lsq_unit_impl.hh revision 9358:aa761458ddcb
1/* 2 * Copyright (c) 2010-2011 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/generic/debugfaults.hh" 45#include "arch/locked_mem.hh" 46#include "base/str.hh" 47#include "config/the_isa.hh" 48#include "cpu/checker/cpu.hh" 49#include "cpu/o3/lsq.hh" 50#include "cpu/o3/lsq_unit.hh" 51#include "debug/Activity.hh" 52#include "debug/IEW.hh" 53#include "debug/LSQUnit.hh" 54#include "mem/packet.hh" 55#include "mem/request.hh" 56 57template<class Impl> 58LSQUnit<Impl>::WritebackEvent::WritebackEvent(DynInstPtr &_inst, PacketPtr _pkt, 59 LSQUnit *lsq_ptr) 60 : Event(Default_Pri, AutoDelete), 61 inst(_inst), pkt(_pkt), lsqPtr(lsq_ptr) 62{ 63} 64 65template<class Impl> 66void 67LSQUnit<Impl>::WritebackEvent::process() 68{ 69 if (!lsqPtr->isSwitchedOut()) { 70 lsqPtr->writeback(inst, pkt); 71 } 72 73 if (pkt->senderState) 74 delete pkt->senderState; 75 76 delete pkt->req; 77 delete pkt; 78} 79 80template<class Impl> 81const char * 82LSQUnit<Impl>::WritebackEvent::description() const 83{ 84 return "Store writeback"; 85} 86 87template<class Impl> 88void 89LSQUnit<Impl>::completeDataAccess(PacketPtr pkt) 90{ 91 LSQSenderState *state = dynamic_cast<LSQSenderState *>(pkt->senderState); 92 DynInstPtr inst = state->inst; 93 DPRINTF(IEW, "Writeback event [sn:%lli].\n", inst->seqNum); 94 DPRINTF(Activity, "Activity: Writeback event [sn:%lli].\n", inst->seqNum); 95 96 //iewStage->ldstQueue.removeMSHR(inst->threadNumber,inst->seqNum); 97 98 // If this is a split access, wait until all packets are received. 99 if (TheISA::HasUnalignedMemAcc && !state->complete()) { 100 delete pkt->req; 101 delete pkt; 102 return; 103 } 104 105 if (isSwitchedOut() || inst->isSquashed()) { 106 iewStage->decrWb(inst->seqNum); 107 } else { 108 if (!state->noWB) { 109 if (!TheISA::HasUnalignedMemAcc || !state->isSplit || 110 !state->isLoad) { 111 writeback(inst, pkt); 112 } else { 113 writeback(inst, state->mainPkt); 114 } 115 } 116 117 if (inst->isStore()) { 118 completeStore(state->idx); 119 } 120 } 121 122 if (TheISA::HasUnalignedMemAcc && state->isSplit && state->isLoad) { 123 delete state->mainPkt->req; 124 delete state->mainPkt; 125 } 126 delete state; 127 delete pkt->req; 128 delete pkt; 129} 130 131template <class Impl> 132LSQUnit<Impl>::LSQUnit() 133 : loads(0), stores(0), storesToWB(0), cacheBlockMask(0), stalled(false), 134 isStoreBlocked(false), isLoadBlocked(false), 135 loadBlockedHandled(false), storeInFlight(false), hasPendingPkt(false) 136{ 137} 138 139template<class Impl> 140void 141LSQUnit<Impl>::init(O3CPU *cpu_ptr, IEW *iew_ptr, DerivO3CPUParams *params, 142 LSQ *lsq_ptr, unsigned maxLQEntries, unsigned maxSQEntries, 143 unsigned id) 144{ 145 cpu = cpu_ptr; 146 iewStage = iew_ptr; 147 148 DPRINTF(LSQUnit, "Creating LSQUnit%i object.\n",id); 149 150 switchedOut = false; 151 152 cacheBlockMask = 0; 153 154 lsq = lsq_ptr; 155 156 lsqID = id; 157 158 // Add 1 for the sentinel entry (they are circular queues). 159 LQEntries = maxLQEntries + 1; 160 SQEntries = maxSQEntries + 1; 161 162 loadQueue.resize(LQEntries); 163 storeQueue.resize(SQEntries); 164 165 depCheckShift = params->LSQDepCheckShift; 166 checkLoads = params->LSQCheckLoads; 167 168 loadHead = loadTail = 0; 169 170 storeHead = storeWBIdx = storeTail = 0; 171 172 usedPorts = 0; 173 cachePorts = params->cachePorts; 174 175 retryPkt = NULL; 176 memDepViolator = NULL; 177 178 blockedLoadSeqNum = 0; 179 needsTSO = params->needsTSO; 180} 181 182template<class Impl> 183std::string 184LSQUnit<Impl>::name() const 185{ 186 if (Impl::MaxThreads == 1) { 187 return iewStage->name() + ".lsq"; 188 } else { 189 return iewStage->name() + ".lsq.thread" + to_string(lsqID); 190 } 191} 192 193template<class Impl> 194void 195LSQUnit<Impl>::regStats() 196{ 197 lsqForwLoads 198 .name(name() + ".forwLoads") 199 .desc("Number of loads that had data forwarded from stores"); 200 201 invAddrLoads 202 .name(name() + ".invAddrLoads") 203 .desc("Number of loads ignored due to an invalid address"); 204 205 lsqSquashedLoads 206 .name(name() + ".squashedLoads") 207 .desc("Number of loads squashed"); 208 209 lsqIgnoredResponses 210 .name(name() + ".ignoredResponses") 211 .desc("Number of memory responses ignored because the instruction is squashed"); 212 213 lsqMemOrderViolation 214 .name(name() + ".memOrderViolation") 215 .desc("Number of memory ordering violations"); 216 217 lsqSquashedStores 218 .name(name() + ".squashedStores") 219 .desc("Number of stores squashed"); 220 221 invAddrSwpfs 222 .name(name() + ".invAddrSwpfs") 223 .desc("Number of software prefetches ignored due to an invalid address"); 224 225 lsqBlockedLoads 226 .name(name() + ".blockedLoads") 227 .desc("Number of blocked loads due to partial load-store forwarding"); 228 229 lsqRescheduledLoads 230 .name(name() + ".rescheduledLoads") 231 .desc("Number of loads that were rescheduled"); 232 233 lsqCacheBlocked 234 .name(name() + ".cacheBlocked") 235 .desc("Number of times an access to memory failed due to the cache being blocked"); 236} 237 238template<class Impl> 239void 240LSQUnit<Impl>::setDcachePort(MasterPort *dcache_port) 241{ 242 dcachePort = dcache_port; 243} 244 245template<class Impl> 246void 247LSQUnit<Impl>::clearLQ() 248{ 249 loadQueue.clear(); 250} 251 252template<class Impl> 253void 254LSQUnit<Impl>::clearSQ() 255{ 256 storeQueue.clear(); 257} 258 259template<class Impl> 260void 261LSQUnit<Impl>::switchOut() 262{ 263 switchedOut = true; 264 for (int i = 0; i < loadQueue.size(); ++i) { 265 assert(!loadQueue[i]); 266 loadQueue[i] = NULL; 267 } 268 269 assert(storesToWB == 0); 270} 271 272template<class Impl> 273void 274LSQUnit<Impl>::takeOverFrom() 275{ 276 switchedOut = false; 277 loads = stores = storesToWB = 0; 278 279 loadHead = loadTail = 0; 280 281 storeHead = storeWBIdx = storeTail = 0; 282 283 usedPorts = 0; 284 285 memDepViolator = NULL; 286 287 blockedLoadSeqNum = 0; 288 289 stalled = false; 290 isLoadBlocked = false; 291 loadBlockedHandled = false; 292 293 // Just incase the memory system changed out from under us 294 cacheBlockMask = 0; 295} 296 297template<class Impl> 298void 299LSQUnit<Impl>::resizeLQ(unsigned size) 300{ 301 unsigned size_plus_sentinel = size + 1; 302 assert(size_plus_sentinel >= LQEntries); 303 304 if (size_plus_sentinel > LQEntries) { 305 while (size_plus_sentinel > loadQueue.size()) { 306 DynInstPtr dummy; 307 loadQueue.push_back(dummy); 308 LQEntries++; 309 } 310 } else { 311 LQEntries = size_plus_sentinel; 312 } 313 314} 315 316template<class Impl> 317void 318LSQUnit<Impl>::resizeSQ(unsigned size) 319{ 320 unsigned size_plus_sentinel = size + 1; 321 if (size_plus_sentinel > SQEntries) { 322 while (size_plus_sentinel > storeQueue.size()) { 323 SQEntry dummy; 324 storeQueue.push_back(dummy); 325 SQEntries++; 326 } 327 } else { 328 SQEntries = size_plus_sentinel; 329 } 330} 331 332template <class Impl> 333void 334LSQUnit<Impl>::insert(DynInstPtr &inst) 335{ 336 assert(inst->isMemRef()); 337 338 assert(inst->isLoad() || inst->isStore()); 339 340 if (inst->isLoad()) { 341 insertLoad(inst); 342 } else { 343 insertStore(inst); 344 } 345 346 inst->setInLSQ(); 347} 348 349template <class Impl> 350void 351LSQUnit<Impl>::insertLoad(DynInstPtr &load_inst) 352{ 353 assert((loadTail + 1) % LQEntries != loadHead); 354 assert(loads < LQEntries); 355 356 DPRINTF(LSQUnit, "Inserting load PC %s, idx:%i [sn:%lli]\n", 357 load_inst->pcState(), loadTail, load_inst->seqNum); 358 359 load_inst->lqIdx = loadTail; 360 361 if (stores == 0) { 362 load_inst->sqIdx = -1; 363 } else { 364 load_inst->sqIdx = storeTail; 365 } 366 367 loadQueue[loadTail] = load_inst; 368 369 incrLdIdx(loadTail); 370 371 ++loads; 372} 373 374template <class Impl> 375void 376LSQUnit<Impl>::insertStore(DynInstPtr &store_inst) 377{ 378 // Make sure it is not full before inserting an instruction. 379 assert((storeTail + 1) % SQEntries != storeHead); 380 assert(stores < SQEntries); 381 382 DPRINTF(LSQUnit, "Inserting store PC %s, idx:%i [sn:%lli]\n", 383 store_inst->pcState(), storeTail, store_inst->seqNum); 384 385 store_inst->sqIdx = storeTail; 386 store_inst->lqIdx = loadTail; 387 388 storeQueue[storeTail] = SQEntry(store_inst); 389 390 incrStIdx(storeTail); 391 392 ++stores; 393} 394 395template <class Impl> 396typename Impl::DynInstPtr 397LSQUnit<Impl>::getMemDepViolator() 398{ 399 DynInstPtr temp = memDepViolator; 400 401 memDepViolator = NULL; 402 403 return temp; 404} 405 406template <class Impl> 407unsigned 408LSQUnit<Impl>::numFreeEntries() 409{ 410 unsigned free_lq_entries = LQEntries - loads; 411 unsigned free_sq_entries = SQEntries - stores; 412 413 // Both the LQ and SQ entries have an extra dummy entry to differentiate 414 // empty/full conditions. Subtract 1 from the free entries. 415 if (free_lq_entries < free_sq_entries) { 416 return free_lq_entries - 1; 417 } else { 418 return free_sq_entries - 1; 419 } 420} 421 422template <class Impl> 423void 424LSQUnit<Impl>::checkSnoop(PacketPtr pkt) 425{ 426 int load_idx = loadHead; 427 428 if (!cacheBlockMask) { 429 assert(dcachePort); 430 Addr bs = dcachePort->peerBlockSize(); 431 432 // Make sure we actually got a size 433 assert(bs != 0); 434 435 cacheBlockMask = ~(bs - 1); 436 } 437 438 // If this is the only load in the LSQ we don't care 439 if (load_idx == loadTail) 440 return; 441 incrLdIdx(load_idx); 442 443 DPRINTF(LSQUnit, "Got snoop for address %#x\n", pkt->getAddr()); 444 Addr invalidate_addr = pkt->getAddr() & cacheBlockMask; 445 while (load_idx != loadTail) { 446 DynInstPtr ld_inst = loadQueue[load_idx]; 447 448 if (!ld_inst->effAddrValid() || ld_inst->uncacheable()) { 449 incrLdIdx(load_idx); 450 continue; 451 } 452 453 Addr load_addr = ld_inst->physEffAddr & cacheBlockMask; 454 DPRINTF(LSQUnit, "-- inst [sn:%lli] load_addr: %#x to pktAddr:%#x\n", 455 ld_inst->seqNum, load_addr, invalidate_addr); 456 457 if (load_addr == invalidate_addr) { 458 if (ld_inst->possibleLoadViolation()) { 459 DPRINTF(LSQUnit, "Conflicting load at addr %#x [sn:%lli]\n", 460 ld_inst->physEffAddr, pkt->getAddr(), ld_inst->seqNum); 461 462 // Mark the load for re-execution 463 ld_inst->fault = new ReExec; 464 } else { 465 // If a older load checks this and it's true 466 // then we might have missed the snoop 467 // in which case we need to invalidate to be sure 468 ld_inst->hitExternalSnoop(true); 469 } 470 } 471 incrLdIdx(load_idx); 472 } 473 return; 474} 475 476template <class Impl> 477Fault 478LSQUnit<Impl>::checkViolations(int load_idx, DynInstPtr &inst) 479{ 480 Addr inst_eff_addr1 = inst->effAddr >> depCheckShift; 481 Addr inst_eff_addr2 = (inst->effAddr + inst->effSize - 1) >> depCheckShift; 482 483 /** @todo in theory you only need to check an instruction that has executed 484 * however, there isn't a good way in the pipeline at the moment to check 485 * all instructions that will execute before the store writes back. Thus, 486 * like the implementation that came before it, we're overly conservative. 487 */ 488 while (load_idx != loadTail) { 489 DynInstPtr ld_inst = loadQueue[load_idx]; 490 if (!ld_inst->effAddrValid() || ld_inst->uncacheable()) { 491 incrLdIdx(load_idx); 492 continue; 493 } 494 495 Addr ld_eff_addr1 = ld_inst->effAddr >> depCheckShift; 496 Addr ld_eff_addr2 = 497 (ld_inst->effAddr + ld_inst->effSize - 1) >> depCheckShift; 498 499 if (inst_eff_addr2 >= ld_eff_addr1 && inst_eff_addr1 <= ld_eff_addr2) { 500 if (inst->isLoad()) { 501 // If this load is to the same block as an external snoop 502 // invalidate that we've observed then the load needs to be 503 // squashed as it could have newer data 504 if (ld_inst->hitExternalSnoop()) { 505 if (!memDepViolator || 506 ld_inst->seqNum < memDepViolator->seqNum) { 507 DPRINTF(LSQUnit, "Detected fault with inst [sn:%lli] " 508 "and [sn:%lli] at address %#x\n", 509 inst->seqNum, ld_inst->seqNum, ld_eff_addr1); 510 memDepViolator = ld_inst; 511 512 ++lsqMemOrderViolation; 513 514 return new GenericISA::M5PanicFault( 515 "Detected fault with inst [sn:%lli] and " 516 "[sn:%lli] at address %#x\n", 517 inst->seqNum, ld_inst->seqNum, ld_eff_addr1); 518 } 519 } 520 521 // Otherwise, mark the load has a possible load violation 522 // and if we see a snoop before it's commited, we need to squash 523 ld_inst->possibleLoadViolation(true); 524 DPRINTF(LSQUnit, "Found possible load violaiton at addr: %#x" 525 " between instructions [sn:%lli] and [sn:%lli]\n", 526 inst_eff_addr1, inst->seqNum, ld_inst->seqNum); 527 } else { 528 // A load/store incorrectly passed this store. 529 // Check if we already have a violator, or if it's newer 530 // squash and refetch. 531 if (memDepViolator && ld_inst->seqNum > memDepViolator->seqNum) 532 break; 533 534 DPRINTF(LSQUnit, "Detected fault with inst [sn:%lli] and " 535 "[sn:%lli] at address %#x\n", 536 inst->seqNum, ld_inst->seqNum, ld_eff_addr1); 537 memDepViolator = ld_inst; 538 539 ++lsqMemOrderViolation; 540 541 return new GenericISA::M5PanicFault("Detected fault with " 542 "inst [sn:%lli] and [sn:%lli] at address %#x\n", 543 inst->seqNum, ld_inst->seqNum, ld_eff_addr1); 544 } 545 } 546 547 incrLdIdx(load_idx); 548 } 549 return NoFault; 550} 551 552 553 554 555template <class Impl> 556Fault 557LSQUnit<Impl>::executeLoad(DynInstPtr &inst) 558{ 559 using namespace TheISA; 560 // Execute a specific load. 561 Fault load_fault = NoFault; 562 563 DPRINTF(LSQUnit, "Executing load PC %s, [sn:%lli]\n", 564 inst->pcState(), inst->seqNum); 565 566 assert(!inst->isSquashed()); 567 568 load_fault = inst->initiateAcc(); 569 570 if (inst->isTranslationDelayed() && 571 load_fault == NoFault) 572 return load_fault; 573 574 // If the instruction faulted or predicated false, then we need to send it 575 // along to commit without the instruction completing. 576 if (load_fault != NoFault || inst->readPredicate() == false) { 577 // Send this instruction to commit, also make sure iew stage 578 // realizes there is activity. 579 // Mark it as executed unless it is an uncached load that 580 // needs to hit the head of commit. 581 if (inst->readPredicate() == false) 582 inst->forwardOldRegs(); 583 DPRINTF(LSQUnit, "Load [sn:%lli] not executed from %s\n", 584 inst->seqNum, 585 (load_fault != NoFault ? "fault" : "predication")); 586 if (!(inst->hasRequest() && inst->uncacheable()) || 587 inst->isAtCommit()) { 588 inst->setExecuted(); 589 } 590 iewStage->instToCommit(inst); 591 iewStage->activityThisCycle(); 592 } else if (!loadBlocked()) { 593 assert(inst->effAddrValid()); 594 int load_idx = inst->lqIdx; 595 incrLdIdx(load_idx); 596 597 if (checkLoads) 598 return checkViolations(load_idx, inst); 599 } 600 601 return load_fault; 602} 603 604template <class Impl> 605Fault 606LSQUnit<Impl>::executeStore(DynInstPtr &store_inst) 607{ 608 using namespace TheISA; 609 // Make sure that a store exists. 610 assert(stores != 0); 611 612 int store_idx = store_inst->sqIdx; 613 614 DPRINTF(LSQUnit, "Executing store PC %s [sn:%lli]\n", 615 store_inst->pcState(), store_inst->seqNum); 616 617 assert(!store_inst->isSquashed()); 618 619 // Check the recently completed loads to see if any match this store's 620 // address. If so, then we have a memory ordering violation. 621 int load_idx = store_inst->lqIdx; 622 623 Fault store_fault = store_inst->initiateAcc(); 624 625 if (store_inst->isTranslationDelayed() && 626 store_fault == NoFault) 627 return store_fault; 628 629 if (store_inst->readPredicate() == false) 630 store_inst->forwardOldRegs(); 631 632 if (storeQueue[store_idx].size == 0) { 633 DPRINTF(LSQUnit,"Fault on Store PC %s, [sn:%lli], Size = 0\n", 634 store_inst->pcState(), store_inst->seqNum); 635 636 return store_fault; 637 } else if (store_inst->readPredicate() == false) { 638 DPRINTF(LSQUnit, "Store [sn:%lli] not executed from predication\n", 639 store_inst->seqNum); 640 return store_fault; 641 } 642 643 assert(store_fault == NoFault); 644 645 if (store_inst->isStoreConditional()) { 646 // Store conditionals need to set themselves as able to 647 // writeback if we haven't had a fault by here. 648 storeQueue[store_idx].canWB = true; 649 650 ++storesToWB; 651 } 652 653 return checkViolations(load_idx, store_inst); 654 655} 656 657template <class Impl> 658void 659LSQUnit<Impl>::commitLoad() 660{ 661 assert(loadQueue[loadHead]); 662 663 DPRINTF(LSQUnit, "Committing head load instruction, PC %s\n", 664 loadQueue[loadHead]->pcState()); 665 666 loadQueue[loadHead] = NULL; 667 668 incrLdIdx(loadHead); 669 670 --loads; 671} 672 673template <class Impl> 674void 675LSQUnit<Impl>::commitLoads(InstSeqNum &youngest_inst) 676{ 677 assert(loads == 0 || loadQueue[loadHead]); 678 679 while (loads != 0 && loadQueue[loadHead]->seqNum <= youngest_inst) { 680 commitLoad(); 681 } 682} 683 684template <class Impl> 685void 686LSQUnit<Impl>::commitStores(InstSeqNum &youngest_inst) 687{ 688 assert(stores == 0 || storeQueue[storeHead].inst); 689 690 int store_idx = storeHead; 691 692 while (store_idx != storeTail) { 693 assert(storeQueue[store_idx].inst); 694 // Mark any stores that are now committed and have not yet 695 // been marked as able to write back. 696 if (!storeQueue[store_idx].canWB) { 697 if (storeQueue[store_idx].inst->seqNum > youngest_inst) { 698 break; 699 } 700 DPRINTF(LSQUnit, "Marking store as able to write back, PC " 701 "%s [sn:%lli]\n", 702 storeQueue[store_idx].inst->pcState(), 703 storeQueue[store_idx].inst->seqNum); 704 705 storeQueue[store_idx].canWB = true; 706 707 ++storesToWB; 708 } 709 710 incrStIdx(store_idx); 711 } 712} 713 714template <class Impl> 715void 716LSQUnit<Impl>::writebackPendingStore() 717{ 718 if (hasPendingPkt) { 719 assert(pendingPkt != NULL); 720 721 // If the cache is blocked, this will store the packet for retry. 722 if (sendStore(pendingPkt)) { 723 storePostSend(pendingPkt); 724 } 725 pendingPkt = NULL; 726 hasPendingPkt = false; 727 } 728} 729 730template <class Impl> 731void 732LSQUnit<Impl>::writebackStores() 733{ 734 // First writeback the second packet from any split store that didn't 735 // complete last cycle because there weren't enough cache ports available. 736 if (TheISA::HasUnalignedMemAcc) { 737 writebackPendingStore(); 738 } 739 740 while (storesToWB > 0 && 741 storeWBIdx != storeTail && 742 storeQueue[storeWBIdx].inst && 743 storeQueue[storeWBIdx].canWB && 744 ((!needsTSO) || (!storeInFlight)) && 745 usedPorts < cachePorts) { 746 747 if (isStoreBlocked || lsq->cacheBlocked()) { 748 DPRINTF(LSQUnit, "Unable to write back any more stores, cache" 749 " is blocked!\n"); 750 break; 751 } 752 753 // Store didn't write any data so no need to write it back to 754 // memory. 755 if (storeQueue[storeWBIdx].size == 0) { 756 completeStore(storeWBIdx); 757 758 incrStIdx(storeWBIdx); 759 760 continue; 761 } 762 763 ++usedPorts; 764 765 if (storeQueue[storeWBIdx].inst->isDataPrefetch()) { 766 incrStIdx(storeWBIdx); 767 768 continue; 769 } 770 771 assert(storeQueue[storeWBIdx].req); 772 assert(!storeQueue[storeWBIdx].committed); 773 774 if (TheISA::HasUnalignedMemAcc && storeQueue[storeWBIdx].isSplit) { 775 assert(storeQueue[storeWBIdx].sreqLow); 776 assert(storeQueue[storeWBIdx].sreqHigh); 777 } 778 779 DynInstPtr inst = storeQueue[storeWBIdx].inst; 780 781 Request *req = storeQueue[storeWBIdx].req; 782 RequestPtr sreqLow = storeQueue[storeWBIdx].sreqLow; 783 RequestPtr sreqHigh = storeQueue[storeWBIdx].sreqHigh; 784 785 storeQueue[storeWBIdx].committed = true; 786 787 assert(!inst->memData); 788 inst->memData = new uint8_t[64]; 789 790 memcpy(inst->memData, storeQueue[storeWBIdx].data, req->getSize()); 791 792 MemCmd command = 793 req->isSwap() ? MemCmd::SwapReq : 794 (req->isLLSC() ? MemCmd::StoreCondReq : MemCmd::WriteReq); 795 PacketPtr data_pkt; 796 PacketPtr snd_data_pkt = NULL; 797 798 LSQSenderState *state = new LSQSenderState; 799 state->isLoad = false; 800 state->idx = storeWBIdx; 801 state->inst = inst; 802 803 if (!TheISA::HasUnalignedMemAcc || !storeQueue[storeWBIdx].isSplit) { 804 805 // Build a single data packet if the store isn't split. 806 data_pkt = new Packet(req, command); 807 data_pkt->dataStatic(inst->memData); 808 data_pkt->senderState = state; 809 } else { 810 // Create two packets if the store is split in two. 811 data_pkt = new Packet(sreqLow, command); 812 snd_data_pkt = new Packet(sreqHigh, command); 813 814 data_pkt->dataStatic(inst->memData); 815 snd_data_pkt->dataStatic(inst->memData + sreqLow->getSize()); 816 817 data_pkt->senderState = state; 818 snd_data_pkt->senderState = state; 819 820 state->isSplit = true; 821 state->outstanding = 2; 822 823 // Can delete the main request now. 824 delete req; 825 req = sreqLow; 826 } 827 828 DPRINTF(LSQUnit, "D-Cache: Writing back store idx:%i PC:%s " 829 "to Addr:%#x, data:%#x [sn:%lli]\n", 830 storeWBIdx, inst->pcState(), 831 req->getPaddr(), (int)*(inst->memData), 832 inst->seqNum); 833 834 // @todo: Remove this SC hack once the memory system handles it. 835 if (inst->isStoreConditional()) { 836 assert(!storeQueue[storeWBIdx].isSplit); 837 // Disable recording the result temporarily. Writing to 838 // misc regs normally updates the result, but this is not 839 // the desired behavior when handling store conditionals. 840 inst->recordResult(false); 841 bool success = TheISA::handleLockedWrite(inst.get(), req); 842 inst->recordResult(true); 843 844 if (!success) { 845 // Instantly complete this store. 846 DPRINTF(LSQUnit, "Store conditional [sn:%lli] failed. " 847 "Instantly completing it.\n", 848 inst->seqNum); 849 WritebackEvent *wb = new WritebackEvent(inst, data_pkt, this); 850 cpu->schedule(wb, curTick() + 1); 851 if (cpu->checker) { 852 // Make sure to set the LLSC data for verification 853 // if checker is loaded 854 inst->reqToVerify->setExtraData(0); 855 inst->completeAcc(data_pkt); 856 } 857 completeStore(storeWBIdx); 858 incrStIdx(storeWBIdx); 859 continue; 860 } 861 } else { 862 // Non-store conditionals do not need a writeback. 863 state->noWB = true; 864 } 865 866 bool split = 867 TheISA::HasUnalignedMemAcc && storeQueue[storeWBIdx].isSplit; 868 869 ThreadContext *thread = cpu->tcBase(lsqID); 870 871 if (req->isMmappedIpr()) { 872 assert(!inst->isStoreConditional()); 873 TheISA::handleIprWrite(thread, data_pkt); 874 delete data_pkt; 875 if (split) { 876 assert(snd_data_pkt->req->isMmappedIpr()); 877 TheISA::handleIprWrite(thread, snd_data_pkt); 878 delete snd_data_pkt; 879 delete sreqLow; 880 delete sreqHigh; 881 } 882 delete state; 883 delete req; 884 completeStore(storeWBIdx); 885 incrStIdx(storeWBIdx); 886 } else if (!sendStore(data_pkt)) { 887 DPRINTF(IEW, "D-Cache became blocked when writing [sn:%lli], will" 888 "retry later\n", 889 inst->seqNum); 890 891 // Need to store the second packet, if split. 892 if (split) { 893 state->pktToSend = true; 894 state->pendingPacket = snd_data_pkt; 895 } 896 } else { 897 898 // If split, try to send the second packet too 899 if (split) { 900 assert(snd_data_pkt); 901 902 // Ensure there are enough ports to use. 903 if (usedPorts < cachePorts) { 904 ++usedPorts; 905 if (sendStore(snd_data_pkt)) { 906 storePostSend(snd_data_pkt); 907 } else { 908 DPRINTF(IEW, "D-Cache became blocked when writing" 909 " [sn:%lli] second packet, will retry later\n", 910 inst->seqNum); 911 } 912 } else { 913 914 // Store the packet for when there's free ports. 915 assert(pendingPkt == NULL); 916 pendingPkt = snd_data_pkt; 917 hasPendingPkt = true; 918 } 919 } else { 920 921 // Not a split store. 922 storePostSend(data_pkt); 923 } 924 } 925 } 926 927 // Not sure this should set it to 0. 928 usedPorts = 0; 929 930 assert(stores >= 0 && storesToWB >= 0); 931} 932 933/*template <class Impl> 934void 935LSQUnit<Impl>::removeMSHR(InstSeqNum seqNum) 936{ 937 list<InstSeqNum>::iterator mshr_it = find(mshrSeqNums.begin(), 938 mshrSeqNums.end(), 939 seqNum); 940 941 if (mshr_it != mshrSeqNums.end()) { 942 mshrSeqNums.erase(mshr_it); 943 DPRINTF(LSQUnit, "Removing MSHR. count = %i\n",mshrSeqNums.size()); 944 } 945}*/ 946 947template <class Impl> 948void 949LSQUnit<Impl>::squash(const InstSeqNum &squashed_num) 950{ 951 DPRINTF(LSQUnit, "Squashing until [sn:%lli]!" 952 "(Loads:%i Stores:%i)\n", squashed_num, loads, stores); 953 954 int load_idx = loadTail; 955 decrLdIdx(load_idx); 956 957 while (loads != 0 && loadQueue[load_idx]->seqNum > squashed_num) { 958 DPRINTF(LSQUnit,"Load Instruction PC %s squashed, " 959 "[sn:%lli]\n", 960 loadQueue[load_idx]->pcState(), 961 loadQueue[load_idx]->seqNum); 962 963 if (isStalled() && load_idx == stallingLoadIdx) { 964 stalled = false; 965 stallingStoreIsn = 0; 966 stallingLoadIdx = 0; 967 } 968 969 // Clear the smart pointer to make sure it is decremented. 970 loadQueue[load_idx]->setSquashed(); 971 loadQueue[load_idx] = NULL; 972 --loads; 973 974 // Inefficient! 975 loadTail = load_idx; 976 977 decrLdIdx(load_idx); 978 ++lsqSquashedLoads; 979 } 980 981 if (isLoadBlocked) { 982 if (squashed_num < blockedLoadSeqNum) { 983 isLoadBlocked = false; 984 loadBlockedHandled = false; 985 blockedLoadSeqNum = 0; 986 } 987 } 988 989 if (memDepViolator && squashed_num < memDepViolator->seqNum) { 990 memDepViolator = NULL; 991 } 992 993 int store_idx = storeTail; 994 decrStIdx(store_idx); 995 996 while (stores != 0 && 997 storeQueue[store_idx].inst->seqNum > squashed_num) { 998 // Instructions marked as can WB are already committed. 999 if (storeQueue[store_idx].canWB) { 1000 break; 1001 } 1002 1003 DPRINTF(LSQUnit,"Store Instruction PC %s squashed, " 1004 "idx:%i [sn:%lli]\n", 1005 storeQueue[store_idx].inst->pcState(), 1006 store_idx, storeQueue[store_idx].inst->seqNum); 1007 1008 // I don't think this can happen. It should have been cleared 1009 // by the stalling load. 1010 if (isStalled() && 1011 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) { 1012 panic("Is stalled should have been cleared by stalling load!\n"); 1013 stalled = false; 1014 stallingStoreIsn = 0; 1015 } 1016 1017 // Clear the smart pointer to make sure it is decremented. 1018 storeQueue[store_idx].inst->setSquashed(); 1019 storeQueue[store_idx].inst = NULL; 1020 storeQueue[store_idx].canWB = 0; 1021 1022 // Must delete request now that it wasn't handed off to 1023 // memory. This is quite ugly. @todo: Figure out the proper 1024 // place to really handle request deletes. 1025 delete storeQueue[store_idx].req; 1026 if (TheISA::HasUnalignedMemAcc && storeQueue[store_idx].isSplit) { 1027 delete storeQueue[store_idx].sreqLow; 1028 delete storeQueue[store_idx].sreqHigh; 1029 1030 storeQueue[store_idx].sreqLow = NULL; 1031 storeQueue[store_idx].sreqHigh = NULL; 1032 } 1033 1034 storeQueue[store_idx].req = NULL; 1035 --stores; 1036 1037 // Inefficient! 1038 storeTail = store_idx; 1039 1040 decrStIdx(store_idx); 1041 ++lsqSquashedStores; 1042 } 1043} 1044 1045template <class Impl> 1046void 1047LSQUnit<Impl>::storePostSend(PacketPtr pkt) 1048{ 1049 if (isStalled() && 1050 storeQueue[storeWBIdx].inst->seqNum == stallingStoreIsn) { 1051 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] " 1052 "load idx:%i\n", 1053 stallingStoreIsn, stallingLoadIdx); 1054 stalled = false; 1055 stallingStoreIsn = 0; 1056 iewStage->replayMemInst(loadQueue[stallingLoadIdx]); 1057 } 1058 1059 if (!storeQueue[storeWBIdx].inst->isStoreConditional()) { 1060 // The store is basically completed at this time. This 1061 // only works so long as the checker doesn't try to 1062 // verify the value in memory for stores. 1063 storeQueue[storeWBIdx].inst->setCompleted(); 1064 1065 if (cpu->checker) { 1066 cpu->checker->verify(storeQueue[storeWBIdx].inst); 1067 } 1068 } 1069 1070 if (needsTSO) { 1071 storeInFlight = true; 1072 } 1073 1074 incrStIdx(storeWBIdx); 1075} 1076 1077template <class Impl> 1078void 1079LSQUnit<Impl>::writeback(DynInstPtr &inst, PacketPtr pkt) 1080{ 1081 iewStage->wakeCPU(); 1082 1083 // Squashed instructions do not need to complete their access. 1084 if (inst->isSquashed()) { 1085 iewStage->decrWb(inst->seqNum); 1086 assert(!inst->isStore()); 1087 ++lsqIgnoredResponses; 1088 return; 1089 } 1090 1091 if (!inst->isExecuted()) { 1092 inst->setExecuted(); 1093 1094 // Complete access to copy data to proper place. 1095 inst->completeAcc(pkt); 1096 } 1097 1098 // Need to insert instruction into queue to commit 1099 iewStage->instToCommit(inst); 1100 1101 iewStage->activityThisCycle(); 1102 1103 // see if this load changed the PC 1104 iewStage->checkMisprediction(inst); 1105} 1106 1107template <class Impl> 1108void 1109LSQUnit<Impl>::completeStore(int store_idx) 1110{ 1111 assert(storeQueue[store_idx].inst); 1112 storeQueue[store_idx].completed = true; 1113 --storesToWB; 1114 // A bit conservative because a store completion may not free up entries, 1115 // but hopefully avoids two store completions in one cycle from making 1116 // the CPU tick twice. 1117 cpu->wakeCPU(); 1118 cpu->activityThisCycle(); 1119 1120 if (store_idx == storeHead) { 1121 do { 1122 incrStIdx(storeHead); 1123 1124 --stores; 1125 } while (storeQueue[storeHead].completed && 1126 storeHead != storeTail); 1127 1128 iewStage->updateLSQNextCycle = true; 1129 } 1130 1131 DPRINTF(LSQUnit, "Completing store [sn:%lli], idx:%i, store head " 1132 "idx:%i\n", 1133 storeQueue[store_idx].inst->seqNum, store_idx, storeHead); 1134 1135 if (isStalled() && 1136 storeQueue[store_idx].inst->seqNum == stallingStoreIsn) { 1137 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] " 1138 "load idx:%i\n", 1139 stallingStoreIsn, stallingLoadIdx); 1140 stalled = false; 1141 stallingStoreIsn = 0; 1142 iewStage->replayMemInst(loadQueue[stallingLoadIdx]); 1143 } 1144 1145 storeQueue[store_idx].inst->setCompleted(); 1146 1147 if (needsTSO) { 1148 storeInFlight = false; 1149 } 1150 1151 // Tell the checker we've completed this instruction. Some stores 1152 // may get reported twice to the checker, but the checker can 1153 // handle that case. 1154 if (cpu->checker) { 1155 cpu->checker->verify(storeQueue[store_idx].inst); 1156 } 1157} 1158 1159template <class Impl> 1160bool 1161LSQUnit<Impl>::sendStore(PacketPtr data_pkt) 1162{ 1163 if (!dcachePort->sendTimingReq(data_pkt)) { 1164 // Need to handle becoming blocked on a store. 1165 isStoreBlocked = true; 1166 ++lsqCacheBlocked; 1167 assert(retryPkt == NULL); 1168 retryPkt = data_pkt; 1169 lsq->setRetryTid(lsqID); 1170 return false; 1171 } 1172 return true; 1173} 1174 1175template <class Impl> 1176void 1177LSQUnit<Impl>::recvRetry() 1178{ 1179 if (isStoreBlocked) { 1180 DPRINTF(LSQUnit, "Receiving retry: store blocked\n"); 1181 assert(retryPkt != NULL); 1182 1183 LSQSenderState *state = 1184 dynamic_cast<LSQSenderState *>(retryPkt->senderState); 1185 1186 if (dcachePort->sendTimingReq(retryPkt)) { 1187 // Don't finish the store unless this is the last packet. 1188 if (!TheISA::HasUnalignedMemAcc || !state->pktToSend || 1189 state->pendingPacket == retryPkt) { 1190 state->pktToSend = false; 1191 storePostSend(retryPkt); 1192 } 1193 retryPkt = NULL; 1194 isStoreBlocked = false; 1195 lsq->setRetryTid(InvalidThreadID); 1196 1197 // Send any outstanding packet. 1198 if (TheISA::HasUnalignedMemAcc && state->pktToSend) { 1199 assert(state->pendingPacket); 1200 if (sendStore(state->pendingPacket)) { 1201 storePostSend(state->pendingPacket); 1202 } 1203 } 1204 } else { 1205 // Still blocked! 1206 ++lsqCacheBlocked; 1207 lsq->setRetryTid(lsqID); 1208 } 1209 } else if (isLoadBlocked) { 1210 DPRINTF(LSQUnit, "Loads squash themselves and all younger insts, " 1211 "no need to resend packet.\n"); 1212 } else { 1213 DPRINTF(LSQUnit, "Retry received but LSQ is no longer blocked.\n"); 1214 } 1215} 1216 1217template <class Impl> 1218inline void 1219LSQUnit<Impl>::incrStIdx(int &store_idx) 1220{ 1221 if (++store_idx >= SQEntries) 1222 store_idx = 0; 1223} 1224 1225template <class Impl> 1226inline void 1227LSQUnit<Impl>::decrStIdx(int &store_idx) 1228{ 1229 if (--store_idx < 0) 1230 store_idx += SQEntries; 1231} 1232 1233template <class Impl> 1234inline void 1235LSQUnit<Impl>::incrLdIdx(int &load_idx) 1236{ 1237 if (++load_idx >= LQEntries) 1238 load_idx = 0; 1239} 1240 1241template <class Impl> 1242inline void 1243LSQUnit<Impl>::decrLdIdx(int &load_idx) 1244{ 1245 if (--load_idx < 0) 1246 load_idx += LQEntries; 1247} 1248 1249template <class Impl> 1250void 1251LSQUnit<Impl>::dumpInsts() 1252{ 1253 cprintf("Load store queue: Dumping instructions.\n"); 1254 cprintf("Load queue size: %i\n", loads); 1255 cprintf("Load queue: "); 1256 1257 int load_idx = loadHead; 1258 1259 while (load_idx != loadTail && loadQueue[load_idx]) { 1260 cprintf("%s ", loadQueue[load_idx]->pcState()); 1261 1262 incrLdIdx(load_idx); 1263 } 1264 1265 cprintf("Store queue size: %i\n", stores); 1266 cprintf("Store queue: "); 1267 1268 int store_idx = storeHead; 1269 1270 while (store_idx != storeTail && storeQueue[store_idx].inst) { 1271 cprintf("%s ", storeQueue[store_idx].inst->pcState()); 1272 1273 incrStIdx(store_idx); 1274 } 1275 1276 cprintf("\n"); 1277} 1278