lsq_unit_impl.hh revision 13429:a1e199fd8122
1 2/* 3 * Copyright (c) 2010-2014, 2017 ARM Limited 4 * Copyright (c) 2013 Advanced Micro Devices, Inc. 5 * All rights reserved 6 * 7 * The license below extends only to copyright in the software and shall 8 * not be construed as granting a license to any other intellectual 9 * property including but not limited to intellectual property relating 10 * to a hardware implementation of the functionality of the software 11 * licensed hereunder. You may use the software subject to the license 12 * terms below provided that you ensure that this notice is replicated 13 * unmodified and in its entirety in all distributions of the software, 14 * modified or unmodified, in source code or in binary form. 15 * 16 * Copyright (c) 2004-2005 The Regents of The University of Michigan 17 * All rights reserved. 18 * 19 * Redistribution and use in source and binary forms, with or without 20 * modification, are permitted provided that the following conditions are 21 * met: redistributions of source code must retain the above copyright 22 * notice, this list of conditions and the following disclaimer; 23 * redistributions in binary form must reproduce the above copyright 24 * notice, this list of conditions and the following disclaimer in the 25 * documentation and/or other materials provided with the distribution; 26 * neither the name of the copyright holders nor the names of its 27 * contributors may be used to endorse or promote products derived from 28 * this software without specific prior written permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 31 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 32 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 33 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 34 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 35 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 36 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 37 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 38 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 39 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 40 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 41 * 42 * Authors: Kevin Lim 43 * Korey Sewell 44 */ 45 46#ifndef __CPU_O3_LSQ_UNIT_IMPL_HH__ 47#define __CPU_O3_LSQ_UNIT_IMPL_HH__ 48 49#include "arch/generic/debugfaults.hh" 50#include "arch/locked_mem.hh" 51#include "base/str.hh" 52#include "config/the_isa.hh" 53#include "cpu/checker/cpu.hh" 54#include "cpu/o3/lsq.hh" 55#include "cpu/o3/lsq_unit.hh" 56#include "debug/Activity.hh" 57#include "debug/IEW.hh" 58#include "debug/LSQUnit.hh" 59#include "debug/O3PipeView.hh" 60#include "mem/packet.hh" 61#include "mem/request.hh" 62 63template<class Impl> 64LSQUnit<Impl>::WritebackEvent::WritebackEvent(const DynInstPtr &_inst, 65 PacketPtr _pkt, LSQUnit *lsq_ptr) 66 : Event(Default_Pri, AutoDelete), 67 inst(_inst), pkt(_pkt), lsqPtr(lsq_ptr) 68{ 69} 70 71template<class Impl> 72void 73LSQUnit<Impl>::WritebackEvent::process() 74{ 75 assert(!lsqPtr->cpu->switchedOut()); 76 77 lsqPtr->writeback(inst, pkt); 78 79 if (pkt->senderState) 80 delete pkt->senderState; 81 82 delete pkt; 83} 84 85template<class Impl> 86const char * 87LSQUnit<Impl>::WritebackEvent::description() const 88{ 89 return "Store writeback"; 90} 91 92template<class Impl> 93void 94LSQUnit<Impl>::completeDataAccess(PacketPtr pkt) 95{ 96 LSQSenderState *state = dynamic_cast<LSQSenderState *>(pkt->senderState); 97 DynInstPtr inst = state->inst; 98 DPRINTF(IEW, "Writeback event [sn:%lli].\n", inst->seqNum); 99 DPRINTF(Activity, "Activity: Writeback event [sn:%lli].\n", inst->seqNum); 100 101 if (state->cacheBlocked) { 102 // This is the first half of a previous split load, 103 // where the 2nd half blocked, ignore this response 104 DPRINTF(IEW, "[sn:%lli]: Response from first half of earlier " 105 "blocked split load recieved. Ignoring.\n", inst->seqNum); 106 delete state; 107 return; 108 } 109 110 // If this is a split access, wait until all packets are received. 111 if (TheISA::HasUnalignedMemAcc && !state->complete()) { 112 return; 113 } 114 115 assert(!cpu->switchedOut()); 116 if (!inst->isSquashed()) { 117 if (!state->noWB) { 118 // Only loads and store conditionals perform the writeback 119 // after receving the response from the memory 120 assert(inst->isLoad() || inst->isStoreConditional()); 121 if (!TheISA::HasUnalignedMemAcc || !state->isSplit || 122 !state->isLoad) { 123 writeback(inst, pkt); 124 } else { 125 writeback(inst, state->mainPkt); 126 } 127 } 128 129 if (inst->isStore()) { 130 completeStore(state->idx); 131 } 132 } 133 134 if (TheISA::HasUnalignedMemAcc && state->isSplit && state->isLoad) { 135 delete state->mainPkt; 136 } 137 138 pkt->req->setAccessLatency(); 139 cpu->ppDataAccessComplete->notify(std::make_pair(inst, pkt)); 140 141 delete state; 142} 143 144template <class Impl> 145LSQUnit<Impl>::LSQUnit() 146 : loads(0), stores(0), storesToWB(0), cacheBlockMask(0), stalled(false), 147 isStoreBlocked(false), storeInFlight(false), hasPendingPkt(false), 148 pendingPkt(nullptr) 149{ 150} 151 152template<class Impl> 153void 154LSQUnit<Impl>::init(O3CPU *cpu_ptr, IEW *iew_ptr, DerivO3CPUParams *params, 155 LSQ *lsq_ptr, unsigned maxLQEntries, unsigned maxSQEntries, 156 unsigned id) 157{ 158 cpu = cpu_ptr; 159 iewStage = iew_ptr; 160 161 lsq = lsq_ptr; 162 163 lsqID = id; 164 165 DPRINTF(LSQUnit, "Creating LSQUnit%i object.\n",id); 166 167 // Add 1 for the sentinel entry (they are circular queues). 168 LQEntries = maxLQEntries + 1; 169 SQEntries = maxSQEntries + 1; 170 171 //Due to uint8_t index in LSQSenderState 172 assert(LQEntries <= 256); 173 assert(SQEntries <= 256); 174 175 loadQueue.resize(LQEntries); 176 storeQueue.resize(SQEntries); 177 178 depCheckShift = params->LSQDepCheckShift; 179 checkLoads = params->LSQCheckLoads; 180 cacheStorePorts = params->cacheStorePorts; 181 needsTSO = params->needsTSO; 182 183 resetState(); 184} 185 186 187template<class Impl> 188void 189LSQUnit<Impl>::resetState() 190{ 191 loads = stores = storesToWB = 0; 192 193 loadHead = loadTail = 0; 194 195 storeHead = storeWBIdx = storeTail = 0; 196 197 usedStorePorts = 0; 198 199 retryPkt = NULL; 200 memDepViolator = NULL; 201 202 stalled = false; 203 204 cacheBlockMask = ~(cpu->cacheLineSize() - 1); 205} 206 207template<class Impl> 208std::string 209LSQUnit<Impl>::name() const 210{ 211 if (Impl::MaxThreads == 1) { 212 return iewStage->name() + ".lsq"; 213 } else { 214 return iewStage->name() + ".lsq.thread" + std::to_string(lsqID); 215 } 216} 217 218template<class Impl> 219void 220LSQUnit<Impl>::regStats() 221{ 222 lsqForwLoads 223 .name(name() + ".forwLoads") 224 .desc("Number of loads that had data forwarded from stores"); 225 226 invAddrLoads 227 .name(name() + ".invAddrLoads") 228 .desc("Number of loads ignored due to an invalid address"); 229 230 lsqSquashedLoads 231 .name(name() + ".squashedLoads") 232 .desc("Number of loads squashed"); 233 234 lsqIgnoredResponses 235 .name(name() + ".ignoredResponses") 236 .desc("Number of memory responses ignored because the instruction is squashed"); 237 238 lsqMemOrderViolation 239 .name(name() + ".memOrderViolation") 240 .desc("Number of memory ordering violations"); 241 242 lsqSquashedStores 243 .name(name() + ".squashedStores") 244 .desc("Number of stores squashed"); 245 246 invAddrSwpfs 247 .name(name() + ".invAddrSwpfs") 248 .desc("Number of software prefetches ignored due to an invalid address"); 249 250 lsqBlockedLoads 251 .name(name() + ".blockedLoads") 252 .desc("Number of blocked loads due to partial load-store forwarding"); 253 254 lsqRescheduledLoads 255 .name(name() + ".rescheduledLoads") 256 .desc("Number of loads that were rescheduled"); 257 258 lsqCacheBlocked 259 .name(name() + ".cacheBlocked") 260 .desc("Number of times an access to memory failed due to the cache being blocked"); 261} 262 263template<class Impl> 264void 265LSQUnit<Impl>::setDcachePort(MasterPort *dcache_port) 266{ 267 dcachePort = dcache_port; 268} 269 270template<class Impl> 271void 272LSQUnit<Impl>::clearLQ() 273{ 274 loadQueue.clear(); 275} 276 277template<class Impl> 278void 279LSQUnit<Impl>::clearSQ() 280{ 281 storeQueue.clear(); 282} 283 284template<class Impl> 285void 286LSQUnit<Impl>::drainSanityCheck() const 287{ 288 for (int i = 0; i < loadQueue.size(); ++i) 289 assert(!loadQueue[i]); 290 291 assert(storesToWB == 0); 292 assert(!retryPkt); 293} 294 295template<class Impl> 296void 297LSQUnit<Impl>::takeOverFrom() 298{ 299 resetState(); 300} 301 302template<class Impl> 303void 304LSQUnit<Impl>::resizeLQ(unsigned size) 305{ 306 unsigned size_plus_sentinel = size + 1; 307 assert(size_plus_sentinel >= LQEntries); 308 309 if (size_plus_sentinel > LQEntries) { 310 while (size_plus_sentinel > loadQueue.size()) { 311 DynInstPtr dummy; 312 loadQueue.push_back(dummy); 313 LQEntries++; 314 } 315 } else { 316 LQEntries = size_plus_sentinel; 317 } 318 319 assert(LQEntries <= 256); 320} 321 322template<class Impl> 323void 324LSQUnit<Impl>::resizeSQ(unsigned size) 325{ 326 unsigned size_plus_sentinel = size + 1; 327 if (size_plus_sentinel > SQEntries) { 328 while (size_plus_sentinel > storeQueue.size()) { 329 SQEntry dummy; 330 storeQueue.push_back(dummy); 331 SQEntries++; 332 } 333 } else { 334 SQEntries = size_plus_sentinel; 335 } 336 337 assert(SQEntries <= 256); 338} 339 340template <class Impl> 341void 342LSQUnit<Impl>::insert(const DynInstPtr &inst) 343{ 344 assert(inst->isMemRef()); 345 346 assert(inst->isLoad() || inst->isStore()); 347 348 if (inst->isLoad()) { 349 insertLoad(inst); 350 } else { 351 insertStore(inst); 352 } 353 354 inst->setInLSQ(); 355} 356 357template <class Impl> 358void 359LSQUnit<Impl>::insertLoad(const DynInstPtr &load_inst) 360{ 361 assert((loadTail + 1) % LQEntries != loadHead); 362 assert(loads < LQEntries); 363 364 DPRINTF(LSQUnit, "Inserting load PC %s, idx:%i [sn:%lli]\n", 365 load_inst->pcState(), loadTail, load_inst->seqNum); 366 367 load_inst->lqIdx = loadTail; 368 369 if (stores == 0) { 370 load_inst->sqIdx = -1; 371 } else { 372 load_inst->sqIdx = storeTail; 373 } 374 375 loadQueue[loadTail] = load_inst; 376 377 incrLdIdx(loadTail); 378 379 ++loads; 380} 381 382template <class Impl> 383void 384LSQUnit<Impl>::insertStore(const DynInstPtr &store_inst) 385{ 386 // Make sure it is not full before inserting an instruction. 387 assert((storeTail + 1) % SQEntries != storeHead); 388 assert(stores < SQEntries); 389 390 DPRINTF(LSQUnit, "Inserting store PC %s, idx:%i [sn:%lli]\n", 391 store_inst->pcState(), storeTail, store_inst->seqNum); 392 393 store_inst->sqIdx = storeTail; 394 store_inst->lqIdx = loadTail; 395 396 storeQueue[storeTail] = SQEntry(store_inst); 397 398 incrStIdx(storeTail); 399 400 ++stores; 401} 402 403template <class Impl> 404typename Impl::DynInstPtr 405LSQUnit<Impl>::getMemDepViolator() 406{ 407 DynInstPtr temp = memDepViolator; 408 409 memDepViolator = NULL; 410 411 return temp; 412} 413 414template <class Impl> 415unsigned 416LSQUnit<Impl>::numFreeLoadEntries() 417{ 418 //LQ has an extra dummy entry to differentiate 419 //empty/full conditions. Subtract 1 from the free entries. 420 DPRINTF(LSQUnit, "LQ size: %d, #loads occupied: %d\n", LQEntries, loads); 421 return LQEntries - loads - 1; 422} 423 424template <class Impl> 425unsigned 426LSQUnit<Impl>::numFreeStoreEntries() 427{ 428 //SQ has an extra dummy entry to differentiate 429 //empty/full conditions. Subtract 1 from the free entries. 430 DPRINTF(LSQUnit, "SQ size: %d, #stores occupied: %d\n", SQEntries, stores); 431 return SQEntries - stores - 1; 432 433 } 434 435template <class Impl> 436void 437LSQUnit<Impl>::checkSnoop(PacketPtr pkt) 438{ 439 // Should only ever get invalidations in here 440 assert(pkt->isInvalidate()); 441 442 int load_idx = loadHead; 443 DPRINTF(LSQUnit, "Got snoop for address %#x\n", pkt->getAddr()); 444 445 // Only Invalidate packet calls checkSnoop 446 assert(pkt->isInvalidate()); 447 for (int x = 0; x < cpu->numContexts(); x++) { 448 ThreadContext *tc = cpu->getContext(x); 449 bool no_squash = cpu->thread[x]->noSquashFromTC; 450 cpu->thread[x]->noSquashFromTC = true; 451 TheISA::handleLockedSnoop(tc, pkt, cacheBlockMask); 452 cpu->thread[x]->noSquashFromTC = no_squash; 453 } 454 455 Addr invalidate_addr = pkt->getAddr() & cacheBlockMask; 456 457 DynInstPtr ld_inst = loadQueue[load_idx]; 458 if (ld_inst) { 459 Addr load_addr_low = ld_inst->physEffAddrLow & cacheBlockMask; 460 Addr load_addr_high = ld_inst->physEffAddrHigh & cacheBlockMask; 461 462 // Check that this snoop didn't just invalidate our lock flag 463 if (ld_inst->effAddrValid() && (load_addr_low == invalidate_addr 464 || load_addr_high == invalidate_addr) 465 && ld_inst->memReqFlags & Request::LLSC) 466 TheISA::handleLockedSnoopHit(ld_inst.get()); 467 } 468 469 // If this is the only load in the LSQ we don't care 470 if (load_idx == loadTail) 471 return; 472 473 incrLdIdx(load_idx); 474 475 bool force_squash = false; 476 477 while (load_idx != loadTail) { 478 DynInstPtr ld_inst = loadQueue[load_idx]; 479 480 if (!ld_inst->effAddrValid() || ld_inst->strictlyOrdered()) { 481 incrLdIdx(load_idx); 482 continue; 483 } 484 485 Addr load_addr_low = ld_inst->physEffAddrLow & cacheBlockMask; 486 Addr load_addr_high = ld_inst->physEffAddrHigh & cacheBlockMask; 487 488 DPRINTF(LSQUnit, "-- inst [sn:%lli] load_addr: %#x to pktAddr:%#x\n", 489 ld_inst->seqNum, load_addr_low, invalidate_addr); 490 491 if ((load_addr_low == invalidate_addr 492 || load_addr_high == invalidate_addr) || force_squash) { 493 if (needsTSO) { 494 // If we have a TSO system, as all loads must be ordered with 495 // all other loads, this load as well as *all* subsequent loads 496 // need to be squashed to prevent possible load reordering. 497 force_squash = true; 498 } 499 if (ld_inst->possibleLoadViolation() || force_squash) { 500 DPRINTF(LSQUnit, "Conflicting load at addr %#x [sn:%lli]\n", 501 pkt->getAddr(), ld_inst->seqNum); 502 503 // Mark the load for re-execution 504 ld_inst->fault = std::make_shared<ReExec>(); 505 } else { 506 DPRINTF(LSQUnit, "HitExternal Snoop for addr %#x [sn:%lli]\n", 507 pkt->getAddr(), ld_inst->seqNum); 508 509 // Make sure that we don't lose a snoop hitting a LOCKED 510 // address since the LOCK* flags don't get updated until 511 // commit. 512 if (ld_inst->memReqFlags & Request::LLSC) 513 TheISA::handleLockedSnoopHit(ld_inst.get()); 514 515 // If a older load checks this and it's true 516 // then we might have missed the snoop 517 // in which case we need to invalidate to be sure 518 ld_inst->hitExternalSnoop(true); 519 } 520 } 521 incrLdIdx(load_idx); 522 } 523 return; 524} 525 526template <class Impl> 527Fault 528LSQUnit<Impl>::checkViolations(int load_idx, const DynInstPtr &inst) 529{ 530 Addr inst_eff_addr1 = inst->effAddr >> depCheckShift; 531 Addr inst_eff_addr2 = (inst->effAddr + inst->effSize - 1) >> depCheckShift; 532 533 /** @todo in theory you only need to check an instruction that has executed 534 * however, there isn't a good way in the pipeline at the moment to check 535 * all instructions that will execute before the store writes back. Thus, 536 * like the implementation that came before it, we're overly conservative. 537 */ 538 while (load_idx != loadTail) { 539 DynInstPtr ld_inst = loadQueue[load_idx]; 540 if (!ld_inst->effAddrValid() || ld_inst->strictlyOrdered()) { 541 incrLdIdx(load_idx); 542 continue; 543 } 544 545 Addr ld_eff_addr1 = ld_inst->effAddr >> depCheckShift; 546 Addr ld_eff_addr2 = 547 (ld_inst->effAddr + ld_inst->effSize - 1) >> depCheckShift; 548 549 if (inst_eff_addr2 >= ld_eff_addr1 && inst_eff_addr1 <= ld_eff_addr2) { 550 if (inst->isLoad()) { 551 // If this load is to the same block as an external snoop 552 // invalidate that we've observed then the load needs to be 553 // squashed as it could have newer data 554 if (ld_inst->hitExternalSnoop()) { 555 if (!memDepViolator || 556 ld_inst->seqNum < memDepViolator->seqNum) { 557 DPRINTF(LSQUnit, "Detected fault with inst [sn:%lli] " 558 "and [sn:%lli] at address %#x\n", 559 inst->seqNum, ld_inst->seqNum, ld_eff_addr1); 560 memDepViolator = ld_inst; 561 562 ++lsqMemOrderViolation; 563 564 return std::make_shared<GenericISA::M5PanicFault>( 565 "Detected fault with inst [sn:%lli] and " 566 "[sn:%lli] at address %#x\n", 567 inst->seqNum, ld_inst->seqNum, ld_eff_addr1); 568 } 569 } 570 571 // Otherwise, mark the load has a possible load violation 572 // and if we see a snoop before it's commited, we need to squash 573 ld_inst->possibleLoadViolation(true); 574 DPRINTF(LSQUnit, "Found possible load violation at addr: %#x" 575 " between instructions [sn:%lli] and [sn:%lli]\n", 576 inst_eff_addr1, inst->seqNum, ld_inst->seqNum); 577 } else { 578 // A load/store incorrectly passed this store. 579 // Check if we already have a violator, or if it's newer 580 // squash and refetch. 581 if (memDepViolator && ld_inst->seqNum > memDepViolator->seqNum) 582 break; 583 584 DPRINTF(LSQUnit, "Detected fault with inst [sn:%lli] and " 585 "[sn:%lli] at address %#x\n", 586 inst->seqNum, ld_inst->seqNum, ld_eff_addr1); 587 memDepViolator = ld_inst; 588 589 ++lsqMemOrderViolation; 590 591 return std::make_shared<GenericISA::M5PanicFault>( 592 "Detected fault with " 593 "inst [sn:%lli] and [sn:%lli] at address %#x\n", 594 inst->seqNum, ld_inst->seqNum, ld_eff_addr1); 595 } 596 } 597 598 incrLdIdx(load_idx); 599 } 600 return NoFault; 601} 602 603 604 605 606template <class Impl> 607Fault 608LSQUnit<Impl>::executeLoad(const DynInstPtr &inst) 609{ 610 using namespace TheISA; 611 // Execute a specific load. 612 Fault load_fault = NoFault; 613 614 DPRINTF(LSQUnit, "Executing load PC %s, [sn:%lli]\n", 615 inst->pcState(), inst->seqNum); 616 617 assert(!inst->isSquashed()); 618 619 load_fault = inst->initiateAcc(); 620 621 if (inst->isTranslationDelayed() && 622 load_fault == NoFault) 623 return load_fault; 624 625 // If the instruction faulted or predicated false, then we need to send it 626 // along to commit without the instruction completing. 627 if (load_fault != NoFault || !inst->readPredicate()) { 628 // Send this instruction to commit, also make sure iew stage 629 // realizes there is activity. Mark it as executed unless it 630 // is a strictly ordered load that needs to hit the head of 631 // commit. 632 if (!inst->readPredicate()) 633 inst->forwardOldRegs(); 634 DPRINTF(LSQUnit, "Load [sn:%lli] not executed from %s\n", 635 inst->seqNum, 636 (load_fault != NoFault ? "fault" : "predication")); 637 if (!(inst->hasRequest() && inst->strictlyOrdered()) || 638 inst->isAtCommit()) { 639 inst->setExecuted(); 640 } 641 iewStage->instToCommit(inst); 642 iewStage->activityThisCycle(); 643 } else { 644 assert(inst->effAddrValid()); 645 int load_idx = inst->lqIdx; 646 incrLdIdx(load_idx); 647 648 if (checkLoads) 649 return checkViolations(load_idx, inst); 650 } 651 652 return load_fault; 653} 654 655template <class Impl> 656Fault 657LSQUnit<Impl>::executeStore(const DynInstPtr &store_inst) 658{ 659 using namespace TheISA; 660 // Make sure that a store exists. 661 assert(stores != 0); 662 663 int store_idx = store_inst->sqIdx; 664 665 DPRINTF(LSQUnit, "Executing store PC %s [sn:%lli]\n", 666 store_inst->pcState(), store_inst->seqNum); 667 668 assert(!store_inst->isSquashed()); 669 670 // Check the recently completed loads to see if any match this store's 671 // address. If so, then we have a memory ordering violation. 672 int load_idx = store_inst->lqIdx; 673 674 Fault store_fault = store_inst->initiateAcc(); 675 676 if (store_inst->isTranslationDelayed() && 677 store_fault == NoFault) 678 return store_fault; 679 680 if (!store_inst->readPredicate()) { 681 DPRINTF(LSQUnit, "Store [sn:%lli] not executed from predication\n", 682 store_inst->seqNum); 683 store_inst->forwardOldRegs(); 684 return store_fault; 685 } 686 687 if (storeQueue[store_idx].size == 0) { 688 DPRINTF(LSQUnit,"Fault on Store PC %s, [sn:%lli], Size = 0\n", 689 store_inst->pcState(), store_inst->seqNum); 690 691 return store_fault; 692 } 693 694 assert(store_fault == NoFault); 695 696 if (store_inst->isStoreConditional()) { 697 // Store conditionals need to set themselves as able to 698 // writeback if we haven't had a fault by here. 699 storeQueue[store_idx].canWB = true; 700 701 ++storesToWB; 702 } 703 704 return checkViolations(load_idx, store_inst); 705 706} 707 708template <class Impl> 709void 710LSQUnit<Impl>::commitLoad() 711{ 712 assert(loadQueue[loadHead]); 713 714 DPRINTF(LSQUnit, "Committing head load instruction, PC %s\n", 715 loadQueue[loadHead]->pcState()); 716 717 loadQueue[loadHead] = NULL; 718 719 incrLdIdx(loadHead); 720 721 --loads; 722} 723 724template <class Impl> 725void 726LSQUnit<Impl>::commitLoads(InstSeqNum &youngest_inst) 727{ 728 assert(loads == 0 || loadQueue[loadHead]); 729 730 while (loads != 0 && loadQueue[loadHead]->seqNum <= youngest_inst) { 731 commitLoad(); 732 } 733} 734 735template <class Impl> 736void 737LSQUnit<Impl>::commitStores(InstSeqNum &youngest_inst) 738{ 739 assert(stores == 0 || storeQueue[storeHead].inst); 740 741 int store_idx = storeHead; 742 743 while (store_idx != storeTail) { 744 assert(storeQueue[store_idx].inst); 745 // Mark any stores that are now committed and have not yet 746 // been marked as able to write back. 747 if (!storeQueue[store_idx].canWB) { 748 if (storeQueue[store_idx].inst->seqNum > youngest_inst) { 749 break; 750 } 751 DPRINTF(LSQUnit, "Marking store as able to write back, PC " 752 "%s [sn:%lli]\n", 753 storeQueue[store_idx].inst->pcState(), 754 storeQueue[store_idx].inst->seqNum); 755 756 storeQueue[store_idx].canWB = true; 757 758 ++storesToWB; 759 } 760 761 incrStIdx(store_idx); 762 } 763} 764 765template <class Impl> 766void 767LSQUnit<Impl>::writebackPendingStore() 768{ 769 if (hasPendingPkt) { 770 assert(pendingPkt != NULL); 771 772 // If the cache is blocked, this will store the packet for retry. 773 if (sendStore(pendingPkt)) { 774 storePostSend(pendingPkt); 775 } 776 pendingPkt = NULL; 777 hasPendingPkt = false; 778 } 779} 780 781template <class Impl> 782void 783LSQUnit<Impl>::writebackStores() 784{ 785 // First writeback the second packet from any split store that didn't 786 // complete last cycle because there weren't enough cache ports available. 787 if (TheISA::HasUnalignedMemAcc) { 788 writebackPendingStore(); 789 } 790 791 while (storesToWB > 0 && 792 storeWBIdx != storeTail && 793 storeQueue[storeWBIdx].inst && 794 storeQueue[storeWBIdx].canWB && 795 ((!needsTSO) || (!storeInFlight)) && 796 usedStorePorts < cacheStorePorts) { 797 798 if (isStoreBlocked) { 799 DPRINTF(LSQUnit, "Unable to write back any more stores, cache" 800 " is blocked!\n"); 801 break; 802 } 803 804 // Store didn't write any data so no need to write it back to 805 // memory. 806 if (storeQueue[storeWBIdx].size == 0) { 807 completeStore(storeWBIdx); 808 809 incrStIdx(storeWBIdx); 810 811 continue; 812 } 813 814 ++usedStorePorts; 815 816 if (storeQueue[storeWBIdx].inst->isDataPrefetch()) { 817 incrStIdx(storeWBIdx); 818 819 continue; 820 } 821 822 assert(storeQueue[storeWBIdx].req); 823 assert(!storeQueue[storeWBIdx].committed); 824 825 if (TheISA::HasUnalignedMemAcc && storeQueue[storeWBIdx].isSplit) { 826 assert(storeQueue[storeWBIdx].sreqLow); 827 assert(storeQueue[storeWBIdx].sreqHigh); 828 } 829 830 DynInstPtr inst = storeQueue[storeWBIdx].inst; 831 832 RequestPtr &req = storeQueue[storeWBIdx].req; 833 const RequestPtr &sreqLow = storeQueue[storeWBIdx].sreqLow; 834 const RequestPtr &sreqHigh = storeQueue[storeWBIdx].sreqHigh; 835 836 storeQueue[storeWBIdx].committed = true; 837 838 assert(!inst->memData); 839 inst->memData = new uint8_t[req->getSize()]; 840 841 if (storeQueue[storeWBIdx].isAllZeros) 842 memset(inst->memData, 0, req->getSize()); 843 else 844 memcpy(inst->memData, storeQueue[storeWBIdx].data, req->getSize()); 845 846 PacketPtr data_pkt; 847 PacketPtr snd_data_pkt = NULL; 848 849 LSQSenderState *state = new LSQSenderState; 850 state->isLoad = false; 851 state->idx = storeWBIdx; 852 state->inst = inst; 853 854 if (!TheISA::HasUnalignedMemAcc || !storeQueue[storeWBIdx].isSplit) { 855 856 // Build a single data packet if the store isn't split. 857 data_pkt = Packet::createWrite(req); 858 data_pkt->dataStatic(inst->memData); 859 data_pkt->senderState = state; 860 } else { 861 // Create two packets if the store is split in two. 862 data_pkt = Packet::createWrite(sreqLow); 863 snd_data_pkt = Packet::createWrite(sreqHigh); 864 865 data_pkt->dataStatic(inst->memData); 866 snd_data_pkt->dataStatic(inst->memData + sreqLow->getSize()); 867 868 data_pkt->senderState = state; 869 snd_data_pkt->senderState = state; 870 871 state->isSplit = true; 872 state->outstanding = 2; 873 874 // Can delete the main request now. 875 req = sreqLow; 876 } 877 878 DPRINTF(LSQUnit, "D-Cache: Writing back store idx:%i PC:%s " 879 "to Addr:%#x, data:%#x [sn:%lli]\n", 880 storeWBIdx, inst->pcState(), 881 req->getPaddr(), (int)*(inst->memData), 882 inst->seqNum); 883 884 // @todo: Remove this SC hack once the memory system handles it. 885 if (inst->isStoreConditional()) { 886 assert(!storeQueue[storeWBIdx].isSplit); 887 // Disable recording the result temporarily. Writing to 888 // misc regs normally updates the result, but this is not 889 // the desired behavior when handling store conditionals. 890 inst->recordResult(false); 891 bool success = TheISA::handleLockedWrite(inst.get(), req, cacheBlockMask); 892 inst->recordResult(true); 893 894 if (!success) { 895 // Instantly complete this store. 896 DPRINTF(LSQUnit, "Store conditional [sn:%lli] failed. " 897 "Instantly completing it.\n", 898 inst->seqNum); 899 WritebackEvent *wb = new WritebackEvent(inst, data_pkt, this); 900 cpu->schedule(wb, curTick() + 1); 901 completeStore(storeWBIdx); 902 incrStIdx(storeWBIdx); 903 continue; 904 } 905 } else { 906 // Non-store conditionals do not need a writeback. 907 state->noWB = true; 908 } 909 910 bool split = 911 TheISA::HasUnalignedMemAcc && storeQueue[storeWBIdx].isSplit; 912 913 ThreadContext *thread = cpu->tcBase(lsqID); 914 915 if (req->isMmappedIpr()) { 916 assert(!inst->isStoreConditional()); 917 TheISA::handleIprWrite(thread, data_pkt); 918 delete data_pkt; 919 if (split) { 920 assert(snd_data_pkt->req->isMmappedIpr()); 921 TheISA::handleIprWrite(thread, snd_data_pkt); 922 delete snd_data_pkt; 923 } 924 delete state; 925 completeStore(storeWBIdx); 926 incrStIdx(storeWBIdx); 927 } else if (!sendStore(data_pkt)) { 928 DPRINTF(IEW, "D-Cache became blocked when writing [sn:%lli], will" 929 "retry later\n", 930 inst->seqNum); 931 932 // Need to store the second packet, if split. 933 if (split) { 934 state->pktToSend = true; 935 state->pendingPacket = snd_data_pkt; 936 } 937 } else { 938 939 // If split, try to send the second packet too 940 if (split) { 941 assert(snd_data_pkt); 942 943 // Ensure there are enough ports to use. 944 if (usedStorePorts < cacheStorePorts) { 945 ++usedStorePorts; 946 if (sendStore(snd_data_pkt)) { 947 storePostSend(snd_data_pkt); 948 } else { 949 DPRINTF(IEW, "D-Cache became blocked when writing" 950 " [sn:%lli] second packet, will retry later\n", 951 inst->seqNum); 952 } 953 } else { 954 955 // Store the packet for when there's free ports. 956 assert(pendingPkt == NULL); 957 pendingPkt = snd_data_pkt; 958 hasPendingPkt = true; 959 } 960 } else { 961 962 // Not a split store. 963 storePostSend(data_pkt); 964 } 965 } 966 } 967 968 // Not sure this should set it to 0. 969 usedStorePorts = 0; 970 971 assert(stores >= 0 && storesToWB >= 0); 972} 973 974/*template <class Impl> 975void 976LSQUnit<Impl>::removeMSHR(InstSeqNum seqNum) 977{ 978 list<InstSeqNum>::iterator mshr_it = find(mshrSeqNums.begin(), 979 mshrSeqNums.end(), 980 seqNum); 981 982 if (mshr_it != mshrSeqNums.end()) { 983 mshrSeqNums.erase(mshr_it); 984 DPRINTF(LSQUnit, "Removing MSHR. count = %i\n",mshrSeqNums.size()); 985 } 986}*/ 987 988template <class Impl> 989void 990LSQUnit<Impl>::squash(const InstSeqNum &squashed_num) 991{ 992 DPRINTF(LSQUnit, "Squashing until [sn:%lli]!" 993 "(Loads:%i Stores:%i)\n", squashed_num, loads, stores); 994 995 int load_idx = loadTail; 996 decrLdIdx(load_idx); 997 998 while (loads != 0 && loadQueue[load_idx]->seqNum > squashed_num) { 999 DPRINTF(LSQUnit,"Load Instruction PC %s squashed, " 1000 "[sn:%lli]\n", 1001 loadQueue[load_idx]->pcState(), 1002 loadQueue[load_idx]->seqNum); 1003 1004 if (isStalled() && load_idx == stallingLoadIdx) { 1005 stalled = false; 1006 stallingStoreIsn = 0; 1007 stallingLoadIdx = 0; 1008 } 1009 1010 // Clear the smart pointer to make sure it is decremented. 1011 loadQueue[load_idx]->setSquashed(); 1012 loadQueue[load_idx] = NULL; 1013 --loads; 1014 1015 // Inefficient! 1016 loadTail = load_idx; 1017 1018 decrLdIdx(load_idx); 1019 ++lsqSquashedLoads; 1020 } 1021 1022 if (memDepViolator && squashed_num < memDepViolator->seqNum) { 1023 memDepViolator = NULL; 1024 } 1025 1026 int store_idx = storeTail; 1027 decrStIdx(store_idx); 1028 1029 while (stores != 0 && 1030 storeQueue[store_idx].inst->seqNum > squashed_num) { 1031 // Instructions marked as can WB are already committed. 1032 if (storeQueue[store_idx].canWB) { 1033 break; 1034 } 1035 1036 DPRINTF(LSQUnit,"Store Instruction PC %s squashed, " 1037 "idx:%i [sn:%lli]\n", 1038 storeQueue[store_idx].inst->pcState(), 1039 store_idx, storeQueue[store_idx].inst->seqNum); 1040 1041 // I don't think this can happen. It should have been cleared 1042 // by the stalling load. 1043 if (isStalled() && 1044 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) { 1045 panic("Is stalled should have been cleared by stalling load!\n"); 1046 stalled = false; 1047 stallingStoreIsn = 0; 1048 } 1049 1050 // Clear the smart pointer to make sure it is decremented. 1051 storeQueue[store_idx].inst->setSquashed(); 1052 storeQueue[store_idx].inst = NULL; 1053 storeQueue[store_idx].canWB = 0; 1054 1055 // Must delete request now that it wasn't handed off to 1056 // memory. This is quite ugly. @todo: Figure out the proper 1057 // place to really handle request deletes. 1058 storeQueue[store_idx].req.reset(); 1059 if (TheISA::HasUnalignedMemAcc && storeQueue[store_idx].isSplit) { 1060 storeQueue[store_idx].sreqLow.reset(); 1061 storeQueue[store_idx].sreqHigh.reset(); 1062 } 1063 1064 --stores; 1065 1066 // Inefficient! 1067 storeTail = store_idx; 1068 1069 decrStIdx(store_idx); 1070 ++lsqSquashedStores; 1071 } 1072} 1073 1074template <class Impl> 1075void 1076LSQUnit<Impl>::storePostSend(PacketPtr pkt) 1077{ 1078 if (isStalled() && 1079 storeQueue[storeWBIdx].inst->seqNum == stallingStoreIsn) { 1080 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] " 1081 "load idx:%i\n", 1082 stallingStoreIsn, stallingLoadIdx); 1083 stalled = false; 1084 stallingStoreIsn = 0; 1085 iewStage->replayMemInst(loadQueue[stallingLoadIdx]); 1086 } 1087 1088 if (!storeQueue[storeWBIdx].inst->isStoreConditional()) { 1089 // The store is basically completed at this time. This 1090 // only works so long as the checker doesn't try to 1091 // verify the value in memory for stores. 1092 storeQueue[storeWBIdx].inst->setCompleted(); 1093 1094 if (cpu->checker) { 1095 cpu->checker->verify(storeQueue[storeWBIdx].inst); 1096 } 1097 } 1098 1099 if (needsTSO) { 1100 storeInFlight = true; 1101 } 1102 1103 incrStIdx(storeWBIdx); 1104} 1105 1106template <class Impl> 1107void 1108LSQUnit<Impl>::writeback(const DynInstPtr &inst, PacketPtr pkt) 1109{ 1110 iewStage->wakeCPU(); 1111 1112 // Squashed instructions do not need to complete their access. 1113 if (inst->isSquashed()) { 1114 assert(!inst->isStore()); 1115 ++lsqIgnoredResponses; 1116 return; 1117 } 1118 1119 if (!inst->isExecuted()) { 1120 inst->setExecuted(); 1121 1122 if (inst->fault == NoFault) { 1123 // Complete access to copy data to proper place. 1124 inst->completeAcc(pkt); 1125 } else { 1126 // If the instruction has an outstanding fault, we cannot complete 1127 // the access as this discards the current fault. 1128 1129 // If we have an outstanding fault, the fault should only be of 1130 // type ReExec. 1131 assert(dynamic_cast<ReExec*>(inst->fault.get()) != nullptr); 1132 1133 DPRINTF(LSQUnit, "Not completing instruction [sn:%lli] access " 1134 "due to pending fault.\n", inst->seqNum); 1135 } 1136 } 1137 1138 // Need to insert instruction into queue to commit 1139 iewStage->instToCommit(inst); 1140 1141 iewStage->activityThisCycle(); 1142 1143 // see if this load changed the PC 1144 iewStage->checkMisprediction(inst); 1145} 1146 1147template <class Impl> 1148void 1149LSQUnit<Impl>::completeStore(int store_idx) 1150{ 1151 assert(storeQueue[store_idx].inst); 1152 storeQueue[store_idx].completed = true; 1153 --storesToWB; 1154 // A bit conservative because a store completion may not free up entries, 1155 // but hopefully avoids two store completions in one cycle from making 1156 // the CPU tick twice. 1157 cpu->wakeCPU(); 1158 cpu->activityThisCycle(); 1159 1160 if (store_idx == storeHead) { 1161 do { 1162 incrStIdx(storeHead); 1163 1164 --stores; 1165 } while (storeQueue[storeHead].completed && 1166 storeHead != storeTail); 1167 1168 iewStage->updateLSQNextCycle = true; 1169 } 1170 1171 DPRINTF(LSQUnit, "Completing store [sn:%lli], idx:%i, store head " 1172 "idx:%i\n", 1173 storeQueue[store_idx].inst->seqNum, store_idx, storeHead); 1174 1175#if TRACING_ON 1176 if (DTRACE(O3PipeView)) { 1177 storeQueue[store_idx].inst->storeTick = 1178 curTick() - storeQueue[store_idx].inst->fetchTick; 1179 } 1180#endif 1181 1182 if (isStalled() && 1183 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) { 1184 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] " 1185 "load idx:%i\n", 1186 stallingStoreIsn, stallingLoadIdx); 1187 stalled = false; 1188 stallingStoreIsn = 0; 1189 iewStage->replayMemInst(loadQueue[stallingLoadIdx]); 1190 } 1191 1192 storeQueue[store_idx].inst->setCompleted(); 1193 1194 if (needsTSO) { 1195 storeInFlight = false; 1196 } 1197 1198 // Tell the checker we've completed this instruction. Some stores 1199 // may get reported twice to the checker, but the checker can 1200 // handle that case. 1201 1202 // Store conditionals cannot be sent to the checker yet, they have 1203 // to update the misc registers first which should take place 1204 // when they commit 1205 if (cpu->checker && !storeQueue[store_idx].inst->isStoreConditional()) { 1206 cpu->checker->verify(storeQueue[store_idx].inst); 1207 } 1208} 1209 1210template <class Impl> 1211bool 1212LSQUnit<Impl>::sendStore(PacketPtr data_pkt) 1213{ 1214 if (!dcachePort->sendTimingReq(data_pkt)) { 1215 // Need to handle becoming blocked on a store. 1216 isStoreBlocked = true; 1217 ++lsqCacheBlocked; 1218 assert(retryPkt == NULL); 1219 retryPkt = data_pkt; 1220 return false; 1221 } 1222 return true; 1223} 1224 1225template <class Impl> 1226void 1227LSQUnit<Impl>::recvRetry() 1228{ 1229 if (isStoreBlocked) { 1230 DPRINTF(LSQUnit, "Receiving retry: store blocked\n"); 1231 assert(retryPkt != NULL); 1232 1233 LSQSenderState *state = 1234 dynamic_cast<LSQSenderState *>(retryPkt->senderState); 1235 1236 if (dcachePort->sendTimingReq(retryPkt)) { 1237 // Don't finish the store unless this is the last packet. 1238 if (!TheISA::HasUnalignedMemAcc || !state->pktToSend || 1239 state->pendingPacket == retryPkt) { 1240 state->pktToSend = false; 1241 storePostSend(retryPkt); 1242 } 1243 retryPkt = NULL; 1244 isStoreBlocked = false; 1245 1246 // Send any outstanding packet. 1247 if (TheISA::HasUnalignedMemAcc && state->pktToSend) { 1248 assert(state->pendingPacket); 1249 if (sendStore(state->pendingPacket)) { 1250 storePostSend(state->pendingPacket); 1251 } 1252 } 1253 } else { 1254 // Still blocked! 1255 ++lsqCacheBlocked; 1256 } 1257 } 1258} 1259 1260template <class Impl> 1261inline void 1262LSQUnit<Impl>::incrStIdx(int &store_idx) const 1263{ 1264 if (++store_idx >= SQEntries) 1265 store_idx = 0; 1266} 1267 1268template <class Impl> 1269inline void 1270LSQUnit<Impl>::decrStIdx(int &store_idx) const 1271{ 1272 if (--store_idx < 0) 1273 store_idx += SQEntries; 1274} 1275 1276template <class Impl> 1277inline void 1278LSQUnit<Impl>::incrLdIdx(int &load_idx) const 1279{ 1280 if (++load_idx >= LQEntries) 1281 load_idx = 0; 1282} 1283 1284template <class Impl> 1285inline void 1286LSQUnit<Impl>::decrLdIdx(int &load_idx) const 1287{ 1288 if (--load_idx < 0) 1289 load_idx += LQEntries; 1290} 1291 1292template <class Impl> 1293void 1294LSQUnit<Impl>::dumpInsts() const 1295{ 1296 cprintf("Load store queue: Dumping instructions.\n"); 1297 cprintf("Load queue size: %i\n", loads); 1298 cprintf("Load queue: "); 1299 1300 int load_idx = loadHead; 1301 1302 while (load_idx != loadTail && loadQueue[load_idx]) { 1303 const DynInstPtr &inst(loadQueue[load_idx]); 1304 cprintf("%s.[sn:%i] ", inst->pcState(), inst->seqNum); 1305 1306 incrLdIdx(load_idx); 1307 } 1308 cprintf("\n"); 1309 1310 cprintf("Store queue size: %i\n", stores); 1311 cprintf("Store queue: "); 1312 1313 int store_idx = storeHead; 1314 1315 while (store_idx != storeTail && storeQueue[store_idx].inst) { 1316 const DynInstPtr &inst(storeQueue[store_idx].inst); 1317 cprintf("%s.[sn:%i] ", inst->pcState(), inst->seqNum); 1318 1319 incrStIdx(store_idx); 1320 } 1321 1322 cprintf("\n"); 1323} 1324 1325#endif//__CPU_O3_LSQ_UNIT_IMPL_HH__ 1326