1 2/* 3 * Copyright (c) 2010-2014, 2017-2018 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 assert(_inst->savedReq); 70 _inst->savedReq->writebackScheduled(); 71} 72 73template<class Impl> 74void 75LSQUnit<Impl>::WritebackEvent::process() 76{ 77 assert(!lsqPtr->cpu->switchedOut()); 78 79 lsqPtr->writeback(inst, pkt); 80 81 assert(inst->savedReq); 82 inst->savedReq->writebackDone(); 83 delete pkt; 84} 85 86template<class Impl> 87const char * 88LSQUnit<Impl>::WritebackEvent::description() const 89{ 90 return "Store writeback"; 91} 92 93template <class Impl> 94bool 95LSQUnit<Impl>::recvTimingResp(PacketPtr pkt) 96{ 97 auto senderState = dynamic_cast<LSQSenderState*>(pkt->senderState); 98 LSQRequest* req = senderState->request(); 99 assert(req != nullptr); 100 bool ret = true; 101 /* Check that the request is still alive before any further action. */ 102 if (senderState->alive()) { 103 ret = req->recvTimingResp(pkt); 104 } else { 105 senderState->outstanding--; 106 } 107 return ret; 108 109} 110 111template<class Impl> 112void 113LSQUnit<Impl>::completeDataAccess(PacketPtr pkt) 114{ 115 LSQSenderState *state = dynamic_cast<LSQSenderState *>(pkt->senderState); 116 DynInstPtr inst = state->inst; 117 118 cpu->ppDataAccessComplete->notify(std::make_pair(inst, pkt)); 119 120 /* Notify the sender state that the access is complete (for ownership 121 * tracking). */ 122 state->complete(); 123 124 assert(!cpu->switchedOut()); 125 if (!inst->isSquashed()) { 126 if (state->needWB) { 127 // Only loads, store conditionals and atomics perform the writeback 128 // after receving the response from the memory 129 assert(inst->isLoad() || inst->isStoreConditional() || 130 inst->isAtomic()); 131 writeback(inst, state->request()->mainPacket()); 132 if (inst->isStore() || inst->isAtomic()) { 133 auto ss = dynamic_cast<SQSenderState*>(state); 134 ss->writebackDone(); 135 completeStore(ss->idx); 136 } 137 } else if (inst->isStore()) { 138 // This is a regular store (i.e., not store conditionals and 139 // atomics), so it can complete without writing back 140 completeStore(dynamic_cast<SQSenderState*>(state)->idx); 141 } 142 } 143} 144 145template <class Impl> 146LSQUnit<Impl>::LSQUnit(uint32_t lqEntries, uint32_t sqEntries) 147 : lsqID(-1), storeQueue(sqEntries+1), loadQueue(lqEntries+1), 148 loads(0), stores(0), storesToWB(0), cacheBlockMask(0), stalled(false), 149 isStoreBlocked(false), storeInFlight(false), hasPendingRequest(false), 150 pendingRequest(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 needsTSO = params->needsTSO; 171 172 resetState(); 173} 174 175 176template<class Impl> 177void 178LSQUnit<Impl>::resetState() 179{ 180 loads = stores = storesToWB = 0; 181 182 183 storeWBIt = storeQueue.begin(); 184 185 retryPkt = NULL; 186 memDepViolator = NULL; 187 188 stalled = false; 189 190 cacheBlockMask = ~(cpu->cacheLineSize() - 1); 191} 192 193template<class Impl> 194std::string 195LSQUnit<Impl>::name() const 196{ 197 if (Impl::MaxThreads == 1) { 198 return iewStage->name() + ".lsq"; 199 } else { 200 return iewStage->name() + ".lsq.thread" + std::to_string(lsqID); 201 } 202} 203 204template<class Impl> 205void 206LSQUnit<Impl>::regStats() 207{ 208 lsqForwLoads 209 .name(name() + ".forwLoads") 210 .desc("Number of loads that had data forwarded from stores"); 211 212 invAddrLoads 213 .name(name() + ".invAddrLoads") 214 .desc("Number of loads ignored due to an invalid address"); 215 216 lsqSquashedLoads 217 .name(name() + ".squashedLoads") 218 .desc("Number of loads squashed"); 219 220 lsqIgnoredResponses 221 .name(name() + ".ignoredResponses") 222 .desc("Number of memory responses ignored because the instruction is squashed"); 223 224 lsqMemOrderViolation 225 .name(name() + ".memOrderViolation") 226 .desc("Number of memory ordering violations"); 227 228 lsqSquashedStores 229 .name(name() + ".squashedStores") 230 .desc("Number of stores squashed"); 231 232 invAddrSwpfs 233 .name(name() + ".invAddrSwpfs") 234 .desc("Number of software prefetches ignored due to an invalid address"); 235 236 lsqBlockedLoads 237 .name(name() + ".blockedLoads") 238 .desc("Number of blocked loads due to partial load-store forwarding"); 239 240 lsqRescheduledLoads 241 .name(name() + ".rescheduledLoads") 242 .desc("Number of loads that were rescheduled"); 243 244 lsqCacheBlocked 245 .name(name() + ".cacheBlocked") 246 .desc("Number of times an access to memory failed due to the cache being blocked"); 247} 248 249template<class Impl> 250void 251LSQUnit<Impl>::setDcachePort(MasterPort *dcache_port) 252{ 253 dcachePort = dcache_port; 254} 255 256template<class Impl> 257void 258LSQUnit<Impl>::drainSanityCheck() const 259{ 260 for (int i = 0; i < loadQueue.capacity(); ++i) 261 assert(!loadQueue[i].valid()); 262 263 assert(storesToWB == 0); 264 assert(!retryPkt); 265} 266 267template<class Impl> 268void 269LSQUnit<Impl>::takeOverFrom() 270{ 271 resetState(); 272} 273 274template <class Impl> 275void 276LSQUnit<Impl>::insert(const DynInstPtr &inst) 277{ 278 assert(inst->isMemRef()); 279 280 assert(inst->isLoad() || inst->isStore() || inst->isAtomic()); 281 282 if (inst->isLoad()) { 283 insertLoad(inst); 284 } else { 285 insertStore(inst); 286 } 287 288 inst->setInLSQ(); 289} 290 291template <class Impl> 292void 293LSQUnit<Impl>::insertLoad(const DynInstPtr &load_inst) 294{ 295 assert(!loadQueue.full()); 296 assert(loads < loadQueue.capacity()); 297 298 DPRINTF(LSQUnit, "Inserting load PC %s, idx:%i [sn:%lli]\n", 299 load_inst->pcState(), loadQueue.tail(), load_inst->seqNum); 300 301 /* Grow the queue. */ 302 loadQueue.advance_tail(); 303 304 load_inst->sqIt = storeQueue.end(); 305 306 assert(!loadQueue.back().valid()); 307 loadQueue.back().set(load_inst); 308 load_inst->lqIdx = loadQueue.tail(); 309 load_inst->lqIt = loadQueue.getIterator(load_inst->lqIdx); 310 311 ++loads; 312} 313 314template <class Impl> 315void 316LSQUnit<Impl>::insertStore(const DynInstPtr& store_inst) 317{ 318 // Make sure it is not full before inserting an instruction. 319 assert(!storeQueue.full()); 320 assert(stores < storeQueue.capacity()); 321 322 DPRINTF(LSQUnit, "Inserting store PC %s, idx:%i [sn:%lli]\n", 323 store_inst->pcState(), storeQueue.tail(), store_inst->seqNum); 324 storeQueue.advance_tail(); 325 326 store_inst->sqIdx = storeQueue.tail(); 327 store_inst->lqIdx = loadQueue.moduloAdd(loadQueue.tail(), 1); 328 store_inst->lqIt = loadQueue.end(); 329 330 storeQueue.back().set(store_inst); 331 332 ++stores; 333} 334 335template <class Impl> 336typename Impl::DynInstPtr 337LSQUnit<Impl>::getMemDepViolator() 338{ 339 DynInstPtr temp = memDepViolator; 340 341 memDepViolator = NULL; 342 343 return temp; 344} 345 346template <class Impl> 347unsigned 348LSQUnit<Impl>::numFreeLoadEntries() 349{ 350 //LQ has an extra dummy entry to differentiate 351 //empty/full conditions. Subtract 1 from the free entries. 352 DPRINTF(LSQUnit, "LQ size: %d, #loads occupied: %d\n", 353 1 + loadQueue.capacity(), loads); 354 return loadQueue.capacity() - loads; 355} 356 357template <class Impl> 358unsigned 359LSQUnit<Impl>::numFreeStoreEntries() 360{ 361 //SQ has an extra dummy entry to differentiate 362 //empty/full conditions. Subtract 1 from the free entries. 363 DPRINTF(LSQUnit, "SQ size: %d, #stores occupied: %d\n", 364 1 + storeQueue.capacity(), stores); 365 return storeQueue.capacity() - stores; 366 367 } 368 369template <class Impl> 370void 371LSQUnit<Impl>::checkSnoop(PacketPtr pkt) 372{ 373 // Should only ever get invalidations in here 374 assert(pkt->isInvalidate()); 375 376 DPRINTF(LSQUnit, "Got snoop for address %#x\n", pkt->getAddr()); 377 378 for (int x = 0; x < cpu->numContexts(); x++) { 379 ThreadContext *tc = cpu->getContext(x); 380 bool no_squash = cpu->thread[x]->noSquashFromTC; 381 cpu->thread[x]->noSquashFromTC = true; 382 TheISA::handleLockedSnoop(tc, pkt, cacheBlockMask); 383 cpu->thread[x]->noSquashFromTC = no_squash; 384 } 385 386 if (loadQueue.empty()) 387 return; 388 389 auto iter = loadQueue.begin(); 390 391 Addr invalidate_addr = pkt->getAddr() & cacheBlockMask; 392 393 DynInstPtr ld_inst = iter->instruction(); 394 assert(ld_inst); 395 LSQRequest *req = iter->request(); 396 397 // Check that this snoop didn't just invalidate our lock flag 398 if (ld_inst->effAddrValid() && 399 req->isCacheBlockHit(invalidate_addr, cacheBlockMask) 400 && ld_inst->memReqFlags & Request::LLSC) 401 TheISA::handleLockedSnoopHit(ld_inst.get()); 402 403 bool force_squash = false; 404 405 while (++iter != loadQueue.end()) { 406 ld_inst = iter->instruction(); 407 assert(ld_inst); 408 req = iter->request(); 409 if (!ld_inst->effAddrValid() || ld_inst->strictlyOrdered()) 410 continue; 411 412 DPRINTF(LSQUnit, "-- inst [sn:%lli] to pktAddr:%#x\n", 413 ld_inst->seqNum, invalidate_addr); 414 415 if (force_squash || 416 req->isCacheBlockHit(invalidate_addr, cacheBlockMask)) { 417 if (needsTSO) { 418 // If we have a TSO system, as all loads must be ordered with 419 // all other loads, this load as well as *all* subsequent loads 420 // need to be squashed to prevent possible load reordering. 421 force_squash = true; 422 } 423 if (ld_inst->possibleLoadViolation() || force_squash) { 424 DPRINTF(LSQUnit, "Conflicting load at addr %#x [sn:%lli]\n", 425 pkt->getAddr(), ld_inst->seqNum); 426 427 // Mark the load for re-execution 428 ld_inst->fault = std::make_shared<ReExec>(); 429 } else { 430 DPRINTF(LSQUnit, "HitExternal Snoop for addr %#x [sn:%lli]\n", 431 pkt->getAddr(), ld_inst->seqNum); 432 433 // Make sure that we don't lose a snoop hitting a LOCKED 434 // address since the LOCK* flags don't get updated until 435 // commit. 436 if (ld_inst->memReqFlags & Request::LLSC) 437 TheISA::handleLockedSnoopHit(ld_inst.get()); 438 439 // If a older load checks this and it's true 440 // then we might have missed the snoop 441 // in which case we need to invalidate to be sure 442 ld_inst->hitExternalSnoop(true); 443 } 444 } 445 } 446 return; 447} 448 449template <class Impl> 450Fault 451LSQUnit<Impl>::checkViolations(typename LoadQueue::iterator& loadIt, 452 const DynInstPtr& inst) 453{ 454 Addr inst_eff_addr1 = inst->effAddr >> depCheckShift; 455 Addr inst_eff_addr2 = (inst->effAddr + inst->effSize - 1) >> depCheckShift; 456 457 /** @todo in theory you only need to check an instruction that has executed 458 * however, there isn't a good way in the pipeline at the moment to check 459 * all instructions that will execute before the store writes back. Thus, 460 * like the implementation that came before it, we're overly conservative. 461 */ 462 while (loadIt != loadQueue.end()) { 463 DynInstPtr ld_inst = loadIt->instruction(); 464 if (!ld_inst->effAddrValid() || ld_inst->strictlyOrdered()) { 465 ++loadIt; 466 continue; 467 } 468 469 Addr ld_eff_addr1 = ld_inst->effAddr >> depCheckShift; 470 Addr ld_eff_addr2 = 471 (ld_inst->effAddr + ld_inst->effSize - 1) >> depCheckShift; 472 473 if (inst_eff_addr2 >= ld_eff_addr1 && inst_eff_addr1 <= ld_eff_addr2) { 474 if (inst->isLoad()) { 475 // If this load is to the same block as an external snoop 476 // invalidate that we've observed then the load needs to be 477 // squashed as it could have newer data 478 if (ld_inst->hitExternalSnoop()) { 479 if (!memDepViolator || 480 ld_inst->seqNum < memDepViolator->seqNum) { 481 DPRINTF(LSQUnit, "Detected fault with inst [sn:%lli] " 482 "and [sn:%lli] at address %#x\n", 483 inst->seqNum, ld_inst->seqNum, ld_eff_addr1); 484 memDepViolator = ld_inst; 485 486 ++lsqMemOrderViolation; 487 488 return std::make_shared<GenericISA::M5PanicFault>( 489 "Detected fault with inst [sn:%lli] and " 490 "[sn:%lli] at address %#x\n", 491 inst->seqNum, ld_inst->seqNum, ld_eff_addr1); 492 } 493 } 494 495 // Otherwise, mark the load has a possible load violation 496 // and if we see a snoop before it's commited, we need to squash 497 ld_inst->possibleLoadViolation(true); 498 DPRINTF(LSQUnit, "Found possible load violation at addr: %#x" 499 " between instructions [sn:%lli] and [sn:%lli]\n", 500 inst_eff_addr1, inst->seqNum, ld_inst->seqNum); 501 } else { 502 // A load/store incorrectly passed this store. 503 // Check if we already have a violator, or if it's newer 504 // squash and refetch. 505 if (memDepViolator && ld_inst->seqNum > memDepViolator->seqNum) 506 break; 507 508 DPRINTF(LSQUnit, "Detected fault with inst [sn:%lli] and " 509 "[sn:%lli] at address %#x\n", 510 inst->seqNum, ld_inst->seqNum, ld_eff_addr1); 511 memDepViolator = ld_inst; 512 513 ++lsqMemOrderViolation; 514 515 return std::make_shared<GenericISA::M5PanicFault>( 516 "Detected fault with " 517 "inst [sn:%lli] and [sn:%lli] at address %#x\n", 518 inst->seqNum, ld_inst->seqNum, ld_eff_addr1); 519 } 520 } 521 522 ++loadIt; 523 } 524 return NoFault; 525} 526 527 528 529 530template <class Impl> 531Fault 532LSQUnit<Impl>::executeLoad(const DynInstPtr &inst) 533{ 534 using namespace TheISA; 535 // Execute a specific load. 536 Fault load_fault = NoFault; 537 538 DPRINTF(LSQUnit, "Executing load PC %s, [sn:%lli]\n", 539 inst->pcState(), inst->seqNum); 540 541 assert(!inst->isSquashed()); 542 543 load_fault = inst->initiateAcc(); 544 545 if (load_fault == NoFault && !inst->readMemAccPredicate()) { 546 assert(inst->readPredicate()); 547 inst->setExecuted(); 548 inst->completeAcc(nullptr); 549 iewStage->instToCommit(inst); 550 iewStage->activityThisCycle(); 551 return NoFault; 552 } 553 554 if (inst->isTranslationDelayed() && load_fault == NoFault) 555 return load_fault; 556 557 if (load_fault != NoFault && inst->translationCompleted() && 558 inst->savedReq->isPartialFault() && !inst->savedReq->isComplete()) { 559 assert(inst->savedReq->isSplit()); 560 // If we have a partial fault where the mem access is not complete yet 561 // then the cache must have been blocked. This load will be re-executed 562 // when the cache gets unblocked. We will handle the fault when the 563 // mem access is complete. 564 return NoFault; 565 } 566 567 // If the instruction faulted or predicated false, then we need to send it 568 // along to commit without the instruction completing. 569 if (load_fault != NoFault || !inst->readPredicate()) { 570 // Send this instruction to commit, also make sure iew stage 571 // realizes there is activity. Mark it as executed unless it 572 // is a strictly ordered load that needs to hit the head of 573 // commit. 574 if (!inst->readPredicate()) 575 inst->forwardOldRegs(); 576 DPRINTF(LSQUnit, "Load [sn:%lli] not executed from %s\n", 577 inst->seqNum, 578 (load_fault != NoFault ? "fault" : "predication")); 579 if (!(inst->hasRequest() && inst->strictlyOrdered()) || 580 inst->isAtCommit()) { 581 inst->setExecuted(); 582 } 583 iewStage->instToCommit(inst); 584 iewStage->activityThisCycle(); 585 } else { 586 if (inst->effAddrValid()) { 587 auto it = inst->lqIt; 588 ++it; 589 590 if (checkLoads) 591 return checkViolations(it, inst); 592 } 593 } 594 595 return load_fault; 596} 597 598template <class Impl> 599Fault 600LSQUnit<Impl>::executeStore(const DynInstPtr &store_inst) 601{ 602 using namespace TheISA; 603 // Make sure that a store exists. 604 assert(stores != 0); 605 606 int store_idx = store_inst->sqIdx; 607 608 DPRINTF(LSQUnit, "Executing store PC %s [sn:%lli]\n", 609 store_inst->pcState(), store_inst->seqNum); 610 611 assert(!store_inst->isSquashed()); 612 613 // Check the recently completed loads to see if any match this store's 614 // address. If so, then we have a memory ordering violation. 615 typename LoadQueue::iterator loadIt = store_inst->lqIt; 616 617 Fault store_fault = store_inst->initiateAcc(); 618 619 if (store_inst->isTranslationDelayed() && 620 store_fault == NoFault) 621 return store_fault; 622 623 if (!store_inst->readPredicate()) { 624 DPRINTF(LSQUnit, "Store [sn:%lli] not executed from predication\n", 625 store_inst->seqNum); 626 store_inst->forwardOldRegs(); 627 return store_fault; 628 } 629 630 if (storeQueue[store_idx].size() == 0) { 631 DPRINTF(LSQUnit,"Fault on Store PC %s, [sn:%lli], Size = 0\n", 632 store_inst->pcState(), store_inst->seqNum); 633 634 return store_fault; 635 } 636 637 assert(store_fault == NoFault); 638 639 if (store_inst->isStoreConditional() || store_inst->isAtomic()) { 640 // Store conditionals and Atomics need to set themselves as able to 641 // writeback if we haven't had a fault by here. 642 storeQueue[store_idx].canWB() = true; 643 644 ++storesToWB; 645 } 646 647 return checkViolations(loadIt, store_inst); 648 649} 650 651template <class Impl> 652void 653LSQUnit<Impl>::commitLoad() 654{ 655 assert(loadQueue.front().valid()); 656 657 DPRINTF(LSQUnit, "Committing head load instruction, PC %s\n", 658 loadQueue.front().instruction()->pcState()); 659 660 loadQueue.front().clear(); 661 loadQueue.pop_front(); 662 663 --loads; 664} 665 666template <class Impl> 667void 668LSQUnit<Impl>::commitLoads(InstSeqNum &youngest_inst) 669{ 670 assert(loads == 0 || loadQueue.front().valid()); 671 672 while (loads != 0 && loadQueue.front().instruction()->seqNum 673 <= youngest_inst) { 674 commitLoad(); 675 } 676} 677 678template <class Impl> 679void 680LSQUnit<Impl>::commitStores(InstSeqNum &youngest_inst) 681{ 682 assert(stores == 0 || storeQueue.front().valid()); 683 684 /* Forward iterate the store queue (age order). */ 685 for (auto& x : storeQueue) { 686 assert(x.valid()); 687 // Mark any stores that are now committed and have not yet 688 // been marked as able to write back. 689 if (!x.canWB()) { 690 if (x.instruction()->seqNum > youngest_inst) { 691 break; 692 } 693 DPRINTF(LSQUnit, "Marking store as able to write back, PC " 694 "%s [sn:%lli]\n", 695 x.instruction()->pcState(), 696 x.instruction()->seqNum); 697 698 x.canWB() = true; 699 700 ++storesToWB; 701 } 702 } 703} 704 705template <class Impl> 706void 707LSQUnit<Impl>::writebackBlockedStore() 708{ 709 assert(isStoreBlocked); 710 storeWBIt->request()->sendPacketToCache(); 711 if (storeWBIt->request()->isSent()){ 712 storePostSend(); 713 } 714} 715 716template <class Impl> 717void 718LSQUnit<Impl>::writebackStores() 719{ 720 if (isStoreBlocked) { 721 DPRINTF(LSQUnit, "Writing back blocked store\n"); 722 writebackBlockedStore(); 723 } 724 725 while (storesToWB > 0 && 726 storeWBIt.dereferenceable() && 727 storeWBIt->valid() && 728 storeWBIt->canWB() && 729 ((!needsTSO) || (!storeInFlight)) && 730 lsq->cachePortAvailable(false)) { 731 732 if (isStoreBlocked) { 733 DPRINTF(LSQUnit, "Unable to write back any more stores, cache" 734 " is blocked!\n"); 735 break; 736 } 737 738 // Store didn't write any data so no need to write it back to 739 // memory. 740 if (storeWBIt->size() == 0) { 741 /* It is important that the preincrement happens at (or before) 742 * the call, as the the code of completeStore checks 743 * storeWBIt. */ 744 completeStore(storeWBIt++); 745 continue; 746 } 747 748 if (storeWBIt->instruction()->isDataPrefetch()) { 749 storeWBIt++; 750 continue; 751 } 752 753 assert(storeWBIt->hasRequest()); 754 assert(!storeWBIt->committed()); 755 756 DynInstPtr inst = storeWBIt->instruction(); 757 LSQRequest* req = storeWBIt->request(); 758 storeWBIt->committed() = true; 759 760 assert(!inst->memData); 761 inst->memData = new uint8_t[req->_size]; 762 763 if (storeWBIt->isAllZeros()) 764 memset(inst->memData, 0, req->_size); 765 else 766 memcpy(inst->memData, storeWBIt->data(), req->_size); 767 768 769 if (req->senderState() == nullptr) { 770 SQSenderState *state = new SQSenderState(storeWBIt); 771 state->isLoad = false; 772 state->needWB = false; 773 state->inst = inst; 774 775 req->senderState(state); 776 if (inst->isStoreConditional() || inst->isAtomic()) { 777 /* Only store conditionals and atomics need a writeback. */ 778 state->needWB = true; 779 } 780 } 781 req->buildPackets(); 782 783 DPRINTF(LSQUnit, "D-Cache: Writing back store idx:%i PC:%s " 784 "to Addr:%#x, data:%#x [sn:%lli]\n", 785 storeWBIt.idx(), inst->pcState(), 786 req->request()->getPaddr(), (int)*(inst->memData), 787 inst->seqNum); 788 789 // @todo: Remove this SC hack once the memory system handles it. 790 if (inst->isStoreConditional()) { 791 // Disable recording the result temporarily. Writing to 792 // misc regs normally updates the result, but this is not 793 // the desired behavior when handling store conditionals. 794 inst->recordResult(false); 795 bool success = TheISA::handleLockedWrite(inst.get(), 796 req->request(), cacheBlockMask); 797 inst->recordResult(true); 798 req->packetSent(); 799 800 if (!success) { 801 req->complete(); 802 // Instantly complete this store. 803 DPRINTF(LSQUnit, "Store conditional [sn:%lli] failed. " 804 "Instantly completing it.\n", 805 inst->seqNum); 806 PacketPtr new_pkt = new Packet(*req->packet()); 807 WritebackEvent *wb = new WritebackEvent(inst, 808 new_pkt, this); 809 cpu->schedule(wb, curTick() + 1); 810 completeStore(storeWBIt); 811 if (!storeQueue.empty()) 812 storeWBIt++; 813 else 814 storeWBIt = storeQueue.end(); 815 continue; 816 } 817 } 818 819 if (req->request()->isMmappedIpr()) { 820 assert(!inst->isStoreConditional()); 821 ThreadContext *thread = cpu->tcBase(lsqID); 822 PacketPtr main_pkt = new Packet(req->mainRequest(), 823 MemCmd::WriteReq); 824 main_pkt->dataStatic(inst->memData); 825 req->handleIprWrite(thread, main_pkt); 826 delete main_pkt; 827 completeStore(storeWBIt); 828 storeWBIt++; 829 continue; 830 } 831 /* Send to cache */ 832 req->sendPacketToCache(); 833 834 /* If successful, do the post send */ 835 if (req->isSent()) { 836 storePostSend(); 837 } else { 838 DPRINTF(LSQUnit, "D-Cache became blocked when writing [sn:%lli], " 839 "will retry later\n", 840 inst->seqNum); 841 } 842 } 843 assert(stores >= 0 && storesToWB >= 0); 844} 845 846template <class Impl> 847void 848LSQUnit<Impl>::squash(const InstSeqNum &squashed_num) 849{ 850 DPRINTF(LSQUnit, "Squashing until [sn:%lli]!" 851 "(Loads:%i Stores:%i)\n", squashed_num, loads, stores); 852 853 while (loads != 0 && 854 loadQueue.back().instruction()->seqNum > squashed_num) { 855 DPRINTF(LSQUnit,"Load Instruction PC %s squashed, " 856 "[sn:%lli]\n", 857 loadQueue.back().instruction()->pcState(), 858 loadQueue.back().instruction()->seqNum); 859 860 if (isStalled() && loadQueue.tail() == stallingLoadIdx) { 861 stalled = false; 862 stallingStoreIsn = 0; 863 stallingLoadIdx = 0; 864 } 865 866 // Clear the smart pointer to make sure it is decremented. 867 loadQueue.back().instruction()->setSquashed(); 868 loadQueue.back().clear(); 869 870 --loads; 871 872 loadQueue.pop_back(); 873 ++lsqSquashedLoads; 874 } 875 876 if (memDepViolator && squashed_num < memDepViolator->seqNum) { 877 memDepViolator = NULL; 878 } 879 880 while (stores != 0 && 881 storeQueue.back().instruction()->seqNum > squashed_num) { 882 // Instructions marked as can WB are already committed. 883 if (storeQueue.back().canWB()) { 884 break; 885 } 886 887 DPRINTF(LSQUnit,"Store Instruction PC %s squashed, " 888 "idx:%i [sn:%lli]\n", 889 storeQueue.back().instruction()->pcState(), 890 storeQueue.tail(), storeQueue.back().instruction()->seqNum); 891 892 // I don't think this can happen. It should have been cleared 893 // by the stalling load. 894 if (isStalled() && 895 storeQueue.back().instruction()->seqNum == stallingStoreIsn) { 896 panic("Is stalled should have been cleared by stalling load!\n"); 897 stalled = false; 898 stallingStoreIsn = 0; 899 } 900 901 // Clear the smart pointer to make sure it is decremented. 902 storeQueue.back().instruction()->setSquashed(); 903 904 // Must delete request now that it wasn't handed off to 905 // memory. This is quite ugly. @todo: Figure out the proper 906 // place to really handle request deletes. 907 storeQueue.back().clear(); 908 --stores; 909 910 storeQueue.pop_back(); 911 ++lsqSquashedStores; 912 } 913} 914 915template <class Impl> 916void 917LSQUnit<Impl>::storePostSend() 918{ 919 if (isStalled() && 920 storeWBIt->instruction()->seqNum == stallingStoreIsn) { 921 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] " 922 "load idx:%i\n", 923 stallingStoreIsn, stallingLoadIdx); 924 stalled = false; 925 stallingStoreIsn = 0; 926 iewStage->replayMemInst(loadQueue[stallingLoadIdx].instruction()); 927 } 928 929 if (!storeWBIt->instruction()->isStoreConditional()) { 930 // The store is basically completed at this time. This 931 // only works so long as the checker doesn't try to 932 // verify the value in memory for stores. 933 storeWBIt->instruction()->setCompleted(); 934 935 if (cpu->checker) { 936 cpu->checker->verify(storeWBIt->instruction()); 937 } 938 } 939 940 if (needsTSO) { 941 storeInFlight = true; 942 } 943 944 storeWBIt++; 945} 946 947template <class Impl> 948void 949LSQUnit<Impl>::writeback(const DynInstPtr &inst, PacketPtr pkt) 950{ 951 iewStage->wakeCPU(); 952 953 // Squashed instructions do not need to complete their access. 954 if (inst->isSquashed()) { 955 assert(!inst->isStore()); 956 ++lsqIgnoredResponses; 957 return; 958 } 959 960 if (!inst->isExecuted()) { 961 inst->setExecuted(); 962 963 if (inst->fault == NoFault) { 964 // Complete access to copy data to proper place. 965 inst->completeAcc(pkt); 966 } else { 967 // If the instruction has an outstanding fault, we cannot complete 968 // the access as this discards the current fault. 969 970 // If we have an outstanding fault, the fault should only be of 971 // type ReExec or - in case of a SplitRequest - a partial 972 // translation fault 973 assert(dynamic_cast<ReExec*>(inst->fault.get()) != nullptr || 974 inst->savedReq->isPartialFault()); 975 976 DPRINTF(LSQUnit, "Not completing instruction [sn:%lli] access " 977 "due to pending fault.\n", inst->seqNum); 978 } 979 } 980 981 // Need to insert instruction into queue to commit 982 iewStage->instToCommit(inst); 983 984 iewStage->activityThisCycle(); 985 986 // see if this load changed the PC 987 iewStage->checkMisprediction(inst); 988} 989 990template <class Impl> 991void 992LSQUnit<Impl>::completeStore(typename StoreQueue::iterator store_idx) 993{ 994 assert(store_idx->valid()); 995 store_idx->completed() = true; 996 --storesToWB; 997 // A bit conservative because a store completion may not free up entries, 998 // but hopefully avoids two store completions in one cycle from making 999 // the CPU tick twice. 1000 cpu->wakeCPU(); 1001 cpu->activityThisCycle(); 1002 1003 /* We 'need' a copy here because we may clear the entry from the 1004 * store queue. */ 1005 DynInstPtr store_inst = store_idx->instruction(); 1006 if (store_idx == storeQueue.begin()) { 1007 do { 1008 storeQueue.front().clear(); 1009 storeQueue.pop_front(); 1010 --stores; 1011 } while (storeQueue.front().completed() && 1012 !storeQueue.empty()); 1013 1014 iewStage->updateLSQNextCycle = true; 1015 } 1016 1017 DPRINTF(LSQUnit, "Completing store [sn:%lli], idx:%i, store head " 1018 "idx:%i\n", 1019 store_inst->seqNum, store_idx.idx() - 1, storeQueue.head() - 1); 1020 1021#if TRACING_ON 1022 if (DTRACE(O3PipeView)) { 1023 store_inst->storeTick = 1024 curTick() - store_inst->fetchTick; 1025 } 1026#endif 1027 1028 if (isStalled() && 1029 store_inst->seqNum == stallingStoreIsn) { 1030 DPRINTF(LSQUnit, "Unstalling, stalling store [sn:%lli] " 1031 "load idx:%i\n", 1032 stallingStoreIsn, stallingLoadIdx); 1033 stalled = false; 1034 stallingStoreIsn = 0; 1035 iewStage->replayMemInst(loadQueue[stallingLoadIdx].instruction()); 1036 } 1037 1038 store_inst->setCompleted(); 1039 1040 if (needsTSO) { 1041 storeInFlight = false; 1042 } 1043 1044 // Tell the checker we've completed this instruction. Some stores 1045 // may get reported twice to the checker, but the checker can 1046 // handle that case. 1047 // Store conditionals cannot be sent to the checker yet, they have 1048 // to update the misc registers first which should take place 1049 // when they commit 1050 if (cpu->checker && !store_inst->isStoreConditional()) { 1051 cpu->checker->verify(store_inst); 1052 } 1053} 1054 1055template <class Impl> 1056bool 1057LSQUnit<Impl>::trySendPacket(bool isLoad, PacketPtr data_pkt) 1058{ 1059 bool ret = true; 1060 bool cache_got_blocked = false; 1061 1062 auto state = dynamic_cast<LSQSenderState*>(data_pkt->senderState); 1063 1064 if (!lsq->cacheBlocked() && 1065 lsq->cachePortAvailable(isLoad)) { 1066 if (!dcachePort->sendTimingReq(data_pkt)) { 1067 ret = false; 1068 cache_got_blocked = true; 1069 } 1070 } else { 1071 ret = false; 1072 } 1073 1074 if (ret) { 1075 if (!isLoad) { 1076 isStoreBlocked = false; 1077 } 1078 lsq->cachePortBusy(isLoad); 1079 state->outstanding++; 1080 state->request()->packetSent(); 1081 } else { 1082 if (cache_got_blocked) { 1083 lsq->cacheBlocked(true); 1084 ++lsqCacheBlocked; 1085 } 1086 if (!isLoad) { 1087 assert(state->request() == storeWBIt->request()); 1088 isStoreBlocked = true; 1089 } 1090 state->request()->packetNotSent(); 1091 } 1092 return ret; 1093} 1094 1095template <class Impl> 1096void 1097LSQUnit<Impl>::recvRetry() 1098{ 1099 if (isStoreBlocked) { 1100 DPRINTF(LSQUnit, "Receiving retry: blocked store\n"); 1101 writebackBlockedStore(); 1102 } 1103} 1104 1105template <class Impl> 1106void 1107LSQUnit<Impl>::dumpInsts() const 1108{ 1109 cprintf("Load store queue: Dumping instructions.\n"); 1110 cprintf("Load queue size: %i\n", loads); 1111 cprintf("Load queue: "); 1112 1113 for (const auto& e: loadQueue) { 1114 const DynInstPtr &inst(e.instruction()); 1115 cprintf("%s.[sn:%llu] ", inst->pcState(), inst->seqNum); 1116 } 1117 cprintf("\n"); 1118 1119 cprintf("Store queue size: %i\n", stores); 1120 cprintf("Store queue: "); 1121 1122 for (const auto& e: storeQueue) { 1123 const DynInstPtr &inst(e.instruction()); 1124 cprintf("%s.[sn:%llu] ", inst->pcState(), inst->seqNum); 1125 } 1126 1127 cprintf("\n"); 1128} 1129 1130template <class Impl> 1131unsigned int 1132LSQUnit<Impl>::cacheLineSize() 1133{ 1134 return cpu->cacheLineSize(); 1135} 1136 1137#endif//__CPU_O3_LSQ_UNIT_IMPL_HH__ 1138