lsq_unit_impl.hh revision 8272:82057507f2f9
1/* 2 * Copyright (c) 2010 ARM Limited 3 * All rights reserved 4 * 5 * The license below extends only to copyright in the software and shall 6 * not be construed as granting a license to any other intellectual 7 * property including but not limited to intellectual property relating 8 * to a hardware implementation of the functionality of the software 9 * licensed hereunder. You may use the software subject to the license 10 * terms below provided that you ensure that this notice is replicated 11 * unmodified and in its entirety in all distributions of the software, 12 * modified or unmodified, in source code or in binary form. 13 * 14 * Copyright (c) 2004-2005 The Regents of The University of Michigan 15 * All rights reserved. 16 * 17 * Redistribution and use in source and binary forms, with or without 18 * modification, are permitted provided that the following conditions are 19 * met: redistributions of source code must retain the above copyright 20 * notice, this list of conditions and the following disclaimer; 21 * redistributions in binary form must reproduce the above copyright 22 * notice, this list of conditions and the following disclaimer in the 23 * documentation and/or other materials provided with the distribution; 24 * neither the name of the copyright holders nor the names of its 25 * contributors may be used to endorse or promote products derived from 26 * this software without specific prior written permission. 27 * 28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 39 * 40 * Authors: Kevin Lim 41 * Korey Sewell 42 */ 43 44#include "arch/locked_mem.hh" 45#include "base/str.hh" 46#include "config/the_isa.hh" 47#include "config/use_checker.hh" 48#include "cpu/o3/lsq.hh" 49#include "cpu/o3/lsq_unit.hh" 50#include "debug/Activity.hh" 51#include "debug/IEW.hh" 52#include "debug/LSQUnit.hh" 53#include "mem/packet.hh" 54#include "mem/request.hh" 55 56#if USE_CHECKER 57#include "cpu/checker/cpu.hh" 58#endif 59 60template<class Impl> 61LSQUnit<Impl>::WritebackEvent::WritebackEvent(DynInstPtr &_inst, PacketPtr _pkt, 62 LSQUnit *lsq_ptr) 63 : inst(_inst), pkt(_pkt), lsqPtr(lsq_ptr) 64{ 65 this->setFlags(Event::AutoDelete); 66} 67 68template<class Impl> 69void 70LSQUnit<Impl>::WritebackEvent::process() 71{ 72 if (!lsqPtr->isSwitchedOut()) { 73 lsqPtr->writeback(inst, pkt); 74 } 75 76 if (pkt->senderState) 77 delete pkt->senderState; 78 79 delete pkt->req; 80 delete pkt; 81} 82 83template<class Impl> 84const char * 85LSQUnit<Impl>::WritebackEvent::description() const 86{ 87 return "Store writeback"; 88} 89 90template<class Impl> 91void 92LSQUnit<Impl>::completeDataAccess(PacketPtr pkt) 93{ 94 LSQSenderState *state = dynamic_cast<LSQSenderState *>(pkt->senderState); 95 DynInstPtr inst = state->inst; 96 DPRINTF(IEW, "Writeback event [sn:%lli].\n", inst->seqNum); 97 DPRINTF(Activity, "Activity: Writeback event [sn:%lli].\n", inst->seqNum); 98 99 //iewStage->ldstQueue.removeMSHR(inst->threadNumber,inst->seqNum); 100 101 assert(!pkt->wasNacked()); 102 103 // If this is a split access, wait until all packets are received. 104 if (TheISA::HasUnalignedMemAcc && !state->complete()) { 105 delete pkt->req; 106 delete pkt; 107 return; 108 } 109 110 if (isSwitchedOut() || inst->isSquashed()) { 111 iewStage->decrWb(inst->seqNum); 112 } else { 113 if (!state->noWB) { 114 if (!TheISA::HasUnalignedMemAcc || !state->isSplit || 115 !state->isLoad) { 116 writeback(inst, pkt); 117 } else { 118 writeback(inst, state->mainPkt); 119 } 120 } 121 122 if (inst->isStore()) { 123 completeStore(state->idx); 124 } 125 } 126 127 if (TheISA::HasUnalignedMemAcc && state->isSplit && state->isLoad) { 128 delete state->mainPkt->req; 129 delete state->mainPkt; 130 } 131 delete state; 132 delete pkt->req; 133 delete pkt; 134} 135 136template <class Impl> 137LSQUnit<Impl>::LSQUnit() 138 : loads(0), stores(0), storesToWB(0), stalled(false), 139 isStoreBlocked(false), isLoadBlocked(false), 140 loadBlockedHandled(false), hasPendingPkt(false) 141{ 142} 143 144template<class Impl> 145void 146LSQUnit<Impl>::init(O3CPU *cpu_ptr, IEW *iew_ptr, DerivO3CPUParams *params, 147 LSQ *lsq_ptr, unsigned maxLQEntries, unsigned maxSQEntries, 148 unsigned id) 149{ 150 cpu = cpu_ptr; 151 iewStage = iew_ptr; 152 153 DPRINTF(LSQUnit, "Creating LSQUnit%i object.\n",id); 154 155 switchedOut = false; 156 157 lsq = lsq_ptr; 158 159 lsqID = id; 160 161 // Add 1 for the sentinel entry (they are circular queues). 162 LQEntries = maxLQEntries + 1; 163 SQEntries = maxSQEntries + 1; 164 165 loadQueue.resize(LQEntries); 166 storeQueue.resize(SQEntries); 167 168 depCheckShift = params->LSQDepCheckShift; 169 checkLoads = params->LSQCheckLoads; 170 171 loadHead = loadTail = 0; 172 173 storeHead = storeWBIdx = storeTail = 0; 174 175 usedPorts = 0; 176 cachePorts = params->cachePorts; 177 178 retryPkt = NULL; 179 memDepViolator = NULL; 180 181 blockedLoadSeqNum = 0; 182} 183 184template<class Impl> 185std::string 186LSQUnit<Impl>::name() const 187{ 188 if (Impl::MaxThreads == 1) { 189 return iewStage->name() + ".lsq"; 190 } else { 191 return iewStage->name() + ".lsq.thread" + to_string(lsqID); 192 } 193} 194 195template<class Impl> 196void 197LSQUnit<Impl>::regStats() 198{ 199 lsqForwLoads 200 .name(name() + ".forwLoads") 201 .desc("Number of loads that had data forwarded from stores"); 202 203 invAddrLoads 204 .name(name() + ".invAddrLoads") 205 .desc("Number of loads ignored due to an invalid address"); 206 207 lsqSquashedLoads 208 .name(name() + ".squashedLoads") 209 .desc("Number of loads squashed"); 210 211 lsqIgnoredResponses 212 .name(name() + ".ignoredResponses") 213 .desc("Number of memory responses ignored because the instruction is squashed"); 214 215 lsqMemOrderViolation 216 .name(name() + ".memOrderViolation") 217 .desc("Number of memory ordering violations"); 218 219 lsqSquashedStores 220 .name(name() + ".squashedStores") 221 .desc("Number of stores squashed"); 222 223 invAddrSwpfs 224 .name(name() + ".invAddrSwpfs") 225 .desc("Number of software prefetches ignored due to an invalid address"); 226 227 lsqBlockedLoads 228 .name(name() + ".blockedLoads") 229 .desc("Number of blocked loads due to partial load-store forwarding"); 230 231 lsqRescheduledLoads 232 .name(name() + ".rescheduledLoads") 233 .desc("Number of loads that were rescheduled"); 234 235 lsqCacheBlocked 236 .name(name() + ".cacheBlocked") 237 .desc("Number of times an access to memory failed due to the cache being blocked"); 238} 239 240template<class Impl> 241void 242LSQUnit<Impl>::setDcachePort(Port *dcache_port) 243{ 244 dcachePort = dcache_port; 245 246#if USE_CHECKER 247 if (cpu->checker) { 248 cpu->checker->setDcachePort(dcachePort); 249 } 250#endif 251} 252 253template<class Impl> 254void 255LSQUnit<Impl>::clearLQ() 256{ 257 loadQueue.clear(); 258} 259 260template<class Impl> 261void 262LSQUnit<Impl>::clearSQ() 263{ 264 storeQueue.clear(); 265} 266 267template<class Impl> 268void 269LSQUnit<Impl>::switchOut() 270{ 271 switchedOut = true; 272 for (int i = 0; i < loadQueue.size(); ++i) { 273 assert(!loadQueue[i]); 274 loadQueue[i] = NULL; 275 } 276 277 assert(storesToWB == 0); 278} 279 280template<class Impl> 281void 282LSQUnit<Impl>::takeOverFrom() 283{ 284 switchedOut = false; 285 loads = stores = storesToWB = 0; 286 287 loadHead = loadTail = 0; 288 289 storeHead = storeWBIdx = storeTail = 0; 290 291 usedPorts = 0; 292 293 memDepViolator = NULL; 294 295 blockedLoadSeqNum = 0; 296 297 stalled = false; 298 isLoadBlocked = false; 299 loadBlockedHandled = false; 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} 320 321template<class Impl> 322void 323LSQUnit<Impl>::resizeSQ(unsigned size) 324{ 325 unsigned size_plus_sentinel = size + 1; 326 if (size_plus_sentinel > SQEntries) { 327 while (size_plus_sentinel > storeQueue.size()) { 328 SQEntry dummy; 329 storeQueue.push_back(dummy); 330 SQEntries++; 331 } 332 } else { 333 SQEntries = size_plus_sentinel; 334 } 335} 336 337template <class Impl> 338void 339LSQUnit<Impl>::insert(DynInstPtr &inst) 340{ 341 assert(inst->isMemRef()); 342 343 assert(inst->isLoad() || inst->isStore()); 344 345 if (inst->isLoad()) { 346 insertLoad(inst); 347 } else { 348 insertStore(inst); 349 } 350 351 inst->setInLSQ(); 352} 353 354template <class Impl> 355void 356LSQUnit<Impl>::insertLoad(DynInstPtr &load_inst) 357{ 358 assert((loadTail + 1) % LQEntries != loadHead); 359 assert(loads < LQEntries); 360 361 DPRINTF(LSQUnit, "Inserting load PC %s, idx:%i [sn:%lli]\n", 362 load_inst->pcState(), loadTail, load_inst->seqNum); 363 364 load_inst->lqIdx = loadTail; 365 366 if (stores == 0) { 367 load_inst->sqIdx = -1; 368 } else { 369 load_inst->sqIdx = storeTail; 370 } 371 372 loadQueue[loadTail] = load_inst; 373 374 incrLdIdx(loadTail); 375 376 ++loads; 377} 378 379template <class Impl> 380void 381LSQUnit<Impl>::insertStore(DynInstPtr &store_inst) 382{ 383 // Make sure it is not full before inserting an instruction. 384 assert((storeTail + 1) % SQEntries != storeHead); 385 assert(stores < SQEntries); 386 387 DPRINTF(LSQUnit, "Inserting store PC %s, idx:%i [sn:%lli]\n", 388 store_inst->pcState(), storeTail, store_inst->seqNum); 389 390 store_inst->sqIdx = storeTail; 391 store_inst->lqIdx = loadTail; 392 393 storeQueue[storeTail] = SQEntry(store_inst); 394 395 incrStIdx(storeTail); 396 397 ++stores; 398} 399 400template <class Impl> 401typename Impl::DynInstPtr 402LSQUnit<Impl>::getMemDepViolator() 403{ 404 DynInstPtr temp = memDepViolator; 405 406 memDepViolator = NULL; 407 408 return temp; 409} 410 411template <class Impl> 412unsigned 413LSQUnit<Impl>::numFreeEntries() 414{ 415 unsigned free_lq_entries = LQEntries - loads; 416 unsigned free_sq_entries = SQEntries - stores; 417 418 // Both the LQ and SQ entries have an extra dummy entry to differentiate 419 // empty/full conditions. Subtract 1 from the free entries. 420 if (free_lq_entries < free_sq_entries) { 421 return free_lq_entries - 1; 422 } else { 423 return free_sq_entries - 1; 424 } 425} 426 427template <class Impl> 428int 429LSQUnit<Impl>::numLoadsReady() 430{ 431 int load_idx = loadHead; 432 int retval = 0; 433 434 while (load_idx != loadTail) { 435 assert(loadQueue[load_idx]); 436 437 if (loadQueue[load_idx]->readyToIssue()) { 438 ++retval; 439 } 440 } 441 442 return retval; 443} 444 445template <class Impl> 446Fault 447LSQUnit<Impl>::checkViolations(int load_idx, DynInstPtr &inst) 448{ 449 Addr inst_eff_addr1 = inst->effAddr >> depCheckShift; 450 Addr inst_eff_addr2 = (inst->effAddr + inst->effSize - 1) >> depCheckShift; 451 452 /** @todo in theory you only need to check an instruction that has executed 453 * however, there isn't a good way in the pipeline at the moment to check 454 * all instructions that will execute before the store writes back. Thus, 455 * like the implementation that came before it, we're overly conservative. 456 */ 457 while (load_idx != loadTail) { 458 DynInstPtr ld_inst = loadQueue[load_idx]; 459 if (!ld_inst->effAddrValid || ld_inst->uncacheable()) { 460 incrLdIdx(load_idx); 461 continue; 462 } 463 464 Addr ld_eff_addr1 = ld_inst->effAddr >> depCheckShift; 465 Addr ld_eff_addr2 = 466 (ld_inst->effAddr + ld_inst->effSize - 1) >> depCheckShift; 467 468 if (inst_eff_addr2 >= ld_eff_addr1 && inst_eff_addr1 <= ld_eff_addr2) { 469 // A load/store incorrectly passed this load/store. 470 // Check if we already have a violator, or if it's newer 471 // squash and refetch. 472 if (memDepViolator && ld_inst->seqNum > memDepViolator->seqNum) 473 break; 474 475 DPRINTF(LSQUnit, "Detected fault with inst [sn:%lli] and [sn:%lli]" 476 " at address %#x\n", inst->seqNum, ld_inst->seqNum, 477 ld_eff_addr1); 478 memDepViolator = ld_inst; 479 480 ++lsqMemOrderViolation; 481 482 return TheISA::genMachineCheckFault(); 483 } 484 485 incrLdIdx(load_idx); 486 } 487 return NoFault; 488} 489 490 491 492 493template <class Impl> 494Fault 495LSQUnit<Impl>::executeLoad(DynInstPtr &inst) 496{ 497 using namespace TheISA; 498 // Execute a specific load. 499 Fault load_fault = NoFault; 500 501 DPRINTF(LSQUnit, "Executing load PC %s, [sn:%lli]\n", 502 inst->pcState(), inst->seqNum); 503 504 assert(!inst->isSquashed()); 505 506 load_fault = inst->initiateAcc(); 507 508 if (inst->isTranslationDelayed() && 509 load_fault == NoFault) 510 return load_fault; 511 512 // If the instruction faulted or predicated false, then we need to send it 513 // along to commit without the instruction completing. 514 if (load_fault != NoFault || inst->readPredicate() == false) { 515 // Send this instruction to commit, also make sure iew stage 516 // realizes there is activity. 517 // Mark it as executed unless it is an uncached load that 518 // needs to hit the head of commit. 519 if (inst->readPredicate() == false) 520 inst->forwardOldRegs(); 521 DPRINTF(LSQUnit, "Load [sn:%lli] not executed from %s\n", 522 inst->seqNum, 523 (load_fault != NoFault ? "fault" : "predication")); 524 if (!(inst->hasRequest() && inst->uncacheable()) || 525 inst->isAtCommit()) { 526 inst->setExecuted(); 527 } 528 iewStage->instToCommit(inst); 529 iewStage->activityThisCycle(); 530 } else if (!loadBlocked()) { 531 assert(inst->effAddrValid); 532 int load_idx = inst->lqIdx; 533 incrLdIdx(load_idx); 534 535 if (checkLoads) 536 return checkViolations(load_idx, inst); 537 } 538 539 return load_fault; 540} 541 542template <class Impl> 543Fault 544LSQUnit<Impl>::executeStore(DynInstPtr &store_inst) 545{ 546 using namespace TheISA; 547 // Make sure that a store exists. 548 assert(stores != 0); 549 550 int store_idx = store_inst->sqIdx; 551 552 DPRINTF(LSQUnit, "Executing store PC %s [sn:%lli]\n", 553 store_inst->pcState(), store_inst->seqNum); 554 555 assert(!store_inst->isSquashed()); 556 557 // Check the recently completed loads to see if any match this store's 558 // address. If so, then we have a memory ordering violation. 559 int load_idx = store_inst->lqIdx; 560 561 Fault store_fault = store_inst->initiateAcc(); 562 563 if (store_inst->isTranslationDelayed() && 564 store_fault == NoFault) 565 return store_fault; 566 567 if (store_inst->readPredicate() == false) 568 store_inst->forwardOldRegs(); 569 570 if (storeQueue[store_idx].size == 0) { 571 DPRINTF(LSQUnit,"Fault on Store PC %s, [sn:%lli], Size = 0\n", 572 store_inst->pcState(), store_inst->seqNum); 573 574 return store_fault; 575 } else if (store_inst->readPredicate() == false) { 576 DPRINTF(LSQUnit, "Store [sn:%lli] not executed from predication\n", 577 store_inst->seqNum); 578 return store_fault; 579 } 580 581 assert(store_fault == NoFault); 582 583 if (store_inst->isStoreConditional()) { 584 // Store conditionals need to set themselves as able to 585 // writeback if we haven't had a fault by here. 586 storeQueue[store_idx].canWB = true; 587 588 ++storesToWB; 589 } 590 591 return checkViolations(load_idx, store_inst); 592 593} 594 595template <class Impl> 596void 597LSQUnit<Impl>::commitLoad() 598{ 599 assert(loadQueue[loadHead]); 600 601 DPRINTF(LSQUnit, "Committing head load instruction, PC %s\n", 602 loadQueue[loadHead]->pcState()); 603 604 loadQueue[loadHead] = NULL; 605 606 incrLdIdx(loadHead); 607 608 --loads; 609} 610 611template <class Impl> 612void 613LSQUnit<Impl>::commitLoads(InstSeqNum &youngest_inst) 614{ 615 assert(loads == 0 || loadQueue[loadHead]); 616 617 while (loads != 0 && loadQueue[loadHead]->seqNum <= youngest_inst) { 618 commitLoad(); 619 } 620} 621 622template <class Impl> 623void 624LSQUnit<Impl>::commitStores(InstSeqNum &youngest_inst) 625{ 626 assert(stores == 0 || storeQueue[storeHead].inst); 627 628 int store_idx = storeHead; 629 630 while (store_idx != storeTail) { 631 assert(storeQueue[store_idx].inst); 632 // Mark any stores that are now committed and have not yet 633 // been marked as able to write back. 634 if (!storeQueue[store_idx].canWB) { 635 if (storeQueue[store_idx].inst->seqNum > youngest_inst) { 636 break; 637 } 638 DPRINTF(LSQUnit, "Marking store as able to write back, PC " 639 "%s [sn:%lli]\n", 640 storeQueue[store_idx].inst->pcState(), 641 storeQueue[store_idx].inst->seqNum); 642 643 storeQueue[store_idx].canWB = true; 644 645 ++storesToWB; 646 } 647 648 incrStIdx(store_idx); 649 } 650} 651 652template <class Impl> 653void 654LSQUnit<Impl>::writebackPendingStore() 655{ 656 if (hasPendingPkt) { 657 assert(pendingPkt != NULL); 658 659 // If the cache is blocked, this will store the packet for retry. 660 if (sendStore(pendingPkt)) { 661 storePostSend(pendingPkt); 662 } 663 pendingPkt = NULL; 664 hasPendingPkt = false; 665 } 666} 667 668template <class Impl> 669void 670LSQUnit<Impl>::writebackStores() 671{ 672 // First writeback the second packet from any split store that didn't 673 // complete last cycle because there weren't enough cache ports available. 674 if (TheISA::HasUnalignedMemAcc) { 675 writebackPendingStore(); 676 } 677 678 while (storesToWB > 0 && 679 storeWBIdx != storeTail && 680 storeQueue[storeWBIdx].inst && 681 storeQueue[storeWBIdx].canWB && 682 usedPorts < cachePorts) { 683 684 if (isStoreBlocked || lsq->cacheBlocked()) { 685 DPRINTF(LSQUnit, "Unable to write back any more stores, cache" 686 " is blocked!\n"); 687 break; 688 } 689 690 // Store didn't write any data so no need to write it back to 691 // memory. 692 if (storeQueue[storeWBIdx].size == 0) { 693 completeStore(storeWBIdx); 694 695 incrStIdx(storeWBIdx); 696 697 continue; 698 } 699 700 ++usedPorts; 701 702 if (storeQueue[storeWBIdx].inst->isDataPrefetch()) { 703 incrStIdx(storeWBIdx); 704 705 continue; 706 } 707 708 assert(storeQueue[storeWBIdx].req); 709 assert(!storeQueue[storeWBIdx].committed); 710 711 if (TheISA::HasUnalignedMemAcc && storeQueue[storeWBIdx].isSplit) { 712 assert(storeQueue[storeWBIdx].sreqLow); 713 assert(storeQueue[storeWBIdx].sreqHigh); 714 } 715 716 DynInstPtr inst = storeQueue[storeWBIdx].inst; 717 718 Request *req = storeQueue[storeWBIdx].req; 719 storeQueue[storeWBIdx].committed = true; 720 721 assert(!inst->memData); 722 inst->memData = new uint8_t[64]; 723 724 memcpy(inst->memData, storeQueue[storeWBIdx].data, req->getSize()); 725 726 MemCmd command = 727 req->isSwap() ? MemCmd::SwapReq : 728 (req->isLLSC() ? MemCmd::StoreCondReq : MemCmd::WriteReq); 729 PacketPtr data_pkt; 730 PacketPtr snd_data_pkt = NULL; 731 732 LSQSenderState *state = new LSQSenderState; 733 state->isLoad = false; 734 state->idx = storeWBIdx; 735 state->inst = inst; 736 737 if (!TheISA::HasUnalignedMemAcc || !storeQueue[storeWBIdx].isSplit) { 738 739 // Build a single data packet if the store isn't split. 740 data_pkt = new Packet(req, command, Packet::Broadcast); 741 data_pkt->dataStatic(inst->memData); 742 data_pkt->senderState = state; 743 } else { 744 RequestPtr sreqLow = storeQueue[storeWBIdx].sreqLow; 745 RequestPtr sreqHigh = storeQueue[storeWBIdx].sreqHigh; 746 747 // Create two packets if the store is split in two. 748 data_pkt = new Packet(sreqLow, command, Packet::Broadcast); 749 snd_data_pkt = new Packet(sreqHigh, command, Packet::Broadcast); 750 751 data_pkt->dataStatic(inst->memData); 752 snd_data_pkt->dataStatic(inst->memData + sreqLow->getSize()); 753 754 data_pkt->senderState = state; 755 snd_data_pkt->senderState = state; 756 757 state->isSplit = true; 758 state->outstanding = 2; 759 760 // Can delete the main request now. 761 delete req; 762 req = sreqLow; 763 } 764 765 DPRINTF(LSQUnit, "D-Cache: Writing back store idx:%i PC:%s " 766 "to Addr:%#x, data:%#x [sn:%lli]\n", 767 storeWBIdx, inst->pcState(), 768 req->getPaddr(), (int)*(inst->memData), 769 inst->seqNum); 770 771 // @todo: Remove this SC hack once the memory system handles it. 772 if (inst->isStoreConditional()) { 773 assert(!storeQueue[storeWBIdx].isSplit); 774 // Disable recording the result temporarily. Writing to 775 // misc regs normally updates the result, but this is not 776 // the desired behavior when handling store conditionals. 777 inst->recordResult = false; 778 bool success = TheISA::handleLockedWrite(inst.get(), req); 779 inst->recordResult = true; 780 781 if (!success) { 782 // Instantly complete this store. 783 DPRINTF(LSQUnit, "Store conditional [sn:%lli] failed. " 784 "Instantly completing it.\n", 785 inst->seqNum); 786 WritebackEvent *wb = new WritebackEvent(inst, data_pkt, this); 787 cpu->schedule(wb, curTick() + 1); 788 completeStore(storeWBIdx); 789 incrStIdx(storeWBIdx); 790 continue; 791 } 792 } else { 793 // Non-store conditionals do not need a writeback. 794 state->noWB = true; 795 } 796 797 if (!sendStore(data_pkt)) { 798 DPRINTF(IEW, "D-Cache became blocked when writing [sn:%lli], will" 799 "retry later\n", 800 inst->seqNum); 801 802 // Need to store the second packet, if split. 803 if (TheISA::HasUnalignedMemAcc && storeQueue[storeWBIdx].isSplit) { 804 state->pktToSend = true; 805 state->pendingPacket = snd_data_pkt; 806 } 807 } else { 808 809 // If split, try to send the second packet too 810 if (TheISA::HasUnalignedMemAcc && storeQueue[storeWBIdx].isSplit) { 811 assert(snd_data_pkt); 812 813 // Ensure there are enough ports to use. 814 if (usedPorts < cachePorts) { 815 ++usedPorts; 816 if (sendStore(snd_data_pkt)) { 817 storePostSend(snd_data_pkt); 818 } else { 819 DPRINTF(IEW, "D-Cache became blocked when writing" 820 " [sn:%lli] second packet, will retry later\n", 821 inst->seqNum); 822 } 823 } else { 824 825 // Store the packet for when there's free ports. 826 assert(pendingPkt == NULL); 827 pendingPkt = snd_data_pkt; 828 hasPendingPkt = true; 829 } 830 } else { 831 832 // Not a split store. 833 storePostSend(data_pkt); 834 } 835 } 836 } 837 838 // Not sure this should set it to 0. 839 usedPorts = 0; 840 841 assert(stores >= 0 && storesToWB >= 0); 842} 843 844/*template <class Impl> 845void 846LSQUnit<Impl>::removeMSHR(InstSeqNum seqNum) 847{ 848 list<InstSeqNum>::iterator mshr_it = find(mshrSeqNums.begin(), 849 mshrSeqNums.end(), 850 seqNum); 851 852 if (mshr_it != mshrSeqNums.end()) { 853 mshrSeqNums.erase(mshr_it); 854 DPRINTF(LSQUnit, "Removing MSHR. count = %i\n",mshrSeqNums.size()); 855 } 856}*/ 857 858template <class Impl> 859void 860LSQUnit<Impl>::squash(const InstSeqNum &squashed_num) 861{ 862 DPRINTF(LSQUnit, "Squashing until [sn:%lli]!" 863 "(Loads:%i Stores:%i)\n", squashed_num, loads, stores); 864 865 int load_idx = loadTail; 866 decrLdIdx(load_idx); 867 868 while (loads != 0 && loadQueue[load_idx]->seqNum > squashed_num) { 869 DPRINTF(LSQUnit,"Load Instruction PC %s squashed, " 870 "[sn:%lli]\n", 871 loadQueue[load_idx]->pcState(), 872 loadQueue[load_idx]->seqNum); 873 874 if (isStalled() && load_idx == stallingLoadIdx) { 875 stalled = false; 876 stallingStoreIsn = 0; 877 stallingLoadIdx = 0; 878 } 879 880 // Clear the smart pointer to make sure it is decremented. 881 loadQueue[load_idx]->setSquashed(); 882 loadQueue[load_idx] = NULL; 883 --loads; 884 885 // Inefficient! 886 loadTail = load_idx; 887 888 decrLdIdx(load_idx); 889 ++lsqSquashedLoads; 890 } 891 892 if (isLoadBlocked) { 893 if (squashed_num < blockedLoadSeqNum) { 894 isLoadBlocked = false; 895 loadBlockedHandled = false; 896 blockedLoadSeqNum = 0; 897 } 898 } 899 900 if (memDepViolator && squashed_num < memDepViolator->seqNum) { 901 memDepViolator = NULL; 902 } 903 904 int store_idx = storeTail; 905 decrStIdx(store_idx); 906 907 while (stores != 0 && 908 storeQueue[store_idx].inst->seqNum > squashed_num) { 909 // Instructions marked as can WB are already committed. 910 if (storeQueue[store_idx].canWB) { 911 break; 912 } 913 914 DPRINTF(LSQUnit,"Store Instruction PC %s squashed, " 915 "idx:%i [sn:%lli]\n", 916 storeQueue[store_idx].inst->pcState(), 917 store_idx, storeQueue[store_idx].inst->seqNum); 918 919 // I don't think this can happen. It should have been cleared 920 // by the stalling load. 921 if (isStalled() && 922 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) { 923 panic("Is stalled should have been cleared by stalling load!\n"); 924 stalled = false; 925 stallingStoreIsn = 0; 926 } 927 928 // Clear the smart pointer to make sure it is decremented. 929 storeQueue[store_idx].inst->setSquashed(); 930 storeQueue[store_idx].inst = NULL; 931 storeQueue[store_idx].canWB = 0; 932 933 // Must delete request now that it wasn't handed off to 934 // memory. This is quite ugly. @todo: Figure out the proper 935 // place to really handle request deletes. 936 delete storeQueue[store_idx].req; 937 if (TheISA::HasUnalignedMemAcc && storeQueue[store_idx].isSplit) { 938 delete storeQueue[store_idx].sreqLow; 939 delete storeQueue[store_idx].sreqHigh; 940 941 storeQueue[store_idx].sreqLow = NULL; 942 storeQueue[store_idx].sreqHigh = NULL; 943 } 944 945 storeQueue[store_idx].req = NULL; 946 --stores; 947 948 // Inefficient! 949 storeTail = store_idx; 950 951 decrStIdx(store_idx); 952 ++lsqSquashedStores; 953 } 954} 955 956template <class Impl> 957void 958LSQUnit<Impl>::storePostSend(PacketPtr pkt) 959{ 960 if (isStalled() && 961 storeQueue[storeWBIdx].inst->seqNum == stallingStoreIsn) { 962 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] " 963 "load idx:%i\n", 964 stallingStoreIsn, stallingLoadIdx); 965 stalled = false; 966 stallingStoreIsn = 0; 967 iewStage->replayMemInst(loadQueue[stallingLoadIdx]); 968 } 969 970 if (!storeQueue[storeWBIdx].inst->isStoreConditional()) { 971 // The store is basically completed at this time. This 972 // only works so long as the checker doesn't try to 973 // verify the value in memory for stores. 974 storeQueue[storeWBIdx].inst->setCompleted(); 975#if USE_CHECKER 976 if (cpu->checker) { 977 cpu->checker->verify(storeQueue[storeWBIdx].inst); 978 } 979#endif 980 } 981 982 incrStIdx(storeWBIdx); 983} 984 985template <class Impl> 986void 987LSQUnit<Impl>::writeback(DynInstPtr &inst, PacketPtr pkt) 988{ 989 iewStage->wakeCPU(); 990 991 // Squashed instructions do not need to complete their access. 992 if (inst->isSquashed()) { 993 iewStage->decrWb(inst->seqNum); 994 assert(!inst->isStore()); 995 ++lsqIgnoredResponses; 996 return; 997 } 998 999 if (!inst->isExecuted()) { 1000 inst->setExecuted(); 1001 1002 // Complete access to copy data to proper place. 1003 inst->completeAcc(pkt); 1004 } 1005 1006 // Need to insert instruction into queue to commit 1007 iewStage->instToCommit(inst); 1008 1009 iewStage->activityThisCycle(); 1010 1011 // see if this load changed the PC 1012 iewStage->checkMisprediction(inst); 1013} 1014 1015template <class Impl> 1016void 1017LSQUnit<Impl>::completeStore(int store_idx) 1018{ 1019 assert(storeQueue[store_idx].inst); 1020 storeQueue[store_idx].completed = true; 1021 --storesToWB; 1022 // A bit conservative because a store completion may not free up entries, 1023 // but hopefully avoids two store completions in one cycle from making 1024 // the CPU tick twice. 1025 cpu->wakeCPU(); 1026 cpu->activityThisCycle(); 1027 1028 if (store_idx == storeHead) { 1029 do { 1030 incrStIdx(storeHead); 1031 1032 --stores; 1033 } while (storeQueue[storeHead].completed && 1034 storeHead != storeTail); 1035 1036 iewStage->updateLSQNextCycle = true; 1037 } 1038 1039 DPRINTF(LSQUnit, "Completing store [sn:%lli], idx:%i, store head " 1040 "idx:%i\n", 1041 storeQueue[store_idx].inst->seqNum, store_idx, storeHead); 1042 1043 if (isStalled() && 1044 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) { 1045 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] " 1046 "load idx:%i\n", 1047 stallingStoreIsn, stallingLoadIdx); 1048 stalled = false; 1049 stallingStoreIsn = 0; 1050 iewStage->replayMemInst(loadQueue[stallingLoadIdx]); 1051 } 1052 1053 storeQueue[store_idx].inst->setCompleted(); 1054 1055 // Tell the checker we've completed this instruction. Some stores 1056 // may get reported twice to the checker, but the checker can 1057 // handle that case. 1058#if USE_CHECKER 1059 if (cpu->checker) { 1060 cpu->checker->verify(storeQueue[store_idx].inst); 1061 } 1062#endif 1063} 1064 1065template <class Impl> 1066bool 1067LSQUnit<Impl>::sendStore(PacketPtr data_pkt) 1068{ 1069 if (!dcachePort->sendTiming(data_pkt)) { 1070 // Need to handle becoming blocked on a store. 1071 isStoreBlocked = true; 1072 ++lsqCacheBlocked; 1073 assert(retryPkt == NULL); 1074 retryPkt = data_pkt; 1075 lsq->setRetryTid(lsqID); 1076 return false; 1077 } 1078 return true; 1079} 1080 1081template <class Impl> 1082void 1083LSQUnit<Impl>::recvRetry() 1084{ 1085 if (isStoreBlocked) { 1086 DPRINTF(LSQUnit, "Receiving retry: store blocked\n"); 1087 assert(retryPkt != NULL); 1088 1089 if (dcachePort->sendTiming(retryPkt)) { 1090 LSQSenderState *state = 1091 dynamic_cast<LSQSenderState *>(retryPkt->senderState); 1092 1093 // Don't finish the store unless this is the last packet. 1094 if (!TheISA::HasUnalignedMemAcc || !state->pktToSend || 1095 state->pendingPacket == retryPkt) { 1096 state->pktToSend = false; 1097 storePostSend(retryPkt); 1098 } 1099 retryPkt = NULL; 1100 isStoreBlocked = false; 1101 lsq->setRetryTid(InvalidThreadID); 1102 1103 // Send any outstanding packet. 1104 if (TheISA::HasUnalignedMemAcc && state->pktToSend) { 1105 assert(state->pendingPacket); 1106 if (sendStore(state->pendingPacket)) { 1107 storePostSend(state->pendingPacket); 1108 } 1109 } 1110 } else { 1111 // Still blocked! 1112 ++lsqCacheBlocked; 1113 lsq->setRetryTid(lsqID); 1114 } 1115 } else if (isLoadBlocked) { 1116 DPRINTF(LSQUnit, "Loads squash themselves and all younger insts, " 1117 "no need to resend packet.\n"); 1118 } else { 1119 DPRINTF(LSQUnit, "Retry received but LSQ is no longer blocked.\n"); 1120 } 1121} 1122 1123template <class Impl> 1124inline void 1125LSQUnit<Impl>::incrStIdx(int &store_idx) 1126{ 1127 if (++store_idx >= SQEntries) 1128 store_idx = 0; 1129} 1130 1131template <class Impl> 1132inline void 1133LSQUnit<Impl>::decrStIdx(int &store_idx) 1134{ 1135 if (--store_idx < 0) 1136 store_idx += SQEntries; 1137} 1138 1139template <class Impl> 1140inline void 1141LSQUnit<Impl>::incrLdIdx(int &load_idx) 1142{ 1143 if (++load_idx >= LQEntries) 1144 load_idx = 0; 1145} 1146 1147template <class Impl> 1148inline void 1149LSQUnit<Impl>::decrLdIdx(int &load_idx) 1150{ 1151 if (--load_idx < 0) 1152 load_idx += LQEntries; 1153} 1154 1155template <class Impl> 1156void 1157LSQUnit<Impl>::dumpInsts() 1158{ 1159 cprintf("Load store queue: Dumping instructions.\n"); 1160 cprintf("Load queue size: %i\n", loads); 1161 cprintf("Load queue: "); 1162 1163 int load_idx = loadHead; 1164 1165 while (load_idx != loadTail && loadQueue[load_idx]) { 1166 cprintf("%s ", loadQueue[load_idx]->pcState()); 1167 1168 incrLdIdx(load_idx); 1169 } 1170 1171 cprintf("Store queue size: %i\n", stores); 1172 cprintf("Store queue: "); 1173 1174 int store_idx = storeHead; 1175 1176 while (store_idx != storeTail && storeQueue[store_idx].inst) { 1177 cprintf("%s ", storeQueue[store_idx].inst->pcState()); 1178 1179 incrStIdx(store_idx); 1180 } 1181 1182 cprintf("\n"); 1183} 1184