; 952 if (!stalls[tid].decode && !fetchQueue[tid].empty()) { 953 auto inst = fetchQueue[tid].front(); 954 toDecode->insts[toDecode->size++] = inst; 955 DPRINTF(Fetch, "[tid:%i][sn:%i]: Sending instruction to decode from " 956 "fetch queue. Fetch queue size: %i.\n", 957 tid, inst->seqNum, fetchQueue[tid].size()); 958 959 wroteToTimeBuffer = true; 960 fetchQueue[tid].pop_front(); 961 insts_to_decode++; 962 available_insts--; 963 } 964 965 tid_itr++; 966 // Wrap around if at end of active threads list 967 if (tid_itr == activeThreads->end()) 968 tid_itr = activeThreads->begin(); 969 } 970 971 // If there was activity this cycle, inform the CPU of it. 972 if (wroteToTimeBuffer) { 973 DPRINTF(Activity, "Activity this cycle.\n"); 974 cpu->activityThisCycle(); 975 } 976 977 // Reset the number of the instruction we've fetched. 978 numInst = 0; 979} 980 981template <class Impl> 982bool 983DefaultFetch<Impl>::checkSignalsAndUpdate(ThreadID tid) 984{ 985 // Update the per thread stall statuses. 986 if (fromDecode->decodeBlock[tid]) { 987 stalls[tid].decode = true; 988 } 989 990 if (fromDecode->decodeUnblock[tid]) { 991 assert(stalls[tid].decode); 992 assert(!fromDecode->decodeBlock[tid]); 993 stalls[tid].decode = false; 994 } 995 996 // Check squash signals from commit. 997 if (fromCommit->commitInfo[tid].squash) { 998 999 DPRINTF(Fetch, "[tid:%u]: Squashing instructions due to squash " 1000 "from commit.\n",tid); 1001 // In any case, squash. 1002 squash(fromCommit->commitInfo[tid].pc, 1003 fromCommit->commitInfo[tid].doneSeqNum, 1004 fromCommit->commitInfo[tid].squashInst, tid); 1005 1006 // If it was a branch mispredict on a control instruction, update the 1007 // branch predictor with that instruction, otherwise just kill the 1008 // invalid state we generated in after sequence number 1009 if (fromCommit->commitInfo[tid].mispredictInst && 1010 fromCommit->commitInfo[tid].mispredictInst->isControl()) { 1011 branchPred->squash(fromCommit->commitInfo[tid].doneSeqNum, 1012 fromCommit->commitInfo[tid].pc, 1013 fromCommit->commitInfo[tid].branchTaken, 1014 tid); 1015 } else { 1016 branchPred->squash(fromCommit->commitInfo[tid].doneSeqNum, 1017 tid); 1018 } 1019 1020 return true; 1021 } else if (fromCommit->commitInfo[tid].doneSeqNum) { 1022 // Update the branch predictor if it wasn't a squashed instruction 1023 // that was broadcasted. 1024 branchPred->update(fromCommit->commitInfo[tid].doneSeqNum, tid); 1025 } 1026 1027 // Check squash signals from decode. 1028 if (fromDecode->decodeInfo[tid].squash) { 1029 DPRINTF(Fetch, "[tid:%u]: Squashing instructions due to squash " 1030 "from decode.\n",tid); 1031 1032 // Update the branch predictor. 1033 if (fromDecode->decodeInfo[tid].branchMispredict) { 1034 branchPred->squash(fromDecode->decodeInfo[tid].doneSeqNum, 1035 fromDecode->decodeInfo[tid].nextPC, 1036 fromDecode->decodeInfo[tid].branchTaken, 1037 tid); 1038 } else { 1039 branchPred->squash(fromDecode->decodeInfo[tid].doneSeqNum, 1040 tid); 1041 } 1042 1043 if (fetchStatus[tid] != Squashing) { 1044 1045 DPRINTF(Fetch, "Squashing from decode with PC = %s\n", 1046 fromDecode->decodeInfo[tid].nextPC); 1047 // Squash unless we're already squashing 1048 squashFromDecode(fromDecode->decodeInfo[tid].nextPC, 1049 fromDecode->decodeInfo[tid].squashInst, 1050 fromDecode->decodeInfo[tid].doneSeqNum, 1051 tid); 1052 1053 return true; 1054 } 1055 } 1056 1057 if (checkStall(tid) && 1058 fetchStatus[tid] != IcacheWaitResponse && 1059 fetchStatus[tid] != IcacheWaitRetry && 1060 fetchStatus[tid] != ItlbWait && 1061 fetchStatus[tid] != QuiescePending) { 1062 DPRINTF(Fetch, "[tid:%i]: Setting to blocked\n",tid); 1063 1064 fetchStatus[tid] = Blocked; 1065 1066 return true; 1067 } 1068 1069 if (fetchStatus[tid] == Blocked || 1070 fetchStatus[tid] == Squashing) { 1071 // Switch status to running if fetch isn't being told to block or 1072 // squash this cycle. 1073 DPRINTF(Fetch, "[tid:%i]: Done squashing, switching to running.\n", 1074 tid); 1075 1076 fetchStatus[tid] = Running; 1077 1078 return true; 1079 } 1080 1081 // If we've reached this point, we have not gotten any signals that 1082 // cause fetch to change its status. Fetch remains the same as before. 1083 return false; 1084} 1085 1086template<class Impl> 1087typename Impl::DynInstPtr 1088DefaultFetch<Impl>::buildInst(ThreadID tid, StaticInstPtr staticInst, 1089 StaticInstPtr curMacroop, TheISA::PCState thisPC, 1090 TheISA::PCState nextPC, bool trace) 1091{ 1092 // Get a sequence number. 1093 InstSeqNum seq = cpu->getAndIncrementInstSeq(); 1094 1095 // Create a new DynInst from the instruction fetched. 1096 DynInstPtr instruction = 1097 new DynInst(staticInst, curMacroop, thisPC, nextPC, seq, cpu); 1098 instruction->setTid(tid); 1099 1100 instruction->setASID(tid); 1101 1102 instruction->setThreadState(cpu->thread[tid]); 1103 1104 DPRINTF(Fetch, "[tid:%i]: Instruction PC %#x (%d) created " 1105 "[sn:%lli].\n", tid, thisPC.instAddr(), 1106 thisPC.microPC(), seq); 1107 1108 DPRINTF(Fetch, "[tid:%i]: Instruction is: %s\n", tid, 1109 instruction->staticInst-> 1110 disassemble(thisPC.instAddr())); 1111 1112#if TRACING_ON 1113 if (trace) { 1114 instruction->traceData = 1115 cpu->getTracer()->getInstRecord(curTick(), cpu->tcBase(tid), 1116 instruction->staticInst, thisPC, curMacroop); 1117 } 1118#else 1119 instruction->traceData = NULL; 1120#endif 1121 1122 // Add instruction to the CPU's list of instructions. 1123 instruction->setInstListIt(cpu->addInst(instruction)); 1124 1125 // Write the instruction to the first slot in the queue 1126 // that heads to decode. 1127 assert(numInst < fetchWidth); 1128 fetchQueue[tid].push_back(instruction); 1129 assert(fetchQueue[tid].size() <= fetchQueueSize); 1130 DPRINTF(Fetch, "[tid:%i]: Fetch queue entry created (%i/%i).\n", 1131 tid, fetchQueue[tid].size(), fetchQueueSize); 1132 //toDecode->insts[toDecode->size++] = instruction; 1133 1134 // Keep track of if we can take an interrupt at this boundary 1135 delayedCommit[tid] = instruction->isDelayedCommit(); 1136 1137 return instruction; 1138} 1139 1140template<class Impl> 1141void 1142DefaultFetch<Impl>::fetch(bool &status_change) 1143{ 1144 ////////////////////////////////////////// 1145 // Start actual fetch 1146 ////////////////////////////////////////// 1147 ThreadID tid = getFetchingThread(fetchPolicy); 1148 1149 assert(!cpu->switchedOut()); 1150 1151 if (tid == InvalidThreadID) { 1152 // Breaks looping condition in tick() 1153 threadFetched = numFetchingThreads; 1154 1155 if (numThreads == 1) { // @todo Per-thread stats 1156 profileStall(0); 1157 } 1158 1159 return; 1160 } 1161 1162 DPRINTF(Fetch, "Attempting to fetch from [tid:%i]\n", tid); 1163 1164 // The current PC. 1165 TheISA::PCState thisPC = pc[tid]; 1166 1167 Addr pcOffset = fetchOffset[tid]; 1168 Addr fetchAddr = (thisPC.instAddr() + pcOffset) & BaseCPU::PCMask; 1169 1170 bool inRom = isRomMicroPC(thisPC.microPC()); 1171 1172 // If returning from the delay of a cache miss, then update the status 1173 // to running, otherwise do the cache access. Possibly move this up 1174 // to tick() function. 1175 if (fetchStatus[tid] == IcacheAccessComplete) { 1176 DPRINTF(Fetch, "[tid:%i]: Icache miss is complete.\n", tid); 1177 1178 fetchStatus[tid] = Running; 1179 status_change = true; 1180 } else if (fetchStatus[tid] == Running) { 1181 // Align the fetch PC so its at the start of a fetch buffer segment. 1182 Addr fetchBufferBlockPC = fetchBufferAlignPC(fetchAddr); 1183 1184 // If buffer is no longer valid or fetchAddr has moved to point 1185 // to the next cache block, AND we have no remaining ucode 1186 // from a macro-op, then start fetch from icache. 1187 if (!(fetchBufferValid[tid] && fetchBufferBlockPC == fetchBufferPC[tid]) 1188 && !inRom && !macroop[tid]) { 1189 DPRINTF(Fetch, "[tid:%i]: Attempting to translate and read " 1190 "instruction, starting at PC %s.\n", tid, thisPC); 1191 1192 fetchCacheLine(fetchAddr, tid, thisPC.instAddr()); 1193 1194 if (fetchStatus[tid] == IcacheWaitResponse) 1195 ++icacheStallCycles; 1196 else if (fetchStatus[tid] == ItlbWait) 1197 ++fetchTlbCycles; 1198 else 1199 ++fetchMiscStallCycles; 1200 return; 1201 } else if ((checkInterrupt(thisPC.instAddr()) && !delayedCommit[tid])) { 1202 // Stall CPU if an interrupt is posted and we're not issuing 1203 // an delayed commit micro-op currently (delayed commit instructions 1204 // are not interruptable by interrupts, only faults) 1205 ++fetchMiscStallCycles; 1206 DPRINTF(Fetch, "[tid:%i]: Fetch is stalled!\n", tid); 1207 return; 1208 } 1209 } else { 1210 if (fetchStatus[tid] == Idle) { 1211 ++fetchIdleCycles; 1212 DPRINTF(Fetch, "[tid:%i]: Fetch is idle!\n", tid); 1213 } 1214 1215 // Status is Idle, so fetch should do nothing. 1216 return; 1217 } 1218 1219 ++fetchCycles; 1220 1221 TheISA::PCState nextPC = thisPC; 1222 1223 StaticInstPtr staticInst = NULL; 1224 StaticInstPtr curMacroop = macroop[tid]; 1225 1226 // If the read of the first instruction was successful, then grab the 1227 // instructions from the rest of the cache line and put them into the 1228 // queue heading to decode. 1229 1230 DPRINTF(Fetch, "[tid:%i]: Adding instructions to queue to " 1231 "decode.\n", tid); 1232 1233 // Need to keep track of whether or not a predicted branch 1234 // ended this fetch block. 1235 bool predictedBranch = false; 1236 1237 // Need to halt fetch if quiesce instruction detected 1238 bool quiesce = false; 1239 1240 TheISA::MachInst *cacheInsts = 1241 reinterpret_cast<TheISA::MachInst *>(fetchBuffer[tid]); 1242 1243 const unsigned numInsts = fetchBufferSize / instSize; 1244 unsigned blkOffset = (fetchAddr - fetchBufferPC[tid]) / instSize; 1245 1246 // Loop through instruction memory from the cache. 1247 // Keep issuing while fetchWidth is available and branch is not 1248 // predicted taken 1249 while (numInst < fetchWidth && fetchQueue[tid].size() < fetchQueueSize 1250 && !predictedBranch && !quiesce) { 1251 // We need to process more memory if we aren't going to get a 1252 // StaticInst from the rom, the current macroop, or what's already 1253 // in the decoder. 1254 bool needMem = !inRom && !curMacroop && 1255 !decoder[tid]->instReady(); 1256 fetchAddr = (thisPC.instAddr() + pcOffset) & BaseCPU::PCMask; 1257 Addr fetchBufferBlockPC = fetchBufferAlignPC(fetchAddr); 1258 1259 if (needMem) { 1260 // If buffer is no longer valid or fetchAddr has moved to point 1261 // to the next cache block then start fetch from icache. 1262 if (!fetchBufferValid[tid] || 1263 fetchBufferBlockPC != fetchBufferPC[tid]) 1264 break; 1265 1266 if (blkOffset >= numInsts) { 1267 // We need to process more memory, but we've run out of the 1268 // current block. 1269 break; 1270 } 1271 1272 if (ISA_HAS_DELAY_SLOT && pcOffset == 0) { 1273 // Walk past any annulled delay slot instructions. 1274 Addr pcAddr = thisPC.instAddr() & BaseCPU::PCMask; 1275 while (fetchAddr != pcAddr && blkOffset < numInsts) { 1276 blkOffset++; 1277 fetchAddr += instSize; 1278 } 1279 if (blkOffset >= numInsts) 1280 break; 1281 } 1282 1283 MachInst inst = TheISA::gtoh(cacheInsts[blkOffset]); 1284 decoder[tid]->moreBytes(thisPC, fetchAddr, inst); 1285 1286 if (decoder[tid]->needMoreBytes()) { 1287 blkOffset++; 1288 fetchAddr += instSize; 1289 pcOffset += instSize; 1290 } 1291 } 1292 1293 // Extract as many instructions and/or microops as we can from 1294 // the memory we've processed so far. 1295 do { 1296 if (!(curMacroop || inRom)) { 1297 if (decoder[tid]->instReady()) { 1298 staticInst = decoder[tid]->decode(thisPC); 1299 1300 // Increment stat of fetched instructions. 1301 ++fetchedInsts; 1302 1303 if (staticInst->isMacroop()) { 1304 curMacroop = staticInst; 1305 } else { 1306 pcOffset = 0; 1307 } 1308 } else { 1309 // We need more bytes for this instruction so blkOffset and 1310 // pcOffset will be updated 1311 break; 1312 } 1313 } 1314 // Whether we're moving to a new macroop because we're at the 1315 // end of the current one, or the branch predictor incorrectly 1316 // thinks we are... 1317 bool newMacro = false; 1318 if (curMacroop || inRom) { 1319 if (inRom) { 1320 staticInst = cpu->microcodeRom.fetchMicroop( 1321 thisPC.microPC(), curMacroop); 1322 } else { 1323 staticInst = curMacroop->fetchMicroop(thisPC.microPC()); 1324 } 1325 newMacro |= staticInst->isLastMicroop(); 1326 } 1327 1328 DynInstPtr instruction = 1329 buildInst(tid, staticInst, curMacroop, 1330 thisPC, nextPC, true); 1331 1332 ppFetch->notify(instruction); 1333 numInst++; 1334 1335#if TRACING_ON 1336 if (DTRACE(O3PipeView)) { 1337 instruction->fetchTick = curTick(); 1338 } 1339#endif 1340 1341 nextPC = thisPC; 1342 1343 // If we're branching after this instruction, quit fetching 1344 // from the same block. 1345 predictedBranch |= thisPC.branching(); 1346 predictedBranch |= 1347 lookupAndUpdateNextPC(instruction, nextPC); 1348 if (predictedBranch) { 1349 DPRINTF(Fetch, "Branch detected with PC = %s\n", thisPC); 1350 } 1351 1352 newMacro |= thisPC.instAddr() != nextPC.instAddr(); 1353 1354 // Move to the next instruction, unless we have a branch. 1355 thisPC = nextPC; 1356 inRom = isRomMicroPC(thisPC.microPC()); 1357 1358 if (newMacro) { 1359 fetchAddr = thisPC.instAddr() & BaseCPU::PCMask; 1360 blkOffset = (fetchAddr - fetchBufferPC[tid]) / instSize; 1361 pcOffset = 0; 1362 curMacroop = NULL; 1363 } 1364 1365 if (instruction->isQuiesce()) { 1366 DPRINTF(Fetch, 1367 "Quiesce instruction encountered, halting fetch!\n"); 1368 fetchStatus[tid] = QuiescePending; 1369 status_change = true; 1370 quiesce = true; 1371 break; 1372 } 1373 } while ((curMacroop || decoder[tid]->instReady()) && 1374 numInst < fetchWidth && 1375 fetchQueue[tid].size() < fetchQueueSize); 1376 1377 // Re-evaluate whether the next instruction to fetch is in micro-op ROM 1378 // or not. 1379 inRom = isRomMicroPC(thisPC.microPC()); 1380 } 1381 1382 if (predictedBranch) { 1383 DPRINTF(Fetch, "[tid:%i]: Done fetching, predicted branch " 1384 "instruction encountered.\n", tid); 1385 } else if (numInst >= fetchWidth) { 1386 DPRINTF(Fetch, "[tid:%i]: Done fetching, reached fetch bandwidth " 1387 "for this cycle.\n", tid); 1388 } else if (blkOffset >= fetchBufferSize) { 1389 DPRINTF(Fetch, "[tid:%i]: Done fetching, reached the end of the" 1390 "fetch buffer.\n", tid); 1391 } 1392 1393 macroop[tid] = curMacroop; 1394 fetchOffset[tid] = pcOffset; 1395 1396 if (numInst > 0) { 1397 wroteToTimeBuffer = true; 1398 } 1399 1400 pc[tid] = thisPC; 1401 1402 // pipeline a fetch if we're crossing a fetch buffer boundary and not in 1403 // a state that would preclude fetching 1404 fetchAddr = (thisPC.instAddr() + pcOffset) & BaseCPU::PCMask; 1405 Addr fetchBufferBlockPC = fetchBufferAlignPC(fetchAddr); 1406 issuePipelinedIfetch[tid] = fetchBufferBlockPC != fetchBufferPC[tid] && 1407 fetchStatus[tid] != IcacheWaitResponse && 1408 fetchStatus[tid] != ItlbWait && 1409 fetchStatus[tid] != IcacheWaitRetry && 1410 fetchStatus[tid] != QuiescePending && 1411 !curMacroop; 1412} 1413 1414template<class Impl> 1415void 1416DefaultFetch<Impl>::recvReqRetry() 1417{ 1418 if (retryPkt != NULL) { 1419 assert(cacheBlocked); 1420 assert(retryTid != InvalidThreadID); 1421 assert(fetchStatus[retryTid] == IcacheWaitRetry); 1422 1423 if (cpu->getInstPort().sendTimingReq(retryPkt)) { 1424 fetchStatus[retryTid] = IcacheWaitResponse; 1425 retryPkt = NULL; 1426 retryTid = InvalidThreadID; 1427 cacheBlocked = false; 1428 } 1429 } else { 1430 assert(retryTid == InvalidThreadID); 1431 // Access has been squashed since it was sent out. Just clear 1432 // the cache being blocked. 1433 cacheBlocked = false; 1434 } 1435} 1436 1437/////////////////////////////////////// 1438// // 1439// SMT FETCH POLICY MAINTAINED HERE // 1440// // 1441/////////////////////////////////////// 1442template<class Impl> 1443ThreadID 1444DefaultFetch<Impl>::getFetchingThread(FetchPriority &fetch_priority) 1445{ 1446 if (numThreads > 1) { 1447 switch (fetch_priority) { 1448 1449 case SingleThread: 1450 return 0; 1451 1452 case RoundRobin: 1453 return roundRobin(); 1454 1455 case IQ: 1456 return iqCount(); 1457 1458 case LSQ: 1459 return lsqCount(); 1460 1461 case Branch: 1462 return branchCount(); 1463 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#if 0 1594 list<ThreadID>::iterator thread = activeThreads->begin(); 1595 assert(thread != activeThreads->end()); 1596 ThreadID tid = *thread; 1597#endif 1598 1599 panic("Branch Count Fetch policy unimplemented\n"); 1600 return InvalidThreadID; 1601} 1602 1603template<class Impl> 1604void 1605DefaultFetch<Impl>::pipelineIcacheAccesses(ThreadID tid) 1606{ 1607 if (!issuePipelinedIfetch[tid]) { 1608 return; 1609 } 1610 1611 // The next PC to access. 1612 TheISA::PCState thisPC = pc[tid]; 1613 1614 if (isRomMicroPC(thisPC.microPC())) { 1615 return; 1616 } 1617 1618 Addr pcOffset = fetchOffset[tid]; 1619 Addr fetchAddr = (thisPC.instAddr() + pcOffset) & BaseCPU::PCMask; 1620 1621 // Align the fetch PC so its at the start of a fetch buffer segment. 1622 Addr fetchBufferBlockPC = fetchBufferAlignPC(fetchAddr); 1623 1624 // Unless buffer already got the block, fetch it from icache. 1625 if (!(fetchBufferValid[tid] && fetchBufferBlockPC == fetchBufferPC[tid])) { 1626 DPRINTF(Fetch, "[tid:%i]: Issuing a pipelined I-cache access, " 1627 "starting at PC %s.\n", tid, thisPC); 1628 1629 fetchCacheLine(fetchAddr, tid, thisPC.instAddr()); 1630 } 1631} 1632 1633template<class Impl> 1634void 1635DefaultFetch<Impl>::profileStall(ThreadID tid) { 1636 DPRINTF(Fetch,"There are no more threads available to fetch from.\n"); 1637 1638 // @todo Per-thread stats 1639 1640 if (stalls[tid].drain) { 1641 ++fetchPendingDrainCycles; 1642 DPRINTF(Fetch, "Fetch is waiting for a drain!\n"); 1643 } else if (activeThreads->empty()) { 1644 ++fetchNoActiveThreadStallCycles; 1645 DPRINTF(Fetch, "Fetch has no active thread!\n"); 1646 } else if (fetchStatus[tid] == Blocked) { 1647 ++fetchBlockedCycles; 1648 DPRINTF(Fetch, "[tid:%i]: Fetch is blocked!\n", tid); 1649 } else if (fetchStatus[tid] == Squashing) { 1650 ++fetchSquashCycles; 1651 DPRINTF(Fetch, "[tid:%i]: Fetch is squashing!\n", tid); 1652 } else if (fetchStatus[tid] == IcacheWaitResponse) { 1653 ++icacheStallCycles; 1654 DPRINTF(Fetch, "[tid:%i]: Fetch is waiting cache response!\n", 1655 tid); 1656 } else if (fetchStatus[tid] == ItlbWait) { 1657 ++fetchTlbCycles; 1658 DPRINTF(Fetch, "[tid:%i]: Fetch is waiting ITLB walk to " 1659 "finish!\n", tid); 1660 } else if (fetchStatus[tid] == TrapPending) { 1661 ++fetchPendingTrapStallCycles; 1662 DPRINTF(Fetch, "[tid:%i]: Fetch is waiting for a pending trap!\n", 1663 tid); 1664 } else if (fetchStatus[tid] == QuiescePending) { 1665 ++fetchPendingQuiesceStallCycles; 1666 DPRINTF(Fetch, "[tid:%i]: Fetch is waiting for a pending quiesce " 1667 "instruction!\n", tid); 1668 } else if (fetchStatus[tid] == IcacheWaitRetry) { 1669 ++fetchIcacheWaitRetryStallCycles; 1670 DPRINTF(Fetch, "[tid:%i]: Fetch is waiting for an I-cache retry!\n", 1671 tid); 1672 } else if (fetchStatus[tid] == NoGoodAddr) { 1673 DPRINTF(Fetch, "[tid:%i]: Fetch predicted non-executable address\n", 1674 tid); 1675 } else { 1676 DPRINTF(Fetch, "[tid:%i]: Unexpected fetch stall reason (Status: %i).\n", 1677 tid, fetchStatus[tid]); 1678 } 1679} 1680 1681#endif//__CPU_O3_FETCH_IMPL_HH__ 1682