lsq_unit_impl.hh revision 13472:7ceacede4f1e
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(uint32_t lqEntries, uint32_t sqEntries) 146 : lsqID(-1), storeQueue(sqEntries+1), loadQueue(lqEntries+1), 147 LQEntries(lqEntries+1), SQEntries(lqEntries+1), 148 loads(0), stores(0), storesToWB(0), cacheBlockMask(0), stalled(false), 149 isStoreBlocked(false), storeInFlight(false), hasPendingPkt(false), 150 pendingPkt(nullptr) 151{ 152} 153 154template<class Impl> 155void 156LSQUnit<Impl>::init(O3CPU *cpu_ptr, IEW *iew_ptr, DerivO3CPUParams *params, 157 LSQ *lsq_ptr, unsigned id) 158{ 159 lsqID = id; 160 161 cpu = cpu_ptr; 162 iewStage = iew_ptr; 163 164 lsq = lsq_ptr; 165 166 DPRINTF(LSQUnit, "Creating LSQUnit%i object.\n",lsqID); 167 168 depCheckShift = params->LSQDepCheckShift; 169 checkLoads = params->LSQCheckLoads; 170 cacheStorePorts = params->cacheStorePorts; 171 needsTSO = params->needsTSO; 172 173 resetState(); 174} 175 176 177template<class Impl> 178void 179LSQUnit<Impl>::resetState() 180{ 181 loads = stores = storesToWB = 0; 182 183 loadHead = loadTail = 0; 184 185 storeHead = storeWBIdx = storeTail = 0; 186 187 usedStorePorts = 0; 188 189 retryPkt = NULL; 190 memDepViolator = NULL; 191 192 stalled = false; 193 194 cacheBlockMask = ~(cpu->cacheLineSize() - 1); 195} 196 197template<class Impl> 198std::string 199LSQUnit<Impl>::name() const 200{ 201 if (Impl::MaxThreads == 1) { 202 return iewStage->name() + ".lsq"; 203 } else { 204 return iewStage->name() + ".lsq.thread" + std::to_string(lsqID); 205 } 206} 207 208template<class Impl> 209void 210LSQUnit<Impl>::regStats() 211{ 212 lsqForwLoads 213 .name(name() + ".forwLoads") 214 .desc("Number of loads that had data forwarded from stores"); 215 216 invAddrLoads 217 .name(name() + ".invAddrLoads") 218 .desc("Number of loads ignored due to an invalid address"); 219 220 lsqSquashedLoads 221 .name(name() + ".squashedLoads") 222 .desc("Number of loads squashed"); 223 224 lsqIgnoredResponses 225 .name(name() + ".ignoredResponses") 226 .desc("Number of memory responses ignored because the instruction is squashed"); 227 228 lsqMemOrderViolation 229 .name(name() + ".memOrderViolation") 230 .desc("Number of memory ordering violations"); 231 232 lsqSquashedStores 233 .name(name() + ".squashedStores") 234 .desc("Number of stores squashed"); 235 236 invAddrSwpfs 237 .name(name() + ".invAddrSwpfs") 238 .desc("Number of software prefetches ignored due to an invalid address"); 239 240 lsqBlockedLoads 241 .name(name() + ".blockedLoads") 242 .desc("Number of blocked loads due to partial load-store forwarding"); 243 244 lsqRescheduledLoads 245 .name(name() + ".rescheduledLoads") 246 .desc("Number of loads that were rescheduled"); 247 248 lsqCacheBlocked 249 .name(name() + ".cacheBlocked") 250 .desc("Number of times an access to memory failed due to the cache being blocked"); 251} 252 253template<class Impl> 254void 255LSQUnit<Impl>::setDcachePort(MasterPort *dcache_port) 256{ 257 dcachePort = dcache_port; 258} 259 260template<class Impl> 261void 262LSQUnit<Impl>::clearLQ() 263{ 264 loadQueue.clear(); 265} 266 267template<class Impl> 268void 269LSQUnit<Impl>::clearSQ() 270{ 271 storeQueue.clear(); 272} 273 274template<class Impl> 275void 276LSQUnit<Impl>::drainSanityCheck() const 277{ 278 for (int i = 0; i < loadQueue.size(); ++i) 279 assert(!loadQueue[i]); 280 281 assert(storesToWB == 0); 282 assert(!retryPkt); 283} 284 285template<class Impl> 286void 287LSQUnit<Impl>::takeOverFrom() 288{ 289 resetState(); 290} 291 292template<class Impl> 293void 294LSQUnit<Impl>::resizeLQ(unsigned size) 295{ 296 unsigned size_plus_sentinel = size + 1; 297 assert(size_plus_sentinel >= LQEntries); 298 299 if (size_plus_sentinel > LQEntries) { 300 while (size_plus_sentinel > loadQueue.size()) { 301 DynInstPtr dummy; 302 loadQueue.push_back(dummy); 303 LQEntries++; 304 } 305 } else { 306 LQEntries = size_plus_sentinel; 307 } 308 309 assert(LQEntries <= 256); 310} 311 312template<class Impl> 313void 314LSQUnit<Impl>::resizeSQ(unsigned size) 315{ 316 unsigned size_plus_sentinel = size + 1; 317 if (size_plus_sentinel > SQEntries) { 318 while (size_plus_sentinel > storeQueue.size()) { 319 SQEntry dummy; 320 storeQueue.push_back(dummy); 321 SQEntries++; 322 } 323 } else { 324 SQEntries = size_plus_sentinel; 325 } 326 327 assert(SQEntries <= 256); 328} 329 330template <class Impl> 331void 332LSQUnit<Impl>::insert(const DynInstPtr &inst) 333{ 334 assert(inst->isMemRef()); 335 336 assert(inst->isLoad() || inst->isStore()); 337 338 if (inst->isLoad()) { 339 insertLoad(inst); 340 } else { 341 insertStore(inst); 342 } 343 344 inst->setInLSQ(); 345} 346 347template <class Impl> 348void 349LSQUnit<Impl>::insertLoad(const DynInstPtr &load_inst) 350{ 351 assert((loadTail + 1) % LQEntries != loadHead); 352 assert(loads < LQEntries); 353 354 DPRINTF(LSQUnit, "Inserting load PC %s, idx:%i [sn:%lli]\n", 355 load_inst->pcState(), loadTail, load_inst->seqNum); 356 357 load_inst->lqIdx = loadTail; 358 359 if (stores == 0) { 360 load_inst->sqIdx = -1; 361 } else { 362 load_inst->sqIdx = storeTail; 363 } 364 365 loadQueue[loadTail] = load_inst; 366 367 incrLdIdx(loadTail); 368 369 ++loads; 370} 371 372template <class Impl> 373void 374LSQUnit<Impl>::insertStore(const DynInstPtr &store_inst) 375{ 376 // Make sure it is not full before inserting an instruction. 377 assert((storeTail + 1) % SQEntries != storeHead); 378 assert(stores < SQEntries); 379 380 DPRINTF(LSQUnit, "Inserting store PC %s, idx:%i [sn:%lli]\n", 381 store_inst->pcState(), storeTail, store_inst->seqNum); 382 383 store_inst->sqIdx = storeTail; 384 store_inst->lqIdx = loadTail; 385 386 storeQueue[storeTail] = SQEntry(store_inst); 387 388 incrStIdx(storeTail); 389 390 ++stores; 391} 392 393template <class Impl> 394typename Impl::DynInstPtr 395LSQUnit<Impl>::getMemDepViolator() 396{ 397 DynInstPtr temp = memDepViolator; 398 399 memDepViolator = NULL; 400 401 return temp; 402} 403 404template <class Impl> 405unsigned 406LSQUnit<Impl>::numFreeLoadEntries() 407{ 408 //LQ has an extra dummy entry to differentiate 409 //empty/full conditions. Subtract 1 from the free entries. 410 DPRINTF(LSQUnit, "LQ size: %d, #loads occupied: %d\n", LQEntries, loads); 411 return LQEntries - loads - 1; 412} 413 414template <class Impl> 415unsigned 416LSQUnit<Impl>::numFreeStoreEntries() 417{ 418 //SQ has an extra dummy entry to differentiate 419 //empty/full conditions. Subtract 1 from the free entries. 420 DPRINTF(LSQUnit, "SQ size: %d, #stores occupied: %d\n", SQEntries, stores); 421 return SQEntries - stores - 1; 422 423 } 424 425template <class Impl> 426void 427LSQUnit<Impl>::checkSnoop(PacketPtr pkt) 428{ 429 // Should only ever get invalidations in here 430 assert(pkt->isInvalidate()); 431 432 int load_idx = loadHead; 433 DPRINTF(LSQUnit, "Got snoop for address %#x\n", pkt->getAddr()); 434 435 // Only Invalidate packet calls checkSnoop 436 assert(pkt->isInvalidate()); 437 for (int x = 0; x < cpu->numContexts(); x++) { 438 ThreadContext *tc = cpu->getContext(x); 439 bool no_squash = cpu->thread[x]->noSquashFromTC; 440 cpu->thread[x]->noSquashFromTC = true; 441 TheISA::handleLockedSnoop(tc, pkt, cacheBlockMask); 442 cpu->thread[x]->noSquashFromTC = no_squash; 443 } 444 445 Addr invalidate_addr = pkt->getAddr() & cacheBlockMask; 446 447 DynInstPtr ld_inst = loadQueue[load_idx]; 448 if (ld_inst) { 449 Addr load_addr_low = ld_inst->physEffAddrLow & cacheBlockMask; 450 Addr load_addr_high = ld_inst->physEffAddrHigh & cacheBlockMask; 451 452 // Check that this snoop didn't just invalidate our lock flag 453 if (ld_inst->effAddrValid() && (load_addr_low == invalidate_addr 454 || load_addr_high == invalidate_addr) 455 && ld_inst->memReqFlags & Request::LLSC) 456 TheISA::handleLockedSnoopHit(ld_inst.get()); 457 } 458 459 // If this is the only load in the LSQ we don't care 460 if (load_idx == loadTail) 461 return; 462 463 incrLdIdx(load_idx); 464 465 bool force_squash = false; 466 467 while (load_idx != loadTail) { 468 DynInstPtr ld_inst = loadQueue[load_idx]; 469 470 if (!ld_inst->effAddrValid() || ld_inst->strictlyOrdered()) { 471 incrLdIdx(load_idx); 472 continue; 473 } 474 475 Addr load_addr_low = ld_inst->physEffAddrLow & cacheBlockMask; 476 Addr load_addr_high = ld_inst->physEffAddrHigh & cacheBlockMask; 477 478 DPRINTF(LSQUnit, "-- inst [sn:%lli] load_addr: %#x to pktAddr:%#x\n", 479 ld_inst->seqNum, load_addr_low, invalidate_addr); 480 481 if ((load_addr_low == invalidate_addr 482 || load_addr_high == invalidate_addr) || force_squash) { 483 if (needsTSO) { 484 // If we have a TSO system, as all loads must be ordered with 485 // all other loads, this load as well as *all* subsequent loads 486 // need to be squashed to prevent possible load reordering. 487 force_squash = true; 488 } 489 if (ld_inst->possibleLoadViolation() || force_squash) { 490 DPRINTF(LSQUnit, "Conflicting load at addr %#x [sn:%lli]\n", 491 pkt->getAddr(), ld_inst->seqNum); 492 493 // Mark the load for re-execution 494 ld_inst->fault = std::make_shared<ReExec>(); 495 } else { 496 DPRINTF(LSQUnit, "HitExternal Snoop for addr %#x [sn:%lli]\n", 497 pkt->getAddr(), ld_inst->seqNum); 498 499 // Make sure that we don't lose a snoop hitting a LOCKED 500 // address since the LOCK* flags don't get updated until 501 // commit. 502 if (ld_inst->memReqFlags & Request::LLSC) 503 TheISA::handleLockedSnoopHit(ld_inst.get()); 504 505 // If a older load checks this and it's true 506 // then we might have missed the snoop 507 // in which case we need to invalidate to be sure 508 ld_inst->hitExternalSnoop(true); 509 } 510 } 511 incrLdIdx(load_idx); 512 } 513 return; 514} 515 516template <class Impl> 517Fault 518LSQUnit<Impl>::checkViolations(int load_idx, const DynInstPtr &inst) 519{ 520 Addr inst_eff_addr1 = inst->effAddr >> depCheckShift; 521 Addr inst_eff_addr2 = (inst->effAddr + inst->effSize - 1) >> depCheckShift; 522 523 /** @todo in theory you only need to check an instruction that has executed 524 * however, there isn't a good way in the pipeline at the moment to check 525 * all instructions that will execute before the store writes back. Thus, 526 * like the implementation that came before it, we're overly conservative. 527 */ 528 while (load_idx != loadTail) { 529 DynInstPtr ld_inst = loadQueue[load_idx]; 530 if (!ld_inst->effAddrValid() || ld_inst->strictlyOrdered()) { 531 incrLdIdx(load_idx); 532 continue; 533 } 534 535 Addr ld_eff_addr1 = ld_inst->effAddr >> depCheckShift; 536 Addr ld_eff_addr2 = 537 (ld_inst->effAddr + ld_inst->effSize - 1) >> depCheckShift; 538 539 if (inst_eff_addr2 >= ld_eff_addr1 && inst_eff_addr1 <= ld_eff_addr2) { 540 if (inst->isLoad()) { 541 // If this load is to the same block as an external snoop 542 // invalidate that we've observed then the load needs to be 543 // squashed as it could have newer data 544 if (ld_inst->hitExternalSnoop()) { 545 if (!memDepViolator || 546 ld_inst->seqNum < memDepViolator->seqNum) { 547 DPRINTF(LSQUnit, "Detected fault with inst [sn:%lli] " 548 "and [sn:%lli] at address %#x\n", 549 inst->seqNum, ld_inst->seqNum, ld_eff_addr1); 550 memDepViolator = ld_inst; 551 552 ++lsqMemOrderViolation; 553 554 return std::make_shared<GenericISA::M5PanicFault>( 555 "Detected fault with inst [sn:%lli] and " 556 "[sn:%lli] at address %#x\n", 557 inst->seqNum, ld_inst->seqNum, ld_eff_addr1); 558 } 559 } 560 561 // Otherwise, mark the load has a possible load violation 562 // and if we see a snoop before it's commited, we need to squash 563 ld_inst->possibleLoadViolation(true); 564 DPRINTF(LSQUnit, "Found possible load violation at addr: %#x" 565 " between instructions [sn:%lli] and [sn:%lli]\n", 566 inst_eff_addr1, inst->seqNum, ld_inst->seqNum); 567 } else { 568 // A load/store incorrectly passed this store. 569 // Check if we already have a violator, or if it's newer 570 // squash and refetch. 571 if (memDepViolator && ld_inst->seqNum > memDepViolator->seqNum) 572 break; 573 574 DPRINTF(LSQUnit, "Detected fault with inst [sn:%lli] and " 575 "[sn:%lli] at address %#x\n", 576 inst->seqNum, ld_inst->seqNum, ld_eff_addr1); 577 memDepViolator = ld_inst; 578 579 ++lsqMemOrderViolation; 580 581 return std::make_shared<GenericISA::M5PanicFault>( 582 "Detected fault with " 583 "inst [sn:%lli] and [sn:%lli] at address %#x\n", 584 inst->seqNum, ld_inst->seqNum, ld_eff_addr1); 585 } 586 } 587 588 incrLdIdx(load_idx); 589 } 590 return NoFault; 591} 592 593 594 595 596template <class Impl> 597Fault 598LSQUnit<Impl>::executeLoad(const DynInstPtr &inst) 599{ 600 using namespace TheISA; 601 // Execute a specific load. 602 Fault load_fault = NoFault; 603 604 DPRINTF(LSQUnit, "Executing load PC %s, [sn:%lli]\n", 605 inst->pcState(), inst->seqNum); 606 607 assert(!inst->isSquashed()); 608 609 load_fault = inst->initiateAcc(); 610 611 if (inst->isTranslationDelayed() && 612 load_fault == NoFault) 613 return load_fault; 614 615 // If the instruction faulted or predicated false, then we need to send it 616 // along to commit without the instruction completing. 617 if (load_fault != NoFault || !inst->readPredicate()) { 618 // Send this instruction to commit, also make sure iew stage 619 // realizes there is activity. Mark it as executed unless it 620 // is a strictly ordered load that needs to hit the head of 621 // commit. 622 if (!inst->readPredicate()) 623 inst->forwardOldRegs(); 624 DPRINTF(LSQUnit, "Load [sn:%lli] not executed from %s\n", 625 inst->seqNum, 626 (load_fault != NoFault ? "fault" : "predication")); 627 if (!(inst->hasRequest() && inst->strictlyOrdered()) || 628 inst->isAtCommit()) { 629 inst->setExecuted(); 630 } 631 iewStage->instToCommit(inst); 632 iewStage->activityThisCycle(); 633 } else { 634 assert(inst->effAddrValid()); 635 int load_idx = inst->lqIdx; 636 incrLdIdx(load_idx); 637 638 if (checkLoads) 639 return checkViolations(load_idx, inst); 640 } 641 642 return load_fault; 643} 644 645template <class Impl> 646Fault 647LSQUnit<Impl>::executeStore(const DynInstPtr &store_inst) 648{ 649 using namespace TheISA; 650 // Make sure that a store exists. 651 assert(stores != 0); 652 653 int store_idx = store_inst->sqIdx; 654 655 DPRINTF(LSQUnit, "Executing store PC %s [sn:%lli]\n", 656 store_inst->pcState(), store_inst->seqNum); 657 658 assert(!store_inst->isSquashed()); 659 660 // Check the recently completed loads to see if any match this store's 661 // address. If so, then we have a memory ordering violation. 662 int load_idx = store_inst->lqIdx; 663 664 Fault store_fault = store_inst->initiateAcc(); 665 666 if (store_inst->isTranslationDelayed() && 667 store_fault == NoFault) 668 return store_fault; 669 670 if (!store_inst->readPredicate()) { 671 DPRINTF(LSQUnit, "Store [sn:%lli] not executed from predication\n", 672 store_inst->seqNum); 673 store_inst->forwardOldRegs(); 674 return store_fault; 675 } 676 677 if (storeQueue[store_idx].size == 0) { 678 DPRINTF(LSQUnit,"Fault on Store PC %s, [sn:%lli], Size = 0\n", 679 store_inst->pcState(), store_inst->seqNum); 680 681 return store_fault; 682 } 683 684 assert(store_fault == NoFault); 685 686 if (store_inst->isStoreConditional()) { 687 // Store conditionals need to set themselves as able to 688 // writeback if we haven't had a fault by here. 689 storeQueue[store_idx].canWB = true; 690 691 ++storesToWB; 692 } 693 694 return checkViolations(load_idx, store_inst); 695 696} 697 698template <class Impl> 699void 700LSQUnit<Impl>::commitLoad() 701{ 702 assert(loadQueue[loadHead]); 703 704 DPRINTF(LSQUnit, "Committing head load instruction, PC %s\n", 705 loadQueue[loadHead]->pcState()); 706 707 loadQueue[loadHead] = NULL; 708 709 incrLdIdx(loadHead); 710 711 --loads; 712} 713 714template <class Impl> 715void 716LSQUnit<Impl>::commitLoads(InstSeqNum &youngest_inst) 717{ 718 assert(loads == 0 || loadQueue[loadHead]); 719 720 while (loads != 0 && loadQueue[loadHead]->seqNum <= youngest_inst) { 721 commitLoad(); 722 } 723} 724 725template <class Impl> 726void 727LSQUnit<Impl>::commitStores(InstSeqNum &youngest_inst) 728{ 729 assert(stores == 0 || storeQueue[storeHead].inst); 730 731 int store_idx = storeHead; 732 733 while (store_idx != storeTail) { 734 assert(storeQueue[store_idx].inst); 735 // Mark any stores that are now committed and have not yet 736 // been marked as able to write back. 737 if (!storeQueue[store_idx].canWB) { 738 if (storeQueue[store_idx].inst->seqNum > youngest_inst) { 739 break; 740 } 741 DPRINTF(LSQUnit, "Marking store as able to write back, PC " 742 "%s [sn:%lli]\n", 743 storeQueue[store_idx].inst->pcState(), 744 storeQueue[store_idx].inst->seqNum); 745 746 storeQueue[store_idx].canWB = true; 747 748 ++storesToWB; 749 } 750 751 incrStIdx(store_idx); 752 } 753} 754 755template <class Impl> 756void 757LSQUnit<Impl>::writebackPendingStore() 758{ 759 if (hasPendingPkt) { 760 assert(pendingPkt != NULL); 761 762 // If the cache is blocked, this will store the packet for retry. 763 if (sendStore(pendingPkt)) { 764 storePostSend(pendingPkt); 765 } 766 pendingPkt = NULL; 767 hasPendingPkt = false; 768 } 769} 770 771template <class Impl> 772void 773LSQUnit<Impl>::writebackStores() 774{ 775 // First writeback the second packet from any split store that didn't 776 // complete last cycle because there weren't enough cache ports available. 777 if (TheISA::HasUnalignedMemAcc) { 778 writebackPendingStore(); 779 } 780 781 while (storesToWB > 0 && 782 storeWBIdx != storeTail && 783 storeQueue[storeWBIdx].inst && 784 storeQueue[storeWBIdx].canWB && 785 ((!needsTSO) || (!storeInFlight)) && 786 usedStorePorts < cacheStorePorts) { 787 788 if (isStoreBlocked) { 789 DPRINTF(LSQUnit, "Unable to write back any more stores, cache" 790 " is blocked!\n"); 791 break; 792 } 793 794 // Store didn't write any data so no need to write it back to 795 // memory. 796 if (storeQueue[storeWBIdx].size == 0) { 797 completeStore(storeWBIdx); 798 799 incrStIdx(storeWBIdx); 800 801 continue; 802 } 803 804 ++usedStorePorts; 805 806 if (storeQueue[storeWBIdx].inst->isDataPrefetch()) { 807 incrStIdx(storeWBIdx); 808 809 continue; 810 } 811 812 assert(storeQueue[storeWBIdx].req); 813 assert(!storeQueue[storeWBIdx].committed); 814 815 if (TheISA::HasUnalignedMemAcc && storeQueue[storeWBIdx].isSplit) { 816 assert(storeQueue[storeWBIdx].sreqLow); 817 assert(storeQueue[storeWBIdx].sreqHigh); 818 } 819 820 DynInstPtr inst = storeQueue[storeWBIdx].inst; 821 822 RequestPtr &req = storeQueue[storeWBIdx].req; 823 const RequestPtr &sreqLow = storeQueue[storeWBIdx].sreqLow; 824 const RequestPtr &sreqHigh = storeQueue[storeWBIdx].sreqHigh; 825 826 storeQueue[storeWBIdx].committed = true; 827 828 assert(!inst->memData); 829 inst->memData = new uint8_t[req->getSize()]; 830 831 if (storeQueue[storeWBIdx].isAllZeros) 832 memset(inst->memData, 0, req->getSize()); 833 else 834 memcpy(inst->memData, storeQueue[storeWBIdx].data, req->getSize()); 835 836 PacketPtr data_pkt; 837 PacketPtr snd_data_pkt = NULL; 838 839 LSQSenderState *state = new LSQSenderState; 840 state->isLoad = false; 841 state->idx = storeWBIdx; 842 state->inst = inst; 843 844 if (!TheISA::HasUnalignedMemAcc || !storeQueue[storeWBIdx].isSplit) { 845 846 // Build a single data packet if the store isn't split. 847 data_pkt = Packet::createWrite(req); 848 data_pkt->dataStatic(inst->memData); 849 data_pkt->senderState = state; 850 } else { 851 // Create two packets if the store is split in two. 852 data_pkt = Packet::createWrite(sreqLow); 853 snd_data_pkt = Packet::createWrite(sreqHigh); 854 855 data_pkt->dataStatic(inst->memData); 856 snd_data_pkt->dataStatic(inst->memData + sreqLow->getSize()); 857 858 data_pkt->senderState = state; 859 snd_data_pkt->senderState = state; 860 861 state->isSplit = true; 862 state->outstanding = 2; 863 864 // Can delete the main request now. 865 req = sreqLow; 866 } 867 868 DPRINTF(LSQUnit, "D-Cache: Writing back store idx:%i PC:%s " 869 "to Addr:%#x, data:%#x [sn:%lli]\n", 870 storeWBIdx, inst->pcState(), 871 req->getPaddr(), (int)*(inst->memData), 872 inst->seqNum); 873 874 // @todo: Remove this SC hack once the memory system handles it. 875 if (inst->isStoreConditional()) { 876 assert(!storeQueue[storeWBIdx].isSplit); 877 // Disable recording the result temporarily. Writing to 878 // misc regs normally updates the result, but this is not 879 // the desired behavior when handling store conditionals. 880 inst->recordResult(false); 881 bool success = TheISA::handleLockedWrite(inst.get(), req, cacheBlockMask); 882 inst->recordResult(true); 883 884 if (!success) { 885 // Instantly complete this store. 886 DPRINTF(LSQUnit, "Store conditional [sn:%lli] failed. " 887 "Instantly completing it.\n", 888 inst->seqNum); 889 WritebackEvent *wb = new WritebackEvent(inst, data_pkt, this); 890 cpu->schedule(wb, curTick() + 1); 891 completeStore(storeWBIdx); 892 incrStIdx(storeWBIdx); 893 continue; 894 } 895 } else { 896 // Non-store conditionals do not need a writeback. 897 state->noWB = true; 898 } 899 900 bool split = 901 TheISA::HasUnalignedMemAcc && storeQueue[storeWBIdx].isSplit; 902 903 ThreadContext *thread = cpu->tcBase(lsqID); 904 905 if (req->isMmappedIpr()) { 906 assert(!inst->isStoreConditional()); 907 TheISA::handleIprWrite(thread, data_pkt); 908 delete data_pkt; 909 if (split) { 910 assert(snd_data_pkt->req->isMmappedIpr()); 911 TheISA::handleIprWrite(thread, snd_data_pkt); 912 delete snd_data_pkt; 913 } 914 delete state; 915 completeStore(storeWBIdx); 916 incrStIdx(storeWBIdx); 917 } else if (!sendStore(data_pkt)) { 918 DPRINTF(IEW, "D-Cache became blocked when writing [sn:%lli], will" 919 "retry later\n", 920 inst->seqNum); 921 922 // Need to store the second packet, if split. 923 if (split) { 924 state->pktToSend = true; 925 state->pendingPacket = snd_data_pkt; 926 } 927 } else { 928 929 // If split, try to send the second packet too 930 if (split) { 931 assert(snd_data_pkt); 932 933 // Ensure there are enough ports to use. 934 if (usedStorePorts < cacheStorePorts) { 935 ++usedStorePorts; 936 if (sendStore(snd_data_pkt)) { 937 storePostSend(snd_data_pkt); 938 } else { 939 DPRINTF(IEW, "D-Cache became blocked when writing" 940 " [sn:%lli] second packet, will retry later\n", 941 inst->seqNum); 942 } 943 } else { 944 945 // Store the packet for when there's free ports. 946 assert(pendingPkt == NULL); 947 pendingPkt = snd_data_pkt; 948 hasPendingPkt = true; 949 } 950 } else { 951 952 // Not a split store. 953 storePostSend(data_pkt); 954 } 955 } 956 } 957 958 // Not sure this should set it to 0. 959 usedStorePorts = 0; 960 961 assert(stores >= 0 && storesToWB >= 0); 962} 963 964/*template <class Impl> 965void 966LSQUnit<Impl>::removeMSHR(InstSeqNum seqNum) 967{ 968 list<InstSeqNum>::iterator mshr_it = find(mshrSeqNums.begin(), 969 mshrSeqNums.end(), 970 seqNum); 971 972 if (mshr_it != mshrSeqNums.end()) { 973 mshrSeqNums.erase(mshr_it); 974 DPRINTF(LSQUnit, "Removing MSHR. count = %i\n",mshrSeqNums.size()); 975 } 976}*/ 977 978template <class Impl> 979void 980LSQUnit<Impl>::squash(const InstSeqNum &squashed_num) 981{ 982 DPRINTF(LSQUnit, "Squashing until [sn:%lli]!" 983 "(Loads:%i Stores:%i)\n", squashed_num, loads, stores); 984 985 int load_idx = loadTail; 986 decrLdIdx(load_idx); 987 988 while (loads != 0 && loadQueue[load_idx]->seqNum > squashed_num) { 989 DPRINTF(LSQUnit,"Load Instruction PC %s squashed, " 990 "[sn:%lli]\n", 991 loadQueue[load_idx]->pcState(), 992 loadQueue[load_idx]->seqNum); 993 994 if (isStalled() && load_idx == stallingLoadIdx) { 995 stalled = false; 996 stallingStoreIsn = 0; 997 stallingLoadIdx = 0; 998 } 999 1000 // Clear the smart pointer to make sure it is decremented. 1001 loadQueue[load_idx]->setSquashed(); 1002 loadQueue[load_idx] = NULL; 1003 --loads; 1004 1005 // Inefficient! 1006 loadTail = load_idx; 1007 1008 decrLdIdx(load_idx); 1009 ++lsqSquashedLoads; 1010 } 1011 1012 if (memDepViolator && squashed_num < memDepViolator->seqNum) { 1013 memDepViolator = NULL; 1014 } 1015 1016 int store_idx = storeTail; 1017 decrStIdx(store_idx); 1018 1019 while (stores != 0 && 1020 storeQueue[store_idx].inst->seqNum > squashed_num) { 1021 // Instructions marked as can WB are already committed. 1022 if (storeQueue[store_idx].canWB) { 1023 break; 1024 } 1025 1026 DPRINTF(LSQUnit,"Store Instruction PC %s squashed, " 1027 "idx:%i [sn:%lli]\n", 1028 storeQueue[store_idx].inst->pcState(), 1029 store_idx, storeQueue[store_idx].inst->seqNum); 1030 1031 // I don't think this can happen. It should have been cleared 1032 // by the stalling load. 1033 if (isStalled() && 1034 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) { 1035 panic("Is stalled should have been cleared by stalling load!\n"); 1036 stalled = false; 1037 stallingStoreIsn = 0; 1038 } 1039 1040 // Clear the smart pointer to make sure it is decremented. 1041 storeQueue[store_idx].inst->setSquashed(); 1042 storeQueue[store_idx].inst = NULL; 1043 storeQueue[store_idx].canWB = 0; 1044 1045 // Must delete request now that it wasn't handed off to 1046 // memory. This is quite ugly. @todo: Figure out the proper 1047 // place to really handle request deletes. 1048 storeQueue[store_idx].req.reset(); 1049 if (TheISA::HasUnalignedMemAcc && storeQueue[store_idx].isSplit) { 1050 storeQueue[store_idx].sreqLow.reset(); 1051 storeQueue[store_idx].sreqHigh.reset(); 1052 } 1053 1054 --stores; 1055 1056 // Inefficient! 1057 storeTail = store_idx; 1058 1059 decrStIdx(store_idx); 1060 ++lsqSquashedStores; 1061 } 1062} 1063 1064template <class Impl> 1065void 1066LSQUnit<Impl>::storePostSend(PacketPtr pkt) 1067{ 1068 if (isStalled() && 1069 storeQueue[storeWBIdx].inst->seqNum == stallingStoreIsn) { 1070 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] " 1071 "load idx:%i\n", 1072 stallingStoreIsn, stallingLoadIdx); 1073 stalled = false; 1074 stallingStoreIsn = 0; 1075 iewStage->replayMemInst(loadQueue[stallingLoadIdx]); 1076 } 1077 1078 if (!storeQueue[storeWBIdx].inst->isStoreConditional()) { 1079 // The store is basically completed at this time. This 1080 // only works so long as the checker doesn't try to 1081 // verify the value in memory for stores. 1082 storeQueue[storeWBIdx].inst->setCompleted(); 1083 1084 if (cpu->checker) { 1085 cpu->checker->verify(storeQueue[storeWBIdx].inst); 1086 } 1087 } 1088 1089 if (needsTSO) { 1090 storeInFlight = true; 1091 } 1092 1093 incrStIdx(storeWBIdx); 1094} 1095 1096template <class Impl> 1097void 1098LSQUnit<Impl>::writeback(const DynInstPtr &inst, PacketPtr pkt) 1099{ 1100 iewStage->wakeCPU(); 1101 1102 // Squashed instructions do not need to complete their access. 1103 if (inst->isSquashed()) { 1104 assert(!inst->isStore()); 1105 ++lsqIgnoredResponses; 1106 return; 1107 } 1108 1109 if (!inst->isExecuted()) { 1110 inst->setExecuted(); 1111 1112 if (inst->fault == NoFault) { 1113 // Complete access to copy data to proper place. 1114 inst->completeAcc(pkt); 1115 } else { 1116 // If the instruction has an outstanding fault, we cannot complete 1117 // the access as this discards the current fault. 1118 1119 // If we have an outstanding fault, the fault should only be of 1120 // type ReExec. 1121 assert(dynamic_cast<ReExec*>(inst->fault.get()) != nullptr); 1122 1123 DPRINTF(LSQUnit, "Not completing instruction [sn:%lli] access " 1124 "due to pending fault.\n", inst->seqNum); 1125 } 1126 } 1127 1128 // Need to insert instruction into queue to commit 1129 iewStage->instToCommit(inst); 1130 1131 iewStage->activityThisCycle(); 1132 1133 // see if this load changed the PC 1134 iewStage->checkMisprediction(inst); 1135} 1136 1137template <class Impl> 1138void 1139LSQUnit<Impl>::completeStore(int store_idx) 1140{ 1141 assert(storeQueue[store_idx].inst); 1142 storeQueue[store_idx].completed = true; 1143 --storesToWB; 1144 // A bit conservative because a store completion may not free up entries, 1145 // but hopefully avoids two store completions in one cycle from making 1146 // the CPU tick twice. 1147 cpu->wakeCPU(); 1148 cpu->activityThisCycle(); 1149 1150 if (store_idx == storeHead) { 1151 do { 1152 incrStIdx(storeHead); 1153 1154 --stores; 1155 } while (storeQueue[storeHead].completed && 1156 storeHead != storeTail); 1157 1158 iewStage->updateLSQNextCycle = true; 1159 } 1160 1161 DPRINTF(LSQUnit, "Completing store [sn:%lli], idx:%i, store head " 1162 "idx:%i\n", 1163 storeQueue[store_idx].inst->seqNum, store_idx, storeHead); 1164 1165#if TRACING_ON 1166 if (DTRACE(O3PipeView)) { 1167 storeQueue[store_idx].inst->storeTick = 1168 curTick() - storeQueue[store_idx].inst->fetchTick; 1169 } 1170#endif 1171 1172 if (isStalled() && 1173 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) { 1174 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] " 1175 "load idx:%i\n", 1176 stallingStoreIsn, stallingLoadIdx); 1177 stalled = false; 1178 stallingStoreIsn = 0; 1179 iewStage->replayMemInst(loadQueue[stallingLoadIdx]); 1180 } 1181 1182 storeQueue[store_idx].inst->setCompleted(); 1183 1184 if (needsTSO) { 1185 storeInFlight = false; 1186 } 1187 1188 // Tell the checker we've completed this instruction. Some stores 1189 // may get reported twice to the checker, but the checker can 1190 // handle that case. 1191 1192 // Store conditionals cannot be sent to the checker yet, they have 1193 // to update the misc registers first which should take place 1194 // when they commit 1195 if (cpu->checker && !storeQueue[store_idx].inst->isStoreConditional()) { 1196 cpu->checker->verify(storeQueue[store_idx].inst); 1197 } 1198} 1199 1200template <class Impl> 1201bool 1202LSQUnit<Impl>::sendStore(PacketPtr data_pkt) 1203{ 1204 if (!dcachePort->sendTimingReq(data_pkt)) { 1205 // Need to handle becoming blocked on a store. 1206 isStoreBlocked = true; 1207 ++lsqCacheBlocked; 1208 assert(retryPkt == NULL); 1209 retryPkt = data_pkt; 1210 return false; 1211 } 1212 return true; 1213} 1214 1215template <class Impl> 1216void 1217LSQUnit<Impl>::recvRetry() 1218{ 1219 if (isStoreBlocked) { 1220 DPRINTF(LSQUnit, "Receiving retry: store blocked\n"); 1221 assert(retryPkt != NULL); 1222 1223 LSQSenderState *state = 1224 dynamic_cast<LSQSenderState *>(retryPkt->senderState); 1225 1226 if (dcachePort->sendTimingReq(retryPkt)) { 1227 // Don't finish the store unless this is the last packet. 1228 if (!TheISA::HasUnalignedMemAcc || !state->pktToSend || 1229 state->pendingPacket == retryPkt) { 1230 state->pktToSend = false; 1231 storePostSend(retryPkt); 1232 } 1233 retryPkt = NULL; 1234 isStoreBlocked = false; 1235 1236 // Send any outstanding packet. 1237 if (TheISA::HasUnalignedMemAcc && state->pktToSend) { 1238 assert(state->pendingPacket); 1239 if (sendStore(state->pendingPacket)) { 1240 storePostSend(state->pendingPacket); 1241 } 1242 } 1243 } else { 1244 // Still blocked! 1245 ++lsqCacheBlocked; 1246 } 1247 } 1248} 1249 1250template <class Impl> 1251inline void 1252LSQUnit<Impl>::incrStIdx(int &store_idx) const 1253{ 1254 if (++store_idx >= SQEntries) 1255 store_idx = 0; 1256} 1257 1258template <class Impl> 1259inline void 1260LSQUnit<Impl>::decrStIdx(int &store_idx) const 1261{ 1262 if (--store_idx < 0) 1263 store_idx += SQEntries; 1264} 1265 1266template <class Impl> 1267inline void 1268LSQUnit<Impl>::incrLdIdx(int &load_idx) const 1269{ 1270 if (++load_idx >= LQEntries) 1271 load_idx = 0; 1272} 1273 1274template <class Impl> 1275inline void 1276LSQUnit<Impl>::decrLdIdx(int &load_idx) const 1277{ 1278 if (--load_idx < 0) 1279 load_idx += LQEntries; 1280} 1281 1282template <class Impl> 1283void 1284LSQUnit<Impl>::dumpInsts() const 1285{ 1286 cprintf("Load store queue: Dumping instructions.\n"); 1287 cprintf("Load queue size: %i\n", loads); 1288 cprintf("Load queue: "); 1289 1290 int load_idx = loadHead; 1291 1292 while (load_idx != loadTail && loadQueue[load_idx]) { 1293 const DynInstPtr &inst(loadQueue[load_idx]); 1294 cprintf("%s.[sn:%i] ", inst->pcState(), inst->seqNum); 1295 1296 incrLdIdx(load_idx); 1297 } 1298 cprintf("\n"); 1299 1300 cprintf("Store queue size: %i\n", stores); 1301 cprintf("Store queue: "); 1302 1303 int store_idx = storeHead; 1304 1305 while (store_idx != storeTail && storeQueue[store_idx].inst) { 1306 const DynInstPtr &inst(storeQueue[store_idx].inst); 1307 cprintf("%s.[sn:%i] ", inst->pcState(), inst->seqNum); 1308 1309 incrStIdx(store_idx); 1310 } 1311 1312 cprintf("\n"); 1313} 1314 1315#endif//__CPU_O3_LSQ_UNIT_IMPL_HH__ 1316